2024-09-22 10:50:57,250 INFO [train.py:1266] (2/4) Training started 2024-09-22 10:50:57,250 INFO [train.py:1276] (2/4) Device: cuda:2 2024-09-22 10:50:57,253 INFO [train.py:1307] (2/4) Using dtype=torch.float16 2024-09-22 10:50:57,253 INFO [train.py:1308] (2/4) Use AMP=True 2024-09-22 10:50:57,253 INFO [train.py:1310] (2/4) {'best_train_loss': inf, 'best_valid_loss': inf, 'best_train_epoch': -1, 'best_valid_epoch': -1, 'batch_idx_train': 0, 'log_interval': 50, 'reset_interval': 200, 'valid_interval': 3000, 'feature_dim': 80, 'subsampling_factor': 4, 'ignore_id': -1, 'label_smoothing': 0.1, 'warm_step': 2000, 'env_info': {'k2-version': '1.24.4', 'k2-build-type': 'Release', 'k2-with-cuda': True, 'k2-git-sha1': '44a9d5682af9fd3ef77074777e15278ec6d390eb', 'k2-git-date': 'Wed Sep 27 11:22:55 2023', 'lhotse-version': '1.17.0.dev+git.ccfc5b2c.dirty', 'torch-version': '1.10.0+cu102', 'torch-cuda-available': True, 'torch-cuda-version': '10.2', 'python-version': '3.8', 'icefall-git-branch': 'cr-ctc', 'icefall-git-sha1': 'a6eead6c-clean', 'icefall-git-date': 'Mon Sep 9 10:10:08 2024', 'icefall-path': '/star-zw/workspace/zipformer/icefall_cr_ctc', 'k2-path': '/star-zw/workspace/k2/k2/k2/python/k2/__init__.py', 'lhotse-path': '/star-zw/workspace/lhotse/lhotse/lhotse/__init__.py', 'hostname': 'de-74279-k2-train-7-0905180047-6d6678bc6f-8cwvw', 'IP address': '10.30.5.48'}, 'world_size': 4, 'master_port': 12347, 'tensorboard': True, 'num_epochs': 50, 'start_epoch': 1, 'start_batch': 0, 'exp_dir': PosixPath('zipformer/exp-cr-loss-scale-0.2-time-mask-ratio-2.5-scaled-masked-1-4'), 'bpe_model': 'data/lang_bpe_500/bpe.model', 'base_lr': 0.045, 'lr_batches': 7500, 'lr_epochs': 3.5, 'ref_duration': 600, 'context_size': 2, 'prune_range': 5, 'lm_scale': 0.25, 'am_scale': 0.0, 'simple_loss_scale': 0.5, 'ctc_loss_scale': 1.0, 'cr_loss_scale': 0.2, 'time_mask_ratio': 2.5, 'cr_loss_masked_scale': 1.0, 'attention_decoder_loss_scale': 0.8, 'seed': 42, 'print_diagnostics': False, 'inf_check': False, 'save_every_n': 4000, 'keep_last_k': 30, 'average_period': 200, 'use_fp16': True, 'use_bf16': False, 'num_encoder_layers': '2,2,3,4,3,2', 'downsampling_factor': '1,2,4,8,4,2', 'feedforward_dim': '512,768,1024,1536,1024,768', 'num_heads': '4,4,4,8,4,4', 'encoder_dim': '192,256,384,512,384,256', 'query_head_dim': '32', 'value_head_dim': '12', 'pos_head_dim': '4', 'pos_dim': 48, 'encoder_unmasked_dim': '192,192,256,256,256,192', 'cnn_module_kernel': '31,31,15,15,15,31', 'decoder_dim': 512, 'joiner_dim': 512, 'attention_decoder_dim': 512, 'attention_decoder_num_layers': 6, 'attention_decoder_attention_dim': 512, 'attention_decoder_num_heads': 8, 'attention_decoder_feedforward_dim': 2048, 'causal': False, 'chunk_size': '16,32,64,-1', 'left_context_frames': '64,128,256,-1', 'use_transducer': False, 'use_ctc': True, 'use_attention_decoder': False, 'use_cr_ctc': True, 'full_libri': True, 'mini_libri': False, 'manifest_dir': PosixPath('data/fbank'), 'max_duration': 700, 'bucketing_sampler': True, 'num_buckets': 30, 'concatenate_cuts': False, 'duration_factor': 1.0, 'gap': 1.0, 'on_the_fly_feats': False, 'shuffle': True, 'drop_last': True, 'return_cuts': True, 'num_workers': 2, 'enable_spec_aug': False, 'spec_aug_time_warp_factor': 80, 'enable_musan': True, 'input_strategy': 'PrecomputedFeatures', 'blank_id': 0, 'sos_id': 1, 'eos_id': 1, 'vocab_size': 500, 'dtype': torch.float16, 'use_autocast': True} 2024-09-22 10:50:57,253 INFO [train.py:1312] (2/4) About to create model 2024-09-22 10:50:57,901 INFO [train.py:1316] (2/4) Number of model parameters: 64250603 2024-09-22 10:50:57,902 INFO [train.py:752] (2/4) num_frame_masks: 25, max_frames_mask_fraction: 0.375 2024-09-22 10:51:02,826 INFO [train.py:1338] (2/4) Using DDP 2024-09-22 10:51:03,358 INFO [asr_datamodule.py:436] (2/4) About to get the shuffled train-clean-100, train-clean-360 and train-other-500 cuts 2024-09-22 10:51:03,601 INFO [asr_datamodule.py:232] (2/4) Enable MUSAN 2024-09-22 10:51:03,601 INFO [asr_datamodule.py:233] (2/4) About to get Musan cuts 2024-09-22 10:51:05,254 INFO [asr_datamodule.py:279] (2/4) Disable SpecAugment 2024-09-22 10:51:05,254 INFO [asr_datamodule.py:281] (2/4) About to create train dataset 2024-09-22 10:51:05,254 INFO [asr_datamodule.py:308] (2/4) Using DynamicBucketingSampler. 2024-09-22 10:51:28,215 INFO [asr_datamodule.py:325] (2/4) About to create train dataloader 2024-09-22 10:51:28,216 INFO [asr_datamodule.py:453] (2/4) About to get dev-clean cuts 2024-09-22 10:51:28,217 INFO [asr_datamodule.py:460] (2/4) About to get dev-other cuts 2024-09-22 10:51:28,218 INFO [asr_datamodule.py:356] (2/4) About to create dev dataset 2024-09-22 10:51:28,415 INFO [asr_datamodule.py:373] (2/4) About to create dev dataloader 2024-09-22 10:51:28,415 INFO [train.py:1545] (2/4) Sanity check -- see if any of the batches in epoch 1 would cause OOM. 2024-09-22 10:55:04,879 INFO [train.py:1576] (2/4) Maximum memory allocated so far is 18627MB 2024-09-22 10:55:06,893 INFO [train.py:1576] (2/4) Maximum memory allocated so far is 18627MB 2024-09-22 10:55:09,107 INFO [train.py:1576] (2/4) Maximum memory allocated so far is 18985MB 2024-09-22 10:55:10,930 INFO [train.py:1576] (2/4) Maximum memory allocated so far is 18985MB 2024-09-22 10:55:13,013 INFO [train.py:1576] (2/4) Maximum memory allocated so far is 18985MB 2024-09-22 10:55:15,403 INFO [train.py:1576] (2/4) Maximum memory allocated so far is 18985MB 2024-09-22 10:56:01,485 INFO [train.py:1198] (2/4) Epoch 1, batch 0, loss[loss=4.933, ctc_loss=4.796, cr_loss=0.6836, over 16935.00 frames. ], tot_loss[loss=4.933, ctc_loss=4.796, cr_loss=0.6836, over 16935.00 frames. ], batch size: 42, lr: 2.25e-02, grad_scale: 2.0 2024-09-22 10:56:01,485 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-22 10:56:18,232 INFO [train.py:1230] (2/4) Epoch 1, validation: loss=4.756, ctc_loss=4.756, cr_loss=2.853e-15, over 944034.00 frames. 2024-09-22 10:56:18,233 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 18985MB 2024-09-22 10:56:21,884 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=0.0, ans=0.3 2024-09-22 10:56:22,067 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=0.0, ans=0.1 2024-09-22 10:56:24,780 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.90 vs. limit=5.0 2024-09-22 10:56:32,851 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=0.0, ans=0.3 2024-09-22 10:56:38,254 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.min_positive, batch_count=46.666666666666664, ans=0.24953333333333333 2024-09-22 10:56:39,773 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 2.680e+03 4.888e+03 5.131e+03 6.578e+03 8.849e+03, threshold=2.053e+04, percent-clipped=0.0 2024-09-22 10:56:57,405 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=8.01 vs. limit=4.037333333333334 2024-09-22 10:57:00,471 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.227e+03 2.818e+03 4.846e+03 6.578e+03 1.124e+04, threshold=1.938e+04, percent-clipped=0.0 2024-09-22 10:57:10,397 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=16.74 vs. limit=5.046666666666667 2024-09-22 10:57:25,082 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=6.39 vs. limit=4.056 2024-09-22 10:57:36,832 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 8.493e+02 2.434e+03 3.497e+03 4.961e+03 1.124e+04, threshold=1.399e+04, percent-clipped=0.0 2024-09-22 10:57:37,233 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=186.66666666666666, ans=0.8934666666666667 2024-09-22 10:57:40,031 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=27.93 vs. limit=7.64 2024-09-22 10:57:41,383 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=10.49 vs. limit=4.074666666666666 2024-09-22 10:57:48,197 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.min_positive, batch_count=186.66666666666666, ans=0.09883333333333334 2024-09-22 10:57:51,513 INFO [train.py:1198] (2/4) Epoch 1, batch 50, loss[loss=1.44, ctc_loss=1.369, cr_loss=0.3533, over 17095.00 frames. ], tot_loss[loss=2.3, ctc_loss=2.238, cr_loss=0.31, over 762188.47 frames. ], batch size: 49, lr: 2.48e-02, grad_scale: 0.5 2024-09-22 10:57:52,372 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=36.29 vs. limit=7.5875 2024-09-22 10:57:53,727 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-22 10:58:12,768 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten.whitening_limit, batch_count=280.0, ans=7.605 2024-09-22 10:58:20,345 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=25.18 vs. limit=7.71 2024-09-22 10:58:22,286 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=124.10 vs. limit=7.605 2024-09-22 10:58:23,886 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten.whitening_limit, batch_count=280.0, ans=7.605 2024-09-22 10:58:27,653 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=24.26 vs. limit=7.6225 2024-09-22 10:58:27,932 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=69.11 vs. limit=7.6225 2024-09-22 10:58:29,340 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=16.75 vs. limit=7.745 2024-09-22 10:58:30,158 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=79.09 vs. limit=7.6225 2024-09-22 10:58:42,004 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=326.6666666666667, ans=0.4846875 2024-09-22 10:58:43,739 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=326.6666666666667, ans=0.4846875 2024-09-22 10:58:43,913 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.min_positive, batch_count=326.6666666666667, ans=0.04897916666666667 2024-09-22 10:59:07,022 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn1.whiten, num_groups=1, num_channels=192, metric=47.55 vs. limit=7.815 2024-09-22 10:59:09,884 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=420.0, ans=0.226375 2024-09-22 10:59:24,437 INFO [train.py:1198] (2/4) Epoch 1, batch 100, loss[loss=1.119, ctc_loss=1.091, cr_loss=0.1379, over 17257.00 frames. ], tot_loss[loss=1.721, ctc_loss=1.672, cr_loss=0.2442, over 1341204.75 frames. ], batch size: 42, lr: 2.70e-02, grad_scale: 1.0 2024-09-22 10:59:28,076 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 2.417e+02 5.797e+02 1.227e+03 2.964e+03 1.124e+04, threshold=2.454e+03, percent-clipped=0.0 2024-09-22 10:59:38,059 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=22.60 vs. limit=7.675 2024-09-22 10:59:42,299 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=171.02 vs. limit=7.6925 2024-09-22 10:59:53,503 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=7.32 vs. limit=4.205333333333333 2024-09-22 11:00:00,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=513.3333333333334, ans=0.4759375 2024-09-22 11:00:37,246 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=6.00 vs. limit=4.242666666666667 2024-09-22 11:00:53,076 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=238.12 vs. limit=5.326666666666667 2024-09-22 11:00:54,295 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=18.89 vs. limit=5.326666666666667 2024-09-22 11:00:58,863 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=121.26 vs. limit=5.326666666666667 2024-09-22 11:01:03,163 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=700.0, ans=0.5 2024-09-22 11:01:04,606 INFO [train.py:1198] (2/4) Epoch 1, batch 150, loss[loss=1.271, ctc_loss=1.248, cr_loss=0.1147, over 17017.00 frames. ], tot_loss[loss=1.511, ctc_loss=1.472, cr_loss=0.1957, over 1787767.03 frames. ], batch size: 53, lr: 2.93e-02, grad_scale: 1.0 2024-09-22 11:01:16,680 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=7.82 vs. limit=4.28 2024-09-22 11:01:27,117 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=746.6666666666666, ans=0.46499999999999997 2024-09-22 11:01:29,075 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=68.06 vs. limit=5.373333333333333 2024-09-22 11:01:32,510 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=1.179e-01 2024-09-22 11:01:34,709 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=78.48 vs. limit=7.78 2024-09-22 11:01:38,762 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=18.22 vs. limit=8.06 2024-09-22 11:01:43,550 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=793.3333333333334, ans=0.29206666666666664 2024-09-22 11:01:49,697 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=10.26 vs. limit=4.317333333333333 2024-09-22 11:01:53,844 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=90.42 vs. limit=7.7975 2024-09-22 11:01:55,211 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=49.19 vs. limit=7.7975 2024-09-22 11:01:58,471 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.min_positive, batch_count=840.0, ans=0.09475 2024-09-22 11:01:58,761 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=96.88 vs. limit=7.815 2024-09-22 11:01:59,057 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=34.74 vs. limit=7.815 2024-09-22 11:02:02,919 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=16.02 vs. limit=7.815 2024-09-22 11:02:04,382 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module1.whiten, num_groups=1, num_channels=512, metric=25.13 vs. limit=7.815 2024-09-22 11:02:05,724 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=840.0, ans=0.460625 2024-09-22 11:02:09,584 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=840.0, ans=0.460625 2024-09-22 11:02:15,347 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=6.46 vs. limit=4.336 2024-09-22 11:02:39,844 INFO [train.py:1198] (2/4) Epoch 1, batch 200, loss[loss=1.27, ctc_loss=1.249, cr_loss=0.1043, over 14946.00 frames. ], tot_loss[loss=1.399, ctc_loss=1.366, cr_loss=0.1682, over 2136727.51 frames. ], batch size: 89, lr: 3.15e-02, grad_scale: 2.0 2024-09-22 11:02:42,700 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=8.93 vs. limit=7.85 2024-09-22 11:02:43,594 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.311e+02 2.245e+02 3.033e+02 4.033e+02 1.104e+03, threshold=6.066e+02, percent-clipped=0.0 2024-09-22 11:02:44,640 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=8.69 vs. limit=7.85 2024-09-22 11:02:53,605 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=19.37 vs. limit=8.2 2024-09-22 11:03:00,467 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=980.0, ans=0.2147 2024-09-22 11:03:08,223 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=60.62 vs. limit=7.8675 2024-09-22 11:03:10,121 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=50.16 vs. limit=7.8675 2024-09-22 11:03:14,339 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=51.62 vs. limit=7.8675 2024-09-22 11:03:17,117 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=1026.6666666666667, ans=5.641666666666667 2024-09-22 11:03:21,033 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten.whitening_limit, batch_count=1026.6666666666667, ans=7.885 2024-09-22 11:03:30,527 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=57.37 vs. limit=7.885 2024-09-22 11:03:37,516 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=1073.3333333333333, ans=0.28926666666666667 2024-09-22 11:03:38,129 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=120.12 vs. limit=7.9025 2024-09-22 11:03:41,234 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=1073.3333333333333, ans=0.28926666666666667 2024-09-22 11:03:42,956 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=1073.3333333333333, ans=0.4496875 2024-09-22 11:03:51,056 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=512, metric=16.65 vs. limit=7.9025 2024-09-22 11:03:56,826 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=16.33 vs. limit=7.92 2024-09-22 11:04:03,823 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=19.18 vs. limit=8.34 2024-09-22 11:04:05,268 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=1120.0, ans=0.4475 2024-09-22 11:04:07,007 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=1120.0, ans=0.4475 2024-09-22 11:04:11,440 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=27.74 vs. limit=7.9375 2024-09-22 11:04:12,400 INFO [train.py:1198] (2/4) Epoch 1, batch 250, loss[loss=1.238, ctc_loss=1.213, cr_loss=0.1266, over 16751.00 frames. ], tot_loss[loss=1.337, ctc_loss=1.307, cr_loss=0.153, over 2403018.96 frames. ], batch size: 61, lr: 3.38e-02, grad_scale: 2.0 2024-09-22 11:04:38,211 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.min_positive, batch_count=1213.3333333333333, ans=0.23786666666666667 2024-09-22 11:04:46,625 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=21.49 vs. limit=8.41 2024-09-22 11:04:52,096 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=15.85 vs. limit=8.445 2024-09-22 11:04:53,639 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=42.86 vs. limit=8.445 2024-09-22 11:05:00,660 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=1260.0, ans=0.15275 2024-09-22 11:05:10,448 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=24.79 vs. limit=8.48 2024-09-22 11:05:10,604 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=80.37 vs. limit=7.99 2024-09-22 11:05:12,744 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=27.81 vs. limit=7.99 2024-09-22 11:05:14,373 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=67.65 vs. limit=5.653333333333333 2024-09-22 11:05:35,622 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=8.38 vs. limit=5.676666666666667 2024-09-22 11:05:40,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=1353.3333333333333, ans=0.06955 2024-09-22 11:05:47,708 INFO [train.py:1198] (2/4) Epoch 1, batch 300, loss[loss=1.151, ctc_loss=1.119, cr_loss=0.1616, over 17014.00 frames. ], tot_loss[loss=1.29, ctc_loss=1.26, cr_loss=0.1487, over 2615892.76 frames. ], batch size: 44, lr: 3.60e-02, grad_scale: 4.0 2024-09-22 11:05:48,561 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.10 vs. limit=4.5600000000000005 2024-09-22 11:05:50,492 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=15.33 vs. limit=8.025 2024-09-22 11:05:51,295 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.872e+02 2.621e+02 3.440e+02 6.626e+02, threshold=5.242e+02, percent-clipped=4.0 2024-09-22 11:05:58,787 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=1400.0, ans=0.851 2024-09-22 11:06:04,637 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=29.22 vs. limit=8.0425 2024-09-22 11:06:07,049 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=18.99 vs. limit=8.0425 2024-09-22 11:06:23,253 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=27.38 vs. limit=8.0425 2024-09-22 11:06:40,321 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=11.85 vs. limit=8.06 2024-09-22 11:06:43,118 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=1493.3333333333333, ans=0.22240000000000001 2024-09-22 11:06:43,119 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=1493.3333333333333, ans=0.43 2024-09-22 11:06:52,462 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=1540.0, ans=0.4278125 2024-09-22 11:06:55,244 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.99 vs. limit=5.0 2024-09-22 11:06:56,397 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=22.47 vs. limit=8.0775 2024-09-22 11:07:00,344 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=23.44 vs. limit=8.0775 2024-09-22 11:07:02,106 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.12 vs. limit=4.616 2024-09-22 11:07:25,015 INFO [train.py:1198] (2/4) Epoch 1, batch 350, loss[loss=1.202, ctc_loss=1.16, cr_loss=0.2083, over 17041.00 frames. ], tot_loss[loss=1.262, ctc_loss=1.23, cr_loss=0.1594, over 2774807.82 frames. ], batch size: 56, lr: 3.83e-02, grad_scale: 4.0 2024-09-22 11:07:34,713 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=1633.3333333333333, ans=0.4234375 2024-09-22 11:07:37,069 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=14.19 vs. limit=8.725 2024-09-22 11:07:50,493 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=1680.0, ans=0.035 2024-09-22 11:07:50,508 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=1680.0, ans=0.2832 2024-09-22 11:08:03,473 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=1726.6666666666667, ans=0.4190625 2024-09-22 11:08:17,090 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=9.74 vs. limit=8.1475 2024-09-22 11:08:20,410 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=14.90 vs. limit=8.1475 2024-09-22 11:08:27,509 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=18.93 vs. limit=5.886666666666667 2024-09-22 11:08:38,024 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.max_positive, batch_count=1773.3333333333333, ans=0.7677333333333334 2024-09-22 11:08:40,401 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.93 vs. limit=5.455 2024-09-22 11:08:49,049 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=1820.0, ans=0.2273 2024-09-22 11:09:00,230 INFO [train.py:1198] (2/4) Epoch 1, batch 400, loss[loss=1.183, ctc_loss=1.131, cr_loss=0.2601, over 16881.00 frames. ], tot_loss[loss=1.23, ctc_loss=1.194, cr_loss=0.1785, over 2897835.17 frames. ], batch size: 58, lr: 4.05e-02, grad_scale: 8.0 2024-09-22 11:09:00,589 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=1866.6666666666667, ans=0.13 2024-09-22 11:09:03,789 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.305e+02 2.539e+02 3.204e+02 4.584e+02 1.114e+03, threshold=6.407e+02, percent-clipped=17.0 2024-09-22 11:09:13,865 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=18.44 vs. limit=8.2 2024-09-22 11:09:24,431 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=1913.3333333333333, ans=0.41031249999999997 2024-09-22 11:09:26,855 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=5.33 vs. limit=5.4783333333333335 2024-09-22 11:09:39,801 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=15.51 vs. limit=8.235 2024-09-22 11:09:42,935 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten.whitening_limit, batch_count=1960.0, ans=8.97 2024-09-22 11:09:49,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=1960.0, ans=0.408125 2024-09-22 11:09:49,745 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=1960.0, ans=0.408125 2024-09-22 11:09:50,020 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=12.11 vs. limit=5.98 2024-09-22 11:09:58,954 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=2006.6666666666667, ans=0.4059375 2024-09-22 11:10:06,341 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=2006.6666666666667, ans=0.4059375 2024-09-22 11:10:06,928 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=11.48 vs. limit=9.004999999999999 2024-09-22 11:10:06,941 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=17.09 vs. limit=6.003333333333334 2024-09-22 11:10:31,077 INFO [train.py:1198] (2/4) Epoch 1, batch 450, loss[loss=1.135, ctc_loss=1.075, cr_loss=0.2993, over 17311.00 frames. ], tot_loss[loss=1.202, ctc_loss=1.161, cr_loss=0.2023, over 2998077.05 frames. ], batch size: 51, lr: 4.28e-02, grad_scale: 4.0 2024-09-22 11:10:32,297 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=11.76 vs. limit=9.075 2024-09-22 11:11:07,996 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=2193.3333333333335, ans=0.08629166666666667 2024-09-22 11:11:14,113 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.39 vs. limit=8.3225 2024-09-22 11:11:15,449 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=2193.3333333333335, ans=0.11775 2024-09-22 11:11:30,077 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=2240.0, ans=0.2776 2024-09-22 11:11:36,012 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=9.05 vs. limit=8.34 2024-09-22 11:11:56,162 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=10.54 vs. limit=8.3575 2024-09-22 11:11:56,728 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=9.80 vs. limit=8.3575 2024-09-22 11:11:57,586 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.min_positive, batch_count=2286.6666666666665, ans=0.04285416666666667 2024-09-22 11:12:00,368 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.32 vs. limit=9.215 2024-09-22 11:12:01,623 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=2286.6666666666665, ans=0.11425 2024-09-22 11:12:10,284 INFO [train.py:1198] (2/4) Epoch 1, batch 500, loss[loss=1.028, ctc_loss=0.9631, cr_loss=0.3248, over 17175.00 frames. ], tot_loss[loss=1.167, ctc_loss=1.121, cr_loss=0.2307, over 3075211.66 frames. ], batch size: 45, lr: 4.49e-02, grad_scale: 8.0 2024-09-22 11:12:15,062 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=9.13 vs. limit=8.375 2024-09-22 11:12:15,704 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.330e+02 2.678e+02 3.368e+02 4.496e+02 8.489e+02, threshold=6.737e+02, percent-clipped=3.0 2024-09-22 11:12:18,368 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=9.21 vs. limit=8.375 2024-09-22 11:12:19,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=2333.3333333333335, ans=0.390625 2024-09-22 11:12:27,133 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=9.40 vs. limit=9.285 2024-09-22 11:12:38,311 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=6.80 vs. limit=5.595 2024-09-22 11:12:39,271 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=2380.0, ans=0.3884375 2024-09-22 11:12:50,355 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=7.57 vs. limit=6.213333333333333 2024-09-22 11:12:52,018 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=2426.6666666666665, ans=0.38625 2024-09-22 11:13:04,840 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=2473.3333333333335, ans=0.1908333333333333 2024-09-22 11:13:05,110 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=6.12 vs. limit=5.618333333333333 2024-09-22 11:13:22,602 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=8.95 vs. limit=8.4275 2024-09-22 11:13:38,272 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=9.62 vs. limit=8.445 2024-09-22 11:13:39,951 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=4.74 vs. limit=5.008 2024-09-22 11:13:42,938 INFO [train.py:1198] (2/4) Epoch 1, batch 550, loss[loss=1.049, ctc_loss=0.9768, cr_loss=0.3622, over 16750.00 frames. ], tot_loss[loss=1.129, ctc_loss=1.077, cr_loss=0.2615, over 3146352.14 frames. ], batch size: 61, lr: 4.49e-02, grad_scale: 8.0 2024-09-22 11:13:49,162 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.75 vs. limit=5.641666666666667 2024-09-22 11:13:56,253 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=11.15 vs. limit=8.4625 2024-09-22 11:14:06,409 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=2613.3333333333335, ans=0.102 2024-09-22 11:14:10,414 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=7.26 vs. limit=6.306666666666667 2024-09-22 11:14:13,400 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=2613.3333333333335, ans=0.3775 2024-09-22 11:14:16,989 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=2660.0, ans=0.04015 2024-09-22 11:14:27,652 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=2660.0, ans=0.10024999999999999 2024-09-22 11:14:54,625 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten.whitening_limit, batch_count=2753.3333333333335, ans=8.5325 2024-09-22 11:14:57,485 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=2753.3333333333335, ans=0.8036333333333334 2024-09-22 11:15:01,655 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=10.52 vs. limit=9.565 2024-09-22 11:15:06,365 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=2753.3333333333335, ans=0.27246666666666663 2024-09-22 11:15:11,740 INFO [train.py:1198] (2/4) Epoch 1, batch 600, loss[loss=0.9332, ctc_loss=0.8506, cr_loss=0.4129, over 17302.00 frames. ], tot_loss[loss=1.089, ctc_loss=1.03, cr_loss=0.2922, over 3182852.97 frames. ], batch size: 51, lr: 4.49e-02, grad_scale: 8.0 2024-09-22 11:15:17,227 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.480e+02 2.420e+02 3.324e+02 4.180e+02 8.567e+02, threshold=6.647e+02, percent-clipped=1.0 2024-09-22 11:15:28,544 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=2846.6666666666665, ans=0.09325 2024-09-22 11:15:33,749 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=2846.6666666666665, ans=0.3665625 2024-09-22 11:15:46,495 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=5.04 vs. limit=5.157333333333334 2024-09-22 11:16:42,201 INFO [train.py:1198] (2/4) Epoch 1, batch 650, loss[loss=0.8287, ctc_loss=0.7529, cr_loss=0.3793, over 17246.00 frames. ], tot_loss[loss=1.044, ctc_loss=0.979, cr_loss=0.3229, over 3217539.90 frames. ], batch size: 44, lr: 4.49e-02, grad_scale: 8.0 2024-09-22 11:16:47,709 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=3033.3333333333335, ans=0.08624999999999998 2024-09-22 11:16:53,016 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=2.963e-01 2024-09-22 11:17:07,231 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=5.43 vs. limit=5.232 2024-09-22 11:17:26,434 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=3126.6666666666665, ans=0.2687333333333333 2024-09-22 11:17:40,513 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=3173.3333333333335, ans=0.35125 2024-09-22 11:17:54,358 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=3173.3333333333335, ans=0.10333333333333333 2024-09-22 11:18:13,631 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.min_abs, batch_count=3266.6666666666665, ans=0.249 2024-09-22 11:18:14,946 INFO [train.py:1198] (2/4) Epoch 1, batch 700, loss[loss=0.7151, ctc_loss=0.63, cr_loss=0.4255, over 17050.00 frames. ], tot_loss[loss=0.996, ctc_loss=0.9257, cr_loss=0.3517, over 3244779.29 frames. ], batch size: 39, lr: 4.49e-02, grad_scale: 8.0 2024-09-22 11:18:20,265 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.450e+02 2.357e+02 3.031e+02 4.358e+02 1.002e+03, threshold=6.062e+02, percent-clipped=7.0 2024-09-22 11:18:31,549 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.24 vs. limit=5.828333333333333 2024-09-22 11:19:21,597 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.29 vs. limit=5.362666666666667 2024-09-22 11:19:24,765 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.73 vs. limit=6.703333333333333 2024-09-22 11:19:36,542 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=8.90 vs. limit=8.795 2024-09-22 11:19:44,682 INFO [train.py:1198] (2/4) Epoch 1, batch 750, loss[loss=0.8027, ctc_loss=0.7091, cr_loss=0.4679, over 17302.00 frames. ], tot_loss[loss=0.9504, ctc_loss=0.8753, cr_loss=0.3755, over 3272149.97 frames. ], batch size: 49, lr: 4.49e-02, grad_scale: 8.0 2024-09-22 11:20:04,341 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=3546.6666666666665, ans=0.7758666666666667 2024-09-22 11:20:09,437 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=3546.6666666666665, ans=0.07783333333333334 2024-09-22 11:20:23,018 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=3593.3333333333335, ans=0.26406666666666667 2024-09-22 11:20:28,898 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=5.97 vs. limit=5.898333333333333 2024-09-22 11:20:45,323 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=3640.0, ans=0.329375 2024-09-22 11:20:57,457 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=3686.6666666666665, ans=0.32718749999999996 2024-09-22 11:21:13,169 INFO [train.py:1198] (2/4) Epoch 1, batch 800, loss[loss=0.6853, ctc_loss=0.6179, cr_loss=0.3374, over 17289.00 frames. ], tot_loss[loss=0.8995, ctc_loss=0.8214, cr_loss=0.3901, over 3294944.89 frames. ], batch size: 46, lr: 4.49e-02, grad_scale: 16.0 2024-09-22 11:21:18,256 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.759e+02 2.614e+02 4.071e+02 6.304e+02 1.473e+03, threshold=8.142e+02, percent-clipped=26.0 2024-09-22 11:21:19,324 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=11.16 vs. limit=10.3 2024-09-22 11:21:20,241 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=3733.3333333333335, ans=0.26266666666666666 2024-09-22 11:21:20,513 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=7.38 vs. limit=8.9 2024-09-22 11:21:23,905 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=3733.3333333333335, ans=0.325 2024-09-22 11:22:18,853 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=3873.3333333333335, ans=0.012849999999999986 2024-09-22 11:22:44,172 INFO [train.py:1198] (2/4) Epoch 1, batch 850, loss[loss=0.5951, ctc_loss=0.5193, cr_loss=0.3787, over 17022.00 frames. ], tot_loss[loss=0.8533, ctc_loss=0.7739, cr_loss=0.3966, over 3307235.39 frames. ], batch size: 44, lr: 4.49e-02, grad_scale: 16.0 2024-09-22 11:22:44,600 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=3966.6666666666665, ans=0.3140625 2024-09-22 11:22:50,238 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=4.99 vs. limit=8.9875 2024-09-22 11:23:03,426 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.72 vs. limit=10.51 2024-09-22 11:23:13,485 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=4013.3333333333335, ans=0.311875 2024-09-22 11:23:28,904 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=4060.0, ans=0.25939999999999996 2024-09-22 11:23:32,725 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=8.43 vs. limit=9.0225 2024-09-22 11:23:33,780 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=4106.666666666667, ans=0.009976811594202899 2024-09-22 11:23:45,647 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=4106.666666666667, ans=0.049555555555555554 2024-09-22 11:23:55,137 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=4153.333333333333, ans=0.3053125 2024-09-22 11:23:57,233 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=9.40 vs. limit=9.0575 2024-09-22 11:24:11,781 INFO [train.py:1198] (2/4) Epoch 1, batch 900, loss[loss=0.6081, ctc_loss=0.5216, cr_loss=0.4324, over 17021.00 frames. ], tot_loss[loss=0.8079, ctc_loss=0.7279, cr_loss=0.3999, over 3318514.08 frames. ], batch size: 44, lr: 4.48e-02, grad_scale: 16.0 2024-09-22 11:24:13,753 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=4200.0, ans=0.025 2024-09-22 11:24:16,914 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.753e+02 2.836e+02 3.761e+02 6.423e+02 1.326e+03, threshold=7.521e+02, percent-clipped=10.0 2024-09-22 11:24:55,458 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.00 vs. limit=9.11 2024-09-22 11:25:03,543 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=4340.0, ans=0.2965625 2024-09-22 11:25:12,705 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=8.52 vs. limit=9.1275 2024-09-22 11:25:13,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=4340.0, ans=0.0 2024-09-22 11:25:19,397 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=10.64 vs. limit=10.79 2024-09-22 11:25:22,450 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=4386.666666666667, ans=0.04838888888888889 2024-09-22 11:25:35,989 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=4433.333333333333, ans=0.2665 2024-09-22 11:25:37,367 INFO [train.py:1198] (2/4) Epoch 1, batch 950, loss[loss=0.6868, ctc_loss=0.5998, cr_loss=0.4352, over 17007.00 frames. ], tot_loss[loss=0.7674, ctc_loss=0.6864, cr_loss=0.4048, over 3330502.64 frames. ], batch size: 56, lr: 4.48e-02, grad_scale: 16.0 2024-09-22 11:25:42,822 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=4433.333333333333, ans=0.04819444444444445 2024-09-22 11:25:57,902 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=4480.0, ans=0.29000000000000004 2024-09-22 11:26:06,248 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.min_abs, batch_count=4480.0, ans=0.2672 2024-09-22 11:26:06,264 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=4480.0, ans=0.07200000000000001 2024-09-22 11:26:13,131 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=4526.666666666667, ans=0.7415666666666667 2024-09-22 11:26:31,532 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten.whitening_limit, batch_count=4573.333333333333, ans=9.215 2024-09-22 11:26:34,408 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.min_positive, batch_count=4573.333333333333, ans=0.07141666666666667 2024-09-22 11:27:05,076 INFO [train.py:1198] (2/4) Epoch 1, batch 1000, loss[loss=0.6182, ctc_loss=0.5302, cr_loss=0.4396, over 17199.00 frames. ], tot_loss[loss=0.7293, ctc_loss=0.6477, cr_loss=0.4079, over 3336458.31 frames. ], batch size: 50, lr: 4.48e-02, grad_scale: 16.0 2024-09-22 11:27:09,934 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.873e+02 2.795e+02 3.933e+02 5.449e+02 1.373e+03, threshold=7.866e+02, percent-clipped=13.0 2024-09-22 11:27:42,468 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 11:28:34,215 INFO [train.py:1198] (2/4) Epoch 1, batch 1050, loss[loss=0.526, ctc_loss=0.4486, cr_loss=0.3869, over 17065.00 frames. ], tot_loss[loss=0.6965, ctc_loss=0.6143, cr_loss=0.4112, over 3337088.16 frames. ], batch size: 43, lr: 4.48e-02, grad_scale: 16.0 2024-09-22 11:28:53,031 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.min_positive, batch_count=4946.666666666667, ans=0.034541666666666665 2024-09-22 11:29:01,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=4946.666666666667, ans=0.25053333333333333 2024-09-22 11:29:09,280 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=4993.333333333333, ans=0.045861111111111116 2024-09-22 11:29:51,574 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.max_positive, batch_count=5086.666666666667, ans=0.8008666666666666 2024-09-22 11:29:59,950 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=5133.333333333333, ans=0.009753623188405797 2024-09-22 11:30:01,211 INFO [train.py:1198] (2/4) Epoch 1, batch 1100, loss[loss=0.5305, ctc_loss=0.4412, cr_loss=0.4464, over 17061.00 frames. ], tot_loss[loss=0.6676, ctc_loss=0.5845, cr_loss=0.4159, over 3344643.33 frames. ], batch size: 46, lr: 4.48e-02, grad_scale: 16.0 2024-09-22 11:30:06,284 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.680e+02 2.354e+02 3.241e+02 4.881e+02 1.077e+03, threshold=6.482e+02, percent-clipped=7.0 2024-09-22 11:30:26,713 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=5180.0, ans=0.2571875 2024-09-22 11:30:26,728 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=5180.0, ans=0.2777 2024-09-22 11:30:43,849 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=11.44 vs. limit=11.42 2024-09-22 11:30:43,949 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.38 vs. limit=11.42 2024-09-22 11:31:08,163 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=11.73 vs. limit=11.49 2024-09-22 11:31:14,735 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=12.56 vs. limit=9.495000000000001 2024-09-22 11:31:25,796 INFO [train.py:1198] (2/4) Epoch 1, batch 1150, loss[loss=0.4592, ctc_loss=0.381, cr_loss=0.3907, over 16950.00 frames. ], tot_loss[loss=0.642, ctc_loss=0.5581, cr_loss=0.4196, over 3360563.93 frames. ], batch size: 42, lr: 4.47e-02, grad_scale: 16.0 2024-09-22 11:31:59,117 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=5413.333333333333, ans=0.24625000000000002 2024-09-22 11:32:05,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=5460.0, ans=0.0 2024-09-22 11:32:07,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=5460.0, ans=0.7089000000000001 2024-09-22 11:32:15,890 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=5460.0, ans=0.04391666666666667 2024-09-22 11:32:19,221 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.min_positive, batch_count=5506.666666666667, ans=0.03279166666666666 2024-09-22 11:32:28,218 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=12.84 vs. limit=11.629999999999999 2024-09-22 11:32:29,293 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=5506.666666666667, ans=0.241875 2024-09-22 11:32:58,122 INFO [train.py:1198] (2/4) Epoch 1, batch 1200, loss[loss=0.4683, ctc_loss=0.3916, cr_loss=0.3836, over 17178.00 frames. ], tot_loss[loss=0.6181, ctc_loss=0.5336, cr_loss=0.4225, over 3366006.10 frames. ], batch size: 41, lr: 4.47e-02, grad_scale: 32.0 2024-09-22 11:33:02,960 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.688e+02 2.423e+02 3.283e+02 4.569e+02 8.108e+02, threshold=6.566e+02, percent-clipped=6.0 2024-09-22 11:33:17,268 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.66 vs. limit=11.735 2024-09-22 11:33:43,102 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=5693.333333333333, ans=0.23312500000000003 2024-09-22 11:34:09,444 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=5786.666666666667, ans=0.07 2024-09-22 11:34:23,532 INFO [train.py:1198] (2/4) Epoch 1, batch 1250, loss[loss=0.4874, ctc_loss=0.4027, cr_loss=0.4238, over 17293.00 frames. ], tot_loss[loss=0.5988, ctc_loss=0.5137, cr_loss=0.4254, over 3365104.84 frames. ], batch size: 49, lr: 4.47e-02, grad_scale: 32.0 2024-09-22 11:34:23,863 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=5833.333333333333, ans=0.04236111111111111 2024-09-22 11:34:37,541 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=5833.333333333333, ans=0.2265625 2024-09-22 11:34:39,603 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=9.71 vs. limit=9.705 2024-09-22 11:34:44,191 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=5880.0, ans=0.6942 2024-09-22 11:35:03,129 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=4.67 vs. limit=9.7225 2024-09-22 11:35:20,493 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=5973.333333333333, ans=0.21999999999999997 2024-09-22 11:35:25,895 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.65 vs. limit=11.98 2024-09-22 11:35:30,742 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 11:35:47,097 INFO [train.py:1198] (2/4) Epoch 1, batch 1300, loss[loss=0.5026, ctc_loss=0.4103, cr_loss=0.4615, over 17019.00 frames. ], tot_loss[loss=0.5795, ctc_loss=0.494, cr_loss=0.4275, over 3370885.86 frames. ], batch size: 44, lr: 4.47e-02, grad_scale: 32.0 2024-09-22 11:35:49,535 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=7.51 vs. limit=8.033333333333333 2024-09-22 11:35:52,194 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.553e+02 2.150e+02 2.604e+02 3.500e+02 8.408e+02, threshold=5.208e+02, percent-clipped=5.0 2024-09-22 11:36:54,641 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=6253.333333333333, ans=0.20687499999999998 2024-09-22 11:37:03,157 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=6253.333333333333, ans=0.23746666666666666 2024-09-22 11:37:03,675 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.60 vs. limit=6.5633333333333335 2024-09-22 11:37:12,779 INFO [train.py:1198] (2/4) Epoch 1, batch 1350, loss[loss=0.4818, ctc_loss=0.398, cr_loss=0.4191, over 17019.00 frames. ], tot_loss[loss=0.5643, ctc_loss=0.4785, cr_loss=0.4287, over 3359558.04 frames. ], batch size: 44, lr: 4.46e-02, grad_scale: 32.0 2024-09-22 11:37:14,943 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.52 vs. limit=12.225 2024-09-22 11:37:42,132 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=6346.666666666667, ans=0.009489855072463768 2024-09-22 11:38:14,667 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=6440.0, ans=0.198125 2024-09-22 11:38:21,260 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=6440.0, ans=0.198125 2024-09-22 11:38:41,179 INFO [train.py:1198] (2/4) Epoch 1, batch 1400, loss[loss=0.616, ctc_loss=0.5246, cr_loss=0.4573, over 11543.00 frames. ], tot_loss[loss=0.5507, ctc_loss=0.4647, cr_loss=0.4299, over 3355776.19 frames. ], batch size: 123, lr: 4.46e-02, grad_scale: 32.0 2024-09-22 11:38:46,073 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.832e+02 2.453e+02 3.278e+02 5.014e+02 1.044e+03, threshold=6.556e+02, percent-clipped=21.0 2024-09-22 11:39:02,719 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=6580.0, ans=0.19156250000000002 2024-09-22 11:39:48,081 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=6720.0, ans=0.03866666666666667 2024-09-22 11:39:48,562 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.18 vs. limit=10.02 2024-09-22 11:39:53,630 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=19.39 vs. limit=12.54 2024-09-22 11:40:03,198 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=3.74 vs. limit=10.02 2024-09-22 11:40:05,929 INFO [train.py:1198] (2/4) Epoch 1, batch 1450, loss[loss=0.4718, ctc_loss=0.3819, cr_loss=0.4496, over 17145.00 frames. ], tot_loss[loss=0.5371, ctc_loss=0.451, cr_loss=0.4305, over 3363530.39 frames. ], batch size: 45, lr: 4.46e-02, grad_scale: 32.0 2024-09-22 11:40:25,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=6813.333333333333, ans=0.6615333333333333 2024-09-22 11:40:25,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=6813.333333333333, ans=0.6615333333333333 2024-09-22 11:41:27,762 INFO [train.py:1198] (2/4) Epoch 1, batch 1500, loss[loss=0.4638, ctc_loss=0.3661, cr_loss=0.4884, over 17272.00 frames. ], tot_loss[loss=0.527, ctc_loss=0.4405, cr_loss=0.4324, over 3364446.49 frames. ], batch size: 44, lr: 4.46e-02, grad_scale: 32.0 2024-09-22 11:41:32,711 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.678e+02 2.217e+02 3.113e+02 4.719e+02 9.117e+02, threshold=6.226e+02, percent-clipped=8.0 2024-09-22 11:41:33,022 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=7000.0, ans=0.171875 2024-09-22 11:41:47,767 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=7046.666666666667, ans=0.025 2024-09-22 11:42:24,618 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=7140.0, ans=0.05537500000000001 2024-09-22 11:42:29,685 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=7140.0, ans=0.2286 2024-09-22 11:42:49,780 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=7186.666666666667, ans=0.16312500000000002 2024-09-22 11:42:56,581 INFO [train.py:1198] (2/4) Epoch 1, batch 1550, loss[loss=0.4834, ctc_loss=0.3966, cr_loss=0.4342, over 17029.00 frames. ], tot_loss[loss=0.5169, ctc_loss=0.4302, cr_loss=0.4337, over 3364745.59 frames. ], batch size: 52, lr: 4.45e-02, grad_scale: 32.0 2024-09-22 11:43:23,774 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.83 vs. limit=6.912 2024-09-22 11:43:31,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=7326.666666666667, ans=0.3099 2024-09-22 11:43:39,524 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 11:43:40,057 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.99 vs. limit=6.831666666666667 2024-09-22 11:43:40,217 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=8.08 vs. limit=8.663333333333334 2024-09-22 11:43:41,230 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=7326.666666666667, ans=0.0 2024-09-22 11:43:55,623 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=7373.333333333333, ans=0.035944444444444446 2024-09-22 11:43:57,325 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=7373.333333333333, ans=0.22626666666666667 2024-09-22 11:43:58,860 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=7373.333333333333, ans=0.22626666666666667 2024-09-22 11:44:07,152 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.min_abs, batch_count=7420.0, ans=0.3113 2024-09-22 11:44:17,053 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=7466.666666666667, ans=0.07 2024-09-22 11:44:18,842 INFO [train.py:1198] (2/4) Epoch 1, batch 1600, loss[loss=0.5374, ctc_loss=0.4448, cr_loss=0.4629, over 16895.00 frames. ], tot_loss[loss=0.5065, ctc_loss=0.4197, cr_loss=0.434, over 3372641.13 frames. ], batch size: 58, lr: 4.45e-02, grad_scale: 32.0 2024-09-22 11:44:23,627 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.570e+02 2.041e+02 2.519e+02 3.425e+02 6.490e+02, threshold=5.038e+02, percent-clipped=3.0 2024-09-22 11:44:37,546 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=7513.333333333333, ans=0.6370333333333333 2024-09-22 11:44:41,752 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=6.08 vs. limit=10.317499999999999 2024-09-22 11:44:53,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=7560.0, ans=0.145625 2024-09-22 11:45:18,177 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=7606.666666666667, ans=0.1434375 2024-09-22 11:45:19,711 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=7606.666666666667, ans=0.1434375 2024-09-22 11:45:42,058 INFO [train.py:1198] (2/4) Epoch 1, batch 1650, loss[loss=0.4697, ctc_loss=0.3778, cr_loss=0.4595, over 17300.00 frames. ], tot_loss[loss=0.4986, ctc_loss=0.4116, cr_loss=0.435, over 3376056.75 frames. ], batch size: 46, lr: 4.45e-02, grad_scale: 32.0 2024-09-22 11:45:42,378 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=7700.0, ans=0.009195652173913044 2024-09-22 11:45:43,974 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=7700.0, ans=0.13906249999999998 2024-09-22 11:45:56,888 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=7746.666666666667, ans=0.2225333333333333 2024-09-22 11:46:05,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=7746.666666666667, ans=0.2225333333333333 2024-09-22 11:46:53,157 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=7886.666666666667, ans=0.1303125 2024-09-22 11:46:57,714 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=7886.666666666667, ans=0.22113333333333332 2024-09-22 11:47:05,462 INFO [train.py:1198] (2/4) Epoch 1, batch 1700, loss[loss=0.4199, ctc_loss=0.3347, cr_loss=0.4256, over 17344.00 frames. ], tot_loss[loss=0.4919, ctc_loss=0.4047, cr_loss=0.4358, over 3383796.22 frames. ], batch size: 48, lr: 4.44e-02, grad_scale: 32.0 2024-09-22 11:47:10,185 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.533e+02 2.014e+02 2.718e+02 3.727e+02 5.677e+02, threshold=5.436e+02, percent-clipped=4.0 2024-09-22 11:47:24,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=7980.0, ans=0.2202 2024-09-22 11:47:31,404 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=7980.0, ans=0.12593749999999998 2024-09-22 11:47:46,582 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=8026.666666666667, ans=0.125 2024-09-22 11:48:17,935 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=8120.0, ans=0.6158 2024-09-22 11:48:30,595 INFO [train.py:1198] (2/4) Epoch 1, batch 1750, loss[loss=0.4024, ctc_loss=0.3183, cr_loss=0.4203, over 17280.00 frames. ], tot_loss[loss=0.4844, ctc_loss=0.3974, cr_loss=0.4354, over 3373427.74 frames. ], batch size: 42, lr: 4.44e-02, grad_scale: 32.0 2024-09-22 11:49:18,773 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=8306.666666666666, ans=0.0 2024-09-22 11:49:25,726 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=10.36 vs. limit=10.615 2024-09-22 11:49:25,969 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=5.21 vs. limit=10.615 2024-09-22 11:49:33,719 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.02 vs. limit=7.088333333333333 2024-09-22 11:49:48,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=8353.333333333334, ans=0.03186111111111111 2024-09-22 11:49:53,984 INFO [train.py:1198] (2/4) Epoch 1, batch 1800, loss[loss=0.4907, ctc_loss=0.4027, cr_loss=0.4402, over 16571.00 frames. ], tot_loss[loss=0.4747, ctc_loss=0.388, cr_loss=0.4335, over 3375669.69 frames. ], batch size: 66, lr: 4.44e-02, grad_scale: 32.0 2024-09-22 11:49:58,878 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.382e+02 1.919e+02 2.429e+02 3.208e+02 6.110e+02, threshold=4.858e+02, percent-clipped=5.0 2024-09-22 11:50:02,485 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=8400.0, ans=0.216 2024-09-22 11:50:31,468 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=8493.333333333334, ans=0.6027333333333333 2024-09-22 11:50:36,018 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=8493.333333333334, ans=0.125 2024-09-22 11:50:44,051 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=8540.0, ans=0.2146 2024-09-22 11:50:48,888 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=8540.0, ans=0.031083333333333338 2024-09-22 11:51:05,035 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=8586.666666666666, ans=0.07 2024-09-22 11:51:05,151 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=8586.666666666666, ans=0.030888888888888893 2024-09-22 11:51:13,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=8633.333333333334, ans=0.125 2024-09-22 11:51:14,590 INFO [train.py:1198] (2/4) Epoch 1, batch 1850, loss[loss=0.3812, ctc_loss=0.3092, cr_loss=0.3602, over 16945.00 frames. ], tot_loss[loss=0.4703, ctc_loss=0.3836, cr_loss=0.4336, over 3367148.56 frames. ], batch size: 42, lr: 4.43e-02, grad_scale: 32.0 2024-09-22 11:51:45,467 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.90 vs. limit=14.01 2024-09-22 11:51:50,209 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=15.40 vs. limit=14.044999999999998 2024-09-22 11:51:55,993 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=8726.666666666666, ans=0.125 2024-09-22 11:52:02,507 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=8726.666666666666, ans=0.21273333333333333 2024-09-22 11:52:06,211 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=15.22 vs. limit=14.08 2024-09-22 11:52:11,333 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.40 vs. limit=14.08 2024-09-22 11:52:40,377 INFO [train.py:1198] (2/4) Epoch 1, batch 1900, loss[loss=0.5093, ctc_loss=0.4263, cr_loss=0.4152, over 16732.00 frames. ], tot_loss[loss=0.4662, ctc_loss=0.3793, cr_loss=0.4346, over 3369283.40 frames. ], batch size: 61, lr: 4.43e-02, grad_scale: 32.0 2024-09-22 11:52:44,917 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.532e+02 1.950e+02 2.731e+02 3.550e+02 1.054e+03, threshold=5.462e+02, percent-clipped=8.0 2024-09-22 11:52:50,551 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=11.08 vs. limit=10.825 2024-09-22 11:53:56,755 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=9053.333333333334, ans=0.125 2024-09-22 11:54:02,942 INFO [train.py:1198] (2/4) Epoch 1, batch 1950, loss[loss=0.4306, ctc_loss=0.3497, cr_loss=0.4047, over 17157.00 frames. ], tot_loss[loss=0.4625, ctc_loss=0.3754, cr_loss=0.4359, over 3366267.66 frames. ], batch size: 45, lr: 4.43e-02, grad_scale: 32.0 2024-09-22 11:54:27,498 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=9146.666666666666, ans=0.125 2024-09-22 11:54:36,717 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=13.86 vs. limit=14.395 2024-09-22 11:54:59,340 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=9.90 vs. limit=10.965 2024-09-22 11:55:19,634 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=9286.666666666666, ans=0.027972222222222228 2024-09-22 11:55:26,419 INFO [train.py:1198] (2/4) Epoch 1, batch 2000, loss[loss=0.4217, ctc_loss=0.3365, cr_loss=0.4263, over 17341.00 frames. ], tot_loss[loss=0.459, ctc_loss=0.372, cr_loss=0.4352, over 3359448.06 frames. ], batch size: 48, lr: 4.42e-02, grad_scale: 32.0 2024-09-22 11:55:29,891 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=9333.333333333334, ans=0.125 2024-09-22 11:55:31,236 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.560e+02 1.918e+02 2.442e+02 3.354e+02 7.763e+02, threshold=4.883e+02, percent-clipped=5.0 2024-09-22 11:55:35,143 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=2.93 vs. limit=11.0 2024-09-22 11:55:59,160 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=9426.666666666666, ans=0.125 2024-09-22 11:56:00,610 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=9426.666666666666, ans=0.125 2024-09-22 11:56:26,221 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=9473.333333333334, ans=0.20526666666666665 2024-09-22 11:56:38,091 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten.whitening_limit, batch_count=9520.0, ans=11.07 2024-09-22 11:56:41,914 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=11.29 vs. limit=11.07 2024-09-22 11:56:49,511 INFO [train.py:1198] (2/4) Epoch 1, batch 2050, loss[loss=0.4324, ctc_loss=0.3529, cr_loss=0.3973, over 17267.00 frames. ], tot_loss[loss=0.4545, ctc_loss=0.3674, cr_loss=0.4354, over 3352039.58 frames. ], batch size: 44, lr: 4.42e-02, grad_scale: 32.0 2024-09-22 11:56:59,803 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.32 vs. limit=9.783333333333333 2024-09-22 11:57:00,943 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=9566.666666666666, ans=0.008789855072463769 2024-09-22 11:57:23,610 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=9660.0, ans=0.025 2024-09-22 11:57:43,972 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=9706.666666666666, ans=0.125 2024-09-22 11:58:03,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=9753.333333333334, ans=0.00874927536231884 2024-09-22 11:58:14,743 INFO [train.py:1198] (2/4) Epoch 1, batch 2100, loss[loss=0.4249, ctc_loss=0.339, cr_loss=0.4298, over 17226.00 frames. ], tot_loss[loss=0.4495, ctc_loss=0.3623, cr_loss=0.4359, over 3359780.94 frames. ], batch size: 50, lr: 4.42e-02, grad_scale: 32.0 2024-09-22 11:58:19,529 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.503e+02 1.883e+02 2.281e+02 3.077e+02 7.464e+02, threshold=4.562e+02, percent-clipped=6.0 2024-09-22 11:59:11,130 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=9940.0, ans=0.125 2024-09-22 11:59:20,900 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.min_positive, batch_count=9986.666666666666, ans=0.025 2024-09-22 11:59:37,505 INFO [train.py:1198] (2/4) Epoch 1, batch 2150, loss[loss=0.4044, ctc_loss=0.3154, cr_loss=0.4447, over 17015.00 frames. ], tot_loss[loss=0.4468, ctc_loss=0.3596, cr_loss=0.4362, over 3350014.35 frames. ], batch size: 44, lr: 4.41e-02, grad_scale: 32.0 2024-09-22 11:59:39,309 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=10033.333333333334, ans=0.125 2024-09-22 12:00:14,420 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=10126.666666666666, ans=0.19873333333333332 2024-09-22 12:00:15,948 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=10126.666666666666, ans=0.125 2024-09-22 12:00:40,037 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=10220.0, ans=0.025 2024-09-22 12:00:44,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=10220.0, ans=0.125 2024-09-22 12:00:45,517 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=11.05 vs. limit=11.3325 2024-09-22 12:00:58,121 INFO [train.py:1198] (2/4) Epoch 1, batch 2200, loss[loss=0.3863, ctc_loss=0.3024, cr_loss=0.4194, over 17036.00 frames. ], tot_loss[loss=0.443, ctc_loss=0.356, cr_loss=0.4352, over 3348989.84 frames. ], batch size: 39, lr: 4.41e-02, grad_scale: 32.0 2024-09-22 12:01:02,862 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.617e+02 2.024e+02 2.529e+02 3.777e+02 5.736e+02, threshold=5.059e+02, percent-clipped=14.0 2024-09-22 12:01:12,975 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=10313.333333333334, ans=0.19686666666666666 2024-09-22 12:01:27,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=10313.333333333334, ans=0.19686666666666666 2024-09-22 12:01:29,366 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=10360.0, ans=0.125 2024-09-22 12:01:36,666 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=10360.0, ans=0.125 2024-09-22 12:01:51,690 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.86 vs. limit=10.203333333333333 2024-09-22 12:02:20,135 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.max_abs, batch_count=10500.0, ans=10.0 2024-09-22 12:02:21,445 INFO [train.py:1198] (2/4) Epoch 1, batch 2250, loss[loss=0.401, ctc_loss=0.3154, cr_loss=0.4278, over 17169.00 frames. ], tot_loss[loss=0.4402, ctc_loss=0.3532, cr_loss=0.4351, over 3347049.86 frames. ], batch size: 41, lr: 4.40e-02, grad_scale: 32.0 2024-09-22 12:02:48,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=10546.666666666666, ans=0.5308666666666667 2024-09-22 12:03:03,684 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=10593.333333333334, ans=0.125 2024-09-22 12:03:08,716 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=6.87 vs. limit=11.4725 2024-09-22 12:03:24,341 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=10640.0, ans=0.125 2024-09-22 12:03:46,762 INFO [train.py:1198] (2/4) Epoch 1, batch 2300, loss[loss=0.4033, ctc_loss=0.3184, cr_loss=0.4242, over 17277.00 frames. ], tot_loss[loss=0.434, ctc_loss=0.3471, cr_loss=0.4344, over 3359638.21 frames. ], batch size: 42, lr: 4.40e-02, grad_scale: 32.0 2024-09-22 12:03:47,122 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 12:03:51,576 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.433e+02 1.850e+02 2.386e+02 2.971e+02 5.038e+02, threshold=4.772e+02, percent-clipped=0.0 2024-09-22 12:04:47,655 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=10873.333333333334, ans=0.125 2024-09-22 12:05:00,353 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=10920.0, ans=0.04949747468305833 2024-09-22 12:05:00,522 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=10920.0, ans=0.125 2024-09-22 12:05:09,818 INFO [train.py:1198] (2/4) Epoch 1, batch 2350, loss[loss=0.394, ctc_loss=0.3132, cr_loss=0.4039, over 16727.00 frames. ], tot_loss[loss=0.4307, ctc_loss=0.3439, cr_loss=0.4338, over 3362117.57 frames. ], batch size: 37, lr: 4.40e-02, grad_scale: 32.0 2024-09-22 12:05:15,417 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=16.39 vs. limit=15.725 2024-09-22 12:05:32,175 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=11013.333333333334, ans=0.125 2024-09-22 12:05:54,875 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=11060.0, ans=0.125 2024-09-22 12:06:13,375 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.94 vs. limit=4.673 2024-09-22 12:06:18,803 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=11153.333333333334, ans=0.008444927536231883 2024-09-22 12:06:30,103 INFO [train.py:1198] (2/4) Epoch 1, batch 2400, loss[loss=0.4164, ctc_loss=0.3332, cr_loss=0.4162, over 16536.00 frames. ], tot_loss[loss=0.4285, ctc_loss=0.3417, cr_loss=0.4337, over 3362162.46 frames. ], batch size: 66, lr: 4.39e-02, grad_scale: 32.0 2024-09-22 12:06:37,258 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.394e+02 1.777e+02 2.025e+02 2.571e+02 5.493e+02, threshold=4.051e+02, percent-clipped=2.0 2024-09-22 12:06:37,689 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=11200.0, ans=0.125 2024-09-22 12:06:40,807 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=11200.0, ans=0.125 2024-09-22 12:06:44,623 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.42 vs. limit=4.68 2024-09-22 12:07:07,872 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=11293.333333333334, ans=0.008414492753623189 2024-09-22 12:07:34,410 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=11340.0, ans=0.07 2024-09-22 12:07:41,769 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=11386.666666666666, ans=0.125 2024-09-22 12:07:57,472 INFO [train.py:1198] (2/4) Epoch 1, batch 2450, loss[loss=0.3568, ctc_loss=0.275, cr_loss=0.409, over 17271.00 frames. ], tot_loss[loss=0.4288, ctc_loss=0.342, cr_loss=0.4342, over 3352169.64 frames. ], batch size: 42, lr: 4.39e-02, grad_scale: 64.0 2024-09-22 12:08:05,329 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.17 vs. limit=4.715 2024-09-22 12:08:38,465 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer_ff3.min_abs, batch_count=11526.666666666666, ans=0.2 2024-09-22 12:08:43,300 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=11526.666666666666, ans=0.125 2024-09-22 12:09:15,285 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=11620.0, ans=0.008343478260869565 2024-09-22 12:09:18,266 INFO [train.py:1198] (2/4) Epoch 1, batch 2500, loss[loss=0.3717, ctc_loss=0.2885, cr_loss=0.4157, over 16941.00 frames. ], tot_loss[loss=0.4266, ctc_loss=0.3397, cr_loss=0.4347, over 3354931.25 frames. ], batch size: 42, lr: 4.38e-02, grad_scale: 64.0 2024-09-22 12:09:22,934 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.438e+02 2.074e+02 2.928e+02 4.593e+02 9.871e+02, threshold=5.856e+02, percent-clipped=30.0 2024-09-22 12:09:36,035 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=4.50 vs. limit=8.685333333333332 2024-09-22 12:09:43,975 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.54 vs. limit=4.757 2024-09-22 12:09:48,448 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=11713.333333333334, ans=0.008323188405797101 2024-09-22 12:09:58,105 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=11760.0, ans=0.1824 2024-09-22 12:10:02,903 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=11760.0, ans=0.1824 2024-09-22 12:10:40,845 INFO [train.py:1198] (2/4) Epoch 1, batch 2550, loss[loss=0.3766, ctc_loss=0.2949, cr_loss=0.4085, over 17017.00 frames. ], tot_loss[loss=0.4226, ctc_loss=0.336, cr_loss=0.4332, over 3359887.73 frames. ], batch size: 44, lr: 4.38e-02, grad_scale: 64.0 2024-09-22 12:10:41,881 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=18.62 vs. limit=16.425 2024-09-22 12:10:46,424 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=17.34 vs. limit=16.425 2024-09-22 12:10:48,001 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.89 vs. limit=8.76 2024-09-22 12:11:23,217 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=11993.333333333334, ans=0.008262318840579711 2024-09-22 12:11:27,939 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=12040.0, ans=0.025 2024-09-22 12:11:29,382 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=12040.0, ans=0.17959999999999998 2024-09-22 12:11:30,083 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.07 vs. limit=12.015 2024-09-22 12:11:36,761 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=12040.0, ans=0.125 2024-09-22 12:11:38,521 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=12040.0, ans=0.125 2024-09-22 12:11:56,361 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=12086.666666666666, ans=0.01630555555555556 2024-09-22 12:11:59,490 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=12086.666666666666, ans=0.17913333333333334 2024-09-22 12:12:04,476 INFO [train.py:1198] (2/4) Epoch 1, batch 2600, loss[loss=0.4015, ctc_loss=0.3206, cr_loss=0.4048, over 17258.00 frames. ], tot_loss[loss=0.4199, ctc_loss=0.3333, cr_loss=0.433, over 3364728.84 frames. ], batch size: 44, lr: 4.37e-02, grad_scale: 64.0 2024-09-22 12:12:09,352 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.462e+02 1.984e+02 2.668e+02 3.339e+02 5.918e+02, threshold=5.335e+02, percent-clipped=1.0 2024-09-22 12:12:14,465 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=12133.333333333334, ans=0.025 2024-09-22 12:12:50,067 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=12226.666666666666, ans=0.125 2024-09-22 12:12:56,146 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=12273.333333333334, ans=0.125 2024-09-22 12:12:56,229 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=12273.333333333334, ans=0.125 2024-09-22 12:13:23,601 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=12320.0, ans=0.008191304347826087 2024-09-22 12:13:29,612 INFO [train.py:1198] (2/4) Epoch 1, batch 2650, loss[loss=0.4183, ctc_loss=0.3248, cr_loss=0.4675, over 17347.00 frames. ], tot_loss[loss=0.418, ctc_loss=0.3314, cr_loss=0.4327, over 3368169.83 frames. ], batch size: 48, lr: 4.37e-02, grad_scale: 64.0 2024-09-22 12:13:44,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=12413.333333333334, ans=0.01494444444444444 2024-09-22 12:13:59,085 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=12413.333333333334, ans=0.125 2024-09-22 12:14:03,860 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=12460.0, ans=0.014750000000000006 2024-09-22 12:14:43,386 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=12553.333333333334, ans=0.008140579710144927 2024-09-22 12:14:52,682 INFO [train.py:1198] (2/4) Epoch 1, batch 2700, loss[loss=0.3646, ctc_loss=0.2874, cr_loss=0.386, over 17018.00 frames. ], tot_loss[loss=0.4184, ctc_loss=0.3317, cr_loss=0.4336, over 3355437.37 frames. ], batch size: 39, lr: 4.36e-02, grad_scale: 64.0 2024-09-22 12:14:57,514 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.460e+02 1.915e+02 2.536e+02 3.410e+02 5.700e+02, threshold=5.072e+02, percent-clipped=2.0 2024-09-22 12:15:02,641 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=12600.0, ans=0.174 2024-09-22 12:15:25,116 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=12693.333333333334, ans=0.025 2024-09-22 12:15:46,154 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=7.36 vs. limit=8.185 2024-09-22 12:15:56,902 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=12786.666666666666, ans=0.008089855072463768 2024-09-22 12:16:10,984 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=12833.333333333334, ans=0.008079710144927536 2024-09-22 12:16:12,417 INFO [train.py:1198] (2/4) Epoch 1, batch 2750, loss[loss=0.404, ctc_loss=0.3188, cr_loss=0.4261, over 17044.00 frames. ], tot_loss[loss=0.4149, ctc_loss=0.3286, cr_loss=0.4315, over 3353754.43 frames. ], batch size: 52, lr: 4.36e-02, grad_scale: 64.0 2024-09-22 12:16:47,436 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=12926.666666666666, ans=0.17073333333333335 2024-09-22 12:16:51,002 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=5.59 vs. limit=8.231666666666666 2024-09-22 12:16:53,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=12926.666666666666, ans=0.125 2024-09-22 12:17:09,747 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=12973.333333333334, ans=0.125 2024-09-22 12:17:14,315 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=12973.333333333334, ans=0.012611111111111108 2024-09-22 12:17:17,603 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=12973.333333333334, ans=0.125 2024-09-22 12:17:27,079 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=13020.0, ans=0.00803913043478261 2024-09-22 12:17:32,833 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=13020.0, ans=0.125 2024-09-22 12:17:40,889 INFO [train.py:1198] (2/4) Epoch 1, batch 2800, loss[loss=0.4133, ctc_loss=0.3262, cr_loss=0.4357, over 16008.00 frames. ], tot_loss[loss=0.4123, ctc_loss=0.3259, cr_loss=0.4319, over 3361100.29 frames. ], batch size: 74, lr: 4.36e-02, grad_scale: 64.0 2024-09-22 12:17:45,584 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.462e+02 1.867e+02 2.104e+02 2.772e+02 6.258e+02, threshold=4.209e+02, percent-clipped=2.0 2024-09-22 12:18:50,666 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=4.16 vs. limit=9.301333333333334 2024-09-22 12:19:00,770 INFO [train.py:1198] (2/4) Epoch 1, batch 2850, loss[loss=0.3798, ctc_loss=0.2936, cr_loss=0.4309, over 17347.00 frames. ], tot_loss[loss=0.4125, ctc_loss=0.326, cr_loss=0.4326, over 3352364.89 frames. ], batch size: 48, lr: 4.35e-02, grad_scale: 32.0 2024-09-22 12:19:12,777 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.00 vs. limit=12.4875 2024-09-22 12:19:20,155 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=13346.666666666666, ans=0.16653333333333334 2024-09-22 12:19:31,505 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=15.72 vs. limit=17.509999999999998 2024-09-22 12:19:35,745 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.min_positive, batch_count=13393.333333333334, ans=0.05 2024-09-22 12:19:53,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=13440.0, ans=0.1656 2024-09-22 12:20:21,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=13486.666666666666, ans=0.125 2024-09-22 12:20:24,534 INFO [train.py:1198] (2/4) Epoch 1, batch 2900, loss[loss=0.4114, ctc_loss=0.3192, cr_loss=0.461, over 17011.00 frames. ], tot_loss[loss=0.4115, ctc_loss=0.3249, cr_loss=0.4327, over 3359238.16 frames. ], batch size: 44, lr: 4.35e-02, grad_scale: 32.0 2024-09-22 12:20:29,810 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=13533.333333333334, ans=0.125 2024-09-22 12:20:31,042 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.490e+02 1.863e+02 2.342e+02 3.249e+02 5.939e+02, threshold=4.685e+02, percent-clipped=7.0 2024-09-22 12:20:59,549 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=11.14 vs. limit=12.61 2024-09-22 12:21:29,361 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=13673.333333333334, ans=0.125 2024-09-22 12:21:37,391 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=13720.0, ans=0.125 2024-09-22 12:21:48,333 INFO [train.py:1198] (2/4) Epoch 1, batch 2950, loss[loss=0.4537, ctc_loss=0.3574, cr_loss=0.4815, over 17290.00 frames. ], tot_loss[loss=0.4104, ctc_loss=0.3238, cr_loss=0.4326, over 3352450.86 frames. ], batch size: 51, lr: 4.34e-02, grad_scale: 32.0 2024-09-22 12:22:08,973 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=13813.333333333334, ans=0.009111111111111105 2024-09-22 12:22:33,972 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=13860.0, ans=0.125 2024-09-22 12:22:34,388 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=19.81 vs. limit=17.895 2024-09-22 12:22:43,680 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=13906.666666666666, ans=0.16093333333333334 2024-09-22 12:22:45,960 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=5.97 vs. limit=9.562666666666665 2024-09-22 12:23:13,840 INFO [train.py:1198] (2/4) Epoch 1, batch 3000, loss[loss=0.3845, ctc_loss=0.2971, cr_loss=0.4371, over 17041.00 frames. ], tot_loss[loss=0.4084, ctc_loss=0.322, cr_loss=0.4317, over 3360899.75 frames. ], batch size: 39, lr: 4.34e-02, grad_scale: 32.0 2024-09-22 12:23:13,840 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-22 12:23:29,184 INFO [train.py:1230] (2/4) Epoch 1, validation: loss=0.1235, ctc_loss=0.1235, cr_loss=7.044e-15, over 944034.00 frames. 2024-09-22 12:23:29,184 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-22 12:23:35,639 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.419e+02 1.886e+02 2.276e+02 2.833e+02 5.148e+02, threshold=4.553e+02, percent-clipped=2.0 2024-09-22 12:24:00,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=14093.333333333334, ans=0.125 2024-09-22 12:24:40,810 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=19.90 vs. limit=18.14 2024-09-22 12:24:46,464 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=14233.333333333334, ans=0.06834166666666666 2024-09-22 12:24:47,935 INFO [train.py:1198] (2/4) Epoch 1, batch 3050, loss[loss=0.4297, ctc_loss=0.3355, cr_loss=0.4708, over 17228.00 frames. ], tot_loss[loss=0.4039, ctc_loss=0.3179, cr_loss=0.4302, over 3368701.82 frames. ], batch size: 55, lr: 4.33e-02, grad_scale: 32.0 2024-09-22 12:25:22,188 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=4.11 vs. limit=9.730666666666666 2024-09-22 12:25:27,973 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=14326.666666666666, ans=0.15673333333333334 2024-09-22 12:25:51,739 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=14420.0, ans=0.007734782608695652 2024-09-22 12:26:09,729 INFO [train.py:1198] (2/4) Epoch 1, batch 3100, loss[loss=0.4321, ctc_loss=0.3417, cr_loss=0.4519, over 16404.00 frames. ], tot_loss[loss=0.403, ctc_loss=0.3167, cr_loss=0.4311, over 3375125.04 frames. ], batch size: 66, lr: 4.33e-02, grad_scale: 32.0 2024-09-22 12:26:15,852 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.374e+02 1.870e+02 2.371e+02 3.057e+02 5.717e+02, threshold=4.743e+02, percent-clipped=5.0 2024-09-22 12:26:22,741 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=12.73 vs. limit=12.925 2024-09-22 12:26:24,201 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=14513.333333333334, ans=0.125 2024-09-22 12:26:30,203 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=14513.333333333334, ans=0.125 2024-09-22 12:26:38,171 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=14513.333333333334, ans=0.15486666666666665 2024-09-22 12:26:46,135 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=14560.0, ans=0.39039999999999997 2024-09-22 12:26:50,794 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=14560.0, ans=0.125 2024-09-22 12:27:11,636 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=14653.333333333334, ans=0.15346666666666667 2024-09-22 12:27:16,530 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=14653.333333333334, ans=0.04949747468305833 2024-09-22 12:27:17,282 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn1.whiten, num_groups=1, num_channels=192, metric=11.07 vs. limit=18.490000000000002 2024-09-22 12:27:18,127 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=14653.333333333334, ans=0.125 2024-09-22 12:27:23,189 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=512, metric=3.44 vs. limit=12.995000000000001 2024-09-22 12:27:24,099 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=14653.333333333334, ans=0.025 2024-09-22 12:27:28,657 INFO [train.py:1198] (2/4) Epoch 1, batch 3150, loss[loss=0.398, ctc_loss=0.318, cr_loss=0.4, over 17006.00 frames. ], tot_loss[loss=0.4023, ctc_loss=0.3161, cr_loss=0.4313, over 3374604.52 frames. ], batch size: 53, lr: 4.32e-02, grad_scale: 32.0 2024-09-22 12:27:34,995 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=14700.0, ans=0.38550000000000006 2024-09-22 12:28:22,774 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.78 vs. limit=13.065000000000001 2024-09-22 12:28:42,816 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.29 vs. limit=5.2330000000000005 2024-09-22 12:28:47,524 INFO [train.py:1198] (2/4) Epoch 1, batch 3200, loss[loss=0.4706, ctc_loss=0.376, cr_loss=0.4732, over 16809.00 frames. ], tot_loss[loss=0.3993, ctc_loss=0.3132, cr_loss=0.4304, over 3373203.85 frames. ], batch size: 61, lr: 4.32e-02, grad_scale: 32.0 2024-09-22 12:28:53,525 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.398e+02 1.832e+02 2.280e+02 2.861e+02 6.877e+02, threshold=4.560e+02, percent-clipped=3.0 2024-09-22 12:28:59,959 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=14933.333333333334, ans=0.004444444444444438 2024-09-22 12:29:13,053 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=14980.0, ans=0.125 2024-09-22 12:29:33,544 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=15073.333333333334, ans=0.125 2024-09-22 12:29:50,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=15120.0, ans=0.125 2024-09-22 12:29:55,403 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=15120.0, ans=0.007582608695652174 2024-09-22 12:30:06,100 INFO [train.py:1198] (2/4) Epoch 1, batch 3250, loss[loss=0.3529, ctc_loss=0.2696, cr_loss=0.4166, over 17058.00 frames. ], tot_loss[loss=0.3996, ctc_loss=0.3134, cr_loss=0.4312, over 3368887.13 frames. ], batch size: 39, lr: 4.31e-02, grad_scale: 32.0 2024-09-22 12:30:14,295 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=15166.666666666666, ans=0.0034722222222222238 2024-09-22 12:30:41,377 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=15260.0, ans=0.125 2024-09-22 12:30:48,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=15260.0, ans=0.0030833333333333338 2024-09-22 12:30:58,938 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=10.97 vs. limit=13.24 2024-09-22 12:31:06,371 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=15306.666666666666, ans=0.125 2024-09-22 12:31:10,318 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys.whitening_limit, batch_count=15353.333333333334, ans=5.303 2024-09-22 12:31:10,319 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=5.25 vs. limit=5.303 2024-09-22 12:31:17,703 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=15353.333333333334, ans=0.125 2024-09-22 12:31:22,339 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=15353.333333333334, ans=0.007531884057971014 2024-09-22 12:31:26,912 INFO [train.py:1198] (2/4) Epoch 1, batch 3300, loss[loss=0.4001, ctc_loss=0.3086, cr_loss=0.4574, over 16910.00 frames. ], tot_loss[loss=0.3978, ctc_loss=0.3118, cr_loss=0.4299, over 3365574.87 frames. ], batch size: 58, lr: 4.31e-02, grad_scale: 32.0 2024-09-22 12:31:33,367 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.515e+02 1.835e+02 2.401e+02 3.313e+02 5.174e+02, threshold=4.802e+02, percent-clipped=5.0 2024-09-22 12:31:54,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=15446.666666666666, ans=0.125 2024-09-22 12:31:55,612 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=10.13 vs. limit=13.2925 2024-09-22 12:32:10,736 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-22 12:32:31,007 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=15586.666666666666, ans=0.3544666666666667 2024-09-22 12:32:39,104 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=15586.666666666666, ans=0.125 2024-09-22 12:32:50,476 INFO [train.py:1198] (2/4) Epoch 1, batch 3350, loss[loss=0.3375, ctc_loss=0.267, cr_loss=0.3524, over 17019.00 frames. ], tot_loss[loss=0.3968, ctc_loss=0.3109, cr_loss=0.4294, over 3368255.40 frames. ], batch size: 39, lr: 4.30e-02, grad_scale: 32.0 2024-09-22 12:32:54,617 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=7.10 vs. limit=13.3625 2024-09-22 12:32:58,891 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=8.96 vs. limit=10.253333333333334 2024-09-22 12:33:16,184 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=15680.0, ans=0.14320000000000002 2024-09-22 12:33:17,600 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=15680.0, ans=0.125 2024-09-22 12:33:39,386 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=15773.333333333334, ans=0.025 2024-09-22 12:34:04,315 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=15820.0, ans=0.025 2024-09-22 12:34:09,473 INFO [train.py:1198] (2/4) Epoch 1, batch 3400, loss[loss=0.3678, ctc_loss=0.2896, cr_loss=0.3908, over 17043.00 frames. ], tot_loss[loss=0.3966, ctc_loss=0.3107, cr_loss=0.4295, over 3356180.59 frames. ], batch size: 39, lr: 4.29e-02, grad_scale: 32.0 2024-09-22 12:34:13,157 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-22 12:34:15,798 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.414e+02 1.769e+02 2.096e+02 2.628e+02 4.837e+02, threshold=4.193e+02, percent-clipped=1.0 2024-09-22 12:34:17,997 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=6.45 vs. limit=10.346666666666668 2024-09-22 12:34:20,718 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=15866.666666666666, ans=0.125 2024-09-22 12:34:33,711 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=15913.333333333334, ans=0.00036111111111111066 2024-09-22 12:34:38,556 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=15913.333333333334, ans=0.14086666666666667 2024-09-22 12:34:40,039 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=15960.0, ans=0.00016666666666666913 2024-09-22 12:34:41,532 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=15960.0, ans=0.0074 2024-09-22 12:34:52,477 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=15960.0, ans=0.9096 2024-09-22 12:35:03,793 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=4.52 vs. limit=13.502500000000001 2024-09-22 12:35:19,154 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=16053.333333333334, ans=0.3381333333333333 2024-09-22 12:35:28,155 INFO [train.py:1198] (2/4) Epoch 1, batch 3450, loss[loss=0.4005, ctc_loss=0.3144, cr_loss=0.4307, over 15887.00 frames. ], tot_loss[loss=0.396, ctc_loss=0.3102, cr_loss=0.429, over 3352471.97 frames. ], batch size: 74, lr: 4.29e-02, grad_scale: 32.0 2024-09-22 12:35:49,016 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=16146.666666666666, ans=0.08853333333333332 2024-09-22 12:36:14,342 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=16193.333333333334, ans=0.0 2024-09-22 12:36:14,344 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=16193.333333333334, ans=0.125 2024-09-22 12:36:17,534 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=16240.0, ans=0.0 2024-09-22 12:36:19,303 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=16240.0, ans=0.125 2024-09-22 12:36:20,718 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=16240.0, ans=0.125 2024-09-22 12:36:38,172 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.51 vs. limit=13.6075 2024-09-22 12:36:45,835 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=16286.666666666666, ans=0.04949747468305833 2024-09-22 12:36:47,332 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=16333.333333333334, ans=0.0 2024-09-22 12:36:48,696 INFO [train.py:1198] (2/4) Epoch 1, batch 3500, loss[loss=0.3639, ctc_loss=0.2834, cr_loss=0.4022, over 17208.00 frames. ], tot_loss[loss=0.3943, ctc_loss=0.3087, cr_loss=0.4284, over 3354872.36 frames. ], batch size: 41, lr: 4.28e-02, grad_scale: 32.0 2024-09-22 12:36:54,800 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.402e+02 1.781e+02 2.318e+02 3.100e+02 5.527e+02, threshold=4.636e+02, percent-clipped=9.0 2024-09-22 12:37:06,088 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=16380.0, ans=0.125 2024-09-22 12:37:21,693 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=16426.666666666668, ans=0.125 2024-09-22 12:37:26,560 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=14.39 vs. limit=13.66 2024-09-22 12:37:52,661 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=16520.0, ans=0.0 2024-09-22 12:37:53,420 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=6.15 vs. limit=9.129999999999999 2024-09-22 12:38:06,622 INFO [train.py:1198] (2/4) Epoch 1, batch 3550, loss[loss=0.4246, ctc_loss=0.3322, cr_loss=0.4619, over 17020.00 frames. ], tot_loss[loss=0.3927, ctc_loss=0.3069, cr_loss=0.4286, over 3363338.35 frames. ], batch size: 52, lr: 4.28e-02, grad_scale: 32.0 2024-09-22 12:38:09,945 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=16566.666666666668, ans=0.007268115942028985 2024-09-22 12:38:44,107 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=16660.0, ans=0.125 2024-09-22 12:38:47,364 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 12:38:48,797 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=16660.0, ans=0.13340000000000002 2024-09-22 12:38:56,648 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=16706.666666666668, ans=0.125 2024-09-22 12:39:10,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.min_positive, batch_count=16753.333333333332, ans=0.08246666666666669 2024-09-22 12:39:24,819 INFO [train.py:1198] (2/4) Epoch 1, batch 3600, loss[loss=0.4224, ctc_loss=0.3286, cr_loss=0.4689, over 16091.00 frames. ], tot_loss[loss=0.3916, ctc_loss=0.3059, cr_loss=0.4285, over 3360963.34 frames. ], batch size: 75, lr: 4.27e-02, grad_scale: 32.0 2024-09-22 12:39:29,581 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=16800.0, ans=0.125 2024-09-22 12:39:30,801 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.429e+02 1.842e+02 2.057e+02 2.677e+02 5.057e+02, threshold=4.115e+02, percent-clipped=2.0 2024-09-22 12:39:37,277 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=16800.0, ans=0.025 2024-09-22 12:39:46,486 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=16846.666666666668, ans=0.125 2024-09-22 12:39:46,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=16846.666666666668, ans=0.3103666666666667 2024-09-22 12:39:49,494 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=16846.666666666668, ans=0.1315333333333333 2024-09-22 12:39:54,246 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=16893.333333333332, ans=0.125 2024-09-22 12:40:43,993 INFO [train.py:1198] (2/4) Epoch 1, batch 3650, loss[loss=0.407, ctc_loss=0.3149, cr_loss=0.4601, over 17223.00 frames. ], tot_loss[loss=0.3917, ctc_loss=0.306, cr_loss=0.4283, over 3355899.21 frames. ], batch size: 47, lr: 4.27e-02, grad_scale: 32.0 2024-09-22 12:40:47,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=17033.333333333332, ans=0.12966666666666668 2024-09-22 12:40:59,093 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=5.19 vs. limit=9.27 2024-09-22 12:41:38,390 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.92 vs. limit=5.5760000000000005 2024-09-22 12:41:50,825 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=17220.0, ans=0.0 2024-09-22 12:42:00,567 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=4.12 vs. limit=10.888 2024-09-22 12:42:05,821 INFO [train.py:1198] (2/4) Epoch 1, batch 3700, loss[loss=0.3635, ctc_loss=0.2795, cr_loss=0.4198, over 17239.00 frames. ], tot_loss[loss=0.3892, ctc_loss=0.3037, cr_loss=0.4275, over 3349252.30 frames. ], batch size: 44, lr: 4.26e-02, grad_scale: 32.0 2024-09-22 12:42:12,057 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.410e+02 1.892e+02 2.660e+02 3.633e+02 5.715e+02, threshold=5.320e+02, percent-clipped=15.0 2024-09-22 12:42:15,526 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=17266.666666666668, ans=0.0 2024-09-22 12:42:26,680 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.87 vs. limit=13.656666666666666 2024-09-22 12:42:29,371 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=17313.333333333332, ans=0.125 2024-09-22 12:43:18,126 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=17453.333333333332, ans=0.0 2024-09-22 12:43:22,576 INFO [train.py:1198] (2/4) Epoch 1, batch 3750, loss[loss=0.4037, ctc_loss=0.3156, cr_loss=0.4404, over 16138.00 frames. ], tot_loss[loss=0.3903, ctc_loss=0.3046, cr_loss=0.4283, over 3330617.36 frames. ], batch size: 74, lr: 4.26e-02, grad_scale: 32.0 2024-09-22 12:43:24,274 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=17500.0, ans=0.125 2024-09-22 12:43:34,847 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=17500.0, ans=0.125 2024-09-22 12:43:41,080 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=17546.666666666668, ans=0.125 2024-09-22 12:44:14,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=17640.0, ans=0.125 2024-09-22 12:44:21,941 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=6.38 vs. limit=14.115 2024-09-22 12:44:22,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=17686.666666666668, ans=0.007024637681159421 2024-09-22 12:44:26,034 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 12:44:40,334 INFO [train.py:1198] (2/4) Epoch 1, batch 3800, loss[loss=0.3898, ctc_loss=0.3102, cr_loss=0.3978, over 14825.00 frames. ], tot_loss[loss=0.3921, ctc_loss=0.3062, cr_loss=0.4292, over 3315922.83 frames. ], batch size: 89, lr: 4.25e-02, grad_scale: 32.0 2024-09-22 12:44:46,416 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.363e+02 1.780e+02 2.356e+02 3.200e+02 5.376e+02, threshold=4.713e+02, percent-clipped=1.0 2024-09-22 12:45:03,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=17780.0, ans=0.27770000000000006 2024-09-22 12:45:07,208 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=17780.0, ans=0.125 2024-09-22 12:45:33,037 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=17873.333333333332, ans=0.125 2024-09-22 12:45:50,861 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=5.18 vs. limit=9.48 2024-09-22 12:45:55,166 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=17920.0, ans=0.006973913043478261 2024-09-22 12:45:59,543 INFO [train.py:1198] (2/4) Epoch 1, batch 3850, loss[loss=0.4734, ctc_loss=0.3874, cr_loss=0.4302, over 11999.00 frames. ], tot_loss[loss=0.3941, ctc_loss=0.3082, cr_loss=0.4295, over 3280231.97 frames. ], batch size: 123, lr: 4.24e-02, grad_scale: 32.0 2024-09-22 12:46:01,376 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=17966.666666666668, ans=0.125 2024-09-22 12:46:09,246 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=17966.666666666668, ans=0.006963768115942029 2024-09-22 12:46:10,790 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=17966.666666666668, ans=0.125 2024-09-22 12:46:12,301 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=17966.666666666668, ans=0.12033333333333332 2024-09-22 12:46:45,518 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=18106.666666666668, ans=0.125 2024-09-22 12:47:54,283 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-22 12:48:02,933 INFO [train.py:1198] (2/4) Epoch 2, batch 0, loss[loss=0.3723, ctc_loss=0.2975, cr_loss=0.3741, over 17098.00 frames. ], tot_loss[loss=0.3723, ctc_loss=0.2975, cr_loss=0.3741, over 17098.00 frames. ], batch size: 40, lr: 4.16e-02, grad_scale: 32.0 2024-09-22 12:48:02,933 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-22 12:48:18,075 INFO [train.py:1230] (2/4) Epoch 2, validation: loss=0.1169, ctc_loss=0.1169, cr_loss=1.034e-14, over 944034.00 frames. 2024-09-22 12:48:18,076 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-22 12:48:30,952 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.378e+02 1.944e+02 2.365e+02 3.007e+02 5.794e+02, threshold=4.731e+02, percent-clipped=1.0 2024-09-22 12:48:38,500 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.max_positive, batch_count=18228.0, ans=0.93228 2024-09-22 12:48:48,055 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=18228.0, ans=0.0 2024-09-22 12:48:49,474 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.min_abs, batch_count=18228.0, ans=0.47342 2024-09-22 12:49:05,127 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=18274.666666666668, ans=0.0 2024-09-22 12:49:19,316 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=18321.333333333332, ans=0.125 2024-09-22 12:49:25,071 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module1.whiten, num_groups=1, num_channels=192, metric=5.28 vs. limit=14.388 2024-09-22 12:49:32,127 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=18368.0, ans=0.125 2024-09-22 12:49:33,567 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=18368.0, ans=0.0 2024-09-22 12:49:39,679 INFO [train.py:1198] (2/4) Epoch 2, batch 50, loss[loss=0.3626, ctc_loss=0.281, cr_loss=0.4079, over 17299.00 frames. ], tot_loss[loss=0.3845, ctc_loss=0.2992, cr_loss=0.4266, over 751960.83 frames. ], batch size: 46, lr: 4.15e-02, grad_scale: 32.0 2024-09-22 12:49:49,582 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=18414.666666666668, ans=0.07 2024-09-22 12:50:05,392 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=18461.333333333332, ans=0.11538666666666669 2024-09-22 12:50:12,045 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 12:50:19,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=18508.0, ans=0.006846086956521739 2024-09-22 12:50:34,108 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=18554.666666666668, ans=0.125 2024-09-22 12:50:35,849 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=18554.666666666668, ans=0.125 2024-09-22 12:50:53,073 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=18601.333333333332, ans=0.125 2024-09-22 12:50:59,412 INFO [train.py:1198] (2/4) Epoch 2, batch 100, loss[loss=0.3683, ctc_loss=0.2864, cr_loss=0.4091, over 17226.00 frames. ], tot_loss[loss=0.381, ctc_loss=0.296, cr_loss=0.4251, over 1327381.03 frames. ], batch size: 47, lr: 4.15e-02, grad_scale: 32.0 2024-09-22 12:51:19,088 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.460e+02 1.796e+02 2.128e+02 2.839e+02 5.119e+02, threshold=4.256e+02, percent-clipped=1.0 2024-09-22 12:51:29,675 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=6.19 vs. limit=14.5105 2024-09-22 12:51:53,067 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=18788.0, ans=0.125 2024-09-22 12:51:56,273 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=18788.0, ans=0.125 2024-09-22 12:52:26,281 INFO [train.py:1198] (2/4) Epoch 2, batch 150, loss[loss=0.3414, ctc_loss=0.2605, cr_loss=0.4048, over 17121.00 frames. ], tot_loss[loss=0.3796, ctc_loss=0.2943, cr_loss=0.4265, over 1779317.77 frames. ], batch size: 40, lr: 4.14e-02, grad_scale: 32.0 2024-09-22 12:52:37,100 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=18881.333333333332, ans=0.04949747468305833 2024-09-22 12:53:17,184 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=19021.333333333332, ans=0.125 2024-09-22 12:53:51,164 INFO [train.py:1198] (2/4) Epoch 2, batch 200, loss[loss=0.3762, ctc_loss=0.2912, cr_loss=0.4253, over 17295.00 frames. ], tot_loss[loss=0.3771, ctc_loss=0.292, cr_loss=0.4254, over 2133118.97 frames. ], batch size: 51, lr: 4.13e-02, grad_scale: 32.0 2024-09-22 12:53:57,559 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=19114.666666666668, ans=0.0 2024-09-22 12:54:03,746 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.386e+02 1.786e+02 2.214e+02 2.969e+02 6.338e+02, threshold=4.427e+02, percent-clipped=7.0 2024-09-22 12:54:05,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=19161.333333333332, ans=0.125 2024-09-22 12:54:13,654 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=19161.333333333332, ans=0.10838666666666669 2024-09-22 12:54:40,489 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=19254.666666666668, ans=0.125 2024-09-22 12:55:04,264 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=19301.333333333332, ans=0.125 2024-09-22 12:55:10,170 INFO [train.py:1198] (2/4) Epoch 2, batch 250, loss[loss=0.3278, ctc_loss=0.2503, cr_loss=0.3872, over 17069.00 frames. ], tot_loss[loss=0.3772, ctc_loss=0.2919, cr_loss=0.4262, over 2402781.15 frames. ], batch size: 43, lr: 4.13e-02, grad_scale: 32.0 2024-09-22 12:55:21,666 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=19348.0, ans=0.125 2024-09-22 12:55:24,904 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=19394.666666666668, ans=0.0 2024-09-22 12:55:53,386 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=19441.333333333332, ans=0.006643188405797102 2024-09-22 12:56:35,317 INFO [train.py:1198] (2/4) Epoch 2, batch 300, loss[loss=0.366, ctc_loss=0.277, cr_loss=0.445, over 17258.00 frames. ], tot_loss[loss=0.3783, ctc_loss=0.293, cr_loss=0.4266, over 2613893.18 frames. ], batch size: 44, lr: 4.12e-02, grad_scale: 32.0 2024-09-22 12:56:48,372 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.373e+02 1.759e+02 2.124e+02 2.853e+02 4.892e+02, threshold=4.248e+02, percent-clipped=3.0 2024-09-22 12:57:04,476 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=19628.0, ans=0.125 2024-09-22 12:57:21,907 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=19721.333333333332, ans=0.125 2024-09-22 12:57:42,254 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=19768.0, ans=0.10232000000000002 2024-09-22 12:57:53,319 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=19768.0, ans=0.0 2024-09-22 12:57:57,762 INFO [train.py:1198] (2/4) Epoch 2, batch 350, loss[loss=0.4007, ctc_loss=0.3172, cr_loss=0.4175, over 16511.00 frames. ], tot_loss[loss=0.3772, ctc_loss=0.2921, cr_loss=0.4254, over 2781057.18 frames. ], batch size: 66, lr: 4.12e-02, grad_scale: 32.0 2024-09-22 12:58:02,678 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=19814.666666666668, ans=0.125 2024-09-22 12:58:31,053 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=19908.0, ans=0.0 2024-09-22 12:58:41,098 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=5.29 vs. limit=11.9632 2024-09-22 12:58:58,540 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=24.77 vs. limit=22.466 2024-09-22 12:59:09,165 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=20001.333333333332, ans=0.125 2024-09-22 12:59:09,167 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=20001.333333333332, ans=0.125 2024-09-22 12:59:20,246 INFO [train.py:1198] (2/4) Epoch 2, batch 400, loss[loss=0.3338, ctc_loss=0.2514, cr_loss=0.4121, over 16942.00 frames. ], tot_loss[loss=0.3768, ctc_loss=0.2916, cr_loss=0.4256, over 2912717.02 frames. ], batch size: 42, lr: 4.11e-02, grad_scale: 32.0 2024-09-22 12:59:20,479 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=20048.0, ans=0.125 2024-09-22 12:59:22,614 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=19.70 vs. limit=22.5 2024-09-22 12:59:32,897 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.308e+02 1.825e+02 2.194e+02 2.995e+02 5.365e+02, threshold=4.388e+02, percent-clipped=4.0 2024-09-22 12:59:39,486 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=20094.666666666668, ans=0.125 2024-09-22 12:59:50,403 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=20141.333333333332, ans=0.006491014492753624 2024-09-22 12:59:58,804 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.82 vs. limit=15.0 2024-09-22 13:00:01,948 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=23.25 vs. limit=15.0 2024-09-22 13:00:09,676 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=20188.0, ans=0.1 2024-09-22 13:00:39,701 INFO [train.py:1198] (2/4) Epoch 2, batch 450, loss[loss=0.4035, ctc_loss=0.3069, cr_loss=0.483, over 17148.00 frames. ], tot_loss[loss=0.3763, ctc_loss=0.2911, cr_loss=0.426, over 3015023.59 frames. ], batch size: 48, lr: 4.10e-02, grad_scale: 32.0 2024-09-22 13:00:40,110 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=20281.333333333332, ans=0.125 2024-09-22 13:00:48,453 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=20.92 vs. limit=22.5 2024-09-22 13:01:05,325 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-22 13:01:25,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.min_positive, batch_count=20374.666666666668, ans=0.05 2024-09-22 13:01:38,699 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=7.04 vs. limit=15.0 2024-09-22 13:01:40,213 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.50 vs. limit=10.0 2024-09-22 13:01:57,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=20468.0, ans=0.125 2024-09-22 13:02:04,846 INFO [train.py:1198] (2/4) Epoch 2, batch 500, loss[loss=0.3185, ctc_loss=0.2444, cr_loss=0.3704, over 17072.00 frames. ], tot_loss[loss=0.3749, ctc_loss=0.2899, cr_loss=0.4253, over 3091342.03 frames. ], batch size: 43, lr: 4.10e-02, grad_scale: 32.0 2024-09-22 13:02:05,645 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.38 vs. limit=12.0 2024-09-22 13:02:14,036 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=14.45 vs. limit=15.0 2024-09-22 13:02:17,796 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.414e+02 1.804e+02 2.205e+02 2.881e+02 5.655e+02, threshold=4.410e+02, percent-clipped=3.0 2024-09-22 13:02:24,399 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=20561.333333333332, ans=0.0 2024-09-22 13:02:28,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=20561.333333333332, ans=0.04949747468305833 2024-09-22 13:02:58,642 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 13:03:21,368 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 13:03:29,948 INFO [train.py:1198] (2/4) Epoch 2, batch 550, loss[loss=0.3517, ctc_loss=0.2694, cr_loss=0.4117, over 17033.00 frames. ], tot_loss[loss=0.3746, ctc_loss=0.2895, cr_loss=0.4256, over 3147133.76 frames. ], batch size: 44, lr: 4.09e-02, grad_scale: 32.0 2024-09-22 13:03:38,365 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=20748.0, ans=0.125 2024-09-22 13:04:13,582 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=20841.333333333332, ans=0.0 2024-09-22 13:04:32,553 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=20934.666666666668, ans=0.125 2024-09-22 13:04:34,578 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=13.44 vs. limit=15.0 2024-09-22 13:04:49,740 INFO [train.py:1198] (2/4) Epoch 2, batch 600, loss[loss=0.4076, ctc_loss=0.3218, cr_loss=0.429, over 16881.00 frames. ], tot_loss[loss=0.3736, ctc_loss=0.2884, cr_loss=0.4258, over 3197338.24 frames. ], batch size: 58, lr: 4.09e-02, grad_scale: 32.0 2024-09-22 13:04:54,874 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=20981.333333333332, ans=0.2 2024-09-22 13:04:59,750 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=20981.333333333332, ans=0.0 2024-09-22 13:05:02,738 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.397e+02 1.714e+02 2.171e+02 2.480e+02 4.403e+02, threshold=4.342e+02, percent-clipped=0.0 2024-09-22 13:05:17,413 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=21028.0, ans=0.125 2024-09-22 13:05:43,366 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=2.85 vs. limit=15.0 2024-09-22 13:05:54,034 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=21168.0, ans=0.0 2024-09-22 13:06:15,351 INFO [train.py:1198] (2/4) Epoch 2, batch 650, loss[loss=0.361, ctc_loss=0.2683, cr_loss=0.4633, over 17202.00 frames. ], tot_loss[loss=0.3726, ctc_loss=0.2873, cr_loss=0.4265, over 3240691.76 frames. ], batch size: 47, lr: 4.08e-02, grad_scale: 32.0 2024-09-22 13:06:33,177 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=21261.333333333332, ans=0.1 2024-09-22 13:06:47,814 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=21308.0, ans=0.0 2024-09-22 13:07:01,859 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.min_positive, batch_count=21354.666666666668, ans=0.025 2024-09-22 13:07:03,604 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=21354.666666666668, ans=0.125 2024-09-22 13:07:36,315 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=21448.0, ans=0.015 2024-09-22 13:07:37,761 INFO [train.py:1198] (2/4) Epoch 2, batch 700, loss[loss=0.3808, ctc_loss=0.2971, cr_loss=0.4189, over 16532.00 frames. ], tot_loss[loss=0.3718, ctc_loss=0.2866, cr_loss=0.4258, over 3266395.98 frames. ], batch size: 66, lr: 4.07e-02, grad_scale: 32.0 2024-09-22 13:07:50,030 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=12.51 vs. limit=15.0 2024-09-22 13:07:50,782 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.427e+02 1.762e+02 2.307e+02 2.806e+02 6.388e+02, threshold=4.614e+02, percent-clipped=12.0 2024-09-22 13:07:59,716 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=11.28 vs. limit=15.0 2024-09-22 13:08:13,215 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-22 13:08:29,999 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=21588.0, ans=0.006176521739130435 2024-09-22 13:08:59,998 INFO [train.py:1198] (2/4) Epoch 2, batch 750, loss[loss=0.4272, ctc_loss=0.337, cr_loss=0.4511, over 15020.00 frames. ], tot_loss[loss=0.3714, ctc_loss=0.2863, cr_loss=0.4252, over 3293197.54 frames. ], batch size: 89, lr: 4.07e-02, grad_scale: 32.0 2024-09-22 13:09:04,030 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=22.91 vs. limit=22.5 2024-09-22 13:09:36,655 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=21774.666666666668, ans=0.1 2024-09-22 13:10:19,246 INFO [train.py:1198] (2/4) Epoch 2, batch 800, loss[loss=0.4066, ctc_loss=0.3187, cr_loss=0.4394, over 16846.00 frames. ], tot_loss[loss=0.373, ctc_loss=0.2876, cr_loss=0.4266, over 3303594.35 frames. ], batch size: 58, lr: 4.06e-02, grad_scale: 32.0 2024-09-22 13:10:32,122 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.379e+02 1.875e+02 2.284e+02 2.792e+02 4.521e+02, threshold=4.569e+02, percent-clipped=0.0 2024-09-22 13:10:37,045 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=21961.333333333332, ans=0.0 2024-09-22 13:11:33,156 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=22101.333333333332, ans=0.1 2024-09-22 13:11:43,999 INFO [train.py:1198] (2/4) Epoch 2, batch 850, loss[loss=0.3841, ctc_loss=0.2938, cr_loss=0.4517, over 17050.00 frames. ], tot_loss[loss=0.3731, ctc_loss=0.2877, cr_loss=0.4272, over 3316406.70 frames. ], batch size: 52, lr: 4.06e-02, grad_scale: 32.0 2024-09-22 13:11:56,955 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=22148.0, ans=0.125 2024-09-22 13:12:00,162 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=22194.666666666668, ans=0.09899494936611666 2024-09-22 13:12:06,516 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=22194.666666666668, ans=0.125 2024-09-22 13:12:13,047 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=22194.666666666668, ans=0.025 2024-09-22 13:12:22,457 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=12.27 vs. limit=15.0 2024-09-22 13:13:05,748 INFO [train.py:1198] (2/4) Epoch 2, batch 900, loss[loss=0.3585, ctc_loss=0.2715, cr_loss=0.4353, over 17131.00 frames. ], tot_loss[loss=0.3734, ctc_loss=0.288, cr_loss=0.427, over 3320213.20 frames. ], batch size: 48, lr: 4.05e-02, grad_scale: 32.0 2024-09-22 13:13:06,060 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=22381.333333333332, ans=0.025 2024-09-22 13:13:15,902 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=22381.333333333332, ans=0.0 2024-09-22 13:13:21,505 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.391e+02 1.677e+02 2.003e+02 2.906e+02 6.404e+02, threshold=4.006e+02, percent-clipped=3.0 2024-09-22 13:13:26,924 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=24.93 vs. limit=22.5 2024-09-22 13:13:31,273 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=22428.0, ans=0.2 2024-09-22 13:13:55,255 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=22521.333333333332, ans=0.1 2024-09-22 13:14:06,393 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=22521.333333333332, ans=0.125 2024-09-22 13:14:27,566 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=22614.666666666668, ans=0.125 2024-09-22 13:14:28,988 INFO [train.py:1198] (2/4) Epoch 2, batch 950, loss[loss=0.3275, ctc_loss=0.2528, cr_loss=0.374, over 17098.00 frames. ], tot_loss[loss=0.3717, ctc_loss=0.2865, cr_loss=0.426, over 3335902.91 frames. ], batch size: 43, lr: 4.04e-02, grad_scale: 64.0 2024-09-22 13:15:05,894 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=22708.0, ans=0.125 2024-09-22 13:15:48,455 INFO [train.py:1198] (2/4) Epoch 2, batch 1000, loss[loss=0.3867, ctc_loss=0.2919, cr_loss=0.4739, over 17173.00 frames. ], tot_loss[loss=0.3717, ctc_loss=0.2864, cr_loss=0.4266, over 3341280.49 frames. ], batch size: 55, lr: 4.04e-02, grad_scale: 64.0 2024-09-22 13:16:07,762 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.375e+02 1.793e+02 2.170e+02 2.654e+02 4.100e+02, threshold=4.339e+02, percent-clipped=2.0 2024-09-22 13:17:03,924 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=23034.666666666668, ans=0.005862028985507246 2024-09-22 13:17:15,714 INFO [train.py:1198] (2/4) Epoch 2, batch 1050, loss[loss=0.339, ctc_loss=0.2625, cr_loss=0.3826, over 17248.00 frames. ], tot_loss[loss=0.3705, ctc_loss=0.2855, cr_loss=0.4252, over 3340395.57 frames. ], batch size: 44, lr: 4.03e-02, grad_scale: 32.0 2024-09-22 13:17:29,119 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=23081.333333333332, ans=0.125 2024-09-22 13:17:38,609 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=23128.0, ans=0.1 2024-09-22 13:17:38,727 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=23128.0, ans=0.125 2024-09-22 13:17:47,442 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.whiten_keys.whitening_limit, batch_count=23174.666666666668, ans=6.0 2024-09-22 13:17:50,084 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=23174.666666666668, ans=0.125 2024-09-22 13:17:58,278 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=23174.666666666668, ans=0.2 2024-09-22 13:18:07,762 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=23221.333333333332, ans=0.125 2024-09-22 13:18:18,619 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=5.46 vs. limit=6.0 2024-09-22 13:18:30,833 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=23268.0, ans=0.0 2024-09-22 13:18:38,529 INFO [train.py:1198] (2/4) Epoch 2, batch 1100, loss[loss=0.3812, ctc_loss=0.2943, cr_loss=0.4349, over 17091.00 frames. ], tot_loss[loss=0.371, ctc_loss=0.2858, cr_loss=0.4257, over 3338395.82 frames. ], batch size: 49, lr: 4.03e-02, grad_scale: 32.0 2024-09-22 13:18:53,229 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.326e+02 1.707e+02 2.038e+02 2.702e+02 5.370e+02, threshold=4.076e+02, percent-clipped=2.0 2024-09-22 13:19:47,530 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=23501.333333333332, ans=0.125 2024-09-22 13:19:58,506 INFO [train.py:1198] (2/4) Epoch 2, batch 1150, loss[loss=0.3266, ctc_loss=0.2415, cr_loss=0.4254, over 17103.00 frames. ], tot_loss[loss=0.3708, ctc_loss=0.2856, cr_loss=0.426, over 3336122.49 frames. ], batch size: 40, lr: 4.02e-02, grad_scale: 32.0 2024-09-22 13:19:58,757 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=23548.0, ans=0.2 2024-09-22 13:20:05,239 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=23548.0, ans=0.07 2024-09-22 13:20:09,770 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=23548.0, ans=0.1 2024-09-22 13:20:36,234 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=8.77 vs. limit=15.0 2024-09-22 13:20:45,016 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=23688.0, ans=0.0 2024-09-22 13:21:02,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=23688.0, ans=0.125 2024-09-22 13:21:23,360 INFO [train.py:1198] (2/4) Epoch 2, batch 1200, loss[loss=0.3674, ctc_loss=0.2838, cr_loss=0.4177, over 17209.00 frames. ], tot_loss[loss=0.3726, ctc_loss=0.2872, cr_loss=0.427, over 3335412.56 frames. ], batch size: 50, lr: 4.01e-02, grad_scale: 32.0 2024-09-22 13:21:35,109 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=23781.333333333332, ans=0.125 2024-09-22 13:21:37,785 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.332e+02 1.890e+02 2.251e+02 2.751e+02 4.727e+02, threshold=4.502e+02, percent-clipped=4.0 2024-09-22 13:21:51,373 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.78 vs. limit=10.0 2024-09-22 13:21:54,287 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=23874.666666666668, ans=0.025 2024-09-22 13:22:03,831 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=23874.666666666668, ans=0.005679420289855072 2024-09-22 13:22:12,650 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=23921.333333333332, ans=0.0 2024-09-22 13:22:40,108 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.max_positive, batch_count=23968.0, ans=0.95 2024-09-22 13:22:46,210 INFO [train.py:1198] (2/4) Epoch 2, batch 1250, loss[loss=0.3878, ctc_loss=0.2994, cr_loss=0.4421, over 16973.00 frames. ], tot_loss[loss=0.3713, ctc_loss=0.2862, cr_loss=0.4255, over 3332336.89 frames. ], batch size: 53, lr: 4.01e-02, grad_scale: 32.0 2024-09-22 13:23:09,928 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=24061.333333333332, ans=0.125 2024-09-22 13:23:13,179 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=24061.333333333332, ans=0.005638840579710145 2024-09-22 13:23:31,035 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=24108.0, ans=0.1 2024-09-22 13:23:40,612 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=24154.666666666668, ans=0.05 2024-09-22 13:23:48,291 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=24154.666666666668, ans=0.125 2024-09-22 13:23:54,947 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.06 vs. limit=15.0 2024-09-22 13:24:08,731 INFO [train.py:1198] (2/4) Epoch 2, batch 1300, loss[loss=0.3238, ctc_loss=0.2438, cr_loss=0.3999, over 17084.00 frames. ], tot_loss[loss=0.3697, ctc_loss=0.2846, cr_loss=0.4254, over 3346912.66 frames. ], batch size: 43, lr: 4.00e-02, grad_scale: 32.0 2024-09-22 13:24:09,002 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=24248.0, ans=0.2 2024-09-22 13:24:22,207 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=24248.0, ans=0.2 2024-09-22 13:24:23,423 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.394e+02 1.885e+02 2.250e+02 3.074e+02 5.750e+02, threshold=4.500e+02, percent-clipped=7.0 2024-09-22 13:24:36,672 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=24294.666666666668, ans=0.1 2024-09-22 13:24:42,785 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 13:24:46,058 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=24341.333333333332, ans=0.005577971014492754 2024-09-22 13:25:25,812 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-22 13:25:28,592 INFO [train.py:1198] (2/4) Epoch 2, batch 1350, loss[loss=0.2984, ctc_loss=0.2215, cr_loss=0.3845, over 16395.00 frames. ], tot_loss[loss=0.3666, ctc_loss=0.2819, cr_loss=0.4235, over 3358873.75 frames. ], batch size: 36, lr: 3.99e-02, grad_scale: 32.0 2024-09-22 13:25:58,120 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=24528.0, ans=0.125 2024-09-22 13:26:12,589 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=24574.666666666668, ans=0.125 2024-09-22 13:26:14,081 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=24574.666666666668, ans=0.2 2024-09-22 13:26:27,573 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=18.15 vs. limit=22.5 2024-09-22 13:26:45,003 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=24668.0, ans=0.025 2024-09-22 13:26:54,050 INFO [train.py:1198] (2/4) Epoch 2, batch 1400, loss[loss=0.4076, ctc_loss=0.3142, cr_loss=0.467, over 17212.00 frames. ], tot_loss[loss=0.3664, ctc_loss=0.2816, cr_loss=0.4237, over 3366920.15 frames. ], batch size: 50, lr: 3.99e-02, grad_scale: 32.0 2024-09-22 13:27:10,325 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=21.20 vs. limit=22.5 2024-09-22 13:27:11,157 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.488e+02 2.006e+02 2.479e+02 3.166e+02 4.715e+02, threshold=4.958e+02, percent-clipped=3.0 2024-09-22 13:27:17,082 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=11.79 vs. limit=15.0 2024-09-22 13:27:35,549 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=24808.0, ans=0.005476521739130435 2024-09-22 13:27:50,180 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=24854.666666666668, ans=0.0 2024-09-22 13:27:56,444 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=24854.666666666668, ans=0.0 2024-09-22 13:28:19,235 INFO [train.py:1198] (2/4) Epoch 2, batch 1450, loss[loss=0.3413, ctc_loss=0.2633, cr_loss=0.39, over 17098.00 frames. ], tot_loss[loss=0.3639, ctc_loss=0.2793, cr_loss=0.4228, over 3368878.68 frames. ], batch size: 49, lr: 3.98e-02, grad_scale: 32.0 2024-09-22 13:28:58,162 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=21.89 vs. limit=22.5 2024-09-22 13:29:19,669 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=25088.0, ans=0.125 2024-09-22 13:29:21,412 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=25134.666666666668, ans=0.0 2024-09-22 13:29:26,966 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=11.04 vs. limit=15.0 2024-09-22 13:29:38,894 INFO [train.py:1198] (2/4) Epoch 2, batch 1500, loss[loss=0.3168, ctc_loss=0.2439, cr_loss=0.3644, over 16955.00 frames. ], tot_loss[loss=0.3625, ctc_loss=0.2782, cr_loss=0.4217, over 3364691.31 frames. ], batch size: 42, lr: 3.98e-02, grad_scale: 32.0 2024-09-22 13:29:40,740 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=25181.333333333332, ans=0.125 2024-09-22 13:29:53,536 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.354e+02 1.701e+02 2.138e+02 2.969e+02 5.127e+02, threshold=4.276e+02, percent-clipped=1.0 2024-09-22 13:30:01,917 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=25228.0, ans=0.125 2024-09-22 13:30:19,264 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=25274.666666666668, ans=0.2 2024-09-22 13:30:19,286 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=25274.666666666668, ans=0.125 2024-09-22 13:30:28,684 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=25321.333333333332, ans=0.0 2024-09-22 13:30:39,082 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.48 vs. limit=15.0 2024-09-22 13:31:03,514 INFO [train.py:1198] (2/4) Epoch 2, batch 1550, loss[loss=0.4151, ctc_loss=0.3266, cr_loss=0.4423, over 15857.00 frames. ], tot_loss[loss=0.3624, ctc_loss=0.2781, cr_loss=0.4218, over 3365814.45 frames. ], batch size: 74, lr: 3.97e-02, grad_scale: 32.0 2024-09-22 13:31:22,987 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=25461.333333333332, ans=0.125 2024-09-22 13:31:27,961 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=25461.333333333332, ans=0.125 2024-09-22 13:31:51,422 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=25554.666666666668, ans=0.1 2024-09-22 13:32:09,802 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=25601.333333333332, ans=0.125 2024-09-22 13:32:25,136 INFO [train.py:1198] (2/4) Epoch 2, batch 1600, loss[loss=0.3567, ctc_loss=0.2788, cr_loss=0.3895, over 17312.00 frames. ], tot_loss[loss=0.36, ctc_loss=0.2758, cr_loss=0.4207, over 3376599.34 frames. ], batch size: 46, lr: 3.96e-02, grad_scale: 32.0 2024-09-22 13:32:30,113 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=25648.0, ans=0.125 2024-09-22 13:32:39,455 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.451e+02 1.757e+02 1.985e+02 2.306e+02 3.282e+02, threshold=3.970e+02, percent-clipped=0.0 2024-09-22 13:33:47,217 INFO [train.py:1198] (2/4) Epoch 2, batch 1650, loss[loss=0.3702, ctc_loss=0.2836, cr_loss=0.4327, over 17144.00 frames. ], tot_loss[loss=0.3602, ctc_loss=0.2758, cr_loss=0.4217, over 3381244.01 frames. ], batch size: 48, lr: 3.96e-02, grad_scale: 32.0 2024-09-22 13:34:00,426 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=25881.333333333332, ans=0.1 2024-09-22 13:34:05,322 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=25928.0, ans=0.0 2024-09-22 13:34:32,731 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=25974.666666666668, ans=0.1 2024-09-22 13:34:34,295 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 13:34:47,335 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 13:34:55,190 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.max_positive, batch_count=26068.0, ans=0.95 2024-09-22 13:35:07,607 INFO [train.py:1198] (2/4) Epoch 2, batch 1700, loss[loss=0.3862, ctc_loss=0.2921, cr_loss=0.4705, over 17212.00 frames. ], tot_loss[loss=0.3631, ctc_loss=0.2783, cr_loss=0.4237, over 3374314.31 frames. ], batch size: 47, lr: 3.95e-02, grad_scale: 32.0 2024-09-22 13:35:14,562 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=26114.666666666668, ans=0.125 2024-09-22 13:35:22,186 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.426e+02 1.761e+02 2.144e+02 2.600e+02 4.141e+02, threshold=4.288e+02, percent-clipped=2.0 2024-09-22 13:36:23,829 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=26301.333333333332, ans=0.1 2024-09-22 13:36:33,101 INFO [train.py:1198] (2/4) Epoch 2, batch 1750, loss[loss=0.3737, ctc_loss=0.2903, cr_loss=0.4172, over 17190.00 frames. ], tot_loss[loss=0.3635, ctc_loss=0.2787, cr_loss=0.4242, over 3375875.05 frames. ], batch size: 55, lr: 3.94e-02, grad_scale: 32.0 2024-09-22 13:37:18,882 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.max_positive, batch_count=26441.333333333332, ans=0.95 2024-09-22 13:37:28,525 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=26488.0, ans=0.125 2024-09-22 13:37:28,538 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=26488.0, ans=0.125 2024-09-22 13:37:57,654 INFO [train.py:1198] (2/4) Epoch 2, batch 1800, loss[loss=0.3393, ctc_loss=0.2594, cr_loss=0.3997, over 17296.00 frames. ], tot_loss[loss=0.3637, ctc_loss=0.2787, cr_loss=0.4247, over 3378141.44 frames. ], batch size: 46, lr: 3.94e-02, grad_scale: 32.0 2024-09-22 13:38:10,525 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=26581.333333333332, ans=0.2 2024-09-22 13:38:11,940 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.435e+02 1.770e+02 2.078e+02 2.651e+02 4.856e+02, threshold=4.156e+02, percent-clipped=2.0 2024-09-22 13:38:37,886 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=26674.666666666668, ans=0.125 2024-09-22 13:38:59,940 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.min_positive, batch_count=26768.0, ans=0.025 2024-09-22 13:39:04,996 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=26768.0, ans=0.125 2024-09-22 13:39:11,230 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=26768.0, ans=0.125 2024-09-22 13:39:12,893 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=26768.0, ans=0.95 2024-09-22 13:39:17,324 INFO [train.py:1198] (2/4) Epoch 2, batch 1850, loss[loss=0.3784, ctc_loss=0.2954, cr_loss=0.4149, over 16820.00 frames. ], tot_loss[loss=0.3631, ctc_loss=0.2782, cr_loss=0.4244, over 3377717.35 frames. ], batch size: 61, lr: 3.93e-02, grad_scale: 32.0 2024-09-22 13:39:24,070 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=26814.666666666668, ans=0.1 2024-09-22 13:39:24,226 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=26814.666666666668, ans=0.125 2024-09-22 13:39:54,483 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.84 vs. limit=15.0 2024-09-22 13:39:55,708 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=26908.0, ans=0.0 2024-09-22 13:40:18,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=26954.666666666668, ans=0.125 2024-09-22 13:40:39,945 INFO [train.py:1198] (2/4) Epoch 2, batch 1900, loss[loss=0.3755, ctc_loss=0.292, cr_loss=0.4176, over 17309.00 frames. ], tot_loss[loss=0.3617, ctc_loss=0.2769, cr_loss=0.4238, over 3375242.12 frames. ], batch size: 51, lr: 3.92e-02, grad_scale: 32.0 2024-09-22 13:40:57,183 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.334e+02 1.864e+02 2.204e+02 2.855e+02 4.990e+02, threshold=4.407e+02, percent-clipped=2.0 2024-09-22 13:40:58,121 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.94 vs. limit=15.0 2024-09-22 13:41:36,485 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=27188.0, ans=0.0 2024-09-22 13:42:05,866 INFO [train.py:1198] (2/4) Epoch 2, batch 1950, loss[loss=0.3599, ctc_loss=0.272, cr_loss=0.4395, over 16647.00 frames. ], tot_loss[loss=0.3602, ctc_loss=0.2756, cr_loss=0.4228, over 3376537.78 frames. ], batch size: 61, lr: 3.92e-02, grad_scale: 32.0 2024-09-22 13:42:22,216 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 13:42:28,996 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.68 vs. limit=6.0 2024-09-22 13:42:33,405 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=27328.0, ans=0.004928695652173913 2024-09-22 13:42:38,079 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=27374.666666666668, ans=0.125 2024-09-22 13:42:44,217 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=27374.666666666668, ans=0.5 2024-09-22 13:42:44,329 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=27374.666666666668, ans=0.1 2024-09-22 13:42:47,795 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=7.46 vs. limit=15.0 2024-09-22 13:42:56,818 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.24 vs. limit=10.0 2024-09-22 13:43:00,968 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=27421.333333333332, ans=0.1 2024-09-22 13:43:03,405 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=6.48 vs. limit=10.0 2024-09-22 13:43:07,201 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=27421.333333333332, ans=0.125 2024-09-22 13:43:15,244 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=27468.0, ans=0.125 2024-09-22 13:43:27,962 INFO [train.py:1198] (2/4) Epoch 2, batch 2000, loss[loss=0.3661, ctc_loss=0.2768, cr_loss=0.4467, over 17234.00 frames. ], tot_loss[loss=0.36, ctc_loss=0.2754, cr_loss=0.4229, over 3381230.76 frames. ], batch size: 55, lr: 3.91e-02, grad_scale: 32.0 2024-09-22 13:43:42,254 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.401e+02 1.881e+02 2.197e+02 2.842e+02 5.136e+02, threshold=4.393e+02, percent-clipped=2.0 2024-09-22 13:43:56,811 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=27561.333333333332, ans=0.0 2024-09-22 13:44:00,893 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=7.41 vs. limit=10.0 2024-09-22 13:44:17,856 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=27654.666666666668, ans=0.0048576811594202894 2024-09-22 13:44:47,945 INFO [train.py:1198] (2/4) Epoch 2, batch 2050, loss[loss=0.3453, ctc_loss=0.2636, cr_loss=0.4084, over 17218.00 frames. ], tot_loss[loss=0.3599, ctc_loss=0.2751, cr_loss=0.4238, over 3373394.11 frames. ], batch size: 50, lr: 3.91e-02, grad_scale: 32.0 2024-09-22 13:44:48,331 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=27748.0, ans=0.125 2024-09-22 13:44:49,741 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=27748.0, ans=0.0 2024-09-22 13:45:51,342 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=27888.0, ans=0.125 2024-09-22 13:45:56,510 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.92 vs. limit=6.0 2024-09-22 13:46:05,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=27934.666666666668, ans=0.125 2024-09-22 13:46:08,973 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=27934.666666666668, ans=10.0 2024-09-22 13:46:13,303 INFO [train.py:1198] (2/4) Epoch 2, batch 2100, loss[loss=0.3127, ctc_loss=0.2328, cr_loss=0.3997, over 17159.00 frames. ], tot_loss[loss=0.3616, ctc_loss=0.2767, cr_loss=0.4244, over 3355062.76 frames. ], batch size: 45, lr: 3.90e-02, grad_scale: 32.0 2024-09-22 13:46:15,919 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=11.57 vs. limit=15.0 2024-09-22 13:46:27,943 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.428e+02 1.746e+02 2.180e+02 2.623e+02 4.533e+02, threshold=4.360e+02, percent-clipped=2.0 2024-09-22 13:46:47,583 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=28074.666666666668, ans=0.1 2024-09-22 13:47:03,104 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=28121.333333333332, ans=0.0047562318840579714 2024-09-22 13:47:04,501 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=28121.333333333332, ans=0.0047562318840579714 2024-09-22 13:47:07,666 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=28121.333333333332, ans=0.1 2024-09-22 13:47:14,187 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=28121.333333333332, ans=0.125 2024-09-22 13:47:23,590 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.min_positive, batch_count=28168.0, ans=0.025 2024-09-22 13:47:31,678 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=28168.0, ans=0.0 2024-09-22 13:47:31,690 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=28168.0, ans=0.0047460869565217395 2024-09-22 13:47:36,047 INFO [train.py:1198] (2/4) Epoch 2, batch 2150, loss[loss=0.3791, ctc_loss=0.2879, cr_loss=0.456, over 17046.00 frames. ], tot_loss[loss=0.3602, ctc_loss=0.2756, cr_loss=0.423, over 3359147.26 frames. ], batch size: 52, lr: 3.89e-02, grad_scale: 32.0 2024-09-22 13:47:37,921 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=28214.666666666668, ans=0.125 2024-09-22 13:48:04,450 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=28261.333333333332, ans=0.2 2024-09-22 13:48:18,596 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=28308.0, ans=0.07 2024-09-22 13:48:20,854 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=5.02 vs. limit=15.0 2024-09-22 13:48:37,887 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=28354.666666666668, ans=0.125 2024-09-22 13:48:44,520 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=28401.333333333332, ans=0.125 2024-09-22 13:48:58,600 INFO [train.py:1198] (2/4) Epoch 2, batch 2200, loss[loss=0.4211, ctc_loss=0.333, cr_loss=0.4402, over 14892.00 frames. ], tot_loss[loss=0.3608, ctc_loss=0.2761, cr_loss=0.4236, over 3357836.55 frames. ], batch size: 89, lr: 3.89e-02, grad_scale: 32.0 2024-09-22 13:49:12,809 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.310e+02 1.723e+02 2.085e+02 2.498e+02 4.255e+02, threshold=4.169e+02, percent-clipped=0.0 2024-09-22 13:49:26,669 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=3.64 vs. limit=15.0 2024-09-22 13:49:37,553 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=512, metric=3.32 vs. limit=15.0 2024-09-22 13:49:42,112 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.44 vs. limit=15.0 2024-09-22 13:49:46,720 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=28588.0, ans=0.2 2024-09-22 13:49:57,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=28588.0, ans=0.2 2024-09-22 13:50:05,826 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=28634.666666666668, ans=0.125 2024-09-22 13:50:18,353 INFO [train.py:1198] (2/4) Epoch 2, batch 2250, loss[loss=0.3469, ctc_loss=0.2587, cr_loss=0.441, over 17018.00 frames. ], tot_loss[loss=0.3604, ctc_loss=0.2758, cr_loss=0.4233, over 3356890.32 frames. ], batch size: 44, lr: 3.88e-02, grad_scale: 32.0 2024-09-22 13:50:25,100 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=28681.333333333332, ans=0.004634492753623189 2024-09-22 13:50:37,439 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=28728.0, ans=0.125 2024-09-22 13:50:40,550 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=28728.0, ans=0.125 2024-09-22 13:51:43,311 INFO [train.py:1198] (2/4) Epoch 2, batch 2300, loss[loss=0.3287, ctc_loss=0.2469, cr_loss=0.4091, over 17059.00 frames. ], tot_loss[loss=0.358, ctc_loss=0.2737, cr_loss=0.4216, over 3369653.49 frames. ], batch size: 46, lr: 3.87e-02, grad_scale: 32.0 2024-09-22 13:52:00,470 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.412e+02 1.812e+02 2.268e+02 2.899e+02 4.767e+02, threshold=4.537e+02, percent-clipped=4.0 2024-09-22 13:52:22,230 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.whiten.whitening_limit, batch_count=29008.0, ans=15.0 2024-09-22 13:52:24,641 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=29008.0, ans=0.07 2024-09-22 13:52:27,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=29008.0, ans=0.1 2024-09-22 13:52:39,123 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=29054.666666666668, ans=0.0 2024-09-22 13:52:51,858 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=29101.333333333332, ans=0.125 2024-09-22 13:53:02,438 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=29101.333333333332, ans=0.07 2024-09-22 13:53:02,968 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=6.23 vs. limit=15.0 2024-09-22 13:53:04,929 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.92 vs. limit=6.0 2024-09-22 13:53:08,675 INFO [train.py:1198] (2/4) Epoch 2, batch 2350, loss[loss=0.3067, ctc_loss=0.2336, cr_loss=0.3658, over 17152.00 frames. ], tot_loss[loss=0.3589, ctc_loss=0.2746, cr_loss=0.4214, over 3360372.44 frames. ], batch size: 48, lr: 3.87e-02, grad_scale: 32.0 2024-09-22 13:53:18,956 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=19.29 vs. limit=22.5 2024-09-22 13:53:21,903 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=29148.0, ans=0.2 2024-09-22 13:53:32,936 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=29194.666666666668, ans=0.125 2024-09-22 13:54:27,693 INFO [train.py:1198] (2/4) Epoch 2, batch 2400, loss[loss=0.2986, ctc_loss=0.2228, cr_loss=0.3794, over 17110.00 frames. ], tot_loss[loss=0.3587, ctc_loss=0.274, cr_loss=0.4234, over 3371437.59 frames. ], batch size: 40, lr: 3.86e-02, grad_scale: 32.0 2024-09-22 13:54:41,835 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.409e+02 1.714e+02 1.995e+02 2.694e+02 4.976e+02, threshold=3.990e+02, percent-clipped=1.0 2024-09-22 13:54:53,500 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=29428.0, ans=0.125 2024-09-22 13:55:15,542 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=29521.333333333332, ans=0.004451884057971015 2024-09-22 13:55:34,075 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 13:55:52,213 INFO [train.py:1198] (2/4) Epoch 2, batch 2450, loss[loss=0.3388, ctc_loss=0.2548, cr_loss=0.4201, over 16653.00 frames. ], tot_loss[loss=0.3583, ctc_loss=0.2736, cr_loss=0.4233, over 3367318.11 frames. ], batch size: 37, lr: 3.86e-02, grad_scale: 32.0 2024-09-22 13:56:27,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=29708.0, ans=0.004411304347826088 2024-09-22 13:56:46,847 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=29754.666666666668, ans=0.04949747468305833 2024-09-22 13:56:57,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=29801.333333333332, ans=0.004391014492753624 2024-09-22 13:57:02,298 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=29801.333333333332, ans=0.0 2024-09-22 13:57:14,734 INFO [train.py:1198] (2/4) Epoch 2, batch 2500, loss[loss=0.3898, ctc_loss=0.3046, cr_loss=0.4264, over 17149.00 frames. ], tot_loss[loss=0.3587, ctc_loss=0.2738, cr_loss=0.4246, over 3372529.11 frames. ], batch size: 48, lr: 3.85e-02, grad_scale: 32.0 2024-09-22 13:57:18,290 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=29848.0, ans=0.0 2024-09-22 13:57:26,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=29848.0, ans=0.025 2024-09-22 13:57:29,030 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.436e+02 1.845e+02 2.309e+02 3.069e+02 4.385e+02, threshold=4.618e+02, percent-clipped=5.0 2024-09-22 13:57:29,384 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=29894.666666666668, ans=0.125 2024-09-22 13:57:32,480 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=29894.666666666668, ans=0.05 2024-09-22 13:58:15,443 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=8.11 vs. limit=15.0 2024-09-22 13:58:18,057 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=29988.0, ans=0.125 2024-09-22 13:58:36,482 INFO [train.py:1198] (2/4) Epoch 2, batch 2550, loss[loss=0.3806, ctc_loss=0.2899, cr_loss=0.4538, over 17150.00 frames. ], tot_loss[loss=0.3575, ctc_loss=0.2729, cr_loss=0.4232, over 3368942.18 frames. ], batch size: 48, lr: 3.84e-02, grad_scale: 32.0 2024-09-22 13:59:11,878 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=30174.666666666668, ans=0.004309855072463768 2024-09-22 13:59:15,066 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=30174.666666666668, ans=0.2 2024-09-22 13:59:41,610 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=11.04 vs. limit=15.0 2024-09-22 13:59:56,515 INFO [train.py:1198] (2/4) Epoch 2, batch 2600, loss[loss=0.43, ctc_loss=0.3469, cr_loss=0.4154, over 11842.00 frames. ], tot_loss[loss=0.3566, ctc_loss=0.272, cr_loss=0.423, over 3370849.70 frames. ], batch size: 123, lr: 3.84e-02, grad_scale: 32.0 2024-09-22 14:00:06,456 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=30314.666666666668, ans=0.125 2024-09-22 14:00:09,672 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=30314.666666666668, ans=0.125 2024-09-22 14:00:11,073 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.285e+02 1.903e+02 2.207e+02 2.848e+02 4.508e+02, threshold=4.414e+02, percent-clipped=0.0 2024-09-22 14:00:45,430 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=30408.0, ans=0.0 2024-09-22 14:01:09,434 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=30501.333333333332, ans=0.0 2024-09-22 14:01:21,837 INFO [train.py:1198] (2/4) Epoch 2, batch 2650, loss[loss=0.3736, ctc_loss=0.2876, cr_loss=0.4299, over 16924.00 frames. ], tot_loss[loss=0.3566, ctc_loss=0.2719, cr_loss=0.4235, over 3371793.10 frames. ], batch size: 58, lr: 3.83e-02, grad_scale: 32.0 2024-09-22 14:01:23,674 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=30548.0, ans=0.0 2024-09-22 14:01:25,429 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=30548.0, ans=0.0 2024-09-22 14:01:43,418 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=3.44 vs. limit=15.0 2024-09-22 14:01:53,492 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=30594.666666666668, ans=0.025 2024-09-22 14:01:56,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=30641.333333333332, ans=0.0 2024-09-22 14:02:04,669 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=3.62 vs. limit=10.0 2024-09-22 14:02:41,104 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=30734.666666666668, ans=0.0 2024-09-22 14:02:46,374 INFO [train.py:1198] (2/4) Epoch 2, batch 2700, loss[loss=0.3516, ctc_loss=0.2642, cr_loss=0.4368, over 17319.00 frames. ], tot_loss[loss=0.3575, ctc_loss=0.2728, cr_loss=0.4236, over 3365603.50 frames. ], batch size: 46, lr: 3.82e-02, grad_scale: 32.0 2024-09-22 14:03:00,819 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.428e+02 1.814e+02 2.097e+02 2.446e+02 4.164e+02, threshold=4.194e+02, percent-clipped=0.0 2024-09-22 14:03:07,533 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-22 14:03:09,714 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=11.96 vs. limit=15.0 2024-09-22 14:03:16,960 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=30874.666666666668, ans=0.125 2024-09-22 14:03:32,262 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.whiten, num_groups=1, num_channels=192, metric=4.82 vs. limit=12.0 2024-09-22 14:03:39,431 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=30921.333333333332, ans=0.125 2024-09-22 14:03:45,763 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=30921.333333333332, ans=0.0 2024-09-22 14:03:53,680 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=30968.0, ans=0.2 2024-09-22 14:04:01,421 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=30968.0, ans=0.0 2024-09-22 14:04:05,949 INFO [train.py:1198] (2/4) Epoch 2, batch 2750, loss[loss=0.3199, ctc_loss=0.2409, cr_loss=0.3947, over 17115.00 frames. ], tot_loss[loss=0.3582, ctc_loss=0.2732, cr_loss=0.4249, over 3362052.37 frames. ], batch size: 40, lr: 3.82e-02, grad_scale: 32.0 2024-09-22 14:04:10,985 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=31014.666666666668, ans=0.125 2024-09-22 14:04:18,863 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=31014.666666666668, ans=0.0 2024-09-22 14:04:22,221 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=31061.333333333332, ans=0.125 2024-09-22 14:04:38,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=31108.0, ans=0.0 2024-09-22 14:04:49,393 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=31108.0, ans=0.125 2024-09-22 14:05:05,625 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.min_positive, batch_count=31154.666666666668, ans=0.05 2024-09-22 14:05:31,359 INFO [train.py:1198] (2/4) Epoch 2, batch 2800, loss[loss=0.3542, ctc_loss=0.2694, cr_loss=0.4238, over 16994.00 frames. ], tot_loss[loss=0.3571, ctc_loss=0.2724, cr_loss=0.4233, over 3360173.60 frames. ], batch size: 53, lr: 3.81e-02, grad_scale: 32.0 2024-09-22 14:05:33,113 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=31248.0, ans=0.125 2024-09-22 14:05:36,269 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=31248.0, ans=0.125 2024-09-22 14:05:45,623 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.464e+02 1.818e+02 2.147e+02 2.657e+02 4.230e+02, threshold=4.294e+02, percent-clipped=1.0 2024-09-22 14:06:29,178 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=31388.0, ans=0.1 2024-09-22 14:06:53,502 INFO [train.py:1198] (2/4) Epoch 2, batch 2850, loss[loss=0.3032, ctc_loss=0.2283, cr_loss=0.3748, over 17251.00 frames. ], tot_loss[loss=0.3565, ctc_loss=0.2719, cr_loss=0.423, over 3352296.52 frames. ], batch size: 42, lr: 3.80e-02, grad_scale: 32.0 2024-09-22 14:07:12,652 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=31528.0, ans=0.0 2024-09-22 14:07:12,692 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=31528.0, ans=0.125 2024-09-22 14:07:47,538 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.63 vs. limit=6.0 2024-09-22 14:07:48,823 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.11 vs. limit=15.0 2024-09-22 14:08:00,581 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=14.37 vs. limit=22.5 2024-09-22 14:08:06,381 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=31668.0, ans=0.1 2024-09-22 14:08:15,441 INFO [train.py:1198] (2/4) Epoch 2, batch 2900, loss[loss=0.3972, ctc_loss=0.3061, cr_loss=0.4552, over 16915.00 frames. ], tot_loss[loss=0.3562, ctc_loss=0.2716, cr_loss=0.4228, over 3351409.62 frames. ], batch size: 58, lr: 3.80e-02, grad_scale: 32.0 2024-09-22 14:08:20,705 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=31714.666666666668, ans=0.0 2024-09-22 14:08:29,968 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.349e+02 1.781e+02 2.101e+02 2.670e+02 4.501e+02, threshold=4.202e+02, percent-clipped=1.0 2024-09-22 14:08:44,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=31761.333333333332, ans=0.0039649275362318845 2024-09-22 14:08:52,541 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=31808.0, ans=0.125 2024-09-22 14:08:57,451 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=31808.0, ans=0.125 2024-09-22 14:09:03,671 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=31854.666666666668, ans=0.0039446376811594205 2024-09-22 14:09:15,163 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=15.39 vs. limit=15.0 2024-09-22 14:09:35,327 INFO [train.py:1198] (2/4) Epoch 2, batch 2950, loss[loss=0.4021, ctc_loss=0.306, cr_loss=0.4803, over 15138.00 frames. ], tot_loss[loss=0.3557, ctc_loss=0.2712, cr_loss=0.422, over 3349507.94 frames. ], batch size: 90, lr: 3.79e-02, grad_scale: 32.0 2024-09-22 14:09:51,849 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.68 vs. limit=15.0 2024-09-22 14:09:53,140 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=31994.666666666668, ans=0.0 2024-09-22 14:10:02,602 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=31994.666666666668, ans=0.125 2024-09-22 14:10:27,004 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.min_positive, batch_count=32088.0, ans=0.025 2024-09-22 14:10:59,597 INFO [train.py:1198] (2/4) Epoch 2, batch 3000, loss[loss=0.3336, ctc_loss=0.2492, cr_loss=0.4218, over 17191.00 frames. ], tot_loss[loss=0.3555, ctc_loss=0.2709, cr_loss=0.4228, over 3354952.84 frames. ], batch size: 47, lr: 3.79e-02, grad_scale: 32.0 2024-09-22 14:10:59,597 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-22 14:11:15,218 INFO [train.py:1230] (2/4) Epoch 2, validation: loss=0.0967, ctc_loss=0.0967, cr_loss=8.169e-15, over 944034.00 frames. 2024-09-22 14:11:15,219 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-22 14:11:20,374 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=20.72 vs. limit=22.5 2024-09-22 14:11:24,124 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=8.94 vs. limit=10.0 2024-09-22 14:11:29,192 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.469e+02 1.814e+02 2.374e+02 2.965e+02 6.190e+02, threshold=4.748e+02, percent-clipped=4.0 2024-09-22 14:11:32,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=32228.0, ans=0.0 2024-09-22 14:11:41,947 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 14:11:41,955 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=32228.0, ans=0.0 2024-09-22 14:11:44,896 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=32274.666666666668, ans=0.0 2024-09-22 14:12:13,717 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=32321.333333333332, ans=0.125 2024-09-22 14:12:35,204 INFO [train.py:1198] (2/4) Epoch 2, batch 3050, loss[loss=0.3323, ctc_loss=0.2575, cr_loss=0.3742, over 16212.00 frames. ], tot_loss[loss=0.3554, ctc_loss=0.271, cr_loss=0.422, over 3356993.18 frames. ], batch size: 36, lr: 3.78e-02, grad_scale: 32.0 2024-09-22 14:12:35,480 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=32414.666666666668, ans=0.0 2024-09-22 14:12:44,799 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=32414.666666666668, ans=0.0 2024-09-22 14:12:47,908 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=32414.666666666668, ans=0.025 2024-09-22 14:12:52,696 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=32461.333333333332, ans=0.1 2024-09-22 14:13:13,290 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=5.80 vs. limit=10.0 2024-09-22 14:13:20,973 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=22.92 vs. limit=22.5 2024-09-22 14:13:37,499 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 14:13:52,646 INFO [train.py:1198] (2/4) Epoch 2, batch 3100, loss[loss=0.3735, ctc_loss=0.2869, cr_loss=0.4332, over 15840.00 frames. ], tot_loss[loss=0.3554, ctc_loss=0.2711, cr_loss=0.4217, over 3352693.16 frames. ], batch size: 74, lr: 3.77e-02, grad_scale: 32.0 2024-09-22 14:13:57,634 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=32648.0, ans=0.125 2024-09-22 14:13:59,538 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=32648.0, ans=0.003772173913043478 2024-09-22 14:14:10,888 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.354e+02 1.798e+02 2.252e+02 2.859e+02 4.646e+02, threshold=4.504e+02, percent-clipped=0.0 2024-09-22 14:14:14,212 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=32694.666666666668, ans=0.1 2024-09-22 14:14:17,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=32694.666666666668, ans=0.2 2024-09-22 14:14:22,111 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=32694.666666666668, ans=0.125 2024-09-22 14:14:26,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=32741.333333333332, ans=0.025 2024-09-22 14:14:38,974 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=32741.333333333332, ans=0.0 2024-09-22 14:14:46,087 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=10.72 vs. limit=15.0 2024-09-22 14:14:52,843 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=32788.0, ans=0.0 2024-09-22 14:14:53,217 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.75 vs. limit=6.0 2024-09-22 14:15:03,787 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=32834.666666666664, ans=0.025 2024-09-22 14:15:12,719 INFO [train.py:1198] (2/4) Epoch 2, batch 3150, loss[loss=0.3279, ctc_loss=0.2495, cr_loss=0.3918, over 17058.00 frames. ], tot_loss[loss=0.3555, ctc_loss=0.2711, cr_loss=0.4221, over 3359716.25 frames. ], batch size: 46, lr: 3.77e-02, grad_scale: 32.0 2024-09-22 14:15:21,389 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.22 vs. limit=10.0 2024-09-22 14:15:50,774 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=32974.666666666664, ans=0.025 2024-09-22 14:15:55,312 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=32974.666666666664, ans=0.0 2024-09-22 14:16:23,253 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=33068.0, ans=0.0 2024-09-22 14:16:30,738 INFO [train.py:1198] (2/4) Epoch 2, batch 3200, loss[loss=0.3604, ctc_loss=0.2749, cr_loss=0.4274, over 17017.00 frames. ], tot_loss[loss=0.3551, ctc_loss=0.2705, cr_loss=0.4229, over 3363428.04 frames. ], batch size: 51, lr: 3.76e-02, grad_scale: 32.0 2024-09-22 14:16:46,456 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.436e+02 1.721e+02 1.992e+02 2.305e+02 3.966e+02, threshold=3.983e+02, percent-clipped=0.0 2024-09-22 14:17:44,107 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=33301.333333333336, ans=0.0 2024-09-22 14:17:47,230 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=33348.0, ans=0.0 2024-09-22 14:17:48,521 INFO [train.py:1198] (2/4) Epoch 2, batch 3250, loss[loss=0.3679, ctc_loss=0.2755, cr_loss=0.4617, over 17128.00 frames. ], tot_loss[loss=0.3562, ctc_loss=0.2713, cr_loss=0.4249, over 3355010.52 frames. ], batch size: 48, lr: 3.75e-02, grad_scale: 32.0 2024-09-22 14:18:07,315 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=33394.666666666664, ans=0.2 2024-09-22 14:18:25,838 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=33441.333333333336, ans=0.125 2024-09-22 14:18:35,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=33488.0, ans=0.0 2024-09-22 14:19:06,343 INFO [train.py:1198] (2/4) Epoch 2, batch 3300, loss[loss=0.3457, ctc_loss=0.2593, cr_loss=0.432, over 17231.00 frames. ], tot_loss[loss=0.3578, ctc_loss=0.2727, cr_loss=0.4256, over 3348486.60 frames. ], batch size: 47, lr: 3.75e-02, grad_scale: 32.0 2024-09-22 14:19:12,936 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=33581.333333333336, ans=0.1 2024-09-22 14:19:22,038 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.353e+02 1.852e+02 2.253e+02 3.068e+02 5.078e+02, threshold=4.507e+02, percent-clipped=5.0 2024-09-22 14:19:22,412 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=33628.0, ans=0.1 2024-09-22 14:19:25,434 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-22 14:19:44,063 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=33674.666666666664, ans=0.125 2024-09-22 14:19:54,548 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=5.74 vs. limit=10.0 2024-09-22 14:19:55,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=33721.333333333336, ans=0.125 2024-09-22 14:20:13,123 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=33768.0, ans=0.125 2024-09-22 14:20:28,448 INFO [train.py:1198] (2/4) Epoch 2, batch 3350, loss[loss=0.3412, ctc_loss=0.2561, cr_loss=0.4254, over 17197.00 frames. ], tot_loss[loss=0.3565, ctc_loss=0.2713, cr_loss=0.4259, over 3359597.89 frames. ], batch size: 47, lr: 3.74e-02, grad_scale: 32.0 2024-09-22 14:20:42,530 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=33861.333333333336, ans=0.125 2024-09-22 14:20:53,487 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=33861.333333333336, ans=0.125 2024-09-22 14:21:15,203 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=33954.666666666664, ans=0.125 2024-09-22 14:21:20,660 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=21.61 vs. limit=22.5 2024-09-22 14:21:27,760 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=33954.666666666664, ans=0.2 2024-09-22 14:21:34,054 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=34001.333333333336, ans=0.003477971014492753 2024-09-22 14:21:35,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=34001.333333333336, ans=0.125 2024-09-22 14:21:46,546 INFO [train.py:1198] (2/4) Epoch 2, batch 3400, loss[loss=0.3593, ctc_loss=0.2738, cr_loss=0.4276, over 17248.00 frames. ], tot_loss[loss=0.355, ctc_loss=0.27, cr_loss=0.4248, over 3363901.74 frames. ], batch size: 50, lr: 3.74e-02, grad_scale: 32.0 2024-09-22 14:22:00,993 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=34094.666666666664, ans=0.2 2024-09-22 14:22:02,234 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.402e+02 1.761e+02 2.090e+02 2.503e+02 3.941e+02, threshold=4.179e+02, percent-clipped=0.0 2024-09-22 14:22:07,994 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=15.24 vs. limit=15.0 2024-09-22 14:22:08,877 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=34094.666666666664, ans=0.125 2024-09-22 14:22:21,957 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=34141.333333333336, ans=0.125 2024-09-22 14:22:23,500 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=34141.333333333336, ans=0.1 2024-09-22 14:22:28,242 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=34141.333333333336, ans=0.95 2024-09-22 14:23:07,087 INFO [train.py:1198] (2/4) Epoch 2, batch 3450, loss[loss=0.3721, ctc_loss=0.2863, cr_loss=0.429, over 17046.00 frames. ], tot_loss[loss=0.3531, ctc_loss=0.2684, cr_loss=0.4236, over 3367045.16 frames. ], batch size: 56, lr: 3.73e-02, grad_scale: 16.0 2024-09-22 14:23:23,076 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=34328.0, ans=0.2 2024-09-22 14:24:21,493 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=8.66 vs. limit=15.0 2024-09-22 14:24:26,671 INFO [train.py:1198] (2/4) Epoch 2, batch 3500, loss[loss=0.3748, ctc_loss=0.2885, cr_loss=0.4315, over 17314.00 frames. ], tot_loss[loss=0.3528, ctc_loss=0.2684, cr_loss=0.4222, over 3370830.18 frames. ], batch size: 51, lr: 3.72e-02, grad_scale: 16.0 2024-09-22 14:24:36,610 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=34514.666666666664, ans=0.125 2024-09-22 14:24:38,085 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=34514.666666666664, ans=0.125 2024-09-22 14:24:44,121 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.468e+02 1.790e+02 2.190e+02 2.610e+02 4.602e+02, threshold=4.381e+02, percent-clipped=3.0 2024-09-22 14:25:13,966 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=34654.666666666664, ans=0.025 2024-09-22 14:25:14,011 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=34654.666666666664, ans=0.1 2024-09-22 14:25:44,826 INFO [train.py:1198] (2/4) Epoch 2, batch 3550, loss[loss=0.3283, ctc_loss=0.2473, cr_loss=0.4049, over 17156.00 frames. ], tot_loss[loss=0.3528, ctc_loss=0.2684, cr_loss=0.4224, over 3369522.99 frames. ], batch size: 48, lr: 3.72e-02, grad_scale: 16.0 2024-09-22 14:26:03,491 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=34794.666666666664, ans=0.1 2024-09-22 14:26:15,970 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=34841.333333333336, ans=0.125 2024-09-22 14:26:47,376 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=34934.666666666664, ans=0.125 2024-09-22 14:26:58,117 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=34934.666666666664, ans=0.2 2024-09-22 14:27:01,197 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 14:27:02,440 INFO [train.py:1198] (2/4) Epoch 2, batch 3600, loss[loss=0.3424, ctc_loss=0.2614, cr_loss=0.4049, over 17070.00 frames. ], tot_loss[loss=0.3529, ctc_loss=0.2684, cr_loss=0.4226, over 3374007.49 frames. ], batch size: 46, lr: 3.71e-02, grad_scale: 32.0 2024-09-22 14:27:10,972 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.39 vs. limit=6.0 2024-09-22 14:27:19,632 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.393e+02 1.714e+02 2.123e+02 2.581e+02 4.355e+02, threshold=4.245e+02, percent-clipped=0.0 2024-09-22 14:27:38,540 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=35074.666666666664, ans=0.125 2024-09-22 14:28:09,864 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=35168.0, ans=0.125 2024-09-22 14:28:17,646 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=35168.0, ans=0.0032243478260869565 2024-09-22 14:28:20,465 INFO [train.py:1198] (2/4) Epoch 2, batch 3650, loss[loss=0.325, ctc_loss=0.2441, cr_loss=0.405, over 17101.00 frames. ], tot_loss[loss=0.3533, ctc_loss=0.2688, cr_loss=0.4225, over 3376884.27 frames. ], batch size: 43, lr: 3.70e-02, grad_scale: 32.0 2024-09-22 14:28:28,623 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=35214.666666666664, ans=0.125 2024-09-22 14:28:47,519 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=35261.333333333336, ans=0.2 2024-09-22 14:28:56,574 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=35308.0, ans=0.2 2024-09-22 14:29:07,974 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=35354.666666666664, ans=0.0 2024-09-22 14:29:33,520 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=35401.333333333336, ans=0.05 2024-09-22 14:29:40,841 INFO [train.py:1198] (2/4) Epoch 2, batch 3700, loss[loss=0.3512, ctc_loss=0.2649, cr_loss=0.4314, over 17297.00 frames. ], tot_loss[loss=0.3512, ctc_loss=0.2669, cr_loss=0.4215, over 3377340.80 frames. ], batch size: 46, lr: 3.70e-02, grad_scale: 16.0 2024-09-22 14:29:44,400 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.88 vs. limit=15.0 2024-09-22 14:29:45,824 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=35448.0, ans=0.025 2024-09-22 14:29:59,940 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.354e+02 1.763e+02 2.146e+02 2.787e+02 4.998e+02, threshold=4.291e+02, percent-clipped=2.0 2024-09-22 14:30:03,284 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=35494.666666666664, ans=0.125 2024-09-22 14:30:06,578 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=35494.666666666664, ans=0.2 2024-09-22 14:30:15,854 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=35541.333333333336, ans=0.125 2024-09-22 14:30:53,421 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=35634.666666666664, ans=0.125 2024-09-22 14:30:59,290 INFO [train.py:1198] (2/4) Epoch 2, batch 3750, loss[loss=0.2705, ctc_loss=0.2026, cr_loss=0.3394, over 17025.00 frames. ], tot_loss[loss=0.3523, ctc_loss=0.2681, cr_loss=0.4211, over 3351278.57 frames. ], batch size: 39, lr: 3.69e-02, grad_scale: 16.0 2024-09-22 14:31:01,135 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=35681.333333333336, ans=0.025 2024-09-22 14:31:01,235 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 14:31:29,280 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=35774.666666666664, ans=0.125 2024-09-22 14:31:40,134 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=512, metric=23.43 vs. limit=22.5 2024-09-22 14:31:45,871 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=35821.333333333336, ans=0.0 2024-09-22 14:31:55,077 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=35821.333333333336, ans=0.125 2024-09-22 14:32:06,287 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=35868.0, ans=0.2 2024-09-22 14:32:14,281 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=35868.0, ans=0.0 2024-09-22 14:32:14,827 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=9.30 vs. limit=15.0 2024-09-22 14:32:18,707 INFO [train.py:1198] (2/4) Epoch 2, batch 3800, loss[loss=0.3259, ctc_loss=0.2453, cr_loss=0.403, over 17140.00 frames. ], tot_loss[loss=0.3514, ctc_loss=0.2672, cr_loss=0.4207, over 3348631.93 frames. ], batch size: 48, lr: 3.69e-02, grad_scale: 16.0 2024-09-22 14:32:25,893 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=15.36 vs. limit=15.0 2024-09-22 14:32:37,259 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.475e+02 1.782e+02 2.185e+02 2.503e+02 5.708e+02, threshold=4.370e+02, percent-clipped=5.0 2024-09-22 14:32:46,568 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=35961.333333333336, ans=0.025 2024-09-22 14:33:11,497 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=36054.666666666664, ans=0.0 2024-09-22 14:33:23,495 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.min_positive, batch_count=36101.333333333336, ans=0.05 2024-09-22 14:33:28,478 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=14.98 vs. limit=22.5 2024-09-22 14:33:32,806 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=36101.333333333336, ans=0.125 2024-09-22 14:33:32,820 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=36101.333333333336, ans=0.125 2024-09-22 14:33:35,544 INFO [train.py:1198] (2/4) Epoch 2, batch 3850, loss[loss=0.4735, ctc_loss=0.3817, cr_loss=0.4591, over 11831.00 frames. ], tot_loss[loss=0.3551, ctc_loss=0.2709, cr_loss=0.4211, over 3291475.18 frames. ], batch size: 123, lr: 3.68e-02, grad_scale: 16.0 2024-09-22 14:33:43,349 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=36148.0, ans=0.0030113043478260867 2024-09-22 14:34:01,623 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=36194.666666666664, ans=0.1 2024-09-22 14:34:27,176 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=36288.0, ans=0.125 2024-09-22 14:35:27,795 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=36362.666666666664, ans=0.125 2024-09-22 14:35:36,875 INFO [train.py:1198] (2/4) Epoch 3, batch 0, loss[loss=0.4124, ctc_loss=0.3218, cr_loss=0.4534, over 14952.00 frames. ], tot_loss[loss=0.4124, ctc_loss=0.3218, cr_loss=0.4534, over 14952.00 frames. ], batch size: 89, lr: 3.49e-02, grad_scale: 32.0 2024-09-22 14:35:36,876 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-22 14:35:52,229 INFO [train.py:1230] (2/4) Epoch 3, validation: loss=0.1002, ctc_loss=0.1002, cr_loss=7.948e-15, over 944034.00 frames. 2024-09-22 14:35:52,230 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-22 14:36:15,184 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=36409.333333333336, ans=0.125 2024-09-22 14:36:20,866 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.466e+02 1.866e+02 2.182e+02 2.689e+02 4.735e+02, threshold=4.364e+02, percent-clipped=1.0 2024-09-22 14:36:40,452 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=36456.0, ans=0.2 2024-09-22 14:36:51,447 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=36502.666666666664, ans=0.1 2024-09-22 14:37:14,995 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=36549.333333333336, ans=0.2 2024-09-22 14:37:17,946 INFO [train.py:1198] (2/4) Epoch 3, batch 50, loss[loss=0.3448, ctc_loss=0.2617, cr_loss=0.4156, over 17315.00 frames. ], tot_loss[loss=0.3506, ctc_loss=0.2665, cr_loss=0.4202, over 762449.83 frames. ], batch size: 51, lr: 3.49e-02, grad_scale: 32.0 2024-09-22 14:37:31,208 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=36596.0, ans=0.2 2024-09-22 14:37:32,956 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=36596.0, ans=0.0029139130434782615 2024-09-22 14:37:52,594 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=15.43 vs. limit=15.0 2024-09-22 14:38:00,615 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=10.17 vs. limit=15.0 2024-09-22 14:38:30,034 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=36782.666666666664, ans=0.2 2024-09-22 14:38:35,557 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=36782.666666666664, ans=0.125 2024-09-22 14:38:40,466 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=36829.333333333336, ans=0.0 2024-09-22 14:38:41,848 INFO [train.py:1198] (2/4) Epoch 3, batch 100, loss[loss=0.3465, ctc_loss=0.258, cr_loss=0.4424, over 17058.00 frames. ], tot_loss[loss=0.3487, ctc_loss=0.2649, cr_loss=0.419, over 1328486.40 frames. ], batch size: 46, lr: 3.48e-02, grad_scale: 32.0 2024-09-22 14:39:04,319 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 14:39:07,120 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.389e+02 1.767e+02 2.113e+02 2.664e+02 5.595e+02, threshold=4.227e+02, percent-clipped=3.0 2024-09-22 14:39:38,024 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.27 vs. limit=12.0 2024-09-22 14:40:01,179 INFO [train.py:1198] (2/4) Epoch 3, batch 150, loss[loss=0.3603, ctc_loss=0.2721, cr_loss=0.4409, over 16511.00 frames. ], tot_loss[loss=0.3454, ctc_loss=0.2616, cr_loss=0.4193, over 1787957.33 frames. ], batch size: 66, lr: 3.47e-02, grad_scale: 32.0 2024-09-22 14:40:01,775 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.88 vs. limit=15.0 2024-09-22 14:40:14,345 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=37062.666666666664, ans=0.1 2024-09-22 14:40:33,922 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.52 vs. limit=15.0 2024-09-22 14:40:44,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=37156.0, ans=0.2 2024-09-22 14:41:22,659 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.90 vs. limit=6.0 2024-09-22 14:41:23,508 INFO [train.py:1198] (2/4) Epoch 3, batch 200, loss[loss=0.2975, ctc_loss=0.2225, cr_loss=0.375, over 17272.00 frames. ], tot_loss[loss=0.3448, ctc_loss=0.2608, cr_loss=0.4196, over 2139614.41 frames. ], batch size: 42, lr: 3.47e-02, grad_scale: 32.0 2024-09-22 14:41:33,169 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=37296.0, ans=0.025 2024-09-22 14:41:41,011 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.98 vs. limit=15.0 2024-09-22 14:41:51,077 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.278e+02 1.679e+02 1.968e+02 2.454e+02 4.058e+02, threshold=3.935e+02, percent-clipped=0.0 2024-09-22 14:41:56,212 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=37389.333333333336, ans=0.125 2024-09-22 14:42:08,561 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=37389.333333333336, ans=0.002741449275362318 2024-09-22 14:42:27,191 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=37436.0, ans=0.125 2024-09-22 14:42:28,758 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=37436.0, ans=0.125 2024-09-22 14:42:35,739 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=10.96 vs. limit=15.0 2024-09-22 14:42:50,682 INFO [train.py:1198] (2/4) Epoch 3, batch 250, loss[loss=0.4155, ctc_loss=0.3175, cr_loss=0.4897, over 16818.00 frames. ], tot_loss[loss=0.3472, ctc_loss=0.263, cr_loss=0.4209, over 2410794.49 frames. ], batch size: 61, lr: 3.46e-02, grad_scale: 32.0 2024-09-22 14:42:58,957 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=37529.333333333336, ans=0.125 2024-09-22 14:43:19,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=37576.0, ans=0.125 2024-09-22 14:43:19,340 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=37576.0, ans=0.125 2024-09-22 14:43:20,876 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=37622.666666666664, ans=0.125 2024-09-22 14:43:30,427 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=37622.666666666664, ans=0.2 2024-09-22 14:43:42,607 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=37669.333333333336, ans=0.0 2024-09-22 14:43:50,990 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=12.43 vs. limit=15.0 2024-09-22 14:44:01,339 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=37716.0, ans=0.0 2024-09-22 14:44:12,443 INFO [train.py:1198] (2/4) Epoch 3, batch 300, loss[loss=0.3248, ctc_loss=0.2441, cr_loss=0.4035, over 17288.00 frames. ], tot_loss[loss=0.3471, ctc_loss=0.2626, cr_loss=0.4222, over 2615762.32 frames. ], batch size: 49, lr: 3.46e-02, grad_scale: 32.0 2024-09-22 14:44:29,955 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=37809.333333333336, ans=0.125 2024-09-22 14:44:34,791 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=37809.333333333336, ans=0.125 2024-09-22 14:44:37,527 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.369e+02 1.687e+02 1.987e+02 2.495e+02 5.356e+02, threshold=3.975e+02, percent-clipped=4.0 2024-09-22 14:44:52,611 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=12.01 vs. limit=15.0 2024-09-22 14:45:13,067 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=7.01 vs. limit=15.0 2024-09-22 14:45:16,045 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=37949.333333333336, ans=0.125 2024-09-22 14:45:31,397 INFO [train.py:1198] (2/4) Epoch 3, batch 350, loss[loss=0.3677, ctc_loss=0.2867, cr_loss=0.4055, over 16508.00 frames. ], tot_loss[loss=0.3472, ctc_loss=0.2629, cr_loss=0.4218, over 2773637.64 frames. ], batch size: 66, lr: 3.45e-02, grad_scale: 32.0 2024-09-22 14:45:50,960 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=38042.666666666664, ans=0.2 2024-09-22 14:45:53,888 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=38042.666666666664, ans=0.05 2024-09-22 14:46:33,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=38136.0, ans=0.125 2024-09-22 14:46:46,875 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=11.04 vs. limit=15.0 2024-09-22 14:46:49,689 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=38182.666666666664, ans=0.025 2024-09-22 14:46:57,023 INFO [train.py:1198] (2/4) Epoch 3, batch 400, loss[loss=0.3387, ctc_loss=0.2521, cr_loss=0.4331, over 17309.00 frames. ], tot_loss[loss=0.3457, ctc_loss=0.2616, cr_loss=0.4205, over 2905498.16 frames. ], batch size: 49, lr: 3.45e-02, grad_scale: 32.0 2024-09-22 14:46:57,284 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=38229.333333333336, ans=0.05 2024-09-22 14:47:02,091 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=38229.333333333336, ans=0.2 2024-09-22 14:47:08,452 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=38229.333333333336, ans=0.0 2024-09-22 14:47:20,112 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=38276.0, ans=0.0 2024-09-22 14:47:25,983 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.328e+02 1.690e+02 1.971e+02 2.820e+02 5.296e+02, threshold=3.942e+02, percent-clipped=6.0 2024-09-22 14:47:34,267 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=38322.666666666664, ans=0.125 2024-09-22 14:47:58,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=38369.333333333336, ans=0.0 2024-09-22 14:48:05,258 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.92 vs. limit=15.0 2024-09-22 14:48:19,965 INFO [train.py:1198] (2/4) Epoch 3, batch 450, loss[loss=0.352, ctc_loss=0.2591, cr_loss=0.4644, over 17195.00 frames. ], tot_loss[loss=0.345, ctc_loss=0.261, cr_loss=0.4203, over 3005250.71 frames. ], batch size: 50, lr: 3.44e-02, grad_scale: 32.0 2024-09-22 14:48:21,036 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=3.56 vs. limit=15.0 2024-09-22 14:48:25,801 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.61 vs. limit=12.0 2024-09-22 14:49:19,751 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=38602.666666666664, ans=0.125 2024-09-22 14:49:22,970 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=38602.666666666664, ans=0.125 2024-09-22 14:49:37,401 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=38649.333333333336, ans=0.125 2024-09-22 14:49:41,802 INFO [train.py:1198] (2/4) Epoch 3, batch 500, loss[loss=0.3151, ctc_loss=0.2379, cr_loss=0.3859, over 17244.00 frames. ], tot_loss[loss=0.3453, ctc_loss=0.2611, cr_loss=0.4209, over 3077591.56 frames. ], batch size: 44, lr: 3.43e-02, grad_scale: 32.0 2024-09-22 14:50:07,133 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.349e+02 1.795e+02 2.310e+02 2.743e+02 4.759e+02, threshold=4.620e+02, percent-clipped=4.0 2024-09-22 14:50:12,361 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=38789.333333333336, ans=0.1 2024-09-22 14:50:12,375 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=38789.333333333336, ans=0.2 2024-09-22 14:50:26,497 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=38789.333333333336, ans=0.125 2024-09-22 14:50:52,473 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=5.07 vs. limit=12.0 2024-09-22 14:50:52,508 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=512, metric=22.58 vs. limit=22.5 2024-09-22 14:51:01,266 INFO [train.py:1198] (2/4) Epoch 3, batch 550, loss[loss=0.3325, ctc_loss=0.2511, cr_loss=0.407, over 17212.00 frames. ], tot_loss[loss=0.3469, ctc_loss=0.2624, cr_loss=0.4227, over 3136247.64 frames. ], batch size: 47, lr: 3.43e-02, grad_scale: 32.0 2024-09-22 14:51:20,315 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=38976.0, ans=0.025 2024-09-22 14:51:25,197 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer_ff2.min_abs, batch_count=38976.0, ans=0.1 2024-09-22 14:51:28,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=38976.0, ans=0.125 2024-09-22 14:51:37,817 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=39022.666666666664, ans=0.125 2024-09-22 14:51:39,493 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=39022.666666666664, ans=0.125 2024-09-22 14:51:40,968 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=39022.666666666664, ans=0.1 2024-09-22 14:51:54,601 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=39069.333333333336, ans=0.125 2024-09-22 14:51:56,837 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=15.74 vs. limit=15.0 2024-09-22 14:52:10,586 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=39116.0, ans=0.2 2024-09-22 14:52:28,850 INFO [train.py:1198] (2/4) Epoch 3, batch 600, loss[loss=0.3683, ctc_loss=0.2785, cr_loss=0.449, over 17202.00 frames. ], tot_loss[loss=0.3455, ctc_loss=0.2613, cr_loss=0.4212, over 3181340.79 frames. ], batch size: 55, lr: 3.42e-02, grad_scale: 32.0 2024-09-22 14:52:33,985 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=39162.666666666664, ans=0.125 2024-09-22 14:52:53,277 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=39209.333333333336, ans=0.0023457971014492754 2024-09-22 14:52:54,491 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.307e+02 1.738e+02 2.025e+02 2.578e+02 4.577e+02, threshold=4.049e+02, percent-clipped=0.0 2024-09-22 14:53:01,044 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=39256.0, ans=0.125 2024-09-22 14:53:01,149 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=39256.0, ans=0.2 2024-09-22 14:53:14,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=39256.0, ans=0.125 2024-09-22 14:53:25,013 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=15.62 vs. limit=22.5 2024-09-22 14:53:35,751 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=39349.333333333336, ans=0.0 2024-09-22 14:53:40,327 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=39349.333333333336, ans=0.1 2024-09-22 14:53:45,537 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.41 vs. limit=10.0 2024-09-22 14:53:50,600 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.30 vs. limit=15.0 2024-09-22 14:53:51,255 INFO [train.py:1198] (2/4) Epoch 3, batch 650, loss[loss=0.2826, ctc_loss=0.2103, cr_loss=0.3612, over 17035.00 frames. ], tot_loss[loss=0.3476, ctc_loss=0.263, cr_loss=0.4228, over 3203664.86 frames. ], batch size: 39, lr: 3.42e-02, grad_scale: 32.0 2024-09-22 14:53:56,349 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=39396.0, ans=0.0 2024-09-22 14:53:57,006 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=21.05 vs. limit=22.5 2024-09-22 14:54:13,030 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.60 vs. limit=15.0 2024-09-22 14:54:28,348 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=39489.333333333336, ans=0.0 2024-09-22 14:54:36,086 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=39489.333333333336, ans=0.0022849275362318835 2024-09-22 14:55:10,516 INFO [train.py:1198] (2/4) Epoch 3, batch 700, loss[loss=0.4307, ctc_loss=0.3446, cr_loss=0.4308, over 11796.00 frames. ], tot_loss[loss=0.3474, ctc_loss=0.2629, cr_loss=0.4225, over 3225865.01 frames. ], batch size: 123, lr: 3.41e-02, grad_scale: 32.0 2024-09-22 14:55:12,536 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=39629.333333333336, ans=0.0 2024-09-22 14:55:17,732 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.08 vs. limit=10.0 2024-09-22 14:55:35,724 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.431e+02 1.707e+02 1.931e+02 2.238e+02 4.906e+02, threshold=3.863e+02, percent-clipped=1.0 2024-09-22 14:55:44,615 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=5.14 vs. limit=6.0 2024-09-22 14:56:26,754 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=3.99 vs. limit=15.0 2024-09-22 14:56:32,764 INFO [train.py:1198] (2/4) Epoch 3, batch 750, loss[loss=0.3339, ctc_loss=0.252, cr_loss=0.4091, over 17080.00 frames. ], tot_loss[loss=0.3457, ctc_loss=0.2613, cr_loss=0.422, over 3251418.34 frames. ], batch size: 46, lr: 3.41e-02, grad_scale: 32.0 2024-09-22 14:56:50,197 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=7.52 vs. limit=15.0 2024-09-22 14:57:05,876 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=39956.0, ans=0.025 2024-09-22 14:57:44,616 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=40049.333333333336, ans=0.0021631884057971007 2024-09-22 14:57:57,149 INFO [train.py:1198] (2/4) Epoch 3, batch 800, loss[loss=0.3721, ctc_loss=0.2851, cr_loss=0.4346, over 16855.00 frames. ], tot_loss[loss=0.3446, ctc_loss=0.2603, cr_loss=0.4215, over 3275694.39 frames. ], batch size: 58, lr: 3.40e-02, grad_scale: 32.0 2024-09-22 14:57:59,195 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 14:58:00,805 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=40096.0, ans=0.125 2024-09-22 14:58:08,831 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.31 vs. limit=15.0 2024-09-22 14:58:13,442 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=40142.666666666664, ans=0.125 2024-09-22 14:58:14,998 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=40142.666666666664, ans=0.125 2024-09-22 14:58:25,181 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.263e+02 1.611e+02 1.885e+02 2.376e+02 4.057e+02, threshold=3.771e+02, percent-clipped=1.0 2024-09-22 14:58:30,303 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=40189.333333333336, ans=0.0021327536231884048 2024-09-22 14:59:11,133 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=40282.666666666664, ans=0.07 2024-09-22 14:59:17,524 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=40329.333333333336, ans=0.025 2024-09-22 14:59:18,681 INFO [train.py:1198] (2/4) Epoch 3, batch 850, loss[loss=0.3135, ctc_loss=0.2271, cr_loss=0.4321, over 17293.00 frames. ], tot_loss[loss=0.3445, ctc_loss=0.2602, cr_loss=0.4216, over 3298394.42 frames. ], batch size: 46, lr: 3.39e-02, grad_scale: 32.0 2024-09-22 14:59:33,247 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=40376.0, ans=0.07 2024-09-22 14:59:43,139 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.47 vs. limit=15.0 2024-09-22 14:59:46,548 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.44 vs. limit=10.0 2024-09-22 15:00:19,374 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=40469.333333333336, ans=0.125 2024-09-22 15:00:37,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=40562.666666666664, ans=0.1 2024-09-22 15:00:38,409 INFO [train.py:1198] (2/4) Epoch 3, batch 900, loss[loss=0.369, ctc_loss=0.2732, cr_loss=0.479, over 16727.00 frames. ], tot_loss[loss=0.3438, ctc_loss=0.2595, cr_loss=0.4217, over 3314636.53 frames. ], batch size: 61, lr: 3.39e-02, grad_scale: 32.0 2024-09-22 15:00:41,957 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 15:00:43,910 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=12.56 vs. limit=15.0 2024-09-22 15:01:06,321 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.272e+02 1.646e+02 1.943e+02 2.321e+02 3.880e+02, threshold=3.887e+02, percent-clipped=1.0 2024-09-22 15:01:17,843 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=40656.0, ans=0.1 2024-09-22 15:01:40,374 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=19.34 vs. limit=22.5 2024-09-22 15:01:49,437 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=40749.333333333336, ans=0.125 2024-09-22 15:02:03,456 INFO [train.py:1198] (2/4) Epoch 3, batch 950, loss[loss=0.3652, ctc_loss=0.281, cr_loss=0.4207, over 17051.00 frames. ], tot_loss[loss=0.3421, ctc_loss=0.2579, cr_loss=0.4205, over 3320180.85 frames. ], batch size: 56, lr: 3.38e-02, grad_scale: 32.0 2024-09-22 15:02:04,045 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=12.98 vs. limit=15.0 2024-09-22 15:02:29,374 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=13.83 vs. limit=15.0 2024-09-22 15:02:52,872 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=40936.0, ans=0.001970434782608695 2024-09-22 15:03:14,513 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=7.48 vs. limit=15.0 2024-09-22 15:03:28,683 INFO [train.py:1198] (2/4) Epoch 3, batch 1000, loss[loss=0.3166, ctc_loss=0.236, cr_loss=0.4033, over 17138.00 frames. ], tot_loss[loss=0.3405, ctc_loss=0.2568, cr_loss=0.4186, over 3324294.45 frames. ], batch size: 48, lr: 3.38e-02, grad_scale: 32.0 2024-09-22 15:03:30,637 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=41029.333333333336, ans=0.125 2024-09-22 15:03:54,234 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.429e+02 1.784e+02 2.139e+02 2.624e+02 4.654e+02, threshold=4.278e+02, percent-clipped=1.0 2024-09-22 15:04:11,995 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-22 15:04:27,645 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=41169.333333333336, ans=0.07 2024-09-22 15:04:48,223 INFO [train.py:1198] (2/4) Epoch 3, batch 1050, loss[loss=0.308, ctc_loss=0.2303, cr_loss=0.3884, over 16924.00 frames. ], tot_loss[loss=0.3412, ctc_loss=0.2573, cr_loss=0.4195, over 3331415.46 frames. ], batch size: 42, lr: 3.37e-02, grad_scale: 32.0 2024-09-22 15:04:48,573 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=41262.666666666664, ans=0.1 2024-09-22 15:04:56,247 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=41262.666666666664, ans=0.0 2024-09-22 15:05:20,464 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=41356.0, ans=0.125 2024-09-22 15:05:41,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=41402.666666666664, ans=0.125 2024-09-22 15:06:10,764 INFO [train.py:1198] (2/4) Epoch 3, batch 1100, loss[loss=0.3858, ctc_loss=0.2938, cr_loss=0.46, over 15075.00 frames. ], tot_loss[loss=0.34, ctc_loss=0.2562, cr_loss=0.4191, over 3335302.71 frames. ], batch size: 89, lr: 3.37e-02, grad_scale: 32.0 2024-09-22 15:06:16,050 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=11.88 vs. limit=15.0 2024-09-22 15:06:33,168 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=41542.666666666664, ans=0.001838550724637681 2024-09-22 15:06:35,964 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.301e+02 1.612e+02 1.917e+02 2.437e+02 4.278e+02, threshold=3.834e+02, percent-clipped=2.0 2024-09-22 15:06:49,715 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=41589.333333333336, ans=0.1 2024-09-22 15:06:58,119 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.26 vs. limit=12.0 2024-09-22 15:07:00,841 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=41636.0, ans=0.125 2024-09-22 15:07:00,874 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=41636.0, ans=0.125 2024-09-22 15:07:19,691 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=41682.666666666664, ans=0.125 2024-09-22 15:07:29,574 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=41682.666666666664, ans=0.125 2024-09-22 15:07:35,513 INFO [train.py:1198] (2/4) Epoch 3, batch 1150, loss[loss=0.3693, ctc_loss=0.2756, cr_loss=0.4687, over 17234.00 frames. ], tot_loss[loss=0.3383, ctc_loss=0.2546, cr_loss=0.4185, over 3347134.40 frames. ], batch size: 55, lr: 3.36e-02, grad_scale: 32.0 2024-09-22 15:08:08,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=41822.666666666664, ans=0.2 2024-09-22 15:08:12,828 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.min_positive, batch_count=41822.666666666664, ans=0.05 2024-09-22 15:08:36,487 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=41869.333333333336, ans=0.0 2024-09-22 15:08:57,144 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-22 15:08:58,268 INFO [train.py:1198] (2/4) Epoch 3, batch 1200, loss[loss=0.3485, ctc_loss=0.2622, cr_loss=0.4313, over 17008.00 frames. ], tot_loss[loss=0.3378, ctc_loss=0.2542, cr_loss=0.4179, over 3353345.36 frames. ], batch size: 44, lr: 3.36e-02, grad_scale: 32.0 2024-09-22 15:09:18,108 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=512, metric=2.91 vs. limit=15.0 2024-09-22 15:09:23,540 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.305e+02 1.644e+02 1.881e+02 2.304e+02 4.141e+02, threshold=3.762e+02, percent-clipped=3.0 2024-09-22 15:09:46,529 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.36 vs. limit=15.0 2024-09-22 15:09:52,477 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=42102.666666666664, ans=0.07 2024-09-22 15:10:17,333 INFO [train.py:1198] (2/4) Epoch 3, batch 1250, loss[loss=0.3175, ctc_loss=0.2391, cr_loss=0.3921, over 16971.00 frames. ], tot_loss[loss=0.3384, ctc_loss=0.2546, cr_loss=0.4187, over 3357268.04 frames. ], batch size: 42, lr: 3.35e-02, grad_scale: 32.0 2024-09-22 15:10:17,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=42196.0, ans=0.0016965217391304351 2024-09-22 15:10:32,164 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=42242.666666666664, ans=0.125 2024-09-22 15:10:58,423 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=42289.333333333336, ans=0.0016762318840579712 2024-09-22 15:11:17,866 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=8.60 vs. limit=15.0 2024-09-22 15:11:41,929 INFO [train.py:1198] (2/4) Epoch 3, batch 1300, loss[loss=0.3892, ctc_loss=0.3005, cr_loss=0.4433, over 17142.00 frames. ], tot_loss[loss=0.3367, ctc_loss=0.2532, cr_loss=0.4174, over 3364776.47 frames. ], batch size: 48, lr: 3.34e-02, grad_scale: 32.0 2024-09-22 15:12:04,333 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=42476.0, ans=0.125 2024-09-22 15:12:05,807 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=42476.0, ans=0.1 2024-09-22 15:12:09,822 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.394e+02 1.777e+02 1.981e+02 2.466e+02 4.544e+02, threshold=3.962e+02, percent-clipped=3.0 2024-09-22 15:12:38,700 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=42569.333333333336, ans=0.0016153623188405793 2024-09-22 15:12:51,767 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=21.04 vs. limit=22.5 2024-09-22 15:13:06,078 INFO [train.py:1198] (2/4) Epoch 3, batch 1350, loss[loss=0.3243, ctc_loss=0.2494, cr_loss=0.3742, over 17044.00 frames. ], tot_loss[loss=0.3382, ctc_loss=0.2546, cr_loss=0.4182, over 3355365.21 frames. ], batch size: 44, lr: 3.34e-02, grad_scale: 32.0 2024-09-22 15:13:44,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=42756.0, ans=0.0015747826086956532 2024-09-22 15:14:21,495 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=42849.333333333336, ans=0.125 2024-09-22 15:14:25,893 INFO [train.py:1198] (2/4) Epoch 3, batch 1400, loss[loss=0.3051, ctc_loss=0.2325, cr_loss=0.363, over 17259.00 frames. ], tot_loss[loss=0.3379, ctc_loss=0.2544, cr_loss=0.4175, over 3355790.14 frames. ], batch size: 42, lr: 3.33e-02, grad_scale: 32.0 2024-09-22 15:14:26,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=42896.0, ans=0.1 2024-09-22 15:14:51,780 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.368e+02 1.674e+02 1.940e+02 2.218e+02 3.917e+02, threshold=3.881e+02, percent-clipped=0.0 2024-09-22 15:15:08,010 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=42989.333333333336, ans=0.125 2024-09-22 15:15:34,032 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=43082.666666666664, ans=0.125 2024-09-22 15:15:48,006 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 15:15:48,032 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=43129.333333333336, ans=0.1 2024-09-22 15:15:49,145 INFO [train.py:1198] (2/4) Epoch 3, batch 1450, loss[loss=0.3719, ctc_loss=0.2777, cr_loss=0.4714, over 17029.00 frames. ], tot_loss[loss=0.3401, ctc_loss=0.2561, cr_loss=0.4198, over 3350551.92 frames. ], batch size: 56, lr: 3.33e-02, grad_scale: 32.0 2024-09-22 15:15:49,447 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=43129.333333333336, ans=0.125 2024-09-22 15:15:52,749 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=43129.333333333336, ans=0.125 2024-09-22 15:15:59,237 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=43129.333333333336, ans=0.0 2024-09-22 15:16:04,024 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=43176.0, ans=0.2 2024-09-22 15:16:16,926 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=43176.0, ans=0.125 2024-09-22 15:16:23,205 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=43222.666666666664, ans=0.125 2024-09-22 15:17:10,702 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.min_positive, batch_count=43316.0, ans=0.025 2024-09-22 15:17:12,965 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=15.56 vs. limit=15.0 2024-09-22 15:17:13,618 INFO [train.py:1198] (2/4) Epoch 3, batch 1500, loss[loss=0.3305, ctc_loss=0.2468, cr_loss=0.4184, over 17170.00 frames. ], tot_loss[loss=0.3401, ctc_loss=0.2563, cr_loss=0.4187, over 3340883.33 frames. ], batch size: 45, lr: 3.32e-02, grad_scale: 32.0 2024-09-22 15:17:21,973 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=43362.666666666664, ans=0.0 2024-09-22 15:17:38,773 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.398e+02 1.638e+02 2.002e+02 2.354e+02 3.823e+02, threshold=4.005e+02, percent-clipped=0.0 2024-09-22 15:18:06,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=43502.666666666664, ans=0.0 2024-09-22 15:18:35,371 INFO [train.py:1198] (2/4) Epoch 3, batch 1550, loss[loss=0.328, ctc_loss=0.2458, cr_loss=0.4111, over 17021.00 frames. ], tot_loss[loss=0.3387, ctc_loss=0.2552, cr_loss=0.4178, over 3339312.45 frames. ], batch size: 44, lr: 3.32e-02, grad_scale: 32.0 2024-09-22 15:19:01,340 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=43642.666666666664, ans=0.0 2024-09-22 15:19:55,280 INFO [train.py:1198] (2/4) Epoch 3, batch 1600, loss[loss=0.3588, ctc_loss=0.2706, cr_loss=0.4413, over 17037.00 frames. ], tot_loss[loss=0.337, ctc_loss=0.2538, cr_loss=0.4161, over 3339909.98 frames. ], batch size: 52, lr: 3.31e-02, grad_scale: 32.0 2024-09-22 15:20:03,987 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten.whitening_limit, batch_count=43829.333333333336, ans=15.0 2024-09-22 15:20:06,635 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.min_positive, batch_count=43829.333333333336, ans=0.025 2024-09-22 15:20:20,844 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.341e+02 1.634e+02 1.960e+02 2.382e+02 4.201e+02, threshold=3.920e+02, percent-clipped=2.0 2024-09-22 15:20:39,352 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=43922.666666666664, ans=0.125 2024-09-22 15:20:47,119 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.min_positive, batch_count=43969.333333333336, ans=0.025 2024-09-22 15:21:00,058 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=44016.0, ans=0.125 2024-09-22 15:21:16,292 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.66 vs. limit=6.0 2024-09-22 15:21:17,086 INFO [train.py:1198] (2/4) Epoch 3, batch 1650, loss[loss=0.3272, ctc_loss=0.249, cr_loss=0.3909, over 16963.00 frames. ], tot_loss[loss=0.3353, ctc_loss=0.2523, cr_loss=0.4152, over 3348949.90 frames. ], batch size: 42, lr: 3.31e-02, grad_scale: 32.0 2024-09-22 15:21:42,070 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=44109.333333333336, ans=0.2 2024-09-22 15:21:50,909 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module1.whiten, num_groups=1, num_channels=192, metric=8.46 vs. limit=15.0 2024-09-22 15:21:53,250 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=44156.0, ans=0.125 2024-09-22 15:21:56,472 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=44156.0, ans=0.09899494936611666 2024-09-22 15:22:34,427 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=44249.333333333336, ans=0.125 2024-09-22 15:22:41,997 INFO [train.py:1198] (2/4) Epoch 3, batch 1700, loss[loss=0.3101, ctc_loss=0.237, cr_loss=0.3652, over 17033.00 frames. ], tot_loss[loss=0.3374, ctc_loss=0.2541, cr_loss=0.4169, over 3342958.42 frames. ], batch size: 44, lr: 3.30e-02, grad_scale: 32.0 2024-09-22 15:23:09,724 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.265e+02 1.715e+02 2.359e+02 2.941e+02 4.631e+02, threshold=4.717e+02, percent-clipped=4.0 2024-09-22 15:23:59,333 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=44482.666666666664, ans=0.0 2024-09-22 15:24:03,723 INFO [train.py:1198] (2/4) Epoch 3, batch 1750, loss[loss=0.2831, ctc_loss=0.2063, cr_loss=0.3837, over 17053.00 frames. ], tot_loss[loss=0.338, ctc_loss=0.2545, cr_loss=0.4178, over 3351869.11 frames. ], batch size: 39, lr: 3.30e-02, grad_scale: 32.0 2024-09-22 15:24:17,584 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.whiten, num_groups=1, num_channels=192, metric=5.24 vs. limit=12.0 2024-09-22 15:24:40,537 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=44622.666666666664, ans=0.125 2024-09-22 15:25:25,619 INFO [train.py:1198] (2/4) Epoch 3, batch 1800, loss[loss=0.3042, ctc_loss=0.2267, cr_loss=0.3878, over 17260.00 frames. ], tot_loss[loss=0.3397, ctc_loss=0.256, cr_loss=0.4185, over 3331271.66 frames. ], batch size: 44, lr: 3.29e-02, grad_scale: 64.0 2024-09-22 15:25:53,003 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.338e+02 1.793e+02 2.251e+02 2.697e+02 4.483e+02, threshold=4.502e+02, percent-clipped=0.0 2024-09-22 15:26:07,521 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=44856.0, ans=0.125 2024-09-22 15:26:34,909 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=44949.333333333336, ans=0.125 2024-09-22 15:26:46,242 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=44996.0, ans=0.125 2024-09-22 15:26:47,445 INFO [train.py:1198] (2/4) Epoch 3, batch 1850, loss[loss=0.3993, ctc_loss=0.3055, cr_loss=0.469, over 17295.00 frames. ], tot_loss[loss=0.3398, ctc_loss=0.2559, cr_loss=0.4199, over 3346228.80 frames. ], batch size: 49, lr: 3.29e-02, grad_scale: 32.0 2024-09-22 15:27:53,276 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=45136.0, ans=0.001057391304347826 2024-09-22 15:28:12,189 INFO [train.py:1198] (2/4) Epoch 3, batch 1900, loss[loss=0.2884, ctc_loss=0.2171, cr_loss=0.3563, over 16693.00 frames. ], tot_loss[loss=0.3384, ctc_loss=0.2547, cr_loss=0.4185, over 3353988.73 frames. ], batch size: 37, lr: 3.28e-02, grad_scale: 32.0 2024-09-22 15:28:34,573 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=45276.0, ans=0.0 2024-09-22 15:28:38,955 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.328e+02 1.740e+02 2.164e+02 2.812e+02 4.193e+02, threshold=4.328e+02, percent-clipped=0.0 2024-09-22 15:29:14,993 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.16 vs. limit=6.0 2024-09-22 15:29:27,687 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=45416.0, ans=0.0009965217391304342 2024-09-22 15:29:29,154 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=45416.0, ans=0.125 2024-09-22 15:29:32,029 INFO [train.py:1198] (2/4) Epoch 3, batch 1950, loss[loss=0.3504, ctc_loss=0.2595, cr_loss=0.4549, over 17133.00 frames. ], tot_loss[loss=0.3374, ctc_loss=0.2537, cr_loss=0.4185, over 3361344.17 frames. ], batch size: 48, lr: 3.27e-02, grad_scale: 32.0 2024-09-22 15:29:33,931 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=45462.666666666664, ans=0.000986376811594204 2024-09-22 15:29:35,491 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=45462.666666666664, ans=0.0 2024-09-22 15:30:06,085 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=16.51 vs. limit=22.5 2024-09-22 15:30:11,205 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten.whitening_limit, batch_count=45556.0, ans=15.0 2024-09-22 15:30:12,347 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 15:30:27,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=45602.666666666664, ans=0.1 2024-09-22 15:30:52,084 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.36 vs. limit=15.0 2024-09-22 15:30:54,280 INFO [train.py:1198] (2/4) Epoch 3, batch 2000, loss[loss=0.2938, ctc_loss=0.2181, cr_loss=0.3784, over 17242.00 frames. ], tot_loss[loss=0.3366, ctc_loss=0.2528, cr_loss=0.4186, over 3361780.21 frames. ], batch size: 42, lr: 3.27e-02, grad_scale: 32.0 2024-09-22 15:31:04,015 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=45696.0, ans=0.125 2024-09-22 15:31:15,561 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=45742.666666666664, ans=0.0 2024-09-22 15:31:23,957 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.430e+02 1.768e+02 1.993e+02 2.472e+02 5.161e+02, threshold=3.986e+02, percent-clipped=2.0 2024-09-22 15:31:27,621 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 15:31:29,241 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=45789.333333333336, ans=0.07 2024-09-22 15:31:29,602 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=3.20 vs. limit=15.0 2024-09-22 15:31:49,385 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.57 vs. limit=6.0 2024-09-22 15:32:02,640 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=4.88 vs. limit=12.0 2024-09-22 15:32:19,441 INFO [train.py:1198] (2/4) Epoch 3, batch 2050, loss[loss=0.343, ctc_loss=0.2561, cr_loss=0.4346, over 17020.00 frames. ], tot_loss[loss=0.3362, ctc_loss=0.2526, cr_loss=0.4179, over 3356517.01 frames. ], batch size: 52, lr: 3.26e-02, grad_scale: 32.0 2024-09-22 15:32:27,751 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=45929.333333333336, ans=0.125 2024-09-22 15:33:03,380 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=46022.666666666664, ans=0.025 2024-09-22 15:33:04,195 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=12.26 vs. limit=15.0 2024-09-22 15:33:13,617 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=11.37 vs. limit=15.0 2024-09-22 15:33:24,748 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=4.83 vs. limit=12.0 2024-09-22 15:33:41,356 INFO [train.py:1198] (2/4) Epoch 3, batch 2100, loss[loss=0.3324, ctc_loss=0.2441, cr_loss=0.4417, over 17100.00 frames. ], tot_loss[loss=0.3358, ctc_loss=0.2523, cr_loss=0.4175, over 3350328.23 frames. ], batch size: 49, lr: 3.26e-02, grad_scale: 32.0 2024-09-22 15:33:48,123 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=46162.666666666664, ans=0.125 2024-09-22 15:33:54,428 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=46162.666666666664, ans=0.0008342028985507243 2024-09-22 15:34:08,496 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.358e+02 1.827e+02 2.117e+02 2.620e+02 4.403e+02, threshold=4.235e+02, percent-clipped=1.0 2024-09-22 15:34:43,819 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=46349.333333333336, ans=0.0 2024-09-22 15:35:01,056 INFO [train.py:1198] (2/4) Epoch 3, batch 2150, loss[loss=0.3459, ctc_loss=0.2552, cr_loss=0.4534, over 17021.00 frames. ], tot_loss[loss=0.3363, ctc_loss=0.2526, cr_loss=0.4185, over 3355019.00 frames. ], batch size: 44, lr: 3.25e-02, grad_scale: 32.0 2024-09-22 15:35:06,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=46396.0, ans=0.125 2024-09-22 15:35:30,866 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=46442.666666666664, ans=0.0007733333333333325 2024-09-22 15:35:32,458 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=46442.666666666664, ans=0.025 2024-09-22 15:35:32,489 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=46442.666666666664, ans=0.125 2024-09-22 15:35:36,143 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=4.82 vs. limit=12.0 2024-09-22 15:35:37,192 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=46489.333333333336, ans=0.125 2024-09-22 15:36:00,977 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=46536.0, ans=0.1 2024-09-22 15:36:25,581 INFO [train.py:1198] (2/4) Epoch 3, batch 2200, loss[loss=0.278, ctc_loss=0.2048, cr_loss=0.3661, over 17012.00 frames. ], tot_loss[loss=0.3368, ctc_loss=0.2527, cr_loss=0.4204, over 3362835.65 frames. ], batch size: 39, lr: 3.25e-02, grad_scale: 32.0 2024-09-22 15:36:33,773 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=46629.333333333336, ans=0.0 2024-09-22 15:36:42,044 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=46676.0, ans=0.1 2024-09-22 15:36:43,762 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=46676.0, ans=0.025 2024-09-22 15:36:55,287 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.307e+02 1.636e+02 2.012e+02 2.486e+02 4.697e+02, threshold=4.025e+02, percent-clipped=4.0 2024-09-22 15:37:07,165 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=46722.666666666664, ans=0.1 2024-09-22 15:37:32,395 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=46816.0, ans=0.125 2024-09-22 15:37:49,529 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=46862.666666666664, ans=0.125 2024-09-22 15:37:50,688 INFO [train.py:1198] (2/4) Epoch 3, batch 2250, loss[loss=0.3256, ctc_loss=0.2416, cr_loss=0.4197, over 17018.00 frames. ], tot_loss[loss=0.3359, ctc_loss=0.252, cr_loss=0.4198, over 3369508.68 frames. ], batch size: 56, lr: 3.24e-02, grad_scale: 32.0 2024-09-22 15:38:01,963 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=46862.666666666664, ans=0.125 2024-09-22 15:38:03,889 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=8.35 vs. limit=15.0 2024-09-22 15:38:46,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=47002.666666666664, ans=0.125 2024-09-22 15:39:07,719 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=47049.333333333336, ans=0.125 2024-09-22 15:39:10,365 INFO [train.py:1198] (2/4) Epoch 3, batch 2300, loss[loss=0.3355, ctc_loss=0.2492, cr_loss=0.4315, over 17014.00 frames. ], tot_loss[loss=0.3366, ctc_loss=0.2524, cr_loss=0.4207, over 3370393.04 frames. ], batch size: 51, lr: 3.24e-02, grad_scale: 32.0 2024-09-22 15:39:33,221 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=47142.666666666664, ans=0.0006211594202898546 2024-09-22 15:39:37,770 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.373e+02 1.712e+02 2.236e+02 2.750e+02 4.925e+02, threshold=4.473e+02, percent-clipped=7.0 2024-09-22 15:39:47,821 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=47189.333333333336, ans=0.0006110144927536226 2024-09-22 15:40:18,092 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=3.68 vs. limit=15.0 2024-09-22 15:40:33,161 INFO [train.py:1198] (2/4) Epoch 3, batch 2350, loss[loss=0.2914, ctc_loss=0.2158, cr_loss=0.378, over 17068.00 frames. ], tot_loss[loss=0.3341, ctc_loss=0.2505, cr_loss=0.4183, over 3369459.16 frames. ], batch size: 46, lr: 3.23e-02, grad_scale: 32.0 2024-09-22 15:40:43,379 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten.whitening_limit, batch_count=47329.333333333336, ans=15.0 2024-09-22 15:40:46,243 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=47329.333333333336, ans=0.2 2024-09-22 15:40:57,626 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=47376.0, ans=0.125 2024-09-22 15:41:02,958 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=23.46 vs. limit=22.5 2024-09-22 15:41:26,676 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=11.05 vs. limit=15.0 2024-09-22 15:41:27,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=47469.333333333336, ans=0.125 2024-09-22 15:41:42,844 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=47516.0, ans=0.1 2024-09-22 15:41:58,588 INFO [train.py:1198] (2/4) Epoch 3, batch 2400, loss[loss=0.2832, ctc_loss=0.2068, cr_loss=0.382, over 17026.00 frames. ], tot_loss[loss=0.3339, ctc_loss=0.2504, cr_loss=0.4176, over 3363529.74 frames. ], batch size: 39, lr: 3.23e-02, grad_scale: 32.0 2024-09-22 15:42:25,880 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.420e+02 1.787e+02 2.027e+02 2.376e+02 4.296e+02, threshold=4.054e+02, percent-clipped=0.0 2024-09-22 15:42:51,441 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=47702.666666666664, ans=0.125 2024-09-22 15:42:53,010 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.min_positive, batch_count=47702.666666666664, ans=0.025 2024-09-22 15:43:07,924 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.77 vs. limit=6.0 2024-09-22 15:43:15,530 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=47749.333333333336, ans=0.125 2024-09-22 15:43:21,634 INFO [train.py:1198] (2/4) Epoch 3, batch 2450, loss[loss=0.3675, ctc_loss=0.2839, cr_loss=0.4179, over 16910.00 frames. ], tot_loss[loss=0.3334, ctc_loss=0.25, cr_loss=0.4169, over 3367955.40 frames. ], batch size: 58, lr: 3.22e-02, grad_scale: 32.0 2024-09-22 15:43:26,726 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=47796.0, ans=10.0 2024-09-22 15:43:33,038 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=47796.0, ans=0.125 2024-09-22 15:43:53,520 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=47889.333333333336, ans=0.125 2024-09-22 15:43:53,526 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=47889.333333333336, ans=0.125 2024-09-22 15:44:04,843 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=47889.333333333336, ans=0.0 2024-09-22 15:44:12,879 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 15:44:41,201 INFO [train.py:1198] (2/4) Epoch 3, batch 2500, loss[loss=0.3322, ctc_loss=0.2521, cr_loss=0.4003, over 17018.00 frames. ], tot_loss[loss=0.3328, ctc_loss=0.2496, cr_loss=0.4161, over 3371584.90 frames. ], batch size: 53, lr: 3.22e-02, grad_scale: 32.0 2024-09-22 15:44:53,238 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=512, metric=2.90 vs. limit=15.0 2024-09-22 15:45:07,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=48076.0, ans=0.0 2024-09-22 15:45:08,320 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.319e+02 1.703e+02 1.872e+02 2.246e+02 3.567e+02, threshold=3.744e+02, percent-clipped=0.0 2024-09-22 15:45:27,650 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.17 vs. limit=6.0 2024-09-22 15:45:48,324 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=48216.0, ans=0.0 2024-09-22 15:46:03,938 INFO [train.py:1198] (2/4) Epoch 3, batch 2550, loss[loss=0.2662, ctc_loss=0.1978, cr_loss=0.3419, over 16265.00 frames. ], tot_loss[loss=0.3321, ctc_loss=0.2489, cr_loss=0.4161, over 3377595.62 frames. ], batch size: 36, lr: 3.21e-02, grad_scale: 32.0 2024-09-22 15:46:04,341 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=48262.666666666664, ans=0.0 2024-09-22 15:46:14,845 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=48262.666666666664, ans=0.1 2024-09-22 15:46:18,161 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=48262.666666666664, ans=0.125 2024-09-22 15:46:30,657 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=48309.333333333336, ans=0.0 2024-09-22 15:46:38,583 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=48356.0, ans=0.2 2024-09-22 15:47:10,172 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=48402.666666666664, ans=0.125 2024-09-22 15:47:22,765 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=48449.333333333336, ans=0.125 2024-09-22 15:47:31,517 INFO [train.py:1198] (2/4) Epoch 3, batch 2600, loss[loss=0.3404, ctc_loss=0.2524, cr_loss=0.4398, over 17345.00 frames. ], tot_loss[loss=0.3307, ctc_loss=0.2476, cr_loss=0.4153, over 3375882.09 frames. ], batch size: 48, lr: 3.21e-02, grad_scale: 32.0 2024-09-22 15:47:42,961 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=48496.0, ans=0.0 2024-09-22 15:47:58,749 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.408e+02 1.709e+02 1.993e+02 2.394e+02 4.094e+02, threshold=3.987e+02, percent-clipped=1.0 2024-09-22 15:48:00,506 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=48542.666666666664, ans=0.1 2024-09-22 15:48:39,235 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 15:48:46,224 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.28 vs. limit=15.0 2024-09-22 15:48:51,831 INFO [train.py:1198] (2/4) Epoch 3, batch 2650, loss[loss=0.335, ctc_loss=0.2519, cr_loss=0.4159, over 17021.00 frames. ], tot_loss[loss=0.3334, ctc_loss=0.2501, cr_loss=0.4167, over 3346481.07 frames. ], batch size: 51, lr: 3.20e-02, grad_scale: 32.0 2024-09-22 15:49:36,193 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=17.83 vs. limit=22.5 2024-09-22 15:50:14,727 INFO [train.py:1198] (2/4) Epoch 3, batch 2700, loss[loss=0.3323, ctc_loss=0.2509, cr_loss=0.4068, over 17208.00 frames. ], tot_loss[loss=0.3344, ctc_loss=0.2509, cr_loss=0.4175, over 3351357.13 frames. ], batch size: 47, lr: 3.20e-02, grad_scale: 32.0 2024-09-22 15:50:24,706 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=48962.666666666664, ans=0.025 2024-09-22 15:50:41,576 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.331e+02 1.782e+02 2.096e+02 2.443e+02 4.661e+02, threshold=4.192e+02, percent-clipped=1.0 2024-09-22 15:50:43,675 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=49009.333333333336, ans=0.125 2024-09-22 15:50:50,016 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=49056.0, ans=0.0 2024-09-22 15:50:50,023 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=49056.0, ans=0.125 2024-09-22 15:51:23,114 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=49149.333333333336, ans=0.125 2024-09-22 15:51:25,182 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=512, metric=2.95 vs. limit=15.0 2024-09-22 15:51:25,262 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=20.95 vs. limit=22.5 2024-09-22 15:51:39,723 INFO [train.py:1198] (2/4) Epoch 3, batch 2750, loss[loss=0.3427, ctc_loss=0.2517, cr_loss=0.4547, over 17210.00 frames. ], tot_loss[loss=0.3353, ctc_loss=0.2514, cr_loss=0.4195, over 3353947.71 frames. ], batch size: 47, lr: 3.19e-02, grad_scale: 32.0 2024-09-22 15:51:41,563 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=49196.0, ans=0.125 2024-09-22 15:51:43,419 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.42 vs. limit=15.0 2024-09-22 15:51:44,703 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=49196.0, ans=0.1 2024-09-22 15:52:08,782 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=49242.666666666664, ans=0.1 2024-09-22 15:52:25,736 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=49289.333333333336, ans=0.0001544927536231873 2024-09-22 15:53:01,835 INFO [train.py:1198] (2/4) Epoch 3, batch 2800, loss[loss=0.3472, ctc_loss=0.262, cr_loss=0.4258, over 17025.00 frames. ], tot_loss[loss=0.3351, ctc_loss=0.2511, cr_loss=0.4201, over 3362187.36 frames. ], batch size: 52, lr: 3.19e-02, grad_scale: 32.0 2024-09-22 15:53:23,270 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=49476.0, ans=0.125 2024-09-22 15:53:24,930 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=49476.0, ans=0.1 2024-09-22 15:53:27,100 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=9.37 vs. limit=15.0 2024-09-22 15:53:29,475 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.403e+02 1.759e+02 2.001e+02 2.340e+02 4.757e+02, threshold=4.003e+02, percent-clipped=1.0 2024-09-22 15:53:44,309 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.23 vs. limit=12.0 2024-09-22 15:53:52,055 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.max_abs, batch_count=49569.333333333336, ans=10.0 2024-09-22 15:54:06,544 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=49616.0, ans=0.1 2024-09-22 15:54:11,369 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=49616.0, ans=0.125 2024-09-22 15:54:14,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=49616.0, ans=0.2 2024-09-22 15:54:22,390 INFO [train.py:1198] (2/4) Epoch 3, batch 2850, loss[loss=0.3634, ctc_loss=0.275, cr_loss=0.4418, over 17057.00 frames. ], tot_loss[loss=0.3356, ctc_loss=0.2514, cr_loss=0.421, over 3364373.38 frames. ], batch size: 46, lr: 3.18e-02, grad_scale: 32.0 2024-09-22 15:54:27,600 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=49662.666666666664, ans=0.0 2024-09-22 15:54:42,672 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.61 vs. limit=15.0 2024-09-22 15:54:56,424 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=49756.0, ans=0.125 2024-09-22 15:54:58,264 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=49756.0, ans=0.125 2024-09-22 15:55:09,224 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.42 vs. limit=15.0 2024-09-22 15:55:12,177 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 15:55:26,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=49802.666666666664, ans=0.0 2024-09-22 15:55:26,233 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=49802.666666666664, ans=0.125 2024-09-22 15:55:45,084 INFO [train.py:1198] (2/4) Epoch 3, batch 2900, loss[loss=0.3717, ctc_loss=0.2863, cr_loss=0.4268, over 16864.00 frames. ], tot_loss[loss=0.3349, ctc_loss=0.251, cr_loss=0.4199, over 3368944.04 frames. ], batch size: 58, lr: 3.18e-02, grad_scale: 32.0 2024-09-22 15:56:11,522 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=49942.666666666664, ans=0.1 2024-09-22 15:56:14,387 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.376e+02 1.745e+02 2.067e+02 2.607e+02 4.355e+02, threshold=4.133e+02, percent-clipped=1.0 2024-09-22 15:56:14,768 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=49942.666666666664, ans=0.125 2024-09-22 15:56:49,326 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=50036.0, ans=0.125 2024-09-22 15:57:09,736 INFO [train.py:1198] (2/4) Epoch 3, batch 2950, loss[loss=0.3194, ctc_loss=0.2386, cr_loss=0.404, over 17213.00 frames. ], tot_loss[loss=0.3355, ctc_loss=0.2515, cr_loss=0.4198, over 3356617.06 frames. ], batch size: 47, lr: 3.17e-02, grad_scale: 32.0 2024-09-22 15:57:20,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=50129.333333333336, ans=0.125 2024-09-22 15:57:26,983 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=50176.0, ans=0.125 2024-09-22 15:57:36,529 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=50176.0, ans=0.1 2024-09-22 15:57:53,828 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=50222.666666666664, ans=0.1 2024-09-22 15:58:04,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=50269.333333333336, ans=10.0 2024-09-22 15:58:15,853 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=50316.0, ans=0.1 2024-09-22 15:58:30,063 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=50362.666666666664, ans=0.125 2024-09-22 15:58:31,272 INFO [train.py:1198] (2/4) Epoch 3, batch 3000, loss[loss=0.3434, ctc_loss=0.2541, cr_loss=0.4467, over 17219.00 frames. ], tot_loss[loss=0.3345, ctc_loss=0.2506, cr_loss=0.4195, over 3359501.57 frames. ], batch size: 47, lr: 3.17e-02, grad_scale: 32.0 2024-09-22 15:58:31,272 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-22 15:58:46,487 INFO [train.py:1230] (2/4) Epoch 3, validation: loss=0.08436, ctc_loss=0.08436, cr_loss=7.957e-15, over 944034.00 frames. 2024-09-22 15:58:46,488 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-22 15:58:46,741 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=50362.666666666664, ans=0.04949747468305833 2024-09-22 15:58:57,400 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=50362.666666666664, ans=0.0 2024-09-22 15:58:59,794 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=10.72 vs. limit=15.0 2024-09-22 15:59:01,054 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.53 vs. limit=12.0 2024-09-22 15:59:09,935 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=50409.333333333336, ans=0.125 2024-09-22 15:59:12,853 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.406e+02 1.636e+02 1.924e+02 2.171e+02 3.615e+02, threshold=3.848e+02, percent-clipped=0.0 2024-09-22 15:59:18,579 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.94 vs. limit=6.0 2024-09-22 15:59:21,109 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 15:59:33,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=50502.666666666664, ans=0.125 2024-09-22 15:59:40,712 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=7.47 vs. limit=15.0 2024-09-22 16:00:04,588 INFO [train.py:1198] (2/4) Epoch 3, batch 3050, loss[loss=0.2811, ctc_loss=0.2061, cr_loss=0.3751, over 16955.00 frames. ], tot_loss[loss=0.3338, ctc_loss=0.2499, cr_loss=0.4195, over 3362094.56 frames. ], batch size: 42, lr: 3.16e-02, grad_scale: 32.0 2024-09-22 16:00:13,088 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.68 vs. limit=15.0 2024-09-22 16:00:24,910 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=50642.666666666664, ans=0.125 2024-09-22 16:00:28,086 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=50642.666666666664, ans=0.1 2024-09-22 16:00:34,335 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=50689.333333333336, ans=0.0 2024-09-22 16:00:42,116 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=50689.333333333336, ans=0.125 2024-09-22 16:01:15,083 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=50782.666666666664, ans=0.0 2024-09-22 16:01:17,150 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=22.04 vs. limit=22.5 2024-09-22 16:01:22,485 INFO [train.py:1198] (2/4) Epoch 3, batch 3100, loss[loss=0.3346, ctc_loss=0.2483, cr_loss=0.4317, over 17363.00 frames. ], tot_loss[loss=0.3329, ctc_loss=0.2491, cr_loss=0.4193, over 3363939.66 frames. ], batch size: 48, lr: 3.16e-02, grad_scale: 32.0 2024-09-22 16:01:42,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=50876.0, ans=0.025 2024-09-22 16:01:42,996 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=50876.0, ans=0.125 2024-09-22 16:01:49,066 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.310e+02 1.630e+02 1.881e+02 2.229e+02 3.253e+02, threshold=3.762e+02, percent-clipped=0.0 2024-09-22 16:01:54,958 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=50922.666666666664, ans=0.125 2024-09-22 16:02:02,440 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=50922.666666666664, ans=0.0 2024-09-22 16:02:19,871 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=50969.333333333336, ans=0.0 2024-09-22 16:02:29,467 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=51016.0, ans=0.125 2024-09-22 16:02:32,462 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=51016.0, ans=0.125 2024-09-22 16:02:43,156 INFO [train.py:1198] (2/4) Epoch 3, batch 3150, loss[loss=0.3447, ctc_loss=0.2598, cr_loss=0.4242, over 16988.00 frames. ], tot_loss[loss=0.333, ctc_loss=0.2491, cr_loss=0.4192, over 3359536.04 frames. ], batch size: 53, lr: 3.15e-02, grad_scale: 32.0 2024-09-22 16:02:54,796 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.38 vs. limit=12.0 2024-09-22 16:02:55,896 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=51062.666666666664, ans=0.125 2024-09-22 16:03:16,282 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=51156.0, ans=0.1 2024-09-22 16:03:51,720 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.57 vs. limit=15.0 2024-09-22 16:03:55,835 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=51249.333333333336, ans=0.125 2024-09-22 16:04:03,519 INFO [train.py:1198] (2/4) Epoch 3, batch 3200, loss[loss=0.3279, ctc_loss=0.2403, cr_loss=0.4377, over 17153.00 frames. ], tot_loss[loss=0.3343, ctc_loss=0.2502, cr_loss=0.4204, over 3361648.68 frames. ], batch size: 45, lr: 3.15e-02, grad_scale: 32.0 2024-09-22 16:04:03,786 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=51296.0, ans=0.025 2024-09-22 16:04:27,434 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-22 16:04:30,052 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.332e+02 1.638e+02 1.865e+02 2.186e+02 5.181e+02, threshold=3.729e+02, percent-clipped=1.0 2024-09-22 16:04:41,657 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.65 vs. limit=15.0 2024-09-22 16:04:45,713 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=51389.333333333336, ans=0.0 2024-09-22 16:04:55,028 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=51436.0, ans=0.04949747468305833 2024-09-22 16:05:04,933 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=23.26 vs. limit=22.5 2024-09-22 16:05:05,278 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=12.82 vs. limit=15.0 2024-09-22 16:05:07,923 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten.whitening_limit, batch_count=51482.666666666664, ans=22.5 2024-09-22 16:05:15,368 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=51482.666666666664, ans=0.1 2024-09-22 16:05:20,601 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=51482.666666666664, ans=0.0 2024-09-22 16:05:23,508 INFO [train.py:1198] (2/4) Epoch 3, batch 3250, loss[loss=0.3487, ctc_loss=0.2553, cr_loss=0.4669, over 15950.00 frames. ], tot_loss[loss=0.3326, ctc_loss=0.2488, cr_loss=0.419, over 3363168.95 frames. ], batch size: 74, lr: 3.14e-02, grad_scale: 32.0 2024-09-22 16:05:23,783 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=51529.333333333336, ans=0.0 2024-09-22 16:05:26,935 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=51529.333333333336, ans=0.1 2024-09-22 16:05:54,939 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=51622.666666666664, ans=0.0 2024-09-22 16:05:58,612 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=7.32 vs. limit=15.0 2024-09-22 16:06:08,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=51669.333333333336, ans=0.025 2024-09-22 16:06:24,734 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.52 vs. limit=15.0 2024-09-22 16:06:43,309 INFO [train.py:1198] (2/4) Epoch 3, batch 3300, loss[loss=0.2969, ctc_loss=0.2181, cr_loss=0.3941, over 17181.00 frames. ], tot_loss[loss=0.3327, ctc_loss=0.2489, cr_loss=0.4189, over 3363150.35 frames. ], batch size: 45, lr: 3.14e-02, grad_scale: 32.0 2024-09-22 16:07:09,978 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.328e+02 1.647e+02 1.953e+02 2.348e+02 4.155e+02, threshold=3.905e+02, percent-clipped=1.0 2024-09-22 16:07:27,291 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=51856.0, ans=0.1 2024-09-22 16:07:35,443 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=3.63 vs. limit=10.0 2024-09-22 16:07:46,032 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=51949.333333333336, ans=0.0 2024-09-22 16:07:55,383 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 16:08:01,372 INFO [train.py:1198] (2/4) Epoch 3, batch 3350, loss[loss=0.3684, ctc_loss=0.2719, cr_loss=0.4826, over 17103.00 frames. ], tot_loss[loss=0.3324, ctc_loss=0.2486, cr_loss=0.4194, over 3363269.38 frames. ], batch size: 49, lr: 3.13e-02, grad_scale: 32.0 2024-09-22 16:08:04,822 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=51996.0, ans=0.125 2024-09-22 16:08:35,903 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=52089.333333333336, ans=0.0 2024-09-22 16:08:36,004 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=52089.333333333336, ans=0.125 2024-09-22 16:08:56,216 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=52136.0, ans=0.125 2024-09-22 16:09:19,277 INFO [train.py:1198] (2/4) Epoch 3, batch 3400, loss[loss=0.3088, ctc_loss=0.2259, cr_loss=0.4141, over 17270.00 frames. ], tot_loss[loss=0.3319, ctc_loss=0.2481, cr_loss=0.4187, over 3363367.89 frames. ], batch size: 42, lr: 3.13e-02, grad_scale: 32.0 2024-09-22 16:09:25,882 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=52229.333333333336, ans=0.125 2024-09-22 16:09:27,532 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=52229.333333333336, ans=0.125 2024-09-22 16:09:33,882 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=52276.0, ans=0.09899494936611666 2024-09-22 16:09:40,022 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=52276.0, ans=0.125 2024-09-22 16:09:45,958 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.323e+02 1.677e+02 1.939e+02 2.464e+02 4.534e+02, threshold=3.878e+02, percent-clipped=3.0 2024-09-22 16:10:37,704 INFO [train.py:1198] (2/4) Epoch 3, batch 3450, loss[loss=0.3672, ctc_loss=0.2824, cr_loss=0.4238, over 17037.00 frames. ], tot_loss[loss=0.3312, ctc_loss=0.2475, cr_loss=0.4184, over 3366202.67 frames. ], batch size: 52, lr: 3.12e-02, grad_scale: 32.0 2024-09-22 16:10:55,249 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=52509.333333333336, ans=0.1 2024-09-22 16:11:19,970 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=52556.0, ans=0.125 2024-09-22 16:11:27,959 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=52602.666666666664, ans=0.2 2024-09-22 16:11:37,367 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-22 16:11:38,759 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=52649.333333333336, ans=0.0 2024-09-22 16:11:46,443 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=52649.333333333336, ans=0.125 2024-09-22 16:11:47,889 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=52649.333333333336, ans=0.125 2024-09-22 16:11:57,412 INFO [train.py:1198] (2/4) Epoch 3, batch 3500, loss[loss=0.307, ctc_loss=0.23, cr_loss=0.385, over 17028.00 frames. ], tot_loss[loss=0.3329, ctc_loss=0.2491, cr_loss=0.4188, over 3358771.55 frames. ], batch size: 44, lr: 3.12e-02, grad_scale: 32.0 2024-09-22 16:12:22,922 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=52742.666666666664, ans=0.025 2024-09-22 16:12:24,122 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.446e+02 1.808e+02 2.110e+02 2.675e+02 4.151e+02, threshold=4.220e+02, percent-clipped=2.0 2024-09-22 16:12:36,897 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=52789.333333333336, ans=0.125 2024-09-22 16:12:43,127 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=52836.0, ans=0.125 2024-09-22 16:13:04,988 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=52882.666666666664, ans=0.125 2024-09-22 16:13:08,069 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=52882.666666666664, ans=0.035 2024-09-22 16:13:15,557 INFO [train.py:1198] (2/4) Epoch 3, batch 3550, loss[loss=0.3037, ctc_loss=0.2255, cr_loss=0.3905, over 16950.00 frames. ], tot_loss[loss=0.3325, ctc_loss=0.2487, cr_loss=0.4187, over 3359931.54 frames. ], batch size: 42, lr: 3.11e-02, grad_scale: 32.0 2024-09-22 16:13:34,977 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.07 vs. limit=10.0 2024-09-22 16:13:52,486 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.29 vs. limit=12.0 2024-09-22 16:14:07,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=53069.333333333336, ans=0.125 2024-09-22 16:14:20,180 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=53116.0, ans=0.0 2024-09-22 16:14:27,900 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=53116.0, ans=0.04949747468305833 2024-09-22 16:14:35,690 INFO [train.py:1198] (2/4) Epoch 3, batch 3600, loss[loss=0.3334, ctc_loss=0.2517, cr_loss=0.4084, over 17285.00 frames. ], tot_loss[loss=0.3307, ctc_loss=0.2471, cr_loss=0.4178, over 3366604.32 frames. ], batch size: 51, lr: 3.11e-02, grad_scale: 32.0 2024-09-22 16:14:38,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=53162.666666666664, ans=0.125 2024-09-22 16:14:48,525 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=53162.666666666664, ans=0.125 2024-09-22 16:14:50,512 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=53162.666666666664, ans=0.0 2024-09-22 16:15:04,297 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.276e+02 1.804e+02 2.185e+02 2.634e+02 3.942e+02, threshold=4.371e+02, percent-clipped=0.0 2024-09-22 16:15:07,782 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=53256.0, ans=0.125 2024-09-22 16:15:22,238 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=16.68 vs. limit=22.5 2024-09-22 16:15:29,646 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=53302.666666666664, ans=0.1 2024-09-22 16:15:30,239 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.34 vs. limit=6.0 2024-09-22 16:15:57,743 INFO [train.py:1198] (2/4) Epoch 3, batch 3650, loss[loss=0.3442, ctc_loss=0.2602, cr_loss=0.42, over 15822.00 frames. ], tot_loss[loss=0.3303, ctc_loss=0.2468, cr_loss=0.4175, over 3363146.66 frames. ], batch size: 74, lr: 3.10e-02, grad_scale: 32.0 2024-09-22 16:15:58,070 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=53396.0, ans=0.125 2024-09-22 16:16:05,733 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=53396.0, ans=0.2 2024-09-22 16:16:07,190 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=53396.0, ans=0.1 2024-09-22 16:17:03,863 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=53582.666666666664, ans=0.125 2024-09-22 16:17:16,109 INFO [train.py:1198] (2/4) Epoch 3, batch 3700, loss[loss=0.2848, ctc_loss=0.2075, cr_loss=0.3867, over 16711.00 frames. ], tot_loss[loss=0.3318, ctc_loss=0.248, cr_loss=0.4187, over 3357142.75 frames. ], batch size: 37, lr: 3.10e-02, grad_scale: 32.0 2024-09-22 16:17:16,313 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=53629.333333333336, ans=0.0 2024-09-22 16:17:30,110 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=53676.0, ans=0.2 2024-09-22 16:17:42,218 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.264e+02 1.642e+02 1.878e+02 2.249e+02 4.018e+02, threshold=3.757e+02, percent-clipped=0.0 2024-09-22 16:18:29,507 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer_na.min_abs, batch_count=53816.0, ans=0.02 2024-09-22 16:18:34,046 INFO [train.py:1198] (2/4) Epoch 3, batch 3750, loss[loss=0.2788, ctc_loss=0.2049, cr_loss=0.3696, over 16708.00 frames. ], tot_loss[loss=0.3319, ctc_loss=0.2482, cr_loss=0.4184, over 3342081.49 frames. ], batch size: 37, lr: 3.10e-02, grad_scale: 32.0 2024-09-22 16:18:47,070 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=18.21 vs. limit=22.5 2024-09-22 16:18:50,247 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=15.94 vs. limit=15.0 2024-09-22 16:18:53,748 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=17.14 vs. limit=22.5 2024-09-22 16:18:57,620 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 16:19:01,139 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=12.44 vs. limit=15.0 2024-09-22 16:19:19,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=54002.666666666664, ans=0.0 2024-09-22 16:19:37,876 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 16:19:51,719 INFO [train.py:1198] (2/4) Epoch 3, batch 3800, loss[loss=0.3814, ctc_loss=0.288, cr_loss=0.467, over 15042.00 frames. ], tot_loss[loss=0.3344, ctc_loss=0.2505, cr_loss=0.4194, over 3315285.01 frames. ], batch size: 89, lr: 3.09e-02, grad_scale: 32.0 2024-09-22 16:19:54,985 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=54096.0, ans=0.125 2024-09-22 16:19:58,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=54096.0, ans=0.125 2024-09-22 16:19:58,301 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=54096.0, ans=0.0 2024-09-22 16:20:01,228 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=54096.0, ans=0.125 2024-09-22 16:20:17,980 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.396e+02 1.642e+02 1.883e+02 2.367e+02 4.025e+02, threshold=3.766e+02, percent-clipped=5.0 2024-09-22 16:20:34,162 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=54189.333333333336, ans=0.0 2024-09-22 16:20:42,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=54236.0, ans=0.0 2024-09-22 16:21:10,249 INFO [train.py:1198] (2/4) Epoch 3, batch 3850, loss[loss=0.2774, ctc_loss=0.2043, cr_loss=0.3653, over 16972.00 frames. ], tot_loss[loss=0.3344, ctc_loss=0.2506, cr_loss=0.4191, over 3305403.98 frames. ], batch size: 42, lr: 3.09e-02, grad_scale: 64.0 2024-09-22 16:21:17,246 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=10.73 vs. limit=15.0 2024-09-22 16:21:30,498 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=54376.0, ans=0.125 2024-09-22 16:21:40,499 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=3.78 vs. limit=15.0 2024-09-22 16:21:55,275 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=54469.333333333336, ans=10.0 2024-09-22 16:22:01,275 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=54469.333333333336, ans=0.0 2024-09-22 16:22:01,753 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=10.73 vs. limit=15.0 2024-09-22 16:22:12,188 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.99 vs. limit=6.0 2024-09-22 16:22:13,211 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=54516.0, ans=0.025 2024-09-22 16:22:18,274 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=13.60 vs. limit=15.0 2024-09-22 16:23:12,171 INFO [train.py:1198] (2/4) Epoch 4, batch 0, loss[loss=0.3368, ctc_loss=0.2506, cr_loss=0.4306, over 17225.00 frames. ], tot_loss[loss=0.3368, ctc_loss=0.2506, cr_loss=0.4306, over 17225.00 frames. ], batch size: 47, lr: 2.88e-02, grad_scale: 32.0 2024-09-22 16:23:12,171 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-22 16:23:27,757 INFO [train.py:1230] (2/4) Epoch 4, validation: loss=0.08466, ctc_loss=0.08466, cr_loss=9.003e-15, over 944034.00 frames. 2024-09-22 16:23:27,758 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-22 16:24:06,065 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.466e+02 1.944e+02 2.324e+02 2.751e+02 6.786e+02, threshold=4.649e+02, percent-clipped=3.0 2024-09-22 16:24:07,868 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=54637.333333333336, ans=0.0 2024-09-22 16:24:26,833 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=54684.0, ans=0.04949747468305833 2024-09-22 16:24:33,200 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=54730.666666666664, ans=0.125 2024-09-22 16:24:50,274 INFO [train.py:1198] (2/4) Epoch 4, batch 50, loss[loss=0.3345, ctc_loss=0.2494, cr_loss=0.4254, over 17037.00 frames. ], tot_loss[loss=0.3334, ctc_loss=0.2493, cr_loss=0.4208, over 762081.68 frames. ], batch size: 52, lr: 2.88e-02, grad_scale: 32.0 2024-09-22 16:25:12,804 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=54824.0, ans=0.2 2024-09-22 16:25:20,545 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=54870.666666666664, ans=0.2 2024-09-22 16:25:21,021 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.29 vs. limit=15.0 2024-09-22 16:25:39,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=54917.333333333336, ans=0.2 2024-09-22 16:25:49,212 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=54917.333333333336, ans=0.125 2024-09-22 16:25:52,466 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=54964.0, ans=0.0 2024-09-22 16:26:12,371 INFO [train.py:1198] (2/4) Epoch 4, batch 100, loss[loss=0.3435, ctc_loss=0.2578, cr_loss=0.4286, over 17102.00 frames. ], tot_loss[loss=0.3267, ctc_loss=0.2434, cr_loss=0.4167, over 1328612.19 frames. ], batch size: 49, lr: 2.87e-02, grad_scale: 32.0 2024-09-22 16:26:23,936 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=6.68 vs. limit=15.0 2024-09-22 16:26:43,224 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.59 vs. limit=6.0 2024-09-22 16:26:50,366 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.359e+02 1.670e+02 1.866e+02 2.190e+02 3.249e+02, threshold=3.731e+02, percent-clipped=0.0 2024-09-22 16:27:04,089 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.64 vs. limit=15.0 2024-09-22 16:27:08,540 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=55150.666666666664, ans=0.125 2024-09-22 16:27:24,284 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=55197.333333333336, ans=0.125 2024-09-22 16:27:29,132 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=55197.333333333336, ans=0.125 2024-09-22 16:27:32,714 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.75 vs. limit=6.0 2024-09-22 16:27:35,224 INFO [train.py:1198] (2/4) Epoch 4, batch 150, loss[loss=0.3056, ctc_loss=0.2225, cr_loss=0.4155, over 17074.00 frames. ], tot_loss[loss=0.3227, ctc_loss=0.2402, cr_loss=0.4125, over 1775469.75 frames. ], batch size: 46, lr: 2.87e-02, grad_scale: 32.0 2024-09-22 16:27:46,483 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=55244.0, ans=0.0 2024-09-22 16:28:10,572 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=55337.333333333336, ans=0.125 2024-09-22 16:28:22,519 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.20 vs. limit=6.0 2024-09-22 16:28:56,167 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=55430.666666666664, ans=0.0 2024-09-22 16:28:58,934 INFO [train.py:1198] (2/4) Epoch 4, batch 200, loss[loss=0.3069, ctc_loss=0.2251, cr_loss=0.409, over 17100.00 frames. ], tot_loss[loss=0.3255, ctc_loss=0.2424, cr_loss=0.4157, over 2124400.94 frames. ], batch size: 43, lr: 2.86e-02, grad_scale: 32.0 2024-09-22 16:29:02,381 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=55477.333333333336, ans=0.025 2024-09-22 16:29:18,573 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.82 vs. limit=10.0 2024-09-22 16:29:30,014 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.whiten, num_groups=1, num_channels=192, metric=4.15 vs. limit=12.0 2024-09-22 16:29:32,431 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=55570.666666666664, ans=0.125 2024-09-22 16:29:33,618 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.233e+02 1.559e+02 1.682e+02 1.957e+02 2.989e+02, threshold=3.363e+02, percent-clipped=0.0 2024-09-22 16:29:46,738 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 16:29:51,415 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=55617.333333333336, ans=0.1 2024-09-22 16:30:10,085 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=55664.0, ans=0.0 2024-09-22 16:30:17,535 INFO [train.py:1198] (2/4) Epoch 4, batch 250, loss[loss=0.374, ctc_loss=0.2798, cr_loss=0.471, over 17218.00 frames. ], tot_loss[loss=0.3222, ctc_loss=0.2394, cr_loss=0.4137, over 2407810.37 frames. ], batch size: 55, lr: 2.86e-02, grad_scale: 32.0 2024-09-22 16:30:17,746 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=55710.666666666664, ans=0.125 2024-09-22 16:30:52,676 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=55804.0, ans=0.125 2024-09-22 16:30:58,903 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=55804.0, ans=0.125 2024-09-22 16:31:14,650 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=55850.666666666664, ans=0.125 2024-09-22 16:31:27,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=55897.333333333336, ans=0.125 2024-09-22 16:31:29,120 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=55897.333333333336, ans=0.1 2024-09-22 16:31:42,160 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.15 vs. limit=15.0 2024-09-22 16:31:43,124 INFO [train.py:1198] (2/4) Epoch 4, batch 300, loss[loss=0.2973, ctc_loss=0.2263, cr_loss=0.3554, over 17059.00 frames. ], tot_loss[loss=0.3231, ctc_loss=0.2403, cr_loss=0.4139, over 2614827.03 frames. ], batch size: 46, lr: 2.86e-02, grad_scale: 32.0 2024-09-22 16:31:55,953 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=55944.0, ans=0.125 2024-09-22 16:32:20,153 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.310e+02 1.659e+02 1.864e+02 2.226e+02 3.223e+02, threshold=3.728e+02, percent-clipped=0.0 2024-09-22 16:32:22,082 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=56037.333333333336, ans=0.025 2024-09-22 16:33:07,320 INFO [train.py:1198] (2/4) Epoch 4, batch 350, loss[loss=0.3386, ctc_loss=0.2493, cr_loss=0.4466, over 17269.00 frames. ], tot_loss[loss=0.3231, ctc_loss=0.2404, cr_loss=0.4133, over 2764136.12 frames. ], batch size: 46, lr: 2.85e-02, grad_scale: 32.0 2024-09-22 16:33:54,060 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=9.56 vs. limit=15.0 2024-09-22 16:34:15,799 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=56364.0, ans=0.125 2024-09-22 16:34:29,677 INFO [train.py:1198] (2/4) Epoch 4, batch 400, loss[loss=0.2604, ctc_loss=0.1887, cr_loss=0.3587, over 16957.00 frames. ], tot_loss[loss=0.3246, ctc_loss=0.2416, cr_loss=0.4149, over 2896204.56 frames. ], batch size: 42, lr: 2.85e-02, grad_scale: 32.0 2024-09-22 16:34:33,846 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.87 vs. limit=6.0 2024-09-22 16:34:41,196 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=56410.666666666664, ans=0.0 2024-09-22 16:34:42,230 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=9.17 vs. limit=15.0 2024-09-22 16:34:46,299 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=56457.333333333336, ans=0.2 2024-09-22 16:34:49,378 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=56457.333333333336, ans=0.125 2024-09-22 16:34:55,573 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=56457.333333333336, ans=0.0 2024-09-22 16:35:04,961 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.343e+02 1.655e+02 1.851e+02 2.343e+02 4.879e+02, threshold=3.703e+02, percent-clipped=2.0 2024-09-22 16:35:15,013 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=56504.0, ans=0.0 2024-09-22 16:35:52,501 INFO [train.py:1198] (2/4) Epoch 4, batch 450, loss[loss=0.4086, ctc_loss=0.3177, cr_loss=0.4545, over 12075.00 frames. ], tot_loss[loss=0.3253, ctc_loss=0.2422, cr_loss=0.4156, over 2996667.62 frames. ], batch size: 123, lr: 2.84e-02, grad_scale: 32.0 2024-09-22 16:36:08,310 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=56644.0, ans=0.125 2024-09-22 16:36:12,989 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=56690.666666666664, ans=0.07 2024-09-22 16:36:15,111 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=17.54 vs. limit=22.5 2024-09-22 16:36:27,548 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=56737.333333333336, ans=0.125 2024-09-22 16:36:32,228 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=56737.333333333336, ans=0.2 2024-09-22 16:36:48,136 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=56784.0, ans=0.1 2024-09-22 16:36:55,479 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=11.69 vs. limit=15.0 2024-09-22 16:37:01,012 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=56830.666666666664, ans=0.125 2024-09-22 16:37:14,899 INFO [train.py:1198] (2/4) Epoch 4, batch 500, loss[loss=0.3144, ctc_loss=0.2318, cr_loss=0.4132, over 17292.00 frames. ], tot_loss[loss=0.3237, ctc_loss=0.2407, cr_loss=0.4148, over 3086418.61 frames. ], batch size: 46, lr: 2.84e-02, grad_scale: 32.0 2024-09-22 16:37:36,405 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=56924.0, ans=0.125 2024-09-22 16:37:37,998 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=56924.0, ans=0.125 2024-09-22 16:37:53,150 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.289e+02 1.629e+02 1.949e+02 2.165e+02 3.477e+02, threshold=3.897e+02, percent-clipped=0.0 2024-09-22 16:38:02,973 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=56970.666666666664, ans=0.125 2024-09-22 16:38:18,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=57017.333333333336, ans=0.025 2024-09-22 16:38:39,921 INFO [train.py:1198] (2/4) Epoch 4, batch 550, loss[loss=0.3347, ctc_loss=0.2505, cr_loss=0.4212, over 16889.00 frames. ], tot_loss[loss=0.3217, ctc_loss=0.2391, cr_loss=0.4132, over 3151708.55 frames. ], batch size: 58, lr: 2.83e-02, grad_scale: 32.0 2024-09-22 16:38:51,936 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.48 vs. limit=6.0 2024-09-22 16:38:52,826 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=57110.666666666664, ans=0.025 2024-09-22 16:39:29,453 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=57250.666666666664, ans=0.0 2024-09-22 16:39:43,766 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=57297.333333333336, ans=0.125 2024-09-22 16:39:44,286 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=15.99 vs. limit=15.0 2024-09-22 16:39:59,407 INFO [train.py:1198] (2/4) Epoch 4, batch 600, loss[loss=0.3142, ctc_loss=0.2307, cr_loss=0.4174, over 17056.00 frames. ], tot_loss[loss=0.3227, ctc_loss=0.2397, cr_loss=0.4147, over 3202434.91 frames. ], batch size: 46, lr: 2.83e-02, grad_scale: 32.0 2024-09-22 16:40:02,833 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 16:40:02,837 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=57344.0, ans=0.1 2024-09-22 16:40:04,430 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-22 16:40:19,159 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=23.02 vs. limit=22.5 2024-09-22 16:40:25,128 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=57390.666666666664, ans=0.125 2024-09-22 16:40:34,388 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.289e+02 1.589e+02 1.770e+02 2.208e+02 4.389e+02, threshold=3.540e+02, percent-clipped=1.0 2024-09-22 16:40:41,103 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=57437.333333333336, ans=0.09899494936611666 2024-09-22 16:40:49,925 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=57484.0, ans=0.2 2024-09-22 16:40:49,937 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=57484.0, ans=0.125 2024-09-22 16:40:52,169 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=7.08 vs. limit=10.0 2024-09-22 16:41:04,101 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.03 vs. limit=15.0 2024-09-22 16:41:24,122 INFO [train.py:1198] (2/4) Epoch 4, batch 650, loss[loss=0.3154, ctc_loss=0.2349, cr_loss=0.4025, over 17081.00 frames. ], tot_loss[loss=0.3219, ctc_loss=0.2389, cr_loss=0.4151, over 3245762.34 frames. ], batch size: 49, lr: 2.83e-02, grad_scale: 32.0 2024-09-22 16:41:41,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.min_abs, batch_count=57624.0, ans=0.5 2024-09-22 16:41:43,271 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=57624.0, ans=0.2 2024-09-22 16:42:27,500 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=57764.0, ans=0.2 2024-09-22 16:42:33,967 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=57764.0, ans=0.125 2024-09-22 16:42:42,720 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=57764.0, ans=0.125 2024-09-22 16:42:45,631 INFO [train.py:1198] (2/4) Epoch 4, batch 700, loss[loss=0.3072, ctc_loss=0.2259, cr_loss=0.4065, over 17150.00 frames. ], tot_loss[loss=0.3215, ctc_loss=0.2385, cr_loss=0.4153, over 3275582.52 frames. ], batch size: 48, lr: 2.82e-02, grad_scale: 32.0 2024-09-22 16:43:10,079 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=57857.333333333336, ans=0.2 2024-09-22 16:43:23,507 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.245e+02 1.612e+02 1.882e+02 2.294e+02 3.695e+02, threshold=3.764e+02, percent-clipped=3.0 2024-09-22 16:43:36,890 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=57950.666666666664, ans=0.07 2024-09-22 16:43:57,944 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.70 vs. limit=10.0 2024-09-22 16:44:00,767 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=57997.333333333336, ans=0.1 2024-09-22 16:44:08,398 INFO [train.py:1198] (2/4) Epoch 4, batch 750, loss[loss=0.3142, ctc_loss=0.235, cr_loss=0.3958, over 17264.00 frames. ], tot_loss[loss=0.3212, ctc_loss=0.2382, cr_loss=0.4146, over 3292403.70 frames. ], batch size: 44, lr: 2.82e-02, grad_scale: 32.0 2024-09-22 16:44:12,470 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.33 vs. limit=6.0 2024-09-22 16:44:14,873 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=58044.0, ans=0.1 2024-09-22 16:44:30,619 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=58090.666666666664, ans=0.125 2024-09-22 16:44:37,108 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.27 vs. limit=6.0 2024-09-22 16:44:44,746 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=58137.333333333336, ans=0.125 2024-09-22 16:44:44,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=58137.333333333336, ans=0.0 2024-09-22 16:44:46,293 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=58137.333333333336, ans=0.125 2024-09-22 16:44:55,780 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=58184.0, ans=0.2 2024-09-22 16:45:25,239 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=8.38 vs. limit=15.0 2024-09-22 16:45:27,577 INFO [train.py:1198] (2/4) Epoch 4, batch 800, loss[loss=0.2627, ctc_loss=0.191, cr_loss=0.3581, over 17172.00 frames. ], tot_loss[loss=0.323, ctc_loss=0.2399, cr_loss=0.4154, over 3290260.60 frames. ], batch size: 41, lr: 2.81e-02, grad_scale: 32.0 2024-09-22 16:46:07,411 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.352e+02 1.727e+02 1.869e+02 2.216e+02 3.268e+02, threshold=3.738e+02, percent-clipped=0.0 2024-09-22 16:46:20,409 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=58417.333333333336, ans=0.125 2024-09-22 16:46:36,932 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=8.10 vs. limit=15.0 2024-09-22 16:46:45,128 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=12.24 vs. limit=22.5 2024-09-22 16:46:50,642 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=58510.666666666664, ans=0.5 2024-09-22 16:46:51,992 INFO [train.py:1198] (2/4) Epoch 4, batch 850, loss[loss=0.3739, ctc_loss=0.2817, cr_loss=0.4607, over 16947.00 frames. ], tot_loss[loss=0.3211, ctc_loss=0.2383, cr_loss=0.4142, over 3300204.11 frames. ], batch size: 58, lr: 2.81e-02, grad_scale: 32.0 2024-09-22 16:47:08,658 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=9.63 vs. limit=15.0 2024-09-22 16:47:40,072 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=58650.666666666664, ans=0.125 2024-09-22 16:48:02,340 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=21.16 vs. limit=22.5 2024-09-22 16:48:04,078 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=4.15 vs. limit=12.0 2024-09-22 16:48:09,607 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=58697.333333333336, ans=0.1 2024-09-22 16:48:13,693 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.78 vs. limit=6.0 2024-09-22 16:48:16,702 INFO [train.py:1198] (2/4) Epoch 4, batch 900, loss[loss=0.2896, ctc_loss=0.2124, cr_loss=0.3858, over 17193.00 frames. ], tot_loss[loss=0.3205, ctc_loss=0.2376, cr_loss=0.4143, over 3318401.83 frames. ], batch size: 41, lr: 2.81e-02, grad_scale: 32.0 2024-09-22 16:48:25,001 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=58744.0, ans=0.2 2024-09-22 16:48:32,988 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=58790.666666666664, ans=0.035 2024-09-22 16:48:42,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=58790.666666666664, ans=0.0 2024-09-22 16:48:46,585 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module1.whiten, num_groups=1, num_channels=512, metric=7.88 vs. limit=15.0 2024-09-22 16:48:50,929 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=58837.333333333336, ans=0.125 2024-09-22 16:48:52,230 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.310e+02 1.664e+02 1.989e+02 2.596e+02 4.339e+02, threshold=3.979e+02, percent-clipped=2.0 2024-09-22 16:49:08,603 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=58884.0, ans=0.0 2024-09-22 16:49:18,319 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=58884.0, ans=0.1 2024-09-22 16:49:21,596 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=58930.666666666664, ans=0.125 2024-09-22 16:49:36,961 INFO [train.py:1198] (2/4) Epoch 4, batch 950, loss[loss=0.3475, ctc_loss=0.2602, cr_loss=0.4365, over 17001.00 frames. ], tot_loss[loss=0.3185, ctc_loss=0.2359, cr_loss=0.4132, over 3338326.16 frames. ], batch size: 56, lr: 2.80e-02, grad_scale: 32.0 2024-09-22 16:50:18,285 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=59070.666666666664, ans=0.125 2024-09-22 16:50:44,881 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=59164.0, ans=0.025 2024-09-22 16:51:01,450 INFO [train.py:1198] (2/4) Epoch 4, batch 1000, loss[loss=0.3937, ctc_loss=0.2977, cr_loss=0.4801, over 15851.00 frames. ], tot_loss[loss=0.3194, ctc_loss=0.2366, cr_loss=0.4142, over 3342534.51 frames. ], batch size: 74, lr: 2.80e-02, grad_scale: 32.0 2024-09-22 16:51:04,800 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=59210.666666666664, ans=0.2 2024-09-22 16:51:36,100 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.266e+02 1.574e+02 1.735e+02 2.105e+02 3.870e+02, threshold=3.470e+02, percent-clipped=0.0 2024-09-22 16:51:36,459 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=59304.0, ans=0.1 2024-09-22 16:51:44,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=59304.0, ans=0.125 2024-09-22 16:51:58,556 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=59350.666666666664, ans=0.0 2024-09-22 16:52:19,040 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=59444.0, ans=0.5 2024-09-22 16:52:19,089 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=59444.0, ans=0.125 2024-09-22 16:52:20,426 INFO [train.py:1198] (2/4) Epoch 4, batch 1050, loss[loss=0.2863, ctc_loss=0.2092, cr_loss=0.3854, over 17363.00 frames. ], tot_loss[loss=0.3187, ctc_loss=0.236, cr_loss=0.4131, over 3351833.09 frames. ], batch size: 48, lr: 2.79e-02, grad_scale: 32.0 2024-09-22 16:52:39,450 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=59490.666666666664, ans=0.1 2024-09-22 16:53:09,453 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=59584.0, ans=0.0 2024-09-22 16:53:16,667 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=59584.0, ans=0.0 2024-09-22 16:53:44,831 INFO [train.py:1198] (2/4) Epoch 4, batch 1100, loss[loss=0.3166, ctc_loss=0.2317, cr_loss=0.4243, over 17355.00 frames. ], tot_loss[loss=0.3187, ctc_loss=0.236, cr_loss=0.4132, over 3359246.06 frames. ], batch size: 48, lr: 2.79e-02, grad_scale: 32.0 2024-09-22 16:53:48,357 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=59677.333333333336, ans=0.2 2024-09-22 16:53:56,632 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=512, metric=3.63 vs. limit=15.0 2024-09-22 16:54:02,621 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=59724.0, ans=0.025 2024-09-22 16:54:17,092 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=59770.666666666664, ans=0.0 2024-09-22 16:54:20,014 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.292e+02 1.576e+02 1.850e+02 2.274e+02 3.544e+02, threshold=3.699e+02, percent-clipped=1.0 2024-09-22 16:54:28,481 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=59770.666666666664, ans=0.125 2024-09-22 16:55:04,528 INFO [train.py:1198] (2/4) Epoch 4, batch 1150, loss[loss=0.3142, ctc_loss=0.2306, cr_loss=0.4182, over 17357.00 frames. ], tot_loss[loss=0.3167, ctc_loss=0.2344, cr_loss=0.4112, over 3363626.10 frames. ], batch size: 48, lr: 2.78e-02, grad_scale: 32.0 2024-09-22 16:55:09,692 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=59910.666666666664, ans=0.125 2024-09-22 16:55:40,660 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=60004.0, ans=0.0 2024-09-22 16:55:47,359 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=4.87 vs. limit=12.0 2024-09-22 16:56:06,953 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=60050.666666666664, ans=0.0 2024-09-22 16:56:28,658 INFO [train.py:1198] (2/4) Epoch 4, batch 1200, loss[loss=0.2726, ctc_loss=0.1987, cr_loss=0.3696, over 17244.00 frames. ], tot_loss[loss=0.3167, ctc_loss=0.2344, cr_loss=0.4114, over 3364248.20 frames. ], batch size: 42, lr: 2.78e-02, grad_scale: 32.0 2024-09-22 16:56:47,448 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=9.50 vs. limit=15.0 2024-09-22 16:57:03,616 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.315e+02 1.578e+02 1.757e+02 2.030e+02 3.618e+02, threshold=3.514e+02, percent-clipped=0.0 2024-09-22 16:57:21,287 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=60284.0, ans=0.015 2024-09-22 16:57:39,972 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=60330.666666666664, ans=0.125 2024-09-22 16:57:51,021 INFO [train.py:1198] (2/4) Epoch 4, batch 1250, loss[loss=0.3193, ctc_loss=0.235, cr_loss=0.4214, over 16967.00 frames. ], tot_loss[loss=0.3167, ctc_loss=0.2345, cr_loss=0.4109, over 3352842.65 frames. ], batch size: 42, lr: 2.78e-02, grad_scale: 32.0 2024-09-22 16:58:12,838 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=60424.0, ans=0.125 2024-09-22 16:58:24,873 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn2.whiten, num_groups=1, num_channels=192, metric=10.67 vs. limit=22.5 2024-09-22 16:59:00,549 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=60564.0, ans=0.2 2024-09-22 16:59:05,251 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=60564.0, ans=0.0 2024-09-22 16:59:10,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=60564.0, ans=0.125 2024-09-22 16:59:12,919 INFO [train.py:1198] (2/4) Epoch 4, batch 1300, loss[loss=0.3474, ctc_loss=0.2586, cr_loss=0.4439, over 17052.00 frames. ], tot_loss[loss=0.3169, ctc_loss=0.2345, cr_loss=0.4118, over 3363369.74 frames. ], batch size: 56, lr: 2.77e-02, grad_scale: 32.0 2024-09-22 16:59:42,176 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=60657.333333333336, ans=0.125 2024-09-22 16:59:45,198 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=60704.0, ans=0.0 2024-09-22 16:59:48,139 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.293e+02 1.594e+02 1.780e+02 2.076e+02 4.160e+02, threshold=3.560e+02, percent-clipped=3.0 2024-09-22 17:00:31,113 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=60844.0, ans=0.125 2024-09-22 17:00:32,385 INFO [train.py:1198] (2/4) Epoch 4, batch 1350, loss[loss=0.3332, ctc_loss=0.2455, cr_loss=0.4385, over 17050.00 frames. ], tot_loss[loss=0.3169, ctc_loss=0.2345, cr_loss=0.4116, over 3369432.89 frames. ], batch size: 52, lr: 2.77e-02, grad_scale: 32.0 2024-09-22 17:00:34,231 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=60844.0, ans=0.1 2024-09-22 17:00:35,767 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=60844.0, ans=0.125 2024-09-22 17:00:43,202 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=60844.0, ans=0.1 2024-09-22 17:00:47,820 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=60844.0, ans=0.125 2024-09-22 17:00:47,943 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=60844.0, ans=0.125 2024-09-22 17:01:03,428 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=60890.666666666664, ans=0.0 2024-09-22 17:01:12,939 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=60937.333333333336, ans=0.125 2024-09-22 17:01:21,306 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=22.25 vs. limit=22.5 2024-09-22 17:01:57,166 INFO [train.py:1198] (2/4) Epoch 4, batch 1400, loss[loss=0.3515, ctc_loss=0.2661, cr_loss=0.4271, over 16732.00 frames. ], tot_loss[loss=0.3178, ctc_loss=0.2354, cr_loss=0.412, over 3354378.02 frames. ], batch size: 61, lr: 2.76e-02, grad_scale: 32.0 2024-09-22 17:02:34,721 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.284e+02 1.619e+02 1.850e+02 2.266e+02 3.949e+02, threshold=3.701e+02, percent-clipped=2.0 2024-09-22 17:03:16,345 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 17:03:22,329 INFO [train.py:1198] (2/4) Epoch 4, batch 1450, loss[loss=0.3775, ctc_loss=0.2915, cr_loss=0.4303, over 12165.00 frames. ], tot_loss[loss=0.3201, ctc_loss=0.2373, cr_loss=0.4141, over 3342115.22 frames. ], batch size: 123, lr: 2.76e-02, grad_scale: 32.0 2024-09-22 17:03:46,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=61357.333333333336, ans=0.0 2024-09-22 17:04:03,764 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=61404.0, ans=0.0 2024-09-22 17:04:12,182 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=4.98 vs. limit=12.0 2024-09-22 17:04:41,900 INFO [train.py:1198] (2/4) Epoch 4, batch 1500, loss[loss=0.2616, ctc_loss=0.1904, cr_loss=0.3559, over 16956.00 frames. ], tot_loss[loss=0.3192, ctc_loss=0.2365, cr_loss=0.4136, over 3341424.98 frames. ], batch size: 42, lr: 2.76e-02, grad_scale: 32.0 2024-09-22 17:04:50,510 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=13.23 vs. limit=15.0 2024-09-22 17:05:17,065 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.343e+02 1.547e+02 1.744e+02 2.028e+02 3.491e+02, threshold=3.489e+02, percent-clipped=0.0 2024-09-22 17:05:47,500 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=61684.0, ans=0.0 2024-09-22 17:06:03,748 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=9.73 vs. limit=15.0 2024-09-22 17:06:06,037 INFO [train.py:1198] (2/4) Epoch 4, batch 1550, loss[loss=0.2688, ctc_loss=0.1986, cr_loss=0.3512, over 17122.00 frames. ], tot_loss[loss=0.3183, ctc_loss=0.2358, cr_loss=0.4124, over 3340912.70 frames. ], batch size: 40, lr: 2.75e-02, grad_scale: 32.0 2024-09-22 17:06:14,489 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=61777.333333333336, ans=0.125 2024-09-22 17:06:23,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=61824.0, ans=0.0 2024-09-22 17:06:27,016 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=61824.0, ans=0.125 2024-09-22 17:06:52,460 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=61917.333333333336, ans=0.125 2024-09-22 17:07:26,974 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.whiten.whitening_limit, batch_count=62010.666666666664, ans=15.0 2024-09-22 17:07:27,833 INFO [train.py:1198] (2/4) Epoch 4, batch 1600, loss[loss=0.3165, ctc_loss=0.2309, cr_loss=0.4278, over 17299.00 frames. ], tot_loss[loss=0.3185, ctc_loss=0.2359, cr_loss=0.413, over 3349867.69 frames. ], batch size: 46, lr: 2.75e-02, grad_scale: 32.0 2024-09-22 17:07:39,649 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=62010.666666666664, ans=0.125 2024-09-22 17:07:42,894 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 17:07:42,895 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=62057.333333333336, ans=0.125 2024-09-22 17:08:05,555 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.293e+02 1.652e+02 1.885e+02 2.249e+02 4.170e+02, threshold=3.770e+02, percent-clipped=2.0 2024-09-22 17:08:49,401 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=3.12 vs. limit=15.0 2024-09-22 17:08:50,127 INFO [train.py:1198] (2/4) Epoch 4, batch 1650, loss[loss=0.3505, ctc_loss=0.2636, cr_loss=0.4342, over 17304.00 frames. ], tot_loss[loss=0.3161, ctc_loss=0.2338, cr_loss=0.4113, over 3363702.10 frames. ], batch size: 49, lr: 2.75e-02, grad_scale: 32.0 2024-09-22 17:08:52,080 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=62244.0, ans=0.125 2024-09-22 17:09:30,024 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=62337.333333333336, ans=0.1 2024-09-22 17:09:53,740 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 17:09:58,615 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=62430.666666666664, ans=0.2 2024-09-22 17:10:09,473 INFO [train.py:1198] (2/4) Epoch 4, batch 1700, loss[loss=0.3601, ctc_loss=0.2697, cr_loss=0.4523, over 16621.00 frames. ], tot_loss[loss=0.3194, ctc_loss=0.2367, cr_loss=0.4135, over 3355032.80 frames. ], batch size: 66, lr: 2.74e-02, grad_scale: 32.0 2024-09-22 17:10:14,882 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=6.90 vs. limit=15.0 2024-09-22 17:10:22,258 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=62477.333333333336, ans=0.125 2024-09-22 17:10:36,458 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer_na.min_abs, batch_count=62524.0, ans=0.02 2024-09-22 17:10:49,656 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.211e+02 1.564e+02 1.856e+02 2.208e+02 3.257e+02, threshold=3.711e+02, percent-clipped=0.0 2024-09-22 17:11:04,220 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=62617.333333333336, ans=0.125 2024-09-22 17:11:15,303 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=62617.333333333336, ans=0.025 2024-09-22 17:11:18,665 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=62664.0, ans=0.0 2024-09-22 17:11:34,080 INFO [train.py:1198] (2/4) Epoch 4, batch 1750, loss[loss=0.3174, ctc_loss=0.2315, cr_loss=0.4293, over 17027.00 frames. ], tot_loss[loss=0.3195, ctc_loss=0.2369, cr_loss=0.4132, over 3348496.62 frames. ], batch size: 51, lr: 2.74e-02, grad_scale: 32.0 2024-09-22 17:11:53,468 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=62757.333333333336, ans=0.0 2024-09-22 17:11:55,664 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=5.78 vs. limit=10.0 2024-09-22 17:11:56,641 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=62757.333333333336, ans=0.125 2024-09-22 17:12:04,733 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=62804.0, ans=0.0 2024-09-22 17:12:14,004 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=62804.0, ans=0.2 2024-09-22 17:12:40,526 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=62897.333333333336, ans=0.0 2024-09-22 17:12:42,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=62897.333333333336, ans=0.125 2024-09-22 17:12:43,948 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=62897.333333333336, ans=0.125 2024-09-22 17:12:54,482 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=62897.333333333336, ans=0.0 2024-09-22 17:12:58,852 INFO [train.py:1198] (2/4) Epoch 4, batch 1800, loss[loss=0.242, ctc_loss=0.177, cr_loss=0.3251, over 17032.00 frames. ], tot_loss[loss=0.3174, ctc_loss=0.2351, cr_loss=0.4115, over 3348220.42 frames. ], batch size: 39, lr: 2.73e-02, grad_scale: 32.0 2024-09-22 17:13:02,354 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 17:13:02,419 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer_ff3.min_abs, batch_count=62944.0, ans=0.2 2024-09-22 17:13:22,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=62990.666666666664, ans=0.2 2024-09-22 17:13:33,467 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.277e+02 1.628e+02 1.997e+02 2.584e+02 3.622e+02, threshold=3.995e+02, percent-clipped=0.0 2024-09-22 17:13:38,753 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=13.66 vs. limit=15.0 2024-09-22 17:13:40,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=63037.333333333336, ans=0.0 2024-09-22 17:13:48,318 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.23 vs. limit=6.0 2024-09-22 17:13:54,096 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=63084.0, ans=0.125 2024-09-22 17:14:17,812 INFO [train.py:1198] (2/4) Epoch 4, batch 1850, loss[loss=0.3283, ctc_loss=0.2431, cr_loss=0.4257, over 17040.00 frames. ], tot_loss[loss=0.3165, ctc_loss=0.2342, cr_loss=0.4118, over 3352723.61 frames. ], batch size: 52, lr: 2.73e-02, grad_scale: 32.0 2024-09-22 17:14:18,044 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=63177.333333333336, ans=0.0 2024-09-22 17:14:40,656 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.74 vs. limit=15.0 2024-09-22 17:14:49,615 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=63270.666666666664, ans=0.2 2024-09-22 17:14:51,242 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=63270.666666666664, ans=0.2 2024-09-22 17:15:05,326 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=63317.333333333336, ans=0.2 2024-09-22 17:15:08,599 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 17:15:09,037 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=20.56 vs. limit=22.5 2024-09-22 17:15:26,850 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=63364.0, ans=0.0 2024-09-22 17:15:41,416 INFO [train.py:1198] (2/4) Epoch 4, batch 1900, loss[loss=0.2776, ctc_loss=0.2014, cr_loss=0.3807, over 17044.00 frames. ], tot_loss[loss=0.3173, ctc_loss=0.2347, cr_loss=0.4131, over 3360359.07 frames. ], batch size: 39, lr: 2.73e-02, grad_scale: 32.0 2024-09-22 17:15:45,428 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=10.64 vs. limit=15.0 2024-09-22 17:16:12,004 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=63504.0, ans=0.0 2024-09-22 17:16:16,406 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.329e+02 1.606e+02 1.832e+02 2.161e+02 3.717e+02, threshold=3.664e+02, percent-clipped=0.0 2024-09-22 17:16:26,344 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=63504.0, ans=0.125 2024-09-22 17:16:26,365 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=63504.0, ans=0.125 2024-09-22 17:16:42,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=63550.666666666664, ans=0.0 2024-09-22 17:16:48,928 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=63597.333333333336, ans=0.0 2024-09-22 17:16:51,213 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.79 vs. limit=12.0 2024-09-22 17:16:58,482 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=63597.333333333336, ans=0.0 2024-09-22 17:17:01,342 INFO [train.py:1198] (2/4) Epoch 4, batch 1950, loss[loss=0.351, ctc_loss=0.2613, cr_loss=0.4483, over 14877.00 frames. ], tot_loss[loss=0.3192, ctc_loss=0.2362, cr_loss=0.4153, over 3366757.42 frames. ], batch size: 89, lr: 2.72e-02, grad_scale: 32.0 2024-09-22 17:17:16,394 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=63690.666666666664, ans=0.07 2024-09-22 17:17:21,437 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=7.06 vs. limit=8.0 2024-09-22 17:17:23,743 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=63690.666666666664, ans=0.09899494936611666 2024-09-22 17:17:38,089 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=63737.333333333336, ans=0.0 2024-09-22 17:17:41,933 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=14.65 vs. limit=15.0 2024-09-22 17:17:42,215 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.17 vs. limit=6.0 2024-09-22 17:17:51,961 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=13.58 vs. limit=15.0 2024-09-22 17:18:05,996 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=12.22 vs. limit=15.0 2024-09-22 17:18:07,113 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=63784.0, ans=0.0 2024-09-22 17:18:10,356 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.min_positive, batch_count=63830.666666666664, ans=0.025 2024-09-22 17:18:20,013 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=63830.666666666664, ans=0.0 2024-09-22 17:18:26,131 INFO [train.py:1198] (2/4) Epoch 4, batch 2000, loss[loss=0.3407, ctc_loss=0.2569, cr_loss=0.4191, over 14782.00 frames. ], tot_loss[loss=0.3191, ctc_loss=0.2362, cr_loss=0.4146, over 3363386.84 frames. ], batch size: 88, lr: 2.72e-02, grad_scale: 64.0 2024-09-22 17:18:37,862 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=63877.333333333336, ans=0.2 2024-09-22 17:18:58,383 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=63970.666666666664, ans=0.0 2024-09-22 17:19:01,203 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.280e+02 1.569e+02 1.789e+02 2.368e+02 3.802e+02, threshold=3.577e+02, percent-clipped=1.0 2024-09-22 17:19:04,763 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 17:19:34,776 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=64064.0, ans=0.0 2024-09-22 17:19:34,921 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=64064.0, ans=0.025 2024-09-22 17:19:38,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=64064.0, ans=0.1 2024-09-22 17:19:45,616 INFO [train.py:1198] (2/4) Epoch 4, batch 2050, loss[loss=0.3346, ctc_loss=0.2486, cr_loss=0.43, over 17359.00 frames. ], tot_loss[loss=0.3194, ctc_loss=0.2363, cr_loss=0.4156, over 3369944.31 frames. ], batch size: 48, lr: 2.71e-02, grad_scale: 64.0 2024-09-22 17:19:47,969 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=16.96 vs. limit=22.5 2024-09-22 17:20:29,974 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.min_positive, batch_count=64204.0, ans=0.025 2024-09-22 17:20:30,081 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=64204.0, ans=0.0 2024-09-22 17:20:33,267 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=64204.0, ans=10.0 2024-09-22 17:20:37,973 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.min_positive, batch_count=64250.666666666664, ans=0.025 2024-09-22 17:20:45,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=64250.666666666664, ans=0.0 2024-09-22 17:20:45,665 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=64250.666666666664, ans=0.0 2024-09-22 17:20:45,668 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=64250.666666666664, ans=0.125 2024-09-22 17:21:07,689 INFO [train.py:1198] (2/4) Epoch 4, batch 2100, loss[loss=0.3493, ctc_loss=0.2625, cr_loss=0.4344, over 17029.00 frames. ], tot_loss[loss=0.3195, ctc_loss=0.2364, cr_loss=0.4152, over 3360652.98 frames. ], batch size: 52, lr: 2.71e-02, grad_scale: 32.0 2024-09-22 17:21:09,595 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=64344.0, ans=0.1 2024-09-22 17:21:19,036 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=64344.0, ans=0.1 2024-09-22 17:21:40,073 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=64437.333333333336, ans=0.125 2024-09-22 17:21:44,599 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.356e+02 1.633e+02 1.973e+02 2.304e+02 3.408e+02, threshold=3.946e+02, percent-clipped=0.0 2024-09-22 17:21:54,421 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=64484.0, ans=0.0 2024-09-22 17:21:57,518 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=64484.0, ans=0.125 2024-09-22 17:22:21,125 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=64530.666666666664, ans=0.1 2024-09-22 17:22:33,062 INFO [train.py:1198] (2/4) Epoch 4, batch 2150, loss[loss=0.3595, ctc_loss=0.2701, cr_loss=0.4468, over 17027.00 frames. ], tot_loss[loss=0.3183, ctc_loss=0.2355, cr_loss=0.4142, over 3363871.49 frames. ], batch size: 56, lr: 2.71e-02, grad_scale: 32.0 2024-09-22 17:23:09,869 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=64670.666666666664, ans=0.125 2024-09-22 17:23:19,920 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=13.22 vs. limit=15.0 2024-09-22 17:23:51,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=64810.666666666664, ans=0.0 2024-09-22 17:23:52,466 INFO [train.py:1198] (2/4) Epoch 4, batch 2200, loss[loss=0.3531, ctc_loss=0.2644, cr_loss=0.4437, over 16994.00 frames. ], tot_loss[loss=0.3167, ctc_loss=0.2342, cr_loss=0.4129, over 3361975.24 frames. ], batch size: 53, lr: 2.70e-02, grad_scale: 32.0 2024-09-22 17:24:23,225 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=64904.0, ans=0.1 2024-09-22 17:24:29,093 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.320e+02 1.691e+02 2.009e+02 2.410e+02 3.639e+02, threshold=4.017e+02, percent-clipped=0.0 2024-09-22 17:25:11,834 INFO [train.py:1198] (2/4) Epoch 4, batch 2250, loss[loss=0.3272, ctc_loss=0.2468, cr_loss=0.402, over 17024.00 frames. ], tot_loss[loss=0.3158, ctc_loss=0.2334, cr_loss=0.4117, over 3358695.31 frames. ], batch size: 51, lr: 2.70e-02, grad_scale: 32.0 2024-09-22 17:25:12,371 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=14.26 vs. limit=15.0 2024-09-22 17:25:42,342 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=23.70 vs. limit=22.5 2024-09-22 17:25:52,722 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=65137.333333333336, ans=0.025 2024-09-22 17:25:53,186 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.69 vs. limit=15.0 2024-09-22 17:26:11,975 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=65184.0, ans=0.0 2024-09-22 17:26:33,875 INFO [train.py:1198] (2/4) Epoch 4, batch 2300, loss[loss=0.338, ctc_loss=0.2519, cr_loss=0.4307, over 17336.00 frames. ], tot_loss[loss=0.3169, ctc_loss=0.2342, cr_loss=0.4133, over 3353822.34 frames. ], batch size: 48, lr: 2.70e-02, grad_scale: 32.0 2024-09-22 17:27:03,200 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.84 vs. limit=15.0 2024-09-22 17:27:10,450 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=65370.666666666664, ans=0.0 2024-09-22 17:27:13,235 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.343e+02 1.664e+02 1.971e+02 2.314e+02 3.882e+02, threshold=3.942e+02, percent-clipped=0.0 2024-09-22 17:27:58,573 INFO [train.py:1198] (2/4) Epoch 4, batch 2350, loss[loss=0.3209, ctc_loss=0.2372, cr_loss=0.4183, over 17147.00 frames. ], tot_loss[loss=0.3159, ctc_loss=0.2332, cr_loss=0.4136, over 3358925.88 frames. ], batch size: 48, lr: 2.69e-02, grad_scale: 32.0 2024-09-22 17:28:12,193 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=7.61 vs. limit=15.0 2024-09-22 17:28:26,429 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=11.95 vs. limit=15.0 2024-09-22 17:28:32,614 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=65604.0, ans=0.125 2024-09-22 17:28:45,335 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-22 17:28:49,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=65650.66666666667, ans=0.0 2024-09-22 17:28:56,362 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=65650.66666666667, ans=0.2 2024-09-22 17:29:18,256 INFO [train.py:1198] (2/4) Epoch 4, batch 2400, loss[loss=0.276, ctc_loss=0.2015, cr_loss=0.3722, over 17200.00 frames. ], tot_loss[loss=0.315, ctc_loss=0.2325, cr_loss=0.4125, over 3364176.29 frames. ], batch size: 41, lr: 2.69e-02, grad_scale: 32.0 2024-09-22 17:29:28,222 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=65744.0, ans=0.2 2024-09-22 17:29:54,854 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.273e+02 1.593e+02 1.794e+02 2.177e+02 3.793e+02, threshold=3.589e+02, percent-clipped=0.0 2024-09-22 17:29:59,917 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=65837.33333333333, ans=0.1 2024-09-22 17:30:10,598 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=65884.0, ans=0.0 2024-09-22 17:30:21,751 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=65884.0, ans=0.125 2024-09-22 17:30:36,168 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.36 vs. limit=10.0 2024-09-22 17:30:40,198 INFO [train.py:1198] (2/4) Epoch 4, batch 2450, loss[loss=0.2764, ctc_loss=0.2049, cr_loss=0.3575, over 17089.00 frames. ], tot_loss[loss=0.3138, ctc_loss=0.2314, cr_loss=0.4123, over 3369480.92 frames. ], batch size: 43, lr: 2.68e-02, grad_scale: 32.0 2024-09-22 17:31:00,487 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=4.36 vs. limit=15.0 2024-09-22 17:31:28,295 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=66117.33333333333, ans=0.0 2024-09-22 17:31:42,724 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=66164.0, ans=0.2 2024-09-22 17:31:42,730 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=66164.0, ans=0.09899494936611666 2024-09-22 17:31:49,414 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=15.52 vs. limit=22.5 2024-09-22 17:32:02,189 INFO [train.py:1198] (2/4) Epoch 4, batch 2500, loss[loss=0.3973, ctc_loss=0.3127, cr_loss=0.4231, over 11786.00 frames. ], tot_loss[loss=0.315, ctc_loss=0.2324, cr_loss=0.4128, over 3361572.63 frames. ], batch size: 123, lr: 2.68e-02, grad_scale: 32.0 2024-09-22 17:32:17,866 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 17:32:19,444 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=66257.33333333333, ans=0.125 2024-09-22 17:32:41,646 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.315e+02 1.715e+02 1.996e+02 2.438e+02 3.886e+02, threshold=3.992e+02, percent-clipped=3.0 2024-09-22 17:32:41,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=66304.0, ans=0.1 2024-09-22 17:32:41,930 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=66304.0, ans=0.125 2024-09-22 17:32:46,642 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=66304.0, ans=0.1 2024-09-22 17:33:01,613 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=6.42 vs. limit=15.0 2024-09-22 17:33:16,838 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=66397.33333333333, ans=0.2 2024-09-22 17:33:24,679 INFO [train.py:1198] (2/4) Epoch 4, batch 2550, loss[loss=0.3116, ctc_loss=0.231, cr_loss=0.4031, over 17235.00 frames. ], tot_loss[loss=0.314, ctc_loss=0.2316, cr_loss=0.4123, over 3360577.89 frames. ], batch size: 55, lr: 2.68e-02, grad_scale: 32.0 2024-09-22 17:33:28,170 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=66444.0, ans=0.125 2024-09-22 17:34:44,461 INFO [train.py:1198] (2/4) Epoch 4, batch 2600, loss[loss=0.3259, ctc_loss=0.2422, cr_loss=0.4184, over 17008.00 frames. ], tot_loss[loss=0.3145, ctc_loss=0.232, cr_loss=0.4125, over 3366592.56 frames. ], batch size: 51, lr: 2.67e-02, grad_scale: 32.0 2024-09-22 17:34:55,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=66677.33333333333, ans=0.0 2024-09-22 17:35:10,214 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.max_positive, batch_count=66724.0, ans=0.95 2024-09-22 17:35:25,830 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.648e+02 1.831e+02 2.212e+02 5.606e+02, threshold=3.662e+02, percent-clipped=1.0 2024-09-22 17:35:32,528 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=66770.66666666667, ans=0.125 2024-09-22 17:35:34,100 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=66770.66666666667, ans=0.025 2024-09-22 17:35:46,720 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=66817.33333333333, ans=0.2 2024-09-22 17:36:00,960 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=66864.0, ans=0.125 2024-09-22 17:36:08,650 INFO [train.py:1198] (2/4) Epoch 4, batch 2650, loss[loss=0.3022, ctc_loss=0.2195, cr_loss=0.4136, over 17288.00 frames. ], tot_loss[loss=0.3134, ctc_loss=0.2311, cr_loss=0.4117, over 3365461.19 frames. ], batch size: 46, lr: 2.67e-02, grad_scale: 32.0 2024-09-22 17:36:23,222 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=66957.33333333333, ans=0.2 2024-09-22 17:36:37,948 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.45 vs. limit=6.0 2024-09-22 17:36:52,691 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=67004.0, ans=0.025 2024-09-22 17:36:54,290 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=67004.0, ans=0.125 2024-09-22 17:37:06,413 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=67050.66666666667, ans=0.125 2024-09-22 17:37:18,624 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=67097.33333333333, ans=0.07 2024-09-22 17:37:32,620 INFO [train.py:1198] (2/4) Epoch 4, batch 2700, loss[loss=0.2608, ctc_loss=0.1884, cr_loss=0.3619, over 17031.00 frames. ], tot_loss[loss=0.3136, ctc_loss=0.2313, cr_loss=0.4119, over 3363902.44 frames. ], batch size: 39, lr: 2.67e-02, grad_scale: 32.0 2024-09-22 17:38:01,519 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=67190.66666666667, ans=0.2 2024-09-22 17:38:09,012 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.345e+02 1.678e+02 1.945e+02 2.373e+02 3.767e+02, threshold=3.890e+02, percent-clipped=1.0 2024-09-22 17:38:52,039 INFO [train.py:1198] (2/4) Epoch 4, batch 2750, loss[loss=0.3348, ctc_loss=0.2489, cr_loss=0.4295, over 16625.00 frames. ], tot_loss[loss=0.3153, ctc_loss=0.2326, cr_loss=0.4136, over 3362481.25 frames. ], batch size: 66, lr: 2.66e-02, grad_scale: 32.0 2024-09-22 17:38:58,882 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=67377.33333333333, ans=0.125 2024-09-22 17:39:29,088 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=67470.66666666667, ans=0.2 2024-09-22 17:39:45,350 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.64 vs. limit=10.0 2024-09-22 17:39:48,450 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=12.48 vs. limit=15.0 2024-09-22 17:39:52,892 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=67517.33333333333, ans=0.125 2024-09-22 17:40:04,720 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=67564.0, ans=0.0 2024-09-22 17:40:17,122 INFO [train.py:1198] (2/4) Epoch 4, batch 2800, loss[loss=0.3409, ctc_loss=0.2513, cr_loss=0.4482, over 17022.00 frames. ], tot_loss[loss=0.3151, ctc_loss=0.2325, cr_loss=0.4131, over 3367301.36 frames. ], batch size: 53, lr: 2.66e-02, grad_scale: 32.0 2024-09-22 17:40:20,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=67610.66666666667, ans=0.025 2024-09-22 17:40:28,245 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=67610.66666666667, ans=0.1 2024-09-22 17:40:53,564 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.183e+02 1.604e+02 1.847e+02 2.270e+02 3.677e+02, threshold=3.693e+02, percent-clipped=0.0 2024-09-22 17:41:12,981 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.69 vs. limit=6.0 2024-09-22 17:41:22,065 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=67797.33333333333, ans=0.04949747468305833 2024-09-22 17:41:38,543 INFO [train.py:1198] (2/4) Epoch 4, batch 2850, loss[loss=0.4177, ctc_loss=0.3307, cr_loss=0.4351, over 11407.00 frames. ], tot_loss[loss=0.3161, ctc_loss=0.2333, cr_loss=0.4141, over 3366324.04 frames. ], batch size: 124, lr: 2.65e-02, grad_scale: 32.0 2024-09-22 17:41:59,186 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=67890.66666666667, ans=0.0 2024-09-22 17:42:13,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=67937.33333333333, ans=0.125 2024-09-22 17:42:18,549 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=9.01 vs. limit=15.0 2024-09-22 17:42:26,113 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=67937.33333333333, ans=0.0 2024-09-22 17:42:27,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=67984.0, ans=0.09899494936611666 2024-09-22 17:42:29,216 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=67984.0, ans=0.2 2024-09-22 17:42:48,448 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=68030.66666666667, ans=0.125 2024-09-22 17:42:52,280 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.86 vs. limit=12.0 2024-09-22 17:43:00,658 INFO [train.py:1198] (2/4) Epoch 4, batch 2900, loss[loss=0.3026, ctc_loss=0.2195, cr_loss=0.4154, over 16693.00 frames. ], tot_loss[loss=0.3157, ctc_loss=0.2331, cr_loss=0.4134, over 3368328.08 frames. ], batch size: 37, lr: 2.65e-02, grad_scale: 32.0 2024-09-22 17:43:15,940 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=11.66 vs. limit=15.0 2024-09-22 17:43:19,146 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=7.34 vs. limit=15.0 2024-09-22 17:43:37,708 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.237e+02 1.627e+02 1.915e+02 2.361e+02 4.224e+02, threshold=3.831e+02, percent-clipped=1.0 2024-09-22 17:43:52,062 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=68217.33333333333, ans=0.125 2024-09-22 17:43:55,141 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=68217.33333333333, ans=0.0 2024-09-22 17:44:02,965 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 17:44:14,346 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=68264.0, ans=0.125 2024-09-22 17:44:20,377 INFO [train.py:1198] (2/4) Epoch 4, batch 2950, loss[loss=0.3714, ctc_loss=0.292, cr_loss=0.3972, over 11924.00 frames. ], tot_loss[loss=0.3161, ctc_loss=0.2334, cr_loss=0.4135, over 3367205.42 frames. ], batch size: 123, lr: 2.65e-02, grad_scale: 32.0 2024-09-22 17:44:22,502 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=68310.66666666667, ans=0.125 2024-09-22 17:44:31,801 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=68310.66666666667, ans=0.2 2024-09-22 17:44:36,691 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=68357.33333333333, ans=0.125 2024-09-22 17:45:12,799 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten.whitening_limit, batch_count=68450.66666666667, ans=15.0 2024-09-22 17:45:16,971 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=68450.66666666667, ans=0.1 2024-09-22 17:45:44,886 INFO [train.py:1198] (2/4) Epoch 4, batch 3000, loss[loss=0.2715, ctc_loss=0.195, cr_loss=0.3827, over 17284.00 frames. ], tot_loss[loss=0.3143, ctc_loss=0.2319, cr_loss=0.412, over 3368174.61 frames. ], batch size: 42, lr: 2.64e-02, grad_scale: 32.0 2024-09-22 17:45:44,886 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-22 17:46:00,411 INFO [train.py:1230] (2/4) Epoch 4, validation: loss=0.07263, ctc_loss=0.07263, cr_loss=7.17e-15, over 944034.00 frames. 2024-09-22 17:46:00,412 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-22 17:46:11,665 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=68544.0, ans=0.1 2024-09-22 17:46:16,377 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=68590.66666666667, ans=0.125 2024-09-22 17:46:36,347 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.234e+02 1.618e+02 1.808e+02 2.126e+02 4.273e+02, threshold=3.616e+02, percent-clipped=2.0 2024-09-22 17:47:19,106 INFO [train.py:1198] (2/4) Epoch 4, batch 3050, loss[loss=0.3298, ctc_loss=0.245, cr_loss=0.4241, over 17234.00 frames. ], tot_loss[loss=0.313, ctc_loss=0.231, cr_loss=0.4104, over 3360256.57 frames. ], batch size: 50, lr: 2.64e-02, grad_scale: 32.0 2024-09-22 17:47:31,926 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=68777.33333333333, ans=0.0 2024-09-22 17:47:39,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=68824.0, ans=0.125 2024-09-22 17:47:49,511 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=3.43 vs. limit=15.0 2024-09-22 17:47:58,196 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=68870.66666666667, ans=0.025 2024-09-22 17:48:21,950 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=68917.33333333333, ans=0.1 2024-09-22 17:48:38,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=68964.0, ans=0.0 2024-09-22 17:48:41,793 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=69010.66666666667, ans=0.025 2024-09-22 17:48:42,985 INFO [train.py:1198] (2/4) Epoch 4, batch 3100, loss[loss=0.3465, ctc_loss=0.2515, cr_loss=0.4751, over 17003.00 frames. ], tot_loss[loss=0.3127, ctc_loss=0.2305, cr_loss=0.411, over 3362941.63 frames. ], batch size: 53, lr: 2.64e-02, grad_scale: 32.0 2024-09-22 17:48:44,791 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=69010.66666666667, ans=0.125 2024-09-22 17:49:04,053 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=69057.33333333333, ans=0.125 2024-09-22 17:49:19,305 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.276e+02 1.562e+02 1.773e+02 2.272e+02 4.016e+02, threshold=3.545e+02, percent-clipped=1.0 2024-09-22 17:49:27,349 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=69104.0, ans=0.07 2024-09-22 17:49:49,453 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=69197.33333333333, ans=0.1 2024-09-22 17:50:01,897 INFO [train.py:1198] (2/4) Epoch 4, batch 3150, loss[loss=0.3095, ctc_loss=0.2297, cr_loss=0.3989, over 17140.00 frames. ], tot_loss[loss=0.3119, ctc_loss=0.2299, cr_loss=0.4102, over 3364464.82 frames. ], batch size: 48, lr: 2.63e-02, grad_scale: 32.0 2024-09-22 17:50:06,020 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.95 vs. limit=15.0 2024-09-22 17:50:17,391 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=69290.66666666667, ans=0.2 2024-09-22 17:50:24,084 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.62 vs. limit=6.0 2024-09-22 17:50:44,411 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module1.whiten, num_groups=1, num_channels=512, metric=7.04 vs. limit=15.0 2024-09-22 17:51:02,832 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=69430.66666666667, ans=0.0 2024-09-22 17:51:04,420 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=69430.66666666667, ans=0.125 2024-09-22 17:51:19,753 INFO [train.py:1198] (2/4) Epoch 4, batch 3200, loss[loss=0.2538, ctc_loss=0.1864, cr_loss=0.337, over 16693.00 frames. ], tot_loss[loss=0.3111, ctc_loss=0.2292, cr_loss=0.4093, over 3362566.15 frames. ], batch size: 37, lr: 2.63e-02, grad_scale: 32.0 2024-09-22 17:51:37,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=69524.0, ans=0.2 2024-09-22 17:51:54,434 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=69570.66666666667, ans=0.1 2024-09-22 17:51:55,750 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.313e+02 1.584e+02 1.798e+02 2.186e+02 3.575e+02, threshold=3.596e+02, percent-clipped=1.0 2024-09-22 17:51:59,082 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=69570.66666666667, ans=0.1 2024-09-22 17:52:00,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=69570.66666666667, ans=0.125 2024-09-22 17:52:18,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=69617.33333333333, ans=0.0 2024-09-22 17:52:21,500 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=12.66 vs. limit=15.0 2024-09-22 17:52:37,931 INFO [train.py:1198] (2/4) Epoch 4, batch 3250, loss[loss=0.3364, ctc_loss=0.241, cr_loss=0.4769, over 17061.00 frames. ], tot_loss[loss=0.3134, ctc_loss=0.2311, cr_loss=0.4117, over 3357715.06 frames. ], batch size: 52, lr: 2.63e-02, grad_scale: 32.0 2024-09-22 17:52:50,239 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=9.48 vs. limit=10.0 2024-09-22 17:52:55,466 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=69757.33333333333, ans=0.125 2024-09-22 17:52:58,397 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=69757.33333333333, ans=0.125 2024-09-22 17:53:00,539 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=5.26 vs. limit=10.0 2024-09-22 17:53:11,124 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=69804.0, ans=0.1 2024-09-22 17:53:23,760 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=69850.66666666667, ans=0.125 2024-09-22 17:53:23,823 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=69850.66666666667, ans=0.1 2024-09-22 17:53:29,799 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=69850.66666666667, ans=0.125 2024-09-22 17:53:32,972 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=69850.66666666667, ans=0.0 2024-09-22 17:53:38,280 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.whiten, num_groups=1, num_channels=192, metric=4.10 vs. limit=12.0 2024-09-22 17:53:44,236 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.01 vs. limit=15.0 2024-09-22 17:53:56,009 INFO [train.py:1198] (2/4) Epoch 4, batch 3300, loss[loss=0.2984, ctc_loss=0.2178, cr_loss=0.4028, over 17038.00 frames. ], tot_loss[loss=0.3142, ctc_loss=0.2318, cr_loss=0.4123, over 3351512.91 frames. ], batch size: 44, lr: 2.62e-02, grad_scale: 32.0 2024-09-22 17:54:29,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=70037.33333333333, ans=0.2 2024-09-22 17:54:32,122 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.316e+02 1.544e+02 1.731e+02 2.121e+02 3.523e+02, threshold=3.462e+02, percent-clipped=0.0 2024-09-22 17:55:19,190 INFO [train.py:1198] (2/4) Epoch 4, batch 3350, loss[loss=0.3603, ctc_loss=0.2722, cr_loss=0.4406, over 16574.00 frames. ], tot_loss[loss=0.3151, ctc_loss=0.2325, cr_loss=0.4131, over 3343206.85 frames. ], batch size: 66, lr: 2.62e-02, grad_scale: 32.0 2024-09-22 17:55:28,614 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=70177.33333333333, ans=0.125 2024-09-22 17:55:36,242 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=70224.0, ans=0.07 2024-09-22 17:55:40,867 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=70224.0, ans=0.2 2024-09-22 17:56:02,719 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=70270.66666666667, ans=0.1 2024-09-22 17:56:36,977 INFO [train.py:1198] (2/4) Epoch 4, batch 3400, loss[loss=0.4176, ctc_loss=0.3287, cr_loss=0.4445, over 11500.00 frames. ], tot_loss[loss=0.3164, ctc_loss=0.2335, cr_loss=0.4141, over 3329735.29 frames. ], batch size: 123, lr: 2.62e-02, grad_scale: 32.0 2024-09-22 17:56:39,031 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=12.61 vs. limit=15.0 2024-09-22 17:56:49,632 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=70410.66666666667, ans=0.1 2024-09-22 17:56:51,686 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=22.12 vs. limit=22.5 2024-09-22 17:56:57,324 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=70457.33333333333, ans=0.0 2024-09-22 17:57:11,110 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=70504.0, ans=0.025 2024-09-22 17:57:12,457 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.333e+02 1.518e+02 1.665e+02 2.014e+02 3.333e+02, threshold=3.330e+02, percent-clipped=0.0 2024-09-22 17:57:21,867 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=70550.66666666667, ans=0.025 2024-09-22 17:57:32,640 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=70550.66666666667, ans=0.125 2024-09-22 17:57:54,208 INFO [train.py:1198] (2/4) Epoch 4, batch 3450, loss[loss=0.3504, ctc_loss=0.2588, cr_loss=0.4577, over 17022.00 frames. ], tot_loss[loss=0.3149, ctc_loss=0.2323, cr_loss=0.4131, over 3338140.76 frames. ], batch size: 51, lr: 2.61e-02, grad_scale: 32.0 2024-09-22 17:58:22,429 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=70690.66666666667, ans=0.125 2024-09-22 17:58:30,371 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=70737.33333333333, ans=0.015 2024-09-22 17:58:33,757 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=70737.33333333333, ans=0.125 2024-09-22 17:58:33,773 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=70737.33333333333, ans=0.125 2024-09-22 17:59:15,855 INFO [train.py:1198] (2/4) Epoch 4, batch 3500, loss[loss=0.3127, ctc_loss=0.2273, cr_loss=0.4272, over 17085.00 frames. ], tot_loss[loss=0.3148, ctc_loss=0.2321, cr_loss=0.4137, over 3344485.60 frames. ], batch size: 49, lr: 2.61e-02, grad_scale: 32.0 2024-09-22 17:59:22,120 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=70877.33333333333, ans=0.1 2024-09-22 17:59:31,379 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=70924.0, ans=0.0 2024-09-22 17:59:31,853 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.09 vs. limit=6.0 2024-09-22 17:59:39,646 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=70924.0, ans=0.125 2024-09-22 17:59:45,800 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=70970.66666666667, ans=0.2 2024-09-22 17:59:47,376 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=70970.66666666667, ans=0.125 2024-09-22 17:59:53,259 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.280e+02 1.551e+02 1.744e+02 2.057e+02 3.215e+02, threshold=3.488e+02, percent-clipped=0.0 2024-09-22 18:00:04,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=71017.33333333333, ans=0.1 2024-09-22 18:00:10,862 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=71017.33333333333, ans=0.125 2024-09-22 18:00:18,771 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=71064.0, ans=0.04949747468305833 2024-09-22 18:00:21,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=71064.0, ans=0.0 2024-09-22 18:00:33,938 INFO [train.py:1198] (2/4) Epoch 4, batch 3550, loss[loss=0.3504, ctc_loss=0.2675, cr_loss=0.4145, over 15053.00 frames. ], tot_loss[loss=0.3134, ctc_loss=0.2309, cr_loss=0.4128, over 3351171.84 frames. ], batch size: 89, lr: 2.61e-02, grad_scale: 16.0 2024-09-22 18:00:40,478 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=71110.66666666667, ans=0.125 2024-09-22 18:01:08,585 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten.whitening_limit, batch_count=71204.0, ans=15.0 2024-09-22 18:01:15,979 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=71204.0, ans=0.125 2024-09-22 18:01:20,844 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=71250.66666666667, ans=0.2 2024-09-22 18:01:51,386 INFO [train.py:1198] (2/4) Epoch 4, batch 3600, loss[loss=0.3235, ctc_loss=0.2432, cr_loss=0.4015, over 14870.00 frames. ], tot_loss[loss=0.3134, ctc_loss=0.2308, cr_loss=0.4129, over 3351120.31 frames. ], batch size: 89, lr: 2.60e-02, grad_scale: 32.0 2024-09-22 18:01:53,681 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=512, metric=4.66 vs. limit=15.0 2024-09-22 18:02:19,657 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=5.39 vs. limit=12.0 2024-09-22 18:02:25,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=71437.33333333333, ans=0.125 2024-09-22 18:02:28,362 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.219e+02 1.639e+02 2.046e+02 2.808e+02 4.325e+02, threshold=4.093e+02, percent-clipped=8.0 2024-09-22 18:02:35,222 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.63 vs. limit=10.0 2024-09-22 18:02:42,984 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.45 vs. limit=12.0 2024-09-22 18:02:47,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=71484.0, ans=0.125 2024-09-22 18:02:55,243 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=71530.66666666667, ans=0.025 2024-09-22 18:03:02,936 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=71530.66666666667, ans=0.0 2024-09-22 18:03:09,092 INFO [train.py:1198] (2/4) Epoch 4, batch 3650, loss[loss=0.3648, ctc_loss=0.273, cr_loss=0.4592, over 16538.00 frames. ], tot_loss[loss=0.3137, ctc_loss=0.2312, cr_loss=0.4121, over 3342189.01 frames. ], batch size: 66, lr: 2.60e-02, grad_scale: 32.0 2024-09-22 18:03:23,294 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=71624.0, ans=0.1 2024-09-22 18:04:24,427 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=71764.0, ans=0.125 2024-09-22 18:04:25,913 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=71764.0, ans=0.125 2024-09-22 18:04:28,660 INFO [train.py:1198] (2/4) Epoch 4, batch 3700, loss[loss=0.3574, ctc_loss=0.2688, cr_loss=0.443, over 17144.00 frames. ], tot_loss[loss=0.3134, ctc_loss=0.231, cr_loss=0.412, over 3349861.87 frames. ], batch size: 48, lr: 2.60e-02, grad_scale: 32.0 2024-09-22 18:04:44,118 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=71857.33333333333, ans=0.125 2024-09-22 18:04:49,783 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.27 vs. limit=6.0 2024-09-22 18:05:07,669 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.267e+02 1.579e+02 1.797e+02 2.044e+02 5.255e+02, threshold=3.594e+02, percent-clipped=1.0 2024-09-22 18:05:33,417 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=13.60 vs. limit=12.0 2024-09-22 18:05:36,839 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.28 vs. limit=12.0 2024-09-22 18:05:48,767 INFO [train.py:1198] (2/4) Epoch 4, batch 3750, loss[loss=0.2985, ctc_loss=0.2139, cr_loss=0.423, over 17169.00 frames. ], tot_loss[loss=0.312, ctc_loss=0.2298, cr_loss=0.4111, over 3347335.51 frames. ], batch size: 45, lr: 2.59e-02, grad_scale: 32.0 2024-09-22 18:06:04,674 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=72090.66666666667, ans=0.2 2024-09-22 18:06:14,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=72090.66666666667, ans=0.125 2024-09-22 18:06:23,332 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=72137.33333333333, ans=0.0 2024-09-22 18:07:05,149 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=72277.33333333333, ans=0.125 2024-09-22 18:07:06,329 INFO [train.py:1198] (2/4) Epoch 4, batch 3800, loss[loss=0.3774, ctc_loss=0.2963, cr_loss=0.4056, over 11735.00 frames. ], tot_loss[loss=0.313, ctc_loss=0.2309, cr_loss=0.4109, over 3334157.43 frames. ], batch size: 123, lr: 2.59e-02, grad_scale: 32.0 2024-09-22 18:07:09,793 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=72277.33333333333, ans=0.125 2024-09-22 18:07:24,140 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.42 vs. limit=12.0 2024-09-22 18:07:25,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=72324.0, ans=0.0 2024-09-22 18:07:31,322 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=72324.0, ans=0.0 2024-09-22 18:07:35,749 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=8.22 vs. limit=15.0 2024-09-22 18:07:44,153 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.309e+02 1.589e+02 1.725e+02 2.068e+02 4.482e+02, threshold=3.450e+02, percent-clipped=2.0 2024-09-22 18:07:55,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=72417.33333333333, ans=0.09899494936611666 2024-09-22 18:07:57,440 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=72417.33333333333, ans=0.2 2024-09-22 18:07:58,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=72417.33333333333, ans=0.95 2024-09-22 18:08:05,115 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=72417.33333333333, ans=0.1 2024-09-22 18:08:14,184 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=72464.0, ans=0.125 2024-09-22 18:08:22,466 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=72464.0, ans=0.0 2024-09-22 18:08:25,284 INFO [train.py:1198] (2/4) Epoch 4, batch 3850, loss[loss=0.2909, ctc_loss=0.2073, cr_loss=0.4175, over 16784.00 frames. ], tot_loss[loss=0.3139, ctc_loss=0.2317, cr_loss=0.4109, over 3299331.90 frames. ], batch size: 37, lr: 2.59e-02, grad_scale: 32.0 2024-09-22 18:08:41,125 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=72557.33333333333, ans=0.125 2024-09-22 18:09:04,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=72604.0, ans=0.125 2024-09-22 18:09:07,722 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-22 18:09:07,732 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-22 18:09:09,784 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.80 vs. limit=6.0 2024-09-22 18:09:33,302 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=72697.33333333333, ans=0.09899494936611666 2024-09-22 18:09:33,330 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=72697.33333333333, ans=0.0 2024-09-22 18:10:26,642 INFO [train.py:1198] (2/4) Epoch 5, batch 0, loss[loss=0.3811, ctc_loss=0.2896, cr_loss=0.4573, over 16558.00 frames. ], tot_loss[loss=0.3811, ctc_loss=0.2896, cr_loss=0.4573, over 16558.00 frames. ], batch size: 66, lr: 2.40e-02, grad_scale: 32.0 2024-09-22 18:10:26,643 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-22 18:10:42,116 INFO [train.py:1230] (2/4) Epoch 5, validation: loss=0.07551, ctc_loss=0.07551, cr_loss=6.915e-15, over 944034.00 frames. 2024-09-22 18:10:42,117 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-22 18:11:17,188 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module1.whiten, num_groups=1, num_channels=192, metric=7.39 vs. limit=15.0 2024-09-22 18:11:27,091 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.296e+02 1.660e+02 1.848e+02 2.232e+02 4.613e+02, threshold=3.696e+02, percent-clipped=4.0 2024-09-22 18:12:02,272 INFO [train.py:1198] (2/4) Epoch 5, batch 50, loss[loss=0.2975, ctc_loss=0.2201, cr_loss=0.3873, over 16722.00 frames. ], tot_loss[loss=0.3149, ctc_loss=0.2315, cr_loss=0.417, over 758849.33 frames. ], batch size: 61, lr: 2.40e-02, grad_scale: 32.0 2024-09-22 18:12:37,471 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=73052.0, ans=0.0 2024-09-22 18:12:39,110 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=73052.0, ans=0.125 2024-09-22 18:12:42,259 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=73052.0, ans=0.0 2024-09-22 18:13:02,782 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=73098.66666666667, ans=0.125 2024-09-22 18:13:26,263 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=73192.0, ans=0.2 2024-09-22 18:13:27,566 INFO [train.py:1198] (2/4) Epoch 5, batch 100, loss[loss=0.3217, ctc_loss=0.2335, cr_loss=0.441, over 16521.00 frames. ], tot_loss[loss=0.3089, ctc_loss=0.2264, cr_loss=0.4122, over 1336034.78 frames. ], batch size: 66, lr: 2.40e-02, grad_scale: 32.0 2024-09-22 18:13:29,741 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=16.80 vs. limit=22.5 2024-09-22 18:13:37,432 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=73192.0, ans=0.125 2024-09-22 18:13:41,247 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.96 vs. limit=10.0 2024-09-22 18:13:46,722 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=73238.66666666667, ans=0.125 2024-09-22 18:14:12,042 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.270e+02 1.507e+02 1.769e+02 2.148e+02 4.396e+02, threshold=3.538e+02, percent-clipped=1.0 2024-09-22 18:14:36,605 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=10.80 vs. limit=15.0 2024-09-22 18:14:46,987 INFO [train.py:1198] (2/4) Epoch 5, batch 150, loss[loss=0.2787, ctc_loss=0.2046, cr_loss=0.3705, over 17308.00 frames. ], tot_loss[loss=0.3055, ctc_loss=0.224, cr_loss=0.4074, over 1785847.87 frames. ], batch size: 49, lr: 2.40e-02, grad_scale: 32.0 2024-09-22 18:14:53,672 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=73425.33333333333, ans=0.025 2024-09-22 18:14:58,640 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=73425.33333333333, ans=0.125 2024-09-22 18:15:00,269 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=73425.33333333333, ans=0.07 2024-09-22 18:15:35,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=73518.66666666667, ans=0.125 2024-09-22 18:15:52,509 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=73565.33333333333, ans=0.0 2024-09-22 18:16:02,823 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=7.74 vs. limit=15.0 2024-09-22 18:16:11,572 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 18:16:12,903 INFO [train.py:1198] (2/4) Epoch 5, batch 200, loss[loss=0.299, ctc_loss=0.2185, cr_loss=0.4026, over 17285.00 frames. ], tot_loss[loss=0.3067, ctc_loss=0.2248, cr_loss=0.4092, over 2143253.16 frames. ], batch size: 46, lr: 2.39e-02, grad_scale: 32.0 2024-09-22 18:16:23,139 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.23 vs. limit=6.0 2024-09-22 18:16:24,307 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=73658.66666666667, ans=0.125 2024-09-22 18:16:57,603 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.221e+02 1.454e+02 1.719e+02 2.223e+02 3.373e+02, threshold=3.438e+02, percent-clipped=0.0 2024-09-22 18:17:12,224 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 18:17:13,827 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=73798.66666666667, ans=0.0 2024-09-22 18:17:32,285 INFO [train.py:1198] (2/4) Epoch 5, batch 250, loss[loss=0.2833, ctc_loss=0.2018, cr_loss=0.4077, over 17027.00 frames. ], tot_loss[loss=0.306, ctc_loss=0.2241, cr_loss=0.4098, over 2416356.57 frames. ], batch size: 44, lr: 2.39e-02, grad_scale: 32.0 2024-09-22 18:18:57,276 INFO [train.py:1198] (2/4) Epoch 5, batch 300, loss[loss=0.2957, ctc_loss=0.2221, cr_loss=0.3677, over 16712.00 frames. ], tot_loss[loss=0.3056, ctc_loss=0.2238, cr_loss=0.4092, over 2627581.52 frames. ], batch size: 61, lr: 2.39e-02, grad_scale: 32.0 2024-09-22 18:19:16,578 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=74172.0, ans=0.0 2024-09-22 18:19:20,329 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.09 vs. limit=15.0 2024-09-22 18:19:36,106 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=10.67 vs. limit=15.0 2024-09-22 18:19:38,965 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=74218.66666666667, ans=0.0 2024-09-22 18:19:40,469 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=74218.66666666667, ans=0.0 2024-09-22 18:19:41,767 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.262e+02 1.584e+02 1.929e+02 2.265e+02 4.720e+02, threshold=3.859e+02, percent-clipped=2.0 2024-09-22 18:20:19,148 INFO [train.py:1198] (2/4) Epoch 5, batch 350, loss[loss=0.2957, ctc_loss=0.2157, cr_loss=0.3998, over 17076.00 frames. ], tot_loss[loss=0.3074, ctc_loss=0.2253, cr_loss=0.4105, over 2787088.82 frames. ], batch size: 39, lr: 2.38e-02, grad_scale: 32.0 2024-09-22 18:20:27,443 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=74358.66666666667, ans=0.1 2024-09-22 18:21:18,118 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=74498.66666666667, ans=0.0 2024-09-22 18:21:27,360 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=74545.33333333333, ans=0.125 2024-09-22 18:21:41,486 INFO [train.py:1198] (2/4) Epoch 5, batch 400, loss[loss=0.3303, ctc_loss=0.2453, cr_loss=0.4248, over 17006.00 frames. ], tot_loss[loss=0.3064, ctc_loss=0.2246, cr_loss=0.4092, over 2917930.55 frames. ], batch size: 56, lr: 2.38e-02, grad_scale: 32.0 2024-09-22 18:21:46,436 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=74592.0, ans=0.125 2024-09-22 18:22:12,582 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=74638.66666666667, ans=0.125 2024-09-22 18:22:28,036 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.233e+02 1.516e+02 1.701e+02 1.953e+02 3.306e+02, threshold=3.402e+02, percent-clipped=0.0 2024-09-22 18:22:34,715 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=74732.0, ans=0.125 2024-09-22 18:22:54,716 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=74778.66666666667, ans=0.125 2024-09-22 18:23:05,459 INFO [train.py:1198] (2/4) Epoch 5, batch 450, loss[loss=0.2739, ctc_loss=0.1986, cr_loss=0.3765, over 17111.00 frames. ], tot_loss[loss=0.3048, ctc_loss=0.2233, cr_loss=0.4075, over 3021566.46 frames. ], batch size: 43, lr: 2.38e-02, grad_scale: 32.0 2024-09-22 18:23:17,983 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=74825.33333333333, ans=0.1 2024-09-22 18:23:21,146 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=74825.33333333333, ans=0.125 2024-09-22 18:23:21,695 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=3.33 vs. limit=15.0 2024-09-22 18:23:26,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=74872.0, ans=0.1 2024-09-22 18:23:57,634 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=74965.33333333333, ans=0.125 2024-09-22 18:24:08,483 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=74965.33333333333, ans=0.0 2024-09-22 18:24:27,402 INFO [train.py:1198] (2/4) Epoch 5, batch 500, loss[loss=0.2697, ctc_loss=0.1963, cr_loss=0.3669, over 17081.00 frames. ], tot_loss[loss=0.3038, ctc_loss=0.2223, cr_loss=0.4077, over 3102916.10 frames. ], batch size: 40, lr: 2.37e-02, grad_scale: 32.0 2024-09-22 18:24:27,765 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=75058.66666666667, ans=0.125 2024-09-22 18:25:02,111 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=75152.0, ans=0.035 2024-09-22 18:25:03,656 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=75152.0, ans=0.1 2024-09-22 18:25:05,458 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=75152.0, ans=0.125 2024-09-22 18:25:13,993 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.212e+02 1.528e+02 1.760e+02 2.096e+02 4.266e+02, threshold=3.519e+02, percent-clipped=4.0 2024-09-22 18:25:17,834 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=7.47 vs. limit=12.0 2024-09-22 18:25:33,688 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=11.59 vs. limit=15.0 2024-09-22 18:25:47,574 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.whiten.whitening_limit, batch_count=75245.33333333333, ans=15.0 2024-09-22 18:25:51,520 INFO [train.py:1198] (2/4) Epoch 5, batch 550, loss[loss=0.3151, ctc_loss=0.2305, cr_loss=0.4232, over 17085.00 frames. ], tot_loss[loss=0.302, ctc_loss=0.2207, cr_loss=0.4065, over 3170387.72 frames. ], batch size: 49, lr: 2.37e-02, grad_scale: 32.0 2024-09-22 18:26:03,059 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=75292.0, ans=0.1 2024-09-22 18:27:10,587 INFO [train.py:1198] (2/4) Epoch 5, batch 600, loss[loss=0.3045, ctc_loss=0.2223, cr_loss=0.4108, over 17308.00 frames. ], tot_loss[loss=0.3005, ctc_loss=0.2193, cr_loss=0.4058, over 3223059.71 frames. ], batch size: 49, lr: 2.37e-02, grad_scale: 32.0 2024-09-22 18:27:20,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=75525.33333333333, ans=0.07 2024-09-22 18:27:23,568 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=75525.33333333333, ans=0.0 2024-09-22 18:27:40,358 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=6.34 vs. limit=15.0 2024-09-22 18:27:53,216 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=75618.66666666667, ans=0.125 2024-09-22 18:27:57,743 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.285e+02 1.538e+02 1.800e+02 2.212e+02 3.356e+02, threshold=3.600e+02, percent-clipped=0.0 2024-09-22 18:28:35,277 INFO [train.py:1198] (2/4) Epoch 5, batch 650, loss[loss=0.2994, ctc_loss=0.2221, cr_loss=0.3867, over 17145.00 frames. ], tot_loss[loss=0.3019, ctc_loss=0.2205, cr_loss=0.407, over 3256189.10 frames. ], batch size: 48, lr: 2.36e-02, grad_scale: 32.0 2024-09-22 18:28:38,888 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=75758.66666666667, ans=0.125 2024-09-22 18:28:45,574 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=12.12 vs. limit=15.0 2024-09-22 18:28:59,275 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=75805.33333333333, ans=0.125 2024-09-22 18:29:04,125 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=75805.33333333333, ans=0.2 2024-09-22 18:29:10,325 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=75852.0, ans=0.07 2024-09-22 18:29:24,589 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=75898.66666666667, ans=0.09899494936611666 2024-09-22 18:29:32,373 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=75898.66666666667, ans=0.125 2024-09-22 18:29:34,214 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=13.45 vs. limit=15.0 2024-09-22 18:29:54,259 INFO [train.py:1198] (2/4) Epoch 5, batch 700, loss[loss=0.3286, ctc_loss=0.2407, cr_loss=0.4394, over 17217.00 frames. ], tot_loss[loss=0.3011, ctc_loss=0.2197, cr_loss=0.407, over 3290742.02 frames. ], batch size: 50, lr: 2.36e-02, grad_scale: 32.0 2024-09-22 18:30:36,527 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=76085.33333333333, ans=0.2 2024-09-22 18:30:38,521 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=14.70 vs. limit=22.5 2024-09-22 18:30:43,603 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.226e+02 1.521e+02 1.705e+02 2.105e+02 2.881e+02, threshold=3.410e+02, percent-clipped=0.0 2024-09-22 18:31:03,080 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=76178.66666666667, ans=0.125 2024-09-22 18:31:18,062 INFO [train.py:1198] (2/4) Epoch 5, batch 750, loss[loss=0.3155, ctc_loss=0.2328, cr_loss=0.4134, over 17166.00 frames. ], tot_loss[loss=0.3019, ctc_loss=0.2206, cr_loss=0.4065, over 3290345.28 frames. ], batch size: 45, lr: 2.36e-02, grad_scale: 32.0 2024-09-22 18:31:24,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=76225.33333333333, ans=0.0 2024-09-22 18:31:48,357 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=76318.66666666667, ans=0.125 2024-09-22 18:32:03,752 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.58 vs. limit=5.0 2024-09-22 18:32:15,406 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=76365.33333333333, ans=0.1 2024-09-22 18:32:23,348 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=76412.0, ans=0.125 2024-09-22 18:32:37,281 INFO [train.py:1198] (2/4) Epoch 5, batch 800, loss[loss=0.3097, ctc_loss=0.2301, cr_loss=0.3978, over 17009.00 frames. ], tot_loss[loss=0.3031, ctc_loss=0.2214, cr_loss=0.4085, over 3308300.58 frames. ], batch size: 51, lr: 2.36e-02, grad_scale: 32.0 2024-09-22 18:32:57,557 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=76505.33333333333, ans=0.125 2024-09-22 18:33:26,740 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.316e+02 1.675e+02 1.893e+02 2.259e+02 3.192e+02, threshold=3.786e+02, percent-clipped=0.0 2024-09-22 18:33:35,069 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=76598.66666666667, ans=0.125 2024-09-22 18:33:39,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.min_positive, batch_count=76598.66666666667, ans=0.025 2024-09-22 18:33:46,208 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=76645.33333333333, ans=0.125 2024-09-22 18:33:46,854 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.57 vs. limit=12.0 2024-09-22 18:34:01,757 INFO [train.py:1198] (2/4) Epoch 5, batch 850, loss[loss=0.2922, ctc_loss=0.2097, cr_loss=0.4124, over 17226.00 frames. ], tot_loss[loss=0.3036, ctc_loss=0.2219, cr_loss=0.4087, over 3313089.19 frames. ], batch size: 50, lr: 2.35e-02, grad_scale: 32.0 2024-09-22 18:34:05,324 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=76692.0, ans=0.025 2024-09-22 18:34:25,087 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn2.whiten, num_groups=1, num_channels=192, metric=10.04 vs. limit=22.5 2024-09-22 18:34:33,938 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=76785.33333333333, ans=0.0 2024-09-22 18:34:34,541 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=19.00 vs. limit=22.5 2024-09-22 18:35:23,903 INFO [train.py:1198] (2/4) Epoch 5, batch 900, loss[loss=0.2962, ctc_loss=0.2179, cr_loss=0.3914, over 17083.00 frames. ], tot_loss[loss=0.3017, ctc_loss=0.2206, cr_loss=0.4055, over 3322038.07 frames. ], batch size: 43, lr: 2.35e-02, grad_scale: 32.0 2024-09-22 18:35:44,531 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.40 vs. limit=6.0 2024-09-22 18:36:00,138 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=77018.66666666667, ans=0.125 2024-09-22 18:36:03,272 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=77018.66666666667, ans=0.125 2024-09-22 18:36:06,374 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=77018.66666666667, ans=0.125 2024-09-22 18:36:10,876 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.238e+02 1.424e+02 1.574e+02 1.806e+02 2.908e+02, threshold=3.147e+02, percent-clipped=0.0 2024-09-22 18:36:30,979 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.42 vs. limit=12.0 2024-09-22 18:36:38,847 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=8.03 vs. limit=15.0 2024-09-22 18:36:44,633 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=77158.66666666667, ans=0.1 2024-09-22 18:36:46,127 INFO [train.py:1198] (2/4) Epoch 5, batch 950, loss[loss=0.2778, ctc_loss=0.2033, cr_loss=0.3727, over 17101.00 frames. ], tot_loss[loss=0.3008, ctc_loss=0.2197, cr_loss=0.4056, over 3333372.40 frames. ], batch size: 40, lr: 2.35e-02, grad_scale: 32.0 2024-09-22 18:36:48,113 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=77158.66666666667, ans=0.025 2024-09-22 18:36:51,934 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=7.41 vs. limit=8.0 2024-09-22 18:36:55,608 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=77158.66666666667, ans=0.025 2024-09-22 18:36:58,902 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=77158.66666666667, ans=0.125 2024-09-22 18:37:39,808 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=77298.66666666667, ans=0.1 2024-09-22 18:38:10,579 INFO [train.py:1198] (2/4) Epoch 5, batch 1000, loss[loss=0.2672, ctc_loss=0.1909, cr_loss=0.3814, over 16943.00 frames. ], tot_loss[loss=0.302, ctc_loss=0.2207, cr_loss=0.4064, over 3332225.54 frames. ], batch size: 42, lr: 2.34e-02, grad_scale: 32.0 2024-09-22 18:38:18,840 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=77392.0, ans=0.0 2024-09-22 18:38:23,548 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 18:38:25,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=77438.66666666667, ans=0.0 2024-09-22 18:38:40,997 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 18:38:55,016 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.192e+02 1.650e+02 1.785e+02 2.141e+02 3.125e+02, threshold=3.569e+02, percent-clipped=0.0 2024-09-22 18:39:17,996 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=21.91 vs. limit=22.5 2024-09-22 18:39:27,485 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=7.81 vs. limit=15.0 2024-09-22 18:39:30,025 INFO [train.py:1198] (2/4) Epoch 5, batch 1050, loss[loss=0.3383, ctc_loss=0.2518, cr_loss=0.4322, over 17213.00 frames. ], tot_loss[loss=0.3018, ctc_loss=0.2206, cr_loss=0.4059, over 3334588.88 frames. ], batch size: 55, lr: 2.34e-02, grad_scale: 32.0 2024-09-22 18:39:33,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=77625.33333333333, ans=0.025 2024-09-22 18:39:35,604 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.94 vs. limit=15.0 2024-09-22 18:40:11,203 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=77718.66666666667, ans=0.125 2024-09-22 18:40:11,295 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=77718.66666666667, ans=0.0 2024-09-22 18:40:47,106 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=77812.0, ans=0.1 2024-09-22 18:40:48,725 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=77812.0, ans=0.125 2024-09-22 18:40:54,538 INFO [train.py:1198] (2/4) Epoch 5, batch 1100, loss[loss=0.297, ctc_loss=0.213, cr_loss=0.42, over 17004.00 frames. ], tot_loss[loss=0.3013, ctc_loss=0.2202, cr_loss=0.4054, over 3350037.54 frames. ], batch size: 51, lr: 2.34e-02, grad_scale: 16.0 2024-09-22 18:41:10,563 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=77905.33333333333, ans=0.125 2024-09-22 18:41:12,169 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=77905.33333333333, ans=0.2 2024-09-22 18:41:39,248 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=77952.0, ans=0.1 2024-09-22 18:41:40,427 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.256e+02 1.520e+02 1.761e+02 2.071e+02 3.558e+02, threshold=3.523e+02, percent-clipped=0.0 2024-09-22 18:42:01,645 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=78045.33333333333, ans=0.025 2024-09-22 18:42:07,716 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=78045.33333333333, ans=0.1 2024-09-22 18:42:13,886 INFO [train.py:1198] (2/4) Epoch 5, batch 1150, loss[loss=0.2998, ctc_loss=0.2224, cr_loss=0.3873, over 17013.00 frames. ], tot_loss[loss=0.3014, ctc_loss=0.2202, cr_loss=0.4059, over 3353998.37 frames. ], batch size: 51, lr: 2.34e-02, grad_scale: 16.0 2024-09-22 18:43:02,114 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=34.31 vs. limit=15.0 2024-09-22 18:43:37,774 INFO [train.py:1198] (2/4) Epoch 5, batch 1200, loss[loss=0.3072, ctc_loss=0.2245, cr_loss=0.4132, over 15948.00 frames. ], tot_loss[loss=0.3014, ctc_loss=0.2201, cr_loss=0.4063, over 3355911.62 frames. ], batch size: 74, lr: 2.33e-02, grad_scale: 32.0 2024-09-22 18:43:41,231 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=78325.33333333333, ans=0.0 2024-09-22 18:43:57,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.out_whiten.whitening_limit, batch_count=78372.0, ans=15.0 2024-09-22 18:44:00,432 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=78372.0, ans=0.125 2024-09-22 18:44:08,574 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=78418.66666666667, ans=0.125 2024-09-22 18:44:24,262 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.229e+02 1.563e+02 1.694e+02 1.943e+02 2.938e+02, threshold=3.387e+02, percent-clipped=0.0 2024-09-22 18:44:39,068 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.min_positive, batch_count=78465.33333333333, ans=0.05 2024-09-22 18:44:40,437 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=78512.0, ans=0.125 2024-09-22 18:44:48,786 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.42 vs. limit=12.0 2024-09-22 18:44:57,612 INFO [train.py:1198] (2/4) Epoch 5, batch 1250, loss[loss=0.3911, ctc_loss=0.3091, cr_loss=0.4096, over 11580.00 frames. ], tot_loss[loss=0.3013, ctc_loss=0.22, cr_loss=0.4066, over 3363009.98 frames. ], batch size: 123, lr: 2.33e-02, grad_scale: 32.0 2024-09-22 18:45:19,555 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=78605.33333333333, ans=0.0 2024-09-22 18:45:50,490 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=78698.66666666667, ans=0.125 2024-09-22 18:45:56,797 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=78698.66666666667, ans=0.0 2024-09-22 18:46:22,171 INFO [train.py:1198] (2/4) Epoch 5, batch 1300, loss[loss=0.3111, ctc_loss=0.225, cr_loss=0.4305, over 17047.00 frames. ], tot_loss[loss=0.3023, ctc_loss=0.2208, cr_loss=0.4078, over 3350985.21 frames. ], batch size: 52, lr: 2.33e-02, grad_scale: 32.0 2024-09-22 18:46:28,715 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=78792.0, ans=0.015 2024-09-22 18:46:28,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=78792.0, ans=0.2 2024-09-22 18:46:53,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=78885.33333333333, ans=0.0 2024-09-22 18:47:00,632 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=78885.33333333333, ans=0.0 2024-09-22 18:47:08,164 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.258e+02 1.495e+02 1.790e+02 2.182e+02 4.439e+02, threshold=3.579e+02, percent-clipped=1.0 2024-09-22 18:47:14,671 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=78932.0, ans=0.2 2024-09-22 18:47:21,523 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=3.80 vs. limit=15.0 2024-09-22 18:47:43,980 INFO [train.py:1198] (2/4) Epoch 5, batch 1350, loss[loss=0.3299, ctc_loss=0.24, cr_loss=0.4494, over 17030.00 frames. ], tot_loss[loss=0.304, ctc_loss=0.222, cr_loss=0.41, over 3356524.41 frames. ], batch size: 56, lr: 2.32e-02, grad_scale: 32.0 2024-09-22 18:47:57,923 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=79025.33333333333, ans=0.1 2024-09-22 18:48:01,175 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=79072.0, ans=0.125 2024-09-22 18:48:13,862 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=20.42 vs. limit=22.5 2024-09-22 18:48:16,527 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=79118.66666666667, ans=0.125 2024-09-22 18:48:39,061 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=79165.33333333333, ans=0.125 2024-09-22 18:48:59,059 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.42 vs. limit=15.0 2024-09-22 18:49:05,830 INFO [train.py:1198] (2/4) Epoch 5, batch 1400, loss[loss=0.3758, ctc_loss=0.2946, cr_loss=0.4058, over 11714.00 frames. ], tot_loss[loss=0.3032, ctc_loss=0.2215, cr_loss=0.4084, over 3343111.86 frames. ], batch size: 124, lr: 2.32e-02, grad_scale: 32.0 2024-09-22 18:49:31,634 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=79305.33333333333, ans=0.2 2024-09-22 18:49:54,449 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.244e+02 1.617e+02 1.825e+02 2.247e+02 3.972e+02, threshold=3.649e+02, percent-clipped=1.0 2024-09-22 18:49:59,429 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=79398.66666666667, ans=0.0 2024-09-22 18:50:02,697 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=79398.66666666667, ans=0.125 2024-09-22 18:50:05,829 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=79398.66666666667, ans=0.125 2024-09-22 18:50:12,258 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=79445.33333333333, ans=0.125 2024-09-22 18:50:30,040 INFO [train.py:1198] (2/4) Epoch 5, batch 1450, loss[loss=0.3165, ctc_loss=0.23, cr_loss=0.4325, over 17307.00 frames. ], tot_loss[loss=0.3047, ctc_loss=0.2228, cr_loss=0.4098, over 3336744.95 frames. ], batch size: 49, lr: 2.32e-02, grad_scale: 32.0 2024-09-22 18:50:36,738 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 18:50:53,312 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=22.74 vs. limit=22.5 2024-09-22 18:50:56,002 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 18:50:59,175 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=79538.66666666667, ans=0.0 2024-09-22 18:51:43,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=79678.66666666667, ans=0.125 2024-09-22 18:51:44,274 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=2.93 vs. limit=15.0 2024-09-22 18:51:49,951 INFO [train.py:1198] (2/4) Epoch 5, batch 1500, loss[loss=0.3121, ctc_loss=0.2299, cr_loss=0.4113, over 17108.00 frames. ], tot_loss[loss=0.303, ctc_loss=0.2213, cr_loss=0.4085, over 3343320.03 frames. ], batch size: 49, lr: 2.32e-02, grad_scale: 32.0 2024-09-22 18:51:53,406 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=79725.33333333333, ans=0.125 2024-09-22 18:51:55,375 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.38 vs. limit=10.0 2024-09-22 18:51:56,461 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.min_positive, batch_count=79725.33333333333, ans=0.025 2024-09-22 18:52:12,269 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 18:52:22,773 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=79818.66666666667, ans=0.0 2024-09-22 18:52:24,593 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.69 vs. limit=15.0 2024-09-22 18:52:26,400 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=21.51 vs. limit=22.5 2024-09-22 18:52:40,768 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.149e+02 1.573e+02 1.905e+02 2.453e+02 3.573e+02, threshold=3.810e+02, percent-clipped=0.0 2024-09-22 18:53:09,543 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=79912.0, ans=0.07 2024-09-22 18:53:09,999 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=13.97 vs. limit=15.0 2024-09-22 18:53:14,048 INFO [train.py:1198] (2/4) Epoch 5, batch 1550, loss[loss=0.257, ctc_loss=0.184, cr_loss=0.365, over 17091.00 frames. ], tot_loss[loss=0.3023, ctc_loss=0.2207, cr_loss=0.4081, over 3350235.68 frames. ], batch size: 40, lr: 2.31e-02, grad_scale: 32.0 2024-09-22 18:53:14,902 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=21.65 vs. limit=22.5 2024-09-22 18:53:19,714 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.61 vs. limit=6.0 2024-09-22 18:53:44,713 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=80052.0, ans=0.0 2024-09-22 18:53:45,315 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.93 vs. limit=10.0 2024-09-22 18:53:52,541 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=80052.0, ans=0.125 2024-09-22 18:54:08,879 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=7.19 vs. limit=12.0 2024-09-22 18:54:17,622 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=7.42 vs. limit=8.0 2024-09-22 18:54:30,845 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=80145.33333333333, ans=0.2 2024-09-22 18:54:33,898 INFO [train.py:1198] (2/4) Epoch 5, batch 1600, loss[loss=0.3129, ctc_loss=0.2285, cr_loss=0.4222, over 17219.00 frames. ], tot_loss[loss=0.302, ctc_loss=0.2204, cr_loss=0.4078, over 3354512.40 frames. ], batch size: 50, lr: 2.31e-02, grad_scale: 32.0 2024-09-22 18:54:46,546 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=21.67 vs. limit=22.5 2024-09-22 18:54:49,108 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=80192.0, ans=0.0 2024-09-22 18:55:08,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=80285.33333333333, ans=0.1 2024-09-22 18:55:22,231 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.274e+02 1.455e+02 1.591e+02 1.855e+02 3.051e+02, threshold=3.183e+02, percent-clipped=0.0 2024-09-22 18:55:28,597 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=20.16 vs. limit=22.5 2024-09-22 18:55:34,448 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=80332.0, ans=0.0 2024-09-22 18:55:58,301 INFO [train.py:1198] (2/4) Epoch 5, batch 1650, loss[loss=0.3006, ctc_loss=0.2221, cr_loss=0.3922, over 17088.00 frames. ], tot_loss[loss=0.3018, ctc_loss=0.2202, cr_loss=0.4081, over 3357446.56 frames. ], batch size: 43, lr: 2.31e-02, grad_scale: 32.0 2024-09-22 18:56:00,206 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=80425.33333333333, ans=0.07 2024-09-22 18:56:14,427 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=80472.0, ans=0.2 2024-09-22 18:56:18,284 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=20.94 vs. limit=22.5 2024-09-22 18:56:18,331 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=3.59 vs. limit=15.0 2024-09-22 18:56:20,764 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=80472.0, ans=0.0 2024-09-22 18:56:35,058 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=80518.66666666667, ans=0.125 2024-09-22 18:57:01,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.min_positive, batch_count=80612.0, ans=0.025 2024-09-22 18:57:01,919 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=80612.0, ans=0.0 2024-09-22 18:57:05,172 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=80612.0, ans=0.0 2024-09-22 18:57:07,410 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=17.09 vs. limit=22.5 2024-09-22 18:57:19,846 INFO [train.py:1198] (2/4) Epoch 5, batch 1700, loss[loss=0.3489, ctc_loss=0.2622, cr_loss=0.4337, over 15091.00 frames. ], tot_loss[loss=0.3012, ctc_loss=0.2196, cr_loss=0.4076, over 3356619.46 frames. ], batch size: 88, lr: 2.30e-02, grad_scale: 32.0 2024-09-22 18:57:29,638 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=80658.66666666667, ans=0.2 2024-09-22 18:57:31,243 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.max_positive, batch_count=80658.66666666667, ans=0.95 2024-09-22 18:57:31,306 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=80658.66666666667, ans=0.125 2024-09-22 18:57:44,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=80705.33333333333, ans=0.1 2024-09-22 18:57:49,624 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=80705.33333333333, ans=0.2 2024-09-22 18:58:08,258 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.290e+02 1.556e+02 1.832e+02 2.128e+02 4.427e+02, threshold=3.664e+02, percent-clipped=3.0 2024-09-22 18:58:14,848 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.min_positive, batch_count=80798.66666666667, ans=0.025 2024-09-22 18:58:35,922 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=80845.33333333333, ans=0.125 2024-09-22 18:58:41,926 INFO [train.py:1198] (2/4) Epoch 5, batch 1750, loss[loss=0.3204, ctc_loss=0.2366, cr_loss=0.419, over 17041.00 frames. ], tot_loss[loss=0.3007, ctc_loss=0.2192, cr_loss=0.4077, over 3366390.99 frames. ], batch size: 52, lr: 2.30e-02, grad_scale: 32.0 2024-09-22 18:58:42,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=80892.0, ans=0.125 2024-09-22 18:58:43,740 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=80892.0, ans=0.125 2024-09-22 18:59:10,757 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=80938.66666666667, ans=0.125 2024-09-22 18:59:23,491 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=80985.33333333333, ans=0.125 2024-09-22 19:00:03,798 INFO [train.py:1198] (2/4) Epoch 5, batch 1800, loss[loss=0.327, ctc_loss=0.2393, cr_loss=0.4387, over 17105.00 frames. ], tot_loss[loss=0.3009, ctc_loss=0.2192, cr_loss=0.4082, over 3371781.18 frames. ], batch size: 49, lr: 2.30e-02, grad_scale: 32.0 2024-09-22 19:00:09,713 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=14.80 vs. limit=22.5 2024-09-22 19:00:15,088 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=81125.33333333333, ans=0.0 2024-09-22 19:00:21,409 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=81172.0, ans=0.125 2024-09-22 19:00:49,506 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=81218.66666666667, ans=0.1 2024-09-22 19:00:52,230 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.234e+02 1.571e+02 1.779e+02 2.119e+02 4.116e+02, threshold=3.559e+02, percent-clipped=1.0 2024-09-22 19:01:00,641 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=81265.33333333333, ans=0.125 2024-09-22 19:01:03,879 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=81265.33333333333, ans=0.1 2024-09-22 19:01:04,534 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.86 vs. limit=15.0 2024-09-22 19:01:13,048 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=81312.0, ans=0.125 2024-09-22 19:01:13,384 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=13.05 vs. limit=22.5 2024-09-22 19:01:17,858 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=81312.0, ans=0.2 2024-09-22 19:01:17,965 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=81312.0, ans=0.95 2024-09-22 19:01:25,611 INFO [train.py:1198] (2/4) Epoch 5, batch 1850, loss[loss=0.2847, ctc_loss=0.2066, cr_loss=0.3904, over 17098.00 frames. ], tot_loss[loss=0.3015, ctc_loss=0.2197, cr_loss=0.4089, over 3369147.90 frames. ], batch size: 40, lr: 2.30e-02, grad_scale: 32.0 2024-09-22 19:01:27,792 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.01 vs. limit=15.0 2024-09-22 19:01:29,744 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=19.38 vs. limit=22.5 2024-09-22 19:01:54,749 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=81405.33333333333, ans=0.1 2024-09-22 19:01:54,795 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=81405.33333333333, ans=0.125 2024-09-22 19:01:56,429 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=81452.0, ans=0.125 2024-09-22 19:02:08,554 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module1.whiten, num_groups=1, num_channels=192, metric=7.38 vs. limit=15.0 2024-09-22 19:02:10,731 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=81452.0, ans=0.0 2024-09-22 19:02:32,120 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=81545.33333333333, ans=0.2 2024-09-22 19:02:50,453 INFO [train.py:1198] (2/4) Epoch 5, batch 1900, loss[loss=0.2393, ctc_loss=0.1692, cr_loss=0.3505, over 17166.00 frames. ], tot_loss[loss=0.2999, ctc_loss=0.2186, cr_loss=0.4067, over 3367487.68 frames. ], batch size: 41, lr: 2.29e-02, grad_scale: 32.0 2024-09-22 19:03:29,256 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.13 vs. limit=10.0 2024-09-22 19:03:35,164 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=81685.33333333333, ans=0.2 2024-09-22 19:03:36,543 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.294e+02 1.506e+02 1.844e+02 2.270e+02 3.771e+02, threshold=3.688e+02, percent-clipped=2.0 2024-09-22 19:03:52,914 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=81778.66666666667, ans=0.125 2024-09-22 19:04:07,784 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=23.24 vs. limit=22.5 2024-09-22 19:04:10,002 INFO [train.py:1198] (2/4) Epoch 5, batch 1950, loss[loss=0.2472, ctc_loss=0.1778, cr_loss=0.3469, over 17190.00 frames. ], tot_loss[loss=0.2997, ctc_loss=0.2183, cr_loss=0.4072, over 3371290.94 frames. ], batch size: 41, lr: 2.29e-02, grad_scale: 32.0 2024-09-22 19:04:20,574 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.11 vs. limit=15.0 2024-09-22 19:04:34,610 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=81872.0, ans=0.2 2024-09-22 19:04:57,619 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=81918.66666666667, ans=0.0 2024-09-22 19:05:00,811 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=81965.33333333333, ans=0.0 2024-09-22 19:05:18,277 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=82012.0, ans=0.125 2024-09-22 19:05:28,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=82012.0, ans=0.015 2024-09-22 19:05:34,769 INFO [train.py:1198] (2/4) Epoch 5, batch 2000, loss[loss=0.2899, ctc_loss=0.2075, cr_loss=0.4118, over 17166.00 frames. ], tot_loss[loss=0.2996, ctc_loss=0.2181, cr_loss=0.4075, over 3372423.06 frames. ], batch size: 45, lr: 2.29e-02, grad_scale: 32.0 2024-09-22 19:05:35,943 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.91 vs. limit=15.0 2024-09-22 19:06:02,524 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=82105.33333333333, ans=0.125 2024-09-22 19:06:07,351 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=82152.0, ans=0.2 2024-09-22 19:06:18,421 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=82152.0, ans=0.125 2024-09-22 19:06:21,337 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.273e+02 1.502e+02 1.753e+02 2.170e+02 3.000e+02, threshold=3.507e+02, percent-clipped=0.0 2024-09-22 19:06:21,639 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=82198.66666666667, ans=0.125 2024-09-22 19:06:39,046 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=82245.33333333333, ans=0.2 2024-09-22 19:06:54,492 INFO [train.py:1198] (2/4) Epoch 5, batch 2050, loss[loss=0.2562, ctc_loss=0.1854, cr_loss=0.354, over 17265.00 frames. ], tot_loss[loss=0.3001, ctc_loss=0.2185, cr_loss=0.4078, over 3372721.38 frames. ], batch size: 44, lr: 2.28e-02, grad_scale: 32.0 2024-09-22 19:07:01,672 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.40 vs. limit=15.0 2024-09-22 19:07:13,160 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=82338.66666666667, ans=0.125 2024-09-22 19:07:16,256 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=82338.66666666667, ans=0.0 2024-09-22 19:07:22,615 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=82338.66666666667, ans=0.125 2024-09-22 19:07:41,206 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=82385.33333333333, ans=0.125 2024-09-22 19:07:42,757 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 19:07:47,522 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=82432.0, ans=0.0 2024-09-22 19:07:48,986 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=82432.0, ans=0.0 2024-09-22 19:07:52,694 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.40 vs. limit=15.0 2024-09-22 19:08:18,626 INFO [train.py:1198] (2/4) Epoch 5, batch 2100, loss[loss=0.2724, ctc_loss=0.2006, cr_loss=0.3591, over 17299.00 frames. ], tot_loss[loss=0.2991, ctc_loss=0.2178, cr_loss=0.4063, over 3373059.49 frames. ], batch size: 46, lr: 2.28e-02, grad_scale: 32.0 2024-09-22 19:08:56,893 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=82618.66666666667, ans=0.025 2024-09-22 19:09:04,410 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.206e+02 1.583e+02 1.843e+02 2.179e+02 3.617e+02, threshold=3.686e+02, percent-clipped=2.0 2024-09-22 19:09:11,184 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=82665.33333333333, ans=0.125 2024-09-22 19:09:12,666 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=82665.33333333333, ans=0.0 2024-09-22 19:09:20,598 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=82712.0, ans=0.1 2024-09-22 19:09:28,498 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=82712.0, ans=0.125 2024-09-22 19:09:39,788 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=9.69 vs. limit=15.0 2024-09-22 19:09:40,263 INFO [train.py:1198] (2/4) Epoch 5, batch 2150, loss[loss=0.3202, ctc_loss=0.2334, cr_loss=0.4339, over 17096.00 frames. ], tot_loss[loss=0.2988, ctc_loss=0.2177, cr_loss=0.4057, over 3368258.66 frames. ], batch size: 49, lr: 2.28e-02, grad_scale: 32.0 2024-09-22 19:09:46,773 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=82758.66666666667, ans=0.0 2024-09-22 19:09:53,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=82758.66666666667, ans=0.07 2024-09-22 19:09:59,592 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=82805.33333333333, ans=0.0 2024-09-22 19:10:01,965 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=8.29 vs. limit=15.0 2024-09-22 19:10:04,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=82805.33333333333, ans=0.125 2024-09-22 19:10:09,551 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=8.18 vs. limit=15.0 2024-09-22 19:10:10,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=82852.0, ans=0.1 2024-09-22 19:10:46,992 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten.whitening_limit, batch_count=82945.33333333333, ans=15.0 2024-09-22 19:10:57,694 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=82945.33333333333, ans=0.125 2024-09-22 19:11:02,149 INFO [train.py:1198] (2/4) Epoch 5, batch 2200, loss[loss=0.3055, ctc_loss=0.2267, cr_loss=0.3944, over 17047.00 frames. ], tot_loss[loss=0.2986, ctc_loss=0.2175, cr_loss=0.4055, over 3362099.08 frames. ], batch size: 52, lr: 2.28e-02, grad_scale: 32.0 2024-09-22 19:11:11,806 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=82992.0, ans=0.025 2024-09-22 19:11:26,431 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=83038.66666666667, ans=0.1 2024-09-22 19:11:41,067 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=83085.33333333333, ans=0.125 2024-09-22 19:11:48,591 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.270e+02 1.603e+02 1.776e+02 2.386e+02 3.569e+02, threshold=3.552e+02, percent-clipped=0.0 2024-09-22 19:11:49,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten.whitening_limit, batch_count=83132.0, ans=15.0 2024-09-22 19:12:15,362 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=83178.66666666667, ans=0.025 2024-09-22 19:12:24,734 INFO [train.py:1198] (2/4) Epoch 5, batch 2250, loss[loss=0.307, ctc_loss=0.223, cr_loss=0.42, over 17028.00 frames. ], tot_loss[loss=0.2978, ctc_loss=0.2169, cr_loss=0.4045, over 3370967.55 frames. ], batch size: 56, lr: 2.27e-02, grad_scale: 32.0 2024-09-22 19:12:29,788 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=83225.33333333333, ans=0.0 2024-09-22 19:12:30,099 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=11.90 vs. limit=15.0 2024-09-22 19:12:42,138 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=83272.0, ans=0.1 2024-09-22 19:12:53,740 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=13.39 vs. limit=22.5 2024-09-22 19:13:00,621 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=83318.66666666667, ans=0.1 2024-09-22 19:13:05,639 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=83318.66666666667, ans=0.0 2024-09-22 19:13:08,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=83318.66666666667, ans=0.1 2024-09-22 19:13:16,143 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.60 vs. limit=5.0 2024-09-22 19:13:18,409 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=83365.33333333333, ans=0.1 2024-09-22 19:13:19,825 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.min_abs, batch_count=83365.33333333333, ans=0.5 2024-09-22 19:13:27,759 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=83365.33333333333, ans=0.125 2024-09-22 19:13:38,195 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.13 vs. limit=15.0 2024-09-22 19:13:46,566 INFO [train.py:1198] (2/4) Epoch 5, batch 2300, loss[loss=0.299, ctc_loss=0.2202, cr_loss=0.3939, over 16942.00 frames. ], tot_loss[loss=0.2974, ctc_loss=0.2165, cr_loss=0.4043, over 3374130.27 frames. ], batch size: 58, lr: 2.27e-02, grad_scale: 32.0 2024-09-22 19:14:27,782 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=83552.0, ans=0.125 2024-09-22 19:14:34,650 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.240e+02 1.515e+02 1.751e+02 2.046e+02 3.052e+02, threshold=3.503e+02, percent-clipped=0.0 2024-09-22 19:14:38,196 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=83598.66666666667, ans=0.125 2024-09-22 19:15:07,863 INFO [train.py:1198] (2/4) Epoch 5, batch 2350, loss[loss=0.317, ctc_loss=0.2346, cr_loss=0.4123, over 17030.00 frames. ], tot_loss[loss=0.2973, ctc_loss=0.2163, cr_loss=0.4046, over 3372127.75 frames. ], batch size: 56, lr: 2.27e-02, grad_scale: 32.0 2024-09-22 19:16:26,403 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=5.46 vs. limit=10.0 2024-09-22 19:16:27,587 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=83878.66666666667, ans=0.0 2024-09-22 19:16:30,648 INFO [train.py:1198] (2/4) Epoch 5, batch 2400, loss[loss=0.2902, ctc_loss=0.2072, cr_loss=0.4147, over 17298.00 frames. ], tot_loss[loss=0.2987, ctc_loss=0.2175, cr_loss=0.4058, over 3359032.15 frames. ], batch size: 49, lr: 2.27e-02, grad_scale: 32.0 2024-09-22 19:16:48,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=83972.0, ans=0.125 2024-09-22 19:16:53,715 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.51 vs. limit=12.0 2024-09-22 19:17:19,651 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.205e+02 1.495e+02 1.658e+02 1.967e+02 2.763e+02, threshold=3.315e+02, percent-clipped=0.0 2024-09-22 19:17:20,066 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=84065.33333333333, ans=0.2 2024-09-22 19:17:50,965 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=84112.0, ans=0.125 2024-09-22 19:17:55,425 INFO [train.py:1198] (2/4) Epoch 5, batch 2450, loss[loss=0.3199, ctc_loss=0.2311, cr_loss=0.4439, over 16986.00 frames. ], tot_loss[loss=0.3005, ctc_loss=0.219, cr_loss=0.4073, over 3351233.95 frames. ], batch size: 53, lr: 2.26e-02, grad_scale: 32.0 2024-09-22 19:18:14,260 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=9.36 vs. limit=10.0 2024-09-22 19:18:17,074 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.23 vs. limit=10.0 2024-09-22 19:18:46,788 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=84298.66666666667, ans=0.125 2024-09-22 19:19:09,099 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=84345.33333333333, ans=0.125 2024-09-22 19:19:15,415 INFO [train.py:1198] (2/4) Epoch 5, batch 2500, loss[loss=0.2552, ctc_loss=0.1815, cr_loss=0.3687, over 17037.00 frames. ], tot_loss[loss=0.3006, ctc_loss=0.2192, cr_loss=0.4072, over 3350786.43 frames. ], batch size: 39, lr: 2.26e-02, grad_scale: 32.0 2024-09-22 19:19:15,818 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=84392.0, ans=0.125 2024-09-22 19:19:34,374 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=84438.66666666667, ans=0.125 2024-09-22 19:20:04,526 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.268e+02 1.488e+02 1.690e+02 1.912e+02 3.424e+02, threshold=3.381e+02, percent-clipped=1.0 2024-09-22 19:20:21,085 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.42 vs. limit=6.0 2024-09-22 19:20:40,872 INFO [train.py:1198] (2/4) Epoch 5, batch 2550, loss[loss=0.2545, ctc_loss=0.1843, cr_loss=0.3509, over 17126.00 frames. ], tot_loss[loss=0.3009, ctc_loss=0.2194, cr_loss=0.4077, over 3350772.58 frames. ], batch size: 40, lr: 2.26e-02, grad_scale: 32.0 2024-09-22 19:20:52,308 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=84625.33333333333, ans=0.0 2024-09-22 19:21:25,547 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=84718.66666666667, ans=0.1 2024-09-22 19:22:03,009 INFO [train.py:1198] (2/4) Epoch 5, batch 2600, loss[loss=0.321, ctc_loss=0.2352, cr_loss=0.4293, over 16538.00 frames. ], tot_loss[loss=0.3006, ctc_loss=0.219, cr_loss=0.4081, over 3356145.89 frames. ], batch size: 66, lr: 2.25e-02, grad_scale: 32.0 2024-09-22 19:22:09,610 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 19:22:16,381 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.26 vs. limit=10.0 2024-09-22 19:22:24,329 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.whiten, num_groups=1, num_channels=512, metric=4.88 vs. limit=12.0 2024-09-22 19:22:31,537 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=84905.33333333333, ans=0.1 2024-09-22 19:22:51,550 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.221e+02 1.583e+02 1.835e+02 2.132e+02 3.981e+02, threshold=3.669e+02, percent-clipped=1.0 2024-09-22 19:22:56,610 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=84998.66666666667, ans=0.0 2024-09-22 19:23:24,863 INFO [train.py:1198] (2/4) Epoch 5, batch 2650, loss[loss=0.2613, ctc_loss=0.1847, cr_loss=0.3832, over 17006.00 frames. ], tot_loss[loss=0.2994, ctc_loss=0.218, cr_loss=0.4068, over 3343330.61 frames. ], batch size: 39, lr: 2.25e-02, grad_scale: 32.0 2024-09-22 19:23:36,397 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=85092.0, ans=0.025 2024-09-22 19:23:53,689 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=85138.66666666667, ans=0.125 2024-09-22 19:24:18,046 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=85232.0, ans=0.05 2024-09-22 19:24:31,570 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=5.55 vs. limit=10.0 2024-09-22 19:24:41,735 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=85278.66666666667, ans=0.025 2024-09-22 19:24:44,862 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=85325.33333333333, ans=0.125 2024-09-22 19:24:46,094 INFO [train.py:1198] (2/4) Epoch 5, batch 2700, loss[loss=0.2498, ctc_loss=0.1799, cr_loss=0.3495, over 17275.00 frames. ], tot_loss[loss=0.2993, ctc_loss=0.218, cr_loss=0.4065, over 3340869.24 frames. ], batch size: 42, lr: 2.25e-02, grad_scale: 32.0 2024-09-22 19:24:46,439 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=85325.33333333333, ans=0.1 2024-09-22 19:24:48,091 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=85325.33333333333, ans=0.025 2024-09-22 19:25:02,591 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=4.90 vs. limit=12.0 2024-09-22 19:25:30,951 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.01 vs. limit=15.0 2024-09-22 19:25:34,609 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.200e+02 1.500e+02 1.639e+02 1.809e+02 2.394e+02, threshold=3.278e+02, percent-clipped=0.0 2024-09-22 19:25:35,564 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=23.92 vs. limit=22.5 2024-09-22 19:25:57,140 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=85512.0, ans=0.125 2024-09-22 19:26:08,136 INFO [train.py:1198] (2/4) Epoch 5, batch 2750, loss[loss=0.3254, ctc_loss=0.239, cr_loss=0.4321, over 17039.00 frames. ], tot_loss[loss=0.2989, ctc_loss=0.2176, cr_loss=0.4064, over 3348569.89 frames. ], batch size: 56, lr: 2.25e-02, grad_scale: 32.0 2024-09-22 19:26:27,242 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=85605.33333333333, ans=0.1 2024-09-22 19:26:55,027 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=85652.0, ans=0.04949747468305833 2024-09-22 19:27:10,654 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=85698.66666666667, ans=0.125 2024-09-22 19:27:24,182 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=85745.33333333333, ans=0.125 2024-09-22 19:27:25,890 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=85745.33333333333, ans=0.1 2024-09-22 19:27:31,978 INFO [train.py:1198] (2/4) Epoch 5, batch 2800, loss[loss=0.3121, ctc_loss=0.2267, cr_loss=0.427, over 16442.00 frames. ], tot_loss[loss=0.2992, ctc_loss=0.2179, cr_loss=0.4065, over 3342926.62 frames. ], batch size: 66, lr: 2.24e-02, grad_scale: 32.0 2024-09-22 19:27:41,944 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=85792.0, ans=0.0 2024-09-22 19:28:18,240 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.220e+02 1.484e+02 1.665e+02 1.912e+02 3.153e+02, threshold=3.329e+02, percent-clipped=0.0 2024-09-22 19:28:29,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=85932.0, ans=0.125 2024-09-22 19:28:32,866 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=85932.0, ans=0.1 2024-09-22 19:28:43,911 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=85978.66666666667, ans=0.0 2024-09-22 19:28:47,261 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=85978.66666666667, ans=0.1 2024-09-22 19:28:51,698 INFO [train.py:1198] (2/4) Epoch 5, batch 2850, loss[loss=0.3107, ctc_loss=0.2293, cr_loss=0.4071, over 16741.00 frames. ], tot_loss[loss=0.2992, ctc_loss=0.2179, cr_loss=0.4069, over 3346654.65 frames. ], batch size: 61, lr: 2.24e-02, grad_scale: 32.0 2024-09-22 19:29:29,495 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=86118.66666666667, ans=0.125 2024-09-22 19:29:42,116 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=86165.33333333333, ans=0.0 2024-09-22 19:29:45,376 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=86165.33333333333, ans=0.04949747468305833 2024-09-22 19:30:10,545 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=512, metric=2.59 vs. limit=15.0 2024-09-22 19:30:16,018 INFO [train.py:1198] (2/4) Epoch 5, batch 2900, loss[loss=0.3149, ctc_loss=0.2273, cr_loss=0.4381, over 17198.00 frames. ], tot_loss[loss=0.2997, ctc_loss=0.2182, cr_loss=0.4074, over 3341000.46 frames. ], batch size: 50, lr: 2.24e-02, grad_scale: 32.0 2024-09-22 19:30:22,841 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=86258.66666666667, ans=0.125 2024-09-22 19:30:24,309 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=86258.66666666667, ans=0.125 2024-09-22 19:30:29,107 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=86258.66666666667, ans=0.125 2024-09-22 19:30:32,765 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=14.28 vs. limit=15.0 2024-09-22 19:30:36,133 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2.whitening_limit, batch_count=86305.33333333333, ans=15.0 2024-09-22 19:31:01,933 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.291e+02 1.588e+02 1.741e+02 2.202e+02 4.410e+02, threshold=3.483e+02, percent-clipped=1.0 2024-09-22 19:31:03,794 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=86398.66666666667, ans=0.0 2024-09-22 19:31:25,698 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=86445.33333333333, ans=0.2 2024-09-22 19:31:27,268 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=86445.33333333333, ans=0.2 2024-09-22 19:31:35,099 INFO [train.py:1198] (2/4) Epoch 5, batch 2950, loss[loss=0.249, ctc_loss=0.1767, cr_loss=0.3618, over 17036.00 frames. ], tot_loss[loss=0.2985, ctc_loss=0.2173, cr_loss=0.4063, over 3339102.82 frames. ], batch size: 39, lr: 2.24e-02, grad_scale: 32.0 2024-09-22 19:31:52,355 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=86538.66666666667, ans=0.125 2024-09-22 19:32:14,527 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.38 vs. limit=15.0 2024-09-22 19:32:26,606 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=86632.0, ans=0.1 2024-09-22 19:32:29,576 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=86632.0, ans=0.2 2024-09-22 19:32:42,296 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-22 19:32:59,035 INFO [train.py:1198] (2/4) Epoch 5, batch 3000, loss[loss=0.3297, ctc_loss=0.2442, cr_loss=0.4275, over 16561.00 frames. ], tot_loss[loss=0.299, ctc_loss=0.2176, cr_loss=0.4068, over 3341287.68 frames. ], batch size: 66, lr: 2.23e-02, grad_scale: 32.0 2024-09-22 19:32:59,036 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-22 19:33:14,591 INFO [train.py:1230] (2/4) Epoch 5, validation: loss=0.06642, ctc_loss=0.06642, cr_loss=7.381e-15, over 944034.00 frames. 2024-09-22 19:33:14,592 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-22 19:33:22,582 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=86725.33333333333, ans=0.2 2024-09-22 19:33:39,871 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=86772.0, ans=0.05 2024-09-22 19:33:46,263 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=86818.66666666667, ans=0.0 2024-09-22 19:34:00,034 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.256e+02 1.524e+02 1.804e+02 2.236e+02 6.139e+02, threshold=3.607e+02, percent-clipped=4.0 2024-09-22 19:34:11,382 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=86865.33333333333, ans=0.125 2024-09-22 19:34:20,677 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=86912.0, ans=0.2 2024-09-22 19:34:35,391 INFO [train.py:1198] (2/4) Epoch 5, batch 3050, loss[loss=0.3101, ctc_loss=0.227, cr_loss=0.4157, over 17223.00 frames. ], tot_loss[loss=0.3003, ctc_loss=0.2187, cr_loss=0.408, over 3349970.64 frames. ], batch size: 55, lr: 2.23e-02, grad_scale: 32.0 2024-09-22 19:34:43,443 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=86958.66666666667, ans=0.125 2024-09-22 19:35:08,101 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=87052.0, ans=0.0 2024-09-22 19:35:18,911 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=87052.0, ans=0.125 2024-09-22 19:35:28,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=87098.66666666667, ans=0.125 2024-09-22 19:35:29,949 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=87098.66666666667, ans=0.125 2024-09-22 19:35:48,636 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=87145.33333333333, ans=0.125 2024-09-22 19:35:53,027 INFO [train.py:1198] (2/4) Epoch 5, batch 3100, loss[loss=0.3212, ctc_loss=0.2371, cr_loss=0.4205, over 17132.00 frames. ], tot_loss[loss=0.2996, ctc_loss=0.218, cr_loss=0.4082, over 3361681.23 frames. ], batch size: 48, lr: 2.23e-02, grad_scale: 64.0 2024-09-22 19:36:37,948 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.213e+02 1.555e+02 1.774e+02 2.101e+02 3.567e+02, threshold=3.549e+02, percent-clipped=0.0 2024-09-22 19:36:47,539 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=87332.0, ans=0.0 2024-09-22 19:36:52,901 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=87332.0, ans=0.0 2024-09-22 19:37:13,091 INFO [train.py:1198] (2/4) Epoch 5, batch 3150, loss[loss=0.3169, ctc_loss=0.2325, cr_loss=0.4219, over 17223.00 frames. ], tot_loss[loss=0.2991, ctc_loss=0.2176, cr_loss=0.4076, over 3370076.15 frames. ], batch size: 47, lr: 2.23e-02, grad_scale: 64.0 2024-09-22 19:37:24,353 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=87425.33333333333, ans=0.0 2024-09-22 19:37:38,637 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=87472.0, ans=0.0 2024-09-22 19:38:14,427 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=87612.0, ans=0.0 2024-09-22 19:38:31,402 INFO [train.py:1198] (2/4) Epoch 5, batch 3200, loss[loss=0.2424, ctc_loss=0.1725, cr_loss=0.3498, over 17198.00 frames. ], tot_loss[loss=0.2997, ctc_loss=0.2181, cr_loss=0.4079, over 3366480.54 frames. ], batch size: 41, lr: 2.22e-02, grad_scale: 32.0 2024-09-22 19:38:42,665 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=87658.66666666667, ans=0.2 2024-09-22 19:38:59,760 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=87705.33333333333, ans=0.125 2024-09-22 19:39:14,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=87752.0, ans=0.0 2024-09-22 19:39:15,967 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.56 vs. limit=15.0 2024-09-22 19:39:18,460 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.243e+02 1.520e+02 1.734e+02 1.974e+02 3.517e+02, threshold=3.467e+02, percent-clipped=0.0 2024-09-22 19:39:27,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=87798.66666666667, ans=0.125 2024-09-22 19:39:48,360 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=87892.0, ans=0.125 2024-09-22 19:39:49,577 INFO [train.py:1198] (2/4) Epoch 5, batch 3250, loss[loss=0.266, ctc_loss=0.1928, cr_loss=0.3656, over 16943.00 frames. ], tot_loss[loss=0.2985, ctc_loss=0.2171, cr_loss=0.4069, over 3360234.74 frames. ], batch size: 42, lr: 2.22e-02, grad_scale: 32.0 2024-09-22 19:39:49,808 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=87892.0, ans=0.125 2024-09-22 19:40:31,934 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=87985.33333333333, ans=0.2 2024-09-22 19:40:47,011 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=12.14 vs. limit=15.0 2024-09-22 19:41:09,579 INFO [train.py:1198] (2/4) Epoch 5, batch 3300, loss[loss=0.2729, ctc_loss=0.1991, cr_loss=0.369, over 17300.00 frames. ], tot_loss[loss=0.2972, ctc_loss=0.2159, cr_loss=0.4064, over 3363128.67 frames. ], batch size: 46, lr: 2.22e-02, grad_scale: 32.0 2024-09-22 19:41:30,275 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=88172.0, ans=0.0 2024-09-22 19:41:41,997 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.82 vs. limit=15.0 2024-09-22 19:41:44,539 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=88218.66666666667, ans=0.125 2024-09-22 19:41:52,051 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=88218.66666666667, ans=0.2 2024-09-22 19:41:58,205 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.266e+02 1.544e+02 1.772e+02 2.234e+02 4.094e+02, threshold=3.543e+02, percent-clipped=4.0 2024-09-22 19:42:17,107 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=88312.0, ans=0.125 2024-09-22 19:42:29,375 INFO [train.py:1198] (2/4) Epoch 5, batch 3350, loss[loss=0.2573, ctc_loss=0.1848, cr_loss=0.3626, over 17202.00 frames. ], tot_loss[loss=0.298, ctc_loss=0.2166, cr_loss=0.407, over 3358097.86 frames. ], batch size: 41, lr: 2.22e-02, grad_scale: 32.0 2024-09-22 19:42:32,623 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=88358.66666666667, ans=0.0 2024-09-22 19:43:38,465 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 19:43:47,476 INFO [train.py:1198] (2/4) Epoch 5, batch 3400, loss[loss=0.2969, ctc_loss=0.2206, cr_loss=0.3815, over 17292.00 frames. ], tot_loss[loss=0.2951, ctc_loss=0.2141, cr_loss=0.405, over 3370741.81 frames. ], batch size: 51, lr: 2.21e-02, grad_scale: 32.0 2024-09-22 19:44:07,506 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=15.62 vs. limit=15.0 2024-09-22 19:44:22,313 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=88685.33333333333, ans=0.05 2024-09-22 19:44:26,958 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=88685.33333333333, ans=0.125 2024-09-22 19:44:26,974 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=88685.33333333333, ans=0.1 2024-09-22 19:44:34,720 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.286e+02 1.562e+02 1.776e+02 2.153e+02 3.268e+02, threshold=3.552e+02, percent-clipped=0.0 2024-09-22 19:44:47,534 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=16.55 vs. limit=22.5 2024-09-22 19:45:01,445 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 19:45:01,487 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=88778.66666666667, ans=0.0 2024-09-22 19:45:04,902 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=11.26 vs. limit=15.0 2024-09-22 19:45:05,685 INFO [train.py:1198] (2/4) Epoch 5, batch 3450, loss[loss=0.2626, ctc_loss=0.1873, cr_loss=0.3763, over 16927.00 frames. ], tot_loss[loss=0.2949, ctc_loss=0.2138, cr_loss=0.4057, over 3375175.53 frames. ], batch size: 42, lr: 2.21e-02, grad_scale: 32.0 2024-09-22 19:45:12,606 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=88825.33333333333, ans=0.0 2024-09-22 19:45:12,608 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=88825.33333333333, ans=0.125 2024-09-22 19:45:17,356 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=88825.33333333333, ans=0.0 2024-09-22 19:45:17,393 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=88825.33333333333, ans=0.09899494936611666 2024-09-22 19:45:25,255 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer_na.min_abs, batch_count=88872.0, ans=0.02 2024-09-22 19:46:02,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=88965.33333333333, ans=0.125 2024-09-22 19:46:03,123 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=10.89 vs. limit=15.0 2024-09-22 19:46:16,608 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=89012.0, ans=0.0 2024-09-22 19:46:25,704 INFO [train.py:1198] (2/4) Epoch 5, batch 3500, loss[loss=0.3048, ctc_loss=0.2226, cr_loss=0.4108, over 17280.00 frames. ], tot_loss[loss=0.2954, ctc_loss=0.2143, cr_loss=0.4056, over 3364020.69 frames. ], batch size: 46, lr: 2.21e-02, grad_scale: 32.0 2024-09-22 19:46:30,545 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=89058.66666666667, ans=0.125 2024-09-22 19:46:38,033 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=89058.66666666667, ans=0.0 2024-09-22 19:46:38,185 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=89058.66666666667, ans=0.0 2024-09-22 19:47:14,315 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 1.466e+02 1.595e+02 1.829e+02 3.245e+02, threshold=3.189e+02, percent-clipped=0.0 2024-09-22 19:47:19,914 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=4.93 vs. limit=15.0 2024-09-22 19:47:45,031 INFO [train.py:1198] (2/4) Epoch 5, batch 3550, loss[loss=0.3356, ctc_loss=0.246, cr_loss=0.448, over 17024.00 frames. ], tot_loss[loss=0.2954, ctc_loss=0.2143, cr_loss=0.4055, over 3352028.19 frames. ], batch size: 52, lr: 2.21e-02, grad_scale: 32.0 2024-09-22 19:47:57,996 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=89292.0, ans=0.09899494936611666 2024-09-22 19:48:03,364 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=10.63 vs. limit=15.0 2024-09-22 19:48:13,498 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=89338.66666666667, ans=0.0 2024-09-22 19:48:17,968 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=89385.33333333333, ans=0.2 2024-09-22 19:48:21,091 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=89385.33333333333, ans=0.125 2024-09-22 19:48:21,236 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=89385.33333333333, ans=0.07 2024-09-22 19:48:25,904 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=89385.33333333333, ans=0.125 2024-09-22 19:48:45,944 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=89478.66666666667, ans=0.125 2024-09-22 19:49:02,835 INFO [train.py:1198] (2/4) Epoch 5, batch 3600, loss[loss=0.3139, ctc_loss=0.2327, cr_loss=0.4058, over 15029.00 frames. ], tot_loss[loss=0.2963, ctc_loss=0.2151, cr_loss=0.406, over 3340043.17 frames. ], batch size: 89, lr: 2.20e-02, grad_scale: 32.0 2024-09-22 19:49:07,870 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=89525.33333333333, ans=0.125 2024-09-22 19:49:09,311 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=89525.33333333333, ans=0.125 2024-09-22 19:49:12,933 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.30 vs. limit=6.0 2024-09-22 19:49:16,278 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.44 vs. limit=6.0 2024-09-22 19:49:20,060 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=89572.0, ans=0.1 2024-09-22 19:49:21,675 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=89572.0, ans=0.1 2024-09-22 19:49:26,480 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=89572.0, ans=0.125 2024-09-22 19:49:49,381 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.188e+02 1.445e+02 1.594e+02 1.731e+02 2.971e+02, threshold=3.187e+02, percent-clipped=0.0 2024-09-22 19:49:52,849 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=89665.33333333333, ans=0.0 2024-09-22 19:50:01,309 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.02 vs. limit=6.0 2024-09-22 19:50:13,686 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.33 vs. limit=15.0 2024-09-22 19:50:16,138 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=89712.0, ans=0.1 2024-09-22 19:50:22,564 INFO [train.py:1198] (2/4) Epoch 5, batch 3650, loss[loss=0.3805, ctc_loss=0.2951, cr_loss=0.4273, over 11861.00 frames. ], tot_loss[loss=0.2957, ctc_loss=0.2147, cr_loss=0.4051, over 3343882.02 frames. ], batch size: 123, lr: 2.20e-02, grad_scale: 32.0 2024-09-22 19:51:28,450 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=89945.33333333333, ans=0.0 2024-09-22 19:51:28,528 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=89945.33333333333, ans=0.0 2024-09-22 19:51:43,156 INFO [train.py:1198] (2/4) Epoch 5, batch 3700, loss[loss=0.2816, ctc_loss=0.2072, cr_loss=0.372, over 17241.00 frames. ], tot_loss[loss=0.2971, ctc_loss=0.216, cr_loss=0.4059, over 3344625.36 frames. ], batch size: 50, lr: 2.20e-02, grad_scale: 32.0 2024-09-22 19:52:08,278 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=90038.66666666667, ans=0.0 2024-09-22 19:52:14,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=90085.33333333333, ans=0.125 2024-09-22 19:52:26,848 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=90085.33333333333, ans=0.1 2024-09-22 19:52:29,540 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.256e+02 1.562e+02 1.758e+02 2.028e+02 3.638e+02, threshold=3.517e+02, percent-clipped=1.0 2024-09-22 19:52:29,786 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.min_positive, batch_count=90132.0, ans=0.025 2024-09-22 19:52:56,572 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=90178.66666666667, ans=0.125 2024-09-22 19:53:01,155 INFO [train.py:1198] (2/4) Epoch 5, batch 3750, loss[loss=0.2593, ctc_loss=0.1848, cr_loss=0.3724, over 17028.00 frames. ], tot_loss[loss=0.2974, ctc_loss=0.2162, cr_loss=0.406, over 3339543.46 frames. ], batch size: 44, lr: 2.20e-02, grad_scale: 32.0 2024-09-22 19:53:06,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=90225.33333333333, ans=0.05 2024-09-22 19:53:09,225 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=90225.33333333333, ans=0.0 2024-09-22 19:53:11,063 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=90225.33333333333, ans=0.125 2024-09-22 19:53:19,126 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=90272.0, ans=0.125 2024-09-22 19:53:52,182 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=90365.33333333333, ans=0.125 2024-09-22 19:54:19,781 INFO [train.py:1198] (2/4) Epoch 5, batch 3800, loss[loss=0.3161, ctc_loss=0.2317, cr_loss=0.4221, over 17031.00 frames. ], tot_loss[loss=0.2996, ctc_loss=0.2181, cr_loss=0.4076, over 3321546.72 frames. ], batch size: 56, lr: 2.19e-02, grad_scale: 32.0 2024-09-22 19:54:24,595 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=90458.66666666667, ans=0.2 2024-09-22 19:54:30,053 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=6.20 vs. limit=15.0 2024-09-22 19:54:32,227 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=90458.66666666667, ans=0.0 2024-09-22 19:54:41,294 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=512, metric=2.63 vs. limit=15.0 2024-09-22 19:54:47,075 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=90505.33333333333, ans=0.0 2024-09-22 19:55:01,029 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=90552.0, ans=0.0 2024-09-22 19:55:06,884 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.208e+02 1.578e+02 1.852e+02 2.152e+02 4.120e+02, threshold=3.704e+02, percent-clipped=2.0 2024-09-22 19:55:16,290 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=90598.66666666667, ans=0.0 2024-09-22 19:55:30,991 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=90645.33333333333, ans=0.0 2024-09-22 19:55:38,741 INFO [train.py:1198] (2/4) Epoch 5, batch 3850, loss[loss=0.3774, ctc_loss=0.2876, cr_loss=0.4489, over 11727.00 frames. ], tot_loss[loss=0.3017, ctc_loss=0.2202, cr_loss=0.4078, over 3266589.30 frames. ], batch size: 123, lr: 2.19e-02, grad_scale: 32.0 2024-09-22 19:55:51,185 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=90692.0, ans=0.1 2024-09-22 19:55:54,453 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=15.19 vs. limit=15.0 2024-09-22 19:56:06,777 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=90738.66666666667, ans=0.2 2024-09-22 19:56:23,165 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=90832.0, ans=0.2 2024-09-22 19:56:42,251 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=11.93 vs. limit=22.5 2024-09-22 19:57:39,896 INFO [train.py:1198] (2/4) Epoch 6, batch 0, loss[loss=0.2843, ctc_loss=0.2052, cr_loss=0.3953, over 17092.00 frames. ], tot_loss[loss=0.2843, ctc_loss=0.2052, cr_loss=0.3953, over 17092.00 frames. ], batch size: 43, lr: 2.04e-02, grad_scale: 32.0 2024-09-22 19:57:39,897 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-22 19:57:55,107 INFO [train.py:1230] (2/4) Epoch 6, validation: loss=0.06886, ctc_loss=0.06886, cr_loss=9.986e-15, over 944034.00 frames. 2024-09-22 19:57:55,107 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-22 19:58:00,424 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=90906.66666666667, ans=0.025 2024-09-22 19:58:11,606 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=90953.33333333333, ans=0.125 2024-09-22 19:58:29,584 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys.whitening_limit, batch_count=91000.0, ans=6.0 2024-09-22 19:58:30,685 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=91000.0, ans=0.125 2024-09-22 19:58:38,534 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=91000.0, ans=0.04949747468305833 2024-09-22 19:58:50,053 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=91046.66666666667, ans=0.0 2024-09-22 19:58:51,239 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.268e+02 1.570e+02 1.853e+02 2.174e+02 4.194e+02, threshold=3.706e+02, percent-clipped=2.0 2024-09-22 19:58:53,101 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=91046.66666666667, ans=0.125 2024-09-22 19:58:54,738 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=91046.66666666667, ans=0.0 2024-09-22 19:59:06,754 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=91093.33333333333, ans=0.0 2024-09-22 19:59:19,094 INFO [train.py:1198] (2/4) Epoch 6, batch 50, loss[loss=0.3142, ctc_loss=0.2318, cr_loss=0.4118, over 17003.00 frames. ], tot_loss[loss=0.2989, ctc_loss=0.2173, cr_loss=0.408, over 751284.08 frames. ], batch size: 51, lr: 2.04e-02, grad_scale: 32.0 2024-09-22 19:59:20,973 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=91140.0, ans=0.1 2024-09-22 19:59:32,925 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=91140.0, ans=0.1 2024-09-22 19:59:32,967 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=91140.0, ans=0.125 2024-09-22 19:59:39,493 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=91186.66666666667, ans=0.1 2024-09-22 20:00:06,461 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=91233.33333333333, ans=0.125 2024-09-22 20:00:15,873 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=91280.0, ans=0.0 2024-09-22 20:00:26,292 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.whiten, num_groups=1, num_channels=192, metric=4.15 vs. limit=12.0 2024-09-22 20:00:41,134 INFO [train.py:1198] (2/4) Epoch 6, batch 100, loss[loss=0.2695, ctc_loss=0.1926, cr_loss=0.385, over 17094.00 frames. ], tot_loss[loss=0.2943, ctc_loss=0.2132, cr_loss=0.4054, over 1333845.47 frames. ], batch size: 49, lr: 2.04e-02, grad_scale: 32.0 2024-09-22 20:01:02,013 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=91420.0, ans=0.0 2024-09-22 20:01:03,763 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=91420.0, ans=0.125 2024-09-22 20:01:04,271 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten.whitening_limit, batch_count=91420.0, ans=15.0 2024-09-22 20:01:18,359 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=91466.66666666667, ans=0.125 2024-09-22 20:01:27,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=91513.33333333333, ans=0.125 2024-09-22 20:01:35,327 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.210e+02 1.399e+02 1.624e+02 1.941e+02 3.446e+02, threshold=3.247e+02, percent-clipped=0.0 2024-09-22 20:02:00,844 INFO [train.py:1198] (2/4) Epoch 6, batch 150, loss[loss=0.3242, ctc_loss=0.2381, cr_loss=0.4302, over 17020.00 frames. ], tot_loss[loss=0.2913, ctc_loss=0.2106, cr_loss=0.4033, over 1788408.16 frames. ], batch size: 52, lr: 2.04e-02, grad_scale: 32.0 2024-09-22 20:02:28,241 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=91653.33333333333, ans=0.2 2024-09-22 20:02:33,486 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.12 vs. limit=15.0 2024-09-22 20:02:33,529 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=11.98 vs. limit=15.0 2024-09-22 20:02:50,273 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=91746.66666666667, ans=0.1 2024-09-22 20:03:02,918 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.94 vs. limit=10.0 2024-09-22 20:03:03,052 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.61 vs. limit=15.0 2024-09-22 20:03:13,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=91793.33333333333, ans=0.0 2024-09-22 20:03:25,936 INFO [train.py:1198] (2/4) Epoch 6, batch 200, loss[loss=0.3258, ctc_loss=0.2358, cr_loss=0.4501, over 16520.00 frames. ], tot_loss[loss=0.2938, ctc_loss=0.2126, cr_loss=0.4057, over 2137728.39 frames. ], batch size: 66, lr: 2.03e-02, grad_scale: 32.0 2024-09-22 20:03:29,995 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=15.14 vs. limit=15.0 2024-09-22 20:03:42,362 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=91886.66666666667, ans=0.1 2024-09-22 20:03:51,998 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=91886.66666666667, ans=0.2 2024-09-22 20:03:59,933 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 20:04:08,998 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=91933.33333333333, ans=0.125 2024-09-22 20:04:16,962 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=91980.0, ans=0.09899494936611666 2024-09-22 20:04:24,703 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=91980.0, ans=0.125 2024-09-22 20:04:25,624 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.298e+02 1.556e+02 1.826e+02 2.254e+02 3.362e+02, threshold=3.652e+02, percent-clipped=2.0 2024-09-22 20:04:35,519 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=92026.66666666667, ans=0.0 2024-09-22 20:04:45,985 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=16.79 vs. limit=22.5 2024-09-22 20:04:51,205 INFO [train.py:1198] (2/4) Epoch 6, batch 250, loss[loss=0.2768, ctc_loss=0.2006, cr_loss=0.3812, over 17172.00 frames. ], tot_loss[loss=0.2924, ctc_loss=0.2114, cr_loss=0.4048, over 2407667.06 frames. ], batch size: 41, lr: 2.03e-02, grad_scale: 32.0 2024-09-22 20:05:38,903 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=92213.33333333333, ans=0.125 2024-09-22 20:05:44,125 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.73 vs. limit=15.0 2024-09-22 20:05:48,388 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=92213.33333333333, ans=0.125 2024-09-22 20:06:00,069 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.29 vs. limit=15.0 2024-09-22 20:06:06,789 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=8.54 vs. limit=15.0 2024-09-22 20:06:10,521 INFO [train.py:1198] (2/4) Epoch 6, batch 300, loss[loss=0.2594, ctc_loss=0.1878, cr_loss=0.3578, over 17177.00 frames. ], tot_loss[loss=0.2935, ctc_loss=0.2125, cr_loss=0.4051, over 2612790.69 frames. ], batch size: 41, lr: 2.03e-02, grad_scale: 32.0 2024-09-22 20:06:35,954 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=92353.33333333333, ans=0.125 2024-09-22 20:07:04,351 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.190e+02 1.501e+02 1.679e+02 1.995e+02 3.588e+02, threshold=3.358e+02, percent-clipped=0.0 2024-09-22 20:07:22,045 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=92493.33333333333, ans=0.0 2024-09-22 20:07:24,510 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.whiten, num_groups=1, num_channels=192, metric=4.33 vs. limit=12.0 2024-09-22 20:07:32,317 INFO [train.py:1198] (2/4) Epoch 6, batch 350, loss[loss=0.2563, ctc_loss=0.1821, cr_loss=0.3712, over 17069.00 frames. ], tot_loss[loss=0.291, ctc_loss=0.2103, cr_loss=0.4037, over 2788825.62 frames. ], batch size: 46, lr: 2.03e-02, grad_scale: 32.0 2024-09-22 20:08:29,170 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=92680.0, ans=0.125 2024-09-22 20:08:46,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=92726.66666666667, ans=0.07 2024-09-22 20:08:53,270 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.06 vs. limit=15.0 2024-09-22 20:08:55,868 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=92773.33333333333, ans=0.1 2024-09-22 20:08:57,314 INFO [train.py:1198] (2/4) Epoch 6, batch 400, loss[loss=0.2647, ctc_loss=0.1871, cr_loss=0.3882, over 17024.00 frames. ], tot_loss[loss=0.2895, ctc_loss=0.2091, cr_loss=0.4018, over 2914595.64 frames. ], batch size: 44, lr: 2.02e-02, grad_scale: 32.0 2024-09-22 20:09:05,618 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=92773.33333333333, ans=0.125 2024-09-22 20:09:19,473 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-22 20:09:21,325 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=13.76 vs. limit=22.5 2024-09-22 20:09:54,060 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.272e+02 1.481e+02 1.649e+02 1.890e+02 2.985e+02, threshold=3.299e+02, percent-clipped=0.0 2024-09-22 20:10:13,751 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=92960.0, ans=0.2 2024-09-22 20:10:19,708 INFO [train.py:1198] (2/4) Epoch 6, batch 450, loss[loss=0.2862, ctc_loss=0.2, cr_loss=0.4313, over 17075.00 frames. ], tot_loss[loss=0.2901, ctc_loss=0.2096, cr_loss=0.4025, over 3004405.19 frames. ], batch size: 46, lr: 2.02e-02, grad_scale: 32.0 2024-09-22 20:11:15,614 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=93146.66666666667, ans=0.125 2024-09-22 20:11:22,112 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=93193.33333333333, ans=0.125 2024-09-22 20:11:39,001 INFO [train.py:1198] (2/4) Epoch 6, batch 500, loss[loss=0.3075, ctc_loss=0.222, cr_loss=0.4279, over 16884.00 frames. ], tot_loss[loss=0.2903, ctc_loss=0.2097, cr_loss=0.4029, over 3083533.94 frames. ], batch size: 58, lr: 2.02e-02, grad_scale: 32.0 2024-09-22 20:11:39,329 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=93240.0, ans=0.1 2024-09-22 20:11:42,316 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=93240.0, ans=0.0 2024-09-22 20:11:51,690 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=93240.0, ans=0.0 2024-09-22 20:12:05,611 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=93286.66666666667, ans=0.0 2024-09-22 20:12:29,841 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=93380.0, ans=0.125 2024-09-22 20:12:31,403 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=93380.0, ans=0.0 2024-09-22 20:12:32,921 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.min_abs, batch_count=93380.0, ans=0.5 2024-09-22 20:12:37,312 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.206e+02 1.459e+02 1.650e+02 1.980e+02 2.967e+02, threshold=3.301e+02, percent-clipped=0.0 2024-09-22 20:12:43,569 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=93380.0, ans=0.0 2024-09-22 20:13:05,202 INFO [train.py:1198] (2/4) Epoch 6, batch 550, loss[loss=0.3089, ctc_loss=0.2256, cr_loss=0.4164, over 17036.00 frames. ], tot_loss[loss=0.2902, ctc_loss=0.2097, cr_loss=0.4024, over 3147339.22 frames. ], batch size: 52, lr: 2.02e-02, grad_scale: 32.0 2024-09-22 20:13:29,278 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=93520.0, ans=0.125 2024-09-22 20:13:49,695 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=93566.66666666667, ans=10.0 2024-09-22 20:13:59,203 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=93613.33333333333, ans=0.125 2024-09-22 20:14:30,621 INFO [train.py:1198] (2/4) Epoch 6, batch 600, loss[loss=0.2626, ctc_loss=0.1865, cr_loss=0.3808, over 17192.00 frames. ], tot_loss[loss=0.2895, ctc_loss=0.209, cr_loss=0.4021, over 3198985.11 frames. ], batch size: 41, lr: 2.02e-02, grad_scale: 32.0 2024-09-22 20:14:38,810 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=93706.66666666667, ans=0.125 2024-09-22 20:14:51,656 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=93753.33333333333, ans=0.1 2024-09-22 20:15:17,427 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.73 vs. limit=15.0 2024-09-22 20:15:21,996 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=93846.66666666667, ans=0.125 2024-09-22 20:15:24,959 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.069e+02 1.461e+02 1.595e+02 1.846e+02 3.407e+02, threshold=3.191e+02, percent-clipped=1.0 2024-09-22 20:15:45,071 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=512, metric=19.65 vs. limit=22.5 2024-09-22 20:15:46,205 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=93893.33333333333, ans=0.2 2024-09-22 20:15:47,635 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=93893.33333333333, ans=0.1 2024-09-22 20:15:50,710 INFO [train.py:1198] (2/4) Epoch 6, batch 650, loss[loss=0.3002, ctc_loss=0.2187, cr_loss=0.4078, over 16996.00 frames. ], tot_loss[loss=0.287, ctc_loss=0.2071, cr_loss=0.3995, over 3237048.11 frames. ], batch size: 56, lr: 2.01e-02, grad_scale: 32.0 2024-09-22 20:16:27,384 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=94033.33333333333, ans=0.125 2024-09-22 20:16:47,964 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=94080.0, ans=0.125 2024-09-22 20:17:09,771 INFO [train.py:1198] (2/4) Epoch 6, batch 700, loss[loss=0.3078, ctc_loss=0.2251, cr_loss=0.4136, over 17141.00 frames. ], tot_loss[loss=0.2876, ctc_loss=0.2075, cr_loss=0.4005, over 3269742.92 frames. ], batch size: 48, lr: 2.01e-02, grad_scale: 32.0 2024-09-22 20:17:55,970 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=6.38 vs. limit=15.0 2024-09-22 20:18:01,430 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=94313.33333333333, ans=0.0 2024-09-22 20:18:09,123 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.224e+02 1.435e+02 1.629e+02 1.890e+02 2.825e+02, threshold=3.258e+02, percent-clipped=0.0 2024-09-22 20:18:19,128 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=94360.0, ans=0.0 2024-09-22 20:18:23,846 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=94360.0, ans=0.0 2024-09-22 20:18:27,156 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=94360.0, ans=0.0 2024-09-22 20:18:34,803 INFO [train.py:1198] (2/4) Epoch 6, batch 750, loss[loss=0.2556, ctc_loss=0.1847, cr_loss=0.3543, over 17031.00 frames. ], tot_loss[loss=0.2885, ctc_loss=0.2082, cr_loss=0.4016, over 3286659.00 frames. ], batch size: 44, lr: 2.01e-02, grad_scale: 32.0 2024-09-22 20:18:38,194 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer_ff2.min_abs, batch_count=94406.66666666667, ans=0.1 2024-09-22 20:18:55,009 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=94453.33333333333, ans=0.0 2024-09-22 20:19:06,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=94453.33333333333, ans=0.125 2024-09-22 20:19:14,930 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=94500.0, ans=0.025 2024-09-22 20:19:18,177 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=94500.0, ans=0.1 2024-09-22 20:19:22,574 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=94500.0, ans=0.125 2024-09-22 20:19:58,979 INFO [train.py:1198] (2/4) Epoch 6, batch 800, loss[loss=0.3252, ctc_loss=0.2393, cr_loss=0.4294, over 15981.00 frames. ], tot_loss[loss=0.2884, ctc_loss=0.2081, cr_loss=0.4014, over 3303207.83 frames. ], batch size: 74, lr: 2.01e-02, grad_scale: 32.0 2024-09-22 20:20:11,794 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=94640.0, ans=0.125 2024-09-22 20:20:18,041 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=94686.66666666667, ans=0.0 2024-09-22 20:20:21,216 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=94686.66666666667, ans=0.0 2024-09-22 20:20:36,107 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.78 vs. limit=15.0 2024-09-22 20:20:38,910 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=94733.33333333333, ans=0.0 2024-09-22 20:20:40,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=94733.33333333333, ans=0.125 2024-09-22 20:20:42,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.min_positive, batch_count=94733.33333333333, ans=0.025 2024-09-22 20:20:51,780 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=94780.0, ans=0.125 2024-09-22 20:20:53,105 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.186e+02 1.473e+02 1.596e+02 1.875e+02 3.402e+02, threshold=3.192e+02, percent-clipped=2.0 2024-09-22 20:21:15,851 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=94826.66666666667, ans=0.0 2024-09-22 20:21:18,626 INFO [train.py:1198] (2/4) Epoch 6, batch 850, loss[loss=0.3199, ctc_loss=0.2278, cr_loss=0.4609, over 17212.00 frames. ], tot_loss[loss=0.2884, ctc_loss=0.2081, cr_loss=0.4017, over 3322874.70 frames. ], batch size: 55, lr: 2.00e-02, grad_scale: 32.0 2024-09-22 20:21:20,621 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=94873.33333333333, ans=0.125 2024-09-22 20:21:28,626 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=94873.33333333333, ans=0.125 2024-09-22 20:21:29,115 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.94 vs. limit=6.0 2024-09-22 20:21:35,058 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=94920.0, ans=0.125 2024-09-22 20:21:54,290 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=94966.66666666667, ans=0.0 2024-09-22 20:22:00,709 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=94966.66666666667, ans=0.125 2024-09-22 20:22:05,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=95013.33333333333, ans=0.07 2024-09-22 20:22:08,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=95013.33333333333, ans=0.125 2024-09-22 20:22:41,603 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=14.07 vs. limit=22.5 2024-09-22 20:22:43,771 INFO [train.py:1198] (2/4) Epoch 6, batch 900, loss[loss=0.3445, ctc_loss=0.2491, cr_loss=0.477, over 16526.00 frames. ], tot_loss[loss=0.2892, ctc_loss=0.2087, cr_loss=0.4024, over 3331875.99 frames. ], batch size: 66, lr: 2.00e-02, grad_scale: 32.0 2024-09-22 20:22:50,721 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=10.52 vs. limit=15.0 2024-09-22 20:23:01,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=95153.33333333333, ans=0.0 2024-09-22 20:23:21,923 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=95200.0, ans=0.0 2024-09-22 20:23:37,720 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.204e+02 1.462e+02 1.630e+02 1.916e+02 2.984e+02, threshold=3.259e+02, percent-clipped=0.0 2024-09-22 20:24:04,715 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=95340.0, ans=0.125 2024-09-22 20:24:05,922 INFO [train.py:1198] (2/4) Epoch 6, batch 950, loss[loss=0.2996, ctc_loss=0.218, cr_loss=0.4076, over 17027.00 frames. ], tot_loss[loss=0.2903, ctc_loss=0.2095, cr_loss=0.404, over 3338743.18 frames. ], batch size: 51, lr: 2.00e-02, grad_scale: 32.0 2024-09-22 20:25:22,399 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=95526.66666666667, ans=0.0 2024-09-22 20:25:28,488 INFO [train.py:1198] (2/4) Epoch 6, batch 1000, loss[loss=0.2717, ctc_loss=0.1974, cr_loss=0.3715, over 17105.00 frames. ], tot_loss[loss=0.2904, ctc_loss=0.2097, cr_loss=0.4035, over 3346572.08 frames. ], batch size: 49, lr: 2.00e-02, grad_scale: 32.0 2024-09-22 20:26:00,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=95666.66666666667, ans=0.1 2024-09-22 20:26:02,386 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=95666.66666666667, ans=0.125 2024-09-22 20:26:13,766 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=95666.66666666667, ans=0.07 2024-09-22 20:26:22,800 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.223e+02 1.403e+02 1.529e+02 1.831e+02 2.517e+02, threshold=3.058e+02, percent-clipped=0.0 2024-09-22 20:26:48,476 INFO [train.py:1198] (2/4) Epoch 6, batch 1050, loss[loss=0.2435, ctc_loss=0.1723, cr_loss=0.3559, over 17085.00 frames. ], tot_loss[loss=0.2886, ctc_loss=0.2084, cr_loss=0.4012, over 3346577.71 frames. ], batch size: 43, lr: 2.00e-02, grad_scale: 32.0 2024-09-22 20:27:04,671 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=95853.33333333333, ans=0.125 2024-09-22 20:27:08,350 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.48 vs. limit=15.0 2024-09-22 20:27:11,351 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=95853.33333333333, ans=0.2 2024-09-22 20:27:38,352 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=95900.0, ans=0.1 2024-09-22 20:27:55,801 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=95993.33333333333, ans=0.04949747468305833 2024-09-22 20:28:00,594 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=95993.33333333333, ans=0.125 2024-09-22 20:28:13,099 INFO [train.py:1198] (2/4) Epoch 6, batch 1100, loss[loss=0.384, ctc_loss=0.308, cr_loss=0.3801, over 11360.00 frames. ], tot_loss[loss=0.2892, ctc_loss=0.2089, cr_loss=0.4015, over 3347335.43 frames. ], batch size: 125, lr: 1.99e-02, grad_scale: 32.0 2024-09-22 20:28:15,106 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=96040.0, ans=0.2 2024-09-22 20:28:19,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=96040.0, ans=0.125 2024-09-22 20:28:26,077 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=96040.0, ans=0.0 2024-09-22 20:28:32,568 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=96086.66666666667, ans=0.125 2024-09-22 20:28:35,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=96086.66666666667, ans=0.1 2024-09-22 20:28:40,479 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=96086.66666666667, ans=0.125 2024-09-22 20:29:04,000 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.max_positive, batch_count=96180.0, ans=0.95 2024-09-22 20:29:12,626 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.277e+02 1.448e+02 1.600e+02 1.818e+02 3.191e+02, threshold=3.201e+02, percent-clipped=3.0 2024-09-22 20:29:25,593 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=96226.66666666667, ans=0.125 2024-09-22 20:29:28,610 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=96226.66666666667, ans=0.125 2024-09-22 20:29:38,029 INFO [train.py:1198] (2/4) Epoch 6, batch 1150, loss[loss=0.2925, ctc_loss=0.206, cr_loss=0.4327, over 17314.00 frames. ], tot_loss[loss=0.2889, ctc_loss=0.2086, cr_loss=0.4015, over 3358815.98 frames. ], batch size: 51, lr: 1.99e-02, grad_scale: 32.0 2024-09-22 20:29:46,222 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=96273.33333333333, ans=0.125 2024-09-22 20:30:02,038 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=96320.0, ans=0.0 2024-09-22 20:30:30,684 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=96413.33333333333, ans=0.125 2024-09-22 20:30:57,273 INFO [train.py:1198] (2/4) Epoch 6, batch 1200, loss[loss=0.2927, ctc_loss=0.2061, cr_loss=0.4332, over 17012.00 frames. ], tot_loss[loss=0.2873, ctc_loss=0.2073, cr_loss=0.4002, over 3354345.68 frames. ], batch size: 51, lr: 1.99e-02, grad_scale: 32.0 2024-09-22 20:30:57,525 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=96506.66666666667, ans=0.125 2024-09-22 20:31:13,907 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=96553.33333333333, ans=0.1 2024-09-22 20:31:28,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=96600.0, ans=0.2 2024-09-22 20:31:40,768 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=96600.0, ans=0.125 2024-09-22 20:31:51,563 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.195e+02 1.470e+02 1.636e+02 2.013e+02 4.309e+02, threshold=3.271e+02, percent-clipped=1.0 2024-09-22 20:31:56,648 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.min_positive, batch_count=96646.66666666667, ans=0.05 2024-09-22 20:32:17,037 INFO [train.py:1198] (2/4) Epoch 6, batch 1250, loss[loss=0.2978, ctc_loss=0.2125, cr_loss=0.4264, over 17301.00 frames. ], tot_loss[loss=0.287, ctc_loss=0.2069, cr_loss=0.4004, over 3359458.33 frames. ], batch size: 51, lr: 1.99e-02, grad_scale: 32.0 2024-09-22 20:32:43,176 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=96786.66666666667, ans=0.2 2024-09-22 20:33:00,695 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=96833.33333333333, ans=0.0 2024-09-22 20:33:08,467 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=96880.0, ans=0.025 2024-09-22 20:33:43,909 INFO [train.py:1198] (2/4) Epoch 6, batch 1300, loss[loss=0.293, ctc_loss=0.2117, cr_loss=0.4068, over 17338.00 frames. ], tot_loss[loss=0.2871, ctc_loss=0.2069, cr_loss=0.4011, over 3363021.34 frames. ], batch size: 51, lr: 1.99e-02, grad_scale: 32.0 2024-09-22 20:33:53,868 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=96973.33333333333, ans=0.2 2024-09-22 20:34:03,342 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=97020.0, ans=0.125 2024-09-22 20:34:33,420 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=97113.33333333333, ans=0.0 2024-09-22 20:34:41,239 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=97113.33333333333, ans=0.125 2024-09-22 20:34:42,526 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.191e+02 1.454e+02 1.645e+02 1.943e+02 2.545e+02, threshold=3.291e+02, percent-clipped=0.0 2024-09-22 20:34:42,924 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=97113.33333333333, ans=0.0 2024-09-22 20:34:45,881 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=97113.33333333333, ans=0.04949747468305833 2024-09-22 20:34:47,494 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=97113.33333333333, ans=0.125 2024-09-22 20:35:06,834 INFO [train.py:1198] (2/4) Epoch 6, batch 1350, loss[loss=0.3124, ctc_loss=0.223, cr_loss=0.4467, over 16417.00 frames. ], tot_loss[loss=0.2871, ctc_loss=0.2069, cr_loss=0.4009, over 3360593.93 frames. ], batch size: 66, lr: 1.98e-02, grad_scale: 32.0 2024-09-22 20:35:21,395 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=97253.33333333333, ans=0.125 2024-09-22 20:35:35,452 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=97253.33333333333, ans=0.125 2024-09-22 20:35:37,182 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=97300.0, ans=0.0 2024-09-22 20:36:17,885 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=97393.33333333333, ans=0.035 2024-09-22 20:36:25,754 INFO [train.py:1198] (2/4) Epoch 6, batch 1400, loss[loss=0.2644, ctc_loss=0.1884, cr_loss=0.3798, over 17234.00 frames. ], tot_loss[loss=0.2867, ctc_loss=0.2066, cr_loss=0.4004, over 3359842.19 frames. ], batch size: 50, lr: 1.98e-02, grad_scale: 32.0 2024-09-22 20:36:31,031 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=97440.0, ans=0.1 2024-09-22 20:36:39,011 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=97440.0, ans=0.2 2024-09-22 20:37:11,097 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=97533.33333333333, ans=0.125 2024-09-22 20:37:19,112 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=97580.0, ans=0.0 2024-09-22 20:37:24,530 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.201e+02 1.419e+02 1.584e+02 1.972e+02 3.775e+02, threshold=3.168e+02, percent-clipped=1.0 2024-09-22 20:37:39,161 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=97626.66666666667, ans=0.0 2024-09-22 20:37:48,278 INFO [train.py:1198] (2/4) Epoch 6, batch 1450, loss[loss=0.2421, ctc_loss=0.1714, cr_loss=0.3535, over 17294.00 frames. ], tot_loss[loss=0.2859, ctc_loss=0.206, cr_loss=0.3994, over 3359810.79 frames. ], batch size: 42, lr: 1.98e-02, grad_scale: 32.0 2024-09-22 20:37:51,672 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=97673.33333333333, ans=0.0 2024-09-22 20:37:56,557 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=97673.33333333333, ans=0.1 2024-09-22 20:38:12,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=97720.0, ans=0.09899494936611666 2024-09-22 20:38:49,934 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=97813.33333333333, ans=0.125 2024-09-22 20:39:12,837 INFO [train.py:1198] (2/4) Epoch 6, batch 1500, loss[loss=0.2195, ctc_loss=0.1539, cr_loss=0.328, over 17058.00 frames. ], tot_loss[loss=0.286, ctc_loss=0.2061, cr_loss=0.3995, over 3360227.50 frames. ], batch size: 39, lr: 1.98e-02, grad_scale: 32.0 2024-09-22 20:39:24,434 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=97906.66666666667, ans=0.1 2024-09-22 20:39:29,264 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=97953.33333333333, ans=0.0 2024-09-22 20:40:09,174 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.237e+02 1.478e+02 1.585e+02 1.791e+02 2.453e+02, threshold=3.170e+02, percent-clipped=0.0 2024-09-22 20:40:31,917 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.75 vs. limit=10.0 2024-09-22 20:40:33,019 INFO [train.py:1198] (2/4) Epoch 6, batch 1550, loss[loss=0.308, ctc_loss=0.2189, cr_loss=0.4453, over 17051.00 frames. ], tot_loss[loss=0.2874, ctc_loss=0.2072, cr_loss=0.4006, over 3348013.46 frames. ], batch size: 52, lr: 1.98e-02, grad_scale: 32.0 2024-09-22 20:40:34,961 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=98140.0, ans=0.1 2024-09-22 20:40:59,236 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=14.56 vs. limit=15.0 2024-09-22 20:41:01,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=98186.66666666667, ans=0.125 2024-09-22 20:41:08,323 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=98233.33333333333, ans=0.1 2024-09-22 20:41:16,360 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=98233.33333333333, ans=0.125 2024-09-22 20:41:25,567 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=98280.0, ans=0.1 2024-09-22 20:41:25,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=98280.0, ans=0.0 2024-09-22 20:41:27,395 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=98280.0, ans=0.2 2024-09-22 20:41:41,447 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=98326.66666666667, ans=0.125 2024-09-22 20:41:45,957 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=98326.66666666667, ans=0.1 2024-09-22 20:41:51,963 INFO [train.py:1198] (2/4) Epoch 6, batch 1600, loss[loss=0.2567, ctc_loss=0.1821, cr_loss=0.3726, over 17291.00 frames. ], tot_loss[loss=0.2869, ctc_loss=0.2068, cr_loss=0.4005, over 3355399.94 frames. ], batch size: 46, lr: 1.97e-02, grad_scale: 32.0 2024-09-22 20:42:37,814 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=12.93 vs. limit=15.0 2024-09-22 20:42:49,617 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=98513.33333333333, ans=0.2 2024-09-22 20:42:52,417 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.223e+02 1.461e+02 1.657e+02 2.022e+02 3.350e+02, threshold=3.314e+02, percent-clipped=2.0 2024-09-22 20:42:52,664 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=98513.33333333333, ans=0.125 2024-09-22 20:43:00,841 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=98560.0, ans=0.1 2024-09-22 20:43:00,847 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=98560.0, ans=0.1 2024-09-22 20:43:00,875 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=98560.0, ans=0.0 2024-09-22 20:43:02,729 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.22 vs. limit=15.0 2024-09-22 20:43:05,589 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=98560.0, ans=0.125 2024-09-22 20:43:16,606 INFO [train.py:1198] (2/4) Epoch 6, batch 1650, loss[loss=0.2462, ctc_loss=0.1803, cr_loss=0.3298, over 17119.00 frames. ], tot_loss[loss=0.2861, ctc_loss=0.2062, cr_loss=0.3994, over 3360378.00 frames. ], batch size: 40, lr: 1.97e-02, grad_scale: 32.0 2024-09-22 20:43:27,928 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.99 vs. limit=15.0 2024-09-22 20:43:29,766 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn2.whiten, num_groups=1, num_channels=192, metric=9.98 vs. limit=22.5 2024-09-22 20:43:59,458 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=98700.0, ans=0.125 2024-09-22 20:44:04,560 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.71 vs. limit=6.0 2024-09-22 20:44:13,427 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=98746.66666666667, ans=0.125 2024-09-22 20:44:15,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=98746.66666666667, ans=0.1 2024-09-22 20:44:23,145 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=98793.33333333333, ans=0.125 2024-09-22 20:44:30,887 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=98793.33333333333, ans=0.0 2024-09-22 20:44:37,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=98793.33333333333, ans=0.0 2024-09-22 20:44:40,254 INFO [train.py:1198] (2/4) Epoch 6, batch 1700, loss[loss=0.2889, ctc_loss=0.2096, cr_loss=0.3966, over 16892.00 frames. ], tot_loss[loss=0.2858, ctc_loss=0.2059, cr_loss=0.3996, over 3358006.26 frames. ], batch size: 58, lr: 1.97e-02, grad_scale: 32.0 2024-09-22 20:44:58,058 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=98886.66666666667, ans=0.0 2024-09-22 20:45:23,548 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=98933.33333333333, ans=0.0 2024-09-22 20:45:30,227 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.33 vs. limit=15.0 2024-09-22 20:45:36,068 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.207e+02 1.362e+02 1.512e+02 1.797e+02 3.142e+02, threshold=3.023e+02, percent-clipped=0.0 2024-09-22 20:45:58,797 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=99073.33333333333, ans=0.125 2024-09-22 20:46:00,096 INFO [train.py:1198] (2/4) Epoch 6, batch 1750, loss[loss=0.303, ctc_loss=0.2233, cr_loss=0.3981, over 15993.00 frames. ], tot_loss[loss=0.2876, ctc_loss=0.2072, cr_loss=0.4018, over 3359663.03 frames. ], batch size: 74, lr: 1.97e-02, grad_scale: 32.0 2024-09-22 20:46:00,883 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.56 vs. limit=6.0 2024-09-22 20:46:24,454 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=99120.0, ans=0.0 2024-09-22 20:46:32,502 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=99166.66666666667, ans=0.0 2024-09-22 20:46:49,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=99213.33333333333, ans=0.04949747468305833 2024-09-22 20:47:18,722 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=99260.0, ans=0.0 2024-09-22 20:47:24,698 INFO [train.py:1198] (2/4) Epoch 6, batch 1800, loss[loss=0.3887, ctc_loss=0.2996, cr_loss=0.4458, over 11647.00 frames. ], tot_loss[loss=0.2872, ctc_loss=0.207, cr_loss=0.4012, over 3350612.97 frames. ], batch size: 123, lr: 1.96e-02, grad_scale: 32.0 2024-09-22 20:47:26,828 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.79 vs. limit=15.0 2024-09-22 20:48:22,609 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.233e+02 1.401e+02 1.510e+02 1.768e+02 2.829e+02, threshold=3.019e+02, percent-clipped=0.0 2024-09-22 20:48:34,094 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=99493.33333333333, ans=0.2 2024-09-22 20:48:41,888 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=99493.33333333333, ans=0.0 2024-09-22 20:48:46,312 INFO [train.py:1198] (2/4) Epoch 6, batch 1850, loss[loss=0.3093, ctc_loss=0.2232, cr_loss=0.4308, over 16917.00 frames. ], tot_loss[loss=0.2855, ctc_loss=0.2056, cr_loss=0.3997, over 3359963.88 frames. ], batch size: 58, lr: 1.96e-02, grad_scale: 32.0 2024-09-22 20:48:56,070 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=9.45 vs. limit=15.0 2024-09-22 20:48:57,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=99540.0, ans=0.2 2024-09-22 20:49:34,275 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.min_positive, batch_count=99633.33333333333, ans=0.025 2024-09-22 20:49:34,289 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=99633.33333333333, ans=0.05 2024-09-22 20:50:09,094 INFO [train.py:1198] (2/4) Epoch 6, batch 1900, loss[loss=0.2546, ctc_loss=0.1807, cr_loss=0.3695, over 17174.00 frames. ], tot_loss[loss=0.2857, ctc_loss=0.2059, cr_loss=0.3994, over 3348819.00 frames. ], batch size: 41, lr: 1.96e-02, grad_scale: 32.0 2024-09-22 20:50:14,214 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=99773.33333333333, ans=0.0 2024-09-22 20:50:19,499 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=21.42 vs. limit=22.5 2024-09-22 20:51:05,354 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.166e+02 1.450e+02 1.646e+02 2.008e+02 3.199e+02, threshold=3.291e+02, percent-clipped=1.0 2024-09-22 20:51:25,282 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.60 vs. limit=15.0 2024-09-22 20:51:29,421 INFO [train.py:1198] (2/4) Epoch 6, batch 1950, loss[loss=0.2559, ctc_loss=0.1866, cr_loss=0.3465, over 16672.00 frames. ], tot_loss[loss=0.2863, ctc_loss=0.2064, cr_loss=0.3994, over 3335255.25 frames. ], batch size: 37, lr: 1.96e-02, grad_scale: 32.0 2024-09-22 20:51:33,357 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=5.69 vs. limit=10.0 2024-09-22 20:51:42,452 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=100006.66666666667, ans=0.125 2024-09-22 20:52:36,970 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=7.91 vs. limit=15.0 2024-09-22 20:52:53,473 INFO [train.py:1198] (2/4) Epoch 6, batch 2000, loss[loss=0.2921, ctc_loss=0.2066, cr_loss=0.4275, over 17023.00 frames. ], tot_loss[loss=0.2877, ctc_loss=0.2075, cr_loss=0.4011, over 3340154.10 frames. ], batch size: 44, lr: 1.96e-02, grad_scale: 32.0 2024-09-22 20:53:08,857 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=20.03 vs. limit=22.5 2024-09-22 20:53:11,552 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=100286.66666666667, ans=0.0 2024-09-22 20:53:54,074 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.89 vs. limit=15.0 2024-09-22 20:53:54,724 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.289e+02 1.463e+02 1.652e+02 2.000e+02 4.397e+02, threshold=3.304e+02, percent-clipped=3.0 2024-09-22 20:54:18,663 INFO [train.py:1198] (2/4) Epoch 6, batch 2050, loss[loss=0.305, ctc_loss=0.2195, cr_loss=0.4275, over 16747.00 frames. ], tot_loss[loss=0.2867, ctc_loss=0.2067, cr_loss=0.3997, over 3345088.25 frames. ], batch size: 61, lr: 1.95e-02, grad_scale: 32.0 2024-09-22 20:54:30,787 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=21.35 vs. limit=22.5 2024-09-22 20:54:39,633 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=100520.0, ans=0.0 2024-09-22 20:55:03,509 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=100566.66666666667, ans=0.0 2024-09-22 20:55:08,537 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=100613.33333333333, ans=0.125 2024-09-22 20:55:35,376 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=7.33 vs. limit=15.0 2024-09-22 20:55:35,592 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=14.82 vs. limit=22.5 2024-09-22 20:55:37,972 INFO [train.py:1198] (2/4) Epoch 6, batch 2100, loss[loss=0.2436, ctc_loss=0.1734, cr_loss=0.3511, over 17057.00 frames. ], tot_loss[loss=0.2862, ctc_loss=0.2064, cr_loss=0.3993, over 3349275.20 frames. ], batch size: 39, lr: 1.95e-02, grad_scale: 32.0 2024-09-22 20:55:50,019 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=7.41 vs. limit=15.0 2024-09-22 20:56:07,325 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=5.33 vs. limit=12.0 2024-09-22 20:56:10,386 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=100800.0, ans=0.025 2024-09-22 20:56:18,276 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=100800.0, ans=0.1 2024-09-22 20:56:33,956 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.234e+02 1.439e+02 1.570e+02 1.892e+02 4.315e+02, threshold=3.139e+02, percent-clipped=1.0 2024-09-22 20:56:39,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=100846.66666666667, ans=0.125 2024-09-22 20:56:58,026 INFO [train.py:1198] (2/4) Epoch 6, batch 2150, loss[loss=0.2834, ctc_loss=0.2073, cr_loss=0.3802, over 17196.00 frames. ], tot_loss[loss=0.2877, ctc_loss=0.2076, cr_loss=0.4007, over 3348724.94 frames. ], batch size: 55, lr: 1.95e-02, grad_scale: 32.0 2024-09-22 20:57:17,921 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=100986.66666666667, ans=0.1 2024-09-22 20:57:43,426 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=101033.33333333333, ans=0.0 2024-09-22 20:57:56,262 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=101080.0, ans=0.0 2024-09-22 20:57:59,254 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=101080.0, ans=0.0 2024-09-22 20:58:16,019 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=101126.66666666667, ans=0.0 2024-09-22 20:58:17,810 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=101126.66666666667, ans=0.0 2024-09-22 20:58:25,368 INFO [train.py:1198] (2/4) Epoch 6, batch 2200, loss[loss=0.2542, ctc_loss=0.1825, cr_loss=0.3584, over 17195.00 frames. ], tot_loss[loss=0.2877, ctc_loss=0.2074, cr_loss=0.4019, over 3352559.04 frames. ], batch size: 41, lr: 1.95e-02, grad_scale: 32.0 2024-09-22 20:58:30,495 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=101173.33333333333, ans=0.0 2024-09-22 20:58:40,664 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=101173.33333333333, ans=0.125 2024-09-22 20:58:44,245 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=11.71 vs. limit=15.0 2024-09-22 20:58:53,419 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=101220.0, ans=0.2 2024-09-22 20:58:54,950 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=101220.0, ans=0.1 2024-09-22 20:59:08,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.min_abs, batch_count=101266.66666666667, ans=0.5 2024-09-22 20:59:21,769 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 20:59:22,925 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.199e+02 1.501e+02 1.682e+02 2.031e+02 3.137e+02, threshold=3.364e+02, percent-clipped=0.0 2024-09-22 20:59:34,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=101360.0, ans=0.125 2024-09-22 20:59:46,926 INFO [train.py:1198] (2/4) Epoch 6, batch 2250, loss[loss=0.3001, ctc_loss=0.2163, cr_loss=0.4191, over 17205.00 frames. ], tot_loss[loss=0.2878, ctc_loss=0.2073, cr_loss=0.4024, over 3351190.34 frames. ], batch size: 50, lr: 1.95e-02, grad_scale: 32.0 2024-09-22 20:59:50,409 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=101406.66666666667, ans=0.05 2024-09-22 21:00:09,517 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=101453.33333333333, ans=0.125 2024-09-22 21:00:41,318 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=101546.66666666667, ans=0.0 2024-09-22 21:00:54,024 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=101593.33333333333, ans=0.125 2024-09-22 21:01:05,326 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=13.15 vs. limit=15.0 2024-09-22 21:01:06,134 INFO [train.py:1198] (2/4) Epoch 6, batch 2300, loss[loss=0.2678, ctc_loss=0.19, cr_loss=0.3887, over 17009.00 frames. ], tot_loss[loss=0.2885, ctc_loss=0.2079, cr_loss=0.4026, over 3332986.85 frames. ], batch size: 44, lr: 1.94e-02, grad_scale: 32.0 2024-09-22 21:01:08,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=101640.0, ans=0.125 2024-09-22 21:01:33,915 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.07 vs. limit=10.0 2024-09-22 21:01:47,649 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=101733.33333333333, ans=0.125 2024-09-22 21:02:06,859 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.168e+02 1.444e+02 1.711e+02 1.913e+02 2.754e+02, threshold=3.422e+02, percent-clipped=0.0 2024-09-22 21:02:30,422 INFO [train.py:1198] (2/4) Epoch 6, batch 2350, loss[loss=0.2975, ctc_loss=0.2119, cr_loss=0.4281, over 17018.00 frames. ], tot_loss[loss=0.2865, ctc_loss=0.2061, cr_loss=0.4017, over 3345112.73 frames. ], batch size: 52, lr: 1.94e-02, grad_scale: 32.0 2024-09-22 21:02:53,069 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=101920.0, ans=0.0 2024-09-22 21:03:14,683 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=101966.66666666667, ans=0.125 2024-09-22 21:03:21,170 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=102013.33333333333, ans=0.125 2024-09-22 21:03:21,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=102013.33333333333, ans=0.125 2024-09-22 21:03:55,323 INFO [train.py:1198] (2/4) Epoch 6, batch 2400, loss[loss=0.2759, ctc_loss=0.2043, cr_loss=0.3578, over 16943.00 frames. ], tot_loss[loss=0.2851, ctc_loss=0.2052, cr_loss=0.3999, over 3351683.89 frames. ], batch size: 58, lr: 1.94e-02, grad_scale: 32.0 2024-09-22 21:04:25,376 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=102200.0, ans=0.125 2024-09-22 21:04:50,185 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.229e+02 1.414e+02 1.609e+02 1.881e+02 2.919e+02, threshold=3.217e+02, percent-clipped=0.0 2024-09-22 21:05:09,792 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=102293.33333333333, ans=0.125 2024-09-22 21:05:14,226 INFO [train.py:1198] (2/4) Epoch 6, batch 2450, loss[loss=0.3041, ctc_loss=0.2244, cr_loss=0.3988, over 16755.00 frames. ], tot_loss[loss=0.2854, ctc_loss=0.2054, cr_loss=0.4004, over 3356110.66 frames. ], batch size: 61, lr: 1.94e-02, grad_scale: 32.0 2024-09-22 21:05:19,315 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=102340.0, ans=0.125 2024-09-22 21:05:42,207 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=10.41 vs. limit=15.0 2024-09-22 21:06:10,319 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=102480.0, ans=0.0 2024-09-22 21:06:20,324 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=20.61 vs. limit=22.5 2024-09-22 21:06:21,596 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 21:06:24,635 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=102526.66666666667, ans=0.125 2024-09-22 21:06:29,360 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=102526.66666666667, ans=0.125 2024-09-22 21:06:30,859 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=102526.66666666667, ans=0.125 2024-09-22 21:06:33,760 INFO [train.py:1198] (2/4) Epoch 6, batch 2500, loss[loss=0.3032, ctc_loss=0.2172, cr_loss=0.4297, over 17042.00 frames. ], tot_loss[loss=0.2846, ctc_loss=0.2047, cr_loss=0.3998, over 3353311.50 frames. ], batch size: 52, lr: 1.94e-02, grad_scale: 32.0 2024-09-22 21:06:54,254 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=102620.0, ans=0.125 2024-09-22 21:06:54,258 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=102620.0, ans=0.0 2024-09-22 21:06:55,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=102620.0, ans=0.025 2024-09-22 21:07:33,729 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 21:07:34,921 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.228e+02 1.454e+02 1.581e+02 1.860e+02 2.771e+02, threshold=3.162e+02, percent-clipped=0.0 2024-09-22 21:07:52,134 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=14.18 vs. limit=22.5 2024-09-22 21:08:00,067 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=102806.66666666667, ans=0.0 2024-09-22 21:08:01,446 INFO [train.py:1198] (2/4) Epoch 6, batch 2550, loss[loss=0.3317, ctc_loss=0.2441, cr_loss=0.4379, over 17354.00 frames. ], tot_loss[loss=0.2853, ctc_loss=0.2051, cr_loss=0.4013, over 3357862.48 frames. ], batch size: 48, lr: 1.93e-02, grad_scale: 32.0 2024-09-22 21:08:09,612 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=102806.66666666667, ans=0.2 2024-09-22 21:08:20,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=102853.33333333333, ans=0.125 2024-09-22 21:08:26,719 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=102853.33333333333, ans=0.125 2024-09-22 21:08:30,647 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.95 vs. limit=12.0 2024-09-22 21:08:42,059 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=10.98 vs. limit=15.0 2024-09-22 21:08:53,776 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=102946.66666666667, ans=0.125 2024-09-22 21:09:15,924 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=102993.33333333333, ans=0.125 2024-09-22 21:09:16,756 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=7.67 vs. limit=15.0 2024-09-22 21:09:23,605 INFO [train.py:1198] (2/4) Epoch 6, batch 2600, loss[loss=0.2616, ctc_loss=0.1907, cr_loss=0.3543, over 17211.00 frames. ], tot_loss[loss=0.2855, ctc_loss=0.2052, cr_loss=0.4015, over 3359500.73 frames. ], batch size: 47, lr: 1.93e-02, grad_scale: 32.0 2024-09-22 21:09:33,415 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=103040.0, ans=0.125 2024-09-22 21:09:55,365 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=103133.33333333333, ans=0.0 2024-09-22 21:09:58,521 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=103133.33333333333, ans=0.125 2024-09-22 21:10:15,781 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=103180.0, ans=0.125 2024-09-22 21:10:18,671 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.169e+02 1.459e+02 1.706e+02 2.072e+02 3.287e+02, threshold=3.412e+02, percent-clipped=1.0 2024-09-22 21:10:34,344 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=103226.66666666667, ans=0.125 2024-09-22 21:10:42,337 INFO [train.py:1198] (2/4) Epoch 6, batch 2650, loss[loss=0.2513, ctc_loss=0.1803, cr_loss=0.355, over 16997.00 frames. ], tot_loss[loss=0.2853, ctc_loss=0.205, cr_loss=0.4015, over 3362257.06 frames. ], batch size: 44, lr: 1.93e-02, grad_scale: 32.0 2024-09-22 21:10:42,753 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-22 21:10:53,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=103273.33333333333, ans=0.125 2024-09-22 21:11:01,904 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=103320.0, ans=0.2 2024-09-22 21:11:25,925 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=103366.66666666667, ans=0.0 2024-09-22 21:11:27,569 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=103366.66666666667, ans=0.0 2024-09-22 21:11:37,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=103413.33333333333, ans=0.1 2024-09-22 21:11:47,424 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=103460.0, ans=0.125 2024-09-22 21:12:06,888 INFO [train.py:1198] (2/4) Epoch 6, batch 2700, loss[loss=0.2962, ctc_loss=0.2097, cr_loss=0.4328, over 17219.00 frames. ], tot_loss[loss=0.2853, ctc_loss=0.205, cr_loss=0.4015, over 3364779.21 frames. ], batch size: 47, lr: 1.93e-02, grad_scale: 32.0 2024-09-22 21:12:16,854 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=103506.66666666667, ans=0.125 2024-09-22 21:13:05,402 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.175e+02 1.442e+02 1.580e+02 1.786e+02 2.700e+02, threshold=3.159e+02, percent-clipped=0.0 2024-09-22 21:13:05,716 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=103646.66666666667, ans=0.125 2024-09-22 21:13:15,222 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=103693.33333333333, ans=0.035 2024-09-22 21:13:27,214 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=103693.33333333333, ans=0.0 2024-09-22 21:13:31,579 INFO [train.py:1198] (2/4) Epoch 6, batch 2750, loss[loss=0.2714, ctc_loss=0.1959, cr_loss=0.3772, over 16899.00 frames. ], tot_loss[loss=0.2849, ctc_loss=0.2047, cr_loss=0.4011, over 3362080.32 frames. ], batch size: 58, lr: 1.93e-02, grad_scale: 32.0 2024-09-22 21:14:06,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=103833.33333333333, ans=0.0 2024-09-22 21:14:13,901 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.13 vs. limit=10.0 2024-09-22 21:14:15,546 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.37 vs. limit=15.0 2024-09-22 21:14:21,037 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=103880.0, ans=0.125 2024-09-22 21:14:51,040 INFO [train.py:1198] (2/4) Epoch 6, batch 2800, loss[loss=0.2981, ctc_loss=0.2137, cr_loss=0.422, over 16867.00 frames. ], tot_loss[loss=0.2843, ctc_loss=0.2042, cr_loss=0.4005, over 3369136.07 frames. ], batch size: 58, lr: 1.92e-02, grad_scale: 32.0 2024-09-22 21:15:19,927 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=104020.0, ans=0.125 2024-09-22 21:15:45,237 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.max_positive, batch_count=104113.33333333333, ans=0.95 2024-09-22 21:15:46,539 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.163e+02 1.391e+02 1.557e+02 1.769e+02 3.866e+02, threshold=3.114e+02, percent-clipped=1.0 2024-09-22 21:16:10,246 INFO [train.py:1198] (2/4) Epoch 6, batch 2850, loss[loss=0.2816, ctc_loss=0.2008, cr_loss=0.4039, over 17308.00 frames. ], tot_loss[loss=0.2838, ctc_loss=0.2039, cr_loss=0.3993, over 3364430.94 frames. ], batch size: 46, lr: 1.92e-02, grad_scale: 32.0 2024-09-22 21:16:40,053 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=104253.33333333333, ans=0.04949747468305833 2024-09-22 21:16:43,156 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=104300.0, ans=0.0 2024-09-22 21:17:03,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=104346.66666666667, ans=0.1 2024-09-22 21:17:05,461 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.27 vs. limit=15.0 2024-09-22 21:17:22,588 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=104393.33333333333, ans=0.125 2024-09-22 21:17:22,691 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 21:17:33,728 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=104440.0, ans=0.1 2024-09-22 21:17:35,031 INFO [train.py:1198] (2/4) Epoch 6, batch 2900, loss[loss=0.3362, ctc_loss=0.2411, cr_loss=0.4755, over 16775.00 frames. ], tot_loss[loss=0.2839, ctc_loss=0.204, cr_loss=0.3997, over 3367497.62 frames. ], batch size: 61, lr: 1.92e-02, grad_scale: 32.0 2024-09-22 21:17:36,927 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=104440.0, ans=0.1 2024-09-22 21:18:28,351 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=104580.0, ans=0.125 2024-09-22 21:18:35,984 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.209e+02 1.508e+02 1.666e+02 1.915e+02 2.988e+02, threshold=3.332e+02, percent-clipped=0.0 2024-09-22 21:18:53,960 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=104626.66666666667, ans=0.2 2024-09-22 21:18:54,617 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=6.57 vs. limit=12.0 2024-09-22 21:18:59,987 INFO [train.py:1198] (2/4) Epoch 6, batch 2950, loss[loss=0.2867, ctc_loss=0.2049, cr_loss=0.4094, over 17289.00 frames. ], tot_loss[loss=0.284, ctc_loss=0.204, cr_loss=0.4001, over 3374217.73 frames. ], batch size: 46, lr: 1.92e-02, grad_scale: 32.0 2024-09-22 21:19:06,659 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=104673.33333333333, ans=0.2 2024-09-22 21:19:43,513 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=104766.66666666667, ans=0.0 2024-09-22 21:20:11,711 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=104860.0, ans=0.0 2024-09-22 21:20:19,065 INFO [train.py:1198] (2/4) Epoch 6, batch 3000, loss[loss=0.3078, ctc_loss=0.2268, cr_loss=0.4052, over 16034.00 frames. ], tot_loss[loss=0.2863, ctc_loss=0.2059, cr_loss=0.4019, over 3363416.20 frames. ], batch size: 74, lr: 1.92e-02, grad_scale: 32.0 2024-09-22 21:20:19,065 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-22 21:20:30,852 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.4.encoder.layers.2.self_attn_weights, attn_weights_entropy = tensor([3.3065, 2.8413, 3.0236, 3.2183], device='cuda:2') 2024-09-22 21:20:34,498 INFO [train.py:1230] (2/4) Epoch 6, validation: loss=0.06097, ctc_loss=0.06097, cr_loss=6.736e-15, over 944034.00 frames. 2024-09-22 21:20:34,499 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-22 21:20:41,162 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=104906.66666666667, ans=0.0 2024-09-22 21:20:45,611 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.min_positive, batch_count=104906.66666666667, ans=0.025 2024-09-22 21:20:58,260 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=104953.33333333333, ans=0.0 2024-09-22 21:21:02,863 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=104953.33333333333, ans=0.125 2024-09-22 21:21:07,503 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=105000.0, ans=0.0 2024-09-22 21:21:26,372 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=105046.66666666667, ans=0.0 2024-09-22 21:21:29,318 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.397e+02 1.557e+02 1.721e+02 3.093e+02, threshold=3.113e+02, percent-clipped=0.0 2024-09-22 21:21:52,864 INFO [train.py:1198] (2/4) Epoch 6, batch 3050, loss[loss=0.2913, ctc_loss=0.2135, cr_loss=0.3887, over 17021.00 frames. ], tot_loss[loss=0.2854, ctc_loss=0.2054, cr_loss=0.4002, over 3370709.33 frames. ], batch size: 56, lr: 1.92e-02, grad_scale: 32.0 2024-09-22 21:21:53,185 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=105140.0, ans=0.0 2024-09-22 21:22:07,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=105186.66666666667, ans=0.2 2024-09-22 21:22:23,385 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.32 vs. limit=10.0 2024-09-22 21:22:46,327 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=8.83 vs. limit=15.0 2024-09-22 21:22:56,821 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=105326.66666666667, ans=0.125 2024-09-22 21:22:57,302 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=5.31 vs. limit=10.0 2024-09-22 21:23:12,835 INFO [train.py:1198] (2/4) Epoch 6, batch 3100, loss[loss=0.3123, ctc_loss=0.2254, cr_loss=0.4346, over 17143.00 frames. ], tot_loss[loss=0.2856, ctc_loss=0.2055, cr_loss=0.4008, over 3360977.47 frames. ], batch size: 48, lr: 1.91e-02, grad_scale: 32.0 2024-09-22 21:23:22,322 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=105373.33333333333, ans=0.2 2024-09-22 21:23:52,284 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=105466.66666666667, ans=0.0 2024-09-22 21:23:57,644 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=105466.66666666667, ans=0.0 2024-09-22 21:24:09,900 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.162e+02 1.436e+02 1.623e+02 1.870e+02 2.887e+02, threshold=3.246e+02, percent-clipped=0.0 2024-09-22 21:24:10,157 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=105513.33333333333, ans=0.125 2024-09-22 21:24:13,817 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.13 vs. limit=15.0 2024-09-22 21:24:19,546 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=105560.0, ans=0.0 2024-09-22 21:24:33,410 INFO [train.py:1198] (2/4) Epoch 6, batch 3150, loss[loss=0.2528, ctc_loss=0.1782, cr_loss=0.3733, over 17073.00 frames. ], tot_loss[loss=0.2854, ctc_loss=0.2054, cr_loss=0.4001, over 3350431.15 frames. ], batch size: 46, lr: 1.91e-02, grad_scale: 32.0 2024-09-22 21:24:33,957 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.59 vs. limit=15.0 2024-09-22 21:24:57,694 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=5.55 vs. limit=10.0 2024-09-22 21:25:20,223 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=105746.66666666667, ans=0.0 2024-09-22 21:25:26,542 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 21:25:31,284 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=105746.66666666667, ans=0.125 2024-09-22 21:25:44,336 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=105793.33333333333, ans=0.125 2024-09-22 21:25:48,528 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.whiten, num_groups=1, num_channels=192, metric=5.11 vs. limit=12.0 2024-09-22 21:25:53,518 INFO [train.py:1198] (2/4) Epoch 6, batch 3200, loss[loss=0.2566, ctc_loss=0.1807, cr_loss=0.3795, over 17089.00 frames. ], tot_loss[loss=0.2862, ctc_loss=0.2058, cr_loss=0.4015, over 3359317.65 frames. ], batch size: 40, lr: 1.91e-02, grad_scale: 32.0 2024-09-22 21:25:53,841 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=105840.0, ans=0.125 2024-09-22 21:26:23,377 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=105933.33333333333, ans=0.07 2024-09-22 21:26:25,054 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=105933.33333333333, ans=0.1 2024-09-22 21:26:41,177 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.02 vs. limit=15.0 2024-09-22 21:26:50,292 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.272e+02 1.472e+02 1.715e+02 1.976e+02 3.064e+02, threshold=3.429e+02, percent-clipped=0.0 2024-09-22 21:27:01,384 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=106026.66666666667, ans=0.125 2024-09-22 21:27:04,550 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=106026.66666666667, ans=0.125 2024-09-22 21:27:13,660 INFO [train.py:1198] (2/4) Epoch 6, batch 3250, loss[loss=0.2773, ctc_loss=0.1953, cr_loss=0.4102, over 17170.00 frames. ], tot_loss[loss=0.2866, ctc_loss=0.2063, cr_loss=0.4014, over 3344284.27 frames. ], batch size: 45, lr: 1.91e-02, grad_scale: 32.0 2024-09-22 21:27:23,367 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=106073.33333333333, ans=0.04949747468305833 2024-09-22 21:28:07,494 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.31 vs. limit=15.0 2024-09-22 21:28:31,720 INFO [train.py:1198] (2/4) Epoch 6, batch 3300, loss[loss=0.2489, ctc_loss=0.1754, cr_loss=0.3675, over 17030.00 frames. ], tot_loss[loss=0.2864, ctc_loss=0.2062, cr_loss=0.4011, over 3341788.74 frames. ], batch size: 39, lr: 1.91e-02, grad_scale: 64.0 2024-09-22 21:28:31,805 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=106306.66666666667, ans=0.125 2024-09-22 21:28:44,574 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.max_abs, batch_count=106306.66666666667, ans=10.0 2024-09-22 21:29:04,829 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=106400.0, ans=0.0 2024-09-22 21:29:04,927 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=106400.0, ans=0.125 2024-09-22 21:29:20,401 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 21:29:26,377 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.242e+02 1.523e+02 1.757e+02 2.023e+02 3.259e+02, threshold=3.514e+02, percent-clipped=0.0 2024-09-22 21:29:44,659 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=9.95 vs. limit=15.0 2024-09-22 21:29:49,747 INFO [train.py:1198] (2/4) Epoch 6, batch 3350, loss[loss=0.2889, ctc_loss=0.2077, cr_loss=0.4061, over 17305.00 frames. ], tot_loss[loss=0.2848, ctc_loss=0.2049, cr_loss=0.3997, over 3343657.48 frames. ], batch size: 49, lr: 1.90e-02, grad_scale: 32.0 2024-09-22 21:30:18,643 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=106586.66666666667, ans=0.0 2024-09-22 21:30:49,878 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=106680.0, ans=0.05 2024-09-22 21:30:57,798 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 21:31:01,308 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=6.02 vs. limit=15.0 2024-09-22 21:31:08,299 INFO [train.py:1198] (2/4) Epoch 6, batch 3400, loss[loss=0.2762, ctc_loss=0.1945, cr_loss=0.4088, over 17029.00 frames. ], tot_loss[loss=0.2847, ctc_loss=0.2048, cr_loss=0.3994, over 3340286.58 frames. ], batch size: 44, lr: 1.90e-02, grad_scale: 32.0 2024-09-22 21:31:28,786 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=106820.0, ans=0.125 2024-09-22 21:32:04,335 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.185e+02 1.439e+02 1.572e+02 1.824e+02 3.611e+02, threshold=3.144e+02, percent-clipped=1.0 2024-09-22 21:32:26,356 INFO [train.py:1198] (2/4) Epoch 6, batch 3450, loss[loss=0.2338, ctc_loss=0.1675, cr_loss=0.3313, over 17121.00 frames. ], tot_loss[loss=0.2844, ctc_loss=0.2046, cr_loss=0.3991, over 3334650.06 frames. ], batch size: 40, lr: 1.90e-02, grad_scale: 32.0 2024-09-22 21:32:37,469 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=107006.66666666667, ans=0.125 2024-09-22 21:32:54,447 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=107053.33333333333, ans=0.125 2024-09-22 21:33:17,535 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=15.47 vs. limit=22.5 2024-09-22 21:33:26,326 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=107146.66666666667, ans=0.2 2024-09-22 21:33:34,027 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=107193.33333333333, ans=0.2 2024-09-22 21:33:46,060 INFO [train.py:1198] (2/4) Epoch 6, batch 3500, loss[loss=0.2948, ctc_loss=0.2103, cr_loss=0.4226, over 17017.00 frames. ], tot_loss[loss=0.2854, ctc_loss=0.2052, cr_loss=0.4011, over 3341576.32 frames. ], batch size: 44, lr: 1.90e-02, grad_scale: 32.0 2024-09-22 21:33:47,787 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=107240.0, ans=0.125 2024-09-22 21:34:16,419 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=107286.66666666667, ans=0.125 2024-09-22 21:34:44,453 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.243e+02 1.542e+02 1.676e+02 1.907e+02 3.181e+02, threshold=3.352e+02, percent-clipped=1.0 2024-09-22 21:34:50,966 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=107426.66666666667, ans=0.07 2024-09-22 21:34:53,057 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.84 vs. limit=6.0 2024-09-22 21:35:06,138 INFO [train.py:1198] (2/4) Epoch 6, batch 3550, loss[loss=0.3358, ctc_loss=0.2436, cr_loss=0.4607, over 16980.00 frames. ], tot_loss[loss=0.2859, ctc_loss=0.2057, cr_loss=0.4012, over 3347546.32 frames. ], batch size: 56, lr: 1.90e-02, grad_scale: 32.0 2024-09-22 21:35:07,973 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=107473.33333333333, ans=0.0 2024-09-22 21:35:20,318 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.min_positive, batch_count=107520.0, ans=0.05 2024-09-22 21:35:29,039 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.24 vs. limit=15.0 2024-09-22 21:35:49,171 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=3.73 vs. limit=15.0 2024-09-22 21:36:28,359 INFO [train.py:1198] (2/4) Epoch 6, batch 3600, loss[loss=0.315, ctc_loss=0.2273, cr_loss=0.4384, over 17314.00 frames. ], tot_loss[loss=0.286, ctc_loss=0.2058, cr_loss=0.4011, over 3344081.23 frames. ], batch size: 51, lr: 1.89e-02, grad_scale: 32.0 2024-09-22 21:36:28,741 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 21:36:56,820 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=107753.33333333333, ans=0.125 2024-09-22 21:37:12,439 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=107800.0, ans=0.0 2024-09-22 21:37:15,950 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=8.79 vs. limit=15.0 2024-09-22 21:37:24,563 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.276e+02 1.473e+02 1.668e+02 1.932e+02 3.410e+02, threshold=3.336e+02, percent-clipped=1.0 2024-09-22 21:37:46,253 INFO [train.py:1198] (2/4) Epoch 6, batch 3650, loss[loss=0.2978, ctc_loss=0.2075, cr_loss=0.4515, over 17348.00 frames. ], tot_loss[loss=0.285, ctc_loss=0.2049, cr_loss=0.4005, over 3351519.53 frames. ], batch size: 48, lr: 1.89e-02, grad_scale: 32.0 2024-09-22 21:37:49,618 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer_ff2.min_abs, batch_count=107940.0, ans=0.1 2024-09-22 21:37:57,512 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=107940.0, ans=0.125 2024-09-22 21:38:12,956 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=107986.66666666667, ans=0.05 2024-09-22 21:38:23,948 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=108033.33333333333, ans=0.025 2024-09-22 21:38:39,843 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=20.80 vs. limit=22.5 2024-09-22 21:39:03,868 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.63 vs. limit=15.0 2024-09-22 21:39:04,805 INFO [train.py:1198] (2/4) Epoch 6, batch 3700, loss[loss=0.2661, ctc_loss=0.1863, cr_loss=0.3991, over 17183.00 frames. ], tot_loss[loss=0.285, ctc_loss=0.205, cr_loss=0.3999, over 3339204.00 frames. ], batch size: 41, lr: 1.89e-02, grad_scale: 32.0 2024-09-22 21:39:27,309 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=11.70 vs. limit=12.0 2024-09-22 21:39:33,458 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=108220.0, ans=0.0 2024-09-22 21:39:52,395 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=108313.33333333333, ans=0.125 2024-09-22 21:40:01,663 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.177e+02 1.455e+02 1.690e+02 2.004e+02 3.040e+02, threshold=3.380e+02, percent-clipped=0.0 2024-09-22 21:40:14,098 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=108360.0, ans=0.125 2024-09-22 21:40:15,618 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=108360.0, ans=0.2 2024-09-22 21:40:23,105 INFO [train.py:1198] (2/4) Epoch 6, batch 3750, loss[loss=0.241, ctc_loss=0.1712, cr_loss=0.3488, over 17240.00 frames. ], tot_loss[loss=0.284, ctc_loss=0.2042, cr_loss=0.3991, over 3337534.44 frames. ], batch size: 50, lr: 1.89e-02, grad_scale: 32.0 2024-09-22 21:40:31,330 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=108406.66666666667, ans=0.2 2024-09-22 21:40:58,003 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=108500.0, ans=0.0 2024-09-22 21:41:07,684 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=108500.0, ans=0.025 2024-09-22 21:41:17,448 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=7.12 vs. limit=15.0 2024-09-22 21:41:42,108 INFO [train.py:1198] (2/4) Epoch 6, batch 3800, loss[loss=0.2917, ctc_loss=0.2141, cr_loss=0.3881, over 16510.00 frames. ], tot_loss[loss=0.2852, ctc_loss=0.2053, cr_loss=0.3995, over 3315726.61 frames. ], batch size: 66, lr: 1.89e-02, grad_scale: 32.0 2024-09-22 21:41:42,338 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=108640.0, ans=0.1 2024-09-22 21:41:54,946 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=108640.0, ans=0.2 2024-09-22 21:42:07,532 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=108686.66666666667, ans=0.125 2024-09-22 21:42:39,193 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.201e+02 1.543e+02 1.830e+02 2.234e+02 3.927e+02, threshold=3.660e+02, percent-clipped=2.0 2024-09-22 21:42:59,724 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=108873.33333333333, ans=0.125 2024-09-22 21:43:00,878 INFO [train.py:1198] (2/4) Epoch 6, batch 3850, loss[loss=0.3519, ctc_loss=0.2629, cr_loss=0.4451, over 12209.00 frames. ], tot_loss[loss=0.2872, ctc_loss=0.2074, cr_loss=0.3991, over 3257727.63 frames. ], batch size: 123, lr: 1.89e-02, grad_scale: 32.0 2024-09-22 21:43:23,130 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=108920.0, ans=0.125 2024-09-22 21:43:23,641 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=21.25 vs. limit=22.5 2024-09-22 21:43:24,736 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=108920.0, ans=0.2 2024-09-22 21:43:44,166 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=108966.66666666667, ans=0.125 2024-09-22 21:43:50,234 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=109013.33333333333, ans=0.0 2024-09-22 21:45:03,035 INFO [train.py:1198] (2/4) Epoch 7, batch 0, loss[loss=0.2488, ctc_loss=0.1784, cr_loss=0.3523, over 17183.00 frames. ], tot_loss[loss=0.2488, ctc_loss=0.1784, cr_loss=0.3523, over 17183.00 frames. ], batch size: 41, lr: 1.77e-02, grad_scale: 32.0 2024-09-22 21:45:03,035 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-22 21:45:18,425 INFO [train.py:1230] (2/4) Epoch 7, validation: loss=0.06283, ctc_loss=0.06283, cr_loss=7.028e-15, over 944034.00 frames. 2024-09-22 21:45:18,426 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-22 21:45:34,529 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=109134.66666666667, ans=0.2 2024-09-22 21:46:05,384 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=109181.33333333333, ans=0.125 2024-09-22 21:46:15,443 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.87 vs. limit=15.0 2024-09-22 21:46:16,960 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=512, metric=4.14 vs. limit=15.0 2024-09-22 21:46:24,180 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.230e+02 1.560e+02 1.911e+02 2.609e+02 3.890e+02, threshold=3.822e+02, percent-clipped=2.0 2024-09-22 21:46:25,178 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=6.20 vs. limit=12.0 2024-09-22 21:46:37,673 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=10.92 vs. limit=15.0 2024-09-22 21:46:39,893 INFO [train.py:1198] (2/4) Epoch 7, batch 50, loss[loss=0.3236, ctc_loss=0.2344, cr_loss=0.4465, over 17141.00 frames. ], tot_loss[loss=0.2902, ctc_loss=0.2084, cr_loss=0.4087, over 756255.20 frames. ], batch size: 48, lr: 1.76e-02, grad_scale: 32.0 2024-09-22 21:46:40,533 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=15.07 vs. limit=22.5 2024-09-22 21:47:09,922 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=109368.0, ans=0.125 2024-09-22 21:47:55,856 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=109508.0, ans=0.2 2024-09-22 21:48:04,578 INFO [train.py:1198] (2/4) Epoch 7, batch 100, loss[loss=0.2812, ctc_loss=0.2005, cr_loss=0.4034, over 17354.00 frames. ], tot_loss[loss=0.2843, ctc_loss=0.204, cr_loss=0.4017, over 1332073.24 frames. ], batch size: 48, lr: 1.76e-02, grad_scale: 32.0 2024-09-22 21:48:08,585 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.62 vs. limit=6.0 2024-09-22 21:48:12,766 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=109554.66666666667, ans=0.0 2024-09-22 21:48:26,882 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=109601.33333333333, ans=0.025 2024-09-22 21:48:49,019 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=109648.0, ans=0.125 2024-09-22 21:49:07,937 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.339e+02 1.498e+02 1.759e+02 2.443e+02, threshold=2.996e+02, percent-clipped=0.0 2024-09-22 21:49:26,880 INFO [train.py:1198] (2/4) Epoch 7, batch 150, loss[loss=0.2729, ctc_loss=0.1993, cr_loss=0.3682, over 16880.00 frames. ], tot_loss[loss=0.2835, ctc_loss=0.2032, cr_loss=0.4017, over 1775237.18 frames. ], batch size: 58, lr: 1.76e-02, grad_scale: 32.0 2024-09-22 21:49:27,579 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=4.65 vs. limit=12.0 2024-09-22 21:49:52,557 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=109834.66666666667, ans=0.1 2024-09-22 21:49:57,303 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=109881.33333333333, ans=0.125 2024-09-22 21:49:57,336 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=109881.33333333333, ans=0.025 2024-09-22 21:50:29,413 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=109974.66666666667, ans=0.2 2024-09-22 21:50:30,117 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.32 vs. limit=15.0 2024-09-22 21:50:45,086 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=109974.66666666667, ans=0.125 2024-09-22 21:50:48,258 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=110021.33333333333, ans=0.0 2024-09-22 21:50:49,451 INFO [train.py:1198] (2/4) Epoch 7, batch 200, loss[loss=0.2951, ctc_loss=0.2115, cr_loss=0.4178, over 17037.00 frames. ], tot_loss[loss=0.2816, ctc_loss=0.2017, cr_loss=0.3995, over 2124397.39 frames. ], batch size: 53, lr: 1.76e-02, grad_scale: 32.0 2024-09-22 21:50:50,124 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=8.72 vs. limit=15.0 2024-09-22 21:51:28,516 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=110114.66666666667, ans=0.2 2024-09-22 21:51:44,308 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=110161.33333333333, ans=0.125 2024-09-22 21:51:48,124 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=7.52 vs. limit=15.0 2024-09-22 21:51:53,494 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.194e+02 1.411e+02 1.617e+02 1.825e+02 4.000e+02, threshold=3.234e+02, percent-clipped=2.0 2024-09-22 21:51:53,780 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 21:51:53,806 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=110208.0, ans=0.125 2024-09-22 21:51:53,920 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=110208.0, ans=0.125 2024-09-22 21:52:12,070 INFO [train.py:1198] (2/4) Epoch 7, batch 250, loss[loss=0.2983, ctc_loss=0.2123, cr_loss=0.43, over 17011.00 frames. ], tot_loss[loss=0.2799, ctc_loss=0.2003, cr_loss=0.3978, over 2398520.95 frames. ], batch size: 56, lr: 1.76e-02, grad_scale: 32.0 2024-09-22 21:52:39,536 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=110301.33333333333, ans=0.125 2024-09-22 21:52:49,671 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=11.62 vs. limit=15.0 2024-09-22 21:53:16,003 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=110394.66666666667, ans=0.125 2024-09-22 21:53:30,426 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=110441.33333333333, ans=0.125 2024-09-22 21:53:34,852 INFO [train.py:1198] (2/4) Epoch 7, batch 300, loss[loss=0.2578, ctc_loss=0.1825, cr_loss=0.3764, over 17123.00 frames. ], tot_loss[loss=0.2785, ctc_loss=0.1991, cr_loss=0.3967, over 2606789.71 frames. ], batch size: 40, lr: 1.76e-02, grad_scale: 32.0 2024-09-22 21:54:19,151 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=110581.33333333333, ans=0.0 2024-09-22 21:54:41,018 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.225e+02 1.446e+02 1.620e+02 1.814e+02 2.683e+02, threshold=3.241e+02, percent-clipped=0.0 2024-09-22 21:54:51,021 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=110674.66666666667, ans=0.125 2024-09-22 21:54:56,992 INFO [train.py:1198] (2/4) Epoch 7, batch 350, loss[loss=0.3744, ctc_loss=0.2823, cr_loss=0.4607, over 11966.00 frames. ], tot_loss[loss=0.2785, ctc_loss=0.1991, cr_loss=0.3971, over 2776310.86 frames. ], batch size: 123, lr: 1.75e-02, grad_scale: 32.0 2024-09-22 21:55:02,106 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.min_positive, batch_count=110721.33333333333, ans=0.05 2024-09-22 21:55:14,734 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=110768.0, ans=0.1 2024-09-22 21:55:16,183 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=110768.0, ans=0.0 2024-09-22 21:55:22,617 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=110768.0, ans=0.1 2024-09-22 21:55:24,270 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=110768.0, ans=0.0 2024-09-22 21:55:47,791 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=110861.33333333333, ans=0.125 2024-09-22 21:56:11,581 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=110908.0, ans=0.2 2024-09-22 21:56:14,516 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=110908.0, ans=0.125 2024-09-22 21:56:14,521 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=110908.0, ans=0.025 2024-09-22 21:56:19,158 INFO [train.py:1198] (2/4) Epoch 7, batch 400, loss[loss=0.299, ctc_loss=0.2158, cr_loss=0.4162, over 16485.00 frames. ], tot_loss[loss=0.2774, ctc_loss=0.1982, cr_loss=0.3957, over 2907831.57 frames. ], batch size: 66, lr: 1.75e-02, grad_scale: 32.0 2024-09-22 21:56:43,708 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=512, metric=18.73 vs. limit=22.5 2024-09-22 21:56:53,036 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=111048.0, ans=0.0 2024-09-22 21:57:02,041 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=111048.0, ans=0.2 2024-09-22 21:57:13,202 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=111094.66666666667, ans=0.125 2024-09-22 21:57:25,761 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.408e+02 1.555e+02 1.806e+02 2.695e+02, threshold=3.109e+02, percent-clipped=0.0 2024-09-22 21:57:41,728 INFO [train.py:1198] (2/4) Epoch 7, batch 450, loss[loss=0.2946, ctc_loss=0.2113, cr_loss=0.4164, over 17028.00 frames. ], tot_loss[loss=0.2793, ctc_loss=0.2, cr_loss=0.3963, over 2991531.35 frames. ], batch size: 56, lr: 1.75e-02, grad_scale: 32.0 2024-09-22 21:57:47,964 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=111188.0, ans=0.1 2024-09-22 21:57:48,029 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=111188.0, ans=0.125 2024-09-22 21:57:54,174 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.min_abs, batch_count=111188.0, ans=0.5 2024-09-22 21:58:01,999 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=111234.66666666667, ans=0.0 2024-09-22 21:58:18,233 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=111281.33333333333, ans=0.125 2024-09-22 21:58:21,449 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=111281.33333333333, ans=0.0 2024-09-22 21:58:44,904 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=111328.0, ans=0.125 2024-09-22 21:58:57,923 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=17.65 vs. limit=22.5 2024-09-22 21:58:59,227 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=111374.66666666667, ans=0.125 2024-09-22 21:59:03,788 INFO [train.py:1198] (2/4) Epoch 7, batch 500, loss[loss=0.2365, ctc_loss=0.1652, cr_loss=0.3565, over 16772.00 frames. ], tot_loss[loss=0.2773, ctc_loss=0.1984, cr_loss=0.3946, over 3063068.65 frames. ], batch size: 37, lr: 1.75e-02, grad_scale: 32.0 2024-09-22 21:59:05,719 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=111421.33333333333, ans=0.125 2024-09-22 21:59:08,866 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=111421.33333333333, ans=0.125 2024-09-22 21:59:14,612 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=111421.33333333333, ans=0.125 2024-09-22 22:00:09,622 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.144e+02 1.405e+02 1.626e+02 1.844e+02 3.754e+02, threshold=3.253e+02, percent-clipped=1.0 2024-09-22 22:00:10,352 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=7.64 vs. limit=15.0 2024-09-22 22:00:25,404 INFO [train.py:1198] (2/4) Epoch 7, batch 550, loss[loss=0.3692, ctc_loss=0.2866, cr_loss=0.4133, over 11332.00 frames. ], tot_loss[loss=0.2781, ctc_loss=0.1989, cr_loss=0.3958, over 3129758.50 frames. ], batch size: 124, lr: 1.75e-02, grad_scale: 32.0 2024-09-22 22:00:33,362 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=111654.66666666667, ans=0.125 2024-09-22 22:00:55,499 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=111701.33333333333, ans=0.125 2024-09-22 22:01:22,003 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=19.96 vs. limit=22.5 2024-09-22 22:01:46,995 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.40 vs. limit=6.0 2024-09-22 22:01:47,901 INFO [train.py:1198] (2/4) Epoch 7, batch 600, loss[loss=0.3368, ctc_loss=0.2522, cr_loss=0.4228, over 11877.00 frames. ], tot_loss[loss=0.2775, ctc_loss=0.1984, cr_loss=0.3957, over 3181506.38 frames. ], batch size: 124, lr: 1.75e-02, grad_scale: 32.0 2024-09-22 22:01:49,880 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=111888.0, ans=0.125 2024-09-22 22:02:22,458 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=111981.33333333333, ans=0.0 2024-09-22 22:02:39,378 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=112028.0, ans=0.0 2024-09-22 22:02:54,335 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=112028.0, ans=0.0 2024-09-22 22:02:58,792 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.180e+02 1.416e+02 1.598e+02 2.018e+02 3.504e+02, threshold=3.196e+02, percent-clipped=2.0 2024-09-22 22:03:00,682 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=112074.66666666667, ans=0.1 2024-09-22 22:03:14,671 INFO [train.py:1198] (2/4) Epoch 7, batch 650, loss[loss=0.2594, ctc_loss=0.1835, cr_loss=0.3798, over 17106.00 frames. ], tot_loss[loss=0.2779, ctc_loss=0.1987, cr_loss=0.3961, over 3214835.02 frames. ], batch size: 43, lr: 1.74e-02, grad_scale: 32.0 2024-09-22 22:03:45,144 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=112214.66666666667, ans=0.1 2024-09-22 22:03:47,107 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.68 vs. limit=15.0 2024-09-22 22:03:52,978 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=112214.66666666667, ans=0.05 2024-09-22 22:04:36,942 INFO [train.py:1198] (2/4) Epoch 7, batch 700, loss[loss=0.2306, ctc_loss=0.1607, cr_loss=0.3495, over 17190.00 frames. ], tot_loss[loss=0.2766, ctc_loss=0.1977, cr_loss=0.3946, over 3244567.74 frames. ], batch size: 41, lr: 1.74e-02, grad_scale: 32.0 2024-09-22 22:04:40,287 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=112354.66666666667, ans=0.125 2024-09-22 22:04:43,482 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=112354.66666666667, ans=0.125 2024-09-22 22:04:45,174 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=112354.66666666667, ans=0.125 2024-09-22 22:04:50,072 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=112354.66666666667, ans=0.125 2024-09-22 22:05:13,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=112448.0, ans=0.2 2024-09-22 22:05:42,318 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.167e+02 1.430e+02 1.586e+02 1.856e+02 3.477e+02, threshold=3.173e+02, percent-clipped=1.0 2024-09-22 22:05:44,334 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=112541.33333333333, ans=0.125 2024-09-22 22:05:53,602 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=112541.33333333333, ans=0.1 2024-09-22 22:05:55,260 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=112541.33333333333, ans=0.1 2024-09-22 22:05:56,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=112588.0, ans=0.95 2024-09-22 22:05:58,155 INFO [train.py:1198] (2/4) Epoch 7, batch 750, loss[loss=0.25, ctc_loss=0.1791, cr_loss=0.3547, over 17072.00 frames. ], tot_loss[loss=0.2762, ctc_loss=0.1972, cr_loss=0.3952, over 3277858.18 frames. ], batch size: 43, lr: 1.74e-02, grad_scale: 32.0 2024-09-22 22:06:09,448 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=112588.0, ans=0.2 2024-09-22 22:06:12,705 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 22:06:29,965 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=112681.33333333333, ans=0.1 2024-09-22 22:06:53,149 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=112728.0, ans=0.04949747468305833 2024-09-22 22:07:08,933 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=112774.66666666667, ans=0.125 2024-09-22 22:07:19,626 INFO [train.py:1198] (2/4) Epoch 7, batch 800, loss[loss=0.2695, ctc_loss=0.1953, cr_loss=0.3707, over 17077.00 frames. ], tot_loss[loss=0.2763, ctc_loss=0.1972, cr_loss=0.3955, over 3300895.83 frames. ], batch size: 46, lr: 1.74e-02, grad_scale: 32.0 2024-09-22 22:07:47,901 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=112868.0, ans=0.0 2024-09-22 22:08:07,823 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.26 vs. limit=10.0 2024-09-22 22:08:12,143 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=112961.33333333333, ans=0.0 2024-09-22 22:08:26,195 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.115e+02 1.385e+02 1.523e+02 1.724e+02 2.705e+02, threshold=3.046e+02, percent-clipped=0.0 2024-09-22 22:08:33,535 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.76 vs. limit=15.0 2024-09-22 22:08:41,931 INFO [train.py:1198] (2/4) Epoch 7, batch 850, loss[loss=0.2791, ctc_loss=0.1988, cr_loss=0.4015, over 17314.00 frames. ], tot_loss[loss=0.2764, ctc_loss=0.1973, cr_loss=0.3954, over 3315345.76 frames. ], batch size: 49, lr: 1.74e-02, grad_scale: 32.0 2024-09-22 22:08:46,795 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=113054.66666666667, ans=0.0 2024-09-22 22:08:51,787 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=113054.66666666667, ans=0.125 2024-09-22 22:09:27,321 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=113148.0, ans=0.1 2024-09-22 22:09:35,436 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=113194.66666666667, ans=0.0 2024-09-22 22:09:37,214 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=113194.66666666667, ans=0.125 2024-09-22 22:09:56,165 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=113241.33333333333, ans=0.1 2024-09-22 22:10:03,662 INFO [train.py:1198] (2/4) Epoch 7, batch 900, loss[loss=0.2896, ctc_loss=0.2057, cr_loss=0.4191, over 17224.00 frames. ], tot_loss[loss=0.2781, ctc_loss=0.1987, cr_loss=0.397, over 3327449.53 frames. ], batch size: 50, lr: 1.74e-02, grad_scale: 32.0 2024-09-22 22:10:14,469 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=113288.0, ans=0.125 2024-09-22 22:10:19,122 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=113288.0, ans=0.2 2024-09-22 22:10:20,824 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=113334.66666666667, ans=0.125 2024-09-22 22:10:35,138 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=113334.66666666667, ans=0.125 2024-09-22 22:10:47,002 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.57 vs. limit=15.0 2024-09-22 22:11:03,443 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=113428.0, ans=0.0 2024-09-22 22:11:04,743 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=113428.0, ans=0.125 2024-09-22 22:11:04,884 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=113428.0, ans=0.025 2024-09-22 22:11:09,265 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.239e+02 1.440e+02 1.570e+02 1.832e+02 2.236e+02, threshold=3.140e+02, percent-clipped=0.0 2024-09-22 22:11:25,181 INFO [train.py:1198] (2/4) Epoch 7, batch 950, loss[loss=0.2357, ctc_loss=0.1634, cr_loss=0.3616, over 16957.00 frames. ], tot_loss[loss=0.2763, ctc_loss=0.1972, cr_loss=0.3956, over 3335404.17 frames. ], batch size: 42, lr: 1.73e-02, grad_scale: 32.0 2024-09-22 22:12:12,751 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=113614.66666666667, ans=0.125 2024-09-22 22:12:14,206 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=113661.33333333333, ans=0.025 2024-09-22 22:12:26,386 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer_ff3.min_abs, batch_count=113661.33333333333, ans=0.2 2024-09-22 22:12:37,779 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=113708.0, ans=0.125 2024-09-22 22:12:45,774 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=113708.0, ans=0.1 2024-09-22 22:12:46,157 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=8.80 vs. limit=15.0 2024-09-22 22:12:50,311 INFO [train.py:1198] (2/4) Epoch 7, batch 1000, loss[loss=0.2717, ctc_loss=0.1959, cr_loss=0.3789, over 17217.00 frames. ], tot_loss[loss=0.2756, ctc_loss=0.1966, cr_loss=0.3952, over 3348226.51 frames. ], batch size: 47, lr: 1.73e-02, grad_scale: 32.0 2024-09-22 22:12:50,646 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=113754.66666666667, ans=0.125 2024-09-22 22:13:03,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=113754.66666666667, ans=0.0 2024-09-22 22:13:17,424 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=113801.33333333333, ans=0.025 2024-09-22 22:13:29,168 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=13.01 vs. limit=15.0 2024-09-22 22:13:53,767 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.387e+02 1.549e+02 1.823e+02 4.640e+02, threshold=3.099e+02, percent-clipped=1.0 2024-09-22 22:14:06,169 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=113941.33333333333, ans=0.2 2024-09-22 22:14:12,273 INFO [train.py:1198] (2/4) Epoch 7, batch 1050, loss[loss=0.2753, ctc_loss=0.1976, cr_loss=0.3883, over 17266.00 frames. ], tot_loss[loss=0.2753, ctc_loss=0.1963, cr_loss=0.3949, over 3349775.85 frames. ], batch size: 44, lr: 1.73e-02, grad_scale: 32.0 2024-09-22 22:14:32,090 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.41 vs. limit=6.0 2024-09-22 22:15:14,547 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=5.18 vs. limit=15.0 2024-09-22 22:15:20,688 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=114174.66666666667, ans=0.125 2024-09-22 22:15:34,438 INFO [train.py:1198] (2/4) Epoch 7, batch 1100, loss[loss=0.2383, ctc_loss=0.1673, cr_loss=0.3551, over 17034.00 frames. ], tot_loss[loss=0.2758, ctc_loss=0.1966, cr_loss=0.3958, over 3349051.72 frames. ], batch size: 44, lr: 1.73e-02, grad_scale: 32.0 2024-09-22 22:15:39,587 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=114221.33333333333, ans=0.125 2024-09-22 22:15:41,091 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=114221.33333333333, ans=0.04949747468305833 2024-09-22 22:15:58,681 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=114268.0, ans=0.125 2024-09-22 22:16:20,625 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=114361.33333333333, ans=0.1 2024-09-22 22:16:37,933 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.126e+02 1.457e+02 1.783e+02 2.141e+02 3.294e+02, threshold=3.566e+02, percent-clipped=3.0 2024-09-22 22:16:45,525 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=114408.0, ans=0.125 2024-09-22 22:16:48,659 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=114408.0, ans=0.0 2024-09-22 22:16:56,274 INFO [train.py:1198] (2/4) Epoch 7, batch 1150, loss[loss=0.3128, ctc_loss=0.227, cr_loss=0.4286, over 14910.00 frames. ], tot_loss[loss=0.2763, ctc_loss=0.197, cr_loss=0.3963, over 3346012.62 frames. ], batch size: 88, lr: 1.73e-02, grad_scale: 32.0 2024-09-22 22:16:56,534 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=114454.66666666667, ans=0.0 2024-09-22 22:17:15,368 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=114501.33333333333, ans=0.1 2024-09-22 22:17:29,218 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=114548.0, ans=0.1 2024-09-22 22:17:52,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=114594.66666666667, ans=0.1 2024-09-22 22:18:04,877 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=5.91 vs. limit=10.0 2024-09-22 22:18:17,891 INFO [train.py:1198] (2/4) Epoch 7, batch 1200, loss[loss=0.2501, ctc_loss=0.1764, cr_loss=0.3682, over 17015.00 frames. ], tot_loss[loss=0.2759, ctc_loss=0.1967, cr_loss=0.3958, over 3356547.04 frames. ], batch size: 56, lr: 1.73e-02, grad_scale: 32.0 2024-09-22 22:18:42,273 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=114734.66666666667, ans=0.125 2024-09-22 22:19:04,005 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=114781.33333333333, ans=0.125 2024-09-22 22:19:24,536 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.166e+02 1.404e+02 1.583e+02 1.854e+02 5.575e+02, threshold=3.166e+02, percent-clipped=2.0 2024-09-22 22:19:40,342 INFO [train.py:1198] (2/4) Epoch 7, batch 1250, loss[loss=0.303, ctc_loss=0.2118, cr_loss=0.456, over 17019.00 frames. ], tot_loss[loss=0.2765, ctc_loss=0.1971, cr_loss=0.3973, over 3361583.93 frames. ], batch size: 51, lr: 1.72e-02, grad_scale: 32.0 2024-09-22 22:20:10,263 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=114968.0, ans=0.1 2024-09-22 22:20:38,435 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 22:20:38,448 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=115061.33333333333, ans=0.125 2024-09-22 22:20:42,981 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=115061.33333333333, ans=0.2 2024-09-22 22:20:46,734 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.18 vs. limit=15.0 2024-09-22 22:20:49,254 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=115108.0, ans=0.125 2024-09-22 22:21:00,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=115154.66666666667, ans=0.125 2024-09-22 22:21:01,583 INFO [train.py:1198] (2/4) Epoch 7, batch 1300, loss[loss=0.2479, ctc_loss=0.1797, cr_loss=0.3414, over 17093.00 frames. ], tot_loss[loss=0.2774, ctc_loss=0.1978, cr_loss=0.3979, over 3352098.84 frames. ], batch size: 43, lr: 1.72e-02, grad_scale: 16.0 2024-09-22 22:21:18,989 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=115201.33333333333, ans=0.125 2024-09-22 22:21:25,329 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=115201.33333333333, ans=0.035 2024-09-22 22:21:30,044 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=115201.33333333333, ans=0.125 2024-09-22 22:21:30,047 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=115201.33333333333, ans=0.0 2024-09-22 22:21:50,378 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=115294.66666666667, ans=0.05 2024-09-22 22:22:04,035 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=13.02 vs. limit=22.5 2024-09-22 22:22:09,481 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.087e+02 1.485e+02 1.678e+02 2.013e+02 3.139e+02, threshold=3.356e+02, percent-clipped=0.0 2024-09-22 22:22:19,254 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module1.whiten, num_groups=1, num_channels=512, metric=6.37 vs. limit=15.0 2024-09-22 22:22:20,289 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=115341.33333333333, ans=0.125 2024-09-22 22:22:26,393 INFO [train.py:1198] (2/4) Epoch 7, batch 1350, loss[loss=0.2922, ctc_loss=0.2096, cr_loss=0.4131, over 16914.00 frames. ], tot_loss[loss=0.2773, ctc_loss=0.1977, cr_loss=0.3978, over 3351727.11 frames. ], batch size: 58, lr: 1.72e-02, grad_scale: 16.0 2024-09-22 22:22:37,146 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.26 vs. limit=15.0 2024-09-22 22:22:42,793 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=115434.66666666667, ans=0.0 2024-09-22 22:22:50,564 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=115434.66666666667, ans=0.1 2024-09-22 22:22:55,480 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=115434.66666666667, ans=0.0 2024-09-22 22:23:33,641 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=115574.66666666667, ans=0.0 2024-09-22 22:23:38,586 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 22:23:46,222 INFO [train.py:1198] (2/4) Epoch 7, batch 1400, loss[loss=0.2616, ctc_loss=0.1853, cr_loss=0.3812, over 17222.00 frames. ], tot_loss[loss=0.2768, ctc_loss=0.1973, cr_loss=0.3974, over 3349058.21 frames. ], batch size: 47, lr: 1.72e-02, grad_scale: 16.0 2024-09-22 22:23:58,378 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=115621.33333333333, ans=0.125 2024-09-22 22:24:17,467 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=115668.0, ans=0.125 2024-09-22 22:24:54,139 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.177e+02 1.408e+02 1.629e+02 2.097e+02 4.051e+02, threshold=3.259e+02, percent-clipped=2.0 2024-09-22 22:24:59,933 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=115808.0, ans=0.125 2024-09-22 22:25:08,569 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten.whitening_limit, batch_count=115808.0, ans=22.5 2024-09-22 22:25:10,929 INFO [train.py:1198] (2/4) Epoch 7, batch 1450, loss[loss=0.2762, ctc_loss=0.1974, cr_loss=0.3939, over 17360.00 frames. ], tot_loss[loss=0.2766, ctc_loss=0.1973, cr_loss=0.3969, over 3351160.47 frames. ], batch size: 48, lr: 1.72e-02, grad_scale: 16.0 2024-09-22 22:25:47,367 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=115948.0, ans=0.2 2024-09-22 22:25:48,786 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=115948.0, ans=0.0 2024-09-22 22:26:04,733 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=115994.66666666667, ans=0.125 2024-09-22 22:26:32,314 INFO [train.py:1198] (2/4) Epoch 7, batch 1500, loss[loss=0.2736, ctc_loss=0.193, cr_loss=0.4034, over 17032.00 frames. ], tot_loss[loss=0.2776, ctc_loss=0.1979, cr_loss=0.3982, over 3346684.93 frames. ], batch size: 56, lr: 1.72e-02, grad_scale: 16.0 2024-09-22 22:26:37,333 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=116088.0, ans=0.1 2024-09-22 22:27:14,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=116181.33333333333, ans=0.1 2024-09-22 22:27:16,597 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=116181.33333333333, ans=0.1 2024-09-22 22:27:18,184 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=116181.33333333333, ans=0.09899494936611666 2024-09-22 22:27:27,882 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=116228.0, ans=0.125 2024-09-22 22:27:40,531 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.248e+02 1.482e+02 1.664e+02 1.936e+02 3.285e+02, threshold=3.328e+02, percent-clipped=1.0 2024-09-22 22:27:54,707 INFO [train.py:1198] (2/4) Epoch 7, batch 1550, loss[loss=0.2848, ctc_loss=0.2055, cr_loss=0.3962, over 17240.00 frames. ], tot_loss[loss=0.2775, ctc_loss=0.1979, cr_loss=0.3978, over 3349251.86 frames. ], batch size: 55, lr: 1.71e-02, grad_scale: 16.0 2024-09-22 22:27:59,844 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=116321.33333333333, ans=0.0 2024-09-22 22:28:15,887 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=116368.0, ans=0.1 2024-09-22 22:28:56,321 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=116461.33333333333, ans=0.0 2024-09-22 22:29:04,603 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.96 vs. limit=15.0 2024-09-22 22:29:16,545 INFO [train.py:1198] (2/4) Epoch 7, batch 1600, loss[loss=0.2473, ctc_loss=0.1755, cr_loss=0.3586, over 17169.00 frames. ], tot_loss[loss=0.2767, ctc_loss=0.1974, cr_loss=0.3967, over 3346561.08 frames. ], batch size: 41, lr: 1.71e-02, grad_scale: 32.0 2024-09-22 22:29:38,057 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.71 vs. limit=6.0 2024-09-22 22:30:03,949 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=116648.0, ans=0.125 2024-09-22 22:30:24,104 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.169e+02 1.410e+02 1.579e+02 1.948e+02 3.056e+02, threshold=3.158e+02, percent-clipped=0.0 2024-09-22 22:30:38,421 INFO [train.py:1198] (2/4) Epoch 7, batch 1650, loss[loss=0.2781, ctc_loss=0.199, cr_loss=0.3954, over 17223.00 frames. ], tot_loss[loss=0.2772, ctc_loss=0.1978, cr_loss=0.3967, over 3351635.00 frames. ], batch size: 55, lr: 1.71e-02, grad_scale: 32.0 2024-09-22 22:30:44,099 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=6.82 vs. limit=15.0 2024-09-22 22:30:49,750 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=116788.0, ans=0.2 2024-09-22 22:31:00,750 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=116834.66666666667, ans=0.0 2024-09-22 22:31:41,068 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=116928.0, ans=0.1 2024-09-22 22:31:44,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=116974.66666666667, ans=0.125 2024-09-22 22:31:53,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=116974.66666666667, ans=0.125 2024-09-22 22:31:59,727 INFO [train.py:1198] (2/4) Epoch 7, batch 1700, loss[loss=0.3109, ctc_loss=0.2201, cr_loss=0.4541, over 16526.00 frames. ], tot_loss[loss=0.2757, ctc_loss=0.1966, cr_loss=0.3955, over 3367317.75 frames. ], batch size: 66, lr: 1.71e-02, grad_scale: 32.0 2024-09-22 22:32:15,346 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=117021.33333333333, ans=0.025 2024-09-22 22:32:28,356 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.85 vs. limit=15.0 2024-09-22 22:32:30,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=117068.0, ans=0.2 2024-09-22 22:32:34,142 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=117114.66666666667, ans=0.125 2024-09-22 22:32:38,938 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=117114.66666666667, ans=0.125 2024-09-22 22:33:07,038 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.106e+02 1.332e+02 1.436e+02 1.651e+02 2.321e+02, threshold=2.871e+02, percent-clipped=0.0 2024-09-22 22:33:12,137 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=117208.0, ans=0.125 2024-09-22 22:33:21,093 INFO [train.py:1198] (2/4) Epoch 7, batch 1750, loss[loss=0.3173, ctc_loss=0.229, cr_loss=0.4413, over 16626.00 frames. ], tot_loss[loss=0.2748, ctc_loss=0.196, cr_loss=0.3939, over 3356698.77 frames. ], batch size: 66, lr: 1.71e-02, grad_scale: 32.0 2024-09-22 22:33:27,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=117254.66666666667, ans=0.125 2024-09-22 22:33:29,564 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=117254.66666666667, ans=0.1 2024-09-22 22:34:05,215 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=117348.0, ans=0.0 2024-09-22 22:34:10,059 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=117394.66666666667, ans=0.0 2024-09-22 22:34:18,107 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=117394.66666666667, ans=0.0 2024-09-22 22:34:45,734 INFO [train.py:1198] (2/4) Epoch 7, batch 1800, loss[loss=0.3553, ctc_loss=0.2617, cr_loss=0.4677, over 15107.00 frames. ], tot_loss[loss=0.276, ctc_loss=0.197, cr_loss=0.395, over 3352065.10 frames. ], batch size: 89, lr: 1.71e-02, grad_scale: 32.0 2024-09-22 22:34:52,754 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=5.64 vs. limit=12.0 2024-09-22 22:35:11,136 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=117534.66666666667, ans=0.1 2024-09-22 22:35:19,298 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=117581.33333333333, ans=0.125 2024-09-22 22:35:20,746 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=117581.33333333333, ans=0.125 2024-09-22 22:35:25,617 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=117581.33333333333, ans=0.125 2024-09-22 22:35:43,296 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=117628.0, ans=0.2 2024-09-22 22:35:50,922 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.136e+02 1.391e+02 1.562e+02 1.960e+02 3.440e+02, threshold=3.125e+02, percent-clipped=2.0 2024-09-22 22:36:00,827 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=117674.66666666667, ans=0.125 2024-09-22 22:36:03,897 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=117721.33333333333, ans=0.0 2024-09-22 22:36:04,005 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=117721.33333333333, ans=0.0 2024-09-22 22:36:05,277 INFO [train.py:1198] (2/4) Epoch 7, batch 1850, loss[loss=0.2415, ctc_loss=0.1742, cr_loss=0.3366, over 17259.00 frames. ], tot_loss[loss=0.2774, ctc_loss=0.1981, cr_loss=0.3965, over 3351968.53 frames. ], batch size: 44, lr: 1.71e-02, grad_scale: 32.0 2024-09-22 22:36:19,165 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=117721.33333333333, ans=0.125 2024-09-22 22:36:25,852 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 22:36:27,745 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=12.32 vs. limit=15.0 2024-09-22 22:36:46,431 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=6.09 vs. limit=15.0 2024-09-22 22:36:51,118 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=117814.66666666667, ans=0.1 2024-09-22 22:37:04,840 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=117861.33333333333, ans=0.125 2024-09-22 22:37:13,251 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=7.22 vs. limit=10.0 2024-09-22 22:37:19,189 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=117908.0, ans=0.0 2024-09-22 22:37:29,067 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.57 vs. limit=6.0 2024-09-22 22:37:29,662 INFO [train.py:1198] (2/4) Epoch 7, batch 1900, loss[loss=0.2521, ctc_loss=0.1772, cr_loss=0.3746, over 17035.00 frames. ], tot_loss[loss=0.2774, ctc_loss=0.1983, cr_loss=0.3959, over 3347268.03 frames. ], batch size: 44, lr: 1.70e-02, grad_scale: 32.0 2024-09-22 22:38:07,666 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=118048.0, ans=0.125 2024-09-22 22:38:32,272 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=9.02 vs. limit=15.0 2024-09-22 22:38:37,122 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.224e+02 1.462e+02 1.627e+02 1.880e+02 2.641e+02, threshold=3.255e+02, percent-clipped=0.0 2024-09-22 22:38:51,416 INFO [train.py:1198] (2/4) Epoch 7, batch 1950, loss[loss=0.2325, ctc_loss=0.1585, cr_loss=0.37, over 17119.00 frames. ], tot_loss[loss=0.276, ctc_loss=0.1969, cr_loss=0.3952, over 3355771.04 frames. ], batch size: 40, lr: 1.70e-02, grad_scale: 32.0 2024-09-22 22:38:54,947 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=118188.0, ans=0.0 2024-09-22 22:38:57,246 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=8.01 vs. limit=10.0 2024-09-22 22:39:01,106 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=118188.0, ans=0.125 2024-09-22 22:39:04,267 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=118188.0, ans=0.125 2024-09-22 22:39:09,421 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=118234.66666666667, ans=0.125 2024-09-22 22:39:10,937 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.max_abs, batch_count=118234.66666666667, ans=10.0 2024-09-22 22:39:15,695 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=118234.66666666667, ans=0.1 2024-09-22 22:39:22,062 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=118281.33333333333, ans=0.1 2024-09-22 22:39:32,988 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=118281.33333333333, ans=0.0 2024-09-22 22:40:13,182 INFO [train.py:1198] (2/4) Epoch 7, batch 2000, loss[loss=0.2868, ctc_loss=0.2132, cr_loss=0.368, over 15937.00 frames. ], tot_loss[loss=0.2768, ctc_loss=0.1976, cr_loss=0.3962, over 3336838.91 frames. ], batch size: 74, lr: 1.70e-02, grad_scale: 32.0 2024-09-22 22:40:32,517 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=118468.0, ans=0.0 2024-09-22 22:40:53,731 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=118514.66666666667, ans=0.0 2024-09-22 22:40:59,934 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=118561.33333333333, ans=0.125 2024-09-22 22:41:04,237 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=118561.33333333333, ans=0.2 2024-09-22 22:41:21,329 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.217e+02 1.357e+02 1.496e+02 1.687e+02 2.654e+02, threshold=2.993e+02, percent-clipped=0.0 2024-09-22 22:41:23,245 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=118608.0, ans=0.025 2024-09-22 22:41:35,546 INFO [train.py:1198] (2/4) Epoch 7, batch 2050, loss[loss=0.3066, ctc_loss=0.2185, cr_loss=0.4402, over 16769.00 frames. ], tot_loss[loss=0.2773, ctc_loss=0.1978, cr_loss=0.3976, over 3343286.55 frames. ], batch size: 61, lr: 1.70e-02, grad_scale: 32.0 2024-09-22 22:41:42,258 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=118654.66666666667, ans=0.125 2024-09-22 22:41:43,699 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=118654.66666666667, ans=0.125 2024-09-22 22:41:43,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=118654.66666666667, ans=0.0 2024-09-22 22:42:17,534 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.37 vs. limit=15.0 2024-09-22 22:42:42,436 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=118841.33333333333, ans=0.025 2024-09-22 22:42:42,512 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=118841.33333333333, ans=0.125 2024-09-22 22:42:45,729 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=118841.33333333333, ans=0.1 2024-09-22 22:42:58,045 INFO [train.py:1198] (2/4) Epoch 7, batch 2100, loss[loss=0.2637, ctc_loss=0.1873, cr_loss=0.3823, over 17151.00 frames. ], tot_loss[loss=0.278, ctc_loss=0.1983, cr_loss=0.3985, over 3352282.03 frames. ], batch size: 45, lr: 1.70e-02, grad_scale: 32.0 2024-09-22 22:43:03,113 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=118888.0, ans=0.1 2024-09-22 22:43:26,759 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=118934.66666666667, ans=0.125 2024-09-22 22:43:37,286 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 22:44:06,953 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.173e+02 1.431e+02 1.548e+02 1.785e+02 2.946e+02, threshold=3.097e+02, percent-clipped=0.0 2024-09-22 22:44:08,999 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=119074.66666666667, ans=0.125 2024-09-22 22:44:19,896 INFO [train.py:1198] (2/4) Epoch 7, batch 2150, loss[loss=0.2471, ctc_loss=0.1758, cr_loss=0.3565, over 16972.00 frames. ], tot_loss[loss=0.278, ctc_loss=0.1984, cr_loss=0.3983, over 3340517.71 frames. ], batch size: 42, lr: 1.70e-02, grad_scale: 16.0 2024-09-22 22:44:30,706 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=119121.33333333333, ans=0.1 2024-09-22 22:44:38,830 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.93 vs. limit=15.0 2024-09-22 22:44:46,529 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=119168.0, ans=0.0 2024-09-22 22:44:52,767 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=119214.66666666667, ans=0.125 2024-09-22 22:45:04,157 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.30 vs. limit=6.0 2024-09-22 22:45:16,499 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=119261.33333333333, ans=0.125 2024-09-22 22:45:17,993 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=119261.33333333333, ans=0.1 2024-09-22 22:45:24,119 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=119308.0, ans=0.1 2024-09-22 22:45:41,315 INFO [train.py:1198] (2/4) Epoch 7, batch 2200, loss[loss=0.2576, ctc_loss=0.1822, cr_loss=0.377, over 17148.00 frames. ], tot_loss[loss=0.278, ctc_loss=0.1982, cr_loss=0.3987, over 3343625.34 frames. ], batch size: 48, lr: 1.69e-02, grad_scale: 16.0 2024-09-22 22:45:45,522 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=18.41 vs. limit=22.5 2024-09-22 22:46:01,657 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=119401.33333333333, ans=0.125 2024-09-22 22:46:07,626 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=14.68 vs. limit=22.5 2024-09-22 22:46:29,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=119448.0, ans=0.125 2024-09-22 22:46:40,238 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=119494.66666666667, ans=0.95 2024-09-22 22:46:40,253 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=119494.66666666667, ans=0.04949747468305833 2024-09-22 22:46:53,570 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.144e+02 1.357e+02 1.455e+02 1.682e+02 2.486e+02, threshold=2.909e+02, percent-clipped=0.0 2024-09-22 22:46:55,490 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=119541.33333333333, ans=0.0 2024-09-22 22:47:06,284 INFO [train.py:1198] (2/4) Epoch 7, batch 2250, loss[loss=0.2662, ctc_loss=0.1882, cr_loss=0.3899, over 17141.00 frames. ], tot_loss[loss=0.2781, ctc_loss=0.1984, cr_loss=0.3987, over 3347251.35 frames. ], batch size: 48, lr: 1.69e-02, grad_scale: 16.0 2024-09-22 22:47:16,341 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.83 vs. limit=10.0 2024-09-22 22:47:31,854 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=119634.66666666667, ans=0.125 2024-09-22 22:48:16,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=119774.66666666667, ans=0.125 2024-09-22 22:48:25,672 INFO [train.py:1198] (2/4) Epoch 7, batch 2300, loss[loss=0.2619, ctc_loss=0.1841, cr_loss=0.3891, over 17228.00 frames. ], tot_loss[loss=0.278, ctc_loss=0.1983, cr_loss=0.3988, over 3347679.95 frames. ], batch size: 50, lr: 1.69e-02, grad_scale: 16.0 2024-09-22 22:48:26,314 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=11.94 vs. limit=15.0 2024-09-22 22:48:41,119 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=119821.33333333333, ans=0.125 2024-09-22 22:48:49,146 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=119868.0, ans=0.125 2024-09-22 22:49:01,932 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-22 22:49:16,841 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.43 vs. limit=10.0 2024-09-22 22:49:17,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=119961.33333333333, ans=0.125 2024-09-22 22:49:33,035 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=120008.0, ans=0.0 2024-09-22 22:49:37,565 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.210e+02 1.381e+02 1.499e+02 1.768e+02 3.628e+02, threshold=2.997e+02, percent-clipped=2.0 2024-09-22 22:49:50,250 INFO [train.py:1198] (2/4) Epoch 7, batch 2350, loss[loss=0.2718, ctc_loss=0.1911, cr_loss=0.4036, over 17339.00 frames. ], tot_loss[loss=0.2751, ctc_loss=0.196, cr_loss=0.3957, over 3354193.51 frames. ], batch size: 48, lr: 1.69e-02, grad_scale: 16.0 2024-09-22 22:50:05,007 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=120101.33333333333, ans=0.04949747468305833 2024-09-22 22:50:30,608 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=120148.0, ans=0.0 2024-09-22 22:50:30,968 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=15.88 vs. limit=22.5 2024-09-22 22:50:51,166 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=120194.66666666667, ans=0.0 2024-09-22 22:50:53,259 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.26 vs. limit=10.0 2024-09-22 22:51:12,228 INFO [train.py:1198] (2/4) Epoch 7, batch 2400, loss[loss=0.2698, ctc_loss=0.1909, cr_loss=0.3945, over 16969.00 frames. ], tot_loss[loss=0.2758, ctc_loss=0.1964, cr_loss=0.3967, over 3356087.65 frames. ], batch size: 42, lr: 1.69e-02, grad_scale: 32.0 2024-09-22 22:51:15,670 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=120288.0, ans=0.125 2024-09-22 22:51:50,594 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.08 vs. limit=15.0 2024-09-22 22:52:09,260 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 22:52:21,490 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.189e+02 1.376e+02 1.485e+02 1.691e+02 2.822e+02, threshold=2.971e+02, percent-clipped=0.0 2024-09-22 22:52:34,407 INFO [train.py:1198] (2/4) Epoch 7, batch 2450, loss[loss=0.3019, ctc_loss=0.216, cr_loss=0.4295, over 15038.00 frames. ], tot_loss[loss=0.2753, ctc_loss=0.1961, cr_loss=0.3958, over 3339749.77 frames. ], batch size: 89, lr: 1.69e-02, grad_scale: 32.0 2024-09-22 22:52:44,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=120521.33333333333, ans=0.1 2024-09-22 22:52:45,964 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=120521.33333333333, ans=0.5 2024-09-22 22:53:14,582 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=120614.66666666667, ans=0.125 2024-09-22 22:53:19,316 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=120614.66666666667, ans=0.0 2024-09-22 22:53:53,114 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=12.05 vs. limit=15.0 2024-09-22 22:53:56,719 INFO [train.py:1198] (2/4) Epoch 7, batch 2500, loss[loss=0.2644, ctc_loss=0.1886, cr_loss=0.3787, over 17349.00 frames. ], tot_loss[loss=0.2758, ctc_loss=0.1964, cr_loss=0.3967, over 3336401.34 frames. ], batch size: 48, lr: 1.69e-02, grad_scale: 32.0 2024-09-22 22:54:10,459 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=10.50 vs. limit=15.0 2024-09-22 22:54:25,334 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=120801.33333333333, ans=0.125 2024-09-22 22:54:34,873 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=120848.0, ans=0.125 2024-09-22 22:55:06,269 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.238e+02 1.457e+02 1.729e+02 2.020e+02 3.233e+02, threshold=3.458e+02, percent-clipped=3.0 2024-09-22 22:55:18,950 INFO [train.py:1198] (2/4) Epoch 7, batch 2550, loss[loss=0.2559, ctc_loss=0.1789, cr_loss=0.385, over 17133.00 frames. ], tot_loss[loss=0.2746, ctc_loss=0.1954, cr_loss=0.396, over 3344581.43 frames. ], batch size: 48, lr: 1.68e-02, grad_scale: 32.0 2024-09-22 22:55:36,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=121034.66666666667, ans=0.125 2024-09-22 22:55:54,806 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=121081.33333333333, ans=0.2 2024-09-22 22:55:59,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=121081.33333333333, ans=0.1 2024-09-22 22:56:20,496 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=121128.0, ans=0.0 2024-09-22 22:56:40,342 INFO [train.py:1198] (2/4) Epoch 7, batch 2600, loss[loss=0.2878, ctc_loss=0.2082, cr_loss=0.3979, over 17352.00 frames. ], tot_loss[loss=0.2757, ctc_loss=0.1963, cr_loss=0.397, over 3345101.37 frames. ], batch size: 48, lr: 1.68e-02, grad_scale: 32.0 2024-09-22 22:56:44,856 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=121221.33333333333, ans=0.125 2024-09-22 22:57:16,122 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=121314.66666666667, ans=0.125 2024-09-22 22:57:40,372 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=121361.33333333333, ans=0.125 2024-09-22 22:57:49,628 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.141e+02 1.441e+02 1.588e+02 1.825e+02 2.905e+02, threshold=3.176e+02, percent-clipped=0.0 2024-09-22 22:57:50,058 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=121408.0, ans=0.0 2024-09-22 22:58:02,493 INFO [train.py:1198] (2/4) Epoch 7, batch 2650, loss[loss=0.2756, ctc_loss=0.1918, cr_loss=0.4191, over 17305.00 frames. ], tot_loss[loss=0.2764, ctc_loss=0.197, cr_loss=0.3968, over 3334126.52 frames. ], batch size: 51, lr: 1.68e-02, grad_scale: 32.0 2024-09-22 22:58:07,815 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=10.73 vs. limit=15.0 2024-09-22 22:58:09,138 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=121454.66666666667, ans=0.125 2024-09-22 22:58:23,153 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=121501.33333333333, ans=0.125 2024-09-22 22:58:33,778 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=121501.33333333333, ans=0.0 2024-09-22 22:58:37,100 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=121548.0, ans=0.125 2024-09-22 22:58:46,482 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=121548.0, ans=0.2 2024-09-22 22:58:49,684 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=121548.0, ans=0.2 2024-09-22 22:58:51,682 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.84 vs. limit=10.0 2024-09-22 22:58:57,633 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=121594.66666666667, ans=0.125 2024-09-22 22:58:59,156 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=121594.66666666667, ans=0.0 2024-09-22 22:59:04,314 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=121594.66666666667, ans=0.125 2024-09-22 22:59:27,151 INFO [train.py:1198] (2/4) Epoch 7, batch 2700, loss[loss=0.2754, ctc_loss=0.1959, cr_loss=0.3974, over 17343.00 frames. ], tot_loss[loss=0.276, ctc_loss=0.1966, cr_loss=0.3966, over 3334602.06 frames. ], batch size: 48, lr: 1.68e-02, grad_scale: 32.0 2024-09-22 22:59:27,388 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=121688.0, ans=0.1 2024-09-22 23:00:13,359 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=121828.0, ans=0.125 2024-09-22 23:00:33,554 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.157e+02 1.410e+02 1.530e+02 1.699e+02 3.124e+02, threshold=3.060e+02, percent-clipped=0.0 2024-09-22 23:00:41,781 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=121874.66666666667, ans=0.125 2024-09-22 23:00:47,705 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=4.78 vs. limit=12.0 2024-09-22 23:00:48,618 INFO [train.py:1198] (2/4) Epoch 7, batch 2750, loss[loss=0.2879, ctc_loss=0.2047, cr_loss=0.4159, over 17053.00 frames. ], tot_loss[loss=0.2759, ctc_loss=0.1964, cr_loss=0.3972, over 3346775.36 frames. ], batch size: 46, lr: 1.68e-02, grad_scale: 32.0 2024-09-22 23:00:50,606 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=121921.33333333333, ans=0.0 2024-09-22 23:01:17,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=121968.0, ans=0.125 2024-09-22 23:01:19,542 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.75 vs. limit=15.0 2024-09-22 23:01:39,222 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=122061.33333333333, ans=0.2 2024-09-22 23:02:10,731 INFO [train.py:1198] (2/4) Epoch 7, batch 2800, loss[loss=0.3154, ctc_loss=0.2288, cr_loss=0.4331, over 16883.00 frames. ], tot_loss[loss=0.2765, ctc_loss=0.1969, cr_loss=0.3984, over 3351120.84 frames. ], batch size: 58, lr: 1.68e-02, grad_scale: 32.0 2024-09-22 23:02:36,722 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 23:02:56,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=122248.0, ans=0.2 2024-09-22 23:03:18,222 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.138e+02 1.349e+02 1.457e+02 1.581e+02 2.828e+02, threshold=2.913e+02, percent-clipped=0.0 2024-09-22 23:03:20,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=122341.33333333333, ans=0.0 2024-09-22 23:03:33,472 INFO [train.py:1198] (2/4) Epoch 7, batch 2850, loss[loss=0.264, ctc_loss=0.1873, cr_loss=0.3839, over 17156.00 frames. ], tot_loss[loss=0.2761, ctc_loss=0.1966, cr_loss=0.3976, over 3356205.91 frames. ], batch size: 48, lr: 1.67e-02, grad_scale: 32.0 2024-09-22 23:03:57,325 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.max_positive, batch_count=122434.66666666667, ans=0.95 2024-09-22 23:04:06,983 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=122481.33333333333, ans=0.125 2024-09-22 23:04:10,316 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=122481.33333333333, ans=0.0 2024-09-22 23:04:12,301 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=12.87 vs. limit=15.0 2024-09-22 23:04:40,267 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=9.61 vs. limit=15.0 2024-09-22 23:04:55,302 INFO [train.py:1198] (2/4) Epoch 7, batch 2900, loss[loss=0.2716, ctc_loss=0.1898, cr_loss=0.4086, over 17063.00 frames. ], tot_loss[loss=0.2763, ctc_loss=0.1966, cr_loss=0.3985, over 3356074.94 frames. ], batch size: 46, lr: 1.67e-02, grad_scale: 16.0 2024-09-22 23:05:37,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=122714.66666666667, ans=0.125 2024-09-22 23:06:05,780 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.190e+02 1.397e+02 1.534e+02 1.877e+02 3.701e+02, threshold=3.067e+02, percent-clipped=2.0 2024-09-22 23:06:16,923 INFO [train.py:1198] (2/4) Epoch 7, batch 2950, loss[loss=0.2633, ctc_loss=0.1873, cr_loss=0.3802, over 17072.00 frames. ], tot_loss[loss=0.2752, ctc_loss=0.1958, cr_loss=0.3967, over 3367098.84 frames. ], batch size: 46, lr: 1.67e-02, grad_scale: 16.0 2024-09-22 23:06:43,779 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=122901.33333333333, ans=0.0 2024-09-22 23:06:59,572 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=122948.0, ans=0.125 2024-09-22 23:07:02,861 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=122948.0, ans=0.125 2024-09-22 23:07:16,880 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=122994.66666666667, ans=0.2 2024-09-22 23:07:38,710 INFO [train.py:1198] (2/4) Epoch 7, batch 3000, loss[loss=0.3145, ctc_loss=0.2299, cr_loss=0.4228, over 17310.00 frames. ], tot_loss[loss=0.2738, ctc_loss=0.1946, cr_loss=0.3957, over 3372133.81 frames. ], batch size: 51, lr: 1.67e-02, grad_scale: 16.0 2024-09-22 23:07:38,710 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-22 23:07:54,138 INFO [train.py:1230] (2/4) Epoch 7, validation: loss=0.05688, ctc_loss=0.05688, cr_loss=7.669e-15, over 944034.00 frames. 2024-09-22 23:07:54,139 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-22 23:08:11,411 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=123134.66666666667, ans=0.0 2024-09-22 23:08:16,171 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=123134.66666666667, ans=0.1 2024-09-22 23:08:19,422 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=123134.66666666667, ans=0.0 2024-09-22 23:08:22,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=123134.66666666667, ans=0.0 2024-09-22 23:09:01,890 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.166e+02 1.381e+02 1.504e+02 1.827e+02 5.428e+02, threshold=3.008e+02, percent-clipped=9.0 2024-09-22 23:09:12,840 INFO [train.py:1198] (2/4) Epoch 7, batch 3050, loss[loss=0.2584, ctc_loss=0.1833, cr_loss=0.3751, over 17298.00 frames. ], tot_loss[loss=0.273, ctc_loss=0.194, cr_loss=0.3949, over 3368672.93 frames. ], batch size: 49, lr: 1.67e-02, grad_scale: 16.0 2024-09-22 23:09:15,150 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=512, metric=4.77 vs. limit=15.0 2024-09-22 23:09:24,040 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=123321.33333333333, ans=10.0 2024-09-22 23:09:46,798 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=123414.66666666667, ans=0.125 2024-09-22 23:10:26,319 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn1.whiten.whitening_limit, batch_count=123508.0, ans=22.5 2024-09-22 23:10:33,467 INFO [train.py:1198] (2/4) Epoch 7, batch 3100, loss[loss=0.2947, ctc_loss=0.2076, cr_loss=0.4352, over 16438.00 frames. ], tot_loss[loss=0.2741, ctc_loss=0.1948, cr_loss=0.3964, over 3373764.23 frames. ], batch size: 66, lr: 1.67e-02, grad_scale: 16.0 2024-09-22 23:10:47,621 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=123601.33333333333, ans=0.125 2024-09-22 23:11:08,439 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=14.14 vs. limit=15.0 2024-09-22 23:11:18,926 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=123694.66666666667, ans=0.125 2024-09-22 23:11:35,417 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=123694.66666666667, ans=0.125 2024-09-22 23:11:42,860 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.191e+02 1.455e+02 1.593e+02 1.857e+02 3.038e+02, threshold=3.186e+02, percent-clipped=1.0 2024-09-22 23:11:53,642 INFO [train.py:1198] (2/4) Epoch 7, batch 3150, loss[loss=0.3147, ctc_loss=0.2246, cr_loss=0.4509, over 17004.00 frames. ], tot_loss[loss=0.2753, ctc_loss=0.1957, cr_loss=0.3979, over 3371058.13 frames. ], batch size: 53, lr: 1.67e-02, grad_scale: 16.0 2024-09-22 23:13:12,036 INFO [train.py:1198] (2/4) Epoch 7, batch 3200, loss[loss=0.257, ctc_loss=0.184, cr_loss=0.365, over 17097.00 frames. ], tot_loss[loss=0.2743, ctc_loss=0.195, cr_loss=0.3966, over 3368861.79 frames. ], batch size: 43, lr: 1.66e-02, grad_scale: 32.0 2024-09-22 23:13:51,409 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=124114.66666666667, ans=0.2 2024-09-22 23:13:59,488 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=5.49 vs. limit=15.0 2024-09-22 23:14:10,228 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=124161.33333333333, ans=0.0 2024-09-22 23:14:19,024 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.163e+02 1.425e+02 1.561e+02 1.765e+02 3.696e+02, threshold=3.122e+02, percent-clipped=1.0 2024-09-22 23:14:29,938 INFO [train.py:1198] (2/4) Epoch 7, batch 3250, loss[loss=0.2756, ctc_loss=0.1949, cr_loss=0.4034, over 17210.00 frames. ], tot_loss[loss=0.2727, ctc_loss=0.1937, cr_loss=0.3947, over 3377052.83 frames. ], batch size: 50, lr: 1.66e-02, grad_scale: 32.0 2024-09-22 23:14:32,250 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=11.65 vs. limit=15.0 2024-09-22 23:14:45,395 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=6.45 vs. limit=15.0 2024-09-22 23:14:51,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=124301.33333333333, ans=0.025 2024-09-22 23:14:57,213 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=124301.33333333333, ans=0.1 2024-09-22 23:15:00,258 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=124301.33333333333, ans=0.1 2024-09-22 23:15:13,447 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=15.50 vs. limit=22.5 2024-09-22 23:15:18,927 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=124394.66666666667, ans=0.125 2024-09-22 23:15:28,462 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 23:15:42,220 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=124441.33333333333, ans=0.0 2024-09-22 23:15:49,579 INFO [train.py:1198] (2/4) Epoch 7, batch 3300, loss[loss=0.2928, ctc_loss=0.2145, cr_loss=0.3917, over 16021.00 frames. ], tot_loss[loss=0.2723, ctc_loss=0.1935, cr_loss=0.3939, over 3366694.86 frames. ], batch size: 74, lr: 1.66e-02, grad_scale: 32.0 2024-09-22 23:16:05,097 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=124534.66666666667, ans=0.125 2024-09-22 23:16:06,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer_ff2.min_abs, batch_count=124534.66666666667, ans=0.1 2024-09-22 23:16:15,191 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=124534.66666666667, ans=0.025 2024-09-22 23:16:21,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=124581.33333333333, ans=0.0 2024-09-22 23:16:55,373 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=124674.66666666667, ans=0.2 2024-09-22 23:16:58,143 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.199e+02 1.418e+02 1.603e+02 1.871e+02 3.678e+02, threshold=3.206e+02, percent-clipped=3.0 2024-09-22 23:17:03,691 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=10.94 vs. limit=15.0 2024-09-22 23:17:06,747 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=6.53 vs. limit=15.0 2024-09-22 23:17:09,115 INFO [train.py:1198] (2/4) Epoch 7, batch 3350, loss[loss=0.2539, ctc_loss=0.1844, cr_loss=0.3472, over 17283.00 frames. ], tot_loss[loss=0.2732, ctc_loss=0.1941, cr_loss=0.3952, over 3369693.63 frames. ], batch size: 42, lr: 1.66e-02, grad_scale: 32.0 2024-09-22 23:17:17,837 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=11.10 vs. limit=15.0 2024-09-22 23:17:36,005 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=10.34 vs. limit=12.0 2024-09-22 23:18:13,005 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=124908.0, ans=10.0 2024-09-22 23:18:26,927 INFO [train.py:1198] (2/4) Epoch 7, batch 3400, loss[loss=0.3121, ctc_loss=0.226, cr_loss=0.431, over 16923.00 frames. ], tot_loss[loss=0.2733, ctc_loss=0.1941, cr_loss=0.3957, over 3369139.22 frames. ], batch size: 58, lr: 1.66e-02, grad_scale: 32.0 2024-09-22 23:18:28,778 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=124954.66666666667, ans=0.0 2024-09-22 23:18:36,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=124954.66666666667, ans=0.0 2024-09-22 23:18:45,045 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=5.20 vs. limit=10.0 2024-09-22 23:18:48,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten.whitening_limit, batch_count=125001.33333333333, ans=22.5 2024-09-22 23:18:50,076 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=13.86 vs. limit=15.0 2024-09-22 23:19:29,361 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=13.15 vs. limit=15.0 2024-09-22 23:19:34,687 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.374e+02 1.483e+02 1.750e+02 2.564e+02, threshold=2.966e+02, percent-clipped=0.0 2024-09-22 23:19:38,511 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=14.07 vs. limit=15.0 2024-09-22 23:19:45,700 INFO [train.py:1198] (2/4) Epoch 7, batch 3450, loss[loss=0.2901, ctc_loss=0.2086, cr_loss=0.4076, over 17109.00 frames. ], tot_loss[loss=0.2738, ctc_loss=0.1946, cr_loss=0.3957, over 3359838.31 frames. ], batch size: 49, lr: 1.66e-02, grad_scale: 32.0 2024-09-22 23:19:49,085 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.max_abs, batch_count=125188.0, ans=10.0 2024-09-22 23:19:55,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=125188.0, ans=0.2 2024-09-22 23:20:13,139 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=125234.66666666667, ans=0.125 2024-09-22 23:20:42,635 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=125328.0, ans=0.125 2024-09-22 23:21:05,780 INFO [train.py:1198] (2/4) Epoch 7, batch 3500, loss[loss=0.2457, ctc_loss=0.1739, cr_loss=0.3591, over 17109.00 frames. ], tot_loss[loss=0.2731, ctc_loss=0.1941, cr_loss=0.3952, over 3366065.06 frames. ], batch size: 49, lr: 1.66e-02, grad_scale: 32.0 2024-09-22 23:21:29,748 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=125468.0, ans=0.0 2024-09-22 23:22:02,260 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=125561.33333333333, ans=0.025 2024-09-22 23:22:03,955 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=125561.33333333333, ans=0.04949747468305833 2024-09-22 23:22:05,439 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer_ff3.min_abs, batch_count=125561.33333333333, ans=0.2 2024-09-22 23:22:11,641 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=125608.0, ans=0.0 2024-09-22 23:22:14,393 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.203e+02 1.377e+02 1.536e+02 1.902e+02 3.624e+02, threshold=3.071e+02, percent-clipped=2.0 2024-09-22 23:22:17,696 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=125608.0, ans=0.2 2024-09-22 23:22:19,306 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=125608.0, ans=0.125 2024-09-22 23:22:20,542 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=125608.0, ans=0.1 2024-09-22 23:22:25,092 INFO [train.py:1198] (2/4) Epoch 7, batch 3550, loss[loss=0.2926, ctc_loss=0.2148, cr_loss=0.389, over 16553.00 frames. ], tot_loss[loss=0.274, ctc_loss=0.1948, cr_loss=0.3959, over 3363651.07 frames. ], batch size: 66, lr: 1.65e-02, grad_scale: 32.0 2024-09-22 23:22:37,067 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=3.74 vs. limit=15.0 2024-09-22 23:23:02,130 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=125748.0, ans=0.0 2024-09-22 23:23:05,286 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=125748.0, ans=0.125 2024-09-22 23:23:08,206 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=125748.0, ans=0.125 2024-09-22 23:23:33,300 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=125841.33333333333, ans=0.2 2024-09-22 23:23:34,682 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=125841.33333333333, ans=0.2 2024-09-22 23:23:42,344 INFO [train.py:1198] (2/4) Epoch 7, batch 3600, loss[loss=0.3219, ctc_loss=0.2374, cr_loss=0.4225, over 14787.00 frames. ], tot_loss[loss=0.2737, ctc_loss=0.1946, cr_loss=0.3957, over 3362899.80 frames. ], batch size: 89, lr: 1.65e-02, grad_scale: 32.0 2024-09-22 23:23:50,447 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=125888.0, ans=0.125 2024-09-22 23:23:54,169 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=125888.0, ans=0.125 2024-09-22 23:23:57,116 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=125888.0, ans=0.025 2024-09-22 23:24:13,003 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.33 vs. limit=15.0 2024-09-22 23:24:41,610 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=126028.0, ans=0.0 2024-09-22 23:24:50,666 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.222e+02 1.445e+02 1.618e+02 1.998e+02 3.129e+02, threshold=3.236e+02, percent-clipped=1.0 2024-09-22 23:24:52,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=126074.66666666667, ans=0.0 2024-09-22 23:25:00,287 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=126121.33333333333, ans=0.125 2024-09-22 23:25:00,296 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=126121.33333333333, ans=0.1 2024-09-22 23:25:01,498 INFO [train.py:1198] (2/4) Epoch 7, batch 3650, loss[loss=0.3024, ctc_loss=0.2148, cr_loss=0.4378, over 17013.00 frames. ], tot_loss[loss=0.2727, ctc_loss=0.1939, cr_loss=0.3944, over 3355967.86 frames. ], batch size: 51, lr: 1.65e-02, grad_scale: 32.0 2024-09-22 23:25:04,780 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=126121.33333333333, ans=0.1 2024-09-22 23:25:07,244 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn1.whiten, num_groups=1, num_channels=192, metric=9.67 vs. limit=22.5 2024-09-22 23:25:59,555 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=11.60 vs. limit=15.0 2024-09-22 23:26:16,384 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=126308.0, ans=0.0 2024-09-22 23:26:19,898 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=12.68 vs. limit=15.0 2024-09-22 23:26:20,853 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=126354.66666666667, ans=0.125 2024-09-22 23:26:22,287 INFO [train.py:1198] (2/4) Epoch 7, batch 3700, loss[loss=0.3167, ctc_loss=0.2275, cr_loss=0.4463, over 16587.00 frames. ], tot_loss[loss=0.2737, ctc_loss=0.1946, cr_loss=0.3957, over 3358448.13 frames. ], batch size: 66, lr: 1.65e-02, grad_scale: 32.0 2024-09-22 23:26:35,303 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=126354.66666666667, ans=0.2 2024-09-22 23:26:38,308 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.max_positive, batch_count=126401.33333333333, ans=0.95 2024-09-22 23:27:08,452 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=126494.66666666667, ans=0.125 2024-09-22 23:27:30,062 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.388e+02 1.512e+02 1.895e+02 2.751e+02, threshold=3.024e+02, percent-clipped=0.0 2024-09-22 23:27:30,950 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=21.79 vs. limit=22.5 2024-09-22 23:27:31,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=126541.33333333333, ans=0.125 2024-09-22 23:27:41,021 INFO [train.py:1198] (2/4) Epoch 7, batch 3750, loss[loss=0.2772, ctc_loss=0.2004, cr_loss=0.3839, over 17295.00 frames. ], tot_loss[loss=0.2744, ctc_loss=0.1951, cr_loss=0.3962, over 3345192.41 frames. ], batch size: 46, lr: 1.65e-02, grad_scale: 32.0 2024-09-22 23:27:53,506 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=126588.0, ans=0.1 2024-09-22 23:27:55,033 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=126634.66666666667, ans=0.125 2024-09-22 23:27:59,890 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 23:28:15,456 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=126681.33333333333, ans=0.125 2024-09-22 23:28:32,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=126728.0, ans=0.1 2024-09-22 23:28:42,108 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-22 23:28:59,018 INFO [train.py:1198] (2/4) Epoch 7, batch 3800, loss[loss=0.2427, ctc_loss=0.1683, cr_loss=0.3721, over 17019.00 frames. ], tot_loss[loss=0.2747, ctc_loss=0.1955, cr_loss=0.3961, over 3332702.69 frames. ], batch size: 44, lr: 1.65e-02, grad_scale: 32.0 2024-09-22 23:29:16,293 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=126868.0, ans=0.1 2024-09-22 23:29:33,528 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=126914.66666666667, ans=0.07 2024-09-22 23:29:37,606 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=126914.66666666667, ans=0.125 2024-09-22 23:29:39,264 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=126914.66666666667, ans=0.125 2024-09-22 23:30:05,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=127008.0, ans=0.04949747468305833 2024-09-22 23:30:07,020 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.225e+02 1.460e+02 1.621e+02 1.904e+02 2.959e+02, threshold=3.241e+02, percent-clipped=0.0 2024-09-22 23:30:17,775 INFO [train.py:1198] (2/4) Epoch 7, batch 3850, loss[loss=0.3321, ctc_loss=0.237, cr_loss=0.4753, over 14871.00 frames. ], tot_loss[loss=0.2773, ctc_loss=0.1978, cr_loss=0.3972, over 3297157.46 frames. ], batch size: 89, lr: 1.65e-02, grad_scale: 32.0 2024-09-22 23:30:24,846 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.18 vs. limit=15.0 2024-09-22 23:30:25,732 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=127054.66666666667, ans=0.125 2024-09-22 23:30:29,699 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.49 vs. limit=6.0 2024-09-22 23:30:52,562 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=127148.0, ans=0.0 2024-09-22 23:31:01,397 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=127148.0, ans=0.0 2024-09-22 23:31:13,771 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=127194.66666666667, ans=0.125 2024-09-22 23:31:25,570 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=127241.33333333333, ans=0.1 2024-09-22 23:32:20,221 INFO [train.py:1198] (2/4) Epoch 8, batch 0, loss[loss=0.2578, ctc_loss=0.1843, cr_loss=0.3677, over 17241.00 frames. ], tot_loss[loss=0.2578, ctc_loss=0.1843, cr_loss=0.3677, over 17241.00 frames. ], batch size: 50, lr: 1.55e-02, grad_scale: 32.0 2024-09-22 23:32:20,221 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-22 23:32:35,550 INFO [train.py:1230] (2/4) Epoch 8, validation: loss=0.05692, ctc_loss=0.05692, cr_loss=7.316e-15, over 944034.00 frames. 2024-09-22 23:32:35,551 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-22 23:32:37,716 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.81 vs. limit=15.0 2024-09-22 23:33:27,843 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=7.12 vs. limit=15.0 2024-09-22 23:33:28,617 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=127409.33333333333, ans=0.2 2024-09-22 23:33:39,979 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=127456.0, ans=10.0 2024-09-22 23:33:52,392 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.139e+02 1.394e+02 1.752e+02 2.098e+02 6.301e+02, threshold=3.504e+02, percent-clipped=3.0 2024-09-22 23:33:55,627 INFO [train.py:1198] (2/4) Epoch 8, batch 50, loss[loss=0.2869, ctc_loss=0.1966, cr_loss=0.4513, over 16742.00 frames. ], tot_loss[loss=0.2759, ctc_loss=0.1959, cr_loss=0.3998, over 764351.73 frames. ], batch size: 61, lr: 1.55e-02, grad_scale: 32.0 2024-09-22 23:34:20,705 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.90 vs. limit=15.0 2024-09-22 23:34:27,822 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=127596.0, ans=0.125 2024-09-22 23:35:02,466 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=127689.33333333333, ans=0.1 2024-09-22 23:35:19,390 INFO [train.py:1198] (2/4) Epoch 8, batch 100, loss[loss=0.2735, ctc_loss=0.191, cr_loss=0.4125, over 17053.00 frames. ], tot_loss[loss=0.2712, ctc_loss=0.192, cr_loss=0.396, over 1350411.89 frames. ], batch size: 46, lr: 1.55e-02, grad_scale: 32.0 2024-09-22 23:35:24,471 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=127736.0, ans=0.125 2024-09-22 23:35:59,135 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=127829.33333333333, ans=0.1 2024-09-22 23:36:23,526 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=127922.66666666667, ans=0.125 2024-09-22 23:36:34,539 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=127922.66666666667, ans=0.1 2024-09-22 23:36:34,576 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=127922.66666666667, ans=0.125 2024-09-22 23:36:34,647 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=127922.66666666667, ans=0.125 2024-09-22 23:36:37,479 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.179e+02 1.356e+02 1.486e+02 1.737e+02 3.200e+02, threshold=2.973e+02, percent-clipped=0.0 2024-09-22 23:36:40,697 INFO [train.py:1198] (2/4) Epoch 8, batch 150, loss[loss=0.2576, ctc_loss=0.1795, cr_loss=0.3905, over 17354.00 frames. ], tot_loss[loss=0.2685, ctc_loss=0.1897, cr_loss=0.3942, over 1804054.39 frames. ], batch size: 48, lr: 1.55e-02, grad_scale: 32.0 2024-09-22 23:37:01,212 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.81 vs. limit=6.0 2024-09-22 23:37:06,442 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=128016.0, ans=0.125 2024-09-22 23:37:45,167 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=11.52 vs. limit=15.0 2024-09-22 23:37:56,114 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=128156.0, ans=0.0 2024-09-22 23:37:56,148 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=128156.0, ans=0.0 2024-09-22 23:38:02,290 INFO [train.py:1198] (2/4) Epoch 8, batch 200, loss[loss=0.3308, ctc_loss=0.248, cr_loss=0.4141, over 11416.00 frames. ], tot_loss[loss=0.2702, ctc_loss=0.1911, cr_loss=0.3956, over 2153517.70 frames. ], batch size: 123, lr: 1.54e-02, grad_scale: 32.0 2024-09-22 23:38:02,647 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=128202.66666666667, ans=0.0 2024-09-22 23:39:20,518 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.202e+02 1.362e+02 1.519e+02 1.695e+02 2.654e+02, threshold=3.037e+02, percent-clipped=0.0 2024-09-22 23:39:23,708 INFO [train.py:1198] (2/4) Epoch 8, batch 250, loss[loss=0.2683, ctc_loss=0.1911, cr_loss=0.3857, over 16942.00 frames. ], tot_loss[loss=0.2686, ctc_loss=0.19, cr_loss=0.3935, over 2425493.71 frames. ], batch size: 58, lr: 1.54e-02, grad_scale: 32.0 2024-09-22 23:39:59,721 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=8.82 vs. limit=15.0 2024-09-22 23:40:02,411 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=128529.33333333333, ans=0.125 2024-09-22 23:40:13,236 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=128576.0, ans=0.125 2024-09-22 23:40:26,136 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-22 23:40:34,070 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=128622.66666666667, ans=0.125 2024-09-22 23:40:46,249 INFO [train.py:1198] (2/4) Epoch 8, batch 300, loss[loss=0.3172, ctc_loss=0.2279, cr_loss=0.4463, over 16204.00 frames. ], tot_loss[loss=0.2697, ctc_loss=0.1907, cr_loss=0.3947, over 2626020.11 frames. ], batch size: 74, lr: 1.54e-02, grad_scale: 32.0 2024-09-22 23:41:11,443 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=128716.0, ans=0.125 2024-09-22 23:41:11,467 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=128716.0, ans=0.125 2024-09-22 23:41:18,022 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=128716.0, ans=0.0 2024-09-22 23:41:21,135 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=128762.66666666667, ans=0.125 2024-09-22 23:42:07,483 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.170e+02 1.437e+02 1.612e+02 1.809e+02 3.626e+02, threshold=3.224e+02, percent-clipped=3.0 2024-09-22 23:42:09,370 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=128902.66666666667, ans=0.1 2024-09-22 23:42:10,765 INFO [train.py:1198] (2/4) Epoch 8, batch 350, loss[loss=0.3025, ctc_loss=0.2127, cr_loss=0.4492, over 17039.00 frames. ], tot_loss[loss=0.2703, ctc_loss=0.1911, cr_loss=0.3958, over 2793652.90 frames. ], batch size: 52, lr: 1.54e-02, grad_scale: 32.0 2024-09-22 23:42:12,669 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=128902.66666666667, ans=0.125 2024-09-22 23:42:30,813 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=12.30 vs. limit=15.0 2024-09-22 23:42:41,513 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=128996.0, ans=0.1 2024-09-22 23:43:08,272 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=129042.66666666667, ans=0.025 2024-09-22 23:43:17,938 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=129089.33333333333, ans=0.1 2024-09-22 23:43:19,969 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.18 vs. limit=10.0 2024-09-22 23:43:30,392 INFO [train.py:1198] (2/4) Epoch 8, batch 400, loss[loss=0.2696, ctc_loss=0.1924, cr_loss=0.3858, over 17290.00 frames. ], tot_loss[loss=0.2706, ctc_loss=0.1915, cr_loss=0.3956, over 2925849.07 frames. ], batch size: 46, lr: 1.54e-02, grad_scale: 32.0 2024-09-22 23:43:37,481 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.59 vs. limit=6.0 2024-09-22 23:43:51,396 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=129182.66666666667, ans=0.2 2024-09-22 23:44:00,806 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=129229.33333333333, ans=0.125 2024-09-22 23:44:05,669 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=129229.33333333333, ans=0.125 2024-09-22 23:44:08,806 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=129229.33333333333, ans=0.1 2024-09-22 23:44:37,050 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=129322.66666666667, ans=0.125 2024-09-22 23:44:49,603 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.185e+02 1.364e+02 1.519e+02 1.603e+02 2.870e+02, threshold=3.038e+02, percent-clipped=0.0 2024-09-22 23:44:52,557 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=9.26 vs. limit=10.0 2024-09-22 23:44:52,857 INFO [train.py:1198] (2/4) Epoch 8, batch 450, loss[loss=0.262, ctc_loss=0.1849, cr_loss=0.3856, over 17222.00 frames. ], tot_loss[loss=0.2696, ctc_loss=0.1907, cr_loss=0.3945, over 3024639.49 frames. ], batch size: 50, lr: 1.54e-02, grad_scale: 32.0 2024-09-22 23:45:21,124 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=129416.0, ans=0.125 2024-09-22 23:45:40,503 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=129462.66666666667, ans=0.2 2024-09-22 23:45:51,587 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=129509.33333333333, ans=0.0 2024-09-22 23:46:18,032 INFO [train.py:1198] (2/4) Epoch 8, batch 500, loss[loss=0.3039, ctc_loss=0.2194, cr_loss=0.4226, over 15130.00 frames. ], tot_loss[loss=0.2675, ctc_loss=0.1892, cr_loss=0.3917, over 3108977.07 frames. ], batch size: 89, lr: 1.54e-02, grad_scale: 32.0 2024-09-22 23:46:47,152 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=129649.33333333333, ans=0.1 2024-09-22 23:46:49,236 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.18 vs. limit=10.0 2024-09-22 23:47:36,030 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.188e+02 1.377e+02 1.524e+02 1.726e+02 2.894e+02, threshold=3.047e+02, percent-clipped=0.0 2024-09-22 23:47:39,313 INFO [train.py:1198] (2/4) Epoch 8, batch 550, loss[loss=0.2311, ctc_loss=0.1626, cr_loss=0.3428, over 17305.00 frames. ], tot_loss[loss=0.2701, ctc_loss=0.1913, cr_loss=0.3938, over 3152473.76 frames. ], batch size: 46, lr: 1.54e-02, grad_scale: 32.0 2024-09-22 23:47:47,625 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=129836.0, ans=0.0 2024-09-22 23:47:49,604 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=19.46 vs. limit=22.5 2024-09-22 23:48:00,393 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=129882.66666666667, ans=0.2 2024-09-22 23:48:14,874 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-22 23:48:25,910 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=129976.0, ans=0.125 2024-09-22 23:48:30,949 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.24 vs. limit=6.0 2024-09-22 23:48:58,802 INFO [train.py:1198] (2/4) Epoch 8, batch 600, loss[loss=0.2979, ctc_loss=0.2116, cr_loss=0.4317, over 17300.00 frames. ], tot_loss[loss=0.2711, ctc_loss=0.1922, cr_loss=0.3945, over 3182857.86 frames. ], batch size: 51, lr: 1.53e-02, grad_scale: 32.0 2024-09-22 23:49:00,624 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=130069.33333333333, ans=0.2 2024-09-22 23:50:07,866 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=130256.0, ans=0.125 2024-09-22 23:50:20,278 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.107e+02 1.396e+02 1.531e+02 1.850e+02 5.586e+02, threshold=3.061e+02, percent-clipped=2.0 2024-09-22 23:50:23,504 INFO [train.py:1198] (2/4) Epoch 8, batch 650, loss[loss=0.2318, ctc_loss=0.1602, cr_loss=0.3577, over 17100.00 frames. ], tot_loss[loss=0.2697, ctc_loss=0.191, cr_loss=0.3938, over 3228544.55 frames. ], batch size: 43, lr: 1.53e-02, grad_scale: 32.0 2024-09-22 23:50:28,585 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=130302.66666666667, ans=0.035 2024-09-22 23:50:35,850 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=13.54 vs. limit=22.5 2024-09-22 23:50:43,227 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=130349.33333333333, ans=0.125 2024-09-22 23:51:19,039 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=130442.66666666667, ans=0.0 2024-09-22 23:51:48,928 INFO [train.py:1198] (2/4) Epoch 8, batch 700, loss[loss=0.3112, ctc_loss=0.2242, cr_loss=0.4351, over 17034.00 frames. ], tot_loss[loss=0.2695, ctc_loss=0.1908, cr_loss=0.3936, over 3268088.69 frames. ], batch size: 52, lr: 1.53e-02, grad_scale: 32.0 2024-09-22 23:51:58,654 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=130536.0, ans=0.125 2024-09-22 23:52:09,633 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=130582.66666666667, ans=0.0 2024-09-22 23:52:29,607 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.96 vs. limit=15.0 2024-09-22 23:52:35,897 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=130629.33333333333, ans=0.0 2024-09-22 23:52:37,473 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=130676.0, ans=0.0 2024-09-22 23:53:07,233 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.368e+02 1.561e+02 1.822e+02 3.141e+02, threshold=3.123e+02, percent-clipped=1.0 2024-09-22 23:53:09,207 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=130769.33333333333, ans=0.2 2024-09-22 23:53:10,432 INFO [train.py:1198] (2/4) Epoch 8, batch 750, loss[loss=0.317, ctc_loss=0.2335, cr_loss=0.4173, over 12083.00 frames. ], tot_loss[loss=0.2713, ctc_loss=0.1923, cr_loss=0.3951, over 3277875.74 frames. ], batch size: 123, lr: 1.53e-02, grad_scale: 32.0 2024-09-22 23:53:36,103 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=130816.0, ans=0.0 2024-09-22 23:53:46,005 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.47 vs. limit=15.0 2024-09-22 23:53:52,069 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=130862.66666666667, ans=0.0 2024-09-22 23:54:23,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=130956.0, ans=0.2 2024-09-22 23:54:33,311 INFO [train.py:1198] (2/4) Epoch 8, batch 800, loss[loss=0.244, ctc_loss=0.1764, cr_loss=0.3376, over 17005.00 frames. ], tot_loss[loss=0.2697, ctc_loss=0.191, cr_loss=0.3937, over 3298685.94 frames. ], batch size: 39, lr: 1.53e-02, grad_scale: 32.0 2024-09-22 23:55:06,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=131096.0, ans=0.0 2024-09-22 23:55:08,695 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=7.38 vs. limit=8.0 2024-09-22 23:55:28,726 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=131142.66666666666, ans=0.2 2024-09-22 23:55:41,322 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=131189.33333333334, ans=0.025 2024-09-22 23:55:46,574 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=512, metric=4.17 vs. limit=15.0 2024-09-22 23:55:54,935 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.155e+02 1.424e+02 1.641e+02 1.911e+02 2.980e+02, threshold=3.282e+02, percent-clipped=0.0 2024-09-22 23:55:58,176 INFO [train.py:1198] (2/4) Epoch 8, batch 850, loss[loss=0.2746, ctc_loss=0.1924, cr_loss=0.411, over 17034.00 frames. ], tot_loss[loss=0.27, ctc_loss=0.1912, cr_loss=0.394, over 3307833.46 frames. ], batch size: 51, lr: 1.53e-02, grad_scale: 32.0 2024-09-22 23:56:16,194 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=131282.66666666666, ans=0.0 2024-09-22 23:56:30,510 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=131282.66666666666, ans=0.1 2024-09-22 23:56:45,677 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.83 vs. limit=12.0 2024-09-22 23:57:07,295 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=131422.66666666666, ans=0.125 2024-09-22 23:57:07,648 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.17 vs. limit=6.0 2024-09-22 23:57:10,468 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=131422.66666666666, ans=0.0 2024-09-22 23:57:21,042 INFO [train.py:1198] (2/4) Epoch 8, batch 900, loss[loss=0.2804, ctc_loss=0.1991, cr_loss=0.4068, over 17081.00 frames. ], tot_loss[loss=0.2696, ctc_loss=0.1908, cr_loss=0.3938, over 3314359.87 frames. ], batch size: 43, lr: 1.53e-02, grad_scale: 32.0 2024-09-22 23:57:31,025 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=131469.33333333334, ans=0.09899494936611666 2024-09-22 23:57:37,454 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=131516.0, ans=0.1 2024-09-22 23:57:38,942 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=131516.0, ans=0.125 2024-09-22 23:57:58,195 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=131562.66666666666, ans=0.0 2024-09-22 23:57:59,047 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.32 vs. limit=15.0 2024-09-22 23:58:27,153 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=131656.0, ans=0.125 2024-09-22 23:58:37,889 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.169e+02 1.417e+02 1.553e+02 1.784e+02 2.618e+02, threshold=3.106e+02, percent-clipped=0.0 2024-09-22 23:58:38,154 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=131656.0, ans=0.2 2024-09-22 23:58:41,089 INFO [train.py:1198] (2/4) Epoch 8, batch 950, loss[loss=0.2731, ctc_loss=0.1987, cr_loss=0.3719, over 17307.00 frames. ], tot_loss[loss=0.2687, ctc_loss=0.1901, cr_loss=0.3927, over 3323439.46 frames. ], batch size: 49, lr: 1.53e-02, grad_scale: 32.0 2024-09-22 23:58:49,396 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=131702.66666666666, ans=0.0 2024-09-22 23:59:06,405 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=131749.33333333334, ans=0.1 2024-09-22 23:59:19,095 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=131796.0, ans=0.125 2024-09-22 23:59:36,191 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=131842.66666666666, ans=0.0 2024-09-23 00:00:05,377 INFO [train.py:1198] (2/4) Epoch 8, batch 1000, loss[loss=0.2645, ctc_loss=0.1873, cr_loss=0.3856, over 17038.00 frames. ], tot_loss[loss=0.2687, ctc_loss=0.1901, cr_loss=0.3926, over 3334078.47 frames. ], batch size: 44, lr: 1.52e-02, grad_scale: 32.0 2024-09-23 00:00:25,065 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=9.70 vs. limit=15.0 2024-09-23 00:00:31,040 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=131982.66666666666, ans=0.125 2024-09-23 00:00:31,322 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.56 vs. limit=15.0 2024-09-23 00:00:48,646 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=2.82 vs. limit=15.0 2024-09-23 00:00:59,195 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer_na.min_abs, batch_count=132076.0, ans=0.02 2024-09-23 00:01:16,589 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=132122.66666666666, ans=0.025 2024-09-23 00:01:26,819 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.157e+02 1.410e+02 1.544e+02 1.747e+02 2.558e+02, threshold=3.087e+02, percent-clipped=0.0 2024-09-23 00:01:29,992 INFO [train.py:1198] (2/4) Epoch 8, batch 1050, loss[loss=0.3179, ctc_loss=0.2385, cr_loss=0.3974, over 11842.00 frames. ], tot_loss[loss=0.2686, ctc_loss=0.1899, cr_loss=0.3935, over 3335979.57 frames. ], batch size: 123, lr: 1.52e-02, grad_scale: 32.0 2024-09-23 00:02:46,245 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=132356.0, ans=0.2 2024-09-23 00:02:49,158 INFO [train.py:1198] (2/4) Epoch 8, batch 1100, loss[loss=0.2525, ctc_loss=0.1795, cr_loss=0.3648, over 17099.00 frames. ], tot_loss[loss=0.2697, ctc_loss=0.1907, cr_loss=0.3949, over 3346439.92 frames. ], batch size: 49, lr: 1.52e-02, grad_scale: 32.0 2024-09-23 00:03:02,323 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=132402.66666666666, ans=0.0 2024-09-23 00:03:30,824 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=132496.0, ans=0.1 2024-09-23 00:04:08,242 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.144e+02 1.381e+02 1.495e+02 1.697e+02 2.213e+02, threshold=2.991e+02, percent-clipped=0.0 2024-09-23 00:04:11,435 INFO [train.py:1198] (2/4) Epoch 8, batch 1150, loss[loss=0.2374, ctc_loss=0.1657, cr_loss=0.3586, over 17274.00 frames. ], tot_loss[loss=0.2703, ctc_loss=0.1912, cr_loss=0.3951, over 3346017.52 frames. ], batch size: 44, lr: 1.52e-02, grad_scale: 32.0 2024-09-23 00:04:42,406 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=132729.33333333334, ans=0.125 2024-09-23 00:04:43,854 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=132729.33333333334, ans=0.125 2024-09-23 00:04:58,015 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=14.94 vs. limit=15.0 2024-09-23 00:05:10,349 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 00:05:33,701 INFO [train.py:1198] (2/4) Epoch 8, batch 1200, loss[loss=0.2319, ctc_loss=0.1623, cr_loss=0.3476, over 17073.00 frames. ], tot_loss[loss=0.2694, ctc_loss=0.1904, cr_loss=0.3946, over 3348175.78 frames. ], batch size: 46, lr: 1.52e-02, grad_scale: 32.0 2024-09-23 00:05:37,135 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=132869.33333333334, ans=0.2 2024-09-23 00:05:57,206 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=132916.0, ans=0.1 2024-09-23 00:06:07,437 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.74 vs. limit=15.0 2024-09-23 00:06:19,225 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=132962.66666666666, ans=0.2 2024-09-23 00:06:20,843 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=132962.66666666666, ans=0.0 2024-09-23 00:06:38,683 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=133009.33333333334, ans=0.0 2024-09-23 00:06:57,396 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.195e+02 1.352e+02 1.465e+02 1.619e+02 2.564e+02, threshold=2.930e+02, percent-clipped=0.0 2024-09-23 00:06:58,990 INFO [train.py:1198] (2/4) Epoch 8, batch 1250, loss[loss=0.2554, ctc_loss=0.1755, cr_loss=0.3995, over 17314.00 frames. ], tot_loss[loss=0.2696, ctc_loss=0.1906, cr_loss=0.3949, over 3351621.17 frames. ], batch size: 46, lr: 1.52e-02, grad_scale: 32.0 2024-09-23 00:07:05,700 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=133102.66666666666, ans=0.0 2024-09-23 00:07:15,927 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.83 vs. limit=12.0 2024-09-23 00:07:40,416 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=133196.0, ans=0.125 2024-09-23 00:08:09,108 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=133289.33333333334, ans=0.125 2024-09-23 00:08:14,043 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.69 vs. limit=10.0 2024-09-23 00:08:18,295 INFO [train.py:1198] (2/4) Epoch 8, batch 1300, loss[loss=0.2844, ctc_loss=0.2005, cr_loss=0.4195, over 17035.00 frames. ], tot_loss[loss=0.2696, ctc_loss=0.1906, cr_loss=0.3948, over 3356249.60 frames. ], batch size: 56, lr: 1.52e-02, grad_scale: 32.0 2024-09-23 00:08:45,728 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=133382.66666666666, ans=0.125 2024-09-23 00:08:56,791 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=133429.33333333334, ans=0.1 2024-09-23 00:09:30,255 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=133522.66666666666, ans=0.1 2024-09-23 00:09:38,132 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.229e+02 1.339e+02 1.481e+02 1.642e+02 2.207e+02, threshold=2.961e+02, percent-clipped=0.0 2024-09-23 00:09:39,797 INFO [train.py:1198] (2/4) Epoch 8, batch 1350, loss[loss=0.2318, ctc_loss=0.1582, cr_loss=0.3677, over 16668.00 frames. ], tot_loss[loss=0.2689, ctc_loss=0.1901, cr_loss=0.3941, over 3349195.53 frames. ], batch size: 37, lr: 1.52e-02, grad_scale: 32.0 2024-09-23 00:09:41,615 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=133569.33333333334, ans=0.125 2024-09-23 00:09:52,054 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=133569.33333333334, ans=0.025 2024-09-23 00:10:09,715 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=133616.0, ans=0.1 2024-09-23 00:10:25,595 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=133662.66666666666, ans=0.125 2024-09-23 00:10:28,722 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=133709.33333333334, ans=0.125 2024-09-23 00:11:06,821 INFO [train.py:1198] (2/4) Epoch 8, batch 1400, loss[loss=0.2826, ctc_loss=0.1985, cr_loss=0.4202, over 17076.00 frames. ], tot_loss[loss=0.2695, ctc_loss=0.1906, cr_loss=0.3945, over 3342208.00 frames. ], batch size: 49, lr: 1.51e-02, grad_scale: 32.0 2024-09-23 00:11:13,486 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=133802.66666666666, ans=0.0 2024-09-23 00:11:24,464 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=133849.33333333334, ans=0.125 2024-09-23 00:11:45,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=133896.0, ans=0.125 2024-09-23 00:12:07,281 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=133942.66666666666, ans=0.0 2024-09-23 00:12:11,494 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=7.21 vs. limit=15.0 2024-09-23 00:12:13,682 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=133989.33333333334, ans=0.1 2024-09-23 00:12:15,586 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.90 vs. limit=6.0 2024-09-23 00:12:24,402 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.304e+02 1.414e+02 1.596e+02 2.535e+02, threshold=2.828e+02, percent-clipped=0.0 2024-09-23 00:12:25,993 INFO [train.py:1198] (2/4) Epoch 8, batch 1450, loss[loss=0.2842, ctc_loss=0.2022, cr_loss=0.4101, over 17015.00 frames. ], tot_loss[loss=0.2677, ctc_loss=0.1892, cr_loss=0.3926, over 3350180.41 frames. ], batch size: 52, lr: 1.51e-02, grad_scale: 32.0 2024-09-23 00:12:29,532 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=134036.0, ans=0.0 2024-09-23 00:12:31,195 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=134036.0, ans=0.1 2024-09-23 00:12:32,639 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=134036.0, ans=0.125 2024-09-23 00:12:48,714 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=134082.66666666666, ans=0.2 2024-09-23 00:13:48,010 INFO [train.py:1198] (2/4) Epoch 8, batch 1500, loss[loss=0.2731, ctc_loss=0.1927, cr_loss=0.402, over 17221.00 frames. ], tot_loss[loss=0.2663, ctc_loss=0.1882, cr_loss=0.3904, over 3356763.87 frames. ], batch size: 55, lr: 1.51e-02, grad_scale: 32.0 2024-09-23 00:13:50,022 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=134269.33333333334, ans=0.0 2024-09-23 00:14:01,881 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.whiten, num_groups=1, num_channels=192, metric=4.11 vs. limit=12.0 2024-09-23 00:15:09,201 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.316e+02 1.415e+02 1.558e+02 2.036e+02, threshold=2.830e+02, percent-clipped=0.0 2024-09-23 00:15:10,821 INFO [train.py:1198] (2/4) Epoch 8, batch 1550, loss[loss=0.2779, ctc_loss=0.199, cr_loss=0.3942, over 16116.00 frames. ], tot_loss[loss=0.2664, ctc_loss=0.1883, cr_loss=0.3905, over 3358107.79 frames. ], batch size: 74, lr: 1.51e-02, grad_scale: 32.0 2024-09-23 00:15:11,082 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=134502.66666666666, ans=0.125 2024-09-23 00:15:40,844 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=134549.33333333334, ans=0.1 2024-09-23 00:15:44,228 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=134596.0, ans=0.125 2024-09-23 00:16:13,934 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=134642.66666666666, ans=0.2 2024-09-23 00:16:18,671 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=134689.33333333334, ans=0.0 2024-09-23 00:16:28,116 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=134689.33333333334, ans=0.0 2024-09-23 00:16:35,827 INFO [train.py:1198] (2/4) Epoch 8, batch 1600, loss[loss=0.2872, ctc_loss=0.205, cr_loss=0.4106, over 17313.00 frames. ], tot_loss[loss=0.2656, ctc_loss=0.1876, cr_loss=0.3901, over 3361037.98 frames. ], batch size: 51, lr: 1.51e-02, grad_scale: 32.0 2024-09-23 00:16:51,912 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=134782.66666666666, ans=0.2 2024-09-23 00:17:07,001 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=14.24 vs. limit=22.5 2024-09-23 00:17:35,299 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=134876.0, ans=0.125 2024-09-23 00:17:46,594 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 00:17:54,116 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.144e+02 1.357e+02 1.515e+02 1.820e+02 2.997e+02, threshold=3.030e+02, percent-clipped=2.0 2024-09-23 00:17:54,431 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=134969.33333333334, ans=0.2 2024-09-23 00:17:55,797 INFO [train.py:1198] (2/4) Epoch 8, batch 1650, loss[loss=0.2678, ctc_loss=0.1887, cr_loss=0.3954, over 17041.00 frames. ], tot_loss[loss=0.2642, ctc_loss=0.1864, cr_loss=0.3891, over 3364776.10 frames. ], batch size: 52, lr: 1.51e-02, grad_scale: 32.0 2024-09-23 00:18:04,555 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=21.49 vs. limit=22.5 2024-09-23 00:18:29,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=135062.66666666666, ans=0.0 2024-09-23 00:18:36,999 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=135062.66666666666, ans=0.125 2024-09-23 00:19:00,295 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=135156.0, ans=0.0 2024-09-23 00:19:16,239 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=135202.66666666666, ans=0.2 2024-09-23 00:19:17,537 INFO [train.py:1198] (2/4) Epoch 8, batch 1700, loss[loss=0.2958, ctc_loss=0.211, cr_loss=0.424, over 17042.00 frames. ], tot_loss[loss=0.264, ctc_loss=0.1862, cr_loss=0.3889, over 3372412.66 frames. ], batch size: 52, lr: 1.51e-02, grad_scale: 32.0 2024-09-23 00:19:19,741 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.27 vs. limit=15.0 2024-09-23 00:19:20,970 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=135202.66666666666, ans=0.0 2024-09-23 00:19:23,993 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=135202.66666666666, ans=0.125 2024-09-23 00:19:54,807 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=135296.0, ans=0.0 2024-09-23 00:20:01,268 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=135296.0, ans=0.125 2024-09-23 00:20:14,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=135342.66666666666, ans=0.0 2024-09-23 00:20:20,752 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=135342.66666666666, ans=0.0 2024-09-23 00:20:37,544 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=135389.33333333334, ans=0.125 2024-09-23 00:20:40,291 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.382e+02 1.518e+02 1.704e+02 3.525e+02, threshold=3.037e+02, percent-clipped=1.0 2024-09-23 00:20:41,977 INFO [train.py:1198] (2/4) Epoch 8, batch 1750, loss[loss=0.235, ctc_loss=0.166, cr_loss=0.3453, over 17099.00 frames. ], tot_loss[loss=0.2655, ctc_loss=0.1872, cr_loss=0.3914, over 3375297.74 frames. ], batch size: 49, lr: 1.51e-02, grad_scale: 32.0 2024-09-23 00:21:03,949 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=135482.66666666666, ans=0.0 2024-09-23 00:21:05,534 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=135482.66666666666, ans=0.125 2024-09-23 00:21:10,729 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.96 vs. limit=15.0 2024-09-23 00:21:13,430 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=135482.66666666666, ans=0.1 2024-09-23 00:21:21,154 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=135529.33333333334, ans=0.2 2024-09-23 00:21:33,113 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.31 vs. limit=12.0 2024-09-23 00:21:37,345 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=135576.0, ans=0.09899494936611666 2024-09-23 00:21:46,665 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=135622.66666666666, ans=0.0 2024-09-23 00:22:03,865 INFO [train.py:1198] (2/4) Epoch 8, batch 1800, loss[loss=0.2423, ctc_loss=0.1683, cr_loss=0.3702, over 17251.00 frames. ], tot_loss[loss=0.2647, ctc_loss=0.1865, cr_loss=0.391, over 3369887.02 frames. ], batch size: 44, lr: 1.50e-02, grad_scale: 32.0 2024-09-23 00:22:48,568 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=135762.66666666666, ans=0.2 2024-09-23 00:23:12,245 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=135856.0, ans=0.0 2024-09-23 00:23:21,450 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.146e+02 1.352e+02 1.514e+02 1.696e+02 3.020e+02, threshold=3.028e+02, percent-clipped=0.0 2024-09-23 00:23:23,111 INFO [train.py:1198] (2/4) Epoch 8, batch 1850, loss[loss=0.2776, ctc_loss=0.1946, cr_loss=0.4153, over 17285.00 frames. ], tot_loss[loss=0.2659, ctc_loss=0.1876, cr_loss=0.3917, over 3352498.25 frames. ], batch size: 49, lr: 1.50e-02, grad_scale: 32.0 2024-09-23 00:23:31,829 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=5.62 vs. limit=15.0 2024-09-23 00:24:47,780 INFO [train.py:1198] (2/4) Epoch 8, batch 1900, loss[loss=0.3153, ctc_loss=0.2253, cr_loss=0.4502, over 16989.00 frames. ], tot_loss[loss=0.2655, ctc_loss=0.1872, cr_loss=0.3913, over 3361828.70 frames. ], batch size: 53, lr: 1.50e-02, grad_scale: 32.0 2024-09-23 00:24:51,943 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=3.22 vs. limit=15.0 2024-09-23 00:25:13,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=136182.66666666666, ans=0.2 2024-09-23 00:25:34,507 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=7.10 vs. limit=15.0 2024-09-23 00:26:06,540 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=136322.66666666666, ans=0.0 2024-09-23 00:26:11,059 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.199e+02 1.376e+02 1.564e+02 1.775e+02 2.559e+02, threshold=3.128e+02, percent-clipped=0.0 2024-09-23 00:26:12,646 INFO [train.py:1198] (2/4) Epoch 8, batch 1950, loss[loss=0.308, ctc_loss=0.2274, cr_loss=0.4033, over 11838.00 frames. ], tot_loss[loss=0.2666, ctc_loss=0.188, cr_loss=0.3928, over 3358233.42 frames. ], batch size: 123, lr: 1.50e-02, grad_scale: 32.0 2024-09-23 00:26:21,002 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=136369.33333333334, ans=0.125 2024-09-23 00:26:43,718 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.whiten.whitening_limit, batch_count=136462.66666666666, ans=12.0 2024-09-23 00:26:59,096 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=136509.33333333334, ans=0.125 2024-09-23 00:27:03,889 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=136509.33333333334, ans=0.125 2024-09-23 00:27:15,218 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=136556.0, ans=0.1 2024-09-23 00:27:32,278 INFO [train.py:1198] (2/4) Epoch 8, batch 2000, loss[loss=0.2843, ctc_loss=0.2036, cr_loss=0.4035, over 17007.00 frames. ], tot_loss[loss=0.2665, ctc_loss=0.188, cr_loss=0.3927, over 3364069.84 frames. ], batch size: 51, lr: 1.50e-02, grad_scale: 32.0 2024-09-23 00:27:43,774 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=136602.66666666666, ans=0.125 2024-09-23 00:28:38,131 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=136789.33333333334, ans=0.125 2024-09-23 00:28:53,819 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.191e+02 1.381e+02 1.478e+02 1.758e+02 2.844e+02, threshold=2.957e+02, percent-clipped=0.0 2024-09-23 00:28:55,407 INFO [train.py:1198] (2/4) Epoch 8, batch 2050, loss[loss=0.2486, ctc_loss=0.1691, cr_loss=0.3973, over 17305.00 frames. ], tot_loss[loss=0.2662, ctc_loss=0.1877, cr_loss=0.3928, over 3361868.10 frames. ], batch size: 46, lr: 1.50e-02, grad_scale: 32.0 2024-09-23 00:29:11,313 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=136882.66666666666, ans=0.035 2024-09-23 00:29:47,261 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=136976.0, ans=0.1 2024-09-23 00:30:16,953 INFO [train.py:1198] (2/4) Epoch 8, batch 2100, loss[loss=0.274, ctc_loss=0.1931, cr_loss=0.4043, over 17035.00 frames. ], tot_loss[loss=0.2658, ctc_loss=0.1873, cr_loss=0.3925, over 3362687.44 frames. ], batch size: 51, lr: 1.50e-02, grad_scale: 32.0 2024-09-23 00:30:21,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=137069.33333333334, ans=0.125 2024-09-23 00:30:27,940 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=137069.33333333334, ans=0.1 2024-09-23 00:30:34,065 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=137116.0, ans=0.1 2024-09-23 00:31:12,143 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=137209.33333333334, ans=0.125 2024-09-23 00:31:13,818 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=137209.33333333334, ans=0.125 2024-09-23 00:31:22,325 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=11.83 vs. limit=15.0 2024-09-23 00:31:31,405 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=137256.0, ans=0.125 2024-09-23 00:31:40,396 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.168e+02 1.375e+02 1.520e+02 1.695e+02 2.810e+02, threshold=3.041e+02, percent-clipped=0.0 2024-09-23 00:31:41,916 INFO [train.py:1198] (2/4) Epoch 8, batch 2150, loss[loss=0.2333, ctc_loss=0.1657, cr_loss=0.3378, over 17075.00 frames. ], tot_loss[loss=0.2651, ctc_loss=0.1868, cr_loss=0.3914, over 3360035.81 frames. ], batch size: 40, lr: 1.50e-02, grad_scale: 32.0 2024-09-23 00:31:55,024 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=137302.66666666666, ans=0.1 2024-09-23 00:32:05,361 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=7.62 vs. limit=15.0 2024-09-23 00:32:22,749 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=10.32 vs. limit=15.0 2024-09-23 00:33:00,867 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=11.18 vs. limit=22.5 2024-09-23 00:33:01,804 INFO [train.py:1198] (2/4) Epoch 8, batch 2200, loss[loss=0.2625, ctc_loss=0.1835, cr_loss=0.3951, over 17100.00 frames. ], tot_loss[loss=0.2658, ctc_loss=0.1874, cr_loss=0.392, over 3353948.50 frames. ], batch size: 49, lr: 1.49e-02, grad_scale: 32.0 2024-09-23 00:33:11,537 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=137536.0, ans=0.0 2024-09-23 00:34:08,737 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.24 vs. limit=12.0 2024-09-23 00:34:21,855 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.166e+02 1.429e+02 1.621e+02 1.891e+02 2.462e+02, threshold=3.242e+02, percent-clipped=0.0 2024-09-23 00:34:23,498 INFO [train.py:1198] (2/4) Epoch 8, batch 2250, loss[loss=0.3075, ctc_loss=0.2129, cr_loss=0.4733, over 16407.00 frames. ], tot_loss[loss=0.2673, ctc_loss=0.1886, cr_loss=0.3934, over 3343279.89 frames. ], batch size: 66, lr: 1.49e-02, grad_scale: 32.0 2024-09-23 00:34:51,984 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=137816.0, ans=0.1 2024-09-23 00:35:37,615 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=137956.0, ans=0.2 2024-09-23 00:35:50,811 INFO [train.py:1198] (2/4) Epoch 8, batch 2300, loss[loss=0.2611, ctc_loss=0.1809, cr_loss=0.4014, over 17141.00 frames. ], tot_loss[loss=0.2661, ctc_loss=0.1877, cr_loss=0.3919, over 3343131.49 frames. ], batch size: 48, lr: 1.49e-02, grad_scale: 32.0 2024-09-23 00:35:51,144 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=138002.66666666666, ans=0.1 2024-09-23 00:35:51,178 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=138002.66666666666, ans=0.0 2024-09-23 00:35:52,609 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=138002.66666666666, ans=0.125 2024-09-23 00:36:00,832 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=138002.66666666666, ans=0.5 2024-09-23 00:36:00,918 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=138002.66666666666, ans=0.125 2024-09-23 00:36:15,326 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=138049.33333333334, ans=0.0 2024-09-23 00:36:18,498 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=138049.33333333334, ans=0.1 2024-09-23 00:36:19,987 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=138049.33333333334, ans=0.125 2024-09-23 00:36:40,892 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=138142.66666666666, ans=0.1 2024-09-23 00:36:40,935 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=138142.66666666666, ans=0.0 2024-09-23 00:36:48,754 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=138142.66666666666, ans=0.0 2024-09-23 00:36:50,955 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=5.61 vs. limit=15.0 2024-09-23 00:37:04,688 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=138189.33333333334, ans=0.125 2024-09-23 00:37:09,110 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.147e+02 1.383e+02 1.498e+02 1.682e+02 2.493e+02, threshold=2.995e+02, percent-clipped=0.0 2024-09-23 00:37:10,793 INFO [train.py:1198] (2/4) Epoch 8, batch 2350, loss[loss=0.2857, ctc_loss=0.2003, cr_loss=0.4272, over 17057.00 frames. ], tot_loss[loss=0.2655, ctc_loss=0.1872, cr_loss=0.3914, over 3352164.77 frames. ], batch size: 52, lr: 1.49e-02, grad_scale: 32.0 2024-09-23 00:37:12,503 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=138236.0, ans=0.125 2024-09-23 00:37:18,953 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=138236.0, ans=0.1 2024-09-23 00:37:25,580 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=138282.66666666666, ans=0.125 2024-09-23 00:37:46,942 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=7.81 vs. limit=15.0 2024-09-23 00:37:49,622 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=138329.33333333334, ans=0.1 2024-09-23 00:38:27,238 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.98 vs. limit=22.5 2024-09-23 00:38:32,407 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.02 vs. limit=15.0 2024-09-23 00:38:33,070 INFO [train.py:1198] (2/4) Epoch 8, batch 2400, loss[loss=0.2396, ctc_loss=0.1646, cr_loss=0.3748, over 17088.00 frames. ], tot_loss[loss=0.2651, ctc_loss=0.187, cr_loss=0.3903, over 3344645.77 frames. ], batch size: 43, lr: 1.49e-02, grad_scale: 32.0 2024-09-23 00:38:35,622 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=4.53 vs. limit=15.0 2024-09-23 00:38:39,773 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=138469.33333333334, ans=0.125 2024-09-23 00:38:42,866 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=138469.33333333334, ans=0.0 2024-09-23 00:38:44,587 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=138469.33333333334, ans=0.0 2024-09-23 00:39:10,000 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=138562.66666666666, ans=0.0 2024-09-23 00:39:18,200 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=138562.66666666666, ans=0.1 2024-09-23 00:39:27,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=138609.33333333334, ans=0.125 2024-09-23 00:39:29,733 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=12.96 vs. limit=15.0 2024-09-23 00:39:53,485 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.149e+02 1.369e+02 1.492e+02 1.774e+02 2.629e+02, threshold=2.985e+02, percent-clipped=0.0 2024-09-23 00:39:53,886 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=138702.66666666666, ans=0.125 2024-09-23 00:39:55,152 INFO [train.py:1198] (2/4) Epoch 8, batch 2450, loss[loss=0.2661, ctc_loss=0.1861, cr_loss=0.3997, over 17139.00 frames. ], tot_loss[loss=0.2659, ctc_loss=0.1876, cr_loss=0.3915, over 3346165.65 frames. ], batch size: 48, lr: 1.49e-02, grad_scale: 32.0 2024-09-23 00:40:14,028 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=138749.33333333334, ans=0.0 2024-09-23 00:40:23,593 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=138749.33333333334, ans=0.0 2024-09-23 00:40:27,714 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.whiten, num_groups=1, num_channels=192, metric=4.06 vs. limit=12.0 2024-09-23 00:40:34,873 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=138796.0, ans=0.025 2024-09-23 00:40:50,263 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=138842.66666666666, ans=0.125 2024-09-23 00:41:06,232 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=138889.33333333334, ans=0.125 2024-09-23 00:41:20,429 INFO [train.py:1198] (2/4) Epoch 8, batch 2500, loss[loss=0.2657, ctc_loss=0.1881, cr_loss=0.388, over 16976.00 frames. ], tot_loss[loss=0.2654, ctc_loss=0.1872, cr_loss=0.3905, over 3357239.08 frames. ], batch size: 53, lr: 1.49e-02, grad_scale: 32.0 2024-09-23 00:41:54,724 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.69 vs. limit=10.0 2024-09-23 00:42:11,929 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=139076.0, ans=0.05 2024-09-23 00:42:18,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=139076.0, ans=0.125 2024-09-23 00:42:38,761 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=139169.33333333334, ans=0.0 2024-09-23 00:42:40,013 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.162e+02 1.309e+02 1.454e+02 1.673e+02 2.570e+02, threshold=2.909e+02, percent-clipped=0.0 2024-09-23 00:42:40,037 INFO [train.py:1198] (2/4) Epoch 8, batch 2550, loss[loss=0.2675, ctc_loss=0.1829, cr_loss=0.423, over 17208.00 frames. ], tot_loss[loss=0.2658, ctc_loss=0.1875, cr_loss=0.3913, over 3358975.04 frames. ], batch size: 47, lr: 1.49e-02, grad_scale: 16.0 2024-09-23 00:42:43,920 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.05 vs. limit=15.0 2024-09-23 00:43:09,227 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=139216.0, ans=0.125 2024-09-23 00:43:11,363 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=5.17 vs. limit=10.0 2024-09-23 00:43:19,406 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=139262.66666666666, ans=0.2 2024-09-23 00:43:24,244 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=139262.66666666666, ans=0.2 2024-09-23 00:43:27,456 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=139262.66666666666, ans=0.125 2024-09-23 00:44:02,586 INFO [train.py:1198] (2/4) Epoch 8, batch 2600, loss[loss=0.2466, ctc_loss=0.168, cr_loss=0.3926, over 17019.00 frames. ], tot_loss[loss=0.2651, ctc_loss=0.187, cr_loss=0.3905, over 3357926.33 frames. ], batch size: 44, lr: 1.48e-02, grad_scale: 16.0 2024-09-23 00:44:06,161 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=139402.66666666666, ans=0.0 2024-09-23 00:44:18,695 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=139449.33333333334, ans=0.125 2024-09-23 00:44:19,445 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=4.10 vs. limit=15.0 2024-09-23 00:44:51,739 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=139542.66666666666, ans=0.0 2024-09-23 00:45:04,441 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=139542.66666666666, ans=0.125 2024-09-23 00:45:08,031 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.81 vs. limit=15.0 2024-09-23 00:45:27,487 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.164e+02 1.349e+02 1.476e+02 1.822e+02 2.637e+02, threshold=2.952e+02, percent-clipped=0.0 2024-09-23 00:45:27,512 INFO [train.py:1198] (2/4) Epoch 8, batch 2650, loss[loss=0.2675, ctc_loss=0.1945, cr_loss=0.365, over 16721.00 frames. ], tot_loss[loss=0.2659, ctc_loss=0.1876, cr_loss=0.3915, over 3358006.95 frames. ], batch size: 61, lr: 1.48e-02, grad_scale: 16.0 2024-09-23 00:45:35,033 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=9.77 vs. limit=15.0 2024-09-23 00:45:52,415 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=139682.66666666666, ans=0.2 2024-09-23 00:46:18,192 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=139776.0, ans=0.2 2024-09-23 00:46:24,697 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.min_abs, batch_count=139776.0, ans=0.5 2024-09-23 00:46:49,529 INFO [train.py:1198] (2/4) Epoch 8, batch 2700, loss[loss=0.2564, ctc_loss=0.1755, cr_loss=0.4046, over 17169.00 frames. ], tot_loss[loss=0.2669, ctc_loss=0.1884, cr_loss=0.3925, over 3357162.25 frames. ], batch size: 45, lr: 1.48e-02, grad_scale: 16.0 2024-09-23 00:47:04,560 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=6.07 vs. limit=15.0 2024-09-23 00:47:31,724 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=9.01 vs. limit=15.0 2024-09-23 00:47:59,878 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=140056.0, ans=0.125 2024-09-23 00:48:01,754 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=140056.0, ans=0.0 2024-09-23 00:48:11,966 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.163e+02 1.374e+02 1.576e+02 1.752e+02 3.056e+02, threshold=3.153e+02, percent-clipped=1.0 2024-09-23 00:48:11,990 INFO [train.py:1198] (2/4) Epoch 8, batch 2750, loss[loss=0.2615, ctc_loss=0.1842, cr_loss=0.3867, over 17303.00 frames. ], tot_loss[loss=0.2658, ctc_loss=0.1877, cr_loss=0.3908, over 3358835.03 frames. ], batch size: 46, lr: 1.48e-02, grad_scale: 16.0 2024-09-23 00:48:20,256 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=140102.66666666666, ans=0.125 2024-09-23 00:48:23,586 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=140102.66666666666, ans=0.0 2024-09-23 00:49:03,393 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=140242.66666666666, ans=0.125 2024-09-23 00:49:19,126 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=140289.33333333334, ans=0.125 2024-09-23 00:49:31,646 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.whiten, num_groups=1, num_channels=512, metric=4.64 vs. limit=12.0 2024-09-23 00:49:34,152 INFO [train.py:1198] (2/4) Epoch 8, batch 2800, loss[loss=0.2495, ctc_loss=0.1727, cr_loss=0.3841, over 17216.00 frames. ], tot_loss[loss=0.2664, ctc_loss=0.1881, cr_loss=0.3916, over 3354665.15 frames. ], batch size: 50, lr: 1.48e-02, grad_scale: 32.0 2024-09-23 00:50:13,908 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=140429.33333333334, ans=0.0 2024-09-23 00:50:23,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=140476.0, ans=0.125 2024-09-23 00:50:25,124 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=11.09 vs. limit=15.0 2024-09-23 00:50:44,993 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=140522.66666666666, ans=0.0 2024-09-23 00:50:58,772 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.146e+02 1.406e+02 1.667e+02 1.997e+02 3.785e+02, threshold=3.334e+02, percent-clipped=1.0 2024-09-23 00:50:58,797 INFO [train.py:1198] (2/4) Epoch 8, batch 2850, loss[loss=0.2428, ctc_loss=0.1698, cr_loss=0.3649, over 16955.00 frames. ], tot_loss[loss=0.2648, ctc_loss=0.1868, cr_loss=0.3903, over 3361715.04 frames. ], batch size: 42, lr: 1.48e-02, grad_scale: 32.0 2024-09-23 00:51:38,922 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=140662.66666666666, ans=0.125 2024-09-23 00:51:43,780 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=140662.66666666666, ans=0.125 2024-09-23 00:51:46,906 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=140709.33333333334, ans=0.025 2024-09-23 00:51:58,325 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=140709.33333333334, ans=0.1 2024-09-23 00:52:12,742 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=140756.0, ans=0.0 2024-09-23 00:52:17,516 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=140802.66666666666, ans=0.125 2024-09-23 00:52:18,841 INFO [train.py:1198] (2/4) Epoch 8, batch 2900, loss[loss=0.2425, ctc_loss=0.1681, cr_loss=0.3722, over 17269.00 frames. ], tot_loss[loss=0.2653, ctc_loss=0.1871, cr_loss=0.3911, over 3370337.59 frames. ], batch size: 44, lr: 1.48e-02, grad_scale: 32.0 2024-09-23 00:52:28,880 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=140802.66666666666, ans=0.0 2024-09-23 00:53:07,493 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=140942.66666666666, ans=0.0 2024-09-23 00:53:22,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=140942.66666666666, ans=0.125 2024-09-23 00:53:24,233 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=140989.33333333334, ans=0.2 2024-09-23 00:53:25,683 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=140989.33333333334, ans=0.125 2024-09-23 00:53:29,160 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=4.71 vs. limit=12.0 2024-09-23 00:53:39,784 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=141036.0, ans=0.1 2024-09-23 00:53:41,169 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.147e+02 1.334e+02 1.488e+02 1.595e+02 2.196e+02, threshold=2.977e+02, percent-clipped=0.0 2024-09-23 00:53:41,193 INFO [train.py:1198] (2/4) Epoch 8, batch 2950, loss[loss=0.2673, ctc_loss=0.1874, cr_loss=0.3994, over 17324.00 frames. ], tot_loss[loss=0.2641, ctc_loss=0.1861, cr_loss=0.3896, over 3366423.05 frames. ], batch size: 51, lr: 1.48e-02, grad_scale: 32.0 2024-09-23 00:53:50,966 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=141036.0, ans=0.1 2024-09-23 00:53:53,204 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.90 vs. limit=12.0 2024-09-23 00:54:01,987 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=141082.66666666666, ans=0.1 2024-09-23 00:54:07,652 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=11.63 vs. limit=15.0 2024-09-23 00:54:13,657 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten.whitening_limit, batch_count=141129.33333333334, ans=15.0 2024-09-23 00:54:55,208 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=141222.66666666666, ans=0.125 2024-09-23 00:55:02,757 INFO [train.py:1198] (2/4) Epoch 8, batch 3000, loss[loss=0.2595, ctc_loss=0.1827, cr_loss=0.3839, over 16994.00 frames. ], tot_loss[loss=0.2637, ctc_loss=0.186, cr_loss=0.3885, over 3367737.16 frames. ], batch size: 51, lr: 1.48e-02, grad_scale: 32.0 2024-09-23 00:55:02,757 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 00:55:18,807 INFO [train.py:1230] (2/4) Epoch 8, validation: loss=0.05304, ctc_loss=0.05304, cr_loss=7.247e-15, over 944034.00 frames. 2024-09-23 00:55:18,808 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 00:56:19,833 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=141409.33333333334, ans=0.125 2024-09-23 00:56:20,891 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=8.38 vs. limit=15.0 2024-09-23 00:56:39,932 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.352e+02 1.499e+02 1.786e+02 3.736e+02, threshold=2.997e+02, percent-clipped=4.0 2024-09-23 00:56:39,956 INFO [train.py:1198] (2/4) Epoch 8, batch 3050, loss[loss=0.2955, ctc_loss=0.2125, cr_loss=0.415, over 17204.00 frames. ], tot_loss[loss=0.263, ctc_loss=0.1853, cr_loss=0.3886, over 3364967.16 frames. ], batch size: 55, lr: 1.47e-02, grad_scale: 32.0 2024-09-23 00:57:12,017 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=141596.0, ans=0.0 2024-09-23 00:57:16,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=141596.0, ans=0.025 2024-09-23 00:57:41,750 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=141689.33333333334, ans=0.025 2024-09-23 00:57:42,095 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.23 vs. limit=6.0 2024-09-23 00:57:58,665 INFO [train.py:1198] (2/4) Epoch 8, batch 3100, loss[loss=0.303, ctc_loss=0.2164, cr_loss=0.4331, over 17240.00 frames. ], tot_loss[loss=0.2628, ctc_loss=0.1851, cr_loss=0.3882, over 3370668.02 frames. ], batch size: 50, lr: 1.47e-02, grad_scale: 16.0 2024-09-23 00:58:22,184 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=141782.66666666666, ans=0.2 2024-09-23 00:58:34,378 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=141829.33333333334, ans=0.0 2024-09-23 00:59:04,206 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=141922.66666666666, ans=0.025 2024-09-23 00:59:05,111 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module1.whiten, num_groups=1, num_channels=192, metric=6.89 vs. limit=15.0 2024-09-23 00:59:10,448 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=141922.66666666666, ans=0.125 2024-09-23 00:59:10,553 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=141922.66666666666, ans=0.2 2024-09-23 00:59:16,306 INFO [train.py:1198] (2/4) Epoch 8, batch 3150, loss[loss=0.243, ctc_loss=0.1685, cr_loss=0.3722, over 17045.00 frames. ], tot_loss[loss=0.2625, ctc_loss=0.1849, cr_loss=0.3879, over 3366872.75 frames. ], batch size: 46, lr: 1.47e-02, grad_scale: 16.0 2024-09-23 00:59:17,877 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.325e+02 1.474e+02 1.672e+02 3.223e+02, threshold=2.948e+02, percent-clipped=1.0 2024-09-23 00:59:45,545 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.50 vs. limit=15.0 2024-09-23 00:59:57,436 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=142062.66666666666, ans=0.0 2024-09-23 01:00:09,643 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=142109.33333333334, ans=0.0 2024-09-23 01:00:09,832 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=142109.33333333334, ans=0.125 2024-09-23 01:00:11,250 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=142109.33333333334, ans=0.0 2024-09-23 01:00:29,838 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=142156.0, ans=0.1 2024-09-23 01:00:30,595 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=6.19 vs. limit=12.0 2024-09-23 01:00:34,386 INFO [train.py:1198] (2/4) Epoch 8, batch 3200, loss[loss=0.2524, ctc_loss=0.1781, cr_loss=0.3718, over 17215.00 frames. ], tot_loss[loss=0.2619, ctc_loss=0.1845, cr_loss=0.3871, over 3361705.20 frames. ], batch size: 47, lr: 1.47e-02, grad_scale: 32.0 2024-09-23 01:00:37,833 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=142202.66666666666, ans=0.125 2024-09-23 01:00:47,150 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 01:00:55,020 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=142249.33333333334, ans=0.125 2024-09-23 01:01:36,952 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=142389.33333333334, ans=0.2 2024-09-23 01:01:38,723 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=142389.33333333334, ans=0.0 2024-09-23 01:01:46,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=142389.33333333334, ans=0.125 2024-09-23 01:01:54,764 INFO [train.py:1198] (2/4) Epoch 8, batch 3250, loss[loss=0.2713, ctc_loss=0.1865, cr_loss=0.4239, over 17183.00 frames. ], tot_loss[loss=0.2637, ctc_loss=0.1857, cr_loss=0.3899, over 3362130.68 frames. ], batch size: 45, lr: 1.47e-02, grad_scale: 32.0 2024-09-23 01:01:56,369 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.501e+02 1.666e+02 1.949e+02 2.835e+02, threshold=3.332e+02, percent-clipped=0.0 2024-09-23 01:02:01,160 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=142436.0, ans=0.07 2024-09-23 01:02:18,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=142482.66666666666, ans=0.125 2024-09-23 01:02:23,107 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=142482.66666666666, ans=0.125 2024-09-23 01:02:36,734 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=142529.33333333334, ans=0.125 2024-09-23 01:02:53,960 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=142576.0, ans=0.125 2024-09-23 01:03:12,271 INFO [train.py:1198] (2/4) Epoch 8, batch 3300, loss[loss=0.226, ctc_loss=0.1591, cr_loss=0.3345, over 16964.00 frames. ], tot_loss[loss=0.263, ctc_loss=0.1851, cr_loss=0.3894, over 3363688.87 frames. ], batch size: 42, lr: 1.47e-02, grad_scale: 32.0 2024-09-23 01:03:35,953 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=142716.0, ans=0.1 2024-09-23 01:03:42,606 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=9.41 vs. limit=15.0 2024-09-23 01:03:51,615 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=142762.66666666666, ans=0.125 2024-09-23 01:04:05,071 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=8.81 vs. limit=15.0 2024-09-23 01:04:24,929 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.11 vs. limit=10.0 2024-09-23 01:04:30,462 INFO [train.py:1198] (2/4) Epoch 8, batch 3350, loss[loss=0.247, ctc_loss=0.1729, cr_loss=0.3703, over 17192.00 frames. ], tot_loss[loss=0.2626, ctc_loss=0.1848, cr_loss=0.3891, over 3360932.51 frames. ], batch size: 41, lr: 1.47e-02, grad_scale: 16.0 2024-09-23 01:04:30,792 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=142902.66666666666, ans=0.0 2024-09-23 01:04:33,559 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.197e+02 1.393e+02 1.569e+02 1.774e+02 3.394e+02, threshold=3.137e+02, percent-clipped=1.0 2024-09-23 01:04:33,694 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=142902.66666666666, ans=0.125 2024-09-23 01:04:46,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=142949.33333333334, ans=0.125 2024-09-23 01:04:54,332 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=142949.33333333334, ans=0.125 2024-09-23 01:05:21,524 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=143042.66666666666, ans=0.125 2024-09-23 01:05:40,025 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=143089.33333333334, ans=0.2 2024-09-23 01:05:52,730 INFO [train.py:1198] (2/4) Epoch 8, batch 3400, loss[loss=0.2744, ctc_loss=0.1964, cr_loss=0.3901, over 17287.00 frames. ], tot_loss[loss=0.2642, ctc_loss=0.186, cr_loss=0.3907, over 3361942.12 frames. ], batch size: 49, lr: 1.47e-02, grad_scale: 16.0 2024-09-23 01:05:59,268 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=143136.0, ans=0.0 2024-09-23 01:06:01,270 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.95 vs. limit=6.0 2024-09-23 01:06:19,260 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=143182.66666666666, ans=0.125 2024-09-23 01:06:25,589 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=143229.33333333334, ans=0.125 2024-09-23 01:06:33,694 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.22 vs. limit=15.0 2024-09-23 01:07:06,396 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=143322.66666666666, ans=0.1 2024-09-23 01:07:12,535 INFO [train.py:1198] (2/4) Epoch 8, batch 3450, loss[loss=0.2876, ctc_loss=0.2049, cr_loss=0.4134, over 17137.00 frames. ], tot_loss[loss=0.265, ctc_loss=0.1866, cr_loss=0.3919, over 3359763.84 frames. ], batch size: 48, lr: 1.47e-02, grad_scale: 16.0 2024-09-23 01:07:15,656 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.386e+02 1.521e+02 1.778e+02 2.541e+02, threshold=3.041e+02, percent-clipped=0.0 2024-09-23 01:07:25,301 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=143369.33333333334, ans=0.125 2024-09-23 01:07:25,399 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=143369.33333333334, ans=0.125 2024-09-23 01:07:37,888 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=143416.0, ans=0.125 2024-09-23 01:07:38,367 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=12.31 vs. limit=15.0 2024-09-23 01:07:41,090 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=143416.0, ans=0.125 2024-09-23 01:07:44,314 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=143462.66666666666, ans=0.1 2024-09-23 01:07:52,090 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=143462.66666666666, ans=0.2 2024-09-23 01:07:52,824 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=11.22 vs. limit=15.0 2024-09-23 01:07:59,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=143509.33333333334, ans=0.125 2024-09-23 01:08:30,382 INFO [train.py:1198] (2/4) Epoch 8, batch 3500, loss[loss=0.2295, ctc_loss=0.1571, cr_loss=0.3617, over 17171.00 frames. ], tot_loss[loss=0.2654, ctc_loss=0.187, cr_loss=0.392, over 3346410.86 frames. ], batch size: 45, lr: 1.46e-02, grad_scale: 16.0 2024-09-23 01:08:39,927 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=143602.66666666666, ans=0.05 2024-09-23 01:09:05,854 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=3.33 vs. limit=15.0 2024-09-23 01:09:38,087 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.max_positive, batch_count=143789.33333333334, ans=0.95 2024-09-23 01:09:44,773 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=14.62 vs. limit=22.5 2024-09-23 01:09:48,668 INFO [train.py:1198] (2/4) Epoch 8, batch 3550, loss[loss=0.264, ctc_loss=0.1848, cr_loss=0.3958, over 17198.00 frames. ], tot_loss[loss=0.2658, ctc_loss=0.1873, cr_loss=0.3924, over 3342249.89 frames. ], batch size: 47, lr: 1.46e-02, grad_scale: 16.0 2024-09-23 01:09:51,761 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.159e+02 1.323e+02 1.423e+02 1.598e+02 2.580e+02, threshold=2.846e+02, percent-clipped=0.0 2024-09-23 01:10:13,798 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=143882.66666666666, ans=0.2 2024-09-23 01:10:33,878 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=143976.0, ans=0.2 2024-09-23 01:10:43,856 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=7.56 vs. limit=15.0 2024-09-23 01:10:54,295 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.35 vs. limit=15.0 2024-09-23 01:11:04,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=144069.33333333334, ans=0.125 2024-09-23 01:11:06,357 INFO [train.py:1198] (2/4) Epoch 8, batch 3600, loss[loss=0.2642, ctc_loss=0.1845, cr_loss=0.3985, over 16977.00 frames. ], tot_loss[loss=0.2654, ctc_loss=0.1869, cr_loss=0.3923, over 3347498.26 frames. ], batch size: 53, lr: 1.46e-02, grad_scale: 32.0 2024-09-23 01:11:29,303 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=144116.0, ans=0.1 2024-09-23 01:11:30,591 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=144116.0, ans=0.125 2024-09-23 01:11:38,700 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=144162.66666666666, ans=0.2 2024-09-23 01:11:49,788 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=144162.66666666666, ans=0.0 2024-09-23 01:12:16,405 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=144256.0, ans=0.125 2024-09-23 01:12:27,261 INFO [train.py:1198] (2/4) Epoch 8, batch 3650, loss[loss=0.3112, ctc_loss=0.2346, cr_loss=0.3831, over 12263.00 frames. ], tot_loss[loss=0.2636, ctc_loss=0.1857, cr_loss=0.3895, over 3351774.65 frames. ], batch size: 123, lr: 1.46e-02, grad_scale: 32.0 2024-09-23 01:12:30,476 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.372e+02 1.519e+02 1.739e+02 2.361e+02, threshold=3.037e+02, percent-clipped=0.0 2024-09-23 01:12:43,241 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=144349.33333333334, ans=0.125 2024-09-23 01:12:47,149 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=13.96 vs. limit=22.5 2024-09-23 01:13:07,821 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=144396.0, ans=0.2 2024-09-23 01:13:28,788 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=144489.33333333334, ans=0.2 2024-09-23 01:13:45,448 INFO [train.py:1198] (2/4) Epoch 8, batch 3700, loss[loss=0.2839, ctc_loss=0.2008, cr_loss=0.4159, over 17264.00 frames. ], tot_loss[loss=0.2642, ctc_loss=0.1861, cr_loss=0.3902, over 3350349.64 frames. ], batch size: 44, lr: 1.46e-02, grad_scale: 32.0 2024-09-23 01:14:09,639 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=10.86 vs. limit=15.0 2024-09-23 01:14:12,186 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 01:14:41,150 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module1.whiten, num_groups=1, num_channels=192, metric=8.99 vs. limit=15.0 2024-09-23 01:14:51,665 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=144722.66666666666, ans=0.0 2024-09-23 01:14:57,804 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=144722.66666666666, ans=0.0 2024-09-23 01:15:03,784 INFO [train.py:1198] (2/4) Epoch 8, batch 3750, loss[loss=0.2102, ctc_loss=0.1467, cr_loss=0.3175, over 17204.00 frames. ], tot_loss[loss=0.2632, ctc_loss=0.1854, cr_loss=0.3894, over 3351152.97 frames. ], batch size: 41, lr: 1.46e-02, grad_scale: 32.0 2024-09-23 01:15:03,971 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=144769.33333333334, ans=0.125 2024-09-23 01:15:07,863 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.153e+02 1.371e+02 1.582e+02 1.824e+02 4.757e+02, threshold=3.165e+02, percent-clipped=1.0 2024-09-23 01:15:23,722 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=144816.0, ans=0.0 2024-09-23 01:15:51,727 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=144909.33333333334, ans=0.0 2024-09-23 01:15:54,883 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=144909.33333333334, ans=0.125 2024-09-23 01:16:23,490 INFO [train.py:1198] (2/4) Epoch 8, batch 3800, loss[loss=0.291, ctc_loss=0.2117, cr_loss=0.3961, over 17007.00 frames. ], tot_loss[loss=0.2637, ctc_loss=0.1857, cr_loss=0.3899, over 3356717.28 frames. ], batch size: 51, lr: 1.46e-02, grad_scale: 32.0 2024-09-23 01:16:28,502 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=145002.66666666666, ans=0.0 2024-09-23 01:16:55,130 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=1.95 vs. limit=15.0 2024-09-23 01:17:07,292 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=145096.0, ans=0.0 2024-09-23 01:17:38,316 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.max_abs, batch_count=145189.33333333334, ans=10.0 2024-09-23 01:17:41,069 INFO [train.py:1198] (2/4) Epoch 8, batch 3850, loss[loss=0.229, ctc_loss=0.1637, cr_loss=0.3267, over 16951.00 frames. ], tot_loss[loss=0.2677, ctc_loss=0.1894, cr_loss=0.3919, over 3295482.73 frames. ], batch size: 42, lr: 1.46e-02, grad_scale: 16.0 2024-09-23 01:17:44,817 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=15.28 vs. limit=15.0 2024-09-23 01:17:45,603 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.152e+02 1.439e+02 1.582e+02 1.881e+02 4.076e+02, threshold=3.165e+02, percent-clipped=1.0 2024-09-23 01:18:08,886 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 01:18:27,311 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=145376.0, ans=0.07 2024-09-23 01:18:34,631 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=145376.0, ans=0.2 2024-09-23 01:18:43,771 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=145422.66666666666, ans=0.2 2024-09-23 01:19:42,986 INFO [train.py:1198] (2/4) Epoch 9, batch 0, loss[loss=0.284, ctc_loss=0.2019, cr_loss=0.4105, over 17233.00 frames. ], tot_loss[loss=0.284, ctc_loss=0.2019, cr_loss=0.4105, over 17233.00 frames. ], batch size: 50, lr: 1.38e-02, grad_scale: 32.0 2024-09-23 01:19:42,986 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 01:19:58,858 INFO [train.py:1230] (2/4) Epoch 9, validation: loss=0.05451, ctc_loss=0.05451, cr_loss=7.076e-15, over 944034.00 frames. 2024-09-23 01:19:58,859 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 01:20:23,370 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=145497.33333333334, ans=0.0 2024-09-23 01:20:28,109 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=145497.33333333334, ans=0.2 2024-09-23 01:20:36,809 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=14.90 vs. limit=22.5 2024-09-23 01:20:42,180 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=145544.0, ans=0.0 2024-09-23 01:21:06,604 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.min_positive, batch_count=145637.33333333334, ans=0.05 2024-09-23 01:21:23,269 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=145637.33333333334, ans=0.0 2024-09-23 01:21:26,165 INFO [train.py:1198] (2/4) Epoch 9, batch 50, loss[loss=0.2658, ctc_loss=0.1872, cr_loss=0.3929, over 17016.00 frames. ], tot_loss[loss=0.271, ctc_loss=0.1917, cr_loss=0.3967, over 753583.51 frames. ], batch size: 51, lr: 1.38e-02, grad_scale: 32.0 2024-09-23 01:21:37,360 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.229e+02 1.512e+02 1.709e+02 2.026e+02 3.260e+02, threshold=3.417e+02, percent-clipped=2.0 2024-09-23 01:21:37,945 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=9.81 vs. limit=15.0 2024-09-23 01:21:50,403 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=145730.66666666666, ans=0.95 2024-09-23 01:22:23,415 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=145824.0, ans=0.125 2024-09-23 01:22:27,281 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=7.21 vs. limit=15.0 2024-09-23 01:22:48,404 INFO [train.py:1198] (2/4) Epoch 9, batch 100, loss[loss=0.2556, ctc_loss=0.1792, cr_loss=0.382, over 17301.00 frames. ], tot_loss[loss=0.2633, ctc_loss=0.1854, cr_loss=0.3898, over 1341722.51 frames. ], batch size: 51, lr: 1.38e-02, grad_scale: 32.0 2024-09-23 01:22:56,515 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=145917.33333333334, ans=0.0 2024-09-23 01:23:23,797 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=146010.66666666666, ans=0.025 2024-09-23 01:23:29,107 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=9.06 vs. limit=15.0 2024-09-23 01:23:44,346 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=146057.33333333334, ans=0.5 2024-09-23 01:24:08,182 INFO [train.py:1198] (2/4) Epoch 9, batch 150, loss[loss=0.2419, ctc_loss=0.1663, cr_loss=0.378, over 17314.00 frames. ], tot_loss[loss=0.2628, ctc_loss=0.1847, cr_loss=0.3907, over 1793574.53 frames. ], batch size: 46, lr: 1.37e-02, grad_scale: 32.0 2024-09-23 01:24:19,530 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.178e+02 1.303e+02 1.427e+02 1.679e+02 2.380e+02, threshold=2.853e+02, percent-clipped=0.0 2024-09-23 01:24:43,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=146244.0, ans=0.125 2024-09-23 01:24:44,114 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.12 vs. limit=6.0 2024-09-23 01:24:47,613 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=10.83 vs. limit=15.0 2024-09-23 01:24:48,490 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=146244.0, ans=0.125 2024-09-23 01:25:02,727 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=20.04 vs. limit=22.5 2024-09-23 01:25:05,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=146290.66666666666, ans=0.2 2024-09-23 01:25:23,885 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=146337.33333333334, ans=0.1 2024-09-23 01:25:25,408 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=146337.33333333334, ans=0.125 2024-09-23 01:25:28,599 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=146337.33333333334, ans=0.0 2024-09-23 01:25:30,261 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=146337.33333333334, ans=0.05 2024-09-23 01:25:33,201 INFO [train.py:1198] (2/4) Epoch 9, batch 200, loss[loss=0.2531, ctc_loss=0.174, cr_loss=0.3953, over 17242.00 frames. ], tot_loss[loss=0.2614, ctc_loss=0.1835, cr_loss=0.39, over 2135053.54 frames. ], batch size: 44, lr: 1.37e-02, grad_scale: 32.0 2024-09-23 01:25:51,469 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.08 vs. limit=15.0 2024-09-23 01:25:52,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=146430.66666666666, ans=0.125 2024-09-23 01:25:58,892 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=146430.66666666666, ans=0.07 2024-09-23 01:26:16,359 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=146477.33333333334, ans=0.125 2024-09-23 01:26:19,481 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=146524.0, ans=0.2 2024-09-23 01:26:34,023 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=4.93 vs. limit=12.0 2024-09-23 01:26:40,296 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=146570.66666666666, ans=0.125 2024-09-23 01:26:42,238 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2.whitening_limit, batch_count=146570.66666666666, ans=15.0 2024-09-23 01:26:55,831 INFO [train.py:1198] (2/4) Epoch 9, batch 250, loss[loss=0.2455, ctc_loss=0.1671, cr_loss=0.392, over 17323.00 frames. ], tot_loss[loss=0.2627, ctc_loss=0.1845, cr_loss=0.391, over 2408435.87 frames. ], batch size: 46, lr: 1.37e-02, grad_scale: 32.0 2024-09-23 01:26:56,098 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=146617.33333333334, ans=0.125 2024-09-23 01:27:06,848 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.339e+02 1.502e+02 1.713e+02 2.875e+02, threshold=3.003e+02, percent-clipped=1.0 2024-09-23 01:27:18,947 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=5.54 vs. limit=10.0 2024-09-23 01:27:25,298 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=12.96 vs. limit=22.5 2024-09-23 01:27:29,012 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=146710.66666666666, ans=0.035 2024-09-23 01:27:33,711 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=146710.66666666666, ans=0.1 2024-09-23 01:27:45,870 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=146757.33333333334, ans=0.05 2024-09-23 01:27:45,969 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=146757.33333333334, ans=0.0 2024-09-23 01:27:53,931 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=146757.33333333334, ans=0.125 2024-09-23 01:27:54,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=146757.33333333334, ans=0.0 2024-09-23 01:27:55,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=146757.33333333334, ans=0.025 2024-09-23 01:27:55,634 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=146757.33333333334, ans=0.1 2024-09-23 01:28:17,488 INFO [train.py:1198] (2/4) Epoch 9, batch 300, loss[loss=0.258, ctc_loss=0.1834, cr_loss=0.3732, over 16557.00 frames. ], tot_loss[loss=0.2616, ctc_loss=0.1837, cr_loss=0.3897, over 2618425.42 frames. ], batch size: 66, lr: 1.37e-02, grad_scale: 32.0 2024-09-23 01:28:21,345 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=9.80 vs. limit=15.0 2024-09-23 01:28:48,348 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=146944.0, ans=0.125 2024-09-23 01:29:04,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=146990.66666666666, ans=0.2 2024-09-23 01:29:09,474 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.25 vs. limit=15.0 2024-09-23 01:29:17,142 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=146990.66666666666, ans=0.025 2024-09-23 01:29:18,724 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=146990.66666666666, ans=0.125 2024-09-23 01:29:19,100 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.64 vs. limit=15.0 2024-09-23 01:29:37,417 INFO [train.py:1198] (2/4) Epoch 9, batch 350, loss[loss=0.2597, ctc_loss=0.1794, cr_loss=0.4018, over 17272.00 frames. ], tot_loss[loss=0.2613, ctc_loss=0.1834, cr_loss=0.3895, over 2775910.32 frames. ], batch size: 42, lr: 1.37e-02, grad_scale: 32.0 2024-09-23 01:29:48,846 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.338e+02 1.473e+02 1.743e+02 2.948e+02, threshold=2.946e+02, percent-clipped=0.0 2024-09-23 01:30:42,279 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=147224.0, ans=0.2 2024-09-23 01:30:47,150 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=147270.66666666666, ans=0.2 2024-09-23 01:31:02,654 INFO [train.py:1198] (2/4) Epoch 9, batch 400, loss[loss=0.306, ctc_loss=0.2177, cr_loss=0.4413, over 16096.00 frames. ], tot_loss[loss=0.2603, ctc_loss=0.1827, cr_loss=0.3881, over 2909360.34 frames. ], batch size: 74, lr: 1.37e-02, grad_scale: 32.0 2024-09-23 01:31:40,554 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=147410.66666666666, ans=0.125 2024-09-23 01:31:42,053 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=147410.66666666666, ans=0.2 2024-09-23 01:31:58,318 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=147457.33333333334, ans=0.0 2024-09-23 01:32:25,077 INFO [train.py:1198] (2/4) Epoch 9, batch 450, loss[loss=0.2302, ctc_loss=0.1621, cr_loss=0.3401, over 16685.00 frames. ], tot_loss[loss=0.2605, ctc_loss=0.1828, cr_loss=0.3882, over 3000147.01 frames. ], batch size: 37, lr: 1.37e-02, grad_scale: 32.0 2024-09-23 01:32:38,951 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.069e+02 1.318e+02 1.484e+02 1.790e+02 2.979e+02, threshold=2.969e+02, percent-clipped=1.0 2024-09-23 01:32:54,216 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2.whitening_limit, batch_count=147597.33333333334, ans=15.0 2024-09-23 01:33:33,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=147737.33333333334, ans=0.125 2024-09-23 01:33:46,279 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=147784.0, ans=0.125 2024-09-23 01:33:47,559 INFO [train.py:1198] (2/4) Epoch 9, batch 500, loss[loss=0.2547, ctc_loss=0.1788, cr_loss=0.3796, over 17009.00 frames. ], tot_loss[loss=0.2614, ctc_loss=0.1836, cr_loss=0.3889, over 3081394.87 frames. ], batch size: 44, lr: 1.37e-02, grad_scale: 32.0 2024-09-23 01:34:03,736 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=147830.66666666666, ans=0.0 2024-09-23 01:34:34,277 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=147924.0, ans=0.125 2024-09-23 01:34:35,823 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 01:34:45,346 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=147924.0, ans=0.0 2024-09-23 01:35:09,827 INFO [train.py:1198] (2/4) Epoch 9, batch 550, loss[loss=0.2872, ctc_loss=0.2039, cr_loss=0.4169, over 15901.00 frames. ], tot_loss[loss=0.2611, ctc_loss=0.1834, cr_loss=0.3886, over 3148060.34 frames. ], batch size: 74, lr: 1.37e-02, grad_scale: 32.0 2024-09-23 01:35:23,367 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.335e+02 1.443e+02 1.626e+02 2.688e+02, threshold=2.885e+02, percent-clipped=0.0 2024-09-23 01:35:45,864 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=148110.66666666666, ans=0.1 2024-09-23 01:36:01,649 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=148157.33333333334, ans=0.0 2024-09-23 01:36:23,346 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=148204.0, ans=0.0 2024-09-23 01:36:34,062 INFO [train.py:1198] (2/4) Epoch 9, batch 600, loss[loss=0.2188, ctc_loss=0.1495, cr_loss=0.3465, over 17247.00 frames. ], tot_loss[loss=0.2596, ctc_loss=0.1821, cr_loss=0.3874, over 3202232.29 frames. ], batch size: 42, lr: 1.37e-02, grad_scale: 32.0 2024-09-23 01:36:53,305 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=148297.33333333334, ans=0.0 2024-09-23 01:37:49,096 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=148437.33333333334, ans=0.0 2024-09-23 01:37:56,745 INFO [train.py:1198] (2/4) Epoch 9, batch 650, loss[loss=0.2552, ctc_loss=0.1832, cr_loss=0.3603, over 17288.00 frames. ], tot_loss[loss=0.2601, ctc_loss=0.1826, cr_loss=0.3873, over 3228975.91 frames. ], batch size: 51, lr: 1.36e-02, grad_scale: 16.0 2024-09-23 01:37:56,988 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.min_positive, batch_count=148484.0, ans=0.05 2024-09-23 01:38:09,415 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.167e+02 1.361e+02 1.468e+02 1.616e+02 2.267e+02, threshold=2.935e+02, percent-clipped=0.0 2024-09-23 01:39:15,865 INFO [train.py:1198] (2/4) Epoch 9, batch 700, loss[loss=0.2523, ctc_loss=0.1774, cr_loss=0.3745, over 17051.00 frames. ], tot_loss[loss=0.2609, ctc_loss=0.1832, cr_loss=0.3887, over 3262079.03 frames. ], batch size: 56, lr: 1.36e-02, grad_scale: 16.0 2024-09-23 01:40:10,109 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=148857.33333333334, ans=0.2 2024-09-23 01:40:13,967 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.94 vs. limit=12.0 2024-09-23 01:40:25,949 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=148904.0, ans=0.125 2024-09-23 01:40:37,169 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=148904.0, ans=0.09899494936611666 2024-09-23 01:40:40,041 INFO [train.py:1198] (2/4) Epoch 9, batch 750, loss[loss=0.2552, ctc_loss=0.1774, cr_loss=0.389, over 17343.00 frames. ], tot_loss[loss=0.2604, ctc_loss=0.1827, cr_loss=0.3885, over 3291326.81 frames. ], batch size: 48, lr: 1.36e-02, grad_scale: 16.0 2024-09-23 01:40:52,737 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.115e+02 1.365e+02 1.514e+02 1.779e+02 2.634e+02, threshold=3.027e+02, percent-clipped=0.0 2024-09-23 01:41:29,521 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=149090.66666666666, ans=0.0 2024-09-23 01:42:02,474 INFO [train.py:1198] (2/4) Epoch 9, batch 800, loss[loss=0.2054, ctc_loss=0.1411, cr_loss=0.3212, over 17038.00 frames. ], tot_loss[loss=0.2611, ctc_loss=0.1832, cr_loss=0.3896, over 3304968.41 frames. ], batch size: 39, lr: 1.36e-02, grad_scale: 32.0 2024-09-23 01:42:22,216 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=149230.66666666666, ans=0.2 2024-09-23 01:42:27,911 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=149230.66666666666, ans=0.125 2024-09-23 01:42:47,246 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=149277.33333333334, ans=0.1 2024-09-23 01:43:06,911 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=149324.0, ans=0.0 2024-09-23 01:43:15,044 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=149370.66666666666, ans=0.125 2024-09-23 01:43:27,282 INFO [train.py:1198] (2/4) Epoch 9, batch 850, loss[loss=0.2353, ctc_loss=0.1618, cr_loss=0.3672, over 17266.00 frames. ], tot_loss[loss=0.2604, ctc_loss=0.1828, cr_loss=0.3884, over 3320316.73 frames. ], batch size: 42, lr: 1.36e-02, grad_scale: 32.0 2024-09-23 01:43:38,825 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=149417.33333333334, ans=0.125 2024-09-23 01:43:39,941 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.141e+02 1.394e+02 1.570e+02 1.849e+02 3.011e+02, threshold=3.140e+02, percent-clipped=0.0 2024-09-23 01:44:00,946 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=149510.66666666666, ans=0.1 2024-09-23 01:44:49,131 INFO [train.py:1198] (2/4) Epoch 9, batch 900, loss[loss=0.2365, ctc_loss=0.1592, cr_loss=0.3864, over 16962.00 frames. ], tot_loss[loss=0.2601, ctc_loss=0.1824, cr_loss=0.3888, over 3338720.02 frames. ], batch size: 42, lr: 1.36e-02, grad_scale: 32.0 2024-09-23 01:44:51,096 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=149650.66666666666, ans=0.1 2024-09-23 01:45:06,674 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=149697.33333333334, ans=0.125 2024-09-23 01:45:11,536 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=149697.33333333334, ans=0.125 2024-09-23 01:45:22,800 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=149744.0, ans=0.2 2024-09-23 01:45:24,937 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=3.10 vs. limit=15.0 2024-09-23 01:45:29,270 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=149744.0, ans=0.0 2024-09-23 01:45:37,119 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=149744.0, ans=0.125 2024-09-23 01:45:38,893 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=149790.66666666666, ans=0.125 2024-09-23 01:45:52,888 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=149790.66666666666, ans=0.0 2024-09-23 01:45:52,931 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=149790.66666666666, ans=0.025 2024-09-23 01:46:14,333 INFO [train.py:1198] (2/4) Epoch 9, batch 950, loss[loss=0.2226, ctc_loss=0.1523, cr_loss=0.3515, over 16348.00 frames. ], tot_loss[loss=0.2587, ctc_loss=0.1812, cr_loss=0.3875, over 3352368.18 frames. ], batch size: 36, lr: 1.36e-02, grad_scale: 32.0 2024-09-23 01:46:27,022 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.338e+02 1.417e+02 1.570e+02 2.385e+02, threshold=2.834e+02, percent-clipped=0.0 2024-09-23 01:46:46,575 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=149977.33333333334, ans=0.125 2024-09-23 01:47:17,085 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=150070.66666666666, ans=0.125 2024-09-23 01:47:27,055 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=150070.66666666666, ans=0.125 2024-09-23 01:47:37,173 INFO [train.py:1198] (2/4) Epoch 9, batch 1000, loss[loss=0.3011, ctc_loss=0.2134, cr_loss=0.4387, over 17048.00 frames. ], tot_loss[loss=0.26, ctc_loss=0.1822, cr_loss=0.3888, over 3361485.00 frames. ], batch size: 56, lr: 1.36e-02, grad_scale: 32.0 2024-09-23 01:48:12,862 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.99 vs. limit=15.0 2024-09-23 01:48:20,196 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=150210.66666666666, ans=0.0 2024-09-23 01:48:24,204 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.04 vs. limit=15.0 2024-09-23 01:48:25,012 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=150257.33333333334, ans=0.0 2024-09-23 01:48:28,581 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=150257.33333333334, ans=0.0 2024-09-23 01:48:54,060 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=150304.0, ans=0.0 2024-09-23 01:48:57,104 INFO [train.py:1198] (2/4) Epoch 9, batch 1050, loss[loss=0.2642, ctc_loss=0.1849, cr_loss=0.3962, over 17138.00 frames. ], tot_loss[loss=0.2593, ctc_loss=0.1817, cr_loss=0.3879, over 3363224.65 frames. ], batch size: 48, lr: 1.36e-02, grad_scale: 32.0 2024-09-23 01:49:09,787 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.138e+02 1.329e+02 1.425e+02 1.705e+02 2.859e+02, threshold=2.851e+02, percent-clipped=1.0 2024-09-23 01:49:14,964 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=150397.33333333334, ans=0.0 2024-09-23 01:49:19,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=150397.33333333334, ans=0.0 2024-09-23 01:49:29,752 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=150444.0, ans=0.2 2024-09-23 01:50:19,670 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=150537.33333333334, ans=0.2 2024-09-23 01:50:22,529 INFO [train.py:1198] (2/4) Epoch 9, batch 1100, loss[loss=0.2336, ctc_loss=0.1643, cr_loss=0.3464, over 17019.00 frames. ], tot_loss[loss=0.2576, ctc_loss=0.1803, cr_loss=0.3862, over 3367357.82 frames. ], batch size: 52, lr: 1.36e-02, grad_scale: 32.0 2024-09-23 01:50:35,541 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=150584.0, ans=0.0 2024-09-23 01:50:45,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=150630.66666666666, ans=0.125 2024-09-23 01:51:19,954 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=150724.0, ans=0.2 2024-09-23 01:51:27,799 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=150770.66666666666, ans=0.125 2024-09-23 01:51:45,049 INFO [train.py:1198] (2/4) Epoch 9, batch 1150, loss[loss=0.2176, ctc_loss=0.148, cr_loss=0.3479, over 17036.00 frames. ], tot_loss[loss=0.2565, ctc_loss=0.1795, cr_loss=0.3851, over 3374822.21 frames. ], batch size: 39, lr: 1.35e-02, grad_scale: 32.0 2024-09-23 01:51:57,742 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.135e+02 1.320e+02 1.454e+02 1.675e+02 2.569e+02, threshold=2.907e+02, percent-clipped=0.0 2024-09-23 01:52:01,359 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=150864.0, ans=0.2 2024-09-23 01:52:15,481 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=150910.66666666666, ans=0.1 2024-09-23 01:52:25,238 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.44 vs. limit=12.0 2024-09-23 01:52:33,737 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=150957.33333333334, ans=0.125 2024-09-23 01:52:38,580 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=150957.33333333334, ans=0.125 2024-09-23 01:53:04,408 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 01:53:07,106 INFO [train.py:1198] (2/4) Epoch 9, batch 1200, loss[loss=0.2481, ctc_loss=0.1771, cr_loss=0.355, over 17161.00 frames. ], tot_loss[loss=0.2581, ctc_loss=0.1808, cr_loss=0.3868, over 3369756.98 frames. ], batch size: 45, lr: 1.35e-02, grad_scale: 32.0 2024-09-23 01:54:18,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=151237.33333333334, ans=0.0 2024-09-23 01:54:27,530 INFO [train.py:1198] (2/4) Epoch 9, batch 1250, loss[loss=0.2538, ctc_loss=0.1764, cr_loss=0.387, over 17090.00 frames. ], tot_loss[loss=0.2589, ctc_loss=0.1814, cr_loss=0.3875, over 3363005.00 frames. ], batch size: 49, lr: 1.35e-02, grad_scale: 32.0 2024-09-23 01:54:35,000 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=9.51 vs. limit=10.0 2024-09-23 01:54:37,408 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=151284.0, ans=0.07 2024-09-23 01:54:42,634 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.349e+02 1.478e+02 1.607e+02 2.970e+02, threshold=2.955e+02, percent-clipped=1.0 2024-09-23 01:54:49,504 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=6.12 vs. limit=15.0 2024-09-23 01:54:57,318 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=151330.66666666666, ans=0.1 2024-09-23 01:55:09,239 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=151377.33333333334, ans=0.125 2024-09-23 01:55:21,819 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.max_abs, batch_count=151424.0, ans=10.0 2024-09-23 01:55:50,280 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=151517.33333333334, ans=0.125 2024-09-23 01:55:51,558 INFO [train.py:1198] (2/4) Epoch 9, batch 1300, loss[loss=0.2773, ctc_loss=0.1913, cr_loss=0.4299, over 17218.00 frames. ], tot_loss[loss=0.26, ctc_loss=0.1823, cr_loss=0.3884, over 3357593.61 frames. ], batch size: 50, lr: 1.35e-02, grad_scale: 32.0 2024-09-23 01:56:22,353 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=8.22 vs. limit=15.0 2024-09-23 01:57:13,855 INFO [train.py:1198] (2/4) Epoch 9, batch 1350, loss[loss=0.2236, ctc_loss=0.1522, cr_loss=0.3569, over 17105.00 frames. ], tot_loss[loss=0.2595, ctc_loss=0.1821, cr_loss=0.3873, over 3348135.90 frames. ], batch size: 40, lr: 1.35e-02, grad_scale: 32.0 2024-09-23 01:57:29,083 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.345e+02 1.485e+02 1.651e+02 2.569e+02, threshold=2.970e+02, percent-clipped=0.0 2024-09-23 01:57:37,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=151797.33333333334, ans=0.125 2024-09-23 01:57:39,078 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=151797.33333333334, ans=0.125 2024-09-23 01:57:52,810 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=6.25 vs. limit=15.0 2024-09-23 01:57:58,640 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=11.06 vs. limit=15.0 2024-09-23 01:58:35,901 INFO [train.py:1198] (2/4) Epoch 9, batch 1400, loss[loss=0.2395, ctc_loss=0.1575, cr_loss=0.4099, over 17122.00 frames. ], tot_loss[loss=0.2599, ctc_loss=0.1823, cr_loss=0.388, over 3347780.67 frames. ], batch size: 40, lr: 1.35e-02, grad_scale: 32.0 2024-09-23 01:58:56,843 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=152030.66666666666, ans=0.0 2024-09-23 01:59:20,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=152077.33333333334, ans=0.0 2024-09-23 01:59:31,575 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=152124.0, ans=0.1 2024-09-23 01:59:44,236 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=152170.66666666666, ans=0.0 2024-09-23 02:00:01,048 INFO [train.py:1198] (2/4) Epoch 9, batch 1450, loss[loss=0.2331, ctc_loss=0.1656, cr_loss=0.3377, over 17197.00 frames. ], tot_loss[loss=0.2595, ctc_loss=0.182, cr_loss=0.3877, over 3353174.74 frames. ], batch size: 41, lr: 1.35e-02, grad_scale: 32.0 2024-09-23 02:00:13,885 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.066e+02 1.414e+02 1.556e+02 1.752e+02 2.841e+02, threshold=3.113e+02, percent-clipped=0.0 2024-09-23 02:00:14,276 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=152217.33333333334, ans=0.1 2024-09-23 02:00:15,815 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=152264.0, ans=0.09899494936611666 2024-09-23 02:00:17,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=152264.0, ans=0.0 2024-09-23 02:00:22,212 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=152264.0, ans=0.2 2024-09-23 02:00:22,228 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.min_positive, batch_count=152264.0, ans=0.025 2024-09-23 02:00:33,759 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=14.81 vs. limit=15.0 2024-09-23 02:00:46,674 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten.whitening_limit, batch_count=152310.66666666666, ans=15.0 2024-09-23 02:00:52,529 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=152357.33333333334, ans=0.0 2024-09-23 02:01:06,450 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.24 vs. limit=15.0 2024-09-23 02:01:11,513 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.65 vs. limit=10.0 2024-09-23 02:01:23,505 INFO [train.py:1198] (2/4) Epoch 9, batch 1500, loss[loss=0.3085, ctc_loss=0.2202, cr_loss=0.4418, over 17021.00 frames. ], tot_loss[loss=0.2596, ctc_loss=0.1819, cr_loss=0.3881, over 3343374.55 frames. ], batch size: 56, lr: 1.35e-02, grad_scale: 32.0 2024-09-23 02:01:23,752 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=152450.66666666666, ans=0.0 2024-09-23 02:01:42,962 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.53 vs. limit=12.0 2024-09-23 02:01:58,580 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=152544.0, ans=0.125 2024-09-23 02:02:01,000 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=6.24 vs. limit=15.0 2024-09-23 02:02:12,145 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=152590.66666666666, ans=0.2 2024-09-23 02:02:28,742 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.whiten, num_groups=1, num_channels=192, metric=4.02 vs. limit=12.0 2024-09-23 02:02:29,651 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=152637.33333333334, ans=0.125 2024-09-23 02:02:45,183 INFO [train.py:1198] (2/4) Epoch 9, batch 1550, loss[loss=0.2535, ctc_loss=0.1737, cr_loss=0.3988, over 17009.00 frames. ], tot_loss[loss=0.26, ctc_loss=0.1823, cr_loss=0.3883, over 3347652.39 frames. ], batch size: 44, lr: 1.35e-02, grad_scale: 32.0 2024-09-23 02:02:46,995 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=152684.0, ans=0.1 2024-09-23 02:02:57,972 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.113e+02 1.340e+02 1.461e+02 1.652e+02 2.342e+02, threshold=2.922e+02, percent-clipped=0.0 2024-09-23 02:03:10,992 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=152730.66666666666, ans=0.125 2024-09-23 02:03:14,077 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=152730.66666666666, ans=0.125 2024-09-23 02:03:16,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.whiten.whitening_limit, batch_count=152777.33333333334, ans=15.0 2024-09-23 02:03:57,249 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=152870.66666666666, ans=0.035 2024-09-23 02:04:04,956 INFO [train.py:1198] (2/4) Epoch 9, batch 1600, loss[loss=0.2552, ctc_loss=0.1807, cr_loss=0.3724, over 17145.00 frames. ], tot_loss[loss=0.259, ctc_loss=0.1815, cr_loss=0.3876, over 3358601.71 frames. ], batch size: 48, lr: 1.35e-02, grad_scale: 32.0 2024-09-23 02:04:06,729 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=152917.33333333334, ans=0.015 2024-09-23 02:04:07,354 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=10.74 vs. limit=15.0 2024-09-23 02:04:08,427 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=152917.33333333334, ans=0.0 2024-09-23 02:05:01,836 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=153057.33333333334, ans=0.0 2024-09-23 02:05:30,267 INFO [train.py:1198] (2/4) Epoch 9, batch 1650, loss[loss=0.3488, ctc_loss=0.2647, cr_loss=0.4205, over 11393.00 frames. ], tot_loss[loss=0.2578, ctc_loss=0.1808, cr_loss=0.3853, over 3345515.46 frames. ], batch size: 123, lr: 1.34e-02, grad_scale: 32.0 2024-09-23 02:05:34,127 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=17.53 vs. limit=22.5 2024-09-23 02:05:42,928 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.152e+02 1.315e+02 1.418e+02 1.618e+02 2.447e+02, threshold=2.836e+02, percent-clipped=0.0 2024-09-23 02:06:17,559 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=153244.0, ans=0.125 2024-09-23 02:06:31,852 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=153290.66666666666, ans=0.125 2024-09-23 02:06:43,318 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=153337.33333333334, ans=0.0 2024-09-23 02:06:52,596 INFO [train.py:1198] (2/4) Epoch 9, batch 1700, loss[loss=0.3358, ctc_loss=0.256, cr_loss=0.3987, over 11362.00 frames. ], tot_loss[loss=0.2582, ctc_loss=0.1809, cr_loss=0.386, over 3348567.45 frames. ], batch size: 123, lr: 1.34e-02, grad_scale: 32.0 2024-09-23 02:06:56,063 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=153384.0, ans=0.125 2024-09-23 02:07:04,069 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.min_positive, batch_count=153384.0, ans=0.025 2024-09-23 02:07:17,949 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=153430.66666666666, ans=0.2 2024-09-23 02:07:19,548 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 02:07:27,285 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=153477.33333333334, ans=0.5 2024-09-23 02:07:33,833 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=153477.33333333334, ans=0.1 2024-09-23 02:08:14,978 INFO [train.py:1198] (2/4) Epoch 9, batch 1750, loss[loss=0.2287, ctc_loss=0.157, cr_loss=0.3584, over 17206.00 frames. ], tot_loss[loss=0.2577, ctc_loss=0.1806, cr_loss=0.3854, over 3341335.55 frames. ], batch size: 47, lr: 1.34e-02, grad_scale: 32.0 2024-09-23 02:08:20,080 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=153617.33333333334, ans=0.125 2024-09-23 02:08:20,152 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=153617.33333333334, ans=0.1 2024-09-23 02:08:24,754 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=153617.33333333334, ans=0.125 2024-09-23 02:08:27,630 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.306e+02 1.435e+02 1.601e+02 2.532e+02, threshold=2.871e+02, percent-clipped=0.0 2024-09-23 02:09:32,680 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 02:09:37,154 INFO [train.py:1198] (2/4) Epoch 9, batch 1800, loss[loss=0.2174, ctc_loss=0.1489, cr_loss=0.3425, over 16966.00 frames. ], tot_loss[loss=0.2569, ctc_loss=0.1799, cr_loss=0.3851, over 3348811.72 frames. ], batch size: 42, lr: 1.34e-02, grad_scale: 32.0 2024-09-23 02:09:50,412 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=153850.66666666666, ans=0.0 2024-09-23 02:10:00,807 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=153897.33333333334, ans=0.2 2024-09-23 02:10:19,920 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=153944.0, ans=0.125 2024-09-23 02:10:21,567 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=153944.0, ans=0.125 2024-09-23 02:10:36,322 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=153990.66666666666, ans=0.025 2024-09-23 02:10:39,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=153990.66666666666, ans=0.125 2024-09-23 02:10:56,340 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=154037.33333333334, ans=0.04949747468305833 2024-09-23 02:11:02,433 INFO [train.py:1198] (2/4) Epoch 9, batch 1850, loss[loss=0.2968, ctc_loss=0.2103, cr_loss=0.4327, over 15898.00 frames. ], tot_loss[loss=0.2566, ctc_loss=0.1796, cr_loss=0.385, over 3358403.06 frames. ], batch size: 74, lr: 1.34e-02, grad_scale: 32.0 2024-09-23 02:11:02,840 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=154084.0, ans=0.2 2024-09-23 02:11:15,289 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.166e+02 1.311e+02 1.455e+02 1.599e+02 2.363e+02, threshold=2.909e+02, percent-clipped=0.0 2024-09-23 02:11:26,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=154130.66666666666, ans=0.125 2024-09-23 02:11:44,218 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=154177.33333333334, ans=0.2 2024-09-23 02:12:04,919 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.82 vs. limit=6.0 2024-09-23 02:12:24,532 INFO [train.py:1198] (2/4) Epoch 9, batch 1900, loss[loss=0.2575, ctc_loss=0.1803, cr_loss=0.3859, over 17289.00 frames. ], tot_loss[loss=0.258, ctc_loss=0.1807, cr_loss=0.3865, over 3350963.98 frames. ], batch size: 49, lr: 1.34e-02, grad_scale: 32.0 2024-09-23 02:12:34,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=154317.33333333334, ans=0.125 2024-09-23 02:13:39,286 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=154504.0, ans=0.125 2024-09-23 02:13:43,681 INFO [train.py:1198] (2/4) Epoch 9, batch 1950, loss[loss=0.2131, ctc_loss=0.1473, cr_loss=0.3289, over 16253.00 frames. ], tot_loss[loss=0.2574, ctc_loss=0.1802, cr_loss=0.3862, over 3353296.82 frames. ], batch size: 36, lr: 1.34e-02, grad_scale: 32.0 2024-09-23 02:13:56,415 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.387e+02 1.540e+02 1.769e+02 3.177e+02, threshold=3.081e+02, percent-clipped=1.0 2024-09-23 02:14:28,351 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=154644.0, ans=0.2 2024-09-23 02:14:37,739 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=154690.66666666666, ans=0.0 2024-09-23 02:14:45,939 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=154690.66666666666, ans=0.125 2024-09-23 02:14:58,208 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=154737.33333333334, ans=0.125 2024-09-23 02:15:09,132 INFO [train.py:1198] (2/4) Epoch 9, batch 2000, loss[loss=0.2311, ctc_loss=0.1623, cr_loss=0.3441, over 17038.00 frames. ], tot_loss[loss=0.2571, ctc_loss=0.1799, cr_loss=0.3862, over 3352953.57 frames. ], batch size: 39, lr: 1.34e-02, grad_scale: 32.0 2024-09-23 02:15:17,552 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=154784.0, ans=0.125 2024-09-23 02:15:48,589 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=154877.33333333334, ans=0.125 2024-09-23 02:16:01,287 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=154924.0, ans=0.125 2024-09-23 02:16:26,751 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=154970.66666666666, ans=0.125 2024-09-23 02:16:31,230 INFO [train.py:1198] (2/4) Epoch 9, batch 2050, loss[loss=0.2408, ctc_loss=0.1683, cr_loss=0.3627, over 17293.00 frames. ], tot_loss[loss=0.2578, ctc_loss=0.1804, cr_loss=0.3871, over 3348944.09 frames. ], batch size: 46, lr: 1.34e-02, grad_scale: 32.0 2024-09-23 02:16:41,090 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.min_positive, batch_count=155017.33333333334, ans=0.025 2024-09-23 02:16:43,967 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.064e+02 1.371e+02 1.510e+02 1.685e+02 3.292e+02, threshold=3.020e+02, percent-clipped=1.0 2024-09-23 02:17:06,453 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=11.39 vs. limit=22.5 2024-09-23 02:17:12,888 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.69 vs. limit=15.0 2024-09-23 02:17:26,897 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=155157.33333333334, ans=0.125 2024-09-23 02:17:39,625 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=155204.0, ans=0.1 2024-09-23 02:17:53,734 INFO [train.py:1198] (2/4) Epoch 9, batch 2100, loss[loss=0.2405, ctc_loss=0.1689, cr_loss=0.3584, over 16969.00 frames. ], tot_loss[loss=0.258, ctc_loss=0.1806, cr_loss=0.3872, over 3344997.97 frames. ], batch size: 42, lr: 1.34e-02, grad_scale: 32.0 2024-09-23 02:17:55,999 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=14.77 vs. limit=22.5 2024-09-23 02:18:00,453 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=155250.66666666666, ans=0.0 2024-09-23 02:18:03,674 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=155250.66666666666, ans=0.0 2024-09-23 02:18:05,269 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=155250.66666666666, ans=0.125 2024-09-23 02:18:30,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=155344.0, ans=0.1 2024-09-23 02:19:12,972 INFO [train.py:1198] (2/4) Epoch 9, batch 2150, loss[loss=0.3118, ctc_loss=0.2313, cr_loss=0.4023, over 11538.00 frames. ], tot_loss[loss=0.2583, ctc_loss=0.1809, cr_loss=0.3872, over 3340303.33 frames. ], batch size: 123, lr: 1.34e-02, grad_scale: 32.0 2024-09-23 02:19:23,669 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=155484.0, ans=0.0 2024-09-23 02:19:23,781 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=155484.0, ans=0.025 2024-09-23 02:19:25,327 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=155484.0, ans=0.0 2024-09-23 02:19:28,282 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.177e+02 1.379e+02 1.514e+02 1.800e+02 2.768e+02, threshold=3.028e+02, percent-clipped=0.0 2024-09-23 02:20:07,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=155624.0, ans=0.1 2024-09-23 02:20:09,505 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=155624.0, ans=0.0 2024-09-23 02:20:25,590 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=155670.66666666666, ans=0.2 2024-09-23 02:20:38,243 INFO [train.py:1198] (2/4) Epoch 9, batch 2200, loss[loss=0.2282, ctc_loss=0.1585, cr_loss=0.3485, over 17037.00 frames. ], tot_loss[loss=0.2568, ctc_loss=0.1798, cr_loss=0.3851, over 3337723.84 frames. ], batch size: 39, lr: 1.33e-02, grad_scale: 32.0 2024-09-23 02:20:47,508 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=155717.33333333334, ans=0.125 2024-09-23 02:20:52,244 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=155717.33333333334, ans=0.0 2024-09-23 02:21:19,750 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=155810.66666666666, ans=0.1 2024-09-23 02:21:26,095 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=155810.66666666666, ans=0.125 2024-09-23 02:21:32,693 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=155857.33333333334, ans=0.0 2024-09-23 02:21:43,770 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=155904.0, ans=0.125 2024-09-23 02:22:00,874 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=155904.0, ans=0.2 2024-09-23 02:22:03,746 INFO [train.py:1198] (2/4) Epoch 9, batch 2250, loss[loss=0.2334, ctc_loss=0.1615, cr_loss=0.3595, over 17262.00 frames. ], tot_loss[loss=0.2564, ctc_loss=0.1796, cr_loss=0.3842, over 3342567.30 frames. ], batch size: 44, lr: 1.33e-02, grad_scale: 32.0 2024-09-23 02:22:07,901 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=22.18 vs. limit=22.5 2024-09-23 02:22:16,468 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.340e+02 1.524e+02 1.728e+02 3.121e+02, threshold=3.047e+02, percent-clipped=1.0 2024-09-23 02:22:42,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=156044.0, ans=0.0 2024-09-23 02:22:46,945 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=156044.0, ans=0.125 2024-09-23 02:23:02,847 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer_ff2.min_abs, batch_count=156090.66666666666, ans=0.1 2024-09-23 02:23:06,145 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=156137.33333333334, ans=0.0 2024-09-23 02:23:17,344 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=156137.33333333334, ans=0.04949747468305833 2024-09-23 02:23:23,327 INFO [train.py:1198] (2/4) Epoch 9, batch 2300, loss[loss=0.2096, ctc_loss=0.1413, cr_loss=0.3417, over 17276.00 frames. ], tot_loss[loss=0.2569, ctc_loss=0.18, cr_loss=0.3846, over 3339395.37 frames. ], batch size: 42, lr: 1.33e-02, grad_scale: 32.0 2024-09-23 02:23:33,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=156184.0, ans=0.2 2024-09-23 02:24:10,328 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=156324.0, ans=0.125 2024-09-23 02:24:14,529 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=156324.0, ans=0.0 2024-09-23 02:24:20,633 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=156324.0, ans=0.1 2024-09-23 02:24:22,300 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=156324.0, ans=0.125 2024-09-23 02:24:34,174 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=9.04 vs. limit=15.0 2024-09-23 02:24:37,278 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.66 vs. limit=12.0 2024-09-23 02:24:47,302 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=156417.33333333334, ans=0.0 2024-09-23 02:24:48,596 INFO [train.py:1198] (2/4) Epoch 9, batch 2350, loss[loss=0.2561, ctc_loss=0.1787, cr_loss=0.387, over 17301.00 frames. ], tot_loss[loss=0.2574, ctc_loss=0.1802, cr_loss=0.3858, over 3345283.54 frames. ], batch size: 46, lr: 1.33e-02, grad_scale: 32.0 2024-09-23 02:25:01,204 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.187e+02 1.341e+02 1.440e+02 1.582e+02 2.935e+02, threshold=2.879e+02, percent-clipped=0.0 2024-09-23 02:25:03,883 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.out_whiten.whitening_limit, batch_count=156464.0, ans=15.0 2024-09-23 02:25:18,975 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 02:25:31,633 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=156510.66666666666, ans=0.1 2024-09-23 02:25:42,090 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=156557.33333333334, ans=0.125 2024-09-23 02:25:48,391 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=156557.33333333334, ans=0.125 2024-09-23 02:26:07,350 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=156604.0, ans=0.0 2024-09-23 02:26:10,107 INFO [train.py:1198] (2/4) Epoch 9, batch 2400, loss[loss=0.2327, ctc_loss=0.1629, cr_loss=0.3491, over 17352.00 frames. ], tot_loss[loss=0.2578, ctc_loss=0.1806, cr_loss=0.3859, over 3339917.23 frames. ], batch size: 48, lr: 1.33e-02, grad_scale: 32.0 2024-09-23 02:26:37,511 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=156697.33333333334, ans=0.2 2024-09-23 02:26:52,640 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=156744.0, ans=0.2 2024-09-23 02:27:10,553 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=156790.66666666666, ans=0.0 2024-09-23 02:27:25,097 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=156837.33333333334, ans=0.05 2024-09-23 02:27:28,331 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=156837.33333333334, ans=0.2 2024-09-23 02:27:32,763 INFO [train.py:1198] (2/4) Epoch 9, batch 2450, loss[loss=0.2558, ctc_loss=0.1776, cr_loss=0.3908, over 17150.00 frames. ], tot_loss[loss=0.2591, ctc_loss=0.1816, cr_loss=0.3871, over 3335106.11 frames. ], batch size: 48, lr: 1.33e-02, grad_scale: 32.0 2024-09-23 02:27:45,477 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.155e+02 1.502e+02 1.626e+02 1.858e+02 2.761e+02, threshold=3.252e+02, percent-clipped=0.0 2024-09-23 02:28:00,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=156930.66666666666, ans=0.125 2024-09-23 02:28:06,003 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=12.55 vs. limit=22.5 2024-09-23 02:28:09,974 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=156977.33333333334, ans=0.2 2024-09-23 02:28:17,827 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=156977.33333333334, ans=0.07 2024-09-23 02:28:25,897 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=157024.0, ans=0.125 2024-09-23 02:28:35,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=157070.66666666666, ans=0.1 2024-09-23 02:28:52,842 INFO [train.py:1198] (2/4) Epoch 9, batch 2500, loss[loss=0.2597, ctc_loss=0.1833, cr_loss=0.3821, over 16904.00 frames. ], tot_loss[loss=0.2586, ctc_loss=0.1812, cr_loss=0.3871, over 3348869.52 frames. ], batch size: 58, lr: 1.33e-02, grad_scale: 32.0 2024-09-23 02:28:53,140 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=157117.33333333334, ans=0.0 2024-09-23 02:29:16,664 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten.whitening_limit, batch_count=157164.0, ans=22.5 2024-09-23 02:29:57,588 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.min_positive, batch_count=157257.33333333334, ans=0.025 2024-09-23 02:30:01,458 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.78 vs. limit=15.0 2024-09-23 02:30:03,822 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=157304.0, ans=0.0 2024-09-23 02:30:03,827 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=157304.0, ans=0.125 2024-09-23 02:30:18,043 INFO [train.py:1198] (2/4) Epoch 9, batch 2550, loss[loss=0.2674, ctc_loss=0.1878, cr_loss=0.3977, over 17302.00 frames. ], tot_loss[loss=0.2582, ctc_loss=0.1809, cr_loss=0.3865, over 3346531.83 frames. ], batch size: 46, lr: 1.33e-02, grad_scale: 32.0 2024-09-23 02:30:22,991 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=157350.66666666666, ans=0.2 2024-09-23 02:30:30,729 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.078e+02 1.377e+02 1.519e+02 1.754e+02 2.605e+02, threshold=3.038e+02, percent-clipped=0.0 2024-09-23 02:30:41,590 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=157397.33333333334, ans=0.0 2024-09-23 02:30:51,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=157444.0, ans=0.2 2024-09-23 02:30:54,927 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.17 vs. limit=22.5 2024-09-23 02:31:40,944 INFO [train.py:1198] (2/4) Epoch 9, batch 2600, loss[loss=0.299, ctc_loss=0.2118, cr_loss=0.4363, over 16703.00 frames. ], tot_loss[loss=0.2582, ctc_loss=0.1809, cr_loss=0.3865, over 3340841.36 frames. ], batch size: 61, lr: 1.33e-02, grad_scale: 32.0 2024-09-23 02:31:47,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=157584.0, ans=0.125 2024-09-23 02:31:55,310 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=157584.0, ans=0.0 2024-09-23 02:32:31,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=157724.0, ans=0.125 2024-09-23 02:32:40,260 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=157724.0, ans=0.125 2024-09-23 02:32:45,065 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=157724.0, ans=0.125 2024-09-23 02:33:04,041 INFO [train.py:1198] (2/4) Epoch 9, batch 2650, loss[loss=0.2797, ctc_loss=0.1927, cr_loss=0.4352, over 17349.00 frames. ], tot_loss[loss=0.2598, ctc_loss=0.1821, cr_loss=0.3882, over 3348605.26 frames. ], batch size: 48, lr: 1.33e-02, grad_scale: 64.0 2024-09-23 02:33:04,335 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=157817.33333333334, ans=0.0 2024-09-23 02:33:14,012 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=157817.33333333334, ans=0.125 2024-09-23 02:33:16,814 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.311e+02 1.423e+02 1.645e+02 2.492e+02, threshold=2.847e+02, percent-clipped=0.0 2024-09-23 02:33:20,302 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=157864.0, ans=0.1 2024-09-23 02:34:26,715 INFO [train.py:1198] (2/4) Epoch 9, batch 2700, loss[loss=0.2092, ctc_loss=0.1442, cr_loss=0.3248, over 17253.00 frames. ], tot_loss[loss=0.2585, ctc_loss=0.1811, cr_loss=0.3872, over 3356170.80 frames. ], batch size: 44, lr: 1.32e-02, grad_scale: 64.0 2024-09-23 02:34:30,053 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=158050.66666666666, ans=0.0 2024-09-23 02:34:51,802 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=158097.33333333334, ans=0.125 2024-09-23 02:34:56,727 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=158097.33333333334, ans=0.0 2024-09-23 02:35:02,943 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=158144.0, ans=0.2 2024-09-23 02:35:42,658 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 02:35:47,384 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=158237.33333333334, ans=0.0 2024-09-23 02:35:51,927 INFO [train.py:1198] (2/4) Epoch 9, batch 2750, loss[loss=0.2934, ctc_loss=0.2106, cr_loss=0.4138, over 16710.00 frames. ], tot_loss[loss=0.2586, ctc_loss=0.1812, cr_loss=0.387, over 3355291.00 frames. ], batch size: 61, lr: 1.32e-02, grad_scale: 32.0 2024-09-23 02:36:06,252 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.343e+02 1.485e+02 1.822e+02 3.173e+02, threshold=2.970e+02, percent-clipped=1.0 2024-09-23 02:36:30,493 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=158377.33333333334, ans=0.2 2024-09-23 02:36:56,431 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.whiten.whitening_limit, batch_count=158424.0, ans=15.0 2024-09-23 02:37:08,345 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=158470.66666666666, ans=0.125 2024-09-23 02:37:08,459 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=158470.66666666666, ans=0.1 2024-09-23 02:37:14,489 INFO [train.py:1198] (2/4) Epoch 9, batch 2800, loss[loss=0.2689, ctc_loss=0.1881, cr_loss=0.4039, over 16919.00 frames. ], tot_loss[loss=0.2584, ctc_loss=0.181, cr_loss=0.3872, over 3368234.87 frames. ], batch size: 58, lr: 1.32e-02, grad_scale: 32.0 2024-09-23 02:37:15,456 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=11.90 vs. limit=15.0 2024-09-23 02:37:18,424 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.53 vs. limit=22.5 2024-09-23 02:37:22,988 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=158517.33333333334, ans=0.125 2024-09-23 02:37:24,573 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=158517.33333333334, ans=0.125 2024-09-23 02:37:27,913 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=158517.33333333334, ans=0.125 2024-09-23 02:37:48,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=158610.66666666666, ans=0.125 2024-09-23 02:38:04,453 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=158657.33333333334, ans=0.125 2024-09-23 02:38:25,350 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=158704.0, ans=0.125 2024-09-23 02:38:34,672 INFO [train.py:1198] (2/4) Epoch 9, batch 2850, loss[loss=0.261, ctc_loss=0.1841, cr_loss=0.3844, over 17301.00 frames. ], tot_loss[loss=0.2609, ctc_loss=0.1831, cr_loss=0.3891, over 3345658.55 frames. ], batch size: 49, lr: 1.32e-02, grad_scale: 16.0 2024-09-23 02:38:39,733 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=158750.66666666666, ans=0.125 2024-09-23 02:38:50,463 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.152e+02 1.314e+02 1.411e+02 1.546e+02 2.607e+02, threshold=2.821e+02, percent-clipped=0.0 2024-09-23 02:39:16,027 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=158844.0, ans=0.1 2024-09-23 02:39:17,511 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 02:39:31,894 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.04 vs. limit=6.0 2024-09-23 02:39:43,875 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=158937.33333333334, ans=0.07 2024-09-23 02:39:55,058 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=158937.33333333334, ans=0.0 2024-09-23 02:39:59,459 INFO [train.py:1198] (2/4) Epoch 9, batch 2900, loss[loss=0.2158, ctc_loss=0.1473, cr_loss=0.3427, over 17095.00 frames. ], tot_loss[loss=0.2597, ctc_loss=0.1821, cr_loss=0.388, over 3349891.60 frames. ], batch size: 43, lr: 1.32e-02, grad_scale: 16.0 2024-09-23 02:40:15,748 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=159030.66666666666, ans=0.1 2024-09-23 02:40:34,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=159077.33333333334, ans=0.1 2024-09-23 02:40:34,434 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=10.03 vs. limit=15.0 2024-09-23 02:40:38,705 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=159077.33333333334, ans=0.0 2024-09-23 02:40:40,497 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=159077.33333333334, ans=0.125 2024-09-23 02:40:58,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=159124.0, ans=0.025 2024-09-23 02:41:21,661 INFO [train.py:1198] (2/4) Epoch 9, batch 2950, loss[loss=0.2266, ctc_loss=0.1553, cr_loss=0.3567, over 17111.00 frames. ], tot_loss[loss=0.2592, ctc_loss=0.1816, cr_loss=0.3877, over 3348024.74 frames. ], batch size: 40, lr: 1.32e-02, grad_scale: 16.0 2024-09-23 02:41:38,248 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=9.25 vs. limit=10.0 2024-09-23 02:41:40,276 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.146e+02 1.351e+02 1.516e+02 1.686e+02 2.482e+02, threshold=3.032e+02, percent-clipped=0.0 2024-09-23 02:41:50,363 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=159264.0, ans=0.125 2024-09-23 02:42:04,815 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.67 vs. limit=10.0 2024-09-23 02:42:17,452 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.59 vs. limit=22.5 2024-09-23 02:42:18,498 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=159357.33333333334, ans=0.1 2024-09-23 02:42:18,646 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=159357.33333333334, ans=0.07 2024-09-23 02:42:20,044 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=159357.33333333334, ans=0.1 2024-09-23 02:42:21,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=159357.33333333334, ans=0.125 2024-09-23 02:42:37,483 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=159404.0, ans=0.125 2024-09-23 02:42:43,437 INFO [train.py:1198] (2/4) Epoch 9, batch 3000, loss[loss=0.203, ctc_loss=0.136, cr_loss=0.3347, over 17273.00 frames. ], tot_loss[loss=0.259, ctc_loss=0.1815, cr_loss=0.3877, over 3341603.91 frames. ], batch size: 42, lr: 1.32e-02, grad_scale: 16.0 2024-09-23 02:42:43,437 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 02:42:54,038 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.2.encoder.layers.1.self_attn_weights, attn_weights_entropy = tensor([5.0684, 4.5713, 4.3813, 4.5550], device='cuda:2') 2024-09-23 02:42:59,046 INFO [train.py:1230] (2/4) Epoch 9, validation: loss=0.05024, ctc_loss=0.05024, cr_loss=7.059e-15, over 944034.00 frames. 2024-09-23 02:42:59,047 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 02:43:30,639 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=159544.0, ans=0.0 2024-09-23 02:44:08,689 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=159637.33333333334, ans=0.0 2024-09-23 02:44:17,776 INFO [train.py:1198] (2/4) Epoch 9, batch 3050, loss[loss=0.2281, ctc_loss=0.1587, cr_loss=0.347, over 16963.00 frames. ], tot_loss[loss=0.2589, ctc_loss=0.1813, cr_loss=0.388, over 3346195.10 frames. ], batch size: 42, lr: 1.32e-02, grad_scale: 16.0 2024-09-23 02:44:21,278 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 02:44:32,537 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=159730.66666666666, ans=0.0 2024-09-23 02:44:33,650 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.146e+02 1.314e+02 1.416e+02 1.662e+02 2.316e+02, threshold=2.832e+02, percent-clipped=0.0 2024-09-23 02:44:41,633 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=159730.66666666666, ans=0.2 2024-09-23 02:45:02,888 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=8.96 vs. limit=10.0 2024-09-23 02:45:11,552 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.86 vs. limit=15.0 2024-09-23 02:45:35,916 INFO [train.py:1198] (2/4) Epoch 9, batch 3100, loss[loss=0.2808, ctc_loss=0.1991, cr_loss=0.4083, over 17309.00 frames. ], tot_loss[loss=0.259, ctc_loss=0.1813, cr_loss=0.3886, over 3348799.50 frames. ], batch size: 49, lr: 1.32e-02, grad_scale: 16.0 2024-09-23 02:45:52,716 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=159964.0, ans=0.125 2024-09-23 02:46:00,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=159964.0, ans=0.125 2024-09-23 02:46:03,640 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=159964.0, ans=0.09899494936611666 2024-09-23 02:46:14,597 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=160010.66666666666, ans=0.0 2024-09-23 02:46:58,019 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=160150.66666666666, ans=0.2 2024-09-23 02:46:58,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=160150.66666666666, ans=0.125 2024-09-23 02:46:59,395 INFO [train.py:1198] (2/4) Epoch 9, batch 3150, loss[loss=0.2509, ctc_loss=0.1778, cr_loss=0.3653, over 17142.00 frames. ], tot_loss[loss=0.2582, ctc_loss=0.1806, cr_loss=0.3878, over 3354033.63 frames. ], batch size: 48, lr: 1.32e-02, grad_scale: 16.0 2024-09-23 02:47:15,043 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.206e+02 1.398e+02 1.504e+02 1.697e+02 2.388e+02, threshold=3.008e+02, percent-clipped=0.0 2024-09-23 02:47:26,176 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=160197.33333333334, ans=0.125 2024-09-23 02:47:49,487 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=160290.66666666666, ans=0.1 2024-09-23 02:47:52,878 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=160290.66666666666, ans=0.125 2024-09-23 02:47:54,314 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=160290.66666666666, ans=0.125 2024-09-23 02:48:14,876 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=160337.33333333334, ans=0.2 2024-09-23 02:48:17,649 INFO [train.py:1198] (2/4) Epoch 9, batch 3200, loss[loss=0.288, ctc_loss=0.2026, cr_loss=0.4269, over 17052.00 frames. ], tot_loss[loss=0.259, ctc_loss=0.1813, cr_loss=0.3884, over 3346815.66 frames. ], batch size: 56, lr: 1.32e-02, grad_scale: 32.0 2024-09-23 02:48:24,092 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=160384.0, ans=0.0 2024-09-23 02:48:32,108 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=160430.66666666666, ans=0.025 2024-09-23 02:48:35,823 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.84 vs. limit=10.0 2024-09-23 02:48:36,965 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=160430.66666666666, ans=0.2 2024-09-23 02:48:46,882 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=160430.66666666666, ans=0.1 2024-09-23 02:49:05,652 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 02:49:08,906 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=160524.0, ans=0.035 2024-09-23 02:49:13,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=160524.0, ans=0.2 2024-09-23 02:49:38,713 INFO [train.py:1198] (2/4) Epoch 9, batch 3250, loss[loss=0.2361, ctc_loss=0.1626, cr_loss=0.3676, over 16357.00 frames. ], tot_loss[loss=0.2585, ctc_loss=0.181, cr_loss=0.3873, over 3341480.05 frames. ], batch size: 36, lr: 1.31e-02, grad_scale: 32.0 2024-09-23 02:49:50,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=160617.33333333334, ans=0.125 2024-09-23 02:49:54,281 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.170e+02 1.347e+02 1.464e+02 1.659e+02 3.194e+02, threshold=2.929e+02, percent-clipped=1.0 2024-09-23 02:50:02,534 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=160664.0, ans=0.125 2024-09-23 02:50:19,605 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=160710.66666666666, ans=0.1 2024-09-23 02:50:27,439 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=160757.33333333334, ans=0.2 2024-09-23 02:50:38,222 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=160757.33333333334, ans=0.125 2024-09-23 02:50:41,415 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=160804.0, ans=0.125 2024-09-23 02:50:48,115 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=20.58 vs. limit=22.5 2024-09-23 02:50:50,600 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=160804.0, ans=0.2 2024-09-23 02:50:50,626 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=160804.0, ans=0.07 2024-09-23 02:50:55,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=160850.66666666666, ans=0.0 2024-09-23 02:50:56,580 INFO [train.py:1198] (2/4) Epoch 9, batch 3300, loss[loss=0.2785, ctc_loss=0.1968, cr_loss=0.4084, over 17078.00 frames. ], tot_loss[loss=0.2585, ctc_loss=0.1809, cr_loss=0.3877, over 3354756.05 frames. ], batch size: 49, lr: 1.31e-02, grad_scale: 32.0 2024-09-23 02:51:48,877 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.88 vs. limit=6.0 2024-09-23 02:51:49,735 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=160990.66666666666, ans=0.125 2024-09-23 02:51:52,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=160990.66666666666, ans=0.0 2024-09-23 02:51:55,928 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=160990.66666666666, ans=0.0 2024-09-23 02:52:05,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=161037.33333333334, ans=0.025 2024-09-23 02:52:14,694 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=161084.0, ans=0.0 2024-09-23 02:52:15,710 INFO [train.py:1198] (2/4) Epoch 9, batch 3350, loss[loss=0.2522, ctc_loss=0.1749, cr_loss=0.3867, over 17010.00 frames. ], tot_loss[loss=0.2579, ctc_loss=0.1805, cr_loss=0.3872, over 3356162.17 frames. ], batch size: 56, lr: 1.31e-02, grad_scale: 32.0 2024-09-23 02:52:19,128 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.max_abs, batch_count=161084.0, ans=10.0 2024-09-23 02:52:28,731 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.32 vs. limit=15.0 2024-09-23 02:52:31,264 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.145e+02 1.345e+02 1.546e+02 1.825e+02 3.271e+02, threshold=3.093e+02, percent-clipped=1.0 2024-09-23 02:52:47,135 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=161177.33333333334, ans=0.1 2024-09-23 02:52:47,225 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=161177.33333333334, ans=0.0 2024-09-23 02:53:33,968 INFO [train.py:1198] (2/4) Epoch 9, batch 3400, loss[loss=0.2691, ctc_loss=0.1894, cr_loss=0.3987, over 17306.00 frames. ], tot_loss[loss=0.2568, ctc_loss=0.1796, cr_loss=0.386, over 3360889.67 frames. ], batch size: 46, lr: 1.31e-02, grad_scale: 32.0 2024-09-23 02:53:41,926 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=161317.33333333334, ans=0.025 2024-09-23 02:54:00,715 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=161364.0, ans=0.125 2024-09-23 02:54:02,080 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=161364.0, ans=0.0 2024-09-23 02:54:22,569 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=11.95 vs. limit=15.0 2024-09-23 02:54:30,467 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=161457.33333333334, ans=0.125 2024-09-23 02:54:51,710 INFO [train.py:1198] (2/4) Epoch 9, batch 3450, loss[loss=0.3164, ctc_loss=0.2399, cr_loss=0.3825, over 11313.00 frames. ], tot_loss[loss=0.2575, ctc_loss=0.1803, cr_loss=0.3859, over 3347259.22 frames. ], batch size: 123, lr: 1.31e-02, grad_scale: 16.0 2024-09-23 02:54:54,958 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=161550.66666666666, ans=0.125 2024-09-23 02:55:08,996 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.145e+02 1.381e+02 1.554e+02 1.871e+02 2.951e+02, threshold=3.107e+02, percent-clipped=0.0 2024-09-23 02:55:53,425 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=6.04 vs. limit=15.0 2024-09-23 02:56:12,425 INFO [train.py:1198] (2/4) Epoch 9, batch 3500, loss[loss=0.2463, ctc_loss=0.1754, cr_loss=0.3545, over 17308.00 frames. ], tot_loss[loss=0.2572, ctc_loss=0.1803, cr_loss=0.3849, over 3344013.67 frames. ], batch size: 51, lr: 1.31e-02, grad_scale: 16.0 2024-09-23 02:56:18,903 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=161784.0, ans=0.125 2024-09-23 02:56:20,440 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=161784.0, ans=0.125 2024-09-23 02:56:44,111 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=161877.33333333334, ans=0.125 2024-09-23 02:56:53,094 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=161877.33333333334, ans=0.0 2024-09-23 02:57:18,710 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=161970.66666666666, ans=0.125 2024-09-23 02:57:23,312 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=161970.66666666666, ans=0.09899494936611666 2024-09-23 02:57:32,419 INFO [train.py:1198] (2/4) Epoch 9, batch 3550, loss[loss=0.2427, ctc_loss=0.1695, cr_loss=0.3664, over 17023.00 frames. ], tot_loss[loss=0.2571, ctc_loss=0.1799, cr_loss=0.3857, over 3351768.79 frames. ], batch size: 44, lr: 1.31e-02, grad_scale: 16.0 2024-09-23 02:57:44,105 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.63 vs. limit=6.0 2024-09-23 02:57:49,779 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.397e+02 1.544e+02 1.862e+02 4.630e+02, threshold=3.088e+02, percent-clipped=2.0 2024-09-23 02:57:56,242 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=162064.0, ans=0.125 2024-09-23 02:58:10,416 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=162110.66666666666, ans=0.125 2024-09-23 02:58:19,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=162157.33333333334, ans=0.0 2024-09-23 02:58:41,797 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=162204.0, ans=0.025 2024-09-23 02:58:52,212 INFO [train.py:1198] (2/4) Epoch 9, batch 3600, loss[loss=0.2526, ctc_loss=0.1762, cr_loss=0.382, over 17159.00 frames. ], tot_loss[loss=0.2572, ctc_loss=0.1799, cr_loss=0.3865, over 3351768.59 frames. ], batch size: 45, lr: 1.31e-02, grad_scale: 32.0 2024-09-23 02:59:09,540 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=162297.33333333334, ans=0.125 2024-09-23 02:59:09,644 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=162297.33333333334, ans=0.0 2024-09-23 02:59:12,835 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=162297.33333333334, ans=0.125 2024-09-23 02:59:30,546 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.55 vs. limit=15.0 2024-09-23 02:59:33,495 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=17.07 vs. limit=22.5 2024-09-23 02:59:46,942 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=162390.66666666666, ans=0.125 2024-09-23 03:00:01,115 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=162437.33333333334, ans=0.1 2024-09-23 03:00:10,090 INFO [train.py:1198] (2/4) Epoch 9, batch 3650, loss[loss=0.2763, ctc_loss=0.1949, cr_loss=0.4073, over 17014.00 frames. ], tot_loss[loss=0.2584, ctc_loss=0.1808, cr_loss=0.3878, over 3349906.76 frames. ], batch size: 52, lr: 1.31e-02, grad_scale: 32.0 2024-09-23 03:00:27,444 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.156e+02 1.388e+02 1.522e+02 1.765e+02 2.573e+02, threshold=3.044e+02, percent-clipped=0.0 2024-09-23 03:00:36,374 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=3.17 vs. limit=15.0 2024-09-23 03:01:25,680 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=12.62 vs. limit=15.0 2024-09-23 03:01:30,907 INFO [train.py:1198] (2/4) Epoch 9, batch 3700, loss[loss=0.2252, ctc_loss=0.1554, cr_loss=0.3492, over 17035.00 frames. ], tot_loss[loss=0.257, ctc_loss=0.1796, cr_loss=0.387, over 3361850.27 frames. ], batch size: 39, lr: 1.31e-02, grad_scale: 32.0 2024-09-23 03:01:34,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=162717.33333333334, ans=0.1 2024-09-23 03:01:35,880 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=162717.33333333334, ans=0.2 2024-09-23 03:01:36,310 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.43 vs. limit=6.0 2024-09-23 03:01:39,006 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=162717.33333333334, ans=0.125 2024-09-23 03:01:43,364 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=162717.33333333334, ans=0.1 2024-09-23 03:02:06,828 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=162810.66666666666, ans=0.125 2024-09-23 03:02:08,831 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten.whitening_limit, batch_count=162810.66666666666, ans=15.0 2024-09-23 03:02:11,581 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=162810.66666666666, ans=0.0 2024-09-23 03:02:22,510 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=162857.33333333334, ans=0.125 2024-09-23 03:02:38,397 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.56 vs. limit=15.0 2024-09-23 03:02:48,722 INFO [train.py:1198] (2/4) Epoch 9, batch 3750, loss[loss=0.2358, ctc_loss=0.1617, cr_loss=0.3707, over 17200.00 frames. ], tot_loss[loss=0.2563, ctc_loss=0.1791, cr_loss=0.3861, over 3353499.83 frames. ], batch size: 41, lr: 1.31e-02, grad_scale: 32.0 2024-09-23 03:03:04,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=162997.33333333334, ans=0.1 2024-09-23 03:03:05,821 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.180e+02 1.286e+02 1.440e+02 1.620e+02 2.372e+02, threshold=2.880e+02, percent-clipped=0.0 2024-09-23 03:03:17,068 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=162997.33333333334, ans=0.025 2024-09-23 03:03:19,034 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=13.09 vs. limit=22.5 2024-09-23 03:03:28,012 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=163044.0, ans=0.125 2024-09-23 03:03:31,304 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=163044.0, ans=0.0 2024-09-23 03:03:53,485 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=163137.33333333334, ans=0.09899494936611666 2024-09-23 03:04:07,105 INFO [train.py:1198] (2/4) Epoch 9, batch 3800, loss[loss=0.3078, ctc_loss=0.2178, cr_loss=0.4502, over 17023.00 frames. ], tot_loss[loss=0.2556, ctc_loss=0.1785, cr_loss=0.3851, over 3343924.02 frames. ], batch size: 56, lr: 1.30e-02, grad_scale: 32.0 2024-09-23 03:04:08,902 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 03:04:16,683 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=163184.0, ans=0.125 2024-09-23 03:04:30,787 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=163230.66666666666, ans=0.2 2024-09-23 03:04:34,154 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten.whitening_limit, batch_count=163230.66666666666, ans=15.0 2024-09-23 03:04:40,268 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=163277.33333333334, ans=0.0 2024-09-23 03:04:40,467 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=7.68 vs. limit=12.0 2024-09-23 03:04:56,009 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=163324.0, ans=0.125 2024-09-23 03:05:06,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=163324.0, ans=0.125 2024-09-23 03:05:09,973 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=163370.66666666666, ans=0.125 2024-09-23 03:05:25,695 INFO [train.py:1198] (2/4) Epoch 9, batch 3850, loss[loss=0.242, ctc_loss=0.1678, cr_loss=0.3711, over 16961.00 frames. ], tot_loss[loss=0.2571, ctc_loss=0.18, cr_loss=0.3858, over 3320147.67 frames. ], batch size: 42, lr: 1.30e-02, grad_scale: 32.0 2024-09-23 03:05:41,331 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=163464.0, ans=0.125 2024-09-23 03:05:42,523 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.149e+02 1.391e+02 1.511e+02 1.701e+02 2.274e+02, threshold=3.022e+02, percent-clipped=0.0 2024-09-23 03:05:53,398 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 03:06:01,552 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=163510.66666666666, ans=0.125 2024-09-23 03:06:03,265 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten.whitening_limit, batch_count=163510.66666666666, ans=15.0 2024-09-23 03:06:07,500 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=163510.66666666666, ans=0.1 2024-09-23 03:06:18,128 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=163557.33333333334, ans=0.0 2024-09-23 03:06:32,051 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.94 vs. limit=6.0 2024-09-23 03:07:26,827 INFO [train.py:1198] (2/4) Epoch 10, batch 0, loss[loss=0.2493, ctc_loss=0.1716, cr_loss=0.3887, over 17288.00 frames. ], tot_loss[loss=0.2493, ctc_loss=0.1716, cr_loss=0.3887, over 17288.00 frames. ], batch size: 51, lr: 1.24e-02, grad_scale: 32.0 2024-09-23 03:07:26,827 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 03:07:41,772 INFO [train.py:1230] (2/4) Epoch 10, validation: loss=0.05143, ctc_loss=0.05143, cr_loss=7.705e-15, over 944034.00 frames. 2024-09-23 03:07:41,773 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 03:08:04,336 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=163678.66666666666, ans=0.0 2024-09-23 03:08:34,461 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=163772.0, ans=0.0 2024-09-23 03:08:35,924 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=163772.0, ans=0.125 2024-09-23 03:09:05,465 INFO [train.py:1198] (2/4) Epoch 10, batch 50, loss[loss=0.2811, ctc_loss=0.1998, cr_loss=0.4065, over 16078.00 frames. ], tot_loss[loss=0.2548, ctc_loss=0.1776, cr_loss=0.386, over 757494.03 frames. ], batch size: 74, lr: 1.24e-02, grad_scale: 32.0 2024-09-23 03:09:12,152 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=163865.33333333334, ans=0.125 2024-09-23 03:09:15,424 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=163865.33333333334, ans=0.125 2024-09-23 03:09:29,246 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.138e+02 1.322e+02 1.460e+02 1.757e+02 2.503e+02, threshold=2.921e+02, percent-clipped=0.0 2024-09-23 03:09:51,646 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=164005.33333333334, ans=0.0 2024-09-23 03:10:09,334 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=164052.0, ans=0.05 2024-09-23 03:10:24,753 INFO [train.py:1198] (2/4) Epoch 10, batch 100, loss[loss=0.2253, ctc_loss=0.1567, cr_loss=0.3429, over 17159.00 frames. ], tot_loss[loss=0.255, ctc_loss=0.1777, cr_loss=0.3862, over 1338120.76 frames. ], batch size: 48, lr: 1.24e-02, grad_scale: 32.0 2024-09-23 03:10:25,763 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=22.07 vs. limit=22.5 2024-09-23 03:10:44,336 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=164145.33333333334, ans=0.125 2024-09-23 03:11:07,015 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=164192.0, ans=0.025 2024-09-23 03:11:14,922 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=164238.66666666666, ans=0.0 2024-09-23 03:11:26,777 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.24 vs. limit=10.0 2024-09-23 03:11:46,742 INFO [train.py:1198] (2/4) Epoch 10, batch 150, loss[loss=0.2713, ctc_loss=0.1931, cr_loss=0.3911, over 16910.00 frames. ], tot_loss[loss=0.2543, ctc_loss=0.177, cr_loss=0.3865, over 1796071.24 frames. ], batch size: 58, lr: 1.24e-02, grad_scale: 32.0 2024-09-23 03:12:13,436 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.322e+02 1.418e+02 1.649e+02 2.765e+02, threshold=2.835e+02, percent-clipped=0.0 2024-09-23 03:12:16,992 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.max_abs, batch_count=164378.66666666666, ans=10.0 2024-09-23 03:12:29,685 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=164425.33333333334, ans=0.0 2024-09-23 03:12:36,073 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=164472.0, ans=0.125 2024-09-23 03:13:00,073 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=164518.66666666666, ans=0.1 2024-09-23 03:13:09,147 INFO [train.py:1198] (2/4) Epoch 10, batch 200, loss[loss=0.2428, ctc_loss=0.1702, cr_loss=0.3634, over 17082.00 frames. ], tot_loss[loss=0.2533, ctc_loss=0.1762, cr_loss=0.3854, over 2148643.50 frames. ], batch size: 43, lr: 1.24e-02, grad_scale: 32.0 2024-09-23 03:13:11,050 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=164565.33333333334, ans=0.125 2024-09-23 03:13:15,814 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=164565.33333333334, ans=0.125 2024-09-23 03:13:34,503 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=164612.0, ans=0.0 2024-09-23 03:13:35,931 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=164612.0, ans=0.0 2024-09-23 03:13:37,597 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=164612.0, ans=0.125 2024-09-23 03:13:49,949 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=164658.66666666666, ans=0.125 2024-09-23 03:13:49,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=164658.66666666666, ans=0.0 2024-09-23 03:13:53,388 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=164658.66666666666, ans=0.0 2024-09-23 03:14:01,101 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=164705.33333333334, ans=0.025 2024-09-23 03:14:22,851 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=164752.0, ans=0.0 2024-09-23 03:14:23,254 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=11.53 vs. limit=15.0 2024-09-23 03:14:33,529 INFO [train.py:1198] (2/4) Epoch 10, batch 250, loss[loss=0.2979, ctc_loss=0.207, cr_loss=0.4545, over 16646.00 frames. ], tot_loss[loss=0.2541, ctc_loss=0.1771, cr_loss=0.385, over 2424158.05 frames. ], batch size: 66, lr: 1.23e-02, grad_scale: 32.0 2024-09-23 03:14:48,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=164845.33333333334, ans=0.0 2024-09-23 03:14:55,758 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=164845.33333333334, ans=0.125 2024-09-23 03:14:57,018 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.087e+02 1.299e+02 1.403e+02 1.575e+02 2.434e+02, threshold=2.805e+02, percent-clipped=0.0 2024-09-23 03:15:05,683 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.whiten, num_groups=1, num_channels=512, metric=4.63 vs. limit=12.0 2024-09-23 03:15:24,191 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=164938.66666666666, ans=0.1 2024-09-23 03:15:55,550 INFO [train.py:1198] (2/4) Epoch 10, batch 300, loss[loss=0.2673, ctc_loss=0.1906, cr_loss=0.3835, over 16766.00 frames. ], tot_loss[loss=0.2535, ctc_loss=0.1766, cr_loss=0.3846, over 2636171.58 frames. ], batch size: 61, lr: 1.23e-02, grad_scale: 32.0 2024-09-23 03:15:57,425 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=165032.0, ans=0.125 2024-09-23 03:16:11,422 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=165078.66666666666, ans=0.0 2024-09-23 03:16:43,261 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 03:16:53,468 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=14.47 vs. limit=22.5 2024-09-23 03:17:05,304 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=165218.66666666666, ans=0.125 2024-09-23 03:17:13,456 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=165218.66666666666, ans=0.125 2024-09-23 03:17:14,914 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=165218.66666666666, ans=0.125 2024-09-23 03:17:17,895 INFO [train.py:1198] (2/4) Epoch 10, batch 350, loss[loss=0.2581, ctc_loss=0.1814, cr_loss=0.3833, over 16983.00 frames. ], tot_loss[loss=0.2551, ctc_loss=0.1778, cr_loss=0.3862, over 2792135.76 frames. ], batch size: 53, lr: 1.23e-02, grad_scale: 32.0 2024-09-23 03:17:42,143 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.177e+02 1.372e+02 1.514e+02 1.677e+02 2.269e+02, threshold=3.028e+02, percent-clipped=0.0 2024-09-23 03:18:04,520 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=165405.33333333334, ans=0.0 2024-09-23 03:18:43,353 INFO [train.py:1198] (2/4) Epoch 10, batch 400, loss[loss=0.2143, ctc_loss=0.1456, cr_loss=0.3434, over 15887.00 frames. ], tot_loss[loss=0.2554, ctc_loss=0.178, cr_loss=0.387, over 2915915.80 frames. ], batch size: 35, lr: 1.23e-02, grad_scale: 32.0 2024-09-23 03:18:45,268 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=165498.66666666666, ans=0.125 2024-09-23 03:19:26,956 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=12.49 vs. limit=15.0 2024-09-23 03:19:32,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=165638.66666666666, ans=0.025 2024-09-23 03:19:32,959 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=165638.66666666666, ans=0.125 2024-09-23 03:19:47,223 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=165685.33333333334, ans=0.2 2024-09-23 03:19:55,299 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=165685.33333333334, ans=0.0 2024-09-23 03:20:03,259 INFO [train.py:1198] (2/4) Epoch 10, batch 450, loss[loss=0.2616, ctc_loss=0.1794, cr_loss=0.4113, over 16798.00 frames. ], tot_loss[loss=0.2556, ctc_loss=0.1782, cr_loss=0.387, over 2998863.07 frames. ], batch size: 61, lr: 1.23e-02, grad_scale: 32.0 2024-09-23 03:20:03,492 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=165732.0, ans=0.0 2024-09-23 03:20:13,130 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=165732.0, ans=0.025 2024-09-23 03:20:16,339 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=165732.0, ans=0.025 2024-09-23 03:20:26,966 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.153e+02 1.328e+02 1.495e+02 1.704e+02 3.618e+02, threshold=2.990e+02, percent-clipped=1.0 2024-09-23 03:20:52,882 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.32 vs. limit=12.0 2024-09-23 03:21:11,452 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=165918.66666666666, ans=0.5 2024-09-23 03:21:25,282 INFO [train.py:1198] (2/4) Epoch 10, batch 500, loss[loss=0.2587, ctc_loss=0.1795, cr_loss=0.3962, over 17301.00 frames. ], tot_loss[loss=0.2545, ctc_loss=0.1773, cr_loss=0.3858, over 3072667.73 frames. ], batch size: 46, lr: 1.23e-02, grad_scale: 32.0 2024-09-23 03:21:34,184 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.13 vs. limit=10.0 2024-09-23 03:22:00,637 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=166058.66666666666, ans=0.0 2024-09-23 03:22:08,010 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=166058.66666666666, ans=0.0 2024-09-23 03:22:39,037 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=13.30 vs. limit=15.0 2024-09-23 03:22:47,886 INFO [train.py:1198] (2/4) Epoch 10, batch 550, loss[loss=0.2305, ctc_loss=0.1588, cr_loss=0.3584, over 17111.00 frames. ], tot_loss[loss=0.254, ctc_loss=0.177, cr_loss=0.3847, over 3136191.63 frames. ], batch size: 40, lr: 1.23e-02, grad_scale: 32.0 2024-09-23 03:23:11,787 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.067e+02 1.308e+02 1.376e+02 1.532e+02 2.311e+02, threshold=2.752e+02, percent-clipped=0.0 2024-09-23 03:23:14,890 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=166245.33333333334, ans=0.125 2024-09-23 03:23:26,885 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.39 vs. limit=6.0 2024-09-23 03:23:51,156 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=166338.66666666666, ans=0.1 2024-09-23 03:24:07,067 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=166385.33333333334, ans=0.2 2024-09-23 03:24:13,174 INFO [train.py:1198] (2/4) Epoch 10, batch 600, loss[loss=0.2087, ctc_loss=0.145, cr_loss=0.3185, over 17246.00 frames. ], tot_loss[loss=0.2527, ctc_loss=0.1761, cr_loss=0.383, over 3192830.70 frames. ], batch size: 44, lr: 1.23e-02, grad_scale: 32.0 2024-09-23 03:24:39,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=166478.66666666666, ans=0.125 2024-09-23 03:24:45,420 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer_ff3.min_abs, batch_count=166525.33333333334, ans=0.2 2024-09-23 03:25:03,067 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=166572.0, ans=0.0 2024-09-23 03:25:21,871 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=166618.66666666666, ans=0.0 2024-09-23 03:25:32,728 INFO [train.py:1198] (2/4) Epoch 10, batch 650, loss[loss=0.1956, ctc_loss=0.1339, cr_loss=0.3086, over 17199.00 frames. ], tot_loss[loss=0.2503, ctc_loss=0.1743, cr_loss=0.38, over 3231581.22 frames. ], batch size: 41, lr: 1.23e-02, grad_scale: 32.0 2024-09-23 03:25:43,755 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=166665.33333333334, ans=0.0 2024-09-23 03:25:59,417 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.059e+02 1.346e+02 1.493e+02 1.797e+02 2.927e+02, threshold=2.987e+02, percent-clipped=1.0 2024-09-23 03:26:10,161 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=5.13 vs. limit=10.0 2024-09-23 03:26:32,125 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=166805.33333333334, ans=0.2 2024-09-23 03:26:50,885 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=166852.0, ans=0.0 2024-09-23 03:26:55,505 INFO [train.py:1198] (2/4) Epoch 10, batch 700, loss[loss=0.3295, ctc_loss=0.243, cr_loss=0.4325, over 12007.00 frames. ], tot_loss[loss=0.2509, ctc_loss=0.1747, cr_loss=0.3814, over 3262996.04 frames. ], batch size: 123, lr: 1.23e-02, grad_scale: 32.0 2024-09-23 03:26:59,077 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=166898.66666666666, ans=0.0 2024-09-23 03:26:59,131 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=166898.66666666666, ans=0.0 2024-09-23 03:27:03,949 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=166898.66666666666, ans=0.2 2024-09-23 03:27:22,642 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=13.99 vs. limit=15.0 2024-09-23 03:27:51,432 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=167038.66666666666, ans=0.0 2024-09-23 03:27:53,487 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=21.68 vs. limit=22.5 2024-09-23 03:28:20,933 INFO [train.py:1198] (2/4) Epoch 10, batch 750, loss[loss=0.1997, ctc_loss=0.1367, cr_loss=0.315, over 17230.00 frames. ], tot_loss[loss=0.2501, ctc_loss=0.174, cr_loss=0.3805, over 3281496.95 frames. ], batch size: 42, lr: 1.23e-02, grad_scale: 32.0 2024-09-23 03:28:21,250 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=167132.0, ans=0.0 2024-09-23 03:28:46,465 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=167178.66666666666, ans=0.125 2024-09-23 03:28:47,711 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.321e+02 1.498e+02 1.811e+02 2.765e+02, threshold=2.996e+02, percent-clipped=0.0 2024-09-23 03:28:59,029 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=167225.33333333334, ans=0.125 2024-09-23 03:29:27,895 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=13.07 vs. limit=15.0 2024-09-23 03:29:34,081 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=167318.66666666666, ans=0.125 2024-09-23 03:29:43,236 INFO [train.py:1198] (2/4) Epoch 10, batch 800, loss[loss=0.2494, ctc_loss=0.1715, cr_loss=0.3894, over 17085.00 frames. ], tot_loss[loss=0.2512, ctc_loss=0.1748, cr_loss=0.3818, over 3297564.03 frames. ], batch size: 46, lr: 1.23e-02, grad_scale: 32.0 2024-09-23 03:29:51,451 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=167365.33333333334, ans=0.0 2024-09-23 03:29:56,226 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=167365.33333333334, ans=0.125 2024-09-23 03:30:52,790 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=167552.0, ans=0.0 2024-09-23 03:30:59,423 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=11.82 vs. limit=12.0 2024-09-23 03:31:05,258 INFO [train.py:1198] (2/4) Epoch 10, batch 850, loss[loss=0.2534, ctc_loss=0.179, cr_loss=0.372, over 17303.00 frames. ], tot_loss[loss=0.2514, ctc_loss=0.1751, cr_loss=0.3815, over 3305890.89 frames. ], batch size: 49, lr: 1.22e-02, grad_scale: 32.0 2024-09-23 03:31:07,662 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=512, metric=2.42 vs. limit=15.0 2024-09-23 03:31:16,700 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=167598.66666666666, ans=0.0 2024-09-23 03:31:23,094 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=167645.33333333334, ans=0.125 2024-09-23 03:31:28,120 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=13.11 vs. limit=15.0 2024-09-23 03:31:29,264 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.131e+02 1.338e+02 1.474e+02 1.669e+02 2.399e+02, threshold=2.948e+02, percent-clipped=0.0 2024-09-23 03:31:59,757 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=167738.66666666666, ans=0.2 2024-09-23 03:32:09,494 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=167785.33333333334, ans=0.125 2024-09-23 03:32:23,248 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=167785.33333333334, ans=0.0 2024-09-23 03:32:27,740 INFO [train.py:1198] (2/4) Epoch 10, batch 900, loss[loss=0.2154, ctc_loss=0.1482, cr_loss=0.3361, over 17065.00 frames. ], tot_loss[loss=0.2533, ctc_loss=0.1767, cr_loss=0.383, over 3301971.87 frames. ], batch size: 39, lr: 1.22e-02, grad_scale: 32.0 2024-09-23 03:33:01,639 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.max_abs, batch_count=167925.33333333334, ans=10.0 2024-09-23 03:33:23,250 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=167972.0, ans=0.125 2024-09-23 03:33:55,325 INFO [train.py:1198] (2/4) Epoch 10, batch 950, loss[loss=0.2612, ctc_loss=0.1812, cr_loss=0.3999, over 17026.00 frames. ], tot_loss[loss=0.2537, ctc_loss=0.1769, cr_loss=0.3839, over 3319439.90 frames. ], batch size: 44, lr: 1.22e-02, grad_scale: 32.0 2024-09-23 03:34:18,367 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.46 vs. limit=15.0 2024-09-23 03:34:18,979 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.083e+02 1.276e+02 1.450e+02 1.704e+02 3.070e+02, threshold=2.900e+02, percent-clipped=1.0 2024-09-23 03:34:55,642 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=168205.33333333334, ans=0.125 2024-09-23 03:35:14,461 INFO [train.py:1198] (2/4) Epoch 10, batch 1000, loss[loss=0.3261, ctc_loss=0.2379, cr_loss=0.4406, over 11387.00 frames. ], tot_loss[loss=0.2533, ctc_loss=0.1766, cr_loss=0.3836, over 3323043.00 frames. ], batch size: 123, lr: 1.22e-02, grad_scale: 32.0 2024-09-23 03:35:35,298 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=168345.33333333334, ans=0.125 2024-09-23 03:36:06,544 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=168438.66666666666, ans=0.0 2024-09-23 03:36:20,100 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.75 vs. limit=6.0 2024-09-23 03:36:23,986 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=168485.33333333334, ans=0.0 2024-09-23 03:36:36,363 INFO [train.py:1198] (2/4) Epoch 10, batch 1050, loss[loss=0.2591, ctc_loss=0.1826, cr_loss=0.3822, over 17306.00 frames. ], tot_loss[loss=0.2535, ctc_loss=0.1767, cr_loss=0.384, over 3335384.32 frames. ], batch size: 49, lr: 1.22e-02, grad_scale: 32.0 2024-09-23 03:37:00,705 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.301e+02 1.448e+02 1.658e+02 2.854e+02, threshold=2.897e+02, percent-clipped=0.0 2024-09-23 03:37:30,616 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=168672.0, ans=0.125 2024-09-23 03:37:40,815 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=14.82 vs. limit=22.5 2024-09-23 03:37:41,805 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=168718.66666666666, ans=0.09899494936611666 2024-09-23 03:37:46,918 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=13.86 vs. limit=15.0 2024-09-23 03:37:58,856 INFO [train.py:1198] (2/4) Epoch 10, batch 1100, loss[loss=0.2926, ctc_loss=0.2072, cr_loss=0.4272, over 16480.00 frames. ], tot_loss[loss=0.2535, ctc_loss=0.1767, cr_loss=0.384, over 3337545.80 frames. ], batch size: 66, lr: 1.22e-02, grad_scale: 32.0 2024-09-23 03:38:15,962 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=168812.0, ans=0.0 2024-09-23 03:39:23,799 INFO [train.py:1198] (2/4) Epoch 10, batch 1150, loss[loss=0.2255, ctc_loss=0.1556, cr_loss=0.3496, over 17294.00 frames. ], tot_loss[loss=0.2531, ctc_loss=0.1765, cr_loss=0.3832, over 3333535.20 frames. ], batch size: 49, lr: 1.22e-02, grad_scale: 32.0 2024-09-23 03:39:33,707 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=168998.66666666666, ans=0.1 2024-09-23 03:39:47,598 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.191e+02 1.373e+02 1.514e+02 1.720e+02 2.403e+02, threshold=3.028e+02, percent-clipped=0.0 2024-09-23 03:39:49,623 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=169045.33333333334, ans=0.1 2024-09-23 03:40:04,065 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=169092.0, ans=0.1 2024-09-23 03:40:11,191 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.whiten.whitening_limit, batch_count=169138.66666666666, ans=15.0 2024-09-23 03:40:18,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=169138.66666666666, ans=0.0 2024-09-23 03:40:33,436 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=169185.33333333334, ans=0.1 2024-09-23 03:40:37,465 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.90 vs. limit=15.0 2024-09-23 03:40:45,947 INFO [train.py:1198] (2/4) Epoch 10, batch 1200, loss[loss=0.2727, ctc_loss=0.1919, cr_loss=0.4038, over 16905.00 frames. ], tot_loss[loss=0.2531, ctc_loss=0.1763, cr_loss=0.3836, over 3345685.59 frames. ], batch size: 58, lr: 1.22e-02, grad_scale: 32.0 2024-09-23 03:41:06,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=169278.66666666666, ans=0.0 2024-09-23 03:41:18,077 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=169325.33333333334, ans=0.0 2024-09-23 03:41:29,180 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=169325.33333333334, ans=0.125 2024-09-23 03:41:29,185 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=169325.33333333334, ans=0.125 2024-09-23 03:41:38,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=169372.0, ans=0.125 2024-09-23 03:41:58,753 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=13.88 vs. limit=22.5 2024-09-23 03:42:07,868 INFO [train.py:1198] (2/4) Epoch 10, batch 1250, loss[loss=0.2549, ctc_loss=0.1791, cr_loss=0.3791, over 17234.00 frames. ], tot_loss[loss=0.252, ctc_loss=0.1754, cr_loss=0.3828, over 3348178.61 frames. ], batch size: 50, lr: 1.22e-02, grad_scale: 32.0 2024-09-23 03:42:09,770 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=169465.33333333334, ans=0.07 2024-09-23 03:42:19,333 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=169465.33333333334, ans=0.125 2024-09-23 03:42:31,957 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.271e+02 1.375e+02 1.546e+02 2.384e+02, threshold=2.750e+02, percent-clipped=0.0 2024-09-23 03:42:46,540 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=169558.66666666666, ans=0.125 2024-09-23 03:42:51,349 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=169558.66666666666, ans=0.125 2024-09-23 03:42:52,786 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=169558.66666666666, ans=0.125 2024-09-23 03:43:05,025 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=169605.33333333334, ans=0.0 2024-09-23 03:43:06,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=169605.33333333334, ans=0.125 2024-09-23 03:43:20,047 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.00 vs. limit=15.0 2024-09-23 03:43:29,974 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=169652.0, ans=0.025 2024-09-23 03:43:32,877 INFO [train.py:1198] (2/4) Epoch 10, batch 1300, loss[loss=0.2764, ctc_loss=0.1973, cr_loss=0.3958, over 16050.00 frames. ], tot_loss[loss=0.252, ctc_loss=0.1754, cr_loss=0.3827, over 3347677.35 frames. ], batch size: 74, lr: 1.22e-02, grad_scale: 32.0 2024-09-23 03:43:45,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=169698.66666666666, ans=0.125 2024-09-23 03:43:59,250 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.54 vs. limit=6.0 2024-09-23 03:44:17,699 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=169792.0, ans=0.125 2024-09-23 03:44:29,068 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.77 vs. limit=15.0 2024-09-23 03:44:40,495 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=12.74 vs. limit=15.0 2024-09-23 03:44:44,747 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=169885.33333333334, ans=0.125 2024-09-23 03:44:52,271 INFO [train.py:1198] (2/4) Epoch 10, batch 1350, loss[loss=0.2535, ctc_loss=0.1752, cr_loss=0.3912, over 17090.00 frames. ], tot_loss[loss=0.2501, ctc_loss=0.1741, cr_loss=0.3803, over 3355551.71 frames. ], batch size: 43, lr: 1.22e-02, grad_scale: 32.0 2024-09-23 03:45:16,272 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.144e+02 1.346e+02 1.479e+02 1.695e+02 2.554e+02, threshold=2.958e+02, percent-clipped=0.0 2024-09-23 03:45:23,200 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=170025.33333333334, ans=0.125 2024-09-23 03:45:23,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=170025.33333333334, ans=0.125 2024-09-23 03:45:27,356 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=170025.33333333334, ans=0.0 2024-09-23 03:45:44,911 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=170072.0, ans=0.125 2024-09-23 03:45:47,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=170072.0, ans=0.1 2024-09-23 03:45:51,755 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.75 vs. limit=12.0 2024-09-23 03:46:09,208 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=3.20 vs. limit=15.0 2024-09-23 03:46:14,376 INFO [train.py:1198] (2/4) Epoch 10, batch 1400, loss[loss=0.2636, ctc_loss=0.1813, cr_loss=0.4116, over 17173.00 frames. ], tot_loss[loss=0.2522, ctc_loss=0.1756, cr_loss=0.3834, over 3353271.24 frames. ], batch size: 55, lr: 1.22e-02, grad_scale: 16.0 2024-09-23 03:46:40,856 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=5.10 vs. limit=6.0 2024-09-23 03:47:27,116 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=170352.0, ans=0.125 2024-09-23 03:47:36,188 INFO [train.py:1198] (2/4) Epoch 10, batch 1450, loss[loss=0.2542, ctc_loss=0.1724, cr_loss=0.409, over 17273.00 frames. ], tot_loss[loss=0.252, ctc_loss=0.1753, cr_loss=0.3834, over 3351667.38 frames. ], batch size: 46, lr: 1.22e-02, grad_scale: 16.0 2024-09-23 03:47:52,377 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=170445.33333333334, ans=0.025 2024-09-23 03:47:54,043 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=170445.33333333334, ans=0.125 2024-09-23 03:47:57,004 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=170445.33333333334, ans=0.1 2024-09-23 03:47:57,634 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.81 vs. limit=6.0 2024-09-23 03:48:04,178 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.358e+02 1.487e+02 1.729e+02 2.482e+02, threshold=2.974e+02, percent-clipped=0.0 2024-09-23 03:48:22,706 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=170492.0, ans=0.125 2024-09-23 03:48:34,365 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1.whitening_limit, batch_count=170538.66666666666, ans=10.0 2024-09-23 03:48:48,358 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=170585.33333333334, ans=0.0 2024-09-23 03:48:53,873 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=20.28 vs. limit=22.5 2024-09-23 03:49:00,803 INFO [train.py:1198] (2/4) Epoch 10, batch 1500, loss[loss=0.2496, ctc_loss=0.1714, cr_loss=0.3912, over 17257.00 frames. ], tot_loss[loss=0.2532, ctc_loss=0.1762, cr_loss=0.3854, over 3354427.05 frames. ], batch size: 44, lr: 1.21e-02, grad_scale: 16.0 2024-09-23 03:49:01,150 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=170632.0, ans=0.125 2024-09-23 03:49:40,983 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 03:50:03,397 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=170818.66666666666, ans=0.125 2024-09-23 03:50:22,863 INFO [train.py:1198] (2/4) Epoch 10, batch 1550, loss[loss=0.2396, ctc_loss=0.169, cr_loss=0.3528, over 17177.00 frames. ], tot_loss[loss=0.2529, ctc_loss=0.1759, cr_loss=0.385, over 3355744.15 frames. ], batch size: 41, lr: 1.21e-02, grad_scale: 8.0 2024-09-23 03:50:31,139 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=170865.33333333334, ans=0.125 2024-09-23 03:50:40,362 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=170912.0, ans=0.0 2024-09-23 03:50:49,522 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.180e+02 1.376e+02 1.480e+02 1.669e+02 2.824e+02, threshold=2.960e+02, percent-clipped=0.0 2024-09-23 03:51:01,497 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=7.08 vs. limit=10.0 2024-09-23 03:51:04,161 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=170958.66666666666, ans=0.0 2024-09-23 03:51:13,677 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=171005.33333333334, ans=0.0 2024-09-23 03:51:21,443 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=171005.33333333334, ans=0.0 2024-09-23 03:51:29,659 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer_na.min_abs, batch_count=171052.0, ans=0.02 2024-09-23 03:51:31,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=171052.0, ans=0.125 2024-09-23 03:51:31,318 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=171052.0, ans=0.2 2024-09-23 03:51:38,139 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.13 vs. limit=15.0 2024-09-23 03:51:41,965 INFO [train.py:1198] (2/4) Epoch 10, batch 1600, loss[loss=0.2334, ctc_loss=0.1636, cr_loss=0.349, over 17156.00 frames. ], tot_loss[loss=0.2529, ctc_loss=0.1759, cr_loss=0.3848, over 3354414.01 frames. ], batch size: 40, lr: 1.21e-02, grad_scale: 16.0 2024-09-23 03:52:02,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=171145.33333333334, ans=0.1 2024-09-23 03:52:21,334 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=171192.0, ans=0.1 2024-09-23 03:52:56,103 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=171285.33333333334, ans=0.025 2024-09-23 03:53:01,062 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.74 vs. limit=22.5 2024-09-23 03:53:06,734 INFO [train.py:1198] (2/4) Epoch 10, batch 1650, loss[loss=0.2566, ctc_loss=0.1788, cr_loss=0.389, over 16909.00 frames. ], tot_loss[loss=0.2537, ctc_loss=0.1766, cr_loss=0.3856, over 3355199.26 frames. ], batch size: 58, lr: 1.21e-02, grad_scale: 16.0 2024-09-23 03:53:35,049 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=171378.66666666666, ans=0.0 2024-09-23 03:53:36,437 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.109e+02 1.322e+02 1.432e+02 1.629e+02 2.855e+02, threshold=2.864e+02, percent-clipped=0.0 2024-09-23 03:53:40,666 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=13.44 vs. limit=15.0 2024-09-23 03:53:51,206 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=171425.33333333334, ans=0.0 2024-09-23 03:53:57,849 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=171472.0, ans=0.125 2024-09-23 03:54:00,867 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=171472.0, ans=0.0 2024-09-23 03:54:29,206 INFO [train.py:1198] (2/4) Epoch 10, batch 1700, loss[loss=0.2332, ctc_loss=0.1623, cr_loss=0.3546, over 17197.00 frames. ], tot_loss[loss=0.2545, ctc_loss=0.1772, cr_loss=0.3861, over 3350571.42 frames. ], batch size: 47, lr: 1.21e-02, grad_scale: 16.0 2024-09-23 03:54:34,651 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=13.45 vs. limit=22.5 2024-09-23 03:55:21,115 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=171705.33333333334, ans=0.125 2024-09-23 03:55:45,793 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=20.37 vs. limit=22.5 2024-09-23 03:55:51,290 INFO [train.py:1198] (2/4) Epoch 10, batch 1750, loss[loss=0.2667, ctc_loss=0.1901, cr_loss=0.3828, over 17360.00 frames. ], tot_loss[loss=0.255, ctc_loss=0.1777, cr_loss=0.3865, over 3342719.44 frames. ], batch size: 48, lr: 1.21e-02, grad_scale: 16.0 2024-09-23 03:56:06,007 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=171845.33333333334, ans=0.125 2024-09-23 03:56:18,403 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.152e+02 1.367e+02 1.590e+02 1.877e+02 2.998e+02, threshold=3.180e+02, percent-clipped=1.0 2024-09-23 03:56:20,686 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.69 vs. limit=10.0 2024-09-23 03:56:57,762 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=171985.33333333334, ans=0.2 2024-09-23 03:56:57,772 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=171985.33333333334, ans=0.0 2024-09-23 03:57:08,892 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=171985.33333333334, ans=0.125 2024-09-23 03:57:13,328 INFO [train.py:1198] (2/4) Epoch 10, batch 1800, loss[loss=0.2598, ctc_loss=0.1814, cr_loss=0.3918, over 16909.00 frames. ], tot_loss[loss=0.2543, ctc_loss=0.1772, cr_loss=0.3855, over 3346092.38 frames. ], batch size: 58, lr: 1.21e-02, grad_scale: 16.0 2024-09-23 03:57:29,615 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=172078.66666666666, ans=0.1 2024-09-23 03:57:52,754 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=172125.33333333334, ans=0.0 2024-09-23 03:58:02,124 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=172172.0, ans=0.125 2024-09-23 03:58:15,660 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=172172.0, ans=0.0 2024-09-23 03:58:37,573 INFO [train.py:1198] (2/4) Epoch 10, batch 1850, loss[loss=0.2373, ctc_loss=0.1651, cr_loss=0.3614, over 16303.00 frames. ], tot_loss[loss=0.2538, ctc_loss=0.1769, cr_loss=0.3845, over 3343979.35 frames. ], batch size: 36, lr: 1.21e-02, grad_scale: 16.0 2024-09-23 03:58:40,970 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=172265.33333333334, ans=0.125 2024-09-23 03:58:44,225 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=172265.33333333334, ans=0.125 2024-09-23 03:58:55,007 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=172312.0, ans=0.125 2024-09-23 03:58:57,043 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten.whitening_limit, batch_count=172312.0, ans=15.0 2024-09-23 03:59:04,288 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.087e+02 1.347e+02 1.457e+02 1.655e+02 2.749e+02, threshold=2.915e+02, percent-clipped=0.0 2024-09-23 03:59:35,374 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.39 vs. limit=15.0 2024-09-23 03:59:41,556 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=172452.0, ans=0.025 2024-09-23 03:59:57,221 INFO [train.py:1198] (2/4) Epoch 10, batch 1900, loss[loss=0.2583, ctc_loss=0.1792, cr_loss=0.3951, over 17050.00 frames. ], tot_loss[loss=0.2524, ctc_loss=0.1758, cr_loss=0.3831, over 3349419.88 frames. ], batch size: 39, lr: 1.21e-02, grad_scale: 16.0 2024-09-23 03:59:57,480 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=172498.66666666666, ans=0.125 2024-09-23 04:00:11,800 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=14.43 vs. limit=22.5 2024-09-23 04:00:12,895 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=172498.66666666666, ans=0.0 2024-09-23 04:00:14,478 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=172545.33333333334, ans=0.125 2024-09-23 04:00:24,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=172545.33333333334, ans=0.0 2024-09-23 04:01:11,935 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=172685.33333333334, ans=0.015 2024-09-23 04:01:15,730 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.66 vs. limit=15.0 2024-09-23 04:01:19,643 INFO [train.py:1198] (2/4) Epoch 10, batch 1950, loss[loss=0.253, ctc_loss=0.1728, cr_loss=0.4013, over 17212.00 frames. ], tot_loss[loss=0.2501, ctc_loss=0.1738, cr_loss=0.3814, over 3362778.11 frames. ], batch size: 50, lr: 1.21e-02, grad_scale: 16.0 2024-09-23 04:01:40,380 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=172778.66666666666, ans=0.125 2024-09-23 04:01:48,821 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.191e+02 1.405e+02 1.574e+02 1.755e+02 2.409e+02, threshold=3.148e+02, percent-clipped=0.0 2024-09-23 04:02:05,277 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=172825.33333333334, ans=0.125 2024-09-23 04:02:10,103 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=172872.0, ans=0.0 2024-09-23 04:02:10,646 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=6.15 vs. limit=12.0 2024-09-23 04:02:14,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=172872.0, ans=0.1 2024-09-23 04:02:25,640 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=172918.66666666666, ans=0.125 2024-09-23 04:02:42,801 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=172965.33333333334, ans=0.025 2024-09-23 04:02:42,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=172965.33333333334, ans=0.0 2024-09-23 04:02:43,966 INFO [train.py:1198] (2/4) Epoch 10, batch 2000, loss[loss=0.2974, ctc_loss=0.2165, cr_loss=0.4048, over 12158.00 frames. ], tot_loss[loss=0.2492, ctc_loss=0.1732, cr_loss=0.3801, over 3368262.22 frames. ], batch size: 123, lr: 1.21e-02, grad_scale: 32.0 2024-09-23 04:02:53,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=172965.33333333334, ans=0.0 2024-09-23 04:03:01,940 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 04:03:09,146 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=173012.0, ans=0.125 2024-09-23 04:03:10,875 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=173012.0, ans=0.125 2024-09-23 04:03:23,406 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=173058.66666666666, ans=0.0 2024-09-23 04:03:37,131 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.whiten, num_groups=1, num_channels=192, metric=4.07 vs. limit=12.0 2024-09-23 04:03:43,476 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=3.12 vs. limit=15.0 2024-09-23 04:04:06,412 INFO [train.py:1198] (2/4) Epoch 10, batch 2050, loss[loss=0.2389, ctc_loss=0.1659, cr_loss=0.3649, over 17202.00 frames. ], tot_loss[loss=0.2491, ctc_loss=0.1731, cr_loss=0.3802, over 3367621.95 frames. ], batch size: 47, lr: 1.21e-02, grad_scale: 32.0 2024-09-23 04:04:34,982 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.316e+02 1.437e+02 1.596e+02 3.726e+02, threshold=2.874e+02, percent-clipped=1.0 2024-09-23 04:04:36,964 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=173292.0, ans=0.125 2024-09-23 04:05:09,750 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=173338.66666666666, ans=0.125 2024-09-23 04:05:18,187 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.04 vs. limit=15.0 2024-09-23 04:05:20,814 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.min_positive, batch_count=173385.33333333334, ans=0.025 2024-09-23 04:05:28,497 INFO [train.py:1198] (2/4) Epoch 10, batch 2100, loss[loss=0.2797, ctc_loss=0.1936, cr_loss=0.4302, over 17094.00 frames. ], tot_loss[loss=0.2499, ctc_loss=0.1737, cr_loss=0.3809, over 3357598.42 frames. ], batch size: 49, lr: 1.20e-02, grad_scale: 16.0 2024-09-23 04:05:29,349 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=9.71 vs. limit=15.0 2024-09-23 04:06:47,561 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=3.93 vs. limit=6.0 2024-09-23 04:06:49,963 INFO [train.py:1198] (2/4) Epoch 10, batch 2150, loss[loss=0.2908, ctc_loss=0.1967, cr_loss=0.4706, over 17001.00 frames. ], tot_loss[loss=0.2509, ctc_loss=0.1744, cr_loss=0.382, over 3357505.05 frames. ], batch size: 53, lr: 1.20e-02, grad_scale: 16.0 2024-09-23 04:07:18,784 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.167e+02 1.317e+02 1.458e+02 1.666e+02 3.135e+02, threshold=2.917e+02, percent-clipped=1.0 2024-09-23 04:07:22,430 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=173758.66666666666, ans=0.0 2024-09-23 04:07:34,465 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 04:07:34,520 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=173758.66666666666, ans=0.2 2024-09-23 04:07:45,342 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=173805.33333333334, ans=0.2 2024-09-23 04:07:50,111 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=173805.33333333334, ans=0.125 2024-09-23 04:08:11,717 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=173852.0, ans=0.0 2024-09-23 04:08:12,118 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.07 vs. limit=10.0 2024-09-23 04:08:14,679 INFO [train.py:1198] (2/4) Epoch 10, batch 2200, loss[loss=0.2196, ctc_loss=0.1526, cr_loss=0.3347, over 17166.00 frames. ], tot_loss[loss=0.2516, ctc_loss=0.1749, cr_loss=0.3833, over 3362406.32 frames. ], batch size: 45, lr: 1.20e-02, grad_scale: 16.0 2024-09-23 04:08:19,878 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=173898.66666666666, ans=0.0 2024-09-23 04:08:21,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=173898.66666666666, ans=0.125 2024-09-23 04:08:38,824 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=173945.33333333334, ans=0.125 2024-09-23 04:08:39,616 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=7.85 vs. limit=15.0 2024-09-23 04:08:45,358 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=173992.0, ans=0.0 2024-09-23 04:09:17,844 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.07 vs. limit=15.0 2024-09-23 04:09:20,505 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=174085.33333333334, ans=0.0 2024-09-23 04:09:34,447 INFO [train.py:1198] (2/4) Epoch 10, batch 2250, loss[loss=0.2249, ctc_loss=0.156, cr_loss=0.3446, over 17260.00 frames. ], tot_loss[loss=0.2498, ctc_loss=0.1735, cr_loss=0.3814, over 3364940.39 frames. ], batch size: 42, lr: 1.20e-02, grad_scale: 16.0 2024-09-23 04:09:36,191 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=174132.0, ans=0.125 2024-09-23 04:10:04,609 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=174178.66666666666, ans=0.0 2024-09-23 04:10:05,826 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.156e+02 1.382e+02 1.475e+02 1.727e+02 2.787e+02, threshold=2.949e+02, percent-clipped=0.0 2024-09-23 04:10:07,711 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=174225.33333333334, ans=0.125 2024-09-23 04:10:12,518 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=174225.33333333334, ans=0.1 2024-09-23 04:10:23,884 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=174272.0, ans=0.05 2024-09-23 04:10:27,151 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=174272.0, ans=0.125 2024-09-23 04:10:33,461 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=174272.0, ans=0.07 2024-09-23 04:10:46,594 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.02 vs. limit=10.0 2024-09-23 04:10:56,693 INFO [train.py:1198] (2/4) Epoch 10, batch 2300, loss[loss=0.3084, ctc_loss=0.2233, cr_loss=0.4254, over 11692.00 frames. ], tot_loss[loss=0.2496, ctc_loss=0.1734, cr_loss=0.3813, over 3366151.15 frames. ], batch size: 124, lr: 1.20e-02, grad_scale: 16.0 2024-09-23 04:11:06,407 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=174365.33333333334, ans=0.125 2024-09-23 04:11:14,184 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=174412.0, ans=0.0 2024-09-23 04:11:19,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=174412.0, ans=0.125 2024-09-23 04:11:59,454 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten.whitening_limit, batch_count=174505.33333333334, ans=15.0 2024-09-23 04:12:03,467 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=174552.0, ans=0.125 2024-09-23 04:12:19,122 INFO [train.py:1198] (2/4) Epoch 10, batch 2350, loss[loss=0.2783, ctc_loss=0.1952, cr_loss=0.4152, over 17045.00 frames. ], tot_loss[loss=0.251, ctc_loss=0.1745, cr_loss=0.3826, over 3363577.89 frames. ], batch size: 52, lr: 1.20e-02, grad_scale: 16.0 2024-09-23 04:12:19,486 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=174598.66666666666, ans=0.125 2024-09-23 04:12:50,299 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.141e+02 1.308e+02 1.414e+02 1.626e+02 2.428e+02, threshold=2.828e+02, percent-clipped=0.0 2024-09-23 04:13:17,289 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.whiten, num_groups=1, num_channels=512, metric=4.84 vs. limit=12.0 2024-09-23 04:13:20,090 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=174738.66666666666, ans=0.125 2024-09-23 04:13:23,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=174738.66666666666, ans=0.125 2024-09-23 04:13:29,304 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=174785.33333333334, ans=0.1 2024-09-23 04:13:38,887 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=174785.33333333334, ans=0.1 2024-09-23 04:13:43,455 INFO [train.py:1198] (2/4) Epoch 10, batch 2400, loss[loss=0.2901, ctc_loss=0.2039, cr_loss=0.4312, over 17001.00 frames. ], tot_loss[loss=0.2509, ctc_loss=0.1743, cr_loss=0.3831, over 3358242.95 frames. ], batch size: 52, lr: 1.20e-02, grad_scale: 32.0 2024-09-23 04:14:39,840 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=174972.0, ans=0.125 2024-09-23 04:14:48,090 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=11.18 vs. limit=22.5 2024-09-23 04:14:55,752 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.06 vs. limit=15.0 2024-09-23 04:15:06,053 INFO [train.py:1198] (2/4) Epoch 10, batch 2450, loss[loss=0.3419, ctc_loss=0.2517, cr_loss=0.4511, over 11733.00 frames. ], tot_loss[loss=0.251, ctc_loss=0.1745, cr_loss=0.3826, over 3357462.35 frames. ], batch size: 123, lr: 1.20e-02, grad_scale: 16.0 2024-09-23 04:15:06,276 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=175065.33333333334, ans=0.125 2024-09-23 04:15:11,178 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=175065.33333333334, ans=0.125 2024-09-23 04:15:25,981 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=14.08 vs. limit=15.0 2024-09-23 04:15:27,095 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=175112.0, ans=0.1 2024-09-23 04:15:31,860 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=175112.0, ans=0.125 2024-09-23 04:15:36,388 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.295e+02 1.392e+02 1.667e+02 2.800e+02, threshold=2.783e+02, percent-clipped=0.0 2024-09-23 04:15:37,521 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module1.whiten, num_groups=1, num_channels=192, metric=6.29 vs. limit=15.0 2024-09-23 04:15:54,118 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=175205.33333333334, ans=0.025 2024-09-23 04:16:02,449 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=9.53 vs. limit=15.0 2024-09-23 04:16:07,755 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.90 vs. limit=6.0 2024-09-23 04:16:15,754 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.50 vs. limit=15.0 2024-09-23 04:16:25,985 INFO [train.py:1198] (2/4) Epoch 10, batch 2500, loss[loss=0.2539, ctc_loss=0.1774, cr_loss=0.3821, over 17208.00 frames. ], tot_loss[loss=0.2515, ctc_loss=0.175, cr_loss=0.3828, over 3353004.67 frames. ], batch size: 47, lr: 1.20e-02, grad_scale: 16.0 2024-09-23 04:16:26,410 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=175298.66666666666, ans=0.125 2024-09-23 04:16:51,280 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=175345.33333333334, ans=0.1 2024-09-23 04:17:02,162 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=175392.0, ans=0.125 2024-09-23 04:17:05,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=175392.0, ans=0.0 2024-09-23 04:17:12,034 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.max_positive, batch_count=175392.0, ans=0.95 2024-09-23 04:17:31,321 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=6.25 vs. limit=15.0 2024-09-23 04:17:43,285 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=175485.33333333334, ans=0.125 2024-09-23 04:17:50,877 INFO [train.py:1198] (2/4) Epoch 10, batch 2550, loss[loss=0.2167, ctc_loss=0.1434, cr_loss=0.3666, over 17269.00 frames. ], tot_loss[loss=0.2516, ctc_loss=0.1749, cr_loss=0.3835, over 3360742.67 frames. ], batch size: 42, lr: 1.20e-02, grad_scale: 16.0 2024-09-23 04:18:03,326 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=175532.0, ans=0.0 2024-09-23 04:18:23,452 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.153e+02 1.321e+02 1.441e+02 1.691e+02 2.698e+02, threshold=2.882e+02, percent-clipped=0.0 2024-09-23 04:18:31,682 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=175625.33333333334, ans=0.0 2024-09-23 04:18:38,044 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=175625.33333333334, ans=0.0 2024-09-23 04:18:44,410 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.min_positive, batch_count=175672.0, ans=0.05 2024-09-23 04:18:52,179 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=175672.0, ans=0.0 2024-09-23 04:19:12,531 INFO [train.py:1198] (2/4) Epoch 10, batch 2600, loss[loss=0.2475, ctc_loss=0.1698, cr_loss=0.3884, over 17351.00 frames. ], tot_loss[loss=0.2518, ctc_loss=0.1751, cr_loss=0.3837, over 3358015.09 frames. ], batch size: 48, lr: 1.20e-02, grad_scale: 16.0 2024-09-23 04:19:17,657 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=175765.33333333334, ans=0.125 2024-09-23 04:19:33,795 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 04:19:36,924 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.16 vs. limit=15.0 2024-09-23 04:20:34,898 INFO [train.py:1198] (2/4) Epoch 10, batch 2650, loss[loss=0.2072, ctc_loss=0.1421, cr_loss=0.3256, over 16340.00 frames. ], tot_loss[loss=0.2522, ctc_loss=0.1755, cr_loss=0.3836, over 3357156.31 frames. ], batch size: 36, lr: 1.20e-02, grad_scale: 16.0 2024-09-23 04:20:43,164 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=175998.66666666666, ans=0.125 2024-09-23 04:20:46,219 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=175998.66666666666, ans=0.1 2024-09-23 04:20:49,469 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=176045.33333333334, ans=0.2 2024-09-23 04:20:59,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=176045.33333333334, ans=0.025 2024-09-23 04:21:05,299 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.318e+02 1.384e+02 1.550e+02 2.652e+02, threshold=2.768e+02, percent-clipped=0.0 2024-09-23 04:21:13,507 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=176092.0, ans=0.2 2024-09-23 04:21:24,820 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=176138.66666666666, ans=0.125 2024-09-23 04:21:51,495 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=176185.33333333334, ans=0.2 2024-09-23 04:21:56,233 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=176232.0, ans=0.125 2024-09-23 04:21:57,389 INFO [train.py:1198] (2/4) Epoch 10, batch 2700, loss[loss=0.2824, ctc_loss=0.2023, cr_loss=0.4006, over 15128.00 frames. ], tot_loss[loss=0.2533, ctc_loss=0.1764, cr_loss=0.3846, over 3350610.64 frames. ], batch size: 89, lr: 1.20e-02, grad_scale: 16.0 2024-09-23 04:21:59,313 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=176232.0, ans=0.07 2024-09-23 04:22:02,468 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=176232.0, ans=0.125 2024-09-23 04:22:43,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=176325.33333333334, ans=0.1 2024-09-23 04:23:03,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=176372.0, ans=0.0 2024-09-23 04:23:06,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=176418.66666666666, ans=0.035 2024-09-23 04:23:22,731 INFO [train.py:1198] (2/4) Epoch 10, batch 2750, loss[loss=0.2031, ctc_loss=0.14, cr_loss=0.3155, over 16970.00 frames. ], tot_loss[loss=0.2526, ctc_loss=0.1758, cr_loss=0.3844, over 3357383.30 frames. ], batch size: 42, lr: 1.19e-02, grad_scale: 16.0 2024-09-23 04:23:51,891 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=176512.0, ans=0.025 2024-09-23 04:23:53,132 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.143e+02 1.357e+02 1.501e+02 1.751e+02 2.551e+02, threshold=3.001e+02, percent-clipped=0.0 2024-09-23 04:24:03,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=176558.66666666666, ans=0.0 2024-09-23 04:24:07,691 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=176558.66666666666, ans=0.125 2024-09-23 04:24:27,469 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.68 vs. limit=12.0 2024-09-23 04:24:28,922 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=17.90 vs. limit=22.5 2024-09-23 04:24:42,412 INFO [train.py:1198] (2/4) Epoch 10, batch 2800, loss[loss=0.2458, ctc_loss=0.1702, cr_loss=0.3776, over 17024.00 frames. ], tot_loss[loss=0.252, ctc_loss=0.1751, cr_loss=0.3844, over 3363009.94 frames. ], batch size: 52, lr: 1.19e-02, grad_scale: 32.0 2024-09-23 04:25:31,754 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=176838.66666666666, ans=0.1 2024-09-23 04:25:57,491 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=176885.33333333334, ans=0.125 2024-09-23 04:26:03,795 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=176932.0, ans=0.125 2024-09-23 04:26:05,052 INFO [train.py:1198] (2/4) Epoch 10, batch 2850, loss[loss=0.2127, ctc_loss=0.1473, cr_loss=0.3269, over 17283.00 frames. ], tot_loss[loss=0.2528, ctc_loss=0.1757, cr_loss=0.3853, over 3356870.97 frames. ], batch size: 46, lr: 1.19e-02, grad_scale: 32.0 2024-09-23 04:26:14,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=176932.0, ans=0.125 2024-09-23 04:26:18,112 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=176932.0, ans=0.0 2024-09-23 04:26:21,518 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=176978.66666666666, ans=0.0 2024-09-23 04:26:27,613 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=176978.66666666666, ans=0.2 2024-09-23 04:26:38,102 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.110e+02 1.308e+02 1.482e+02 1.673e+02 2.195e+02, threshold=2.964e+02, percent-clipped=0.0 2024-09-23 04:26:51,553 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=177025.33333333334, ans=0.125 2024-09-23 04:27:20,900 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=177118.66666666666, ans=0.0 2024-09-23 04:27:30,212 INFO [train.py:1198] (2/4) Epoch 10, batch 2900, loss[loss=0.278, ctc_loss=0.1926, cr_loss=0.4268, over 17212.00 frames. ], tot_loss[loss=0.253, ctc_loss=0.1761, cr_loss=0.3848, over 3344929.37 frames. ], batch size: 47, lr: 1.19e-02, grad_scale: 32.0 2024-09-23 04:27:33,727 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=177165.33333333334, ans=0.125 2024-09-23 04:27:56,934 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=177212.0, ans=0.0 2024-09-23 04:27:58,379 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=177212.0, ans=0.1 2024-09-23 04:28:03,342 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=177258.66666666666, ans=0.0 2024-09-23 04:28:17,745 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=177258.66666666666, ans=0.2 2024-09-23 04:28:52,968 INFO [train.py:1198] (2/4) Epoch 10, batch 2950, loss[loss=0.2094, ctc_loss=0.1393, cr_loss=0.3505, over 16949.00 frames. ], tot_loss[loss=0.2535, ctc_loss=0.1765, cr_loss=0.3848, over 3342002.83 frames. ], batch size: 42, lr: 1.19e-02, grad_scale: 32.0 2024-09-23 04:28:53,236 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=177398.66666666666, ans=0.125 2024-09-23 04:28:54,874 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=177398.66666666666, ans=0.125 2024-09-23 04:28:57,898 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=177398.66666666666, ans=0.2 2024-09-23 04:28:57,929 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=177398.66666666666, ans=0.0 2024-09-23 04:29:18,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=177445.33333333334, ans=0.125 2024-09-23 04:29:23,161 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.132e+02 1.333e+02 1.430e+02 1.579e+02 2.314e+02, threshold=2.860e+02, percent-clipped=0.0 2024-09-23 04:30:00,517 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer_ff2.min_abs, batch_count=177585.33333333334, ans=0.1 2024-09-23 04:30:02,105 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer_ff3.min_abs, batch_count=177585.33333333334, ans=0.2 2024-09-23 04:30:14,432 INFO [train.py:1198] (2/4) Epoch 10, batch 3000, loss[loss=0.2831, ctc_loss=0.1985, cr_loss=0.423, over 17168.00 frames. ], tot_loss[loss=0.2544, ctc_loss=0.1771, cr_loss=0.3867, over 3350716.02 frames. ], batch size: 45, lr: 1.19e-02, grad_scale: 32.0 2024-09-23 04:30:14,432 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 04:30:23,795 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.5.encoder.layers.1.self_attn_weights, attn_weights_entropy = tensor([2.3357, 4.3677, 4.1975, 4.4282], device='cuda:2') 2024-09-23 04:30:30,498 INFO [train.py:1230] (2/4) Epoch 10, validation: loss=0.04843, ctc_loss=0.04843, cr_loss=7.942e-15, over 944034.00 frames. 2024-09-23 04:30:30,499 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 04:30:36,061 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.75 vs. limit=15.0 2024-09-23 04:31:00,260 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=177725.33333333334, ans=0.125 2024-09-23 04:31:16,011 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=177772.0, ans=0.125 2024-09-23 04:31:22,040 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=177772.0, ans=0.125 2024-09-23 04:31:23,671 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=177772.0, ans=0.125 2024-09-23 04:31:29,792 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=177772.0, ans=0.125 2024-09-23 04:31:48,170 INFO [train.py:1198] (2/4) Epoch 10, batch 3050, loss[loss=0.287, ctc_loss=0.2012, cr_loss=0.429, over 16542.00 frames. ], tot_loss[loss=0.2539, ctc_loss=0.1768, cr_loss=0.3855, over 3349977.06 frames. ], batch size: 66, lr: 1.19e-02, grad_scale: 32.0 2024-09-23 04:31:53,372 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=177865.33333333334, ans=0.0 2024-09-23 04:31:57,743 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=177865.33333333334, ans=0.125 2024-09-23 04:31:59,532 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=177865.33333333334, ans=0.125 2024-09-23 04:32:02,601 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=177912.0, ans=0.125 2024-09-23 04:32:16,430 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=177912.0, ans=0.125 2024-09-23 04:32:17,672 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.279e+02 1.408e+02 1.568e+02 2.746e+02, threshold=2.816e+02, percent-clipped=0.0 2024-09-23 04:32:33,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=178005.33333333334, ans=0.0 2024-09-23 04:32:42,227 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=13.61 vs. limit=22.5 2024-09-23 04:32:43,230 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=178005.33333333334, ans=0.025 2024-09-23 04:32:50,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=178052.0, ans=0.025 2024-09-23 04:32:50,729 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=178052.0, ans=0.125 2024-09-23 04:32:55,474 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=178052.0, ans=0.1 2024-09-23 04:33:06,402 INFO [train.py:1198] (2/4) Epoch 10, batch 3100, loss[loss=0.2836, ctc_loss=0.1992, cr_loss=0.422, over 17009.00 frames. ], tot_loss[loss=0.2525, ctc_loss=0.1757, cr_loss=0.3842, over 3351926.61 frames. ], batch size: 51, lr: 1.19e-02, grad_scale: 32.0 2024-09-23 04:33:14,392 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=178098.66666666666, ans=0.125 2024-09-23 04:33:43,751 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=178192.0, ans=0.04949747468305833 2024-09-23 04:33:52,854 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=178192.0, ans=0.125 2024-09-23 04:33:59,222 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=178238.66666666666, ans=0.125 2024-09-23 04:34:22,968 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=178285.33333333334, ans=0.0 2024-09-23 04:34:27,442 INFO [train.py:1198] (2/4) Epoch 10, batch 3150, loss[loss=0.2565, ctc_loss=0.1781, cr_loss=0.3924, over 17045.00 frames. ], tot_loss[loss=0.2522, ctc_loss=0.1754, cr_loss=0.3839, over 3344001.63 frames. ], batch size: 52, lr: 1.19e-02, grad_scale: 32.0 2024-09-23 04:34:37,043 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 04:34:56,601 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.289e+02 1.420e+02 1.589e+02 2.247e+02, threshold=2.840e+02, percent-clipped=0.0 2024-09-23 04:34:56,886 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=178425.33333333334, ans=0.1 2024-09-23 04:34:57,371 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=7.83 vs. limit=15.0 2024-09-23 04:35:10,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=178425.33333333334, ans=0.125 2024-09-23 04:35:18,004 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=178472.0, ans=0.09899494936611666 2024-09-23 04:35:49,621 INFO [train.py:1198] (2/4) Epoch 10, batch 3200, loss[loss=0.2318, ctc_loss=0.1615, cr_loss=0.3514, over 17174.00 frames. ], tot_loss[loss=0.2513, ctc_loss=0.1747, cr_loss=0.383, over 3352485.04 frames. ], batch size: 45, lr: 1.19e-02, grad_scale: 32.0 2024-09-23 04:36:02,180 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=178565.33333333334, ans=0.025 2024-09-23 04:36:10,865 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.44 vs. limit=6.0 2024-09-23 04:36:19,345 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=178658.66666666666, ans=0.125 2024-09-23 04:36:50,726 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=178752.0, ans=0.125 2024-09-23 04:37:04,771 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.11 vs. limit=15.0 2024-09-23 04:37:07,405 INFO [train.py:1198] (2/4) Epoch 10, batch 3250, loss[loss=0.2376, ctc_loss=0.1618, cr_loss=0.3789, over 17362.00 frames. ], tot_loss[loss=0.2512, ctc_loss=0.1744, cr_loss=0.3837, over 3357766.16 frames. ], batch size: 48, lr: 1.19e-02, grad_scale: 32.0 2024-09-23 04:37:09,338 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=178798.66666666666, ans=0.125 2024-09-23 04:37:15,358 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=178798.66666666666, ans=0.125 2024-09-23 04:37:27,922 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 04:37:37,060 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.295e+02 1.395e+02 1.531e+02 3.942e+02, threshold=2.791e+02, percent-clipped=1.0 2024-09-23 04:38:16,366 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=178985.33333333334, ans=0.0 2024-09-23 04:38:25,402 INFO [train.py:1198] (2/4) Epoch 10, batch 3300, loss[loss=0.2548, ctc_loss=0.1731, cr_loss=0.4088, over 17019.00 frames. ], tot_loss[loss=0.2518, ctc_loss=0.175, cr_loss=0.3842, over 3350624.09 frames. ], batch size: 44, lr: 1.19e-02, grad_scale: 32.0 2024-09-23 04:39:06,197 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=13.38 vs. limit=15.0 2024-09-23 04:39:19,950 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=179172.0, ans=0.125 2024-09-23 04:39:23,112 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=179172.0, ans=0.2 2024-09-23 04:39:28,143 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=179218.66666666666, ans=0.0 2024-09-23 04:39:29,610 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=179218.66666666666, ans=0.2 2024-09-23 04:39:32,765 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=179218.66666666666, ans=0.2 2024-09-23 04:39:43,186 INFO [train.py:1198] (2/4) Epoch 10, batch 3350, loss[loss=0.2132, ctc_loss=0.1471, cr_loss=0.3306, over 17060.00 frames. ], tot_loss[loss=0.2516, ctc_loss=0.1748, cr_loss=0.3841, over 3357205.01 frames. ], batch size: 40, lr: 1.19e-02, grad_scale: 32.0 2024-09-23 04:40:15,061 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.120e+02 1.385e+02 1.553e+02 1.795e+02 2.936e+02, threshold=3.106e+02, percent-clipped=1.0 2024-09-23 04:40:24,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.max_positive, batch_count=179358.66666666666, ans=0.95 2024-09-23 04:40:35,784 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=179405.33333333334, ans=0.1 2024-09-23 04:40:38,955 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=13.27 vs. limit=15.0 2024-09-23 04:40:45,521 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.62 vs. limit=10.0 2024-09-23 04:40:46,740 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.88 vs. limit=22.5 2024-09-23 04:41:03,095 INFO [train.py:1198] (2/4) Epoch 10, batch 3400, loss[loss=0.2742, ctc_loss=0.1892, cr_loss=0.4254, over 17215.00 frames. ], tot_loss[loss=0.2519, ctc_loss=0.175, cr_loss=0.3843, over 3362913.18 frames. ], batch size: 50, lr: 1.19e-02, grad_scale: 32.0 2024-09-23 04:41:17,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=179545.33333333334, ans=0.125 2024-09-23 04:41:21,203 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module1.whiten, num_groups=1, num_channels=192, metric=7.65 vs. limit=15.0 2024-09-23 04:41:25,028 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=179545.33333333334, ans=0.125 2024-09-23 04:41:27,022 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.35 vs. limit=12.0 2024-09-23 04:41:29,521 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=179545.33333333334, ans=0.2 2024-09-23 04:41:35,966 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=179592.0, ans=0.1 2024-09-23 04:42:21,369 INFO [train.py:1198] (2/4) Epoch 10, batch 3450, loss[loss=0.2651, ctc_loss=0.1822, cr_loss=0.4142, over 17152.00 frames. ], tot_loss[loss=0.252, ctc_loss=0.175, cr_loss=0.3846, over 3360916.77 frames. ], batch size: 45, lr: 1.18e-02, grad_scale: 32.0 2024-09-23 04:42:21,750 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=179732.0, ans=0.1 2024-09-23 04:42:50,603 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.109e+02 1.316e+02 1.451e+02 1.743e+02 2.784e+02, threshold=2.902e+02, percent-clipped=0.0 2024-09-23 04:42:53,956 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=179825.33333333334, ans=0.1 2024-09-23 04:43:07,968 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=179872.0, ans=0.1 2024-09-23 04:43:09,401 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=179872.0, ans=0.125 2024-09-23 04:43:10,947 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=179872.0, ans=0.125 2024-09-23 04:43:22,066 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=179872.0, ans=0.125 2024-09-23 04:43:23,414 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=179918.66666666666, ans=0.125 2024-09-23 04:43:31,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=179918.66666666666, ans=0.125 2024-09-23 04:43:40,111 INFO [train.py:1198] (2/4) Epoch 10, batch 3500, loss[loss=0.2353, ctc_loss=0.1646, cr_loss=0.3538, over 17160.00 frames. ], tot_loss[loss=0.2513, ctc_loss=0.1746, cr_loss=0.3834, over 3354622.78 frames. ], batch size: 48, lr: 1.18e-02, grad_scale: 32.0 2024-09-23 04:43:42,094 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.min_positive, batch_count=179965.33333333334, ans=0.025 2024-09-23 04:43:49,828 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=179965.33333333334, ans=0.0 2024-09-23 04:44:16,739 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.07 vs. limit=6.0 2024-09-23 04:45:02,136 INFO [train.py:1198] (2/4) Epoch 10, batch 3550, loss[loss=0.2342, ctc_loss=0.1634, cr_loss=0.3539, over 17078.00 frames. ], tot_loss[loss=0.2513, ctc_loss=0.1747, cr_loss=0.3831, over 3354266.43 frames. ], batch size: 43, lr: 1.18e-02, grad_scale: 32.0 2024-09-23 04:45:10,741 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.99 vs. limit=6.0 2024-09-23 04:45:24,605 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=180245.33333333334, ans=0.125 2024-09-23 04:45:27,729 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=180245.33333333334, ans=0.2 2024-09-23 04:45:29,254 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=180245.33333333334, ans=0.125 2024-09-23 04:45:32,190 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.126e+02 1.310e+02 1.457e+02 1.783e+02 3.691e+02, threshold=2.913e+02, percent-clipped=1.0 2024-09-23 04:45:52,074 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=13.73 vs. limit=22.5 2024-09-23 04:46:20,829 INFO [train.py:1198] (2/4) Epoch 10, batch 3600, loss[loss=0.2728, ctc_loss=0.1954, cr_loss=0.3874, over 16065.00 frames. ], tot_loss[loss=0.251, ctc_loss=0.1744, cr_loss=0.3832, over 3354815.11 frames. ], batch size: 74, lr: 1.18e-02, grad_scale: 32.0 2024-09-23 04:46:25,586 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=180432.0, ans=0.125 2024-09-23 04:46:25,615 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=180432.0, ans=0.125 2024-09-23 04:46:30,298 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=180432.0, ans=0.125 2024-09-23 04:46:38,875 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=5.03 vs. limit=10.0 2024-09-23 04:46:58,725 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=180525.33333333334, ans=0.125 2024-09-23 04:47:38,885 INFO [train.py:1198] (2/4) Epoch 10, batch 3650, loss[loss=0.2798, ctc_loss=0.1953, cr_loss=0.4222, over 16036.00 frames. ], tot_loss[loss=0.2524, ctc_loss=0.1755, cr_loss=0.3845, over 3347389.29 frames. ], batch size: 74, lr: 1.18e-02, grad_scale: 32.0 2024-09-23 04:48:08,242 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.122e+02 1.307e+02 1.399e+02 1.525e+02 2.249e+02, threshold=2.799e+02, percent-clipped=0.0 2024-09-23 04:48:12,501 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.min_positive, batch_count=180758.66666666666, ans=0.025 2024-09-23 04:48:15,740 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=180758.66666666666, ans=0.1 2024-09-23 04:48:28,133 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=180805.33333333334, ans=0.125 2024-09-23 04:48:29,694 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=180805.33333333334, ans=0.0 2024-09-23 04:48:57,313 INFO [train.py:1198] (2/4) Epoch 10, batch 3700, loss[loss=0.2445, ctc_loss=0.173, cr_loss=0.3576, over 16367.00 frames. ], tot_loss[loss=0.2517, ctc_loss=0.1749, cr_loss=0.3839, over 3355764.78 frames. ], batch size: 36, lr: 1.18e-02, grad_scale: 32.0 2024-09-23 04:49:03,859 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=180898.66666666666, ans=0.125 2024-09-23 04:49:15,384 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=2.71 vs. limit=15.0 2024-09-23 04:49:28,559 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=180992.0, ans=0.125 2024-09-23 04:49:54,951 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=181038.66666666666, ans=0.0 2024-09-23 04:50:12,131 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=181085.33333333334, ans=0.0 2024-09-23 04:50:16,591 INFO [train.py:1198] (2/4) Epoch 10, batch 3750, loss[loss=0.1957, ctc_loss=0.1352, cr_loss=0.3022, over 16681.00 frames. ], tot_loss[loss=0.2498, ctc_loss=0.1736, cr_loss=0.3814, over 3349874.18 frames. ], batch size: 37, lr: 1.18e-02, grad_scale: 32.0 2024-09-23 04:50:29,845 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=18.01 vs. limit=22.5 2024-09-23 04:50:41,733 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=181178.66666666666, ans=0.125 2024-09-23 04:50:46,085 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.196e+02 1.346e+02 1.480e+02 1.709e+02 3.123e+02, threshold=2.960e+02, percent-clipped=1.0 2024-09-23 04:51:31,545 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=181318.66666666666, ans=0.125 2024-09-23 04:51:34,558 INFO [train.py:1198] (2/4) Epoch 10, batch 3800, loss[loss=0.2151, ctc_loss=0.1457, cr_loss=0.3472, over 17178.00 frames. ], tot_loss[loss=0.2483, ctc_loss=0.1723, cr_loss=0.3797, over 3346981.08 frames. ], batch size: 41, lr: 1.18e-02, grad_scale: 32.0 2024-09-23 04:51:37,837 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=181365.33333333334, ans=0.025 2024-09-23 04:52:01,357 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=181412.0, ans=0.125 2024-09-23 04:52:05,998 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=181458.66666666666, ans=0.0 2024-09-23 04:52:18,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=181458.66666666666, ans=0.0 2024-09-23 04:52:35,556 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=181552.0, ans=0.125 2024-09-23 04:52:53,254 INFO [train.py:1198] (2/4) Epoch 10, batch 3850, loss[loss=0.2707, ctc_loss=0.1868, cr_loss=0.4193, over 17052.00 frames. ], tot_loss[loss=0.2477, ctc_loss=0.172, cr_loss=0.3785, over 3340625.18 frames. ], batch size: 52, lr: 1.18e-02, grad_scale: 32.0 2024-09-23 04:53:09,417 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=181645.33333333334, ans=0.125 2024-09-23 04:53:22,407 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.346e+02 1.514e+02 1.745e+02 2.888e+02, threshold=3.027e+02, percent-clipped=0.0 2024-09-23 04:53:36,217 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=181692.0, ans=0.0 2024-09-23 04:53:50,519 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=181738.66666666666, ans=0.125 2024-09-23 04:53:54,820 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=181785.33333333334, ans=0.025 2024-09-23 04:53:59,350 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 04:54:55,685 INFO [train.py:1198] (2/4) Epoch 11, batch 0, loss[loss=0.2662, ctc_loss=0.1879, cr_loss=0.3911, over 17094.00 frames. ], tot_loss[loss=0.2662, ctc_loss=0.1879, cr_loss=0.3911, over 17094.00 frames. ], batch size: 49, lr: 1.12e-02, grad_scale: 32.0 2024-09-23 04:54:55,685 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 04:55:11,307 INFO [train.py:1230] (2/4) Epoch 11, validation: loss=0.04963, ctc_loss=0.04963, cr_loss=7.372e-15, over 944034.00 frames. 2024-09-23 04:55:11,308 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 04:55:19,092 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=181813.33333333334, ans=0.125 2024-09-23 04:55:40,038 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=181860.0, ans=0.0 2024-09-23 04:55:51,216 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=181906.66666666666, ans=0.1 2024-09-23 04:55:52,814 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=181906.66666666666, ans=0.1 2024-09-23 04:55:54,614 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.56 vs. limit=10.0 2024-09-23 04:55:59,223 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=181906.66666666666, ans=0.1 2024-09-23 04:56:29,754 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=182000.0, ans=0.125 2024-09-23 04:56:34,324 INFO [train.py:1198] (2/4) Epoch 11, batch 50, loss[loss=0.2378, ctc_loss=0.1611, cr_loss=0.3832, over 16788.00 frames. ], tot_loss[loss=0.2573, ctc_loss=0.179, cr_loss=0.3917, over 763147.42 frames. ], batch size: 61, lr: 1.12e-02, grad_scale: 32.0 2024-09-23 04:56:34,696 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=182046.66666666666, ans=0.05 2024-09-23 04:56:44,233 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=182046.66666666666, ans=0.125 2024-09-23 04:57:12,613 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.085e+02 1.435e+02 1.595e+02 1.830e+02 2.692e+02, threshold=3.191e+02, percent-clipped=0.0 2024-09-23 04:57:14,479 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=182140.0, ans=0.125 2024-09-23 04:57:40,183 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=10.61 vs. limit=12.0 2024-09-23 04:57:50,966 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=182233.33333333334, ans=0.05 2024-09-23 04:57:53,900 INFO [train.py:1198] (2/4) Epoch 11, batch 100, loss[loss=0.2732, ctc_loss=0.1884, cr_loss=0.4236, over 16678.00 frames. ], tot_loss[loss=0.2519, ctc_loss=0.1744, cr_loss=0.3873, over 1343566.29 frames. ], batch size: 37, lr: 1.12e-02, grad_scale: 16.0 2024-09-23 04:58:16,557 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=182326.66666666666, ans=0.0 2024-09-23 04:58:51,297 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 04:59:10,790 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=7.15 vs. limit=15.0 2024-09-23 04:59:16,233 INFO [train.py:1198] (2/4) Epoch 11, batch 150, loss[loss=0.224, ctc_loss=0.1536, cr_loss=0.3519, over 17292.00 frames. ], tot_loss[loss=0.25, ctc_loss=0.1731, cr_loss=0.3848, over 1791394.55 frames. ], batch size: 46, lr: 1.12e-02, grad_scale: 16.0 2024-09-23 04:59:16,922 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=14.77 vs. limit=15.0 2024-09-23 04:59:24,851 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=11.11 vs. limit=22.5 2024-09-23 04:59:35,815 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=182560.0, ans=0.2 2024-09-23 04:59:36,450 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=8.28 vs. limit=15.0 2024-09-23 04:59:36,579 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=3.25 vs. limit=15.0 2024-09-23 04:59:57,641 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.353e+02 1.511e+02 1.758e+02 2.701e+02, threshold=3.021e+02, percent-clipped=0.0 2024-09-23 05:00:02,762 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 05:00:32,554 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 05:00:35,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=182700.0, ans=0.125 2024-09-23 05:00:37,136 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=182700.0, ans=0.1 2024-09-23 05:00:41,713 INFO [train.py:1198] (2/4) Epoch 11, batch 200, loss[loss=0.249, ctc_loss=0.1725, cr_loss=0.3822, over 17297.00 frames. ], tot_loss[loss=0.2516, ctc_loss=0.1746, cr_loss=0.3852, over 2131749.76 frames. ], batch size: 49, lr: 1.12e-02, grad_scale: 16.0 2024-09-23 05:00:47,171 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=13.15 vs. limit=15.0 2024-09-23 05:01:01,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=182793.33333333334, ans=0.1 2024-09-23 05:01:17,025 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=182840.0, ans=0.125 2024-09-23 05:01:20,271 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=182840.0, ans=0.125 2024-09-23 05:01:28,197 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=182886.66666666666, ans=0.09899494936611666 2024-09-23 05:01:33,020 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=182886.66666666666, ans=0.125 2024-09-23 05:01:36,675 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.17 vs. limit=15.0 2024-09-23 05:01:58,499 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=182933.33333333334, ans=0.125 2024-09-23 05:02:01,466 INFO [train.py:1198] (2/4) Epoch 11, batch 250, loss[loss=0.2683, ctc_loss=0.1861, cr_loss=0.4114, over 17220.00 frames. ], tot_loss[loss=0.2499, ctc_loss=0.1733, cr_loss=0.383, over 2406452.60 frames. ], batch size: 55, lr: 1.12e-02, grad_scale: 16.0 2024-09-23 05:02:01,789 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=182980.0, ans=0.0 2024-09-23 05:02:27,296 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=183026.66666666666, ans=0.125 2024-09-23 05:02:36,924 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=183073.33333333334, ans=0.2 2024-09-23 05:02:39,687 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.107e+02 1.357e+02 1.521e+02 1.743e+02 2.579e+02, threshold=3.043e+02, percent-clipped=0.0 2024-09-23 05:02:45,485 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=14.18 vs. limit=22.5 2024-09-23 05:03:20,846 INFO [train.py:1198] (2/4) Epoch 11, batch 300, loss[loss=0.241, ctc_loss=0.1667, cr_loss=0.3718, over 17022.00 frames. ], tot_loss[loss=0.2473, ctc_loss=0.1712, cr_loss=0.3802, over 2625666.82 frames. ], batch size: 51, lr: 1.12e-02, grad_scale: 16.0 2024-09-23 05:03:59,492 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=183306.66666666666, ans=0.2 2024-09-23 05:04:19,928 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=183353.33333333334, ans=0.0 2024-09-23 05:04:25,200 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=7.84 vs. limit=15.0 2024-09-23 05:04:26,319 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=183400.0, ans=0.0 2024-09-23 05:04:29,383 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=183400.0, ans=0.1 2024-09-23 05:04:43,090 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=183400.0, ans=0.1 2024-09-23 05:04:49,182 INFO [train.py:1198] (2/4) Epoch 11, batch 350, loss[loss=0.2347, ctc_loss=0.1591, cr_loss=0.3781, over 17214.00 frames. ], tot_loss[loss=0.2483, ctc_loss=0.1718, cr_loss=0.3826, over 2793473.11 frames. ], batch size: 47, lr: 1.12e-02, grad_scale: 16.0 2024-09-23 05:04:49,825 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=5.30 vs. limit=12.0 2024-09-23 05:04:53,981 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=183446.66666666666, ans=0.2 2024-09-23 05:05:19,118 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.min_positive, batch_count=183493.33333333334, ans=0.05 2024-09-23 05:05:29,394 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.98 vs. limit=6.0 2024-09-23 05:05:30,065 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.069e+02 1.290e+02 1.386e+02 1.605e+02 2.385e+02, threshold=2.772e+02, percent-clipped=0.0 2024-09-23 05:05:42,858 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.min_positive, batch_count=183586.66666666666, ans=0.05 2024-09-23 05:05:57,104 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=183633.33333333334, ans=0.125 2024-09-23 05:06:11,271 INFO [train.py:1198] (2/4) Epoch 11, batch 400, loss[loss=0.2203, ctc_loss=0.1486, cr_loss=0.3584, over 17003.00 frames. ], tot_loss[loss=0.2488, ctc_loss=0.1722, cr_loss=0.3831, over 2918896.92 frames. ], batch size: 44, lr: 1.12e-02, grad_scale: 32.0 2024-09-23 05:06:13,516 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=13.93 vs. limit=15.0 2024-09-23 05:06:17,197 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=5.37 vs. limit=15.0 2024-09-23 05:06:24,364 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=183680.0, ans=0.0 2024-09-23 05:06:37,197 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=183726.66666666666, ans=0.1 2024-09-23 05:06:38,023 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.81 vs. limit=12.0 2024-09-23 05:06:47,194 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.09 vs. limit=15.0 2024-09-23 05:06:54,594 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=183773.33333333334, ans=0.1 2024-09-23 05:07:31,440 INFO [train.py:1198] (2/4) Epoch 11, batch 450, loss[loss=0.2573, ctc_loss=0.1767, cr_loss=0.4028, over 17227.00 frames. ], tot_loss[loss=0.2488, ctc_loss=0.1721, cr_loss=0.3831, over 3017344.01 frames. ], batch size: 55, lr: 1.12e-02, grad_scale: 32.0 2024-09-23 05:07:32,092 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.06 vs. limit=15.0 2024-09-23 05:08:09,515 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.132e+02 1.312e+02 1.426e+02 1.601e+02 2.161e+02, threshold=2.852e+02, percent-clipped=0.0 2024-09-23 05:08:26,166 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=7.10 vs. limit=15.0 2024-09-23 05:08:33,605 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=184100.0, ans=0.125 2024-09-23 05:08:50,625 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=184100.0, ans=0.1 2024-09-23 05:08:53,365 INFO [train.py:1198] (2/4) Epoch 11, batch 500, loss[loss=0.2357, ctc_loss=0.1617, cr_loss=0.37, over 16971.00 frames. ], tot_loss[loss=0.2479, ctc_loss=0.1715, cr_loss=0.3821, over 3099010.87 frames. ], batch size: 42, lr: 1.12e-02, grad_scale: 32.0 2024-09-23 05:09:36,164 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=184240.0, ans=0.2 2024-09-23 05:09:39,529 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 05:09:39,771 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.67 vs. limit=22.5 2024-09-23 05:09:53,989 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 05:10:00,338 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=184286.66666666666, ans=0.125 2024-09-23 05:10:22,037 INFO [train.py:1198] (2/4) Epoch 11, batch 550, loss[loss=0.2056, ctc_loss=0.1395, cr_loss=0.3304, over 17202.00 frames. ], tot_loss[loss=0.2473, ctc_loss=0.1711, cr_loss=0.3812, over 3160018.28 frames. ], batch size: 41, lr: 1.12e-02, grad_scale: 32.0 2024-09-23 05:10:30,163 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=184380.0, ans=0.025 2024-09-23 05:11:00,280 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.095e+02 1.305e+02 1.444e+02 1.636e+02 3.823e+02, threshold=2.888e+02, percent-clipped=1.0 2024-09-23 05:11:02,338 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=184473.33333333334, ans=0.125 2024-09-23 05:11:13,264 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=184520.0, ans=0.125 2024-09-23 05:11:37,416 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=184566.66666666666, ans=0.125 2024-09-23 05:11:41,804 INFO [train.py:1198] (2/4) Epoch 11, batch 600, loss[loss=0.2619, ctc_loss=0.1814, cr_loss=0.4027, over 16903.00 frames. ], tot_loss[loss=0.2487, ctc_loss=0.1721, cr_loss=0.383, over 3204233.24 frames. ], batch size: 58, lr: 1.12e-02, grad_scale: 32.0 2024-09-23 05:11:42,211 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=184613.33333333334, ans=0.0 2024-09-23 05:11:53,369 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=184613.33333333334, ans=0.125 2024-09-23 05:11:53,816 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=72.05 vs. limit=15.0 2024-09-23 05:12:17,172 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=184706.66666666666, ans=0.1 2024-09-23 05:12:24,995 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=184706.66666666666, ans=0.2 2024-09-23 05:12:52,640 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 05:12:52,885 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.80 vs. limit=15.0 2024-09-23 05:13:01,767 INFO [train.py:1198] (2/4) Epoch 11, batch 650, loss[loss=0.2171, ctc_loss=0.1484, cr_loss=0.3435, over 17295.00 frames. ], tot_loss[loss=0.2485, ctc_loss=0.1719, cr_loss=0.3834, over 3239148.08 frames. ], batch size: 49, lr: 1.12e-02, grad_scale: 32.0 2024-09-23 05:13:10,063 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=184846.66666666666, ans=0.125 2024-09-23 05:13:43,072 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.116e+02 1.331e+02 1.445e+02 1.618e+02 2.572e+02, threshold=2.890e+02, percent-clipped=0.0 2024-09-23 05:14:18,804 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=185033.33333333334, ans=0.125 2024-09-23 05:14:27,588 INFO [train.py:1198] (2/4) Epoch 11, batch 700, loss[loss=0.2303, ctc_loss=0.1583, cr_loss=0.3597, over 17077.00 frames. ], tot_loss[loss=0.2479, ctc_loss=0.1715, cr_loss=0.3821, over 3257578.60 frames. ], batch size: 43, lr: 1.11e-02, grad_scale: 32.0 2024-09-23 05:15:16,070 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=185173.33333333334, ans=0.0 2024-09-23 05:15:22,378 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=185220.0, ans=0.1 2024-09-23 05:15:33,583 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=185220.0, ans=0.125 2024-09-23 05:15:46,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=185266.66666666666, ans=0.1 2024-09-23 05:15:52,062 INFO [train.py:1198] (2/4) Epoch 11, batch 750, loss[loss=0.2727, ctc_loss=0.1921, cr_loss=0.4029, over 16721.00 frames. ], tot_loss[loss=0.2493, ctc_loss=0.1727, cr_loss=0.3833, over 3275250.26 frames. ], batch size: 61, lr: 1.11e-02, grad_scale: 32.0 2024-09-23 05:16:08,476 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=185360.0, ans=0.2 2024-09-23 05:16:30,315 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.321e+02 1.443e+02 1.699e+02 2.904e+02, threshold=2.886e+02, percent-clipped=1.0 2024-09-23 05:16:36,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=185406.66666666666, ans=0.0 2024-09-23 05:16:45,036 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=185453.33333333334, ans=0.0 2024-09-23 05:16:48,160 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=185453.33333333334, ans=0.025 2024-09-23 05:17:11,521 INFO [train.py:1198] (2/4) Epoch 11, batch 800, loss[loss=0.201, ctc_loss=0.1336, cr_loss=0.337, over 17043.00 frames. ], tot_loss[loss=0.2479, ctc_loss=0.1715, cr_loss=0.3819, over 3298328.03 frames. ], batch size: 39, lr: 1.11e-02, grad_scale: 32.0 2024-09-23 05:17:19,986 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.34 vs. limit=15.0 2024-09-23 05:17:32,350 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=185593.33333333334, ans=0.125 2024-09-23 05:17:32,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=185593.33333333334, ans=0.2 2024-09-23 05:17:57,939 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=185686.66666666666, ans=0.025 2024-09-23 05:18:17,031 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=185733.33333333334, ans=0.05 2024-09-23 05:18:17,492 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=11.68 vs. limit=15.0 2024-09-23 05:18:31,203 INFO [train.py:1198] (2/4) Epoch 11, batch 850, loss[loss=0.2928, ctc_loss=0.2096, cr_loss=0.416, over 17018.00 frames. ], tot_loss[loss=0.2489, ctc_loss=0.1723, cr_loss=0.383, over 3303276.02 frames. ], batch size: 56, lr: 1.11e-02, grad_scale: 32.0 2024-09-23 05:18:36,343 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=185780.0, ans=0.1 2024-09-23 05:18:41,215 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.47 vs. limit=6.0 2024-09-23 05:19:12,272 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.355e+02 1.504e+02 1.794e+02 2.469e+02, threshold=3.008e+02, percent-clipped=0.0 2024-09-23 05:19:42,095 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=185966.66666666666, ans=0.125 2024-09-23 05:19:50,103 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=185966.66666666666, ans=0.125 2024-09-23 05:19:56,478 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=185966.66666666666, ans=0.125 2024-09-23 05:19:56,499 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=185966.66666666666, ans=0.2 2024-09-23 05:19:59,415 INFO [train.py:1198] (2/4) Epoch 11, batch 900, loss[loss=0.2321, ctc_loss=0.161, cr_loss=0.3559, over 17206.00 frames. ], tot_loss[loss=0.2486, ctc_loss=0.172, cr_loss=0.383, over 3313987.64 frames. ], batch size: 50, lr: 1.11e-02, grad_scale: 32.0 2024-09-23 05:20:26,504 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=186060.0, ans=0.0 2024-09-23 05:21:22,043 INFO [train.py:1198] (2/4) Epoch 11, batch 950, loss[loss=0.2133, ctc_loss=0.1469, cr_loss=0.332, over 17013.00 frames. ], tot_loss[loss=0.2479, ctc_loss=0.1715, cr_loss=0.3819, over 3317862.75 frames. ], batch size: 39, lr: 1.11e-02, grad_scale: 32.0 2024-09-23 05:21:32,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=186246.66666666666, ans=0.125 2024-09-23 05:21:33,774 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=186246.66666666666, ans=0.125 2024-09-23 05:21:33,885 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=186246.66666666666, ans=0.0 2024-09-23 05:21:35,715 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.89 vs. limit=15.0 2024-09-23 05:21:40,215 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=186293.33333333334, ans=0.0 2024-09-23 05:21:59,531 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=186340.0, ans=0.1 2024-09-23 05:22:00,860 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.078e+02 1.261e+02 1.404e+02 1.572e+02 2.138e+02, threshold=2.809e+02, percent-clipped=0.0 2024-09-23 05:22:18,696 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=186386.66666666666, ans=0.0 2024-09-23 05:22:31,440 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 05:22:42,314 INFO [train.py:1198] (2/4) Epoch 11, batch 1000, loss[loss=0.2447, ctc_loss=0.1665, cr_loss=0.3909, over 17303.00 frames. ], tot_loss[loss=0.2475, ctc_loss=0.1712, cr_loss=0.3816, over 3327490.68 frames. ], batch size: 51, lr: 1.11e-02, grad_scale: 32.0 2024-09-23 05:22:50,684 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=186480.0, ans=0.025 2024-09-23 05:23:13,020 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=186573.33333333334, ans=0.0 2024-09-23 05:23:16,278 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=186573.33333333334, ans=0.0 2024-09-23 05:23:20,896 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=186573.33333333334, ans=0.125 2024-09-23 05:23:30,539 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=186620.0, ans=0.1 2024-09-23 05:23:32,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=186620.0, ans=0.125 2024-09-23 05:23:34,336 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.02 vs. limit=15.0 2024-09-23 05:23:47,218 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=186666.66666666666, ans=0.1 2024-09-23 05:23:59,186 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer_ff2.min_abs, batch_count=186666.66666666666, ans=0.1 2024-09-23 05:24:06,719 INFO [train.py:1198] (2/4) Epoch 11, batch 1050, loss[loss=0.2436, ctc_loss=0.1649, cr_loss=0.3934, over 17120.00 frames. ], tot_loss[loss=0.2481, ctc_loss=0.1716, cr_loss=0.3822, over 3327325.03 frames. ], batch size: 49, lr: 1.11e-02, grad_scale: 32.0 2024-09-23 05:24:13,373 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=186713.33333333334, ans=0.125 2024-09-23 05:24:15,079 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=186713.33333333334, ans=0.0 2024-09-23 05:24:46,146 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=186806.66666666666, ans=0.125 2024-09-23 05:24:50,586 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.157e+02 1.363e+02 1.550e+02 2.011e+02 3.304e+02, threshold=3.099e+02, percent-clipped=2.0 2024-09-23 05:25:22,050 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=6.97 vs. limit=15.0 2024-09-23 05:25:26,516 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 05:25:26,590 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=186900.0, ans=0.125 2024-09-23 05:25:34,180 INFO [train.py:1198] (2/4) Epoch 11, batch 1100, loss[loss=0.2156, ctc_loss=0.1437, cr_loss=0.3598, over 17264.00 frames. ], tot_loss[loss=0.2476, ctc_loss=0.1714, cr_loss=0.3813, over 3326383.88 frames. ], batch size: 42, lr: 1.11e-02, grad_scale: 32.0 2024-09-23 05:25:45,954 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=9.91 vs. limit=15.0 2024-09-23 05:25:48,542 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=186993.33333333334, ans=0.0 2024-09-23 05:26:01,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=186993.33333333334, ans=0.125 2024-09-23 05:26:04,572 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=187040.0, ans=0.0 2024-09-23 05:26:07,790 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=187040.0, ans=0.125 2024-09-23 05:26:12,418 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=187040.0, ans=0.0 2024-09-23 05:26:17,330 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=187040.0, ans=0.1 2024-09-23 05:26:39,903 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=187133.33333333334, ans=0.125 2024-09-23 05:26:48,198 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.49 vs. limit=15.0 2024-09-23 05:26:53,984 INFO [train.py:1198] (2/4) Epoch 11, batch 1150, loss[loss=0.286, ctc_loss=0.1989, cr_loss=0.4354, over 17046.00 frames. ], tot_loss[loss=0.2467, ctc_loss=0.1706, cr_loss=0.3804, over 3328658.35 frames. ], batch size: 52, lr: 1.11e-02, grad_scale: 32.0 2024-09-23 05:27:08,925 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=4.49 vs. limit=12.0 2024-09-23 05:27:18,012 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.max_abs, batch_count=187226.66666666666, ans=10.0 2024-09-23 05:27:32,067 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.352e+02 1.474e+02 1.715e+02 2.168e+02, threshold=2.948e+02, percent-clipped=0.0 2024-09-23 05:27:37,025 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=187273.33333333334, ans=0.125 2024-09-23 05:27:56,316 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=187366.66666666666, ans=0.0 2024-09-23 05:28:02,612 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=187366.66666666666, ans=0.04949747468305833 2024-09-23 05:28:13,186 INFO [train.py:1198] (2/4) Epoch 11, batch 1200, loss[loss=0.2809, ctc_loss=0.196, cr_loss=0.4247, over 17006.00 frames. ], tot_loss[loss=0.2465, ctc_loss=0.1704, cr_loss=0.3804, over 3338337.93 frames. ], batch size: 56, lr: 1.11e-02, grad_scale: 32.0 2024-09-23 05:28:35,906 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=187460.0, ans=0.0 2024-09-23 05:28:36,388 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=5.53 vs. limit=10.0 2024-09-23 05:29:10,307 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=187553.33333333334, ans=0.1 2024-09-23 05:29:18,226 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=187600.0, ans=0.0 2024-09-23 05:29:41,426 INFO [train.py:1198] (2/4) Epoch 11, batch 1250, loss[loss=0.2395, ctc_loss=0.1637, cr_loss=0.3791, over 17235.00 frames. ], tot_loss[loss=0.2462, ctc_loss=0.1702, cr_loss=0.3798, over 3341698.72 frames. ], batch size: 47, lr: 1.11e-02, grad_scale: 32.0 2024-09-23 05:29:42,705 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.75 vs. limit=15.0 2024-09-23 05:30:15,015 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=187740.0, ans=0.125 2024-09-23 05:30:22,691 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.147e+02 1.347e+02 1.488e+02 1.647e+02 3.078e+02, threshold=2.976e+02, percent-clipped=1.0 2024-09-23 05:30:24,891 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.92 vs. limit=15.0 2024-09-23 05:30:42,054 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=187786.66666666666, ans=0.125 2024-09-23 05:30:45,045 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=187786.66666666666, ans=0.2 2024-09-23 05:30:46,795 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=187833.33333333334, ans=0.125 2024-09-23 05:30:54,582 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=187833.33333333334, ans=0.125 2024-09-23 05:31:04,205 INFO [train.py:1198] (2/4) Epoch 11, batch 1300, loss[loss=0.2489, ctc_loss=0.1679, cr_loss=0.4053, over 17191.00 frames. ], tot_loss[loss=0.2457, ctc_loss=0.1699, cr_loss=0.3792, over 3350377.65 frames. ], batch size: 47, lr: 1.11e-02, grad_scale: 32.0 2024-09-23 05:31:07,671 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=187880.0, ans=0.125 2024-09-23 05:31:22,073 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=187926.66666666666, ans=0.2 2024-09-23 05:31:41,032 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=187973.33333333334, ans=0.1 2024-09-23 05:31:57,280 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=8.58 vs. limit=15.0 2024-09-23 05:32:22,352 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.min_abs, batch_count=188113.33333333334, ans=0.5 2024-09-23 05:32:23,725 INFO [train.py:1198] (2/4) Epoch 11, batch 1350, loss[loss=0.2407, ctc_loss=0.1695, cr_loss=0.3559, over 17184.00 frames. ], tot_loss[loss=0.2448, ctc_loss=0.1691, cr_loss=0.3786, over 3359051.81 frames. ], batch size: 45, lr: 1.11e-02, grad_scale: 32.0 2024-09-23 05:32:35,316 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=188113.33333333334, ans=0.1 2024-09-23 05:32:35,743 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.40 vs. limit=15.0 2024-09-23 05:33:02,037 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.065e+02 1.248e+02 1.402e+02 1.590e+02 2.779e+02, threshold=2.805e+02, percent-clipped=0.0 2024-09-23 05:33:45,863 INFO [train.py:1198] (2/4) Epoch 11, batch 1400, loss[loss=0.2577, ctc_loss=0.1773, cr_loss=0.4018, over 17215.00 frames. ], tot_loss[loss=0.2449, ctc_loss=0.1691, cr_loss=0.3786, over 3361292.50 frames. ], batch size: 50, lr: 1.11e-02, grad_scale: 32.0 2024-09-23 05:33:52,333 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=188346.66666666666, ans=0.125 2024-09-23 05:34:01,953 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=188393.33333333334, ans=0.125 2024-09-23 05:34:21,649 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=512, metric=4.59 vs. limit=15.0 2024-09-23 05:34:46,079 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.16 vs. limit=15.0 2024-09-23 05:34:53,483 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=188533.33333333334, ans=0.0 2024-09-23 05:34:54,317 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.07 vs. limit=10.0 2024-09-23 05:35:02,626 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=188533.33333333334, ans=0.0 2024-09-23 05:35:13,556 INFO [train.py:1198] (2/4) Epoch 11, batch 1450, loss[loss=0.2982, ctc_loss=0.2107, cr_loss=0.4376, over 15040.00 frames. ], tot_loss[loss=0.2455, ctc_loss=0.1696, cr_loss=0.3793, over 3366096.74 frames. ], batch size: 89, lr: 1.10e-02, grad_scale: 32.0 2024-09-23 05:35:13,942 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=188580.0, ans=0.04949747468305833 2024-09-23 05:35:23,242 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=188580.0, ans=0.125 2024-09-23 05:35:51,362 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.355e+02 1.506e+02 1.702e+02 2.506e+02, threshold=3.011e+02, percent-clipped=0.0 2024-09-23 05:36:20,404 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=188766.66666666666, ans=0.125 2024-09-23 05:36:26,883 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=188766.66666666666, ans=0.125 2024-09-23 05:36:32,871 INFO [train.py:1198] (2/4) Epoch 11, batch 1500, loss[loss=0.3184, ctc_loss=0.2414, cr_loss=0.3848, over 11439.00 frames. ], tot_loss[loss=0.2463, ctc_loss=0.1702, cr_loss=0.3805, over 3366457.19 frames. ], batch size: 123, lr: 1.10e-02, grad_scale: 32.0 2024-09-23 05:37:32,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=188953.33333333334, ans=0.125 2024-09-23 05:37:40,673 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.86 vs. limit=15.0 2024-09-23 05:37:52,941 INFO [train.py:1198] (2/4) Epoch 11, batch 1550, loss[loss=0.3015, ctc_loss=0.2187, cr_loss=0.4141, over 11309.00 frames. ], tot_loss[loss=0.2457, ctc_loss=0.1698, cr_loss=0.3795, over 3365325.37 frames. ], batch size: 123, lr: 1.10e-02, grad_scale: 32.0 2024-09-23 05:37:57,412 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=8.25 vs. limit=15.0 2024-09-23 05:38:01,356 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=189046.66666666666, ans=0.0 2024-09-23 05:38:15,681 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=189093.33333333334, ans=0.1 2024-09-23 05:38:33,046 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.070e+02 1.268e+02 1.378e+02 1.550e+02 2.066e+02, threshold=2.755e+02, percent-clipped=0.0 2024-09-23 05:38:59,453 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.min_positive, batch_count=189233.33333333334, ans=0.05 2024-09-23 05:39:15,063 INFO [train.py:1198] (2/4) Epoch 11, batch 1600, loss[loss=0.2493, ctc_loss=0.1704, cr_loss=0.3945, over 17045.00 frames. ], tot_loss[loss=0.2475, ctc_loss=0.1711, cr_loss=0.382, over 3351159.25 frames. ], batch size: 52, lr: 1.10e-02, grad_scale: 32.0 2024-09-23 05:40:10,777 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=189420.0, ans=0.2 2024-09-23 05:40:13,101 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.07 vs. limit=15.0 2024-09-23 05:40:42,829 INFO [train.py:1198] (2/4) Epoch 11, batch 1650, loss[loss=0.2985, ctc_loss=0.2037, cr_loss=0.4741, over 16440.00 frames. ], tot_loss[loss=0.2485, ctc_loss=0.172, cr_loss=0.3829, over 3347565.75 frames. ], batch size: 66, lr: 1.10e-02, grad_scale: 32.0 2024-09-23 05:41:05,590 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=189560.0, ans=0.0 2024-09-23 05:41:22,848 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.160e+02 1.323e+02 1.468e+02 1.701e+02 2.632e+02, threshold=2.937e+02, percent-clipped=0.0 2024-09-23 05:41:31,020 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=189653.33333333334, ans=0.125 2024-09-23 05:41:54,750 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=189700.0, ans=0.05 2024-09-23 05:42:02,590 INFO [train.py:1198] (2/4) Epoch 11, batch 1700, loss[loss=0.2294, ctc_loss=0.1567, cr_loss=0.3632, over 17216.00 frames. ], tot_loss[loss=0.2474, ctc_loss=0.1711, cr_loss=0.3818, over 3355759.85 frames. ], batch size: 47, lr: 1.10e-02, grad_scale: 32.0 2024-09-23 05:42:31,789 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=189793.33333333334, ans=0.1 2024-09-23 05:42:34,851 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=189840.0, ans=0.025 2024-09-23 05:42:36,621 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=189840.0, ans=0.125 2024-09-23 05:42:45,981 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=189840.0, ans=0.125 2024-09-23 05:42:55,459 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=189886.66666666666, ans=0.125 2024-09-23 05:42:57,149 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=189886.66666666666, ans=0.125 2024-09-23 05:43:20,991 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=189980.0, ans=0.125 2024-09-23 05:43:21,170 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=189980.0, ans=0.125 2024-09-23 05:43:22,420 INFO [train.py:1198] (2/4) Epoch 11, batch 1750, loss[loss=0.2457, ctc_loss=0.1662, cr_loss=0.3974, over 17192.00 frames. ], tot_loss[loss=0.2483, ctc_loss=0.1718, cr_loss=0.3827, over 3349253.72 frames. ], batch size: 41, lr: 1.10e-02, grad_scale: 32.0 2024-09-23 05:43:39,944 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=190026.66666666666, ans=0.125 2024-09-23 05:44:00,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=190073.33333333334, ans=0.025 2024-09-23 05:44:00,323 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=190073.33333333334, ans=0.0 2024-09-23 05:44:01,966 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=190073.33333333334, ans=0.0 2024-09-23 05:44:04,904 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.149e+02 1.343e+02 1.475e+02 1.671e+02 2.216e+02, threshold=2.950e+02, percent-clipped=0.0 2024-09-23 05:44:23,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=190120.0, ans=0.125 2024-09-23 05:44:38,639 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=190166.66666666666, ans=0.125 2024-09-23 05:44:41,973 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=190166.66666666666, ans=0.125 2024-09-23 05:44:44,184 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.41 vs. limit=12.0 2024-09-23 05:44:45,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=190166.66666666666, ans=0.2 2024-09-23 05:44:52,206 INFO [train.py:1198] (2/4) Epoch 11, batch 1800, loss[loss=0.2485, ctc_loss=0.1753, cr_loss=0.3658, over 17044.00 frames. ], tot_loss[loss=0.2487, ctc_loss=0.172, cr_loss=0.3831, over 3349602.45 frames. ], batch size: 52, lr: 1.10e-02, grad_scale: 32.0 2024-09-23 05:44:56,309 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.28 vs. limit=15.0 2024-09-23 05:45:08,593 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=190260.0, ans=0.0 2024-09-23 05:45:13,239 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=190260.0, ans=10.0 2024-09-23 05:45:22,800 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=190306.66666666666, ans=0.125 2024-09-23 05:45:27,509 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=190306.66666666666, ans=0.05 2024-09-23 05:46:12,735 INFO [train.py:1198] (2/4) Epoch 11, batch 1850, loss[loss=0.2592, ctc_loss=0.1784, cr_loss=0.404, over 17105.00 frames. ], tot_loss[loss=0.2479, ctc_loss=0.1714, cr_loss=0.3823, over 3354926.03 frames. ], batch size: 49, lr: 1.10e-02, grad_scale: 32.0 2024-09-23 05:46:22,388 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=190446.66666666666, ans=0.2 2024-09-23 05:46:36,014 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.07 vs. limit=6.0 2024-09-23 05:46:44,859 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=190540.0, ans=0.125 2024-09-23 05:46:46,899 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=6.49 vs. limit=15.0 2024-09-23 05:46:52,382 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.102e+02 1.336e+02 1.526e+02 1.737e+02 2.806e+02, threshold=3.051e+02, percent-clipped=0.0 2024-09-23 05:46:52,800 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=190540.0, ans=0.2 2024-09-23 05:46:56,711 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.13 vs. limit=10.0 2024-09-23 05:46:58,025 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.37 vs. limit=15.0 2024-09-23 05:47:16,759 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=190633.33333333334, ans=0.2 2024-09-23 05:47:32,658 INFO [train.py:1198] (2/4) Epoch 11, batch 1900, loss[loss=0.2568, ctc_loss=0.1728, cr_loss=0.4201, over 17297.00 frames. ], tot_loss[loss=0.2479, ctc_loss=0.1715, cr_loss=0.3819, over 3339109.07 frames. ], batch size: 49, lr: 1.10e-02, grad_scale: 16.0 2024-09-23 05:47:39,986 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn2.whiten, num_groups=1, num_channels=192, metric=9.31 vs. limit=22.5 2024-09-23 05:47:41,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=190680.0, ans=0.125 2024-09-23 05:47:45,991 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=11.68 vs. limit=15.0 2024-09-23 05:47:47,170 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=190726.66666666666, ans=0.2 2024-09-23 05:47:52,866 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.18 vs. limit=15.0 2024-09-23 05:48:02,334 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.90 vs. limit=6.0 2024-09-23 05:48:26,927 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=190820.0, ans=0.1 2024-09-23 05:48:39,162 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=190866.66666666666, ans=0.125 2024-09-23 05:48:39,265 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=190866.66666666666, ans=0.2 2024-09-23 05:48:47,249 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=190866.66666666666, ans=0.125 2024-09-23 05:48:55,010 INFO [train.py:1198] (2/4) Epoch 11, batch 1950, loss[loss=0.2477, ctc_loss=0.1732, cr_loss=0.3727, over 17367.00 frames. ], tot_loss[loss=0.2481, ctc_loss=0.1717, cr_loss=0.3821, over 3340006.39 frames. ], batch size: 48, lr: 1.10e-02, grad_scale: 16.0 2024-09-23 05:49:01,615 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=190913.33333333334, ans=0.1 2024-09-23 05:49:26,714 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=11.55 vs. limit=22.5 2024-09-23 05:49:27,910 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=190960.0, ans=0.125 2024-09-23 05:49:41,895 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.099e+02 1.367e+02 1.554e+02 1.740e+02 3.825e+02, threshold=3.108e+02, percent-clipped=1.0 2024-09-23 05:49:54,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=191053.33333333334, ans=0.125 2024-09-23 05:50:07,404 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=3.96 vs. limit=15.0 2024-09-23 05:50:22,738 INFO [train.py:1198] (2/4) Epoch 11, batch 2000, loss[loss=0.2395, ctc_loss=0.1623, cr_loss=0.3859, over 17252.00 frames. ], tot_loss[loss=0.2463, ctc_loss=0.1703, cr_loss=0.3803, over 3355215.57 frames. ], batch size: 44, lr: 1.10e-02, grad_scale: 32.0 2024-09-23 05:50:39,956 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.00 vs. limit=6.0 2024-09-23 05:51:34,054 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=191333.33333333334, ans=0.1 2024-09-23 05:51:41,208 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=12.27 vs. limit=15.0 2024-09-23 05:51:43,322 INFO [train.py:1198] (2/4) Epoch 11, batch 2050, loss[loss=0.2736, ctc_loss=0.1884, cr_loss=0.426, over 17105.00 frames. ], tot_loss[loss=0.2473, ctc_loss=0.171, cr_loss=0.3814, over 3355204.50 frames. ], batch size: 49, lr: 1.10e-02, grad_scale: 32.0 2024-09-23 05:51:43,608 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=191380.0, ans=0.05 2024-09-23 05:51:46,893 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=191380.0, ans=0.5 2024-09-23 05:52:07,837 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=9.61 vs. limit=15.0 2024-09-23 05:52:13,683 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=191473.33333333334, ans=0.125 2024-09-23 05:52:18,543 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=191473.33333333334, ans=0.2 2024-09-23 05:52:18,671 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=191473.33333333334, ans=0.125 2024-09-23 05:52:24,542 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.258e+02 1.361e+02 1.544e+02 2.924e+02, threshold=2.721e+02, percent-clipped=0.0 2024-09-23 05:52:47,187 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=191566.66666666666, ans=0.1 2024-09-23 05:53:03,035 INFO [train.py:1198] (2/4) Epoch 11, batch 2100, loss[loss=0.2071, ctc_loss=0.14, cr_loss=0.3351, over 17090.00 frames. ], tot_loss[loss=0.2446, ctc_loss=0.1689, cr_loss=0.3786, over 3358972.47 frames. ], batch size: 40, lr: 1.10e-02, grad_scale: 32.0 2024-09-23 05:53:05,067 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=191613.33333333334, ans=0.0 2024-09-23 05:54:23,684 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=191800.0, ans=0.2 2024-09-23 05:54:30,861 INFO [train.py:1198] (2/4) Epoch 11, batch 2150, loss[loss=0.2177, ctc_loss=0.1489, cr_loss=0.3439, over 17079.00 frames. ], tot_loss[loss=0.2446, ctc_loss=0.1689, cr_loss=0.3785, over 3360098.06 frames. ], batch size: 43, lr: 1.10e-02, grad_scale: 32.0 2024-09-23 05:55:04,424 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.71 vs. limit=15.0 2024-09-23 05:55:14,825 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.163e+02 1.341e+02 1.514e+02 1.705e+02 2.483e+02, threshold=3.028e+02, percent-clipped=0.0 2024-09-23 05:55:21,617 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=191986.66666666666, ans=0.04949747468305833 2024-09-23 05:55:53,281 INFO [train.py:1198] (2/4) Epoch 11, batch 2200, loss[loss=0.2536, ctc_loss=0.1749, cr_loss=0.3934, over 16905.00 frames. ], tot_loss[loss=0.2455, ctc_loss=0.1697, cr_loss=0.379, over 3350865.80 frames. ], batch size: 58, lr: 1.10e-02, grad_scale: 32.0 2024-09-23 05:55:56,848 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=192080.0, ans=0.125 2024-09-23 05:56:40,157 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=192220.0, ans=0.125 2024-09-23 05:56:43,165 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=192220.0, ans=0.1 2024-09-23 05:56:43,499 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.16 vs. limit=22.5 2024-09-23 05:56:44,919 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=192220.0, ans=0.125 2024-09-23 05:56:53,430 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=7.61 vs. limit=15.0 2024-09-23 05:57:08,310 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.57 vs. limit=15.0 2024-09-23 05:57:09,209 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 05:57:13,773 INFO [train.py:1198] (2/4) Epoch 11, batch 2250, loss[loss=0.2995, ctc_loss=0.2151, cr_loss=0.4221, over 14778.00 frames. ], tot_loss[loss=0.2463, ctc_loss=0.1703, cr_loss=0.3803, over 3349957.62 frames. ], batch size: 89, lr: 1.09e-02, grad_scale: 32.0 2024-09-23 05:57:16,471 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.whiten, num_groups=1, num_channels=192, metric=4.03 vs. limit=12.0 2024-09-23 05:57:31,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=192360.0, ans=0.125 2024-09-23 05:57:55,417 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.304e+02 1.378e+02 1.493e+02 3.334e+02, threshold=2.757e+02, percent-clipped=1.0 2024-09-23 05:58:05,573 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.88 vs. limit=15.0 2024-09-23 05:58:23,776 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.75 vs. limit=15.0 2024-09-23 05:58:36,212 INFO [train.py:1198] (2/4) Epoch 11, batch 2300, loss[loss=0.2236, ctc_loss=0.1574, cr_loss=0.3312, over 17055.00 frames. ], tot_loss[loss=0.2468, ctc_loss=0.1707, cr_loss=0.3805, over 3352433.60 frames. ], batch size: 39, lr: 1.09e-02, grad_scale: 32.0 2024-09-23 05:58:42,982 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 05:58:47,718 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=192546.66666666666, ans=0.1 2024-09-23 05:58:49,290 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=192546.66666666666, ans=0.125 2024-09-23 06:00:02,384 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=192780.0, ans=0.0 2024-09-23 06:00:03,694 INFO [train.py:1198] (2/4) Epoch 11, batch 2350, loss[loss=0.2561, ctc_loss=0.1775, cr_loss=0.393, over 17068.00 frames. ], tot_loss[loss=0.248, ctc_loss=0.1718, cr_loss=0.381, over 3342865.55 frames. ], batch size: 46, lr: 1.09e-02, grad_scale: 32.0 2024-09-23 06:00:03,927 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=192780.0, ans=0.1 2024-09-23 06:00:08,882 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=192780.0, ans=0.2 2024-09-23 06:00:15,446 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=9.66 vs. limit=12.0 2024-09-23 06:00:19,631 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=192826.66666666666, ans=0.04949747468305833 2024-09-23 06:00:27,491 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=192826.66666666666, ans=0.125 2024-09-23 06:00:34,394 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.77 vs. limit=15.0 2024-09-23 06:00:43,242 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=192873.33333333334, ans=0.2 2024-09-23 06:00:43,283 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=192873.33333333334, ans=0.125 2024-09-23 06:00:44,552 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.266e+02 1.351e+02 1.522e+02 2.296e+02, threshold=2.702e+02, percent-clipped=0.0 2024-09-23 06:00:46,405 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=192873.33333333334, ans=0.025 2024-09-23 06:00:59,175 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer_ff3.min_abs, batch_count=192920.0, ans=0.2 2024-09-23 06:01:05,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=192966.66666666666, ans=0.125 2024-09-23 06:01:23,281 INFO [train.py:1198] (2/4) Epoch 11, batch 2400, loss[loss=0.2699, ctc_loss=0.1874, cr_loss=0.4123, over 17006.00 frames. ], tot_loss[loss=0.2464, ctc_loss=0.1705, cr_loss=0.3796, over 3350004.32 frames. ], batch size: 56, lr: 1.09e-02, grad_scale: 32.0 2024-09-23 06:01:23,605 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=193013.33333333334, ans=0.1 2024-09-23 06:01:28,405 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=193013.33333333334, ans=0.0 2024-09-23 06:01:30,155 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=193013.33333333334, ans=0.2 2024-09-23 06:01:30,219 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=193013.33333333334, ans=0.1 2024-09-23 06:01:42,936 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=193060.0, ans=0.125 2024-09-23 06:01:42,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=193060.0, ans=0.1 2024-09-23 06:01:43,374 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.29 vs. limit=15.0 2024-09-23 06:02:43,183 INFO [train.py:1198] (2/4) Epoch 11, batch 2450, loss[loss=0.2416, ctc_loss=0.1669, cr_loss=0.3735, over 17350.00 frames. ], tot_loss[loss=0.2462, ctc_loss=0.1701, cr_loss=0.3801, over 3347109.22 frames. ], batch size: 48, lr: 1.09e-02, grad_scale: 32.0 2024-09-23 06:02:56,902 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.whiten.whitening_limit, batch_count=193246.66666666666, ans=12.0 2024-09-23 06:03:10,750 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=193293.33333333334, ans=0.2 2024-09-23 06:03:16,408 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=7.12 vs. limit=15.0 2024-09-23 06:03:24,779 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.115e+02 1.301e+02 1.390e+02 1.544e+02 1.973e+02, threshold=2.781e+02, percent-clipped=0.0 2024-09-23 06:03:37,161 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 06:03:45,518 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.whiten.whitening_limit, batch_count=193386.66666666666, ans=15.0 2024-09-23 06:03:47,053 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=512, metric=15.55 vs. limit=22.5 2024-09-23 06:03:55,921 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=193433.33333333334, ans=0.0 2024-09-23 06:04:02,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=193433.33333333334, ans=0.125 2024-09-23 06:04:02,198 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=193433.33333333334, ans=0.0 2024-09-23 06:04:07,697 INFO [train.py:1198] (2/4) Epoch 11, batch 2500, loss[loss=0.2235, ctc_loss=0.1507, cr_loss=0.3641, over 17160.00 frames. ], tot_loss[loss=0.2444, ctc_loss=0.1687, cr_loss=0.3782, over 3361014.86 frames. ], batch size: 45, lr: 1.09e-02, grad_scale: 32.0 2024-09-23 06:04:25,034 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=193526.66666666666, ans=0.0 2024-09-23 06:04:26,752 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=193526.66666666666, ans=0.2 2024-09-23 06:04:40,808 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=193526.66666666666, ans=0.2 2024-09-23 06:04:41,230 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=17.42 vs. limit=22.5 2024-09-23 06:04:47,937 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=6.22 vs. limit=8.0 2024-09-23 06:04:51,788 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=193573.33333333334, ans=0.1 2024-09-23 06:04:54,920 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=193573.33333333334, ans=0.2 2024-09-23 06:05:03,028 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=193620.0, ans=0.125 2024-09-23 06:05:17,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=193666.66666666666, ans=0.125 2024-09-23 06:05:28,127 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=193666.66666666666, ans=0.125 2024-09-23 06:05:32,500 INFO [train.py:1198] (2/4) Epoch 11, batch 2550, loss[loss=0.2368, ctc_loss=0.1672, cr_loss=0.3482, over 17042.00 frames. ], tot_loss[loss=0.2446, ctc_loss=0.169, cr_loss=0.3783, over 3359355.77 frames. ], batch size: 56, lr: 1.09e-02, grad_scale: 32.0 2024-09-23 06:06:13,861 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.308e+02 1.472e+02 1.682e+02 2.409e+02, threshold=2.944e+02, percent-clipped=0.0 2024-09-23 06:06:18,978 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=193853.33333333334, ans=0.1 2024-09-23 06:06:34,673 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=193900.0, ans=0.0 2024-09-23 06:06:48,962 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=193900.0, ans=0.0 2024-09-23 06:06:51,767 INFO [train.py:1198] (2/4) Epoch 11, batch 2600, loss[loss=0.2159, ctc_loss=0.1461, cr_loss=0.349, over 17015.00 frames. ], tot_loss[loss=0.2441, ctc_loss=0.1686, cr_loss=0.3772, over 3356916.01 frames. ], batch size: 44, lr: 1.09e-02, grad_scale: 32.0 2024-09-23 06:06:55,046 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=193946.66666666666, ans=0.125 2024-09-23 06:06:58,399 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=193946.66666666666, ans=0.1 2024-09-23 06:07:03,186 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=193946.66666666666, ans=0.125 2024-09-23 06:07:38,335 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 06:07:41,880 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.32 vs. limit=15.0 2024-09-23 06:07:48,780 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=13.74 vs. limit=15.0 2024-09-23 06:07:50,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=194086.66666666666, ans=0.1 2024-09-23 06:07:52,593 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=194086.66666666666, ans=0.2 2024-09-23 06:08:11,670 INFO [train.py:1198] (2/4) Epoch 11, batch 2650, loss[loss=0.222, ctc_loss=0.152, cr_loss=0.3496, over 16691.00 frames. ], tot_loss[loss=0.2453, ctc_loss=0.1696, cr_loss=0.3785, over 3343122.03 frames. ], batch size: 37, lr: 1.09e-02, grad_scale: 32.0 2024-09-23 06:08:22,886 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=194180.0, ans=0.0 2024-09-23 06:08:26,109 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=194226.66666666666, ans=0.07 2024-09-23 06:08:29,105 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=194226.66666666666, ans=0.04949747468305833 2024-09-23 06:08:29,132 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=194226.66666666666, ans=0.125 2024-09-23 06:08:52,662 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=194273.33333333334, ans=0.125 2024-09-23 06:08:54,184 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=194273.33333333334, ans=0.0 2024-09-23 06:08:57,081 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.068e+02 1.341e+02 1.529e+02 1.824e+02 3.038e+02, threshold=3.058e+02, percent-clipped=1.0 2024-09-23 06:09:06,360 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=194320.0, ans=0.125 2024-09-23 06:09:11,771 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.61 vs. limit=15.0 2024-09-23 06:09:20,926 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.94 vs. limit=15.0 2024-09-23 06:09:33,952 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=194366.66666666666, ans=0.1 2024-09-23 06:09:40,650 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.63 vs. limit=15.0 2024-09-23 06:09:41,680 INFO [train.py:1198] (2/4) Epoch 11, batch 2700, loss[loss=0.2672, ctc_loss=0.185, cr_loss=0.4109, over 17051.00 frames. ], tot_loss[loss=0.2446, ctc_loss=0.1689, cr_loss=0.3785, over 3353674.82 frames. ], batch size: 52, lr: 1.09e-02, grad_scale: 16.0 2024-09-23 06:09:45,341 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=194413.33333333334, ans=0.125 2024-09-23 06:09:54,745 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=194413.33333333334, ans=0.2 2024-09-23 06:10:04,933 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.52 vs. limit=6.0 2024-09-23 06:10:05,887 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=194460.0, ans=0.025 2024-09-23 06:10:23,432 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=194506.66666666666, ans=0.125 2024-09-23 06:10:23,438 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=194506.66666666666, ans=0.125 2024-09-23 06:10:27,319 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.60 vs. limit=6.0 2024-09-23 06:10:36,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=194553.33333333334, ans=0.1 2024-09-23 06:11:00,036 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=194646.66666666666, ans=0.0 2024-09-23 06:11:01,389 INFO [train.py:1198] (2/4) Epoch 11, batch 2750, loss[loss=0.2378, ctc_loss=0.1627, cr_loss=0.3757, over 17297.00 frames. ], tot_loss[loss=0.245, ctc_loss=0.1692, cr_loss=0.3788, over 3356705.30 frames. ], batch size: 49, lr: 1.09e-02, grad_scale: 16.0 2024-09-23 06:11:06,418 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=194646.66666666666, ans=0.125 2024-09-23 06:11:22,280 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=194693.33333333334, ans=0.125 2024-09-23 06:11:44,296 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.099e+02 1.275e+02 1.375e+02 1.599e+02 2.337e+02, threshold=2.750e+02, percent-clipped=0.0 2024-09-23 06:12:10,712 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=5.11 vs. limit=12.0 2024-09-23 06:12:15,395 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.26 vs. limit=22.5 2024-09-23 06:12:20,385 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.40 vs. limit=12.0 2024-09-23 06:12:21,170 INFO [train.py:1198] (2/4) Epoch 11, batch 2800, loss[loss=0.2636, ctc_loss=0.1867, cr_loss=0.3842, over 16010.00 frames. ], tot_loss[loss=0.2438, ctc_loss=0.1683, cr_loss=0.3776, over 3355247.19 frames. ], batch size: 74, lr: 1.09e-02, grad_scale: 32.0 2024-09-23 06:12:41,104 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=194926.66666666666, ans=0.0 2024-09-23 06:13:14,809 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=195020.0, ans=0.05 2024-09-23 06:13:20,064 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.44 vs. limit=6.0 2024-09-23 06:13:21,396 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 06:13:38,814 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=195066.66666666666, ans=0.0 2024-09-23 06:13:44,868 INFO [train.py:1198] (2/4) Epoch 11, batch 2850, loss[loss=0.2087, ctc_loss=0.1434, cr_loss=0.3264, over 17045.00 frames. ], tot_loss[loss=0.2433, ctc_loss=0.1678, cr_loss=0.3773, over 3355410.92 frames. ], batch size: 39, lr: 1.09e-02, grad_scale: 32.0 2024-09-23 06:13:45,833 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.55 vs. limit=6.0 2024-09-23 06:13:49,072 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.77 vs. limit=12.0 2024-09-23 06:13:52,127 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=5.58 vs. limit=12.0 2024-09-23 06:14:35,500 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.164e+02 1.307e+02 1.410e+02 1.605e+02 2.111e+02, threshold=2.819e+02, percent-clipped=0.0 2024-09-23 06:14:43,602 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=195253.33333333334, ans=0.0 2024-09-23 06:14:55,452 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=8.24 vs. limit=15.0 2024-09-23 06:15:11,916 INFO [train.py:1198] (2/4) Epoch 11, batch 2900, loss[loss=0.2996, ctc_loss=0.2202, cr_loss=0.3971, over 11950.00 frames. ], tot_loss[loss=0.2445, ctc_loss=0.1689, cr_loss=0.3782, over 3341760.61 frames. ], batch size: 123, lr: 1.09e-02, grad_scale: 32.0 2024-09-23 06:15:20,871 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=20.98 vs. limit=22.5 2024-09-23 06:15:37,775 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=195393.33333333334, ans=0.0 2024-09-23 06:16:04,951 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=195486.66666666666, ans=0.0 2024-09-23 06:16:06,593 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=195486.66666666666, ans=0.125 2024-09-23 06:16:08,002 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=195486.66666666666, ans=0.1 2024-09-23 06:16:16,069 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=195533.33333333334, ans=0.04949747468305833 2024-09-23 06:16:31,836 INFO [train.py:1198] (2/4) Epoch 11, batch 2950, loss[loss=0.2743, ctc_loss=0.193, cr_loss=0.4066, over 15993.00 frames. ], tot_loss[loss=0.2449, ctc_loss=0.1692, cr_loss=0.3787, over 3343172.42 frames. ], batch size: 74, lr: 1.09e-02, grad_scale: 32.0 2024-09-23 06:16:35,356 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=195580.0, ans=0.2 2024-09-23 06:16:41,769 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=195580.0, ans=0.1 2024-09-23 06:16:49,665 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.max_abs, batch_count=195626.66666666666, ans=10.0 2024-09-23 06:17:09,440 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=5.04 vs. limit=10.0 2024-09-23 06:17:14,835 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.272e+02 1.364e+02 1.479e+02 2.031e+02, threshold=2.728e+02, percent-clipped=0.0 2024-09-23 06:17:24,631 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=195720.0, ans=0.125 2024-09-23 06:17:24,736 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=195720.0, ans=0.0 2024-09-23 06:17:49,683 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=195813.33333333334, ans=0.1 2024-09-23 06:17:50,960 INFO [train.py:1198] (2/4) Epoch 11, batch 3000, loss[loss=0.256, ctc_loss=0.1808, cr_loss=0.3762, over 17138.00 frames. ], tot_loss[loss=0.246, ctc_loss=0.1701, cr_loss=0.3796, over 3345624.71 frames. ], batch size: 48, lr: 1.08e-02, grad_scale: 32.0 2024-09-23 06:17:50,961 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 06:18:01,192 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.2.encoder.layers.0.self_attn_weights, attn_weights_entropy = tensor([4.9986, 4.8592, 4.6990, 4.3884], device='cuda:2') 2024-09-23 06:18:06,132 INFO [train.py:1230] (2/4) Epoch 11, validation: loss=0.04835, ctc_loss=0.04835, cr_loss=7.412e-15, over 944034.00 frames. 2024-09-23 06:18:06,133 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 06:18:14,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=195813.33333333334, ans=0.125 2024-09-23 06:18:15,656 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=195813.33333333334, ans=0.0 2024-09-23 06:18:42,164 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=195906.66666666666, ans=0.125 2024-09-23 06:18:43,615 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=195906.66666666666, ans=0.1 2024-09-23 06:19:02,969 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.26 vs. limit=6.0 2024-09-23 06:19:10,108 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=196000.0, ans=0.0 2024-09-23 06:19:20,479 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.85 vs. limit=6.0 2024-09-23 06:19:27,162 INFO [train.py:1198] (2/4) Epoch 11, batch 3050, loss[loss=0.2194, ctc_loss=0.1531, cr_loss=0.3317, over 17247.00 frames. ], tot_loss[loss=0.2444, ctc_loss=0.1688, cr_loss=0.3778, over 3352547.04 frames. ], batch size: 44, lr: 1.08e-02, grad_scale: 32.0 2024-09-23 06:19:39,243 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=196046.66666666666, ans=0.125 2024-09-23 06:19:53,625 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=196093.33333333334, ans=0.2 2024-09-23 06:20:11,792 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=196140.0, ans=0.0 2024-09-23 06:20:14,516 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.336e+02 1.544e+02 1.822e+02 2.726e+02, threshold=3.088e+02, percent-clipped=0.0 2024-09-23 06:20:26,917 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=196186.66666666666, ans=0.0 2024-09-23 06:20:50,317 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.41 vs. limit=12.0 2024-09-23 06:20:52,700 INFO [train.py:1198] (2/4) Epoch 11, batch 3100, loss[loss=0.2055, ctc_loss=0.1367, cr_loss=0.3441, over 17043.00 frames. ], tot_loss[loss=0.2453, ctc_loss=0.1695, cr_loss=0.3793, over 3350043.44 frames. ], batch size: 39, lr: 1.08e-02, grad_scale: 32.0 2024-09-23 06:21:24,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.whiten.whitening_limit, batch_count=196373.33333333334, ans=12.0 2024-09-23 06:21:47,844 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=196420.0, ans=0.0 2024-09-23 06:22:00,656 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=13.00 vs. limit=15.0 2024-09-23 06:22:03,479 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=196466.66666666666, ans=0.1 2024-09-23 06:22:11,263 INFO [train.py:1198] (2/4) Epoch 11, batch 3150, loss[loss=0.2648, ctc_loss=0.184, cr_loss=0.4042, over 17219.00 frames. ], tot_loss[loss=0.2448, ctc_loss=0.169, cr_loss=0.3789, over 3355429.55 frames. ], batch size: 55, lr: 1.08e-02, grad_scale: 32.0 2024-09-23 06:22:16,219 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=196513.33333333334, ans=0.2 2024-09-23 06:22:19,239 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=196513.33333333334, ans=0.1 2024-09-23 06:22:51,021 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=11.01 vs. limit=15.0 2024-09-23 06:22:53,533 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.158e+02 1.336e+02 1.469e+02 1.670e+02 2.912e+02, threshold=2.938e+02, percent-clipped=0.0 2024-09-23 06:23:04,101 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=6.15 vs. limit=12.0 2024-09-23 06:23:29,551 INFO [train.py:1198] (2/4) Epoch 11, batch 3200, loss[loss=0.2668, ctc_loss=0.1862, cr_loss=0.4031, over 17217.00 frames. ], tot_loss[loss=0.2452, ctc_loss=0.1693, cr_loss=0.3794, over 3356339.01 frames. ], batch size: 50, lr: 1.08e-02, grad_scale: 32.0 2024-09-23 06:23:52,970 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=196793.33333333334, ans=0.125 2024-09-23 06:24:14,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=196886.66666666666, ans=0.1 2024-09-23 06:24:47,352 INFO [train.py:1198] (2/4) Epoch 11, batch 3250, loss[loss=0.2013, ctc_loss=0.1359, cr_loss=0.3271, over 17010.00 frames. ], tot_loss[loss=0.245, ctc_loss=0.1691, cr_loss=0.3795, over 3360131.55 frames. ], batch size: 44, lr: 1.08e-02, grad_scale: 16.0 2024-09-23 06:25:02,935 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=197026.66666666666, ans=0.125 2024-09-23 06:25:06,529 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=7.13 vs. limit=15.0 2024-09-23 06:25:30,817 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.055e+02 1.294e+02 1.384e+02 1.561e+02 5.237e+02, threshold=2.769e+02, percent-clipped=1.0 2024-09-23 06:25:32,684 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=197120.0, ans=0.125 2024-09-23 06:25:34,292 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=197120.0, ans=0.0 2024-09-23 06:25:49,790 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=197166.66666666666, ans=0.1 2024-09-23 06:25:49,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=197166.66666666666, ans=0.125 2024-09-23 06:26:05,132 INFO [train.py:1198] (2/4) Epoch 11, batch 3300, loss[loss=0.2259, ctc_loss=0.1493, cr_loss=0.3831, over 17174.00 frames. ], tot_loss[loss=0.245, ctc_loss=0.1691, cr_loss=0.3795, over 3355855.30 frames. ], batch size: 45, lr: 1.08e-02, grad_scale: 16.0 2024-09-23 06:26:23,783 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=197260.0, ans=0.1 2024-09-23 06:26:45,613 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=197306.66666666666, ans=0.0 2024-09-23 06:27:13,691 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=197400.0, ans=0.025 2024-09-23 06:27:22,859 INFO [train.py:1198] (2/4) Epoch 11, batch 3350, loss[loss=0.2425, ctc_loss=0.1661, cr_loss=0.382, over 17355.00 frames. ], tot_loss[loss=0.2447, ctc_loss=0.1688, cr_loss=0.3791, over 3358618.46 frames. ], batch size: 48, lr: 1.08e-02, grad_scale: 16.0 2024-09-23 06:27:31,034 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=197446.66666666666, ans=0.0 2024-09-23 06:27:45,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten.whitening_limit, batch_count=197493.33333333334, ans=15.0 2024-09-23 06:28:00,689 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=197540.0, ans=0.125 2024-09-23 06:28:05,426 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=197540.0, ans=0.0 2024-09-23 06:28:06,586 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.135e+02 1.310e+02 1.428e+02 1.594e+02 2.854e+02, threshold=2.856e+02, percent-clipped=1.0 2024-09-23 06:28:10,535 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=19.63 vs. limit=22.5 2024-09-23 06:28:11,768 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=197586.66666666666, ans=0.0 2024-09-23 06:28:31,075 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn1.whiten, num_groups=1, num_channels=192, metric=9.20 vs. limit=22.5 2024-09-23 06:28:41,094 INFO [train.py:1198] (2/4) Epoch 11, batch 3400, loss[loss=0.2143, ctc_loss=0.1434, cr_loss=0.3547, over 17101.00 frames. ], tot_loss[loss=0.2441, ctc_loss=0.1683, cr_loss=0.379, over 3364107.44 frames. ], batch size: 40, lr: 1.08e-02, grad_scale: 16.0 2024-09-23 06:28:56,727 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=197726.66666666666, ans=0.0 2024-09-23 06:29:00,772 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.22 vs. limit=15.0 2024-09-23 06:29:07,987 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=197726.66666666666, ans=0.1 2024-09-23 06:29:18,960 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=197773.33333333334, ans=0.025 2024-09-23 06:29:56,676 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=197866.66666666666, ans=0.125 2024-09-23 06:30:04,326 INFO [train.py:1198] (2/4) Epoch 11, batch 3450, loss[loss=0.2388, ctc_loss=0.1632, cr_loss=0.3778, over 17122.00 frames. ], tot_loss[loss=0.2448, ctc_loss=0.1689, cr_loss=0.3799, over 3354296.94 frames. ], batch size: 48, lr: 1.08e-02, grad_scale: 16.0 2024-09-23 06:30:21,225 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=10.21 vs. limit=15.0 2024-09-23 06:30:28,411 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=197960.0, ans=0.2 2024-09-23 06:30:34,364 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=197960.0, ans=0.1 2024-09-23 06:30:42,063 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.min_positive, batch_count=198006.66666666666, ans=0.05 2024-09-23 06:30:49,691 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.122e+02 1.318e+02 1.398e+02 1.528e+02 2.343e+02, threshold=2.795e+02, percent-clipped=0.0 2024-09-23 06:30:50,457 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.96 vs. limit=6.0 2024-09-23 06:31:26,438 INFO [train.py:1198] (2/4) Epoch 11, batch 3500, loss[loss=0.25, ctc_loss=0.1737, cr_loss=0.381, over 17001.00 frames. ], tot_loss[loss=0.2443, ctc_loss=0.1683, cr_loss=0.3796, over 3363964.78 frames. ], batch size: 56, lr: 1.08e-02, grad_scale: 16.0 2024-09-23 06:31:26,776 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=198146.66666666666, ans=0.0 2024-09-23 06:31:40,860 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=198193.33333333334, ans=0.0 2024-09-23 06:31:46,931 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=198193.33333333334, ans=0.125 2024-09-23 06:31:52,209 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=4.34 vs. limit=15.0 2024-09-23 06:32:21,134 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=198286.66666666666, ans=0.07 2024-09-23 06:32:24,724 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.39 vs. limit=15.0 2024-09-23 06:32:36,811 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.00 vs. limit=15.0 2024-09-23 06:32:41,519 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.74 vs. limit=15.0 2024-09-23 06:32:44,006 INFO [train.py:1198] (2/4) Epoch 11, batch 3550, loss[loss=0.2326, ctc_loss=0.1554, cr_loss=0.3859, over 17344.00 frames. ], tot_loss[loss=0.2444, ctc_loss=0.1684, cr_loss=0.3799, over 3365526.54 frames. ], batch size: 48, lr: 1.08e-02, grad_scale: 16.0 2024-09-23 06:32:50,789 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=198380.0, ans=0.0 2024-09-23 06:33:01,511 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=198426.66666666666, ans=0.0 2024-09-23 06:33:03,268 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=198426.66666666666, ans=0.0 2024-09-23 06:33:06,171 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=198426.66666666666, ans=0.125 2024-09-23 06:33:26,354 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=198473.33333333334, ans=0.2 2024-09-23 06:33:27,727 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.143e+02 1.313e+02 1.389e+02 1.597e+02 2.419e+02, threshold=2.778e+02, percent-clipped=0.0 2024-09-23 06:33:40,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=198520.0, ans=0.125 2024-09-23 06:33:41,950 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=198520.0, ans=0.015 2024-09-23 06:33:45,303 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=198566.66666666666, ans=0.0 2024-09-23 06:33:45,324 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=198566.66666666666, ans=0.125 2024-09-23 06:33:58,475 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.79 vs. limit=6.0 2024-09-23 06:34:02,340 INFO [train.py:1198] (2/4) Epoch 11, batch 3600, loss[loss=0.263, ctc_loss=0.1852, cr_loss=0.3891, over 16891.00 frames. ], tot_loss[loss=0.2457, ctc_loss=0.1696, cr_loss=0.3807, over 3356307.80 frames. ], batch size: 58, lr: 1.08e-02, grad_scale: 32.0 2024-09-23 06:34:19,896 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=198660.0, ans=0.125 2024-09-23 06:35:09,060 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=198800.0, ans=0.0 2024-09-23 06:35:21,044 INFO [train.py:1198] (2/4) Epoch 11, batch 3650, loss[loss=0.2416, ctc_loss=0.1632, cr_loss=0.3924, over 17050.00 frames. ], tot_loss[loss=0.2458, ctc_loss=0.1697, cr_loss=0.3805, over 3361626.29 frames. ], batch size: 52, lr: 1.08e-02, grad_scale: 32.0 2024-09-23 06:35:33,784 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=198846.66666666666, ans=0.2 2024-09-23 06:35:47,769 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=198893.33333333334, ans=0.0 2024-09-23 06:35:50,888 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=198940.0, ans=0.0 2024-09-23 06:36:03,751 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten.whitening_limit, batch_count=198940.0, ans=22.5 2024-09-23 06:36:04,372 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.052e+02 1.307e+02 1.406e+02 1.514e+02 2.450e+02, threshold=2.812e+02, percent-clipped=0.0 2024-09-23 06:36:36,553 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=199033.33333333334, ans=0.125 2024-09-23 06:36:39,290 INFO [train.py:1198] (2/4) Epoch 11, batch 3700, loss[loss=0.2807, ctc_loss=0.1879, cr_loss=0.4637, over 17013.00 frames. ], tot_loss[loss=0.2461, ctc_loss=0.1699, cr_loss=0.3808, over 3354478.69 frames. ], batch size: 53, lr: 1.08e-02, grad_scale: 32.0 2024-09-23 06:37:23,270 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=199173.33333333334, ans=0.0 2024-09-23 06:37:26,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=199220.0, ans=0.1 2024-09-23 06:37:51,585 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=199266.66666666666, ans=10.0 2024-09-23 06:37:57,658 INFO [train.py:1198] (2/4) Epoch 11, batch 3750, loss[loss=0.2709, ctc_loss=0.1878, cr_loss=0.4154, over 17149.00 frames. ], tot_loss[loss=0.2466, ctc_loss=0.1703, cr_loss=0.3815, over 3351227.71 frames. ], batch size: 48, lr: 1.08e-02, grad_scale: 32.0 2024-09-23 06:38:16,971 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.73 vs. limit=22.5 2024-09-23 06:38:25,867 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=199360.0, ans=0.125 2024-09-23 06:38:41,139 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.147e+02 1.417e+02 1.573e+02 1.870e+02 3.069e+02, threshold=3.146e+02, percent-clipped=3.0 2024-09-23 06:38:47,567 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=199453.33333333334, ans=0.125 2024-09-23 06:39:16,468 INFO [train.py:1198] (2/4) Epoch 11, batch 3800, loss[loss=0.2563, ctc_loss=0.1795, cr_loss=0.3839, over 16871.00 frames. ], tot_loss[loss=0.247, ctc_loss=0.1707, cr_loss=0.3815, over 3328394.49 frames. ], batch size: 58, lr: 1.08e-02, grad_scale: 32.0 2024-09-23 06:39:38,432 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=11.03 vs. limit=15.0 2024-09-23 06:40:34,879 INFO [train.py:1198] (2/4) Epoch 11, batch 3850, loss[loss=0.2908, ctc_loss=0.2084, cr_loss=0.4117, over 12310.00 frames. ], tot_loss[loss=0.2492, ctc_loss=0.1726, cr_loss=0.383, over 3294320.49 frames. ], batch size: 123, lr: 1.07e-02, grad_scale: 16.0 2024-09-23 06:40:54,883 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=199826.66666666666, ans=0.1 2024-09-23 06:41:00,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=199826.66666666666, ans=0.025 2024-09-23 06:41:14,442 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=199873.33333333334, ans=0.2 2024-09-23 06:41:18,610 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.202e+02 1.500e+02 1.672e+02 1.857e+02 2.511e+02, threshold=3.343e+02, percent-clipped=0.0 2024-09-23 06:41:20,462 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=199920.0, ans=0.0 2024-09-23 06:41:22,066 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=199920.0, ans=0.1 2024-09-23 06:41:33,986 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=199966.66666666666, ans=0.125 2024-09-23 06:42:27,975 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.40 vs. limit=22.5 2024-09-23 06:42:36,528 INFO [train.py:1198] (2/4) Epoch 12, batch 0, loss[loss=0.2516, ctc_loss=0.1738, cr_loss=0.3889, over 17226.00 frames. ], tot_loss[loss=0.2516, ctc_loss=0.1738, cr_loss=0.3889, over 17226.00 frames. ], batch size: 47, lr: 1.03e-02, grad_scale: 32.0 2024-09-23 06:42:36,528 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 06:42:52,077 INFO [train.py:1230] (2/4) Epoch 12, validation: loss=0.0478, ctc_loss=0.0478, cr_loss=7.52e-15, over 944034.00 frames. 2024-09-23 06:42:52,078 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 06:43:22,678 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=200088.0, ans=0.125 2024-09-23 06:43:24,346 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=200088.0, ans=0.0 2024-09-23 06:43:40,767 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=8.07 vs. limit=15.0 2024-09-23 06:43:41,584 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=200134.66666666666, ans=0.0 2024-09-23 06:43:45,020 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=10.74 vs. limit=15.0 2024-09-23 06:43:57,525 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=200181.33333333334, ans=0.025 2024-09-23 06:44:05,341 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=200181.33333333334, ans=0.1 2024-09-23 06:44:07,075 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=200181.33333333334, ans=0.2 2024-09-23 06:44:11,654 INFO [train.py:1198] (2/4) Epoch 12, batch 50, loss[loss=0.2237, ctc_loss=0.1549, cr_loss=0.3439, over 17218.00 frames. ], tot_loss[loss=0.2437, ctc_loss=0.1681, cr_loss=0.3778, over 757668.90 frames. ], batch size: 55, lr: 1.03e-02, grad_scale: 32.0 2024-09-23 06:44:26,360 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.89 vs. limit=15.0 2024-09-23 06:44:47,699 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=200274.66666666666, ans=0.1 2024-09-23 06:45:06,542 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=200368.0, ans=0.0 2024-09-23 06:45:12,534 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.318e+02 1.438e+02 1.597e+02 2.419e+02, threshold=2.876e+02, percent-clipped=0.0 2024-09-23 06:45:24,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=200414.66666666666, ans=0.0 2024-09-23 06:45:40,749 INFO [train.py:1198] (2/4) Epoch 12, batch 100, loss[loss=0.2537, ctc_loss=0.1714, cr_loss=0.4115, over 17039.00 frames. ], tot_loss[loss=0.2439, ctc_loss=0.168, cr_loss=0.3794, over 1326650.30 frames. ], batch size: 44, lr: 1.03e-02, grad_scale: 16.0 2024-09-23 06:45:51,143 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=13.17 vs. limit=22.5 2024-09-23 06:46:11,612 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=200554.66666666666, ans=0.0 2024-09-23 06:47:00,675 INFO [train.py:1198] (2/4) Epoch 12, batch 150, loss[loss=0.221, ctc_loss=0.1486, cr_loss=0.3618, over 17274.00 frames. ], tot_loss[loss=0.2424, ctc_loss=0.167, cr_loss=0.377, over 1784029.68 frames. ], batch size: 42, lr: 1.03e-02, grad_scale: 16.0 2024-09-23 06:47:05,727 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=200694.66666666666, ans=0.04949747468305833 2024-09-23 06:47:25,416 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=5.52 vs. limit=12.0 2024-09-23 06:47:37,629 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=200788.0, ans=0.125 2024-09-23 06:47:54,508 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.081e+02 1.306e+02 1.461e+02 1.660e+02 2.321e+02, threshold=2.923e+02, percent-clipped=0.0 2024-09-23 06:48:19,935 INFO [train.py:1198] (2/4) Epoch 12, batch 200, loss[loss=0.2688, ctc_loss=0.1867, cr_loss=0.4106, over 16644.00 frames. ], tot_loss[loss=0.2436, ctc_loss=0.1679, cr_loss=0.3783, over 2128724.48 frames. ], batch size: 66, lr: 1.03e-02, grad_scale: 16.0 2024-09-23 06:48:32,782 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=200928.0, ans=0.125 2024-09-23 06:48:35,947 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=200974.66666666666, ans=0.125 2024-09-23 06:49:15,499 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=201068.0, ans=0.0 2024-09-23 06:49:16,229 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.68 vs. limit=15.0 2024-09-23 06:49:25,047 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=201114.66666666666, ans=0.125 2024-09-23 06:49:26,504 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=201114.66666666666, ans=0.2 2024-09-23 06:49:45,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=201114.66666666666, ans=0.07 2024-09-23 06:49:46,493 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=201161.33333333334, ans=0.025 2024-09-23 06:49:47,797 INFO [train.py:1198] (2/4) Epoch 12, batch 250, loss[loss=0.2288, ctc_loss=0.1593, cr_loss=0.3476, over 16942.00 frames. ], tot_loss[loss=0.244, ctc_loss=0.1681, cr_loss=0.3793, over 2405312.81 frames. ], batch size: 42, lr: 1.03e-02, grad_scale: 16.0 2024-09-23 06:50:02,368 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=201208.0, ans=0.125 2024-09-23 06:50:05,703 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 06:50:30,181 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=11.37 vs. limit=22.5 2024-09-23 06:50:45,029 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.112e+02 1.293e+02 1.367e+02 1.506e+02 2.268e+02, threshold=2.734e+02, percent-clipped=0.0 2024-09-23 06:50:48,485 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=201301.33333333334, ans=0.0 2024-09-23 06:50:49,225 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.62 vs. limit=15.0 2024-09-23 06:50:54,960 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=201348.0, ans=0.035 2024-09-23 06:50:55,578 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=12.64 vs. limit=15.0 2024-09-23 06:51:10,809 INFO [train.py:1198] (2/4) Epoch 12, batch 300, loss[loss=0.2434, ctc_loss=0.1699, cr_loss=0.3671, over 17155.00 frames. ], tot_loss[loss=0.2428, ctc_loss=0.1673, cr_loss=0.3774, over 2608326.05 frames. ], batch size: 45, lr: 1.03e-02, grad_scale: 16.0 2024-09-23 06:51:15,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=201394.66666666666, ans=0.125 2024-09-23 06:51:27,315 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.48 vs. limit=6.0 2024-09-23 06:51:53,259 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=21.83 vs. limit=22.5 2024-09-23 06:52:00,654 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=201534.66666666666, ans=0.125 2024-09-23 06:52:21,170 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=201581.33333333334, ans=0.2 2024-09-23 06:52:30,548 INFO [train.py:1198] (2/4) Epoch 12, batch 350, loss[loss=0.2356, ctc_loss=0.1615, cr_loss=0.3704, over 17091.00 frames. ], tot_loss[loss=0.2427, ctc_loss=0.1673, cr_loss=0.3773, over 2768979.54 frames. ], batch size: 49, lr: 1.02e-02, grad_scale: 16.0 2024-09-23 06:53:01,701 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=201721.33333333334, ans=0.125 2024-09-23 06:53:08,849 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=7.40 vs. limit=15.0 2024-09-23 06:53:24,280 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=201768.0, ans=0.125 2024-09-23 06:53:25,561 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.271e+02 1.388e+02 1.541e+02 2.258e+02, threshold=2.777e+02, percent-clipped=0.0 2024-09-23 06:53:43,431 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=201814.66666666666, ans=0.1 2024-09-23 06:53:51,107 INFO [train.py:1198] (2/4) Epoch 12, batch 400, loss[loss=0.2216, ctc_loss=0.1503, cr_loss=0.3563, over 17087.00 frames. ], tot_loss[loss=0.2424, ctc_loss=0.1669, cr_loss=0.3774, over 2904238.86 frames. ], batch size: 43, lr: 1.02e-02, grad_scale: 32.0 2024-09-23 06:54:44,317 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=201954.66666666666, ans=0.0 2024-09-23 06:54:45,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=202001.33333333334, ans=0.125 2024-09-23 06:54:52,225 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.min_positive, batch_count=202001.33333333334, ans=0.05 2024-09-23 06:55:18,833 INFO [train.py:1198] (2/4) Epoch 12, batch 450, loss[loss=0.2463, ctc_loss=0.1684, cr_loss=0.3895, over 17033.00 frames. ], tot_loss[loss=0.2417, ctc_loss=0.1665, cr_loss=0.376, over 2999968.43 frames. ], batch size: 51, lr: 1.02e-02, grad_scale: 32.0 2024-09-23 06:55:20,744 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=202094.66666666666, ans=0.0 2024-09-23 06:56:15,455 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.139e+02 1.330e+02 1.519e+02 1.735e+02 3.092e+02, threshold=3.038e+02, percent-clipped=2.0 2024-09-23 06:56:40,845 INFO [train.py:1198] (2/4) Epoch 12, batch 500, loss[loss=0.2955, ctc_loss=0.2186, cr_loss=0.3842, over 12580.00 frames. ], tot_loss[loss=0.2419, ctc_loss=0.1666, cr_loss=0.3761, over 3079668.25 frames. ], batch size: 123, lr: 1.02e-02, grad_scale: 32.0 2024-09-23 06:57:01,985 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.min_abs, batch_count=202374.66666666666, ans=0.5 2024-09-23 06:57:03,411 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=202374.66666666666, ans=0.0 2024-09-23 06:57:05,200 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=202374.66666666666, ans=0.125 2024-09-23 06:57:11,812 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=17.93 vs. limit=22.5 2024-09-23 06:57:21,452 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.68 vs. limit=6.0 2024-09-23 06:58:00,476 INFO [train.py:1198] (2/4) Epoch 12, batch 550, loss[loss=0.1931, ctc_loss=0.1287, cr_loss=0.3223, over 17087.00 frames. ], tot_loss[loss=0.2413, ctc_loss=0.1661, cr_loss=0.3761, over 3142773.90 frames. ], batch size: 40, lr: 1.02e-02, grad_scale: 32.0 2024-09-23 06:58:08,817 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=202561.33333333334, ans=0.125 2024-09-23 06:58:23,034 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=202608.0, ans=0.1 2024-09-23 06:58:26,307 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=202608.0, ans=0.025 2024-09-23 06:58:45,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=202654.66666666666, ans=0.125 2024-09-23 06:58:50,648 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.88 vs. limit=15.0 2024-09-23 06:58:54,272 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.042e+02 1.321e+02 1.421e+02 1.522e+02 2.085e+02, threshold=2.842e+02, percent-clipped=0.0 2024-09-23 06:59:12,147 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=202748.0, ans=0.125 2024-09-23 06:59:21,039 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=202794.66666666666, ans=0.0 2024-09-23 06:59:22,365 INFO [train.py:1198] (2/4) Epoch 12, batch 600, loss[loss=0.2775, ctc_loss=0.1947, cr_loss=0.4138, over 16741.00 frames. ], tot_loss[loss=0.2429, ctc_loss=0.1672, cr_loss=0.3782, over 3195160.93 frames. ], batch size: 61, lr: 1.02e-02, grad_scale: 32.0 2024-09-23 06:59:53,552 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=202841.33333333334, ans=0.0 2024-09-23 07:00:03,387 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=5.60 vs. limit=15.0 2024-09-23 07:00:04,589 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=202888.0, ans=0.125 2024-09-23 07:00:04,601 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=202888.0, ans=0.025 2024-09-23 07:00:26,783 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=202934.66666666666, ans=0.125 2024-09-23 07:00:34,359 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=202981.33333333334, ans=0.0 2024-09-23 07:00:50,042 INFO [train.py:1198] (2/4) Epoch 12, batch 650, loss[loss=0.2768, ctc_loss=0.1912, cr_loss=0.4282, over 15969.00 frames. ], tot_loss[loss=0.2428, ctc_loss=0.167, cr_loss=0.3786, over 3235671.56 frames. ], batch size: 74, lr: 1.02e-02, grad_scale: 32.0 2024-09-23 07:00:53,665 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=203028.0, ans=0.125 2024-09-23 07:01:09,508 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=203074.66666666666, ans=0.125 2024-09-23 07:01:12,706 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=203074.66666666666, ans=0.1 2024-09-23 07:01:20,515 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=203121.33333333334, ans=0.0 2024-09-23 07:01:25,317 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=203121.33333333334, ans=0.2 2024-09-23 07:01:44,224 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.271e+02 1.387e+02 1.567e+02 2.285e+02, threshold=2.773e+02, percent-clipped=0.0 2024-09-23 07:01:46,121 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 07:01:54,153 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=203214.66666666666, ans=0.125 2024-09-23 07:02:09,751 INFO [train.py:1198] (2/4) Epoch 12, batch 700, loss[loss=0.2317, ctc_loss=0.156, cr_loss=0.3784, over 17219.00 frames. ], tot_loss[loss=0.2431, ctc_loss=0.1673, cr_loss=0.3789, over 3264966.16 frames. ], batch size: 47, lr: 1.02e-02, grad_scale: 32.0 2024-09-23 07:02:50,060 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=203354.66666666666, ans=0.1 2024-09-23 07:03:06,192 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.41 vs. limit=22.5 2024-09-23 07:03:22,911 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.34 vs. limit=22.5 2024-09-23 07:03:28,586 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=203494.66666666666, ans=0.0 2024-09-23 07:03:29,844 INFO [train.py:1198] (2/4) Epoch 12, batch 750, loss[loss=0.2793, ctc_loss=0.2001, cr_loss=0.3959, over 14955.00 frames. ], tot_loss[loss=0.2429, ctc_loss=0.1673, cr_loss=0.378, over 3286239.35 frames. ], batch size: 89, lr: 1.02e-02, grad_scale: 32.0 2024-09-23 07:03:31,905 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=203494.66666666666, ans=0.125 2024-09-23 07:03:47,593 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 07:03:47,599 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=203541.33333333334, ans=0.125 2024-09-23 07:04:21,250 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.53 vs. limit=15.0 2024-09-23 07:04:29,347 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.059e+02 1.298e+02 1.398e+02 1.549e+02 2.740e+02, threshold=2.796e+02, percent-clipped=0.0 2024-09-23 07:04:45,524 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.50 vs. limit=12.0 2024-09-23 07:04:48,237 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=203681.33333333334, ans=0.125 2024-09-23 07:04:57,396 INFO [train.py:1198] (2/4) Epoch 12, batch 800, loss[loss=0.2652, ctc_loss=0.1849, cr_loss=0.4018, over 17231.00 frames. ], tot_loss[loss=0.2421, ctc_loss=0.1668, cr_loss=0.3766, over 3294073.03 frames. ], batch size: 55, lr: 1.02e-02, grad_scale: 32.0 2024-09-23 07:05:21,718 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten.whitening_limit, batch_count=203774.66666666666, ans=15.0 2024-09-23 07:05:28,616 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=203774.66666666666, ans=0.2 2024-09-23 07:05:40,449 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=15.45 vs. limit=15.0 2024-09-23 07:06:01,974 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=5.46 vs. limit=10.0 2024-09-23 07:06:18,738 INFO [train.py:1198] (2/4) Epoch 12, batch 850, loss[loss=0.248, ctc_loss=0.1705, cr_loss=0.3879, over 17309.00 frames. ], tot_loss[loss=0.2425, ctc_loss=0.1671, cr_loss=0.377, over 3290671.23 frames. ], batch size: 49, lr: 1.02e-02, grad_scale: 32.0 2024-09-23 07:06:19,642 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.71 vs. limit=15.0 2024-09-23 07:06:36,629 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 07:06:38,816 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=5.14 vs. limit=10.0 2024-09-23 07:07:12,568 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.085e+02 1.288e+02 1.402e+02 1.556e+02 2.231e+02, threshold=2.804e+02, percent-clipped=0.0 2024-09-23 07:07:22,824 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.13 vs. limit=15.0 2024-09-23 07:07:36,921 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=204194.66666666666, ans=0.0 2024-09-23 07:07:38,074 INFO [train.py:1198] (2/4) Epoch 12, batch 900, loss[loss=0.253, ctc_loss=0.1727, cr_loss=0.4013, over 17358.00 frames. ], tot_loss[loss=0.2423, ctc_loss=0.1668, cr_loss=0.3772, over 3305848.97 frames. ], batch size: 48, lr: 1.02e-02, grad_scale: 32.0 2024-09-23 07:08:02,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=204241.33333333334, ans=0.025 2024-09-23 07:08:29,142 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=204334.66666666666, ans=0.125 2024-09-23 07:08:46,569 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=204381.33333333334, ans=0.0 2024-09-23 07:08:56,511 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=204428.0, ans=0.0 2024-09-23 07:08:57,659 INFO [train.py:1198] (2/4) Epoch 12, batch 950, loss[loss=0.2528, ctc_loss=0.1774, cr_loss=0.3771, over 17260.00 frames. ], tot_loss[loss=0.2407, ctc_loss=0.1656, cr_loss=0.3756, over 3324183.80 frames. ], batch size: 44, lr: 1.02e-02, grad_scale: 32.0 2024-09-23 07:09:01,149 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=204428.0, ans=0.0 2024-09-23 07:09:50,172 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=204521.33333333334, ans=0.125 2024-09-23 07:09:50,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=204521.33333333334, ans=0.125 2024-09-23 07:09:55,259 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=204568.0, ans=0.125 2024-09-23 07:09:59,531 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.109e+02 1.295e+02 1.392e+02 1.548e+02 2.202e+02, threshold=2.785e+02, percent-clipped=0.0 2024-09-23 07:10:06,395 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=204568.0, ans=0.1 2024-09-23 07:10:14,390 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=204614.66666666666, ans=0.125 2024-09-23 07:10:14,804 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.88 vs. limit=15.0 2024-09-23 07:10:16,054 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=204614.66666666666, ans=0.125 2024-09-23 07:10:28,002 INFO [train.py:1198] (2/4) Epoch 12, batch 1000, loss[loss=0.2192, ctc_loss=0.1494, cr_loss=0.3488, over 17255.00 frames. ], tot_loss[loss=0.2419, ctc_loss=0.1665, cr_loss=0.3771, over 3341469.91 frames. ], batch size: 42, lr: 1.02e-02, grad_scale: 32.0 2024-09-23 07:10:39,145 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=204661.33333333334, ans=0.125 2024-09-23 07:10:46,200 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=20.56 vs. limit=22.5 2024-09-23 07:11:12,489 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=204754.66666666666, ans=0.0 2024-09-23 07:11:14,038 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=204801.33333333334, ans=0.125 2024-09-23 07:11:47,280 INFO [train.py:1198] (2/4) Epoch 12, batch 1050, loss[loss=0.2028, ctc_loss=0.1369, cr_loss=0.3295, over 16271.00 frames. ], tot_loss[loss=0.2416, ctc_loss=0.1662, cr_loss=0.3768, over 3351870.54 frames. ], batch size: 36, lr: 1.02e-02, grad_scale: 32.0 2024-09-23 07:12:19,486 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=204988.0, ans=0.125 2024-09-23 07:12:41,477 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.299e+02 1.476e+02 1.734e+02 2.912e+02, threshold=2.951e+02, percent-clipped=1.0 2024-09-23 07:13:07,026 INFO [train.py:1198] (2/4) Epoch 12, batch 1100, loss[loss=0.2063, ctc_loss=0.1376, cr_loss=0.3435, over 17201.00 frames. ], tot_loss[loss=0.2406, ctc_loss=0.1655, cr_loss=0.3754, over 3357759.71 frames. ], batch size: 47, lr: 1.02e-02, grad_scale: 32.0 2024-09-23 07:13:50,516 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=205221.33333333334, ans=0.0 2024-09-23 07:13:56,893 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=205268.0, ans=0.1 2024-09-23 07:14:07,899 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.34 vs. limit=15.0 2024-09-23 07:14:36,514 INFO [train.py:1198] (2/4) Epoch 12, batch 1150, loss[loss=0.2375, ctc_loss=0.1688, cr_loss=0.3434, over 16974.00 frames. ], tot_loss[loss=0.2408, ctc_loss=0.1656, cr_loss=0.3757, over 3355815.20 frames. ], batch size: 56, lr: 1.02e-02, grad_scale: 32.0 2024-09-23 07:15:06,889 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=10.91 vs. limit=15.0 2024-09-23 07:15:24,928 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=205501.33333333334, ans=0.125 2024-09-23 07:15:24,930 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.min_positive, batch_count=205501.33333333334, ans=0.025 2024-09-23 07:15:28,140 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=205501.33333333334, ans=0.125 2024-09-23 07:15:32,629 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.115e+02 1.321e+02 1.590e+02 1.805e+02 2.687e+02, threshold=3.179e+02, percent-clipped=0.0 2024-09-23 07:15:58,028 INFO [train.py:1198] (2/4) Epoch 12, batch 1200, loss[loss=0.2421, ctc_loss=0.1664, cr_loss=0.3788, over 17363.00 frames. ], tot_loss[loss=0.241, ctc_loss=0.1657, cr_loss=0.3769, over 3357983.55 frames. ], batch size: 48, lr: 1.02e-02, grad_scale: 32.0 2024-09-23 07:16:20,613 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=205641.33333333334, ans=0.0 2024-09-23 07:16:43,356 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.52 vs. limit=15.0 2024-09-23 07:16:45,877 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 07:16:53,832 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=205734.66666666666, ans=0.0 2024-09-23 07:17:17,301 INFO [train.py:1198] (2/4) Epoch 12, batch 1250, loss[loss=0.2108, ctc_loss=0.1436, cr_loss=0.3357, over 17073.00 frames. ], tot_loss[loss=0.2398, ctc_loss=0.1647, cr_loss=0.3755, over 3369714.53 frames. ], batch size: 46, lr: 1.01e-02, grad_scale: 32.0 2024-09-23 07:18:00,490 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=205921.33333333334, ans=0.1 2024-09-23 07:18:12,962 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.281e+02 1.384e+02 1.564e+02 2.899e+02, threshold=2.767e+02, percent-clipped=0.0 2024-09-23 07:18:30,860 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=206014.66666666666, ans=0.2 2024-09-23 07:18:36,852 INFO [train.py:1198] (2/4) Epoch 12, batch 1300, loss[loss=0.2385, ctc_loss=0.1641, cr_loss=0.3721, over 17218.00 frames. ], tot_loss[loss=0.2405, ctc_loss=0.1653, cr_loss=0.3758, over 3356311.54 frames. ], batch size: 47, lr: 1.01e-02, grad_scale: 16.0 2024-09-23 07:19:00,115 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=206108.0, ans=0.125 2024-09-23 07:19:05,376 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=6.23 vs. limit=12.0 2024-09-23 07:19:14,432 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=19.53 vs. limit=22.5 2024-09-23 07:19:30,734 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=206201.33333333334, ans=0.0 2024-09-23 07:20:03,897 INFO [train.py:1198] (2/4) Epoch 12, batch 1350, loss[loss=0.2028, ctc_loss=0.1389, cr_loss=0.32, over 17187.00 frames. ], tot_loss[loss=0.2408, ctc_loss=0.1656, cr_loss=0.3761, over 3349096.48 frames. ], batch size: 41, lr: 1.01e-02, grad_scale: 16.0 2024-09-23 07:20:28,853 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=206341.33333333334, ans=0.125 2024-09-23 07:20:45,302 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.55 vs. limit=15.0 2024-09-23 07:21:02,412 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.350e+02 1.483e+02 1.734e+02 2.832e+02, threshold=2.966e+02, percent-clipped=2.0 2024-09-23 07:21:26,076 INFO [train.py:1198] (2/4) Epoch 12, batch 1400, loss[loss=0.2491, ctc_loss=0.1691, cr_loss=0.3999, over 17204.00 frames. ], tot_loss[loss=0.2413, ctc_loss=0.1659, cr_loss=0.3769, over 3349136.78 frames. ], batch size: 47, lr: 1.01e-02, grad_scale: 16.0 2024-09-23 07:21:34,386 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=206528.0, ans=0.125 2024-09-23 07:22:12,889 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=206668.0, ans=0.025 2024-09-23 07:22:13,386 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=12.58 vs. limit=15.0 2024-09-23 07:22:36,895 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=206714.66666666666, ans=0.0 2024-09-23 07:22:36,910 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=206714.66666666666, ans=0.125 2024-09-23 07:22:46,290 INFO [train.py:1198] (2/4) Epoch 12, batch 1450, loss[loss=0.262, ctc_loss=0.1861, cr_loss=0.3795, over 15965.00 frames. ], tot_loss[loss=0.2413, ctc_loss=0.166, cr_loss=0.3763, over 3345962.58 frames. ], batch size: 74, lr: 1.01e-02, grad_scale: 16.0 2024-09-23 07:22:48,172 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=206761.33333333334, ans=0.125 2024-09-23 07:23:00,855 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 07:23:02,610 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=206808.0, ans=0.0 2024-09-23 07:23:21,519 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=206854.66666666666, ans=0.0 2024-09-23 07:23:42,183 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.071e+02 1.321e+02 1.435e+02 1.538e+02 2.142e+02, threshold=2.870e+02, percent-clipped=0.0 2024-09-23 07:23:45,492 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=206901.33333333334, ans=0.125 2024-09-23 07:24:03,753 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=206948.0, ans=0.0 2024-09-23 07:24:10,785 INFO [train.py:1198] (2/4) Epoch 12, batch 1500, loss[loss=0.2525, ctc_loss=0.1781, cr_loss=0.3722, over 17297.00 frames. ], tot_loss[loss=0.2408, ctc_loss=0.1656, cr_loss=0.3761, over 3354016.96 frames. ], batch size: 49, lr: 1.01e-02, grad_scale: 16.0 2024-09-23 07:24:18,910 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=206994.66666666666, ans=0.125 2024-09-23 07:24:27,727 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 07:24:27,780 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=207041.33333333334, ans=0.1 2024-09-23 07:24:43,477 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=207088.0, ans=0.125 2024-09-23 07:25:04,459 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=207134.66666666666, ans=0.2 2024-09-23 07:25:06,265 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=13.24 vs. limit=22.5 2024-09-23 07:25:21,486 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.58 vs. limit=22.5 2024-09-23 07:25:26,133 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=207181.33333333334, ans=0.125 2024-09-23 07:25:35,346 INFO [train.py:1198] (2/4) Epoch 12, batch 1550, loss[loss=0.2512, ctc_loss=0.1713, cr_loss=0.3992, over 17078.00 frames. ], tot_loss[loss=0.2417, ctc_loss=0.1662, cr_loss=0.3773, over 3353793.30 frames. ], batch size: 46, lr: 1.01e-02, grad_scale: 16.0 2024-09-23 07:25:35,680 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=207228.0, ans=0.0 2024-09-23 07:25:37,167 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=207228.0, ans=0.125 2024-09-23 07:25:49,100 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.66 vs. limit=15.0 2024-09-23 07:25:58,067 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=207274.66666666666, ans=0.125 2024-09-23 07:26:09,362 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=207321.33333333334, ans=0.2 2024-09-23 07:26:27,134 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=207368.0, ans=0.125 2024-09-23 07:26:30,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=207368.0, ans=0.125 2024-09-23 07:26:31,711 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.334e+02 1.473e+02 1.633e+02 2.267e+02, threshold=2.947e+02, percent-clipped=0.0 2024-09-23 07:26:55,792 INFO [train.py:1198] (2/4) Epoch 12, batch 1600, loss[loss=0.2613, ctc_loss=0.184, cr_loss=0.3866, over 14928.00 frames. ], tot_loss[loss=0.2422, ctc_loss=0.1668, cr_loss=0.3773, over 3349504.70 frames. ], batch size: 89, lr: 1.01e-02, grad_scale: 32.0 2024-09-23 07:26:59,645 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=5.05 vs. limit=12.0 2024-09-23 07:27:00,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=207461.33333333334, ans=0.0 2024-09-23 07:27:10,502 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=207508.0, ans=0.0 2024-09-23 07:27:19,996 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=207508.0, ans=0.2 2024-09-23 07:27:23,817 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=10.63 vs. limit=15.0 2024-09-23 07:27:32,980 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=12.56 vs. limit=15.0 2024-09-23 07:27:55,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=207601.33333333334, ans=0.125 2024-09-23 07:27:58,304 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=207648.0, ans=0.2 2024-09-23 07:28:14,245 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=207694.66666666666, ans=0.0 2024-09-23 07:28:14,900 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=5.10 vs. limit=10.0 2024-09-23 07:28:15,590 INFO [train.py:1198] (2/4) Epoch 12, batch 1650, loss[loss=0.2069, ctc_loss=0.1393, cr_loss=0.338, over 17037.00 frames. ], tot_loss[loss=0.2408, ctc_loss=0.1656, cr_loss=0.3758, over 3353641.39 frames. ], batch size: 39, lr: 1.01e-02, grad_scale: 32.0 2024-09-23 07:28:17,364 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=207694.66666666666, ans=0.125 2024-09-23 07:28:20,621 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=207694.66666666666, ans=0.0 2024-09-23 07:28:58,377 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 07:28:58,394 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=207788.0, ans=0.1 2024-09-23 07:29:07,276 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=207834.66666666666, ans=0.1 2024-09-23 07:29:11,943 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=207834.66666666666, ans=0.0 2024-09-23 07:29:16,235 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.260e+02 1.353e+02 1.462e+02 2.073e+02, threshold=2.705e+02, percent-clipped=0.0 2024-09-23 07:29:30,165 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=207881.33333333334, ans=0.125 2024-09-23 07:29:42,766 INFO [train.py:1198] (2/4) Epoch 12, batch 1700, loss[loss=0.267, ctc_loss=0.1854, cr_loss=0.4079, over 16985.00 frames. ], tot_loss[loss=0.2417, ctc_loss=0.1663, cr_loss=0.377, over 3341857.70 frames. ], batch size: 53, lr: 1.01e-02, grad_scale: 32.0 2024-09-23 07:30:17,894 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=208021.33333333334, ans=0.2 2024-09-23 07:30:22,831 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=208021.33333333334, ans=0.125 2024-09-23 07:31:03,613 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=12.28 vs. limit=15.0 2024-09-23 07:31:06,203 INFO [train.py:1198] (2/4) Epoch 12, batch 1750, loss[loss=0.2522, ctc_loss=0.1745, cr_loss=0.3884, over 16384.00 frames. ], tot_loss[loss=0.2414, ctc_loss=0.1659, cr_loss=0.3771, over 3347923.91 frames. ], batch size: 66, lr: 1.01e-02, grad_scale: 32.0 2024-09-23 07:31:28,984 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=208208.0, ans=0.125 2024-09-23 07:31:35,891 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=2.87 vs. limit=15.0 2024-09-23 07:32:02,197 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.329e+02 1.442e+02 1.637e+02 2.396e+02, threshold=2.884e+02, percent-clipped=0.0 2024-09-23 07:32:04,011 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=208301.33333333334, ans=0.1 2024-09-23 07:32:12,018 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=208348.0, ans=0.0 2024-09-23 07:32:25,983 INFO [train.py:1198] (2/4) Epoch 12, batch 1800, loss[loss=0.2356, ctc_loss=0.162, cr_loss=0.3679, over 17101.00 frames. ], tot_loss[loss=0.2416, ctc_loss=0.1661, cr_loss=0.3774, over 3342203.75 frames. ], batch size: 49, lr: 1.01e-02, grad_scale: 32.0 2024-09-23 07:32:31,471 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=6.23 vs. limit=15.0 2024-09-23 07:32:45,255 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=208441.33333333334, ans=0.125 2024-09-23 07:33:06,870 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=20.97 vs. limit=22.5 2024-09-23 07:33:42,688 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=208581.33333333334, ans=0.0 2024-09-23 07:33:42,851 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=208581.33333333334, ans=0.125 2024-09-23 07:33:45,651 INFO [train.py:1198] (2/4) Epoch 12, batch 1850, loss[loss=0.1968, ctc_loss=0.1326, cr_loss=0.3211, over 17274.00 frames. ], tot_loss[loss=0.2421, ctc_loss=0.1663, cr_loss=0.3786, over 3354435.81 frames. ], batch size: 42, lr: 1.01e-02, grad_scale: 32.0 2024-09-23 07:34:03,784 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=208628.0, ans=0.125 2024-09-23 07:34:31,832 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 07:34:41,169 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=208768.0, ans=0.0 2024-09-23 07:34:43,074 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.48 vs. limit=15.0 2024-09-23 07:34:45,863 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=208768.0, ans=0.0 2024-09-23 07:34:48,756 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.171e+02 1.290e+02 1.399e+02 1.526e+02 2.337e+02, threshold=2.797e+02, percent-clipped=0.0 2024-09-23 07:35:00,369 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=208814.66666666666, ans=0.2 2024-09-23 07:35:15,123 INFO [train.py:1198] (2/4) Epoch 12, batch 1900, loss[loss=0.3147, ctc_loss=0.2212, cr_loss=0.4677, over 15131.00 frames. ], tot_loss[loss=0.2437, ctc_loss=0.1675, cr_loss=0.3807, over 3352921.33 frames. ], batch size: 89, lr: 1.01e-02, grad_scale: 32.0 2024-09-23 07:35:16,973 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=208861.33333333334, ans=0.125 2024-09-23 07:35:37,499 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=208908.0, ans=0.0 2024-09-23 07:35:40,514 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=208908.0, ans=0.125 2024-09-23 07:35:51,442 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=208954.66666666666, ans=0.025 2024-09-23 07:35:55,432 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=18.99 vs. limit=22.5 2024-09-23 07:36:34,860 INFO [train.py:1198] (2/4) Epoch 12, batch 1950, loss[loss=0.2001, ctc_loss=0.1367, cr_loss=0.3167, over 16944.00 frames. ], tot_loss[loss=0.2438, ctc_loss=0.1678, cr_loss=0.3801, over 3347195.40 frames. ], batch size: 42, lr: 1.01e-02, grad_scale: 16.0 2024-09-23 07:36:44,552 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=209094.66666666666, ans=0.125 2024-09-23 07:37:12,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=209188.0, ans=0.2 2024-09-23 07:37:28,945 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=209234.66666666666, ans=0.125 2024-09-23 07:37:31,804 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.126e+02 1.280e+02 1.393e+02 1.533e+02 3.563e+02, threshold=2.786e+02, percent-clipped=1.0 2024-09-23 07:37:46,731 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=14.99 vs. limit=22.5 2024-09-23 07:37:54,115 INFO [train.py:1198] (2/4) Epoch 12, batch 2000, loss[loss=0.2582, ctc_loss=0.1787, cr_loss=0.3977, over 17375.00 frames. ], tot_loss[loss=0.2433, ctc_loss=0.1675, cr_loss=0.3793, over 3355280.16 frames. ], batch size: 48, lr: 1.01e-02, grad_scale: 32.0 2024-09-23 07:37:56,068 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=209328.0, ans=0.09899494936611666 2024-09-23 07:38:25,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=209421.33333333334, ans=0.1 2024-09-23 07:39:02,214 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.89 vs. limit=15.0 2024-09-23 07:39:07,100 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.66 vs. limit=15.0 2024-09-23 07:39:20,108 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=209561.33333333334, ans=0.1 2024-09-23 07:39:21,409 INFO [train.py:1198] (2/4) Epoch 12, batch 2050, loss[loss=0.2002, ctc_loss=0.1317, cr_loss=0.3428, over 16759.00 frames. ], tot_loss[loss=0.2416, ctc_loss=0.1661, cr_loss=0.3772, over 3359846.89 frames. ], batch size: 37, lr: 1.01e-02, grad_scale: 32.0 2024-09-23 07:39:28,040 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=209561.33333333334, ans=0.125 2024-09-23 07:40:16,727 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=17.70 vs. limit=22.5 2024-09-23 07:40:21,040 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.107e+02 1.313e+02 1.485e+02 1.661e+02 2.450e+02, threshold=2.969e+02, percent-clipped=0.0 2024-09-23 07:40:21,421 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=209701.33333333334, ans=0.125 2024-09-23 07:40:23,035 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=209701.33333333334, ans=0.0 2024-09-23 07:40:25,350 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module1.whiten, num_groups=1, num_channels=192, metric=6.37 vs. limit=15.0 2024-09-23 07:40:43,461 INFO [train.py:1198] (2/4) Epoch 12, batch 2100, loss[loss=0.2819, ctc_loss=0.2, cr_loss=0.4097, over 14987.00 frames. ], tot_loss[loss=0.2406, ctc_loss=0.1654, cr_loss=0.376, over 3362746.87 frames. ], batch size: 89, lr: 1.01e-02, grad_scale: 32.0 2024-09-23 07:40:58,150 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=209841.33333333334, ans=0.1 2024-09-23 07:41:01,391 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=209841.33333333334, ans=0.0 2024-09-23 07:41:07,550 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=209841.33333333334, ans=0.035 2024-09-23 07:41:25,227 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=209888.0, ans=0.125 2024-09-23 07:41:34,685 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=209934.66666666666, ans=0.125 2024-09-23 07:41:41,334 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.74 vs. limit=15.0 2024-09-23 07:41:45,993 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=209981.33333333334, ans=0.025 2024-09-23 07:41:59,018 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=209981.33333333334, ans=0.125 2024-09-23 07:42:02,479 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=210028.0, ans=0.05 2024-09-23 07:42:03,574 INFO [train.py:1198] (2/4) Epoch 12, batch 2150, loss[loss=0.2486, ctc_loss=0.1695, cr_loss=0.3959, over 17080.00 frames. ], tot_loss[loss=0.2395, ctc_loss=0.1645, cr_loss=0.3747, over 3365691.68 frames. ], batch size: 46, lr: 1.00e-02, grad_scale: 32.0 2024-09-23 07:42:21,132 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=210074.66666666666, ans=0.0 2024-09-23 07:42:24,426 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.min_abs, batch_count=210074.66666666666, ans=0.5 2024-09-23 07:42:54,655 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=210168.0, ans=0.1 2024-09-23 07:43:00,522 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.051e+02 1.318e+02 1.400e+02 1.574e+02 2.180e+02, threshold=2.800e+02, percent-clipped=0.0 2024-09-23 07:43:11,212 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn2.whiten, num_groups=1, num_channels=192, metric=9.41 vs. limit=22.5 2024-09-23 07:43:22,704 INFO [train.py:1198] (2/4) Epoch 12, batch 2200, loss[loss=0.2398, ctc_loss=0.1643, cr_loss=0.3778, over 15918.00 frames. ], tot_loss[loss=0.2392, ctc_loss=0.1642, cr_loss=0.3749, over 3361156.71 frames. ], batch size: 74, lr: 1.00e-02, grad_scale: 32.0 2024-09-23 07:43:36,324 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.07 vs. limit=10.0 2024-09-23 07:43:59,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=210354.66666666666, ans=0.125 2024-09-23 07:44:16,746 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=210401.33333333334, ans=0.125 2024-09-23 07:44:16,868 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=210401.33333333334, ans=0.125 2024-09-23 07:44:26,585 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=210401.33333333334, ans=0.125 2024-09-23 07:44:34,508 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 07:44:50,074 INFO [train.py:1198] (2/4) Epoch 12, batch 2250, loss[loss=0.2506, ctc_loss=0.1749, cr_loss=0.3784, over 17216.00 frames. ], tot_loss[loss=0.2401, ctc_loss=0.1648, cr_loss=0.3762, over 3365705.14 frames. ], batch size: 47, lr: 1.00e-02, grad_scale: 32.0 2024-09-23 07:45:02,462 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=210494.66666666666, ans=0.0 2024-09-23 07:45:03,953 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=210494.66666666666, ans=0.0 2024-09-23 07:45:10,278 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=210541.33333333334, ans=0.2 2024-09-23 07:45:49,555 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.113e+02 1.306e+02 1.486e+02 1.708e+02 2.318e+02, threshold=2.971e+02, percent-clipped=0.0 2024-09-23 07:46:04,256 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=210681.33333333334, ans=0.0 2024-09-23 07:46:11,906 INFO [train.py:1198] (2/4) Epoch 12, batch 2300, loss[loss=0.2113, ctc_loss=0.1408, cr_loss=0.3526, over 17272.00 frames. ], tot_loss[loss=0.2413, ctc_loss=0.1658, cr_loss=0.3773, over 3349208.00 frames. ], batch size: 42, lr: 1.00e-02, grad_scale: 32.0 2024-09-23 07:46:13,743 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=210728.0, ans=0.0 2024-09-23 07:46:48,896 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=210821.33333333334, ans=0.05 2024-09-23 07:46:52,618 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.18 vs. limit=12.0 2024-09-23 07:47:18,022 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=210914.66666666666, ans=0.025 2024-09-23 07:47:31,033 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=210961.33333333334, ans=0.1 2024-09-23 07:47:32,370 INFO [train.py:1198] (2/4) Epoch 12, batch 2350, loss[loss=0.275, ctc_loss=0.192, cr_loss=0.4152, over 16610.00 frames. ], tot_loss[loss=0.2414, ctc_loss=0.1659, cr_loss=0.3775, over 3343809.41 frames. ], batch size: 66, lr: 1.00e-02, grad_scale: 32.0 2024-09-23 07:48:29,120 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.329e+02 1.457e+02 1.570e+02 2.040e+02, threshold=2.915e+02, percent-clipped=0.0 2024-09-23 07:48:30,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=211101.33333333334, ans=0.0 2024-09-23 07:48:32,567 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=211101.33333333334, ans=0.0 2024-09-23 07:48:36,410 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=11.92 vs. limit=22.5 2024-09-23 07:48:53,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=211148.0, ans=0.2 2024-09-23 07:48:56,505 INFO [train.py:1198] (2/4) Epoch 12, batch 2400, loss[loss=0.246, ctc_loss=0.1684, cr_loss=0.3881, over 16989.00 frames. ], tot_loss[loss=0.2427, ctc_loss=0.1667, cr_loss=0.3797, over 3348289.94 frames. ], batch size: 53, lr: 1.00e-02, grad_scale: 32.0 2024-09-23 07:50:21,841 INFO [train.py:1198] (2/4) Epoch 12, batch 2450, loss[loss=0.2279, ctc_loss=0.1556, cr_loss=0.3617, over 17016.00 frames. ], tot_loss[loss=0.2409, ctc_loss=0.1653, cr_loss=0.3782, over 3358390.37 frames. ], batch size: 44, lr: 1.00e-02, grad_scale: 32.0 2024-09-23 07:50:22,102 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=211428.0, ans=0.125 2024-09-23 07:51:19,219 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.359e+02 1.540e+02 1.815e+02 2.529e+02, threshold=3.080e+02, percent-clipped=0.0 2024-09-23 07:51:19,379 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=211568.0, ans=0.125 2024-09-23 07:51:22,675 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=211568.0, ans=0.2 2024-09-23 07:51:41,363 INFO [train.py:1198] (2/4) Epoch 12, batch 2500, loss[loss=0.2561, ctc_loss=0.1747, cr_loss=0.407, over 17147.00 frames. ], tot_loss[loss=0.2402, ctc_loss=0.1647, cr_loss=0.3776, over 3360661.19 frames. ], batch size: 48, lr: 1.00e-02, grad_scale: 32.0 2024-09-23 07:52:11,689 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=211754.66666666666, ans=0.2 2024-09-23 07:52:12,373 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.66 vs. limit=12.0 2024-09-23 07:52:44,782 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=211848.0, ans=0.1 2024-09-23 07:52:44,909 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=211848.0, ans=0.125 2024-09-23 07:53:00,935 INFO [train.py:1198] (2/4) Epoch 12, batch 2550, loss[loss=0.2472, ctc_loss=0.1691, cr_loss=0.3909, over 15869.00 frames. ], tot_loss[loss=0.2407, ctc_loss=0.1651, cr_loss=0.3779, over 3357968.39 frames. ], batch size: 74, lr: 1.00e-02, grad_scale: 32.0 2024-09-23 07:53:43,807 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=211988.0, ans=0.125 2024-09-23 07:53:53,149 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=212034.66666666666, ans=0.0 2024-09-23 07:54:00,929 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.326e+02 1.484e+02 1.777e+02 2.748e+02, threshold=2.968e+02, percent-clipped=0.0 2024-09-23 07:54:24,413 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=212128.0, ans=0.125 2024-09-23 07:54:25,809 INFO [train.py:1198] (2/4) Epoch 12, batch 2600, loss[loss=0.2813, ctc_loss=0.1982, cr_loss=0.4157, over 15979.00 frames. ], tot_loss[loss=0.2414, ctc_loss=0.1655, cr_loss=0.3792, over 3360358.45 frames. ], batch size: 74, lr: 1.00e-02, grad_scale: 32.0 2024-09-23 07:54:33,950 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=212128.0, ans=0.0 2024-09-23 07:55:16,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=212268.0, ans=0.1 2024-09-23 07:55:18,114 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=212268.0, ans=0.1 2024-09-23 07:55:47,927 INFO [train.py:1198] (2/4) Epoch 12, batch 2650, loss[loss=0.2221, ctc_loss=0.1487, cr_loss=0.3671, over 17196.00 frames. ], tot_loss[loss=0.242, ctc_loss=0.1661, cr_loss=0.3795, over 3351101.49 frames. ], batch size: 41, lr: 9.99e-03, grad_scale: 32.0 2024-09-23 07:55:57,944 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=212361.33333333334, ans=0.0 2024-09-23 07:56:15,715 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=17.38 vs. limit=22.5 2024-09-23 07:56:23,390 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=212454.66666666666, ans=0.125 2024-09-23 07:56:25,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=212454.66666666666, ans=0.1 2024-09-23 07:56:37,405 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.91 vs. limit=5.0 2024-09-23 07:56:45,680 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.093e+02 1.240e+02 1.338e+02 1.470e+02 2.352e+02, threshold=2.676e+02, percent-clipped=0.0 2024-09-23 07:56:47,570 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=212501.33333333334, ans=0.0 2024-09-23 07:57:00,375 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=212548.0, ans=0.1 2024-09-23 07:57:08,181 INFO [train.py:1198] (2/4) Epoch 12, batch 2700, loss[loss=0.2291, ctc_loss=0.1539, cr_loss=0.3763, over 17233.00 frames. ], tot_loss[loss=0.2408, ctc_loss=0.1651, cr_loss=0.3781, over 3356480.29 frames. ], batch size: 50, lr: 9.99e-03, grad_scale: 32.0 2024-09-23 07:57:24,591 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=212641.33333333334, ans=0.125 2024-09-23 07:57:27,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=212641.33333333334, ans=0.125 2024-09-23 07:57:53,051 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=212688.0, ans=0.125 2024-09-23 07:58:06,224 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.87 vs. limit=22.5 2024-09-23 07:58:07,532 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer_ff3.min_abs, batch_count=212734.66666666666, ans=0.2 2024-09-23 07:58:13,973 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=212781.33333333334, ans=0.125 2024-09-23 07:58:20,661 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=212781.33333333334, ans=0.025 2024-09-23 07:58:22,412 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=212781.33333333334, ans=0.125 2024-09-23 07:58:28,274 INFO [train.py:1198] (2/4) Epoch 12, batch 2750, loss[loss=0.2368, ctc_loss=0.1638, cr_loss=0.365, over 17052.00 frames. ], tot_loss[loss=0.2405, ctc_loss=0.165, cr_loss=0.3778, over 3350336.48 frames. ], batch size: 52, lr: 9.98e-03, grad_scale: 32.0 2024-09-23 07:58:34,481 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=212828.0, ans=0.125 2024-09-23 07:59:33,712 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.136e+02 1.344e+02 1.475e+02 1.726e+02 2.611e+02, threshold=2.950e+02, percent-clipped=0.0 2024-09-23 07:59:46,340 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=213014.66666666666, ans=0.0 2024-09-23 07:59:47,900 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=213014.66666666666, ans=0.0 2024-09-23 07:59:52,731 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=213014.66666666666, ans=0.125 2024-09-23 07:59:58,918 INFO [train.py:1198] (2/4) Epoch 12, batch 2800, loss[loss=0.2379, ctc_loss=0.1611, cr_loss=0.3841, over 17271.00 frames. ], tot_loss[loss=0.241, ctc_loss=0.1655, cr_loss=0.3775, over 3347036.87 frames. ], batch size: 46, lr: 9.98e-03, grad_scale: 32.0 2024-09-23 08:00:07,793 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.06 vs. limit=10.0 2024-09-23 08:00:14,080 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=9.74 vs. limit=15.0 2024-09-23 08:00:24,694 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=213108.0, ans=0.0 2024-09-23 08:00:32,626 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=213154.66666666666, ans=0.2 2024-09-23 08:00:45,296 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=213201.33333333334, ans=0.125 2024-09-23 08:00:59,434 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=213201.33333333334, ans=0.1 2024-09-23 08:01:12,621 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.85 vs. limit=12.0 2024-09-23 08:01:18,197 INFO [train.py:1198] (2/4) Epoch 12, batch 2850, loss[loss=0.2596, ctc_loss=0.1782, cr_loss=0.4071, over 17038.00 frames. ], tot_loss[loss=0.2423, ctc_loss=0.1665, cr_loss=0.3787, over 3346176.42 frames. ], batch size: 52, lr: 9.97e-03, grad_scale: 32.0 2024-09-23 08:01:39,226 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=213341.33333333334, ans=0.125 2024-09-23 08:01:43,028 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.35 vs. limit=12.0 2024-09-23 08:02:03,198 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=213388.0, ans=0.2 2024-09-23 08:02:06,601 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=213434.66666666666, ans=0.125 2024-09-23 08:02:08,105 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 08:02:15,759 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.286e+02 1.451e+02 1.716e+02 2.634e+02, threshold=2.902e+02, percent-clipped=0.0 2024-09-23 08:02:32,003 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=213481.33333333334, ans=0.125 2024-09-23 08:02:35,207 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=213481.33333333334, ans=0.125 2024-09-23 08:02:38,157 INFO [train.py:1198] (2/4) Epoch 12, batch 2900, loss[loss=0.2343, ctc_loss=0.1591, cr_loss=0.3757, over 17064.00 frames. ], tot_loss[loss=0.2421, ctc_loss=0.1663, cr_loss=0.379, over 3344497.24 frames. ], batch size: 43, lr: 9.97e-03, grad_scale: 32.0 2024-09-23 08:02:41,575 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=213528.0, ans=0.2 2024-09-23 08:03:02,350 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=213574.66666666666, ans=0.125 2024-09-23 08:03:05,486 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=213574.66666666666, ans=0.125 2024-09-23 08:03:13,548 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=213621.33333333334, ans=0.125 2024-09-23 08:03:41,666 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=213668.0, ans=0.125 2024-09-23 08:03:41,691 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=213668.0, ans=0.2 2024-09-23 08:03:43,192 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.min_abs, batch_count=213668.0, ans=0.5 2024-09-23 08:04:04,995 INFO [train.py:1198] (2/4) Epoch 12, batch 2950, loss[loss=0.2675, ctc_loss=0.1861, cr_loss=0.4068, over 17145.00 frames. ], tot_loss[loss=0.2413, ctc_loss=0.1656, cr_loss=0.3783, over 3349278.54 frames. ], batch size: 48, lr: 9.96e-03, grad_scale: 32.0 2024-09-23 08:04:06,852 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=213761.33333333334, ans=0.0 2024-09-23 08:04:10,058 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=213761.33333333334, ans=0.125 2024-09-23 08:04:14,904 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=213761.33333333334, ans=0.07 2024-09-23 08:04:43,183 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=213854.66666666666, ans=0.125 2024-09-23 08:05:04,862 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.112e+02 1.287e+02 1.399e+02 1.568e+02 2.905e+02, threshold=2.798e+02, percent-clipped=1.0 2024-09-23 08:05:14,984 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=13.00 vs. limit=22.5 2024-09-23 08:05:26,774 INFO [train.py:1198] (2/4) Epoch 12, batch 3000, loss[loss=0.2602, ctc_loss=0.1812, cr_loss=0.3948, over 16895.00 frames. ], tot_loss[loss=0.2408, ctc_loss=0.1654, cr_loss=0.3774, over 3347963.40 frames. ], batch size: 58, lr: 9.96e-03, grad_scale: 32.0 2024-09-23 08:05:26,775 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 08:05:36,648 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.3.encoder.layers.2.self_attn_weights, attn_weights_entropy = tensor([2.5650, 3.0859, 3.0094, 3.4938, 2.8184, 2.8597, 3.3937, 3.5805], device='cuda:2') 2024-09-23 08:05:42,571 INFO [train.py:1230] (2/4) Epoch 12, validation: loss=0.04588, ctc_loss=0.04588, cr_loss=7.526e-15, over 944034.00 frames. 2024-09-23 08:05:42,571 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 08:06:22,169 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 08:06:28,041 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=214134.66666666666, ans=0.0 2024-09-23 08:07:00,679 INFO [train.py:1198] (2/4) Epoch 12, batch 3050, loss[loss=0.2541, ctc_loss=0.1772, cr_loss=0.3846, over 16916.00 frames. ], tot_loss[loss=0.24, ctc_loss=0.1647, cr_loss=0.3766, over 3352453.80 frames. ], batch size: 58, lr: 9.95e-03, grad_scale: 32.0 2024-09-23 08:07:03,897 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=214228.0, ans=0.125 2024-09-23 08:07:45,484 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=214368.0, ans=0.0 2024-09-23 08:07:53,849 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.41 vs. limit=12.0 2024-09-23 08:07:56,197 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.305e+02 1.426e+02 1.616e+02 2.340e+02, threshold=2.852e+02, percent-clipped=0.0 2024-09-23 08:07:58,546 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=14.68 vs. limit=15.0 2024-09-23 08:08:01,546 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.46 vs. limit=6.0 2024-09-23 08:08:02,709 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=214414.66666666666, ans=0.0 2024-09-23 08:08:17,698 INFO [train.py:1198] (2/4) Epoch 12, batch 3100, loss[loss=0.196, ctc_loss=0.1326, cr_loss=0.3172, over 17042.00 frames. ], tot_loss[loss=0.2416, ctc_loss=0.166, cr_loss=0.3781, over 3355663.23 frames. ], batch size: 39, lr: 9.94e-03, grad_scale: 32.0 2024-09-23 08:08:22,647 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=214461.33333333334, ans=0.1 2024-09-23 08:08:31,122 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.16 vs. limit=10.0 2024-09-23 08:09:21,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=214648.0, ans=0.0 2024-09-23 08:09:36,170 INFO [train.py:1198] (2/4) Epoch 12, batch 3150, loss[loss=0.2528, ctc_loss=0.1754, cr_loss=0.3873, over 17152.00 frames. ], tot_loss[loss=0.2399, ctc_loss=0.1647, cr_loss=0.376, over 3355851.88 frames. ], batch size: 48, lr: 9.94e-03, grad_scale: 32.0 2024-09-23 08:09:41,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.min_positive, batch_count=214694.66666666666, ans=0.05 2024-09-23 08:09:47,466 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=214694.66666666666, ans=0.1 2024-09-23 08:09:55,755 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=5.21 vs. limit=15.0 2024-09-23 08:10:04,428 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=214741.33333333334, ans=0.125 2024-09-23 08:10:06,010 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=214788.0, ans=0.1 2024-09-23 08:10:31,224 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=214834.66666666666, ans=0.0 2024-09-23 08:10:32,319 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.095e+02 1.379e+02 1.486e+02 1.636e+02 2.338e+02, threshold=2.971e+02, percent-clipped=0.0 2024-09-23 08:10:48,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=214881.33333333334, ans=0.2 2024-09-23 08:10:49,616 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=214881.33333333334, ans=0.125 2024-09-23 08:10:54,113 INFO [train.py:1198] (2/4) Epoch 12, batch 3200, loss[loss=0.2565, ctc_loss=0.1708, cr_loss=0.4285, over 17043.00 frames. ], tot_loss[loss=0.2395, ctc_loss=0.1644, cr_loss=0.3757, over 3351804.51 frames. ], batch size: 56, lr: 9.93e-03, grad_scale: 32.0 2024-09-23 08:11:13,227 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.47 vs. limit=6.0 2024-09-23 08:11:23,767 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 08:12:03,373 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=215114.66666666666, ans=0.0 2024-09-23 08:12:16,041 INFO [train.py:1198] (2/4) Epoch 12, batch 3250, loss[loss=0.2375, ctc_loss=0.1656, cr_loss=0.3596, over 17216.00 frames. ], tot_loss[loss=0.2385, ctc_loss=0.1637, cr_loss=0.374, over 3350065.20 frames. ], batch size: 50, lr: 9.93e-03, grad_scale: 16.0 2024-09-23 08:12:23,919 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=215161.33333333334, ans=0.1 2024-09-23 08:12:38,091 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=215208.0, ans=0.125 2024-09-23 08:13:01,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=215301.33333333334, ans=0.09899494936611666 2024-09-23 08:13:04,639 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=215301.33333333334, ans=0.0 2024-09-23 08:13:07,685 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=215301.33333333334, ans=0.125 2024-09-23 08:13:13,440 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.295e+02 1.414e+02 1.544e+02 2.184e+02, threshold=2.828e+02, percent-clipped=0.0 2024-09-23 08:13:14,156 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.05 vs. limit=10.0 2024-09-23 08:13:31,957 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.61 vs. limit=15.0 2024-09-23 08:13:34,637 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=215394.66666666666, ans=0.125 2024-09-23 08:13:35,944 INFO [train.py:1198] (2/4) Epoch 12, batch 3300, loss[loss=0.2103, ctc_loss=0.1397, cr_loss=0.3529, over 16780.00 frames. ], tot_loss[loss=0.2389, ctc_loss=0.1639, cr_loss=0.3751, over 3358108.74 frames. ], batch size: 37, lr: 9.92e-03, grad_scale: 16.0 2024-09-23 08:13:48,632 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=215394.66666666666, ans=0.0 2024-09-23 08:13:51,835 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=215441.33333333334, ans=0.125 2024-09-23 08:13:53,809 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=8.08 vs. limit=15.0 2024-09-23 08:14:05,824 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=215488.0, ans=0.125 2024-09-23 08:14:07,391 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=215488.0, ans=0.125 2024-09-23 08:14:23,011 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=215534.66666666666, ans=0.125 2024-09-23 08:14:34,700 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.34 vs. limit=15.0 2024-09-23 08:14:40,664 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=215581.33333333334, ans=0.0 2024-09-23 08:14:42,120 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=215581.33333333334, ans=0.035 2024-09-23 08:14:45,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=215581.33333333334, ans=0.1 2024-09-23 08:14:50,371 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=215581.33333333334, ans=0.2 2024-09-23 08:14:56,380 INFO [train.py:1198] (2/4) Epoch 12, batch 3350, loss[loss=0.287, ctc_loss=0.1964, cr_loss=0.4529, over 17039.00 frames. ], tot_loss[loss=0.2391, ctc_loss=0.164, cr_loss=0.3755, over 3364051.79 frames. ], batch size: 56, lr: 9.92e-03, grad_scale: 16.0 2024-09-23 08:15:18,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=215674.66666666666, ans=0.07 2024-09-23 08:15:54,078 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.291e+02 1.444e+02 1.665e+02 2.877e+02, threshold=2.888e+02, percent-clipped=1.0 2024-09-23 08:15:55,939 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=215768.0, ans=0.125 2024-09-23 08:16:14,339 INFO [train.py:1198] (2/4) Epoch 12, batch 3400, loss[loss=0.2322, ctc_loss=0.1622, cr_loss=0.35, over 17087.00 frames. ], tot_loss[loss=0.2399, ctc_loss=0.1647, cr_loss=0.3759, over 3362088.26 frames. ], batch size: 49, lr: 9.91e-03, grad_scale: 16.0 2024-09-23 08:16:22,498 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=215861.33333333334, ans=0.125 2024-09-23 08:16:36,428 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=215908.0, ans=0.2 2024-09-23 08:16:53,409 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=215954.66666666666, ans=0.1 2024-09-23 08:17:07,845 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=7.34 vs. limit=15.0 2024-09-23 08:17:09,434 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=5.97 vs. limit=10.0 2024-09-23 08:17:11,977 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 08:17:15,099 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=216048.0, ans=0.125 2024-09-23 08:17:23,027 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=216048.0, ans=0.0 2024-09-23 08:17:28,471 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=5.94 vs. limit=10.0 2024-09-23 08:17:32,368 INFO [train.py:1198] (2/4) Epoch 12, batch 3450, loss[loss=0.2522, ctc_loss=0.1728, cr_loss=0.3969, over 16736.00 frames. ], tot_loss[loss=0.2395, ctc_loss=0.1644, cr_loss=0.3754, over 3358824.24 frames. ], batch size: 61, lr: 9.91e-03, grad_scale: 16.0 2024-09-23 08:17:46,776 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=216141.33333333334, ans=0.0 2024-09-23 08:18:03,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=216188.0, ans=0.025 2024-09-23 08:18:30,276 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.153e+02 1.280e+02 1.416e+02 1.630e+02 3.213e+02, threshold=2.832e+02, percent-clipped=1.0 2024-09-23 08:18:50,513 INFO [train.py:1198] (2/4) Epoch 12, batch 3500, loss[loss=0.2409, ctc_loss=0.1656, cr_loss=0.3761, over 17312.00 frames. ], tot_loss[loss=0.2389, ctc_loss=0.1641, cr_loss=0.3738, over 3361538.05 frames. ], batch size: 51, lr: 9.90e-03, grad_scale: 16.0 2024-09-23 08:19:07,804 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=216374.66666666666, ans=0.0 2024-09-23 08:19:08,792 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=6.89 vs. limit=15.0 2024-09-23 08:19:20,283 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=216421.33333333334, ans=0.125 2024-09-23 08:19:23,374 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=216421.33333333334, ans=0.125 2024-09-23 08:19:36,121 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=7.61 vs. limit=15.0 2024-09-23 08:19:46,725 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=216468.0, ans=0.125 2024-09-23 08:19:48,162 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=216468.0, ans=0.125 2024-09-23 08:19:48,269 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=216468.0, ans=0.2 2024-09-23 08:20:08,728 INFO [train.py:1198] (2/4) Epoch 12, batch 3550, loss[loss=0.2391, ctc_loss=0.1645, cr_loss=0.3728, over 17015.00 frames. ], tot_loss[loss=0.2384, ctc_loss=0.1638, cr_loss=0.373, over 3357246.06 frames. ], batch size: 44, lr: 9.90e-03, grad_scale: 16.0 2024-09-23 08:20:19,742 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=216561.33333333334, ans=0.0 2024-09-23 08:20:39,763 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=216654.66666666666, ans=0.125 2024-09-23 08:21:01,512 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=216701.33333333334, ans=0.125 2024-09-23 08:21:05,845 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.122e+02 1.266e+02 1.416e+02 1.619e+02 2.341e+02, threshold=2.832e+02, percent-clipped=0.0 2024-09-23 08:21:06,076 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=216701.33333333334, ans=0.0 2024-09-23 08:21:28,066 INFO [train.py:1198] (2/4) Epoch 12, batch 3600, loss[loss=0.2315, ctc_loss=0.1579, cr_loss=0.3678, over 17291.00 frames. ], tot_loss[loss=0.2379, ctc_loss=0.1634, cr_loss=0.3726, over 3359961.59 frames. ], batch size: 51, lr: 9.89e-03, grad_scale: 32.0 2024-09-23 08:21:44,316 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.19 vs. limit=6.0 2024-09-23 08:21:55,112 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=216841.33333333334, ans=0.0 2024-09-23 08:22:12,539 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=216888.0, ans=0.125 2024-09-23 08:22:47,136 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=217028.0, ans=0.0 2024-09-23 08:22:48,414 INFO [train.py:1198] (2/4) Epoch 12, batch 3650, loss[loss=0.2122, ctc_loss=0.1407, cr_loss=0.3574, over 16976.00 frames. ], tot_loss[loss=0.2374, ctc_loss=0.163, cr_loss=0.3721, over 3362470.49 frames. ], batch size: 42, lr: 9.89e-03, grad_scale: 32.0 2024-09-23 08:23:26,617 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=18.20 vs. limit=22.5 2024-09-23 08:23:50,114 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.266e+02 1.360e+02 1.445e+02 2.351e+02, threshold=2.719e+02, percent-clipped=0.0 2024-09-23 08:23:50,341 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=217168.0, ans=0.05 2024-09-23 08:23:54,589 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.50 vs. limit=6.0 2024-09-23 08:24:10,932 INFO [train.py:1198] (2/4) Epoch 12, batch 3700, loss[loss=0.2732, ctc_loss=0.1935, cr_loss=0.3983, over 14856.00 frames. ], tot_loss[loss=0.2395, ctc_loss=0.1646, cr_loss=0.3749, over 3358720.37 frames. ], batch size: 89, lr: 9.88e-03, grad_scale: 32.0 2024-09-23 08:24:33,639 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=7.43 vs. limit=10.0 2024-09-23 08:24:42,355 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=217354.66666666666, ans=0.0 2024-09-23 08:24:51,569 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=217354.66666666666, ans=0.09899494936611666 2024-09-23 08:24:56,296 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=217401.33333333334, ans=0.125 2024-09-23 08:24:57,845 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=217401.33333333334, ans=0.0 2024-09-23 08:25:01,142 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=217401.33333333334, ans=0.125 2024-09-23 08:25:05,922 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=217401.33333333334, ans=0.125 2024-09-23 08:25:29,359 INFO [train.py:1198] (2/4) Epoch 12, batch 3750, loss[loss=0.1804, ctc_loss=0.1225, cr_loss=0.2898, over 16664.00 frames. ], tot_loss[loss=0.239, ctc_loss=0.1642, cr_loss=0.374, over 3349966.10 frames. ], batch size: 37, lr: 9.88e-03, grad_scale: 32.0 2024-09-23 08:25:51,263 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=217541.33333333334, ans=0.0 2024-09-23 08:26:16,583 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=4.47 vs. limit=15.0 2024-09-23 08:26:17,698 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=217634.66666666666, ans=0.0 2024-09-23 08:26:26,722 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.365e+02 1.477e+02 1.661e+02 2.472e+02, threshold=2.954e+02, percent-clipped=0.0 2024-09-23 08:26:47,064 INFO [train.py:1198] (2/4) Epoch 12, batch 3800, loss[loss=0.258, ctc_loss=0.1809, cr_loss=0.3852, over 15028.00 frames. ], tot_loss[loss=0.2395, ctc_loss=0.1646, cr_loss=0.3742, over 3333076.16 frames. ], batch size: 89, lr: 9.87e-03, grad_scale: 32.0 2024-09-23 08:26:48,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=217728.0, ans=0.125 2024-09-23 08:27:01,552 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=217774.66666666666, ans=0.1 2024-09-23 08:27:06,363 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=217774.66666666666, ans=0.125 2024-09-23 08:27:16,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=217774.66666666666, ans=0.125 2024-09-23 08:27:24,018 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=217821.33333333334, ans=0.2 2024-09-23 08:27:56,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=217914.66666666666, ans=0.125 2024-09-23 08:28:05,900 INFO [train.py:1198] (2/4) Epoch 12, batch 3850, loss[loss=0.2899, ctc_loss=0.208, cr_loss=0.4093, over 11614.00 frames. ], tot_loss[loss=0.2421, ctc_loss=0.167, cr_loss=0.3754, over 3283930.65 frames. ], batch size: 123, lr: 9.87e-03, grad_scale: 32.0 2024-09-23 08:28:07,852 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=217961.33333333334, ans=0.1 2024-09-23 08:28:17,078 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 08:28:29,226 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=218008.0, ans=0.2 2024-09-23 08:28:50,536 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=218101.33333333334, ans=0.025 2024-09-23 08:29:02,106 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.115e+02 1.424e+02 1.630e+02 1.758e+02 2.384e+02, threshold=3.259e+02, percent-clipped=0.0 2024-09-23 08:29:06,979 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=218148.0, ans=0.0 2024-09-23 08:29:11,411 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=218148.0, ans=0.1 2024-09-23 08:29:13,483 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.whiten, num_groups=1, num_channels=192, metric=4.22 vs. limit=12.0 2024-09-23 08:30:07,238 INFO [train.py:1198] (2/4) Epoch 13, batch 0, loss[loss=0.2579, ctc_loss=0.1726, cr_loss=0.4263, over 17026.00 frames. ], tot_loss[loss=0.2579, ctc_loss=0.1726, cr_loss=0.4263, over 17026.00 frames. ], batch size: 44, lr: 9.48e-03, grad_scale: 32.0 2024-09-23 08:30:07,239 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 08:30:16,568 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.5.encoder.layers.1.self_attn_weights, attn_weights_entropy = tensor([2.2670, 4.2804, 3.9537, 4.4592], device='cuda:2') 2024-09-23 08:30:22,743 INFO [train.py:1230] (2/4) Epoch 13, validation: loss=0.04407, ctc_loss=0.04407, cr_loss=7.62e-15, over 944034.00 frames. 2024-09-23 08:30:22,744 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 08:30:26,390 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=218176.0, ans=0.1 2024-09-23 08:30:27,947 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=218176.0, ans=0.1 2024-09-23 08:30:37,575 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=218222.66666666666, ans=0.125 2024-09-23 08:30:56,444 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=218269.33333333334, ans=0.125 2024-09-23 08:31:06,275 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=218269.33333333334, ans=0.1 2024-09-23 08:31:09,598 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=218316.0, ans=0.0 2024-09-23 08:31:14,977 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=19.51 vs. limit=22.5 2024-09-23 08:31:30,292 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=218362.66666666666, ans=0.2 2024-09-23 08:31:43,033 INFO [train.py:1198] (2/4) Epoch 13, batch 50, loss[loss=0.2333, ctc_loss=0.159, cr_loss=0.3712, over 17177.00 frames. ], tot_loss[loss=0.2425, ctc_loss=0.1669, cr_loss=0.3779, over 751336.25 frames. ], batch size: 45, lr: 9.47e-03, grad_scale: 32.0 2024-09-23 08:31:56,196 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=218409.33333333334, ans=0.125 2024-09-23 08:32:26,370 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=218502.66666666666, ans=0.025 2024-09-23 08:32:32,720 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=218549.33333333334, ans=0.125 2024-09-23 08:32:44,011 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=218549.33333333334, ans=0.09899494936611666 2024-09-23 08:32:45,657 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=218549.33333333334, ans=0.125 2024-09-23 08:32:50,417 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=218596.0, ans=0.125 2024-09-23 08:32:51,665 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.113e+02 1.296e+02 1.397e+02 1.549e+02 2.228e+02, threshold=2.794e+02, percent-clipped=0.0 2024-09-23 08:32:53,554 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=218596.0, ans=0.0 2024-09-23 08:33:05,960 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=218596.0, ans=0.0 2024-09-23 08:33:08,627 INFO [train.py:1198] (2/4) Epoch 13, batch 100, loss[loss=0.2327, ctc_loss=0.1597, cr_loss=0.3648, over 17326.00 frames. ], tot_loss[loss=0.2408, ctc_loss=0.1652, cr_loss=0.3779, over 1336486.28 frames. ], batch size: 51, lr: 9.47e-03, grad_scale: 32.0 2024-09-23 08:33:28,040 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=218689.33333333334, ans=0.0 2024-09-23 08:33:32,905 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=218689.33333333334, ans=0.125 2024-09-23 08:33:37,766 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=218689.33333333334, ans=0.125 2024-09-23 08:33:39,473 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=218736.0, ans=0.125 2024-09-23 08:33:41,710 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=10.32 vs. limit=15.0 2024-09-23 08:33:47,566 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=218736.0, ans=0.0 2024-09-23 08:34:23,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=218829.33333333334, ans=0.2 2024-09-23 08:34:28,452 INFO [train.py:1198] (2/4) Epoch 13, batch 150, loss[loss=0.2338, ctc_loss=0.1589, cr_loss=0.3747, over 17139.00 frames. ], tot_loss[loss=0.2388, ctc_loss=0.1637, cr_loss=0.3753, over 1787597.43 frames. ], batch size: 48, lr: 9.46e-03, grad_scale: 32.0 2024-09-23 08:34:57,440 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=218922.66666666666, ans=0.2 2024-09-23 08:35:05,333 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=218969.33333333334, ans=0.125 2024-09-23 08:35:14,911 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=219016.0, ans=0.05 2024-09-23 08:35:25,795 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=219016.0, ans=0.1 2024-09-23 08:35:27,359 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.min_abs, batch_count=219016.0, ans=0.5 2024-09-23 08:35:36,675 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.058e+02 1.277e+02 1.397e+02 1.518e+02 2.217e+02, threshold=2.795e+02, percent-clipped=0.0 2024-09-23 08:35:51,082 INFO [train.py:1198] (2/4) Epoch 13, batch 200, loss[loss=0.215, ctc_loss=0.1451, cr_loss=0.3496, over 17029.00 frames. ], tot_loss[loss=0.2394, ctc_loss=0.164, cr_loss=0.3769, over 2149121.15 frames. ], batch size: 44, lr: 9.46e-03, grad_scale: 32.0 2024-09-23 08:36:13,696 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=219156.0, ans=0.125 2024-09-23 08:36:31,664 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=219202.66666666666, ans=0.125 2024-09-23 08:36:35,010 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=10.77 vs. limit=22.5 2024-09-23 08:36:53,844 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=219249.33333333334, ans=0.1 2024-09-23 08:36:58,701 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 08:36:58,750 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=219296.0, ans=0.125 2024-09-23 08:37:06,647 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=219296.0, ans=0.125 2024-09-23 08:37:16,176 INFO [train.py:1198] (2/4) Epoch 13, batch 250, loss[loss=0.25, ctc_loss=0.1695, cr_loss=0.4026, over 17039.00 frames. ], tot_loss[loss=0.239, ctc_loss=0.1637, cr_loss=0.3767, over 2419126.42 frames. ], batch size: 52, lr: 9.45e-03, grad_scale: 32.0 2024-09-23 08:37:19,736 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=219342.66666666666, ans=0.125 2024-09-23 08:38:23,001 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=219529.33333333334, ans=0.125 2024-09-23 08:38:24,170 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.243e+02 1.365e+02 1.573e+02 3.010e+02, threshold=2.729e+02, percent-clipped=2.0 2024-09-23 08:38:26,047 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=219529.33333333334, ans=0.0 2024-09-23 08:38:38,515 INFO [train.py:1198] (2/4) Epoch 13, batch 300, loss[loss=0.226, ctc_loss=0.1535, cr_loss=0.3626, over 17091.00 frames. ], tot_loss[loss=0.2381, ctc_loss=0.1629, cr_loss=0.3761, over 2629904.04 frames. ], batch size: 43, lr: 9.45e-03, grad_scale: 32.0 2024-09-23 08:39:30,199 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=15.26 vs. limit=22.5 2024-09-23 08:39:43,287 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=7.86 vs. limit=15.0 2024-09-23 08:39:47,312 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=219762.66666666666, ans=0.125 2024-09-23 08:39:48,926 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=219762.66666666666, ans=0.025 2024-09-23 08:39:53,779 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=219762.66666666666, ans=0.125 2024-09-23 08:39:57,041 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=219809.33333333334, ans=0.125 2024-09-23 08:39:58,340 INFO [train.py:1198] (2/4) Epoch 13, batch 350, loss[loss=0.2308, ctc_loss=0.1582, cr_loss=0.363, over 17150.00 frames. ], tot_loss[loss=0.2377, ctc_loss=0.1629, cr_loss=0.3741, over 2777591.56 frames. ], batch size: 45, lr: 9.44e-03, grad_scale: 32.0 2024-09-23 08:39:58,730 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=219809.33333333334, ans=0.125 2024-09-23 08:40:01,757 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=219809.33333333334, ans=0.025 2024-09-23 08:40:34,511 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=219902.66666666666, ans=0.125 2024-09-23 08:41:06,038 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.107e+02 1.335e+02 1.492e+02 1.717e+02 2.357e+02, threshold=2.983e+02, percent-clipped=0.0 2024-09-23 08:41:11,344 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=219996.0, ans=0.125 2024-09-23 08:41:14,353 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=219996.0, ans=0.125 2024-09-23 08:41:20,270 INFO [train.py:1198] (2/4) Epoch 13, batch 400, loss[loss=0.2185, ctc_loss=0.1507, cr_loss=0.3389, over 17078.00 frames. ], tot_loss[loss=0.2385, ctc_loss=0.1635, cr_loss=0.3751, over 2893821.75 frames. ], batch size: 43, lr: 9.44e-03, grad_scale: 32.0 2024-09-23 08:41:25,230 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=220042.66666666666, ans=0.125 2024-09-23 08:41:50,219 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=220089.33333333334, ans=0.0 2024-09-23 08:42:23,742 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=220182.66666666666, ans=0.2 2024-09-23 08:42:26,951 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=220182.66666666666, ans=0.2 2024-09-23 08:42:45,837 INFO [train.py:1198] (2/4) Epoch 13, batch 450, loss[loss=0.2023, ctc_loss=0.1357, cr_loss=0.3328, over 17186.00 frames. ], tot_loss[loss=0.2372, ctc_loss=0.1626, cr_loss=0.3732, over 3000395.12 frames. ], batch size: 41, lr: 9.43e-03, grad_scale: 32.0 2024-09-23 08:42:46,238 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer_ff2.min_abs, batch_count=220276.0, ans=0.1 2024-09-23 08:43:06,181 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=220322.66666666666, ans=0.125 2024-09-23 08:43:06,250 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=220322.66666666666, ans=0.1 2024-09-23 08:43:19,133 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=220369.33333333334, ans=0.0 2024-09-23 08:43:20,870 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=220369.33333333334, ans=0.125 2024-09-23 08:43:53,978 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.092e+02 1.263e+02 1.351e+02 1.525e+02 2.528e+02, threshold=2.701e+02, percent-clipped=0.0 2024-09-23 08:43:57,438 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=220462.66666666666, ans=0.125 2024-09-23 08:44:03,851 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=220462.66666666666, ans=0.05 2024-09-23 08:44:08,347 INFO [train.py:1198] (2/4) Epoch 13, batch 500, loss[loss=0.2665, ctc_loss=0.1834, cr_loss=0.4155, over 17210.00 frames. ], tot_loss[loss=0.2372, ctc_loss=0.1624, cr_loss=0.3737, over 3081957.64 frames. ], batch size: 55, lr: 9.43e-03, grad_scale: 32.0 2024-09-23 08:44:14,298 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.01 vs. limit=6.0 2024-09-23 08:44:31,673 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=15.58 vs. limit=22.5 2024-09-23 08:44:40,861 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=220602.66666666666, ans=0.125 2024-09-23 08:44:44,112 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=220602.66666666666, ans=0.125 2024-09-23 08:45:13,553 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.17 vs. limit=6.0 2024-09-23 08:45:31,233 INFO [train.py:1198] (2/4) Epoch 13, batch 550, loss[loss=0.2437, ctc_loss=0.169, cr_loss=0.3737, over 17151.00 frames. ], tot_loss[loss=0.2374, ctc_loss=0.1627, cr_loss=0.3735, over 3135388.22 frames. ], batch size: 48, lr: 9.42e-03, grad_scale: 32.0 2024-09-23 08:45:48,804 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=220789.33333333334, ans=0.125 2024-09-23 08:46:09,622 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=220836.0, ans=0.1 2024-09-23 08:46:15,965 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=220836.0, ans=0.025 2024-09-23 08:46:35,368 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=220929.33333333334, ans=0.2 2024-09-23 08:46:36,681 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.066e+02 1.262e+02 1.359e+02 1.486e+02 2.281e+02, threshold=2.719e+02, percent-clipped=0.0 2024-09-23 08:46:41,384 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=220929.33333333334, ans=0.0 2024-09-23 08:46:43,056 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 08:46:56,573 INFO [train.py:1198] (2/4) Epoch 13, batch 600, loss[loss=0.2115, ctc_loss=0.1421, cr_loss=0.3466, over 17319.00 frames. ], tot_loss[loss=0.2373, ctc_loss=0.1626, cr_loss=0.3736, over 3180600.00 frames. ], batch size: 51, lr: 9.42e-03, grad_scale: 32.0 2024-09-23 08:47:06,275 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=220976.0, ans=0.025 2024-09-23 08:47:17,484 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=221022.66666666666, ans=0.0 2024-09-23 08:47:17,602 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=221022.66666666666, ans=0.1 2024-09-23 08:47:55,327 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=221116.0, ans=0.125 2024-09-23 08:48:18,959 INFO [train.py:1198] (2/4) Epoch 13, batch 650, loss[loss=0.2069, ctc_loss=0.144, cr_loss=0.3143, over 17266.00 frames. ], tot_loss[loss=0.237, ctc_loss=0.1624, cr_loss=0.3734, over 3217335.40 frames. ], batch size: 42, lr: 9.41e-03, grad_scale: 32.0 2024-09-23 08:48:37,048 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=221256.0, ans=0.1 2024-09-23 08:49:08,328 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=221349.33333333334, ans=0.125 2024-09-23 08:49:23,122 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.56 vs. limit=6.0 2024-09-23 08:49:23,893 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.284e+02 1.397e+02 1.593e+02 2.300e+02, threshold=2.794e+02, percent-clipped=0.0 2024-09-23 08:49:34,578 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=7.63 vs. limit=15.0 2024-09-23 08:49:38,325 INFO [train.py:1198] (2/4) Epoch 13, batch 700, loss[loss=0.1998, ctc_loss=0.1334, cr_loss=0.3323, over 16258.00 frames. ], tot_loss[loss=0.2365, ctc_loss=0.1619, cr_loss=0.373, over 3253880.43 frames. ], batch size: 36, lr: 9.41e-03, grad_scale: 32.0 2024-09-23 08:49:45,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=221442.66666666666, ans=0.125 2024-09-23 08:49:57,357 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=221489.33333333334, ans=0.125 2024-09-23 08:50:57,035 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=221629.33333333334, ans=0.1 2024-09-23 08:50:59,837 INFO [train.py:1198] (2/4) Epoch 13, batch 750, loss[loss=0.2172, ctc_loss=0.1467, cr_loss=0.3523, over 16952.00 frames. ], tot_loss[loss=0.2368, ctc_loss=0.162, cr_loss=0.374, over 3285781.55 frames. ], batch size: 42, lr: 9.40e-03, grad_scale: 16.0 2024-09-23 08:51:03,311 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=221676.0, ans=0.125 2024-09-23 08:51:19,391 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=221722.66666666666, ans=0.125 2024-09-23 08:51:20,889 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=221722.66666666666, ans=0.125 2024-09-23 08:52:02,708 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=221816.0, ans=0.0 2024-09-23 08:52:11,979 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.098e+02 1.258e+02 1.365e+02 1.480e+02 2.047e+02, threshold=2.729e+02, percent-clipped=0.0 2024-09-23 08:52:15,494 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=221862.66666666666, ans=0.1 2024-09-23 08:52:24,631 INFO [train.py:1198] (2/4) Epoch 13, batch 800, loss[loss=0.2042, ctc_loss=0.1363, cr_loss=0.3395, over 17022.00 frames. ], tot_loss[loss=0.2384, ctc_loss=0.1632, cr_loss=0.3757, over 3296242.25 frames. ], batch size: 44, lr: 9.40e-03, grad_scale: 32.0 2024-09-23 08:52:44,167 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=221956.0, ans=0.125 2024-09-23 08:52:47,213 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=221956.0, ans=0.125 2024-09-23 08:52:52,090 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 08:52:59,492 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 08:53:04,415 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=222002.66666666666, ans=0.0 2024-09-23 08:53:15,456 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=222049.33333333334, ans=0.1 2024-09-23 08:53:22,313 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=14.36 vs. limit=15.0 2024-09-23 08:53:36,192 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=13.18 vs. limit=22.5 2024-09-23 08:53:39,001 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=222096.0, ans=0.125 2024-09-23 08:53:45,804 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=222142.66666666666, ans=0.125 2024-09-23 08:53:47,085 INFO [train.py:1198] (2/4) Epoch 13, batch 850, loss[loss=0.2598, ctc_loss=0.1756, cr_loss=0.4209, over 16881.00 frames. ], tot_loss[loss=0.2393, ctc_loss=0.164, cr_loss=0.3764, over 3303901.08 frames. ], batch size: 58, lr: 9.39e-03, grad_scale: 32.0 2024-09-23 08:53:52,473 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=9.93 vs. limit=15.0 2024-09-23 08:53:55,161 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=222142.66666666666, ans=0.0 2024-09-23 08:54:19,318 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=222236.0, ans=0.0 2024-09-23 08:54:40,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=222282.66666666666, ans=0.125 2024-09-23 08:54:40,211 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=222282.66666666666, ans=0.0 2024-09-23 08:54:52,920 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=222329.33333333334, ans=0.1 2024-09-23 08:54:54,219 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.309e+02 1.415e+02 1.608e+02 2.192e+02, threshold=2.830e+02, percent-clipped=0.0 2024-09-23 08:54:58,279 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=5.13 vs. limit=10.0 2024-09-23 08:55:06,969 INFO [train.py:1198] (2/4) Epoch 13, batch 900, loss[loss=0.2307, ctc_loss=0.1578, cr_loss=0.3645, over 17063.00 frames. ], tot_loss[loss=0.2381, ctc_loss=0.1632, cr_loss=0.3748, over 3316087.88 frames. ], batch size: 46, lr: 9.39e-03, grad_scale: 32.0 2024-09-23 08:55:16,618 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=222376.0, ans=0.0 2024-09-23 08:55:16,743 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=222376.0, ans=0.04949747468305833 2024-09-23 08:55:23,092 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.55 vs. limit=15.0 2024-09-23 08:55:28,234 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=7.05 vs. limit=8.0 2024-09-23 08:55:57,884 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=222516.0, ans=0.1 2024-09-23 08:56:31,997 INFO [train.py:1198] (2/4) Epoch 13, batch 950, loss[loss=0.2591, ctc_loss=0.1756, cr_loss=0.4175, over 17215.00 frames. ], tot_loss[loss=0.2373, ctc_loss=0.1623, cr_loss=0.3752, over 3337513.38 frames. ], batch size: 55, lr: 9.38e-03, grad_scale: 32.0 2024-09-23 08:57:10,336 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=222702.66666666666, ans=0.1 2024-09-23 08:57:41,764 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.336e+02 1.446e+02 1.624e+02 2.624e+02, threshold=2.892e+02, percent-clipped=0.0 2024-09-23 08:57:45,846 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=3.71 vs. limit=15.0 2024-09-23 08:57:57,067 INFO [train.py:1198] (2/4) Epoch 13, batch 1000, loss[loss=0.2497, ctc_loss=0.1693, cr_loss=0.4022, over 16764.00 frames. ], tot_loss[loss=0.2382, ctc_loss=0.163, cr_loss=0.376, over 3337116.76 frames. ], batch size: 61, lr: 9.38e-03, grad_scale: 32.0 2024-09-23 08:57:57,458 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=222842.66666666666, ans=0.0 2024-09-23 08:58:25,543 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=222889.33333333334, ans=0.1 2024-09-23 08:58:27,207 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=222936.0, ans=0.1 2024-09-23 08:58:57,619 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=222982.66666666666, ans=0.125 2024-09-23 08:59:13,848 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=223029.33333333334, ans=0.125 2024-09-23 08:59:16,889 INFO [train.py:1198] (2/4) Epoch 13, batch 1050, loss[loss=0.279, ctc_loss=0.1983, cr_loss=0.4036, over 14984.00 frames. ], tot_loss[loss=0.2374, ctc_loss=0.1623, cr_loss=0.3755, over 3347411.53 frames. ], batch size: 89, lr: 9.37e-03, grad_scale: 32.0 2024-09-23 08:59:17,417 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=7.14 vs. limit=15.0 2024-09-23 09:00:26,978 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.122e+02 1.277e+02 1.397e+02 1.576e+02 3.836e+02, threshold=2.794e+02, percent-clipped=1.0 2024-09-23 09:00:32,060 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=223262.66666666666, ans=0.2 2024-09-23 09:00:39,436 INFO [train.py:1198] (2/4) Epoch 13, batch 1100, loss[loss=0.2767, ctc_loss=0.1989, cr_loss=0.3893, over 17044.00 frames. ], tot_loss[loss=0.2374, ctc_loss=0.1624, cr_loss=0.3751, over 3352860.42 frames. ], batch size: 52, lr: 9.37e-03, grad_scale: 32.0 2024-09-23 09:01:03,428 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=223356.0, ans=0.0 2024-09-23 09:01:03,487 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=223356.0, ans=0.125 2024-09-23 09:01:06,527 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=223356.0, ans=0.125 2024-09-23 09:01:19,325 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=223402.66666666666, ans=0.2 2024-09-23 09:01:53,714 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=223496.0, ans=0.1 2024-09-23 09:02:01,296 INFO [train.py:1198] (2/4) Epoch 13, batch 1150, loss[loss=0.2144, ctc_loss=0.1464, cr_loss=0.3403, over 17099.00 frames. ], tot_loss[loss=0.2358, ctc_loss=0.1611, cr_loss=0.3736, over 3356316.63 frames. ], batch size: 43, lr: 9.37e-03, grad_scale: 32.0 2024-09-23 09:02:12,575 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=223542.66666666666, ans=0.2 2024-09-23 09:02:47,103 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=19.20 vs. limit=22.5 2024-09-23 09:02:52,442 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=223682.66666666666, ans=0.025 2024-09-23 09:03:07,036 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=223729.33333333334, ans=0.0 2024-09-23 09:03:09,132 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.11 vs. limit=15.0 2024-09-23 09:03:11,339 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.252e+02 1.365e+02 1.487e+02 2.591e+02, threshold=2.731e+02, percent-clipped=0.0 2024-09-23 09:03:16,752 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=10.14 vs. limit=15.0 2024-09-23 09:03:19,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=223729.33333333334, ans=0.07 2024-09-23 09:03:23,905 INFO [train.py:1198] (2/4) Epoch 13, batch 1200, loss[loss=0.2507, ctc_loss=0.1751, cr_loss=0.3779, over 16970.00 frames. ], tot_loss[loss=0.2367, ctc_loss=0.1617, cr_loss=0.3749, over 3360562.68 frames. ], batch size: 53, lr: 9.36e-03, grad_scale: 32.0 2024-09-23 09:03:43,108 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=223822.66666666666, ans=0.2 2024-09-23 09:03:54,486 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 09:04:07,017 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=223869.33333333334, ans=0.2 2024-09-23 09:04:16,698 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=223916.0, ans=0.125 2024-09-23 09:04:46,170 INFO [train.py:1198] (2/4) Epoch 13, batch 1250, loss[loss=0.2636, ctc_loss=0.183, cr_loss=0.4028, over 17043.00 frames. ], tot_loss[loss=0.2376, ctc_loss=0.1625, cr_loss=0.3758, over 3350523.65 frames. ], batch size: 52, lr: 9.36e-03, grad_scale: 32.0 2024-09-23 09:05:19,310 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=224102.66666666666, ans=0.125 2024-09-23 09:05:55,490 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.099e+02 1.315e+02 1.381e+02 1.503e+02 2.464e+02, threshold=2.763e+02, percent-clipped=0.0 2024-09-23 09:06:08,064 INFO [train.py:1198] (2/4) Epoch 13, batch 1300, loss[loss=0.2304, ctc_loss=0.1565, cr_loss=0.3691, over 17012.00 frames. ], tot_loss[loss=0.2367, ctc_loss=0.1618, cr_loss=0.3742, over 3350840.52 frames. ], batch size: 51, lr: 9.35e-03, grad_scale: 32.0 2024-09-23 09:06:13,248 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=224242.66666666666, ans=0.1 2024-09-23 09:06:14,674 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=224242.66666666666, ans=0.1 2024-09-23 09:06:16,259 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=224242.66666666666, ans=0.125 2024-09-23 09:06:47,935 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=224336.0, ans=0.0 2024-09-23 09:06:50,161 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.74 vs. limit=12.0 2024-09-23 09:07:32,574 INFO [train.py:1198] (2/4) Epoch 13, batch 1350, loss[loss=0.306, ctc_loss=0.2237, cr_loss=0.4113, over 11839.00 frames. ], tot_loss[loss=0.236, ctc_loss=0.1613, cr_loss=0.3732, over 3351428.45 frames. ], batch size: 123, lr: 9.35e-03, grad_scale: 32.0 2024-09-23 09:08:05,528 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=224569.33333333334, ans=0.0 2024-09-23 09:08:07,247 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=224569.33333333334, ans=0.125 2024-09-23 09:08:07,626 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.00 vs. limit=10.0 2024-09-23 09:08:40,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=224662.66666666666, ans=0.0 2024-09-23 09:08:41,709 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.260e+02 1.380e+02 1.508e+02 2.232e+02, threshold=2.759e+02, percent-clipped=0.0 2024-09-23 09:08:50,070 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=224662.66666666666, ans=0.125 2024-09-23 09:08:54,697 INFO [train.py:1198] (2/4) Epoch 13, batch 1400, loss[loss=0.3127, ctc_loss=0.2227, cr_loss=0.4501, over 11921.00 frames. ], tot_loss[loss=0.236, ctc_loss=0.1615, cr_loss=0.3725, over 3340054.12 frames. ], batch size: 123, lr: 9.34e-03, grad_scale: 32.0 2024-09-23 09:09:06,116 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=224709.33333333334, ans=0.125 2024-09-23 09:09:22,051 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=224756.0, ans=0.1 2024-09-23 09:09:47,604 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=224849.33333333334, ans=0.2 2024-09-23 09:09:55,966 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=10.10 vs. limit=15.0 2024-09-23 09:10:16,111 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=224942.66666666666, ans=0.025 2024-09-23 09:10:17,264 INFO [train.py:1198] (2/4) Epoch 13, batch 1450, loss[loss=0.21, ctc_loss=0.1415, cr_loss=0.3425, over 16360.00 frames. ], tot_loss[loss=0.2368, ctc_loss=0.162, cr_loss=0.3739, over 3341373.89 frames. ], batch size: 36, lr: 9.34e-03, grad_scale: 32.0 2024-09-23 09:10:43,210 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=224989.33333333334, ans=0.1 2024-09-23 09:10:47,040 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=19.40 vs. limit=22.5 2024-09-23 09:11:01,596 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=15.05 vs. limit=22.5 2024-09-23 09:11:29,649 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.289e+02 1.416e+02 1.573e+02 2.089e+02, threshold=2.832e+02, percent-clipped=0.0 2024-09-23 09:11:33,835 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.98 vs. limit=10.0 2024-09-23 09:11:42,311 INFO [train.py:1198] (2/4) Epoch 13, batch 1500, loss[loss=0.2362, ctc_loss=0.1656, cr_loss=0.3527, over 16952.00 frames. ], tot_loss[loss=0.2359, ctc_loss=0.1615, cr_loss=0.3722, over 3346996.13 frames. ], batch size: 58, lr: 9.33e-03, grad_scale: 32.0 2024-09-23 09:11:55,197 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=225176.0, ans=0.0 2024-09-23 09:12:16,036 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=225269.33333333334, ans=0.1 2024-09-23 09:12:32,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=225316.0, ans=0.1 2024-09-23 09:12:54,069 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=225362.66666666666, ans=0.1 2024-09-23 09:13:05,142 INFO [train.py:1198] (2/4) Epoch 13, batch 1550, loss[loss=0.2434, ctc_loss=0.1708, cr_loss=0.3632, over 17211.00 frames. ], tot_loss[loss=0.2352, ctc_loss=0.161, cr_loss=0.3711, over 3352025.39 frames. ], batch size: 50, lr: 9.33e-03, grad_scale: 32.0 2024-09-23 09:13:08,807 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=225409.33333333334, ans=0.0 2024-09-23 09:13:10,396 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=225409.33333333334, ans=0.1 2024-09-23 09:13:16,741 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=225409.33333333334, ans=0.0 2024-09-23 09:13:29,934 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.88 vs. limit=15.0 2024-09-23 09:13:34,597 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=225456.0, ans=0.0 2024-09-23 09:13:36,024 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=225502.66666666666, ans=0.2 2024-09-23 09:13:42,436 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=225502.66666666666, ans=0.1 2024-09-23 09:14:10,055 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=11.10 vs. limit=22.5 2024-09-23 09:14:12,644 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.248e+02 1.360e+02 1.586e+02 2.535e+02, threshold=2.720e+02, percent-clipped=0.0 2024-09-23 09:14:25,419 INFO [train.py:1198] (2/4) Epoch 13, batch 1600, loss[loss=0.2523, ctc_loss=0.1706, cr_loss=0.4085, over 16781.00 frames. ], tot_loss[loss=0.2362, ctc_loss=0.1618, cr_loss=0.3719, over 3334845.35 frames. ], batch size: 61, lr: 9.32e-03, grad_scale: 32.0 2024-09-23 09:15:26,660 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.00 vs. limit=10.0 2024-09-23 09:15:45,461 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=225829.33333333334, ans=0.0 2024-09-23 09:15:48,383 INFO [train.py:1198] (2/4) Epoch 13, batch 1650, loss[loss=0.2302, ctc_loss=0.1557, cr_loss=0.3724, over 17161.00 frames. ], tot_loss[loss=0.2347, ctc_loss=0.1606, cr_loss=0.3705, over 3339687.38 frames. ], batch size: 45, lr: 9.32e-03, grad_scale: 32.0 2024-09-23 09:16:12,947 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=225922.66666666666, ans=0.1 2024-09-23 09:16:30,675 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 09:16:49,611 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=226016.0, ans=0.0 2024-09-23 09:17:00,540 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.122e+02 1.262e+02 1.363e+02 1.503e+02 2.167e+02, threshold=2.727e+02, percent-clipped=0.0 2024-09-23 09:17:13,266 INFO [train.py:1198] (2/4) Epoch 13, batch 1700, loss[loss=0.2673, ctc_loss=0.1824, cr_loss=0.4248, over 16997.00 frames. ], tot_loss[loss=0.235, ctc_loss=0.1608, cr_loss=0.3709, over 3349971.93 frames. ], batch size: 56, lr: 9.31e-03, grad_scale: 32.0 2024-09-23 09:17:57,752 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=226202.66666666666, ans=0.125 2024-09-23 09:18:01,224 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.26 vs. limit=22.5 2024-09-23 09:18:07,363 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=226249.33333333334, ans=0.1 2024-09-23 09:18:13,688 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=226249.33333333334, ans=0.125 2024-09-23 09:18:35,886 INFO [train.py:1198] (2/4) Epoch 13, batch 1750, loss[loss=0.1843, ctc_loss=0.1204, cr_loss=0.3194, over 16963.00 frames. ], tot_loss[loss=0.2358, ctc_loss=0.1614, cr_loss=0.3724, over 3347099.28 frames. ], batch size: 42, lr: 9.31e-03, grad_scale: 32.0 2024-09-23 09:18:58,303 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=226389.33333333334, ans=0.125 2024-09-23 09:18:59,703 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=226389.33333333334, ans=0.0 2024-09-23 09:19:01,304 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=226389.33333333334, ans=0.0 2024-09-23 09:19:08,391 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.07 vs. limit=15.0 2024-09-23 09:19:22,799 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.48 vs. limit=15.0 2024-09-23 09:19:39,636 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=226529.33333333334, ans=0.025 2024-09-23 09:19:42,591 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.092e+02 1.255e+02 1.356e+02 1.518e+02 2.113e+02, threshold=2.712e+02, percent-clipped=0.0 2024-09-23 09:19:55,546 INFO [train.py:1198] (2/4) Epoch 13, batch 1800, loss[loss=0.2086, ctc_loss=0.141, cr_loss=0.3381, over 17203.00 frames. ], tot_loss[loss=0.2359, ctc_loss=0.1614, cr_loss=0.3728, over 3354530.80 frames. ], batch size: 47, lr: 9.30e-03, grad_scale: 32.0 2024-09-23 09:20:04,532 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.84 vs. limit=12.0 2024-09-23 09:20:05,376 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=226576.0, ans=0.025 2024-09-23 09:20:05,494 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=226576.0, ans=0.2 2024-09-23 09:20:07,795 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=14.76 vs. limit=22.5 2024-09-23 09:20:12,670 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=226622.66666666666, ans=0.07 2024-09-23 09:20:23,752 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=226622.66666666666, ans=0.0 2024-09-23 09:20:23,793 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=226622.66666666666, ans=0.2 2024-09-23 09:20:25,291 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=226622.66666666666, ans=0.125 2024-09-23 09:20:35,117 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 09:20:55,566 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=226716.0, ans=0.0 2024-09-23 09:20:58,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=226716.0, ans=0.0 2024-09-23 09:21:00,787 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=512, metric=15.11 vs. limit=22.5 2024-09-23 09:21:05,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=226762.66666666666, ans=0.2 2024-09-23 09:21:22,742 INFO [train.py:1198] (2/4) Epoch 13, batch 1850, loss[loss=0.248, ctc_loss=0.1719, cr_loss=0.3803, over 17310.00 frames. ], tot_loss[loss=0.2367, ctc_loss=0.162, cr_loss=0.3732, over 3349418.38 frames. ], batch size: 51, lr: 9.30e-03, grad_scale: 32.0 2024-09-23 09:21:26,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=226809.33333333334, ans=0.125 2024-09-23 09:21:29,768 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.20 vs. limit=22.5 2024-09-23 09:21:32,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=226809.33333333334, ans=0.125 2024-09-23 09:21:45,308 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=226856.0, ans=0.125 2024-09-23 09:21:56,626 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=226902.66666666666, ans=0.125 2024-09-23 09:22:04,973 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.94 vs. limit=22.5 2024-09-23 09:22:22,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=226949.33333333334, ans=0.125 2024-09-23 09:22:29,439 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.038e+02 1.309e+02 1.425e+02 1.783e+02 3.519e+02, threshold=2.851e+02, percent-clipped=1.0 2024-09-23 09:22:33,867 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.max_abs, batch_count=226996.0, ans=10.0 2024-09-23 09:22:44,629 INFO [train.py:1198] (2/4) Epoch 13, batch 1900, loss[loss=0.2246, ctc_loss=0.1521, cr_loss=0.3625, over 16960.00 frames. ], tot_loss[loss=0.2351, ctc_loss=0.1608, cr_loss=0.3711, over 3350251.20 frames. ], batch size: 42, lr: 9.29e-03, grad_scale: 32.0 2024-09-23 09:23:21,966 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=227136.0, ans=0.125 2024-09-23 09:23:59,601 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=227229.33333333334, ans=0.125 2024-09-23 09:24:04,047 INFO [train.py:1198] (2/4) Epoch 13, batch 1950, loss[loss=0.2121, ctc_loss=0.143, cr_loss=0.3452, over 17260.00 frames. ], tot_loss[loss=0.2357, ctc_loss=0.1613, cr_loss=0.3719, over 3349610.89 frames. ], batch size: 44, lr: 9.29e-03, grad_scale: 32.0 2024-09-23 09:24:27,251 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.72 vs. limit=15.0 2024-09-23 09:24:33,089 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=227322.66666666666, ans=0.125 2024-09-23 09:24:40,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=227369.33333333334, ans=0.0 2024-09-23 09:25:01,823 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=227416.0, ans=0.0 2024-09-23 09:25:07,342 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=227416.0, ans=0.125 2024-09-23 09:25:12,048 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=227462.66666666666, ans=0.125 2024-09-23 09:25:13,356 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.055e+02 1.325e+02 1.427e+02 1.553e+02 2.368e+02, threshold=2.853e+02, percent-clipped=0.0 2024-09-23 09:25:25,898 INFO [train.py:1198] (2/4) Epoch 13, batch 2000, loss[loss=0.2268, ctc_loss=0.1502, cr_loss=0.3829, over 16714.00 frames. ], tot_loss[loss=0.2353, ctc_loss=0.1609, cr_loss=0.372, over 3355577.44 frames. ], batch size: 37, lr: 9.29e-03, grad_scale: 32.0 2024-09-23 09:25:49,970 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=227556.0, ans=0.1 2024-09-23 09:26:51,000 INFO [train.py:1198] (2/4) Epoch 13, batch 2050, loss[loss=0.2046, ctc_loss=0.138, cr_loss=0.3327, over 17097.00 frames. ], tot_loss[loss=0.235, ctc_loss=0.1606, cr_loss=0.3719, over 3366950.39 frames. ], batch size: 43, lr: 9.28e-03, grad_scale: 16.0 2024-09-23 09:26:59,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=227742.66666666666, ans=0.05 2024-09-23 09:27:11,031 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.17 vs. limit=15.0 2024-09-23 09:27:12,794 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=3.39 vs. limit=15.0 2024-09-23 09:27:27,835 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=227836.0, ans=0.125 2024-09-23 09:27:38,977 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=18.60 vs. limit=22.5 2024-09-23 09:27:40,097 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=227882.66666666666, ans=0.025 2024-09-23 09:28:01,913 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.087e+02 1.246e+02 1.339e+02 1.445e+02 2.585e+02, threshold=2.679e+02, percent-clipped=0.0 2024-09-23 09:28:13,095 INFO [train.py:1198] (2/4) Epoch 13, batch 2100, loss[loss=0.2073, ctc_loss=0.1397, cr_loss=0.338, over 17053.00 frames. ], tot_loss[loss=0.2367, ctc_loss=0.1617, cr_loss=0.3747, over 3373259.68 frames. ], batch size: 39, lr: 9.28e-03, grad_scale: 16.0 2024-09-23 09:28:15,019 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=227976.0, ans=0.2 2024-09-23 09:28:44,377 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=3.70 vs. limit=15.0 2024-09-23 09:28:47,080 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=228069.33333333334, ans=0.2 2024-09-23 09:28:50,135 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=228069.33333333334, ans=0.125 2024-09-23 09:29:05,243 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=13.53 vs. limit=22.5 2024-09-23 09:29:12,313 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=228116.0, ans=0.125 2024-09-23 09:29:31,871 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.64 vs. limit=15.0 2024-09-23 09:29:32,532 INFO [train.py:1198] (2/4) Epoch 13, batch 2150, loss[loss=0.2326, ctc_loss=0.1591, cr_loss=0.3673, over 17077.00 frames. ], tot_loss[loss=0.2363, ctc_loss=0.1615, cr_loss=0.3742, over 3365582.20 frames. ], batch size: 43, lr: 9.27e-03, grad_scale: 16.0 2024-09-23 09:29:32,862 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=228209.33333333334, ans=0.125 2024-09-23 09:29:34,470 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=228209.33333333334, ans=0.125 2024-09-23 09:29:34,470 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=228209.33333333334, ans=0.0 2024-09-23 09:29:42,636 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 09:30:06,469 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=228302.66666666666, ans=0.125 2024-09-23 09:30:09,512 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=228302.66666666666, ans=0.1 2024-09-23 09:30:18,392 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=228302.66666666666, ans=0.0 2024-09-23 09:30:19,849 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=228302.66666666666, ans=0.125 2024-09-23 09:30:40,947 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=228396.0, ans=0.0 2024-09-23 09:30:43,654 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.254e+02 1.312e+02 1.458e+02 2.181e+02, threshold=2.624e+02, percent-clipped=0.0 2024-09-23 09:30:45,551 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=228396.0, ans=0.125 2024-09-23 09:30:54,885 INFO [train.py:1198] (2/4) Epoch 13, batch 2200, loss[loss=0.2175, ctc_loss=0.1463, cr_loss=0.356, over 17212.00 frames. ], tot_loss[loss=0.2342, ctc_loss=0.1598, cr_loss=0.372, over 3372241.27 frames. ], batch size: 47, lr: 9.27e-03, grad_scale: 16.0 2024-09-23 09:31:30,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=228536.0, ans=0.125 2024-09-23 09:31:46,216 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 09:32:09,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.min_positive, batch_count=228629.33333333334, ans=0.05 2024-09-23 09:32:19,399 INFO [train.py:1198] (2/4) Epoch 13, batch 2250, loss[loss=0.2074, ctc_loss=0.1383, cr_loss=0.3458, over 17004.00 frames. ], tot_loss[loss=0.2352, ctc_loss=0.1606, cr_loss=0.373, over 3373920.01 frames. ], batch size: 44, lr: 9.26e-03, grad_scale: 16.0 2024-09-23 09:32:19,822 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 09:32:24,959 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module1.whiten, num_groups=1, num_channels=512, metric=6.41 vs. limit=15.0 2024-09-23 09:32:29,692 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=9.34 vs. limit=15.0 2024-09-23 09:32:34,842 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=228676.0, ans=0.025 2024-09-23 09:32:36,262 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=228722.66666666666, ans=0.1 2024-09-23 09:32:45,874 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=228722.66666666666, ans=0.1 2024-09-23 09:32:55,437 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=228769.33333333334, ans=0.0 2024-09-23 09:33:02,494 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.95 vs. limit=15.0 2024-09-23 09:33:20,872 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=228816.0, ans=0.0 2024-09-23 09:33:22,433 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=228816.0, ans=0.125 2024-09-23 09:33:28,947 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=228862.66666666666, ans=0.125 2024-09-23 09:33:30,214 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.142e+02 1.274e+02 1.381e+02 1.487e+02 1.972e+02, threshold=2.763e+02, percent-clipped=0.0 2024-09-23 09:33:32,065 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=228862.66666666666, ans=0.025 2024-09-23 09:33:38,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=228862.66666666666, ans=0.125 2024-09-23 09:33:41,417 INFO [train.py:1198] (2/4) Epoch 13, batch 2300, loss[loss=0.2542, ctc_loss=0.1735, cr_loss=0.4035, over 16991.00 frames. ], tot_loss[loss=0.2364, ctc_loss=0.1615, cr_loss=0.3747, over 3371297.30 frames. ], batch size: 56, lr: 9.26e-03, grad_scale: 16.0 2024-09-23 09:33:46,517 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=228909.33333333334, ans=0.0 2024-09-23 09:33:54,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=228909.33333333334, ans=0.04949747468305833 2024-09-23 09:33:54,473 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=228909.33333333334, ans=0.07 2024-09-23 09:33:56,250 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=228956.0, ans=0.2 2024-09-23 09:34:03,012 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 09:34:09,803 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=7.90 vs. limit=15.0 2024-09-23 09:34:20,510 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=229002.66666666666, ans=0.0 2024-09-23 09:34:42,674 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=229049.33333333334, ans=0.2 2024-09-23 09:35:01,960 INFO [train.py:1198] (2/4) Epoch 13, batch 2350, loss[loss=0.2403, ctc_loss=0.1645, cr_loss=0.3789, over 17218.00 frames. ], tot_loss[loss=0.235, ctc_loss=0.1603, cr_loss=0.3732, over 3380036.25 frames. ], batch size: 47, lr: 9.25e-03, grad_scale: 16.0 2024-09-23 09:35:02,194 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=229142.66666666666, ans=0.0 2024-09-23 09:35:12,821 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=229142.66666666666, ans=0.0 2024-09-23 09:35:23,018 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=7.67 vs. limit=15.0 2024-09-23 09:35:33,575 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=229189.33333333334, ans=0.1 2024-09-23 09:35:38,424 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=229236.0, ans=0.125 2024-09-23 09:35:41,608 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 09:35:53,416 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.60 vs. limit=15.0 2024-09-23 09:36:00,722 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=229282.66666666666, ans=0.0 2024-09-23 09:36:00,763 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=229282.66666666666, ans=0.1 2024-09-23 09:36:15,725 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.173e+02 1.316e+02 1.425e+02 1.605e+02 2.479e+02, threshold=2.851e+02, percent-clipped=0.0 2024-09-23 09:36:19,345 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=229329.33333333334, ans=0.125 2024-09-23 09:36:22,629 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=229329.33333333334, ans=0.05 2024-09-23 09:36:27,010 INFO [train.py:1198] (2/4) Epoch 13, batch 2400, loss[loss=0.2334, ctc_loss=0.1582, cr_loss=0.3759, over 17211.00 frames. ], tot_loss[loss=0.2349, ctc_loss=0.1602, cr_loss=0.3734, over 3380557.45 frames. ], batch size: 47, lr: 9.25e-03, grad_scale: 32.0 2024-09-23 09:37:02,507 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=229469.33333333334, ans=0.125 2024-09-23 09:37:07,270 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=229469.33333333334, ans=0.125 2024-09-23 09:37:10,344 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=229469.33333333334, ans=0.125 2024-09-23 09:37:12,053 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=229469.33333333334, ans=0.0 2024-09-23 09:37:18,355 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=229516.0, ans=0.1 2024-09-23 09:37:36,778 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=229562.66666666666, ans=0.1 2024-09-23 09:37:49,397 INFO [train.py:1198] (2/4) Epoch 13, batch 2450, loss[loss=0.2711, ctc_loss=0.1973, cr_loss=0.3694, over 11965.00 frames. ], tot_loss[loss=0.2341, ctc_loss=0.1597, cr_loss=0.3722, over 3373889.26 frames. ], batch size: 123, lr: 9.24e-03, grad_scale: 32.0 2024-09-23 09:38:37,595 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=229749.33333333334, ans=0.125 2024-09-23 09:38:49,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=229749.33333333334, ans=0.125 2024-09-23 09:38:58,204 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.060e+02 1.292e+02 1.402e+02 1.572e+02 2.224e+02, threshold=2.803e+02, percent-clipped=0.0 2024-09-23 09:39:06,913 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.87 vs. limit=15.0 2024-09-23 09:39:09,495 INFO [train.py:1198] (2/4) Epoch 13, batch 2500, loss[loss=0.2172, ctc_loss=0.1484, cr_loss=0.3442, over 17319.00 frames. ], tot_loss[loss=0.2342, ctc_loss=0.1597, cr_loss=0.3729, over 3382172.16 frames. ], batch size: 46, lr: 9.24e-03, grad_scale: 32.0 2024-09-23 09:39:31,034 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=229889.33333333334, ans=0.0 2024-09-23 09:39:49,277 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=19.05 vs. limit=22.5 2024-09-23 09:39:55,021 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=229936.0, ans=0.1 2024-09-23 09:39:59,766 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=229982.66666666666, ans=0.125 2024-09-23 09:40:13,450 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=229982.66666666666, ans=0.1 2024-09-23 09:40:18,112 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=230029.33333333334, ans=0.0 2024-09-23 09:40:24,469 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=230029.33333333334, ans=0.125 2024-09-23 09:40:32,032 INFO [train.py:1198] (2/4) Epoch 13, batch 2550, loss[loss=0.2271, ctc_loss=0.1531, cr_loss=0.3702, over 17064.00 frames. ], tot_loss[loss=0.2342, ctc_loss=0.1598, cr_loss=0.3721, over 3383238.97 frames. ], batch size: 46, lr: 9.23e-03, grad_scale: 32.0 2024-09-23 09:40:33,875 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=230076.0, ans=0.0 2024-09-23 09:40:57,817 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=230122.66666666666, ans=0.1 2024-09-23 09:41:22,307 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=230169.33333333334, ans=0.125 2024-09-23 09:41:24,379 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.37 vs. limit=6.0 2024-09-23 09:41:36,454 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=230216.0, ans=0.025 2024-09-23 09:41:45,893 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.110e+02 1.282e+02 1.430e+02 1.589e+02 2.134e+02, threshold=2.861e+02, percent-clipped=0.0 2024-09-23 09:41:51,118 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=230262.66666666666, ans=0.04949747468305833 2024-09-23 09:41:57,175 INFO [train.py:1198] (2/4) Epoch 13, batch 2600, loss[loss=0.2586, ctc_loss=0.1756, cr_loss=0.4146, over 17314.00 frames. ], tot_loss[loss=0.2345, ctc_loss=0.16, cr_loss=0.3726, over 3380555.97 frames. ], batch size: 51, lr: 9.23e-03, grad_scale: 32.0 2024-09-23 09:42:06,992 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=230309.33333333334, ans=0.125 2024-09-23 09:42:36,888 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=230402.66666666666, ans=0.125 2024-09-23 09:43:13,651 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=230496.0, ans=0.125 2024-09-23 09:43:13,799 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=230496.0, ans=0.1 2024-09-23 09:43:20,132 INFO [train.py:1198] (2/4) Epoch 13, batch 2650, loss[loss=0.2595, ctc_loss=0.1769, cr_loss=0.4128, over 16985.00 frames. ], tot_loss[loss=0.2362, ctc_loss=0.1615, cr_loss=0.3737, over 3360696.17 frames. ], batch size: 56, lr: 9.23e-03, grad_scale: 32.0 2024-09-23 09:43:29,899 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=230542.66666666666, ans=0.0 2024-09-23 09:43:41,563 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.68 vs. limit=15.0 2024-09-23 09:44:18,497 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.53 vs. limit=15.0 2024-09-23 09:44:28,570 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.270e+02 1.366e+02 1.498e+02 2.280e+02, threshold=2.732e+02, percent-clipped=0.0 2024-09-23 09:44:28,983 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=230729.33333333334, ans=0.125 2024-09-23 09:44:39,785 INFO [train.py:1198] (2/4) Epoch 13, batch 2700, loss[loss=0.2452, ctc_loss=0.1691, cr_loss=0.3806, over 16753.00 frames. ], tot_loss[loss=0.2363, ctc_loss=0.1615, cr_loss=0.3741, over 3364709.85 frames. ], batch size: 61, lr: 9.22e-03, grad_scale: 32.0 2024-09-23 09:44:54,854 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.37 vs. limit=15.0 2024-09-23 09:45:20,935 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=230869.33333333334, ans=0.125 2024-09-23 09:45:38,693 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 09:46:02,466 INFO [train.py:1198] (2/4) Epoch 13, batch 2750, loss[loss=0.2546, ctc_loss=0.1751, cr_loss=0.3972, over 16873.00 frames. ], tot_loss[loss=0.2367, ctc_loss=0.1618, cr_loss=0.3746, over 3366329.35 frames. ], batch size: 58, lr: 9.22e-03, grad_scale: 32.0 2024-09-23 09:46:57,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=231149.33333333334, ans=0.125 2024-09-23 09:47:10,537 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=231196.0, ans=0.0 2024-09-23 09:47:19,273 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.267e+02 1.386e+02 1.565e+02 1.913e+02, threshold=2.771e+02, percent-clipped=0.0 2024-09-23 09:47:19,600 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=231196.0, ans=0.125 2024-09-23 09:47:22,923 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=231196.0, ans=0.05 2024-09-23 09:47:30,586 INFO [train.py:1198] (2/4) Epoch 13, batch 2800, loss[loss=0.2438, ctc_loss=0.1654, cr_loss=0.3919, over 17298.00 frames. ], tot_loss[loss=0.2372, ctc_loss=0.1621, cr_loss=0.3752, over 3373080.35 frames. ], batch size: 49, lr: 9.21e-03, grad_scale: 32.0 2024-09-23 09:47:46,621 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=231289.33333333334, ans=0.125 2024-09-23 09:48:50,273 INFO [train.py:1198] (2/4) Epoch 13, batch 2850, loss[loss=0.2366, ctc_loss=0.1616, cr_loss=0.3751, over 17293.00 frames. ], tot_loss[loss=0.2366, ctc_loss=0.1617, cr_loss=0.3745, over 3375926.83 frames. ], batch size: 49, lr: 9.21e-03, grad_scale: 16.0 2024-09-23 09:48:57,299 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=15.24 vs. limit=22.5 2024-09-23 09:49:15,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=231522.66666666666, ans=0.04949747468305833 2024-09-23 09:49:44,708 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=231616.0, ans=0.0 2024-09-23 09:49:52,810 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=231662.66666666666, ans=0.125 2024-09-23 09:50:00,334 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.288e+02 1.395e+02 1.565e+02 5.212e+02, threshold=2.790e+02, percent-clipped=1.0 2024-09-23 09:50:04,623 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=231662.66666666666, ans=0.0 2024-09-23 09:50:12,523 INFO [train.py:1198] (2/4) Epoch 13, batch 2900, loss[loss=0.2273, ctc_loss=0.156, cr_loss=0.3566, over 16975.00 frames. ], tot_loss[loss=0.2374, ctc_loss=0.1622, cr_loss=0.3757, over 3378632.18 frames. ], batch size: 53, lr: 9.20e-03, grad_scale: 16.0 2024-09-23 09:50:35,771 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.39 vs. limit=15.0 2024-09-23 09:50:36,939 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=231756.0, ans=0.0 2024-09-23 09:50:43,411 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=231802.66666666666, ans=0.125 2024-09-23 09:51:35,581 INFO [train.py:1198] (2/4) Epoch 13, batch 2950, loss[loss=0.2455, ctc_loss=0.1704, cr_loss=0.3758, over 17016.00 frames. ], tot_loss[loss=0.2367, ctc_loss=0.1616, cr_loss=0.3753, over 3376043.41 frames. ], batch size: 51, lr: 9.20e-03, grad_scale: 16.0 2024-09-23 09:51:40,803 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=231942.66666666666, ans=0.0 2024-09-23 09:51:42,273 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=231942.66666666666, ans=0.0 2024-09-23 09:51:42,386 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=231942.66666666666, ans=0.025 2024-09-23 09:52:06,185 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=232036.0, ans=0.0 2024-09-23 09:52:22,988 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=232036.0, ans=0.125 2024-09-23 09:52:26,218 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=232082.66666666666, ans=0.0 2024-09-23 09:52:26,269 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=232082.66666666666, ans=0.2 2024-09-23 09:52:29,456 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 09:52:47,837 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.241e+02 1.342e+02 1.458e+02 2.416e+02, threshold=2.684e+02, percent-clipped=0.0 2024-09-23 09:52:52,998 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=232129.33333333334, ans=0.0 2024-09-23 09:52:55,963 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=232176.0, ans=0.0 2024-09-23 09:52:57,271 INFO [train.py:1198] (2/4) Epoch 13, batch 3000, loss[loss=0.2135, ctc_loss=0.1449, cr_loss=0.3431, over 17036.00 frames. ], tot_loss[loss=0.2365, ctc_loss=0.1616, cr_loss=0.3744, over 3357991.65 frames. ], batch size: 44, lr: 9.19e-03, grad_scale: 16.0 2024-09-23 09:52:57,272 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 09:53:09,524 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.2.encoder.layers.0.self_attn_weights, attn_weights_entropy = tensor([4.8578, 4.6246, 4.5574, 4.2390], device='cuda:2') 2024-09-23 09:53:12,981 INFO [train.py:1230] (2/4) Epoch 13, validation: loss=0.04424, ctc_loss=0.04424, cr_loss=7.269e-15, over 944034.00 frames. 2024-09-23 09:53:12,982 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 09:53:18,980 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.96 vs. limit=15.0 2024-09-23 09:53:28,882 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=232222.66666666666, ans=0.2 2024-09-23 09:53:38,944 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.77 vs. limit=6.0 2024-09-23 09:53:39,792 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=232222.66666666666, ans=0.2 2024-09-23 09:54:28,526 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=232362.66666666666, ans=0.1 2024-09-23 09:54:31,494 INFO [train.py:1198] (2/4) Epoch 13, batch 3050, loss[loss=0.2059, ctc_loss=0.1402, cr_loss=0.3287, over 17300.00 frames. ], tot_loss[loss=0.2351, ctc_loss=0.1604, cr_loss=0.3734, over 3370207.02 frames. ], batch size: 51, lr: 9.19e-03, grad_scale: 16.0 2024-09-23 09:55:06,540 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=8.42 vs. limit=15.0 2024-09-23 09:55:16,949 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=232549.33333333334, ans=0.0 2024-09-23 09:55:39,980 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.113e+02 1.372e+02 1.503e+02 1.600e+02 2.658e+02, threshold=3.005e+02, percent-clipped=0.0 2024-09-23 09:55:48,142 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=232642.66666666666, ans=0.125 2024-09-23 09:55:49,566 INFO [train.py:1198] (2/4) Epoch 13, batch 3100, loss[loss=0.2359, ctc_loss=0.1664, cr_loss=0.3476, over 17351.00 frames. ], tot_loss[loss=0.2362, ctc_loss=0.1614, cr_loss=0.374, over 3356325.92 frames. ], batch size: 48, lr: 9.18e-03, grad_scale: 16.0 2024-09-23 09:55:51,986 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=17.41 vs. limit=22.5 2024-09-23 09:55:59,485 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=232642.66666666666, ans=0.0 2024-09-23 09:56:02,518 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=232642.66666666666, ans=0.125 2024-09-23 09:56:08,809 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=232689.33333333334, ans=0.125 2024-09-23 09:56:24,589 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=232736.0, ans=0.125 2024-09-23 09:57:01,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=232829.33333333334, ans=0.125 2024-09-23 09:57:10,783 INFO [train.py:1198] (2/4) Epoch 13, batch 3150, loss[loss=0.2368, ctc_loss=0.1605, cr_loss=0.3815, over 17000.00 frames. ], tot_loss[loss=0.2363, ctc_loss=0.1615, cr_loss=0.3739, over 3353999.48 frames. ], batch size: 51, lr: 9.18e-03, grad_scale: 16.0 2024-09-23 09:57:24,497 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=5.05 vs. limit=15.0 2024-09-23 09:57:34,737 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=232922.66666666666, ans=0.0 2024-09-23 09:57:36,671 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.84 vs. limit=22.5 2024-09-23 09:57:40,901 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=232969.33333333334, ans=0.125 2024-09-23 09:58:10,696 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=233016.0, ans=0.125 2024-09-23 09:58:12,216 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=233062.66666666666, ans=0.125 2024-09-23 09:58:19,650 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.158e+02 1.308e+02 1.468e+02 1.678e+02 2.567e+02, threshold=2.937e+02, percent-clipped=0.0 2024-09-23 09:58:28,987 INFO [train.py:1198] (2/4) Epoch 13, batch 3200, loss[loss=0.2666, ctc_loss=0.1836, cr_loss=0.4153, over 16887.00 frames. ], tot_loss[loss=0.2364, ctc_loss=0.1615, cr_loss=0.3741, over 3356793.59 frames. ], batch size: 58, lr: 9.18e-03, grad_scale: 32.0 2024-09-23 09:58:46,582 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=233156.0, ans=0.0 2024-09-23 09:59:44,588 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=13.89 vs. limit=15.0 2024-09-23 09:59:49,387 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=233296.0, ans=0.125 2024-09-23 09:59:49,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=233296.0, ans=0.05 2024-09-23 09:59:52,147 INFO [train.py:1198] (2/4) Epoch 13, batch 3250, loss[loss=0.2226, ctc_loss=0.1529, cr_loss=0.3485, over 17235.00 frames. ], tot_loss[loss=0.236, ctc_loss=0.1612, cr_loss=0.3739, over 3362867.61 frames. ], batch size: 44, lr: 9.17e-03, grad_scale: 32.0 2024-09-23 09:59:57,073 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=233342.66666666666, ans=0.2 2024-09-23 09:59:57,111 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=233342.66666666666, ans=0.125 2024-09-23 10:00:00,984 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=5.05 vs. limit=15.0 2024-09-23 10:00:03,856 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.96 vs. limit=12.0 2024-09-23 10:00:08,328 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=233389.33333333334, ans=0.125 2024-09-23 10:00:11,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=233389.33333333334, ans=0.0 2024-09-23 10:00:11,458 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=233389.33333333334, ans=0.125 2024-09-23 10:00:11,963 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=19.55 vs. limit=22.5 2024-09-23 10:00:16,124 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=233389.33333333334, ans=0.125 2024-09-23 10:00:17,609 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=233389.33333333334, ans=0.0 2024-09-23 10:00:22,440 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=233436.0, ans=0.2 2024-09-23 10:01:01,302 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.287e+02 1.416e+02 1.581e+02 2.137e+02, threshold=2.832e+02, percent-clipped=0.0 2024-09-23 10:01:03,147 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=233529.33333333334, ans=0.0 2024-09-23 10:01:10,547 INFO [train.py:1198] (2/4) Epoch 13, batch 3300, loss[loss=0.2227, ctc_loss=0.1536, cr_loss=0.3454, over 17178.00 frames. ], tot_loss[loss=0.2342, ctc_loss=0.1598, cr_loss=0.3723, over 3362746.53 frames. ], batch size: 45, lr: 9.17e-03, grad_scale: 32.0 2024-09-23 10:01:13,946 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=233576.0, ans=0.125 2024-09-23 10:01:26,515 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=233622.66666666666, ans=0.0 2024-09-23 10:01:32,642 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=233622.66666666666, ans=0.07 2024-09-23 10:01:40,386 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 10:02:04,845 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.52 vs. limit=15.0 2024-09-23 10:02:15,157 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=233762.66666666666, ans=0.125 2024-09-23 10:02:22,809 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=233762.66666666666, ans=0.0 2024-09-23 10:02:30,429 INFO [train.py:1198] (2/4) Epoch 13, batch 3350, loss[loss=0.2314, ctc_loss=0.1564, cr_loss=0.375, over 17093.00 frames. ], tot_loss[loss=0.2336, ctc_loss=0.1593, cr_loss=0.3715, over 3369352.96 frames. ], batch size: 49, lr: 9.16e-03, grad_scale: 32.0 2024-09-23 10:02:33,800 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=233809.33333333334, ans=0.2 2024-09-23 10:02:44,720 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=233856.0, ans=0.2 2024-09-23 10:03:00,103 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=233902.66666666666, ans=0.2 2024-09-23 10:03:00,114 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=233902.66666666666, ans=0.2 2024-09-23 10:03:29,940 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=233949.33333333334, ans=0.1 2024-09-23 10:03:38,963 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.032e+02 1.272e+02 1.377e+02 1.520e+02 2.187e+02, threshold=2.755e+02, percent-clipped=0.0 2024-09-23 10:03:45,919 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=11.48 vs. limit=15.0 2024-09-23 10:03:46,953 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=234042.66666666666, ans=0.1 2024-09-23 10:03:48,297 INFO [train.py:1198] (2/4) Epoch 13, batch 3400, loss[loss=0.1812, ctc_loss=0.1195, cr_loss=0.3087, over 17027.00 frames. ], tot_loss[loss=0.2344, ctc_loss=0.1599, cr_loss=0.3723, over 3369170.28 frames. ], batch size: 39, lr: 9.16e-03, grad_scale: 32.0 2024-09-23 10:03:59,194 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.min_positive, batch_count=234042.66666666666, ans=0.025 2024-09-23 10:04:21,087 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.min_positive, batch_count=234136.0, ans=0.025 2024-09-23 10:04:28,986 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=234136.0, ans=0.035 2024-09-23 10:05:06,442 INFO [train.py:1198] (2/4) Epoch 13, batch 3450, loss[loss=0.2206, ctc_loss=0.1489, cr_loss=0.3586, over 17363.00 frames. ], tot_loss[loss=0.2345, ctc_loss=0.1601, cr_loss=0.3724, over 3370207.63 frames. ], batch size: 48, lr: 9.15e-03, grad_scale: 32.0 2024-09-23 10:05:09,004 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.whiten, num_groups=1, num_channels=192, metric=3.91 vs. limit=12.0 2024-09-23 10:06:15,807 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.304e+02 1.412e+02 1.643e+02 2.368e+02, threshold=2.824e+02, percent-clipped=0.0 2024-09-23 10:06:19,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=234462.66666666666, ans=0.1 2024-09-23 10:06:20,748 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=234462.66666666666, ans=0.05 2024-09-23 10:06:25,214 INFO [train.py:1198] (2/4) Epoch 13, batch 3500, loss[loss=0.2449, ctc_loss=0.1697, cr_loss=0.376, over 16937.00 frames. ], tot_loss[loss=0.2343, ctc_loss=0.1599, cr_loss=0.372, over 3367492.81 frames. ], batch size: 58, lr: 9.15e-03, grad_scale: 32.0 2024-09-23 10:06:36,438 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=234509.33333333334, ans=0.125 2024-09-23 10:06:41,301 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=234556.0, ans=0.0 2024-09-23 10:06:52,336 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=234556.0, ans=0.0 2024-09-23 10:07:45,894 INFO [train.py:1198] (2/4) Epoch 13, batch 3550, loss[loss=0.2338, ctc_loss=0.1613, cr_loss=0.3629, over 16607.00 frames. ], tot_loss[loss=0.2342, ctc_loss=0.1599, cr_loss=0.3715, over 3356184.60 frames. ], batch size: 66, lr: 9.14e-03, grad_scale: 32.0 2024-09-23 10:08:39,563 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=234882.66666666666, ans=0.0 2024-09-23 10:08:44,382 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=234882.66666666666, ans=0.1 2024-09-23 10:09:00,082 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.145e+02 1.323e+02 1.492e+02 1.691e+02 2.949e+02, threshold=2.984e+02, percent-clipped=2.0 2024-09-23 10:09:07,582 INFO [train.py:1198] (2/4) Epoch 13, batch 3600, loss[loss=0.2266, ctc_loss=0.1512, cr_loss=0.3774, over 17261.00 frames. ], tot_loss[loss=0.2345, ctc_loss=0.1601, cr_loss=0.3721, over 3352842.73 frames. ], batch size: 44, lr: 9.14e-03, grad_scale: 32.0 2024-09-23 10:09:09,295 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=234976.0, ans=0.025 2024-09-23 10:09:28,114 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.37 vs. limit=15.0 2024-09-23 10:09:57,759 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=235116.0, ans=0.125 2024-09-23 10:10:13,324 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=235162.66666666666, ans=0.125 2024-09-23 10:10:25,895 INFO [train.py:1198] (2/4) Epoch 13, batch 3650, loss[loss=0.2564, ctc_loss=0.1763, cr_loss=0.4007, over 17042.00 frames. ], tot_loss[loss=0.2344, ctc_loss=0.16, cr_loss=0.3717, over 3351335.38 frames. ], batch size: 52, lr: 9.14e-03, grad_scale: 32.0 2024-09-23 10:10:27,784 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 10:10:51,619 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=20.43 vs. limit=22.5 2024-09-23 10:11:24,126 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=235349.33333333334, ans=0.125 2024-09-23 10:11:37,917 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.152e+02 1.270e+02 1.340e+02 1.472e+02 2.759e+02, threshold=2.681e+02, percent-clipped=0.0 2024-09-23 10:11:45,789 INFO [train.py:1198] (2/4) Epoch 13, batch 3700, loss[loss=0.2665, ctc_loss=0.186, cr_loss=0.4022, over 17024.00 frames. ], tot_loss[loss=0.2349, ctc_loss=0.1604, cr_loss=0.3723, over 3351736.87 frames. ], batch size: 53, lr: 9.13e-03, grad_scale: 32.0 2024-09-23 10:11:46,004 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=235442.66666666666, ans=0.2 2024-09-23 10:12:14,943 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=235489.33333333334, ans=0.125 2024-09-23 10:12:43,332 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.96 vs. limit=15.0 2024-09-23 10:12:47,590 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.63 vs. limit=22.5 2024-09-23 10:13:03,920 INFO [train.py:1198] (2/4) Epoch 13, batch 3750, loss[loss=0.2781, ctc_loss=0.1945, cr_loss=0.418, over 16508.00 frames. ], tot_loss[loss=0.2361, ctc_loss=0.1614, cr_loss=0.3735, over 3354071.72 frames. ], batch size: 66, lr: 9.13e-03, grad_scale: 32.0 2024-09-23 10:13:17,086 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=235676.0, ans=0.0 2024-09-23 10:13:43,738 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=235769.33333333334, ans=0.125 2024-09-23 10:13:48,861 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.78 vs. limit=15.0 2024-09-23 10:13:51,330 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 10:13:54,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=235816.0, ans=0.1 2024-09-23 10:13:56,060 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=235816.0, ans=0.0 2024-09-23 10:14:11,467 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=235862.66666666666, ans=0.0 2024-09-23 10:14:14,235 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.047e+02 1.350e+02 1.440e+02 1.635e+02 2.891e+02, threshold=2.880e+02, percent-clipped=1.0 2024-09-23 10:14:22,032 INFO [train.py:1198] (2/4) Epoch 13, batch 3800, loss[loss=0.2752, ctc_loss=0.1952, cr_loss=0.4, over 14920.00 frames. ], tot_loss[loss=0.2367, ctc_loss=0.1621, cr_loss=0.3728, over 3325950.27 frames. ], batch size: 89, lr: 9.12e-03, grad_scale: 32.0 2024-09-23 10:14:22,970 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=8.47 vs. limit=15.0 2024-09-23 10:14:37,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=235956.0, ans=0.2 2024-09-23 10:15:09,043 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.54 vs. limit=15.0 2024-09-23 10:15:11,995 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=236049.33333333334, ans=0.125 2024-09-23 10:15:25,961 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=236096.0, ans=0.125 2024-09-23 10:15:40,299 INFO [train.py:1198] (2/4) Epoch 13, batch 3850, loss[loss=0.2821, ctc_loss=0.2001, cr_loss=0.4099, over 14992.00 frames. ], tot_loss[loss=0.2393, ctc_loss=0.1644, cr_loss=0.3747, over 3268184.31 frames. ], batch size: 89, lr: 9.12e-03, grad_scale: 16.0 2024-09-23 10:15:44,401 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=6.71 vs. limit=15.0 2024-09-23 10:15:45,393 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=236142.66666666666, ans=0.1 2024-09-23 10:16:01,871 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=236189.33333333334, ans=0.025 2024-09-23 10:16:42,046 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten.whitening_limit, batch_count=236329.33333333334, ans=15.0 2024-09-23 10:17:39,608 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.132e+02 1.439e+02 1.592e+02 1.784e+02 2.502e+02, threshold=3.185e+02, percent-clipped=0.0 2024-09-23 10:17:39,632 INFO [train.py:1198] (2/4) Epoch 14, batch 0, loss[loss=0.239, ctc_loss=0.1626, cr_loss=0.382, over 17033.00 frames. ], tot_loss[loss=0.239, ctc_loss=0.1626, cr_loss=0.382, over 17033.00 frames. ], batch size: 44, lr: 8.78e-03, grad_scale: 32.0 2024-09-23 10:17:39,632 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 10:17:55,043 INFO [train.py:1230] (2/4) Epoch 14, validation: loss=0.04435, ctc_loss=0.04435, cr_loss=7.317e-15, over 944034.00 frames. 2024-09-23 10:17:55,043 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 10:18:09,141 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=236357.33333333334, ans=0.1 2024-09-23 10:18:22,626 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.12 vs. limit=15.0 2024-09-23 10:18:33,360 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=236450.66666666666, ans=0.0 2024-09-23 10:19:01,194 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=236497.33333333334, ans=0.1 2024-09-23 10:19:08,664 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=7.50 vs. limit=12.0 2024-09-23 10:19:08,861 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.02 vs. limit=6.0 2024-09-23 10:19:14,677 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=236544.0, ans=0.1 2024-09-23 10:19:19,540 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=236544.0, ans=0.025 2024-09-23 10:19:22,450 INFO [train.py:1198] (2/4) Epoch 14, batch 50, loss[loss=0.2201, ctc_loss=0.1486, cr_loss=0.3574, over 16976.00 frames. ], tot_loss[loss=0.2331, ctc_loss=0.1588, cr_loss=0.3714, over 769924.66 frames. ], batch size: 53, lr: 8.78e-03, grad_scale: 32.0 2024-09-23 10:19:29,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=236590.66666666666, ans=0.125 2024-09-23 10:19:29,649 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=19.13 vs. limit=22.5 2024-09-23 10:19:32,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=236590.66666666666, ans=0.035 2024-09-23 10:19:37,100 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=236637.33333333334, ans=0.2 2024-09-23 10:19:43,422 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=236637.33333333334, ans=0.125 2024-09-23 10:19:43,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=236637.33333333334, ans=0.125 2024-09-23 10:20:00,868 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=236684.0, ans=0.1 2024-09-23 10:20:09,157 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 10:20:26,539 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=236777.33333333334, ans=0.2 2024-09-23 10:20:34,697 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=236777.33333333334, ans=0.025 2024-09-23 10:20:42,383 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.241e+02 1.435e+02 1.718e+02 2.310e+02, threshold=2.871e+02, percent-clipped=0.0 2024-09-23 10:20:42,407 INFO [train.py:1198] (2/4) Epoch 14, batch 100, loss[loss=0.2618, ctc_loss=0.1832, cr_loss=0.3932, over 17041.00 frames. ], tot_loss[loss=0.2331, ctc_loss=0.1587, cr_loss=0.372, over 1348098.42 frames. ], batch size: 52, lr: 8.77e-03, grad_scale: 32.0 2024-09-23 10:21:06,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=236870.66666666666, ans=0.025 2024-09-23 10:21:19,782 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=236917.33333333334, ans=0.0 2024-09-23 10:21:24,545 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=236917.33333333334, ans=0.1 2024-09-23 10:21:26,289 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=236917.33333333334, ans=0.0 2024-09-23 10:21:42,522 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=236964.0, ans=0.1 2024-09-23 10:21:45,625 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=237010.66666666666, ans=0.1 2024-09-23 10:21:49,719 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.71 vs. limit=6.0 2024-09-23 10:22:03,253 INFO [train.py:1198] (2/4) Epoch 14, batch 150, loss[loss=0.1985, ctc_loss=0.1334, cr_loss=0.3253, over 17051.00 frames. ], tot_loss[loss=0.234, ctc_loss=0.1593, cr_loss=0.3736, over 1796928.26 frames. ], batch size: 46, lr: 8.77e-03, grad_scale: 32.0 2024-09-23 10:22:05,227 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=237057.33333333334, ans=0.04949747468305833 2024-09-23 10:22:18,110 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=237104.0, ans=0.025 2024-09-23 10:22:21,805 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.94 vs. limit=15.0 2024-09-23 10:22:33,784 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=237150.66666666666, ans=0.125 2024-09-23 10:22:38,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=237150.66666666666, ans=0.025 2024-09-23 10:22:43,285 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=237150.66666666666, ans=0.0 2024-09-23 10:23:18,298 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.50 vs. limit=15.0 2024-09-23 10:23:21,005 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=237244.0, ans=10.0 2024-09-23 10:23:28,629 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.307e+02 1.446e+02 1.689e+02 2.559e+02, threshold=2.892e+02, percent-clipped=0.0 2024-09-23 10:23:28,653 INFO [train.py:1198] (2/4) Epoch 14, batch 200, loss[loss=0.2449, ctc_loss=0.1661, cr_loss=0.3937, over 17291.00 frames. ], tot_loss[loss=0.2331, ctc_loss=0.1586, cr_loss=0.3728, over 2147717.12 frames. ], batch size: 46, lr: 8.76e-03, grad_scale: 32.0 2024-09-23 10:23:36,873 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=237290.66666666666, ans=0.2 2024-09-23 10:23:54,020 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=237337.33333333334, ans=0.0 2024-09-23 10:24:15,932 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=237384.0, ans=0.125 2024-09-23 10:24:25,511 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer_ff3.min_abs, batch_count=237430.66666666666, ans=0.2 2024-09-23 10:24:36,670 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=237477.33333333334, ans=0.1 2024-09-23 10:24:53,916 INFO [train.py:1198] (2/4) Epoch 14, batch 250, loss[loss=0.2237, ctc_loss=0.1526, cr_loss=0.3555, over 17290.00 frames. ], tot_loss[loss=0.2335, ctc_loss=0.1589, cr_loss=0.3727, over 2408600.50 frames. ], batch size: 49, lr: 8.76e-03, grad_scale: 32.0 2024-09-23 10:24:59,024 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=237524.0, ans=0.125 2024-09-23 10:25:52,959 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=237664.0, ans=0.0 2024-09-23 10:26:07,286 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=237710.66666666666, ans=0.125 2024-09-23 10:26:13,348 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.153e+02 1.272e+02 1.438e+02 1.640e+02 2.630e+02, threshold=2.876e+02, percent-clipped=0.0 2024-09-23 10:26:13,372 INFO [train.py:1198] (2/4) Epoch 14, batch 300, loss[loss=0.2194, ctc_loss=0.1473, cr_loss=0.3601, over 17080.00 frames. ], tot_loss[loss=0.2335, ctc_loss=0.1591, cr_loss=0.3724, over 2605544.04 frames. ], batch size: 43, lr: 8.76e-03, grad_scale: 32.0 2024-09-23 10:26:23,724 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=15.07 vs. limit=22.5 2024-09-23 10:26:28,290 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=5.68 vs. limit=12.0 2024-09-23 10:26:40,456 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=237804.0, ans=0.125 2024-09-23 10:27:01,423 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.29 vs. limit=15.0 2024-09-23 10:27:32,701 INFO [train.py:1198] (2/4) Epoch 14, batch 350, loss[loss=0.2382, ctc_loss=0.1616, cr_loss=0.383, over 17247.00 frames. ], tot_loss[loss=0.2338, ctc_loss=0.1594, cr_loss=0.372, over 2781165.76 frames. ], batch size: 44, lr: 8.75e-03, grad_scale: 32.0 2024-09-23 10:27:41,200 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=237990.66666666666, ans=0.125 2024-09-23 10:28:05,538 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=238037.33333333334, ans=0.125 2024-09-23 10:28:16,646 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=238084.0, ans=0.1 2024-09-23 10:28:17,008 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.94 vs. limit=22.5 2024-09-23 10:28:29,309 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=238130.66666666666, ans=0.1 2024-09-23 10:29:02,515 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.171e+02 1.337e+02 1.482e+02 1.714e+02 2.531e+02, threshold=2.964e+02, percent-clipped=0.0 2024-09-23 10:29:02,539 INFO [train.py:1198] (2/4) Epoch 14, batch 400, loss[loss=0.2165, ctc_loss=0.1482, cr_loss=0.3412, over 17022.00 frames. ], tot_loss[loss=0.234, ctc_loss=0.1596, cr_loss=0.3718, over 2904422.43 frames. ], batch size: 44, lr: 8.75e-03, grad_scale: 32.0 2024-09-23 10:29:16,778 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=238270.66666666666, ans=0.125 2024-09-23 10:29:43,765 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=238317.33333333334, ans=0.1 2024-09-23 10:30:11,347 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=3.93 vs. limit=15.0 2024-09-23 10:30:21,630 INFO [train.py:1198] (2/4) Epoch 14, batch 450, loss[loss=0.244, ctc_loss=0.1657, cr_loss=0.3915, over 16435.00 frames. ], tot_loss[loss=0.233, ctc_loss=0.1589, cr_loss=0.3707, over 3002895.25 frames. ], batch size: 66, lr: 8.74e-03, grad_scale: 32.0 2024-09-23 10:31:39,674 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.min_abs, batch_count=238690.66666666666, ans=0.5 2024-09-23 10:31:39,775 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=238690.66666666666, ans=0.5 2024-09-23 10:31:41,084 INFO [train.py:1198] (2/4) Epoch 14, batch 500, loss[loss=0.1908, ctc_loss=0.1247, cr_loss=0.3302, over 17185.00 frames. ], tot_loss[loss=0.2321, ctc_loss=0.1581, cr_loss=0.3701, over 3084141.40 frames. ], batch size: 41, lr: 8.74e-03, grad_scale: 16.0 2024-09-23 10:31:42,719 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.102e+02 1.208e+02 1.324e+02 1.495e+02 2.086e+02, threshold=2.649e+02, percent-clipped=0.0 2024-09-23 10:31:57,362 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=238737.33333333334, ans=0.125 2024-09-23 10:32:08,314 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=238737.33333333334, ans=0.0 2024-09-23 10:32:19,448 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=238784.0, ans=0.2 2024-09-23 10:32:27,580 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=238830.66666666666, ans=0.2 2024-09-23 10:33:06,646 INFO [train.py:1198] (2/4) Epoch 14, batch 550, loss[loss=0.2285, ctc_loss=0.1582, cr_loss=0.3514, over 17009.00 frames. ], tot_loss[loss=0.2338, ctc_loss=0.1594, cr_loss=0.372, over 3141033.04 frames. ], batch size: 44, lr: 8.74e-03, grad_scale: 16.0 2024-09-23 10:34:05,070 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=239064.0, ans=0.0 2024-09-23 10:34:06,760 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=239064.0, ans=0.125 2024-09-23 10:34:06,811 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=239064.0, ans=0.0 2024-09-23 10:34:11,525 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.min_abs, batch_count=239064.0, ans=0.5 2024-09-23 10:34:17,972 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=239110.66666666666, ans=0.125 2024-09-23 10:34:31,893 INFO [train.py:1198] (2/4) Epoch 14, batch 600, loss[loss=0.2117, ctc_loss=0.14, cr_loss=0.3584, over 17274.00 frames. ], tot_loss[loss=0.2339, ctc_loss=0.1595, cr_loss=0.3719, over 3183489.54 frames. ], batch size: 42, lr: 8.73e-03, grad_scale: 16.0 2024-09-23 10:34:33,454 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.266e+02 1.344e+02 1.475e+02 2.652e+02, threshold=2.689e+02, percent-clipped=1.0 2024-09-23 10:34:35,400 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=239157.33333333334, ans=0.125 2024-09-23 10:34:36,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=239157.33333333334, ans=0.125 2024-09-23 10:34:48,113 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=239204.0, ans=0.0 2024-09-23 10:34:55,863 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=239204.0, ans=0.0 2024-09-23 10:35:21,620 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=11.33 vs. limit=12.0 2024-09-23 10:35:32,369 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=239297.33333333334, ans=0.125 2024-09-23 10:35:38,775 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=239344.0, ans=0.1 2024-09-23 10:35:42,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=239344.0, ans=0.125 2024-09-23 10:35:51,347 INFO [train.py:1198] (2/4) Epoch 14, batch 650, loss[loss=0.183, ctc_loss=0.1219, cr_loss=0.3055, over 16693.00 frames. ], tot_loss[loss=0.2349, ctc_loss=0.1603, cr_loss=0.373, over 3228559.03 frames. ], batch size: 37, lr: 8.73e-03, grad_scale: 16.0 2024-09-23 10:35:54,843 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=239390.66666666666, ans=0.09899494936611666 2024-09-23 10:36:34,694 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=239484.0, ans=0.125 2024-09-23 10:36:37,000 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=11.24 vs. limit=15.0 2024-09-23 10:37:05,099 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=239577.33333333334, ans=0.125 2024-09-23 10:37:11,334 INFO [train.py:1198] (2/4) Epoch 14, batch 700, loss[loss=0.2224, ctc_loss=0.1507, cr_loss=0.3585, over 17048.00 frames. ], tot_loss[loss=0.2351, ctc_loss=0.1604, cr_loss=0.3736, over 3251845.28 frames. ], batch size: 39, lr: 8.72e-03, grad_scale: 16.0 2024-09-23 10:37:11,738 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 10:37:13,000 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.106e+02 1.260e+02 1.373e+02 1.552e+02 2.322e+02, threshold=2.747e+02, percent-clipped=0.0 2024-09-23 10:37:15,023 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=239624.0, ans=0.0 2024-09-23 10:37:49,683 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=3.99 vs. limit=10.0 2024-09-23 10:38:27,627 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=239810.66666666666, ans=0.125 2024-09-23 10:38:29,245 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=239810.66666666666, ans=0.1 2024-09-23 10:38:33,526 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=239810.66666666666, ans=0.125 2024-09-23 10:38:39,754 INFO [train.py:1198] (2/4) Epoch 14, batch 750, loss[loss=0.2085, ctc_loss=0.1395, cr_loss=0.3454, over 17038.00 frames. ], tot_loss[loss=0.234, ctc_loss=0.1594, cr_loss=0.3733, over 3287036.18 frames. ], batch size: 39, lr: 8.72e-03, grad_scale: 16.0 2024-09-23 10:38:49,912 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.45 vs. limit=15.0 2024-09-23 10:39:11,259 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=239904.0, ans=0.125 2024-09-23 10:39:35,829 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=11.69 vs. limit=15.0 2024-09-23 10:39:36,769 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=239997.33333333334, ans=0.125 2024-09-23 10:39:54,395 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=240044.0, ans=0.1 2024-09-23 10:39:54,539 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=240044.0, ans=0.125 2024-09-23 10:39:56,043 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=240044.0, ans=0.0 2024-09-23 10:40:02,417 INFO [train.py:1198] (2/4) Epoch 14, batch 800, loss[loss=0.2007, ctc_loss=0.1356, cr_loss=0.3253, over 17042.00 frames. ], tot_loss[loss=0.2342, ctc_loss=0.1596, cr_loss=0.3734, over 3299250.46 frames. ], batch size: 39, lr: 8.71e-03, grad_scale: 32.0 2024-09-23 10:40:03,951 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.109e+02 1.283e+02 1.393e+02 1.518e+02 3.186e+02, threshold=2.786e+02, percent-clipped=2.0 2024-09-23 10:40:20,622 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=13.76 vs. limit=15.0 2024-09-23 10:40:37,342 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=240184.0, ans=0.025 2024-09-23 10:40:47,066 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=240184.0, ans=0.1 2024-09-23 10:40:48,775 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=240230.66666666666, ans=0.125 2024-09-23 10:40:48,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=240230.66666666666, ans=0.125 2024-09-23 10:40:51,927 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=240230.66666666666, ans=0.035 2024-09-23 10:41:02,316 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.77 vs. limit=12.0 2024-09-23 10:41:03,261 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=240230.66666666666, ans=0.125 2024-09-23 10:41:22,388 INFO [train.py:1198] (2/4) Epoch 14, batch 850, loss[loss=0.2444, ctc_loss=0.1689, cr_loss=0.3779, over 17076.00 frames. ], tot_loss[loss=0.2337, ctc_loss=0.159, cr_loss=0.3732, over 3319256.09 frames. ], batch size: 46, lr: 8.71e-03, grad_scale: 32.0 2024-09-23 10:41:25,788 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=240324.0, ans=0.0 2024-09-23 10:41:56,170 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.max_positive, batch_count=240417.33333333334, ans=0.95 2024-09-23 10:42:01,323 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.52 vs. limit=6.0 2024-09-23 10:42:10,032 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=240464.0, ans=0.125 2024-09-23 10:42:35,774 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 10:42:44,012 INFO [train.py:1198] (2/4) Epoch 14, batch 900, loss[loss=0.2249, ctc_loss=0.1523, cr_loss=0.3629, over 17165.00 frames. ], tot_loss[loss=0.2338, ctc_loss=0.1591, cr_loss=0.3736, over 3335163.60 frames. ], batch size: 45, lr: 8.71e-03, grad_scale: 32.0 2024-09-23 10:42:48,314 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.073e+02 1.265e+02 1.359e+02 1.500e+02 2.203e+02, threshold=2.718e+02, percent-clipped=0.0 2024-09-23 10:42:59,856 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 10:43:20,502 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=240650.66666666666, ans=0.125 2024-09-23 10:43:40,629 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=240697.33333333334, ans=0.1 2024-09-23 10:43:54,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=240744.0, ans=0.125 2024-09-23 10:44:07,061 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=240744.0, ans=0.125 2024-09-23 10:44:11,568 INFO [train.py:1198] (2/4) Epoch 14, batch 950, loss[loss=0.2013, ctc_loss=0.1343, cr_loss=0.3352, over 16989.00 frames. ], tot_loss[loss=0.2353, ctc_loss=0.1603, cr_loss=0.3752, over 3329239.57 frames. ], batch size: 42, lr: 8.70e-03, grad_scale: 32.0 2024-09-23 10:44:23,658 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.79 vs. limit=15.0 2024-09-23 10:44:37,737 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=240837.33333333334, ans=0.2 2024-09-23 10:44:53,727 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=240884.0, ans=0.1 2024-09-23 10:44:55,585 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.97 vs. limit=10.0 2024-09-23 10:45:24,156 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=240977.33333333334, ans=0.0 2024-09-23 10:45:31,844 INFO [train.py:1198] (2/4) Epoch 14, batch 1000, loss[loss=0.2469, ctc_loss=0.1707, cr_loss=0.3807, over 16767.00 frames. ], tot_loss[loss=0.2353, ctc_loss=0.1603, cr_loss=0.3749, over 3334341.55 frames. ], batch size: 61, lr: 8.70e-03, grad_scale: 32.0 2024-09-23 10:45:33,325 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.317e+02 1.495e+02 1.766e+02 2.386e+02, threshold=2.990e+02, percent-clipped=0.0 2024-09-23 10:45:40,063 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=241024.0, ans=0.05 2024-09-23 10:46:18,422 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=241164.0, ans=0.0 2024-09-23 10:46:18,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=241164.0, ans=0.0 2024-09-23 10:46:39,306 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=241210.66666666666, ans=0.125 2024-09-23 10:46:42,292 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=241210.66666666666, ans=0.1 2024-09-23 10:46:44,063 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=241210.66666666666, ans=0.125 2024-09-23 10:46:48,999 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=241210.66666666666, ans=0.0 2024-09-23 10:46:52,003 INFO [train.py:1198] (2/4) Epoch 14, batch 1050, loss[loss=0.2104, ctc_loss=0.1409, cr_loss=0.3474, over 17262.00 frames. ], tot_loss[loss=0.2343, ctc_loss=0.1595, cr_loss=0.374, over 3347254.72 frames. ], batch size: 44, lr: 8.69e-03, grad_scale: 32.0 2024-09-23 10:47:10,538 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=9.43 vs. limit=15.0 2024-09-23 10:47:11,532 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=241304.0, ans=0.1 2024-09-23 10:47:19,594 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=241304.0, ans=0.2 2024-09-23 10:47:26,358 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.55 vs. limit=15.0 2024-09-23 10:47:32,537 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=5.276e-03 2024-09-23 10:48:01,449 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=241444.0, ans=0.125 2024-09-23 10:48:06,230 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=241444.0, ans=0.125 2024-09-23 10:48:15,756 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=241490.66666666666, ans=0.125 2024-09-23 10:48:16,987 INFO [train.py:1198] (2/4) Epoch 14, batch 1100, loss[loss=0.2167, ctc_loss=0.1461, cr_loss=0.3534, over 16306.00 frames. ], tot_loss[loss=0.2341, ctc_loss=0.1594, cr_loss=0.3733, over 3350947.61 frames. ], batch size: 36, lr: 8.69e-03, grad_scale: 32.0 2024-09-23 10:48:18,620 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 1.315e+02 1.444e+02 1.614e+02 2.728e+02, threshold=2.888e+02, percent-clipped=0.0 2024-09-23 10:48:22,267 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=241490.66666666666, ans=0.125 2024-09-23 10:48:29,557 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=241490.66666666666, ans=0.0 2024-09-23 10:48:35,697 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=241537.33333333334, ans=0.0 2024-09-23 10:48:50,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=241537.33333333334, ans=0.125 2024-09-23 10:49:22,613 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=241630.66666666666, ans=0.125 2024-09-23 10:49:25,792 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=241677.33333333334, ans=0.0 2024-09-23 10:49:29,179 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=9.40 vs. limit=15.0 2024-09-23 10:49:35,399 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=241677.33333333334, ans=0.2 2024-09-23 10:49:41,261 INFO [train.py:1198] (2/4) Epoch 14, batch 1150, loss[loss=0.2215, ctc_loss=0.1489, cr_loss=0.3627, over 16950.00 frames. ], tot_loss[loss=0.2347, ctc_loss=0.16, cr_loss=0.3733, over 3342169.77 frames. ], batch size: 42, lr: 8.69e-03, grad_scale: 32.0 2024-09-23 10:49:47,039 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=5.67 vs. limit=12.0 2024-09-23 10:49:56,408 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=241770.66666666666, ans=0.0 2024-09-23 10:49:59,866 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.41 vs. limit=15.0 2024-09-23 10:50:34,482 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=241864.0, ans=0.125 2024-09-23 10:51:01,211 INFO [train.py:1198] (2/4) Epoch 14, batch 1200, loss[loss=0.2157, ctc_loss=0.1421, cr_loss=0.3682, over 17099.00 frames. ], tot_loss[loss=0.2343, ctc_loss=0.1598, cr_loss=0.3725, over 3341067.56 frames. ], batch size: 43, lr: 8.68e-03, grad_scale: 32.0 2024-09-23 10:51:02,261 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=12.67 vs. limit=22.5 2024-09-23 10:51:02,790 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.057e+02 1.307e+02 1.418e+02 1.626e+02 2.907e+02, threshold=2.837e+02, percent-clipped=1.0 2024-09-23 10:51:22,293 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=242004.0, ans=0.125 2024-09-23 10:51:31,786 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=242050.66666666666, ans=0.1 2024-09-23 10:51:36,487 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=242050.66666666666, ans=0.0 2024-09-23 10:52:04,194 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.29 vs. limit=15.0 2024-09-23 10:52:20,958 INFO [train.py:1198] (2/4) Epoch 14, batch 1250, loss[loss=0.2077, ctc_loss=0.1408, cr_loss=0.3345, over 17292.00 frames. ], tot_loss[loss=0.2345, ctc_loss=0.1598, cr_loss=0.3731, over 3341551.09 frames. ], batch size: 46, lr: 8.68e-03, grad_scale: 32.0 2024-09-23 10:52:21,191 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=242190.66666666666, ans=0.0 2024-09-23 10:52:37,177 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=242190.66666666666, ans=0.125 2024-09-23 10:52:53,206 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=242237.33333333334, ans=0.1 2024-09-23 10:53:03,129 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=21.30 vs. limit=22.5 2024-09-23 10:53:05,520 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=242284.0, ans=0.1 2024-09-23 10:53:07,203 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=242284.0, ans=0.125 2024-09-23 10:53:08,191 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.89 vs. limit=15.0 2024-09-23 10:53:16,676 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=242330.66666666666, ans=0.1 2024-09-23 10:53:22,438 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=242330.66666666666, ans=0.125 2024-09-23 10:53:42,247 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=242377.33333333334, ans=0.0 2024-09-23 10:53:45,304 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=242377.33333333334, ans=0.025 2024-09-23 10:53:49,818 INFO [train.py:1198] (2/4) Epoch 14, batch 1300, loss[loss=0.2653, ctc_loss=0.1852, cr_loss=0.4001, over 17210.00 frames. ], tot_loss[loss=0.2348, ctc_loss=0.1602, cr_loss=0.3726, over 3340171.19 frames. ], batch size: 55, lr: 8.67e-03, grad_scale: 32.0 2024-09-23 10:53:51,342 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.122e+02 1.264e+02 1.376e+02 1.514e+02 2.274e+02, threshold=2.753e+02, percent-clipped=0.0 2024-09-23 10:53:51,746 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=242424.0, ans=0.125 2024-09-23 10:54:18,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=242470.66666666666, ans=0.07 2024-09-23 10:54:33,525 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=242517.33333333334, ans=0.1 2024-09-23 10:55:02,169 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=242610.66666666666, ans=0.1 2024-09-23 10:55:06,898 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=242610.66666666666, ans=0.1 2024-09-23 10:55:10,058 INFO [train.py:1198] (2/4) Epoch 14, batch 1350, loss[loss=0.2076, ctc_loss=0.1381, cr_loss=0.3478, over 16968.00 frames. ], tot_loss[loss=0.2353, ctc_loss=0.1607, cr_loss=0.3729, over 3342036.44 frames. ], batch size: 42, lr: 8.67e-03, grad_scale: 32.0 2024-09-23 10:55:36,449 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=242704.0, ans=0.125 2024-09-23 10:55:41,245 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=242704.0, ans=0.2 2024-09-23 10:55:42,819 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=242750.66666666666, ans=0.125 2024-09-23 10:56:32,056 INFO [train.py:1198] (2/4) Epoch 14, batch 1400, loss[loss=0.242, ctc_loss=0.1699, cr_loss=0.3603, over 17084.00 frames. ], tot_loss[loss=0.2345, ctc_loss=0.16, cr_loss=0.3726, over 3356216.71 frames. ], batch size: 49, lr: 8.67e-03, grad_scale: 32.0 2024-09-23 10:56:33,636 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.305e+02 1.425e+02 1.607e+02 2.757e+02, threshold=2.850e+02, percent-clipped=1.0 2024-09-23 10:56:43,404 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=242890.66666666666, ans=0.125 2024-09-23 10:56:47,009 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.28 vs. limit=10.0 2024-09-23 10:57:04,172 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=242984.0, ans=0.07 2024-09-23 10:57:52,267 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=243077.33333333334, ans=0.125 2024-09-23 10:57:56,917 INFO [train.py:1198] (2/4) Epoch 14, batch 1450, loss[loss=0.2479, ctc_loss=0.1711, cr_loss=0.3839, over 16994.00 frames. ], tot_loss[loss=0.2346, ctc_loss=0.16, cr_loss=0.373, over 3349499.98 frames. ], batch size: 53, lr: 8.66e-03, grad_scale: 16.0 2024-09-23 10:58:03,385 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=243124.0, ans=0.025 2024-09-23 10:58:08,038 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=243124.0, ans=0.2 2024-09-23 10:58:38,755 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=243217.33333333334, ans=0.125 2024-09-23 10:58:48,587 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=243264.0, ans=0.2 2024-09-23 10:58:55,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys.whitening_limit, batch_count=243264.0, ans=6.0 2024-09-23 10:59:12,601 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=243310.66666666666, ans=0.0 2024-09-23 10:59:21,663 INFO [train.py:1198] (2/4) Epoch 14, batch 1500, loss[loss=0.2554, ctc_loss=0.1767, cr_loss=0.3931, over 14844.00 frames. ], tot_loss[loss=0.2338, ctc_loss=0.1594, cr_loss=0.3724, over 3343238.90 frames. ], batch size: 89, lr: 8.66e-03, grad_scale: 16.0 2024-09-23 10:59:24,853 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.079e+02 1.258e+02 1.373e+02 1.539e+02 2.095e+02, threshold=2.745e+02, percent-clipped=0.0 2024-09-23 10:59:33,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=243357.33333333334, ans=0.0 2024-09-23 10:59:41,035 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=243404.0, ans=0.2 2024-09-23 11:00:37,737 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=6.57 vs. limit=15.0 2024-09-23 11:00:41,712 INFO [train.py:1198] (2/4) Epoch 14, batch 1550, loss[loss=0.2469, ctc_loss=0.1685, cr_loss=0.3921, over 17029.00 frames. ], tot_loss[loss=0.2348, ctc_loss=0.16, cr_loss=0.374, over 3348585.59 frames. ], batch size: 56, lr: 8.65e-03, grad_scale: 16.0 2024-09-23 11:00:56,771 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=243637.33333333334, ans=0.025 2024-09-23 11:01:07,929 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=243637.33333333334, ans=0.125 2024-09-23 11:01:22,227 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=243684.0, ans=0.125 2024-09-23 11:01:30,879 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.07 vs. limit=15.0 2024-09-23 11:01:33,493 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=243730.66666666666, ans=0.125 2024-09-23 11:01:40,611 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=5.20 vs. limit=10.0 2024-09-23 11:01:49,431 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=243777.33333333334, ans=0.0 2024-09-23 11:02:01,665 INFO [train.py:1198] (2/4) Epoch 14, batch 1600, loss[loss=0.2642, ctc_loss=0.1803, cr_loss=0.4199, over 16940.00 frames. ], tot_loss[loss=0.2353, ctc_loss=0.1603, cr_loss=0.3749, over 3354317.56 frames. ], batch size: 58, lr: 8.65e-03, grad_scale: 32.0 2024-09-23 11:02:04,729 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.106e+02 1.281e+02 1.419e+02 1.549e+02 2.365e+02, threshold=2.838e+02, percent-clipped=0.0 2024-09-23 11:02:30,911 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=243870.66666666666, ans=0.125 2024-09-23 11:03:04,024 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=243964.0, ans=0.025 2024-09-23 11:03:26,470 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=244010.66666666666, ans=0.025 2024-09-23 11:03:30,841 INFO [train.py:1198] (2/4) Epoch 14, batch 1650, loss[loss=0.2126, ctc_loss=0.1396, cr_loss=0.3652, over 16300.00 frames. ], tot_loss[loss=0.2339, ctc_loss=0.1594, cr_loss=0.3729, over 3353896.20 frames. ], batch size: 36, lr: 8.64e-03, grad_scale: 32.0 2024-09-23 11:03:35,210 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.whiten, num_groups=1, num_channels=192, metric=5.09 vs. limit=12.0 2024-09-23 11:03:43,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=244057.33333333334, ans=0.125 2024-09-23 11:04:50,835 INFO [train.py:1198] (2/4) Epoch 14, batch 1700, loss[loss=0.2455, ctc_loss=0.1666, cr_loss=0.3945, over 17056.00 frames. ], tot_loss[loss=0.2337, ctc_loss=0.1592, cr_loss=0.3724, over 3348164.76 frames. ], batch size: 52, lr: 8.64e-03, grad_scale: 32.0 2024-09-23 11:04:54,006 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.254e+02 1.382e+02 1.612e+02 3.536e+02, threshold=2.764e+02, percent-clipped=2.0 2024-09-23 11:04:56,452 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=13.23 vs. limit=15.0 2024-09-23 11:05:04,742 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.73 vs. limit=15.0 2024-09-23 11:05:05,588 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=244337.33333333334, ans=0.1 2024-09-23 11:05:15,233 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=244337.33333333334, ans=0.0 2024-09-23 11:05:20,221 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=6.08 vs. limit=15.0 2024-09-23 11:06:10,497 INFO [train.py:1198] (2/4) Epoch 14, batch 1750, loss[loss=0.21, ctc_loss=0.1372, cr_loss=0.364, over 17114.00 frames. ], tot_loss[loss=0.2339, ctc_loss=0.1593, cr_loss=0.3729, over 3348635.75 frames. ], batch size: 40, lr: 8.64e-03, grad_scale: 32.0 2024-09-23 11:06:16,100 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=244524.0, ans=0.1 2024-09-23 11:06:17,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=244524.0, ans=0.125 2024-09-23 11:06:34,040 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=18.18 vs. limit=22.5 2024-09-23 11:06:43,933 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=15.06 vs. limit=22.5 2024-09-23 11:07:09,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=244664.0, ans=0.0 2024-09-23 11:07:36,212 INFO [train.py:1198] (2/4) Epoch 14, batch 1800, loss[loss=0.2644, ctc_loss=0.1819, cr_loss=0.4126, over 17301.00 frames. ], tot_loss[loss=0.2332, ctc_loss=0.1587, cr_loss=0.3724, over 3352025.81 frames. ], batch size: 49, lr: 8.63e-03, grad_scale: 32.0 2024-09-23 11:07:39,492 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.282e+02 1.372e+02 1.529e+02 2.252e+02, threshold=2.745e+02, percent-clipped=0.0 2024-09-23 11:08:14,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=244850.66666666666, ans=0.0 2024-09-23 11:08:32,897 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=244897.33333333334, ans=0.0 2024-09-23 11:09:01,688 INFO [train.py:1198] (2/4) Epoch 14, batch 1850, loss[loss=0.252, ctc_loss=0.1739, cr_loss=0.3902, over 16978.00 frames. ], tot_loss[loss=0.2325, ctc_loss=0.1583, cr_loss=0.3712, over 3357755.41 frames. ], batch size: 58, lr: 8.63e-03, grad_scale: 32.0 2024-09-23 11:09:05,697 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=6.21 vs. limit=15.0 2024-09-23 11:09:27,661 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=245037.33333333334, ans=0.04949747468305833 2024-09-23 11:09:27,817 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 11:09:27,876 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=245037.33333333334, ans=0.1 2024-09-23 11:09:48,125 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=245130.66666666666, ans=0.125 2024-09-23 11:10:01,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=245130.66666666666, ans=0.125 2024-09-23 11:10:01,305 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=245130.66666666666, ans=0.125 2024-09-23 11:10:05,067 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=23.06 vs. limit=22.5 2024-09-23 11:10:21,848 INFO [train.py:1198] (2/4) Epoch 14, batch 1900, loss[loss=0.1948, ctc_loss=0.1286, cr_loss=0.331, over 16917.00 frames. ], tot_loss[loss=0.2326, ctc_loss=0.1583, cr_loss=0.3714, over 3358746.79 frames. ], batch size: 42, lr: 8.62e-03, grad_scale: 32.0 2024-09-23 11:10:25,081 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.260e+02 1.374e+02 1.529e+02 3.130e+02, threshold=2.747e+02, percent-clipped=1.0 2024-09-23 11:10:38,372 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=245270.66666666666, ans=0.1 2024-09-23 11:11:15,094 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=245364.0, ans=0.125 2024-09-23 11:11:18,192 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=245364.0, ans=0.125 2024-09-23 11:11:27,799 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=245410.66666666666, ans=0.125 2024-09-23 11:11:41,462 INFO [train.py:1198] (2/4) Epoch 14, batch 1950, loss[loss=0.2101, ctc_loss=0.1385, cr_loss=0.3578, over 17046.00 frames. ], tot_loss[loss=0.2323, ctc_loss=0.1581, cr_loss=0.3712, over 3350290.50 frames. ], batch size: 46, lr: 8.62e-03, grad_scale: 32.0 2024-09-23 11:11:50,360 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=245457.33333333334, ans=0.1 2024-09-23 11:12:37,119 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.58 vs. limit=15.0 2024-09-23 11:12:43,072 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=245597.33333333334, ans=0.1 2024-09-23 11:13:09,060 INFO [train.py:1198] (2/4) Epoch 14, batch 2000, loss[loss=0.1957, ctc_loss=0.1324, cr_loss=0.3168, over 17277.00 frames. ], tot_loss[loss=0.2328, ctc_loss=0.1584, cr_loss=0.3721, over 3348733.28 frames. ], batch size: 42, lr: 8.62e-03, grad_scale: 32.0 2024-09-23 11:13:14,678 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.297e+02 1.399e+02 1.635e+02 2.518e+02, threshold=2.799e+02, percent-clipped=0.0 2024-09-23 11:14:16,645 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.64 vs. limit=22.5 2024-09-23 11:14:19,553 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=245877.33333333334, ans=0.2 2024-09-23 11:14:22,684 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=245877.33333333334, ans=0.125 2024-09-23 11:14:31,852 INFO [train.py:1198] (2/4) Epoch 14, batch 2050, loss[loss=0.2289, ctc_loss=0.1517, cr_loss=0.386, over 17030.00 frames. ], tot_loss[loss=0.233, ctc_loss=0.1586, cr_loss=0.3722, over 3343858.45 frames. ], batch size: 51, lr: 8.61e-03, grad_scale: 32.0 2024-09-23 11:14:33,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=245924.0, ans=0.125 2024-09-23 11:14:41,027 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=7.67 vs. limit=15.0 2024-09-23 11:15:52,134 INFO [train.py:1198] (2/4) Epoch 14, batch 2100, loss[loss=0.2229, ctc_loss=0.1531, cr_loss=0.3488, over 17018.00 frames. ], tot_loss[loss=0.2325, ctc_loss=0.1581, cr_loss=0.3721, over 3353034.56 frames. ], batch size: 51, lr: 8.61e-03, grad_scale: 32.0 2024-09-23 11:15:55,403 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.001e+02 1.226e+02 1.315e+02 1.406e+02 3.033e+02, threshold=2.629e+02, percent-clipped=1.0 2024-09-23 11:16:00,496 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=246157.33333333334, ans=0.0 2024-09-23 11:16:10,672 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=6.86 vs. limit=15.0 2024-09-23 11:16:53,613 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=246297.33333333334, ans=0.125 2024-09-23 11:17:07,572 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 11:17:10,588 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=246344.0, ans=0.2 2024-09-23 11:17:15,228 INFO [train.py:1198] (2/4) Epoch 14, batch 2150, loss[loss=0.2145, ctc_loss=0.1464, cr_loss=0.3403, over 17149.00 frames. ], tot_loss[loss=0.2314, ctc_loss=0.1573, cr_loss=0.3708, over 3360667.64 frames. ], batch size: 48, lr: 8.60e-03, grad_scale: 32.0 2024-09-23 11:17:36,364 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=246437.33333333334, ans=0.125 2024-09-23 11:17:44,576 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=246437.33333333334, ans=0.125 2024-09-23 11:17:56,277 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.min_abs, batch_count=246484.0, ans=0.5 2024-09-23 11:18:06,562 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.91 vs. limit=15.0 2024-09-23 11:18:21,584 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=246530.66666666666, ans=0.0 2024-09-23 11:18:41,955 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=246624.0, ans=0.2 2024-09-23 11:18:43,399 INFO [train.py:1198] (2/4) Epoch 14, batch 2200, loss[loss=0.2297, ctc_loss=0.1571, cr_loss=0.3633, over 17175.00 frames. ], tot_loss[loss=0.2316, ctc_loss=0.1575, cr_loss=0.3708, over 3360797.73 frames. ], batch size: 45, lr: 8.60e-03, grad_scale: 32.0 2024-09-23 11:18:45,156 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=246624.0, ans=0.1 2024-09-23 11:18:46,518 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.334e+02 1.425e+02 1.540e+02 2.133e+02, threshold=2.850e+02, percent-clipped=0.0 2024-09-23 11:18:46,809 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=246624.0, ans=0.125 2024-09-23 11:18:55,197 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.77 vs. limit=15.0 2024-09-23 11:18:59,504 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=246670.66666666666, ans=0.0 2024-09-23 11:19:15,563 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=246717.33333333334, ans=0.125 2024-09-23 11:19:21,943 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=246717.33333333334, ans=0.125 2024-09-23 11:19:25,287 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=246717.33333333334, ans=0.0 2024-09-23 11:19:41,242 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=246764.0, ans=0.1 2024-09-23 11:20:02,898 INFO [train.py:1198] (2/4) Epoch 14, batch 2250, loss[loss=0.2455, ctc_loss=0.1685, cr_loss=0.3852, over 17306.00 frames. ], tot_loss[loss=0.2319, ctc_loss=0.1579, cr_loss=0.3705, over 3350881.56 frames. ], batch size: 49, lr: 8.60e-03, grad_scale: 16.0 2024-09-23 11:20:22,255 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=246904.0, ans=0.0 2024-09-23 11:20:25,361 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=246904.0, ans=0.09899494936611666 2024-09-23 11:20:49,460 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys.whitening_limit, batch_count=246997.33333333334, ans=6.0 2024-09-23 11:20:54,277 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.51 vs. limit=6.0 2024-09-23 11:21:22,504 INFO [train.py:1198] (2/4) Epoch 14, batch 2300, loss[loss=0.2526, ctc_loss=0.1739, cr_loss=0.3936, over 14882.00 frames. ], tot_loss[loss=0.2318, ctc_loss=0.1578, cr_loss=0.3701, over 3345753.33 frames. ], batch size: 89, lr: 8.59e-03, grad_scale: 16.0 2024-09-23 11:21:27,261 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.092e+02 1.285e+02 1.398e+02 1.588e+02 2.479e+02, threshold=2.795e+02, percent-clipped=0.0 2024-09-23 11:21:37,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=247137.33333333334, ans=0.0 2024-09-23 11:22:39,248 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=9.06 vs. limit=15.0 2024-09-23 11:22:46,661 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=247324.0, ans=0.0 2024-09-23 11:22:50,511 INFO [train.py:1198] (2/4) Epoch 14, batch 2350, loss[loss=0.2565, ctc_loss=0.1742, cr_loss=0.4115, over 17033.00 frames. ], tot_loss[loss=0.2325, ctc_loss=0.1584, cr_loss=0.3706, over 3341819.93 frames. ], batch size: 52, lr: 8.59e-03, grad_scale: 16.0 2024-09-23 11:22:57,816 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=13.28 vs. limit=15.0 2024-09-23 11:23:28,520 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=247417.33333333334, ans=0.07 2024-09-23 11:23:53,987 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=247464.0, ans=0.0 2024-09-23 11:24:03,470 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.min_positive, batch_count=247510.66666666666, ans=0.05 2024-09-23 11:24:12,731 INFO [train.py:1198] (2/4) Epoch 14, batch 2400, loss[loss=0.2757, ctc_loss=0.1898, cr_loss=0.4293, over 17352.00 frames. ], tot_loss[loss=0.2331, ctc_loss=0.1588, cr_loss=0.3714, over 3342411.15 frames. ], batch size: 48, lr: 8.58e-03, grad_scale: 32.0 2024-09-23 11:24:16,928 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=6.85 vs. limit=15.0 2024-09-23 11:24:17,509 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.081e+02 1.242e+02 1.312e+02 1.459e+02 2.054e+02, threshold=2.624e+02, percent-clipped=0.0 2024-09-23 11:24:17,917 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=247557.33333333334, ans=0.0 2024-09-23 11:24:30,816 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=17.30 vs. limit=22.5 2024-09-23 11:24:30,928 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.42 vs. limit=15.0 2024-09-23 11:24:31,965 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=247604.0, ans=0.0 2024-09-23 11:24:39,934 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=247604.0, ans=0.0 2024-09-23 11:25:11,613 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=247697.33333333334, ans=0.025 2024-09-23 11:25:32,181 INFO [train.py:1198] (2/4) Epoch 14, batch 2450, loss[loss=0.2035, ctc_loss=0.1369, cr_loss=0.3331, over 16934.00 frames. ], tot_loss[loss=0.233, ctc_loss=0.1587, cr_loss=0.3714, over 3352765.82 frames. ], batch size: 42, lr: 8.58e-03, grad_scale: 32.0 2024-09-23 11:25:40,777 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=247790.66666666666, ans=0.0 2024-09-23 11:25:47,108 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=247837.33333333334, ans=0.125 2024-09-23 11:25:50,324 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=247837.33333333334, ans=0.125 2024-09-23 11:25:50,374 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=247837.33333333334, ans=0.125 2024-09-23 11:26:08,071 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.28 vs. limit=15.0 2024-09-23 11:26:14,224 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=247884.0, ans=0.1 2024-09-23 11:26:18,968 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=247930.66666666666, ans=0.125 2024-09-23 11:26:20,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=247930.66666666666, ans=0.125 2024-09-23 11:26:29,135 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=8.95 vs. limit=15.0 2024-09-23 11:26:36,505 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.min_positive, batch_count=247977.33333333334, ans=0.025 2024-09-23 11:26:46,219 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=247977.33333333334, ans=0.125 2024-09-23 11:26:54,747 INFO [train.py:1198] (2/4) Epoch 14, batch 2500, loss[loss=0.2128, ctc_loss=0.1419, cr_loss=0.3548, over 17296.00 frames. ], tot_loss[loss=0.2322, ctc_loss=0.1581, cr_loss=0.3705, over 3350993.25 frames. ], batch size: 46, lr: 8.58e-03, grad_scale: 32.0 2024-09-23 11:26:55,049 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=248024.0, ans=0.0 2024-09-23 11:26:59,531 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.311e+02 1.464e+02 1.674e+02 2.701e+02, threshold=2.928e+02, percent-clipped=1.0 2024-09-23 11:27:02,933 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.min_positive, batch_count=248024.0, ans=0.025 2024-09-23 11:27:54,883 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=3.93 vs. limit=15.0 2024-09-23 11:28:00,182 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 11:28:09,592 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=248210.66666666666, ans=0.125 2024-09-23 11:28:17,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=248210.66666666666, ans=0.0 2024-09-23 11:28:21,882 INFO [train.py:1198] (2/4) Epoch 14, batch 2550, loss[loss=0.2155, ctc_loss=0.148, cr_loss=0.3375, over 17215.00 frames. ], tot_loss[loss=0.2311, ctc_loss=0.1573, cr_loss=0.3693, over 3359817.56 frames. ], batch size: 47, lr: 8.57e-03, grad_scale: 32.0 2024-09-23 11:28:59,016 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=248350.66666666666, ans=0.2 2024-09-23 11:28:59,028 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=248350.66666666666, ans=0.0 2024-09-23 11:29:05,302 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=248350.66666666666, ans=0.07 2024-09-23 11:29:37,526 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=248444.0, ans=0.2 2024-09-23 11:29:42,049 INFO [train.py:1198] (2/4) Epoch 14, batch 2600, loss[loss=0.2104, ctc_loss=0.1444, cr_loss=0.33, over 17002.00 frames. ], tot_loss[loss=0.2309, ctc_loss=0.1571, cr_loss=0.3688, over 3370747.58 frames. ], batch size: 51, lr: 8.57e-03, grad_scale: 32.0 2024-09-23 11:29:43,932 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=248490.66666666666, ans=0.0 2024-09-23 11:29:46,742 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.321e+02 1.442e+02 1.644e+02 2.368e+02, threshold=2.883e+02, percent-clipped=0.0 2024-09-23 11:29:47,165 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 11:29:50,520 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.89 vs. limit=15.0 2024-09-23 11:29:59,823 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=248537.33333333334, ans=0.125 2024-09-23 11:30:15,896 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=248584.0, ans=0.0 2024-09-23 11:31:02,246 INFO [train.py:1198] (2/4) Epoch 14, batch 2650, loss[loss=0.2507, ctc_loss=0.171, cr_loss=0.3985, over 17028.00 frames. ], tot_loss[loss=0.2303, ctc_loss=0.1564, cr_loss=0.3695, over 3375715.15 frames. ], batch size: 56, lr: 8.56e-03, grad_scale: 32.0 2024-09-23 11:31:07,299 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=248724.0, ans=0.125 2024-09-23 11:31:15,687 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=20.16 vs. limit=22.5 2024-09-23 11:31:45,499 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=248817.33333333334, ans=0.125 2024-09-23 11:32:25,978 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=20.97 vs. limit=22.5 2024-09-23 11:32:26,707 INFO [train.py:1198] (2/4) Epoch 14, batch 2700, loss[loss=0.2357, ctc_loss=0.1594, cr_loss=0.3815, over 17018.00 frames. ], tot_loss[loss=0.231, ctc_loss=0.157, cr_loss=0.3703, over 3376799.13 frames. ], batch size: 51, lr: 8.56e-03, grad_scale: 32.0 2024-09-23 11:32:28,444 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=248957.33333333334, ans=0.015 2024-09-23 11:32:31,436 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.165e+02 1.327e+02 1.447e+02 1.619e+02 2.182e+02, threshold=2.895e+02, percent-clipped=0.0 2024-09-23 11:32:36,536 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=248957.33333333334, ans=0.125 2024-09-23 11:33:05,797 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.09 vs. limit=15.0 2024-09-23 11:33:16,803 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=249050.66666666666, ans=0.125 2024-09-23 11:33:18,314 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=249097.33333333334, ans=0.2 2024-09-23 11:33:29,534 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=249097.33333333334, ans=0.125 2024-09-23 11:33:46,879 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=249144.0, ans=0.125 2024-09-23 11:33:48,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=249144.0, ans=0.1 2024-09-23 11:33:51,509 INFO [train.py:1198] (2/4) Epoch 14, batch 2750, loss[loss=0.1931, ctc_loss=0.1278, cr_loss=0.3268, over 16962.00 frames. ], tot_loss[loss=0.2317, ctc_loss=0.1575, cr_loss=0.3709, over 3376546.52 frames. ], batch size: 42, lr: 8.56e-03, grad_scale: 32.0 2024-09-23 11:33:52,400 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=9.85 vs. limit=15.0 2024-09-23 11:34:17,598 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=249237.33333333334, ans=0.2 2024-09-23 11:34:25,519 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=249284.0, ans=0.0 2024-09-23 11:34:28,575 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=249284.0, ans=10.0 2024-09-23 11:34:45,957 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=249330.66666666666, ans=0.125 2024-09-23 11:35:10,882 INFO [train.py:1198] (2/4) Epoch 14, batch 2800, loss[loss=0.2329, ctc_loss=0.1634, cr_loss=0.3473, over 17287.00 frames. ], tot_loss[loss=0.2314, ctc_loss=0.1573, cr_loss=0.3707, over 3376073.50 frames. ], batch size: 46, lr: 8.55e-03, grad_scale: 32.0 2024-09-23 11:35:15,651 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.308e+02 1.382e+02 1.526e+02 2.267e+02, threshold=2.765e+02, percent-clipped=0.0 2024-09-23 11:35:34,951 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=249470.66666666666, ans=0.125 2024-09-23 11:35:37,978 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=249470.66666666666, ans=0.0 2024-09-23 11:35:44,275 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=249517.33333333334, ans=0.125 2024-09-23 11:36:03,912 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=249564.0, ans=0.125 2024-09-23 11:36:12,343 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=3.21 vs. limit=15.0 2024-09-23 11:36:26,495 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=249610.66666666666, ans=0.125 2024-09-23 11:36:31,114 INFO [train.py:1198] (2/4) Epoch 14, batch 2850, loss[loss=0.235, ctc_loss=0.159, cr_loss=0.3798, over 17107.00 frames. ], tot_loss[loss=0.2312, ctc_loss=0.1572, cr_loss=0.3699, over 3366613.79 frames. ], batch size: 49, lr: 8.55e-03, grad_scale: 16.0 2024-09-23 11:36:33,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=249657.33333333334, ans=0.0 2024-09-23 11:36:41,319 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.29 vs. limit=15.0 2024-09-23 11:36:53,276 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=249704.0, ans=0.125 2024-09-23 11:37:13,637 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 11:37:32,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=249797.33333333334, ans=0.125 2024-09-23 11:37:43,583 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=249844.0, ans=0.0 2024-09-23 11:37:58,080 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=12.84 vs. limit=15.0 2024-09-23 11:38:01,839 INFO [train.py:1198] (2/4) Epoch 14, batch 2900, loss[loss=0.2674, ctc_loss=0.1815, cr_loss=0.4293, over 17297.00 frames. ], tot_loss[loss=0.2324, ctc_loss=0.1581, cr_loss=0.3714, over 3351272.91 frames. ], batch size: 49, lr: 8.55e-03, grad_scale: 16.0 2024-09-23 11:38:08,318 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.270e+02 1.418e+02 1.645e+02 2.792e+02, threshold=2.835e+02, percent-clipped=1.0 2024-09-23 11:38:13,317 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=249890.66666666666, ans=0.125 2024-09-23 11:38:49,684 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=250030.66666666666, ans=0.125 2024-09-23 11:38:57,706 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=250030.66666666666, ans=0.125 2024-09-23 11:39:01,776 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.50 vs. limit=15.0 2024-09-23 11:39:15,239 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=250077.33333333334, ans=0.025 2024-09-23 11:39:21,419 INFO [train.py:1198] (2/4) Epoch 14, batch 2950, loss[loss=0.244, ctc_loss=0.1679, cr_loss=0.3805, over 17292.00 frames. ], tot_loss[loss=0.2328, ctc_loss=0.1584, cr_loss=0.3718, over 3355815.28 frames. ], batch size: 49, lr: 8.54e-03, grad_scale: 16.0 2024-09-23 11:40:07,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=250264.0, ans=0.125 2024-09-23 11:40:20,187 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=250264.0, ans=0.125 2024-09-23 11:40:28,699 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.whiten.whitening_limit, batch_count=250310.66666666666, ans=12.0 2024-09-23 11:40:40,235 INFO [train.py:1198] (2/4) Epoch 14, batch 3000, loss[loss=0.2425, ctc_loss=0.1648, cr_loss=0.3888, over 17203.00 frames. ], tot_loss[loss=0.2343, ctc_loss=0.1595, cr_loss=0.3738, over 3350844.09 frames. ], batch size: 47, lr: 8.54e-03, grad_scale: 16.0 2024-09-23 11:40:40,236 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 11:40:55,473 INFO [train.py:1230] (2/4) Epoch 14, validation: loss=0.04331, ctc_loss=0.04331, cr_loss=7.532e-15, over 944034.00 frames. 2024-09-23 11:40:55,474 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 11:41:01,567 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.302e+02 1.389e+02 1.457e+02 1.974e+02, threshold=2.778e+02, percent-clipped=0.0 2024-09-23 11:41:03,341 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=250357.33333333334, ans=0.04949747468305833 2024-09-23 11:41:06,474 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=250357.33333333334, ans=0.04949747468305833 2024-09-23 11:41:43,185 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=250497.33333333334, ans=0.125 2024-09-23 11:41:50,884 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=250497.33333333334, ans=0.125 2024-09-23 11:41:52,498 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer_ff3.min_abs, batch_count=250497.33333333334, ans=0.2 2024-09-23 11:41:53,891 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=250497.33333333334, ans=0.1 2024-09-23 11:42:03,363 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=250544.0, ans=0.0 2024-09-23 11:42:03,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=250544.0, ans=0.125 2024-09-23 11:42:06,642 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=250544.0, ans=0.125 2024-09-23 11:42:13,879 INFO [train.py:1198] (2/4) Epoch 14, batch 3050, loss[loss=0.315, ctc_loss=0.2281, cr_loss=0.4347, over 11894.00 frames. ], tot_loss[loss=0.2325, ctc_loss=0.1582, cr_loss=0.3718, over 3350281.83 frames. ], batch size: 123, lr: 8.53e-03, grad_scale: 16.0 2024-09-23 11:42:46,968 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=250684.0, ans=0.0 2024-09-23 11:43:15,274 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=250777.33333333334, ans=0.0 2024-09-23 11:43:31,714 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=250777.33333333334, ans=0.1 2024-09-23 11:43:34,614 INFO [train.py:1198] (2/4) Epoch 14, batch 3100, loss[loss=0.2179, ctc_loss=0.1459, cr_loss=0.3597, over 16952.00 frames. ], tot_loss[loss=0.2324, ctc_loss=0.1581, cr_loss=0.3713, over 3340291.14 frames. ], batch size: 42, lr: 8.53e-03, grad_scale: 16.0 2024-09-23 11:43:36,416 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.min_positive, batch_count=250824.0, ans=0.025 2024-09-23 11:43:36,507 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=250824.0, ans=0.125 2024-09-23 11:43:40,946 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.263e+02 1.328e+02 1.443e+02 2.080e+02, threshold=2.656e+02, percent-clipped=0.0 2024-09-23 11:44:55,827 INFO [train.py:1198] (2/4) Epoch 14, batch 3150, loss[loss=0.193, ctc_loss=0.1289, cr_loss=0.3207, over 16944.00 frames. ], tot_loss[loss=0.2314, ctc_loss=0.1573, cr_loss=0.3703, over 3345126.01 frames. ], batch size: 42, lr: 8.53e-03, grad_scale: 16.0 2024-09-23 11:45:01,038 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=251057.33333333334, ans=0.125 2024-09-23 11:45:01,041 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=251057.33333333334, ans=0.0 2024-09-23 11:46:18,649 INFO [train.py:1198] (2/4) Epoch 14, batch 3200, loss[loss=0.263, ctc_loss=0.1776, cr_loss=0.4274, over 17223.00 frames. ], tot_loss[loss=0.2316, ctc_loss=0.1575, cr_loss=0.3703, over 3340177.67 frames. ], batch size: 55, lr: 8.52e-03, grad_scale: 32.0 2024-09-23 11:46:20,517 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=251290.66666666666, ans=0.1 2024-09-23 11:46:24,743 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.075e+02 1.267e+02 1.361e+02 1.514e+02 1.918e+02, threshold=2.723e+02, percent-clipped=0.0 2024-09-23 11:47:02,284 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=251384.0, ans=0.0 2024-09-23 11:47:05,390 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=251430.66666666666, ans=0.0 2024-09-23 11:47:25,673 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=251477.33333333334, ans=0.0 2024-09-23 11:47:36,177 INFO [train.py:1198] (2/4) Epoch 14, batch 3250, loss[loss=0.2779, ctc_loss=0.1963, cr_loss=0.4079, over 11393.00 frames. ], tot_loss[loss=0.232, ctc_loss=0.1579, cr_loss=0.3708, over 3341851.00 frames. ], batch size: 123, lr: 8.52e-03, grad_scale: 32.0 2024-09-23 11:48:01,425 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=251570.66666666666, ans=0.0 2024-09-23 11:48:18,669 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=251617.33333333334, ans=0.125 2024-09-23 11:48:38,989 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=251710.66666666666, ans=0.04949747468305833 2024-09-23 11:48:42,040 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=251710.66666666666, ans=0.2 2024-09-23 11:48:46,620 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=251710.66666666666, ans=0.125 2024-09-23 11:48:46,694 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=251710.66666666666, ans=0.0 2024-09-23 11:48:49,927 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=12.28 vs. limit=22.5 2024-09-23 11:48:54,096 INFO [train.py:1198] (2/4) Epoch 14, batch 3300, loss[loss=0.2167, ctc_loss=0.1424, cr_loss=0.3712, over 16927.00 frames. ], tot_loss[loss=0.2315, ctc_loss=0.1574, cr_loss=0.3707, over 3352287.61 frames. ], batch size: 42, lr: 8.51e-03, grad_scale: 32.0 2024-09-23 11:49:00,432 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.139e+02 1.303e+02 1.410e+02 1.606e+02 3.318e+02, threshold=2.819e+02, percent-clipped=1.0 2024-09-23 11:49:00,730 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=251757.33333333334, ans=0.125 2024-09-23 11:49:14,788 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=251804.0, ans=0.0 2024-09-23 11:49:16,368 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=251804.0, ans=0.125 2024-09-23 11:49:19,692 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=251804.0, ans=0.2 2024-09-23 11:49:35,858 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.45 vs. limit=10.0 2024-09-23 11:50:03,232 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=251944.0, ans=0.025 2024-09-23 11:50:12,288 INFO [train.py:1198] (2/4) Epoch 14, batch 3350, loss[loss=0.2375, ctc_loss=0.1608, cr_loss=0.3833, over 17001.00 frames. ], tot_loss[loss=0.2325, ctc_loss=0.1582, cr_loss=0.3716, over 3356018.97 frames. ], batch size: 51, lr: 8.51e-03, grad_scale: 16.0 2024-09-23 11:50:15,853 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=251990.66666666666, ans=0.125 2024-09-23 11:50:36,279 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.min_abs, batch_count=252037.33333333334, ans=0.5 2024-09-23 11:50:39,442 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=252037.33333333334, ans=0.125 2024-09-23 11:51:08,958 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=252130.66666666666, ans=0.1 2024-09-23 11:51:20,633 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=12.03 vs. limit=22.5 2024-09-23 11:51:30,492 INFO [train.py:1198] (2/4) Epoch 14, batch 3400, loss[loss=0.1878, ctc_loss=0.1235, cr_loss=0.3217, over 16953.00 frames. ], tot_loss[loss=0.2326, ctc_loss=0.1582, cr_loss=0.3721, over 3354866.90 frames. ], batch size: 42, lr: 8.51e-03, grad_scale: 16.0 2024-09-23 11:51:31,489 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=13.25 vs. limit=15.0 2024-09-23 11:51:38,161 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.094e+02 1.275e+02 1.402e+02 1.543e+02 4.509e+02, threshold=2.804e+02, percent-clipped=1.0 2024-09-23 11:51:52,739 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=252270.66666666666, ans=0.125 2024-09-23 11:51:52,786 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=252270.66666666666, ans=0.125 2024-09-23 11:52:00,328 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=252317.33333333334, ans=0.125 2024-09-23 11:52:11,277 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=252317.33333333334, ans=0.125 2024-09-23 11:52:48,079 INFO [train.py:1198] (2/4) Epoch 14, batch 3450, loss[loss=0.2385, ctc_loss=0.1593, cr_loss=0.3961, over 17296.00 frames. ], tot_loss[loss=0.2331, ctc_loss=0.1586, cr_loss=0.3726, over 3346412.25 frames. ], batch size: 51, lr: 8.50e-03, grad_scale: 16.0 2024-09-23 11:53:00,773 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=252457.33333333334, ans=0.1 2024-09-23 11:53:08,786 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=252504.0, ans=0.0 2024-09-23 11:53:24,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=252550.66666666666, ans=0.0 2024-09-23 11:53:56,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=252644.0, ans=0.125 2024-09-23 11:54:03,952 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=252644.0, ans=0.0 2024-09-23 11:54:08,250 INFO [train.py:1198] (2/4) Epoch 14, batch 3500, loss[loss=0.2536, ctc_loss=0.175, cr_loss=0.3931, over 16961.00 frames. ], tot_loss[loss=0.2322, ctc_loss=0.1578, cr_loss=0.3719, over 3354141.27 frames. ], batch size: 58, lr: 8.50e-03, grad_scale: 16.0 2024-09-23 11:54:18,110 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.252e+02 1.352e+02 1.458e+02 2.935e+02, threshold=2.705e+02, percent-clipped=1.0 2024-09-23 11:54:24,520 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=252737.33333333334, ans=0.125 2024-09-23 11:54:57,249 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=252830.66666666666, ans=0.05 2024-09-23 11:55:04,105 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=11.58 vs. limit=15.0 2024-09-23 11:55:12,889 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=252877.33333333334, ans=0.1 2024-09-23 11:55:31,506 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.63 vs. limit=6.0 2024-09-23 11:55:32,279 INFO [train.py:1198] (2/4) Epoch 14, batch 3550, loss[loss=0.2765, ctc_loss=0.1899, cr_loss=0.4331, over 17233.00 frames. ], tot_loss[loss=0.2329, ctc_loss=0.1584, cr_loss=0.3728, over 3350802.58 frames. ], batch size: 55, lr: 8.49e-03, grad_scale: 16.0 2024-09-23 11:55:34,238 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=252924.0, ans=0.125 2024-09-23 11:55:51,666 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=252970.66666666666, ans=0.0 2024-09-23 11:56:22,953 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=253064.0, ans=0.125 2024-09-23 11:56:51,145 INFO [train.py:1198] (2/4) Epoch 14, batch 3600, loss[loss=0.2231, ctc_loss=0.1466, cr_loss=0.3826, over 17212.00 frames. ], tot_loss[loss=0.2317, ctc_loss=0.1575, cr_loss=0.3709, over 3354890.37 frames. ], batch size: 47, lr: 8.49e-03, grad_scale: 32.0 2024-09-23 11:56:58,882 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.037e+02 1.239e+02 1.349e+02 1.491e+02 2.999e+02, threshold=2.699e+02, percent-clipped=1.0 2024-09-23 11:56:59,240 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=253157.33333333334, ans=0.125 2024-09-23 11:57:19,122 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer_ff3.min_abs, batch_count=253204.0, ans=0.2 2024-09-23 11:57:25,276 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=253250.66666666666, ans=0.1 2024-09-23 11:58:04,844 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 11:58:06,860 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.79 vs. limit=12.0 2024-09-23 11:58:08,967 INFO [train.py:1198] (2/4) Epoch 14, batch 3650, loss[loss=0.2573, ctc_loss=0.1771, cr_loss=0.4013, over 16401.00 frames. ], tot_loss[loss=0.2318, ctc_loss=0.1576, cr_loss=0.3709, over 3343794.16 frames. ], batch size: 66, lr: 8.49e-03, grad_scale: 32.0 2024-09-23 11:58:15,447 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=253390.66666666666, ans=0.125 2024-09-23 11:58:24,919 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=253437.33333333334, ans=0.125 2024-09-23 11:58:25,032 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=253437.33333333334, ans=0.125 2024-09-23 11:58:28,219 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=253437.33333333334, ans=0.125 2024-09-23 11:58:42,127 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=13.17 vs. limit=15.0 2024-09-23 11:58:51,446 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=6.16 vs. limit=15.0 2024-09-23 11:59:28,443 INFO [train.py:1198] (2/4) Epoch 14, batch 3700, loss[loss=0.209, ctc_loss=0.1386, cr_loss=0.3518, over 16966.00 frames. ], tot_loss[loss=0.23, ctc_loss=0.1562, cr_loss=0.3691, over 3356035.73 frames. ], batch size: 42, lr: 8.48e-03, grad_scale: 32.0 2024-09-23 11:59:36,290 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.294e+02 1.387e+02 1.607e+02 1.987e+02, threshold=2.774e+02, percent-clipped=0.0 2024-09-23 11:59:49,041 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=253670.66666666666, ans=0.1 2024-09-23 12:00:36,066 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=253810.66666666666, ans=0.125 2024-09-23 12:00:46,835 INFO [train.py:1198] (2/4) Epoch 14, batch 3750, loss[loss=0.2105, ctc_loss=0.1423, cr_loss=0.3411, over 17025.00 frames. ], tot_loss[loss=0.2322, ctc_loss=0.158, cr_loss=0.3708, over 3333808.58 frames. ], batch size: 51, lr: 8.48e-03, grad_scale: 32.0 2024-09-23 12:00:58,366 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=253857.33333333334, ans=0.2 2024-09-23 12:01:18,397 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=253950.66666666666, ans=0.0 2024-09-23 12:02:05,058 INFO [train.py:1198] (2/4) Epoch 14, batch 3800, loss[loss=0.2562, ctc_loss=0.1749, cr_loss=0.4069, over 17015.00 frames. ], tot_loss[loss=0.2333, ctc_loss=0.1589, cr_loss=0.372, over 3310971.15 frames. ], batch size: 51, lr: 8.48e-03, grad_scale: 32.0 2024-09-23 12:02:08,471 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=254090.66666666666, ans=0.125 2024-09-23 12:02:13,078 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.174e+02 1.380e+02 1.551e+02 1.777e+02 3.575e+02, threshold=3.102e+02, percent-clipped=2.0 2024-09-23 12:03:01,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=254230.66666666666, ans=0.125 2024-09-23 12:03:02,917 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=254230.66666666666, ans=0.125 2024-09-23 12:03:09,058 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=254277.33333333334, ans=0.0 2024-09-23 12:03:21,388 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=254277.33333333334, ans=0.125 2024-09-23 12:03:24,331 INFO [train.py:1198] (2/4) Epoch 14, batch 3850, loss[loss=0.287, ctc_loss=0.207, cr_loss=0.3999, over 11975.00 frames. ], tot_loss[loss=0.2355, ctc_loss=0.1606, cr_loss=0.3745, over 3290428.58 frames. ], batch size: 125, lr: 8.47e-03, grad_scale: 32.0 2024-09-23 12:03:45,505 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=254370.66666666666, ans=0.125 2024-09-23 12:04:10,250 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=15.59 vs. limit=15.0 2024-09-23 12:04:31,351 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=254510.66666666666, ans=0.0 2024-09-23 12:05:28,602 INFO [train.py:1198] (2/4) Epoch 15, batch 0, loss[loss=0.2229, ctc_loss=0.1518, cr_loss=0.3558, over 17260.00 frames. ], tot_loss[loss=0.2229, ctc_loss=0.1518, cr_loss=0.3558, over 17260.00 frames. ], batch size: 44, lr: 8.18e-03, grad_scale: 32.0 2024-09-23 12:05:28,603 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 12:05:46,326 INFO [train.py:1230] (2/4) Epoch 15, validation: loss=0.0431, ctc_loss=0.0431, cr_loss=7.486e-15, over 944034.00 frames. 2024-09-23 12:05:46,327 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 12:05:46,752 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=254538.66666666666, ans=0.125 2024-09-23 12:05:49,734 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=254538.66666666666, ans=0.1 2024-09-23 12:05:58,013 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=254538.66666666666, ans=0.125 2024-09-23 12:06:00,796 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.387e+02 1.561e+02 1.706e+02 2.670e+02, threshold=3.121e+02, percent-clipped=0.0 2024-09-23 12:06:40,598 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=254678.66666666666, ans=0.07 2024-09-23 12:06:46,904 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=254678.66666666666, ans=0.0 2024-09-23 12:06:58,036 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=254725.33333333334, ans=0.0 2024-09-23 12:07:02,764 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=254725.33333333334, ans=0.2 2024-09-23 12:07:05,543 INFO [train.py:1198] (2/4) Epoch 15, batch 50, loss[loss=0.2243, ctc_loss=0.1534, cr_loss=0.3541, over 17304.00 frames. ], tot_loss[loss=0.2374, ctc_loss=0.1619, cr_loss=0.3777, over 745230.01 frames. ], batch size: 46, lr: 8.18e-03, grad_scale: 32.0 2024-09-23 12:07:24,740 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.50 vs. limit=15.0 2024-09-23 12:07:42,175 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.02 vs. limit=10.0 2024-09-23 12:07:45,037 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=254865.33333333334, ans=0.2 2024-09-23 12:07:46,507 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=254865.33333333334, ans=0.04949747468305833 2024-09-23 12:08:00,969 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.07 vs. limit=6.0 2024-09-23 12:08:01,971 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=254912.0, ans=0.125 2024-09-23 12:08:12,880 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=254958.66666666666, ans=0.125 2024-09-23 12:08:22,443 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=254958.66666666666, ans=0.0 2024-09-23 12:08:28,617 INFO [train.py:1198] (2/4) Epoch 15, batch 100, loss[loss=0.2342, ctc_loss=0.1597, cr_loss=0.3725, over 17216.00 frames. ], tot_loss[loss=0.2332, ctc_loss=0.1584, cr_loss=0.3742, over 1322283.47 frames. ], batch size: 50, lr: 8.17e-03, grad_scale: 32.0 2024-09-23 12:08:42,830 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 9.954e+01 1.228e+02 1.306e+02 1.476e+02 1.867e+02, threshold=2.613e+02, percent-clipped=0.0 2024-09-23 12:08:43,112 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=255052.0, ans=0.125 2024-09-23 12:09:18,155 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=255145.33333333334, ans=0.125 2024-09-23 12:09:47,783 INFO [train.py:1198] (2/4) Epoch 15, batch 150, loss[loss=0.2339, ctc_loss=0.1592, cr_loss=0.3732, over 17161.00 frames. ], tot_loss[loss=0.2334, ctc_loss=0.1586, cr_loss=0.3743, over 1774709.26 frames. ], batch size: 45, lr: 8.17e-03, grad_scale: 32.0 2024-09-23 12:09:48,201 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=255238.66666666666, ans=0.0 2024-09-23 12:10:01,013 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=255238.66666666666, ans=0.125 2024-09-23 12:10:54,104 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=255378.66666666666, ans=0.0 2024-09-23 12:10:54,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=255378.66666666666, ans=0.125 2024-09-23 12:11:14,511 INFO [train.py:1198] (2/4) Epoch 15, batch 200, loss[loss=0.2426, ctc_loss=0.1644, cr_loss=0.3906, over 17210.00 frames. ], tot_loss[loss=0.2321, ctc_loss=0.1575, cr_loss=0.373, over 2129162.69 frames. ], batch size: 50, lr: 8.16e-03, grad_scale: 32.0 2024-09-23 12:11:28,889 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.082e+02 1.243e+02 1.308e+02 1.422e+02 1.839e+02, threshold=2.616e+02, percent-clipped=0.0 2024-09-23 12:11:29,668 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=17.98 vs. limit=22.5 2024-09-23 12:11:34,045 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=255518.66666666666, ans=0.125 2024-09-23 12:11:52,955 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=255565.33333333334, ans=0.125 2024-09-23 12:11:56,694 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=512, metric=2.50 vs. limit=15.0 2024-09-23 12:12:33,973 INFO [train.py:1198] (2/4) Epoch 15, batch 250, loss[loss=0.2286, ctc_loss=0.1573, cr_loss=0.3566, over 17033.00 frames. ], tot_loss[loss=0.2334, ctc_loss=0.1585, cr_loss=0.3741, over 2392829.12 frames. ], batch size: 39, lr: 8.16e-03, grad_scale: 32.0 2024-09-23 12:12:40,545 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=255705.33333333334, ans=0.04949747468305833 2024-09-23 12:12:45,854 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.63 vs. limit=15.0 2024-09-23 12:13:02,738 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=255752.0, ans=0.2 2024-09-23 12:13:21,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=255798.66666666666, ans=0.04949747468305833 2024-09-23 12:13:39,303 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=255892.0, ans=0.1 2024-09-23 12:13:56,717 INFO [train.py:1198] (2/4) Epoch 15, batch 300, loss[loss=0.2382, ctc_loss=0.1616, cr_loss=0.3831, over 17245.00 frames. ], tot_loss[loss=0.2324, ctc_loss=0.1578, cr_loss=0.3728, over 2609541.06 frames. ], batch size: 44, lr: 8.16e-03, grad_scale: 32.0 2024-09-23 12:14:10,832 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.156e+02 1.341e+02 1.470e+02 1.683e+02 2.993e+02, threshold=2.941e+02, percent-clipped=1.0 2024-09-23 12:14:31,494 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=256032.0, ans=0.125 2024-09-23 12:14:56,775 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=256078.66666666666, ans=0.125 2024-09-23 12:15:00,756 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.whiten, num_groups=1, num_channels=192, metric=4.95 vs. limit=12.0 2024-09-23 12:15:04,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=256125.33333333334, ans=0.0 2024-09-23 12:15:21,580 INFO [train.py:1198] (2/4) Epoch 15, batch 350, loss[loss=0.2445, ctc_loss=0.1674, cr_loss=0.3854, over 16935.00 frames. ], tot_loss[loss=0.2335, ctc_loss=0.1588, cr_loss=0.3735, over 2755973.76 frames. ], batch size: 58, lr: 8.15e-03, grad_scale: 32.0 2024-09-23 12:16:03,283 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=11.54 vs. limit=15.0 2024-09-23 12:16:09,818 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.18 vs. limit=15.0 2024-09-23 12:16:12,405 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=256312.0, ans=0.125 2024-09-23 12:16:22,122 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=256312.0, ans=0.2 2024-09-23 12:16:29,893 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=256358.66666666666, ans=0.0 2024-09-23 12:16:44,049 INFO [train.py:1198] (2/4) Epoch 15, batch 400, loss[loss=0.1715, ctc_loss=0.1146, cr_loss=0.2842, over 17119.00 frames. ], tot_loss[loss=0.2327, ctc_loss=0.1582, cr_loss=0.3727, over 2884924.39 frames. ], batch size: 40, lr: 8.15e-03, grad_scale: 32.0 2024-09-23 12:16:52,195 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=256405.33333333334, ans=0.125 2024-09-23 12:16:58,142 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.238e+02 1.377e+02 1.544e+02 2.269e+02, threshold=2.754e+02, percent-clipped=0.0 2024-09-23 12:17:15,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=256498.66666666666, ans=0.125 2024-09-23 12:17:30,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=256545.33333333334, ans=0.1 2024-09-23 12:17:49,521 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=256592.0, ans=0.125 2024-09-23 12:17:54,101 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=256592.0, ans=0.2 2024-09-23 12:17:54,110 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=256592.0, ans=0.0 2024-09-23 12:17:55,620 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=256592.0, ans=0.125 2024-09-23 12:18:06,461 INFO [train.py:1198] (2/4) Epoch 15, batch 450, loss[loss=0.2473, ctc_loss=0.1687, cr_loss=0.3933, over 16913.00 frames. ], tot_loss[loss=0.2322, ctc_loss=0.1576, cr_loss=0.3727, over 2998234.89 frames. ], batch size: 58, lr: 8.15e-03, grad_scale: 32.0 2024-09-23 12:18:36,350 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.02 vs. limit=15.0 2024-09-23 12:18:40,738 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=256732.0, ans=0.95 2024-09-23 12:18:40,771 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=256732.0, ans=0.125 2024-09-23 12:19:04,852 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=256778.66666666666, ans=0.1 2024-09-23 12:19:07,307 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=3.48 vs. limit=15.0 2024-09-23 12:19:17,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=256825.33333333334, ans=0.0 2024-09-23 12:19:27,187 INFO [train.py:1198] (2/4) Epoch 15, batch 500, loss[loss=0.2021, ctc_loss=0.1359, cr_loss=0.331, over 17103.00 frames. ], tot_loss[loss=0.2315, ctc_loss=0.157, cr_loss=0.3727, over 3086308.38 frames. ], batch size: 40, lr: 8.14e-03, grad_scale: 32.0 2024-09-23 12:19:39,045 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 12:19:41,901 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.095e+02 1.305e+02 1.439e+02 1.681e+02 2.242e+02, threshold=2.879e+02, percent-clipped=0.0 2024-09-23 12:20:13,147 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=256965.33333333334, ans=0.1 2024-09-23 12:20:16,337 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=256965.33333333334, ans=0.1 2024-09-23 12:20:16,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=256965.33333333334, ans=0.125 2024-09-23 12:20:19,610 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=257012.0, ans=0.0 2024-09-23 12:20:55,117 INFO [train.py:1198] (2/4) Epoch 15, batch 550, loss[loss=0.3079, ctc_loss=0.224, cr_loss=0.4197, over 11627.00 frames. ], tot_loss[loss=0.2297, ctc_loss=0.1558, cr_loss=0.3699, over 3134008.67 frames. ], batch size: 123, lr: 8.14e-03, grad_scale: 32.0 2024-09-23 12:21:00,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=257105.33333333334, ans=0.125 2024-09-23 12:21:08,002 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=257105.33333333334, ans=0.2 2024-09-23 12:21:12,821 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=257152.0, ans=0.0 2024-09-23 12:21:12,901 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=257152.0, ans=0.2 2024-09-23 12:21:17,733 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=257152.0, ans=0.2 2024-09-23 12:21:33,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=257198.66666666666, ans=0.05 2024-09-23 12:22:01,440 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten.whitening_limit, batch_count=257292.0, ans=15.0 2024-09-23 12:22:15,289 INFO [train.py:1198] (2/4) Epoch 15, batch 600, loss[loss=0.2125, ctc_loss=0.142, cr_loss=0.3525, over 17154.00 frames. ], tot_loss[loss=0.2307, ctc_loss=0.1565, cr_loss=0.3708, over 3175019.51 frames. ], batch size: 45, lr: 8.14e-03, grad_scale: 32.0 2024-09-23 12:22:28,597 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=257338.66666666666, ans=0.125 2024-09-23 12:22:29,755 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.274e+02 1.386e+02 1.572e+02 2.356e+02, threshold=2.773e+02, percent-clipped=0.0 2024-09-23 12:23:18,587 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=8.54 vs. limit=15.0 2024-09-23 12:23:21,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=257525.33333333334, ans=0.125 2024-09-23 12:23:38,341 INFO [train.py:1198] (2/4) Epoch 15, batch 650, loss[loss=0.2581, ctc_loss=0.1736, cr_loss=0.4227, over 17208.00 frames. ], tot_loss[loss=0.2301, ctc_loss=0.156, cr_loss=0.3705, over 3219473.13 frames. ], batch size: 50, lr: 8.13e-03, grad_scale: 32.0 2024-09-23 12:23:56,405 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=257618.66666666666, ans=0.125 2024-09-23 12:23:57,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=257618.66666666666, ans=0.125 2024-09-23 12:23:58,169 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=257618.66666666666, ans=0.125 2024-09-23 12:24:34,940 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=257712.0, ans=0.0 2024-09-23 12:24:57,229 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=257758.66666666666, ans=10.0 2024-09-23 12:25:01,758 INFO [train.py:1198] (2/4) Epoch 15, batch 700, loss[loss=0.2371, ctc_loss=0.1634, cr_loss=0.3686, over 17236.00 frames. ], tot_loss[loss=0.2301, ctc_loss=0.156, cr_loss=0.3705, over 3258602.71 frames. ], batch size: 55, lr: 8.13e-03, grad_scale: 32.0 2024-09-23 12:25:18,901 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.271e+02 1.381e+02 1.537e+02 2.206e+02, threshold=2.762e+02, percent-clipped=0.0 2024-09-23 12:25:32,736 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=257852.0, ans=0.0 2024-09-23 12:25:44,000 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=257898.66666666666, ans=0.0 2024-09-23 12:25:56,768 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=257945.33333333334, ans=0.2 2024-09-23 12:26:09,775 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.24 vs. limit=6.0 2024-09-23 12:26:17,304 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=257992.0, ans=0.0 2024-09-23 12:26:22,137 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=257992.0, ans=0.2 2024-09-23 12:26:26,614 INFO [train.py:1198] (2/4) Epoch 15, batch 750, loss[loss=0.2507, ctc_loss=0.1682, cr_loss=0.4126, over 17216.00 frames. ], tot_loss[loss=0.2301, ctc_loss=0.156, cr_loss=0.3707, over 3277055.02 frames. ], batch size: 50, lr: 8.12e-03, grad_scale: 32.0 2024-09-23 12:26:53,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=258085.33333333334, ans=0.0 2024-09-23 12:26:55,452 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=258085.33333333334, ans=0.125 2024-09-23 12:27:07,073 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.41 vs. limit=15.0 2024-09-23 12:27:13,401 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=10.38 vs. limit=15.0 2024-09-23 12:27:19,425 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=258178.66666666666, ans=0.1 2024-09-23 12:27:29,007 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=258225.33333333334, ans=0.0 2024-09-23 12:27:48,955 INFO [train.py:1198] (2/4) Epoch 15, batch 800, loss[loss=0.2259, ctc_loss=0.1535, cr_loss=0.3624, over 17309.00 frames. ], tot_loss[loss=0.2285, ctc_loss=0.1547, cr_loss=0.3686, over 3304397.34 frames. ], batch size: 46, lr: 8.12e-03, grad_scale: 32.0 2024-09-23 12:28:00,458 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=258272.0, ans=0.025 2024-09-23 12:28:03,180 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.136e+02 1.255e+02 1.392e+02 1.544e+02 3.619e+02, threshold=2.784e+02, percent-clipped=1.0 2024-09-23 12:28:11,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=258318.66666666666, ans=0.0 2024-09-23 12:28:30,369 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=258365.33333333334, ans=0.2 2024-09-23 12:28:38,224 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.min_abs, batch_count=258412.0, ans=0.5 2024-09-23 12:29:08,256 INFO [train.py:1198] (2/4) Epoch 15, batch 850, loss[loss=0.2534, ctc_loss=0.1739, cr_loss=0.3973, over 17238.00 frames. ], tot_loss[loss=0.2297, ctc_loss=0.1558, cr_loss=0.3695, over 3311524.09 frames. ], batch size: 55, lr: 8.12e-03, grad_scale: 32.0 2024-09-23 12:29:20,425 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 12:29:37,605 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=258552.0, ans=0.0 2024-09-23 12:29:57,648 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=258645.33333333334, ans=0.125 2024-09-23 12:30:14,079 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module1.whiten, num_groups=1, num_channels=192, metric=7.81 vs. limit=15.0 2024-09-23 12:30:21,174 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=258692.0, ans=0.0 2024-09-23 12:30:36,041 INFO [train.py:1198] (2/4) Epoch 15, batch 900, loss[loss=0.2389, ctc_loss=0.1591, cr_loss=0.3991, over 17209.00 frames. ], tot_loss[loss=0.229, ctc_loss=0.1553, cr_loss=0.3686, over 3313605.77 frames. ], batch size: 50, lr: 8.11e-03, grad_scale: 32.0 2024-09-23 12:30:36,655 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.34 vs. limit=15.0 2024-09-23 12:30:43,210 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=512, metric=4.13 vs. limit=15.0 2024-09-23 12:30:50,308 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.250e+02 1.335e+02 1.491e+02 2.252e+02, threshold=2.670e+02, percent-clipped=0.0 2024-09-23 12:30:57,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer_ff3.min_abs, batch_count=258785.33333333334, ans=0.2 2024-09-23 12:31:00,185 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=258785.33333333334, ans=0.2 2024-09-23 12:31:00,286 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=258785.33333333334, ans=0.0 2024-09-23 12:31:01,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=258785.33333333334, ans=0.125 2024-09-23 12:31:15,350 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.98 vs. limit=15.0 2024-09-23 12:31:40,552 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.max_positive, batch_count=258925.33333333334, ans=0.95 2024-09-23 12:31:51,683 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=258925.33333333334, ans=0.125 2024-09-23 12:31:55,756 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=13.03 vs. limit=15.0 2024-09-23 12:31:56,134 INFO [train.py:1198] (2/4) Epoch 15, batch 950, loss[loss=0.2257, ctc_loss=0.1504, cr_loss=0.3764, over 16948.00 frames. ], tot_loss[loss=0.2296, ctc_loss=0.1558, cr_loss=0.3692, over 3314371.12 frames. ], batch size: 58, lr: 8.11e-03, grad_scale: 16.0 2024-09-23 12:32:01,148 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=258972.0, ans=0.025 2024-09-23 12:32:02,752 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=258972.0, ans=0.125 2024-09-23 12:32:41,989 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=259065.33333333334, ans=0.125 2024-09-23 12:32:42,022 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=259065.33333333334, ans=0.1 2024-09-23 12:32:49,890 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=259112.0, ans=0.0 2024-09-23 12:32:53,028 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=259112.0, ans=0.0 2024-09-23 12:32:57,683 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=259112.0, ans=0.025 2024-09-23 12:33:17,945 INFO [train.py:1198] (2/4) Epoch 15, batch 1000, loss[loss=0.2675, ctc_loss=0.187, cr_loss=0.4024, over 17031.00 frames. ], tot_loss[loss=0.2297, ctc_loss=0.1559, cr_loss=0.3692, over 3311049.69 frames. ], batch size: 52, lr: 8.11e-03, grad_scale: 16.0 2024-09-23 12:33:33,770 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.079e+02 1.224e+02 1.330e+02 1.426e+02 2.141e+02, threshold=2.660e+02, percent-clipped=0.0 2024-09-23 12:34:04,656 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=259345.33333333334, ans=0.125 2024-09-23 12:34:04,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=259345.33333333334, ans=0.05 2024-09-23 12:34:08,384 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten.whitening_limit, batch_count=259345.33333333334, ans=22.5 2024-09-23 12:34:19,186 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=259345.33333333334, ans=0.0 2024-09-23 12:34:40,677 INFO [train.py:1198] (2/4) Epoch 15, batch 1050, loss[loss=0.2156, ctc_loss=0.1478, cr_loss=0.3393, over 17105.00 frames. ], tot_loss[loss=0.23, ctc_loss=0.1562, cr_loss=0.3692, over 3307584.01 frames. ], batch size: 49, lr: 8.10e-03, grad_scale: 16.0 2024-09-23 12:34:41,090 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=259438.66666666666, ans=0.125 2024-09-23 12:34:48,875 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=259438.66666666666, ans=0.1 2024-09-23 12:35:10,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=259485.33333333334, ans=0.0 2024-09-23 12:35:47,286 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=18.95 vs. limit=22.5 2024-09-23 12:36:00,307 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.01 vs. limit=22.5 2024-09-23 12:36:05,845 INFO [train.py:1198] (2/4) Epoch 15, batch 1100, loss[loss=0.2564, ctc_loss=0.1742, cr_loss=0.4107, over 17150.00 frames. ], tot_loss[loss=0.2299, ctc_loss=0.1559, cr_loss=0.3697, over 3314052.21 frames. ], batch size: 48, lr: 8.10e-03, grad_scale: 16.0 2024-09-23 12:36:17,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=259672.0, ans=0.1 2024-09-23 12:36:21,571 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.305e+02 1.420e+02 1.545e+02 2.157e+02, threshold=2.840e+02, percent-clipped=0.0 2024-09-23 12:36:21,801 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=259718.66666666666, ans=0.1 2024-09-23 12:36:23,492 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.min_abs, batch_count=259718.66666666666, ans=0.5 2024-09-23 12:37:11,635 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=259858.66666666666, ans=0.125 2024-09-23 12:37:16,414 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=259858.66666666666, ans=0.125 2024-09-23 12:37:28,267 INFO [train.py:1198] (2/4) Epoch 15, batch 1150, loss[loss=0.2238, ctc_loss=0.1499, cr_loss=0.3694, over 17304.00 frames. ], tot_loss[loss=0.2298, ctc_loss=0.1558, cr_loss=0.37, over 3333873.13 frames. ], batch size: 51, lr: 8.10e-03, grad_scale: 16.0 2024-09-23 12:37:31,966 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=259905.33333333334, ans=0.0 2024-09-23 12:37:42,940 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=259952.0, ans=0.0 2024-09-23 12:38:12,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=259998.66666666666, ans=0.125 2024-09-23 12:38:13,667 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=259998.66666666666, ans=0.125 2024-09-23 12:38:26,730 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=7.26 vs. limit=15.0 2024-09-23 12:38:48,573 INFO [train.py:1198] (2/4) Epoch 15, batch 1200, loss[loss=0.2001, ctc_loss=0.1341, cr_loss=0.33, over 17185.00 frames. ], tot_loss[loss=0.2298, ctc_loss=0.1558, cr_loss=0.3701, over 3338169.20 frames. ], batch size: 41, lr: 8.09e-03, grad_scale: 32.0 2024-09-23 12:38:52,490 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=10.78 vs. limit=15.0 2024-09-23 12:38:57,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=260138.66666666666, ans=0.0 2024-09-23 12:38:57,877 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=14.16 vs. limit=22.5 2024-09-23 12:39:04,649 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.045e+02 1.247e+02 1.362e+02 1.504e+02 2.311e+02, threshold=2.725e+02, percent-clipped=0.0 2024-09-23 12:39:16,404 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=260185.33333333334, ans=0.0 2024-09-23 12:39:29,099 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=260232.0, ans=0.0 2024-09-23 12:40:13,698 INFO [train.py:1198] (2/4) Epoch 15, batch 1250, loss[loss=0.2352, ctc_loss=0.1569, cr_loss=0.3917, over 17211.00 frames. ], tot_loss[loss=0.2293, ctc_loss=0.1553, cr_loss=0.3701, over 3345343.80 frames. ], batch size: 47, lr: 8.09e-03, grad_scale: 32.0 2024-09-23 12:40:15,928 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=9.31 vs. limit=15.0 2024-09-23 12:40:36,315 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=5.88 vs. limit=12.0 2024-09-23 12:40:45,413 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=260418.66666666666, ans=0.1 2024-09-23 12:41:06,249 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=260512.0, ans=0.125 2024-09-23 12:41:28,981 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.00 vs. limit=15.0 2024-09-23 12:41:35,958 INFO [train.py:1198] (2/4) Epoch 15, batch 1300, loss[loss=0.193, ctc_loss=0.1274, cr_loss=0.3281, over 17302.00 frames. ], tot_loss[loss=0.2289, ctc_loss=0.155, cr_loss=0.3693, over 3352146.89 frames. ], batch size: 51, lr: 8.09e-03, grad_scale: 16.0 2024-09-23 12:41:52,566 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.94 vs. limit=15.0 2024-09-23 12:41:53,392 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.058e+02 1.270e+02 1.373e+02 1.516e+02 2.157e+02, threshold=2.746e+02, percent-clipped=0.0 2024-09-23 12:42:11,609 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=260698.66666666666, ans=0.0 2024-09-23 12:42:26,806 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 12:42:27,135 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=10.77 vs. limit=15.0 2024-09-23 12:42:36,307 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=260745.33333333334, ans=0.125 2024-09-23 12:42:36,333 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=260745.33333333334, ans=0.125 2024-09-23 12:42:58,448 INFO [train.py:1198] (2/4) Epoch 15, batch 1350, loss[loss=0.1815, ctc_loss=0.1201, cr_loss=0.3068, over 17187.00 frames. ], tot_loss[loss=0.2281, ctc_loss=0.1545, cr_loss=0.368, over 3358535.89 frames. ], batch size: 41, lr: 8.08e-03, grad_scale: 16.0 2024-09-23 12:42:58,804 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=260838.66666666666, ans=10.0 2024-09-23 12:43:02,466 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.17 vs. limit=6.0 2024-09-23 12:43:22,632 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=260885.33333333334, ans=0.125 2024-09-23 12:43:28,225 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.40 vs. limit=22.5 2024-09-23 12:43:49,864 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=260978.66666666666, ans=0.025 2024-09-23 12:44:18,618 INFO [train.py:1198] (2/4) Epoch 15, batch 1400, loss[loss=0.187, ctc_loss=0.1243, cr_loss=0.3139, over 16356.00 frames. ], tot_loss[loss=0.2271, ctc_loss=0.1538, cr_loss=0.3663, over 3363012.95 frames. ], batch size: 36, lr: 8.08e-03, grad_scale: 16.0 2024-09-23 12:44:36,421 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.078e+02 1.225e+02 1.306e+02 1.392e+02 2.119e+02, threshold=2.612e+02, percent-clipped=0.0 2024-09-23 12:44:51,318 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=7.23 vs. limit=15.0 2024-09-23 12:44:56,815 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=261165.33333333334, ans=0.0 2024-09-23 12:45:04,826 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=261165.33333333334, ans=0.125 2024-09-23 12:45:13,850 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=261212.0, ans=0.1 2024-09-23 12:45:24,925 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.00 vs. limit=15.0 2024-09-23 12:45:46,016 INFO [train.py:1198] (2/4) Epoch 15, batch 1450, loss[loss=0.2265, ctc_loss=0.1539, cr_loss=0.3633, over 17154.00 frames. ], tot_loss[loss=0.227, ctc_loss=0.1537, cr_loss=0.3666, over 3367276.23 frames. ], batch size: 48, lr: 8.07e-03, grad_scale: 16.0 2024-09-23 12:46:15,660 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=261352.0, ans=0.0 2024-09-23 12:46:26,789 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=261398.66666666666, ans=0.125 2024-09-23 12:46:37,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten.whitening_limit, batch_count=261445.33333333334, ans=15.0 2024-09-23 12:46:42,877 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=261445.33333333334, ans=0.0 2024-09-23 12:46:54,014 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=261492.0, ans=0.025 2024-09-23 12:46:58,974 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=261492.0, ans=0.1 2024-09-23 12:47:06,874 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=261538.66666666666, ans=0.125 2024-09-23 12:47:08,081 INFO [train.py:1198] (2/4) Epoch 15, batch 1500, loss[loss=0.2399, ctc_loss=0.1636, cr_loss=0.3816, over 17016.00 frames. ], tot_loss[loss=0.2263, ctc_loss=0.1532, cr_loss=0.3653, over 3355024.81 frames. ], batch size: 56, lr: 8.07e-03, grad_scale: 16.0 2024-09-23 12:47:21,275 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=261538.66666666666, ans=0.125 2024-09-23 12:47:25,760 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.263e+02 1.374e+02 1.561e+02 5.695e+02, threshold=2.748e+02, percent-clipped=2.0 2024-09-23 12:47:49,225 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=261632.0, ans=0.125 2024-09-23 12:47:49,372 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=261632.0, ans=0.125 2024-09-23 12:47:51,615 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn2.whiten, num_groups=1, num_channels=192, metric=9.08 vs. limit=22.5 2024-09-23 12:48:06,685 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=261678.66666666666, ans=0.0 2024-09-23 12:48:18,017 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=261725.33333333334, ans=0.025 2024-09-23 12:48:22,821 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=261725.33333333334, ans=0.125 2024-09-23 12:48:30,872 INFO [train.py:1198] (2/4) Epoch 15, batch 1550, loss[loss=0.2281, ctc_loss=0.1538, cr_loss=0.3718, over 17150.00 frames. ], tot_loss[loss=0.2255, ctc_loss=0.1525, cr_loss=0.3651, over 3366641.38 frames. ], batch size: 48, lr: 8.07e-03, grad_scale: 16.0 2024-09-23 12:48:39,104 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=261772.0, ans=0.2 2024-09-23 12:48:42,441 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=13.84 vs. limit=15.0 2024-09-23 12:48:51,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=261818.66666666666, ans=0.1 2024-09-23 12:49:08,444 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.85 vs. limit=6.0 2024-09-23 12:49:09,572 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=261865.33333333334, ans=0.09899494936611666 2024-09-23 12:49:53,460 INFO [train.py:1198] (2/4) Epoch 15, batch 1600, loss[loss=0.2272, ctc_loss=0.1517, cr_loss=0.3775, over 17022.00 frames. ], tot_loss[loss=0.2262, ctc_loss=0.153, cr_loss=0.366, over 3374412.15 frames. ], batch size: 52, lr: 8.06e-03, grad_scale: 32.0 2024-09-23 12:49:55,468 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=262005.33333333334, ans=0.0 2024-09-23 12:50:13,468 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.073e+02 1.270e+02 1.410e+02 1.627e+02 2.274e+02, threshold=2.820e+02, percent-clipped=0.0 2024-09-23 12:50:14,076 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=9.43 vs. limit=15.0 2024-09-23 12:50:20,937 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=262052.0, ans=0.0 2024-09-23 12:51:05,625 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=262192.0, ans=0.0 2024-09-23 12:51:10,672 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=11.76 vs. limit=22.5 2024-09-23 12:51:17,820 INFO [train.py:1198] (2/4) Epoch 15, batch 1650, loss[loss=0.2601, ctc_loss=0.1763, cr_loss=0.4188, over 17235.00 frames. ], tot_loss[loss=0.2275, ctc_loss=0.1539, cr_loss=0.3677, over 3366068.85 frames. ], batch size: 50, lr: 8.06e-03, grad_scale: 32.0 2024-09-23 12:51:56,096 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=262332.0, ans=0.2 2024-09-23 12:52:19,953 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=262425.3333333333, ans=0.0 2024-09-23 12:52:39,797 INFO [train.py:1198] (2/4) Epoch 15, batch 1700, loss[loss=0.2398, ctc_loss=0.1613, cr_loss=0.3926, over 16576.00 frames. ], tot_loss[loss=0.2282, ctc_loss=0.1545, cr_loss=0.3683, over 3359268.82 frames. ], batch size: 66, lr: 8.06e-03, grad_scale: 32.0 2024-09-23 12:52:57,187 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.135e+02 1.272e+02 1.393e+02 1.539e+02 2.504e+02, threshold=2.785e+02, percent-clipped=0.0 2024-09-23 12:53:11,691 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=262565.3333333333, ans=0.125 2024-09-23 12:53:19,509 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=262565.3333333333, ans=0.0 2024-09-23 12:53:23,114 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.18 vs. limit=15.0 2024-09-23 12:53:58,953 INFO [train.py:1198] (2/4) Epoch 15, batch 1750, loss[loss=0.2417, ctc_loss=0.1622, cr_loss=0.3971, over 16826.00 frames. ], tot_loss[loss=0.2289, ctc_loss=0.1551, cr_loss=0.3687, over 3358245.05 frames. ], batch size: 61, lr: 8.05e-03, grad_scale: 32.0 2024-09-23 12:53:59,366 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=262705.3333333333, ans=0.125 2024-09-23 12:54:27,824 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=262752.0, ans=0.125 2024-09-23 12:54:34,572 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.68 vs. limit=15.0 2024-09-23 12:54:36,088 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=12.57 vs. limit=12.0 2024-09-23 12:54:44,690 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=262798.6666666667, ans=0.07 2024-09-23 12:54:56,060 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=262845.3333333333, ans=0.125 2024-09-23 12:55:04,938 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=262845.3333333333, ans=0.0 2024-09-23 12:55:11,968 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.74 vs. limit=15.0 2024-09-23 12:55:22,184 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.23 vs. limit=15.0 2024-09-23 12:55:26,237 INFO [train.py:1198] (2/4) Epoch 15, batch 1800, loss[loss=0.2741, ctc_loss=0.1899, cr_loss=0.4207, over 14934.00 frames. ], tot_loss[loss=0.2305, ctc_loss=0.1563, cr_loss=0.371, over 3356013.90 frames. ], batch size: 89, lr: 8.05e-03, grad_scale: 32.0 2024-09-23 12:55:43,903 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.283e+02 1.353e+02 1.483e+02 2.243e+02, threshold=2.705e+02, percent-clipped=0.0 2024-09-23 12:55:44,272 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=262985.3333333333, ans=0.1 2024-09-23 12:55:47,945 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=9.01 vs. limit=15.0 2024-09-23 12:55:49,032 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=262985.3333333333, ans=0.125 2024-09-23 12:55:52,176 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=262985.3333333333, ans=0.025 2024-09-23 12:56:11,311 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=263032.0, ans=0.0 2024-09-23 12:56:14,401 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=263078.6666666667, ans=0.0 2024-09-23 12:56:28,616 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=263125.3333333333, ans=0.0 2024-09-23 12:56:35,172 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 12:56:46,104 INFO [train.py:1198] (2/4) Epoch 15, batch 1850, loss[loss=0.22, ctc_loss=0.1485, cr_loss=0.3571, over 16998.00 frames. ], tot_loss[loss=0.2298, ctc_loss=0.1557, cr_loss=0.3704, over 3365237.69 frames. ], batch size: 44, lr: 8.05e-03, grad_scale: 32.0 2024-09-23 12:56:52,797 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=263172.0, ans=0.09899494936611666 2024-09-23 12:56:59,359 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=263172.0, ans=0.125 2024-09-23 12:57:02,513 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=263218.6666666667, ans=0.04949747468305833 2024-09-23 12:57:08,987 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=263218.6666666667, ans=0.2 2024-09-23 12:57:16,966 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=263265.3333333333, ans=0.0 2024-09-23 12:57:27,311 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=263265.3333333333, ans=0.125 2024-09-23 12:57:27,343 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=263265.3333333333, ans=0.1 2024-09-23 12:57:40,221 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.63 vs. limit=10.0 2024-09-23 12:58:08,513 INFO [train.py:1198] (2/4) Epoch 15, batch 1900, loss[loss=0.2114, ctc_loss=0.1406, cr_loss=0.3538, over 16930.00 frames. ], tot_loss[loss=0.2303, ctc_loss=0.1561, cr_loss=0.3712, over 3356033.28 frames. ], batch size: 42, lr: 8.04e-03, grad_scale: 32.0 2024-09-23 12:58:24,828 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=263452.0, ans=0.2 2024-09-23 12:58:26,131 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.271e+02 1.385e+02 1.549e+02 2.232e+02, threshold=2.769e+02, percent-clipped=0.0 2024-09-23 12:58:34,829 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=5.89 vs. limit=12.0 2024-09-23 12:58:51,676 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 12:59:27,839 INFO [train.py:1198] (2/4) Epoch 15, batch 1950, loss[loss=0.2427, ctc_loss=0.1662, cr_loss=0.3828, over 17044.00 frames. ], tot_loss[loss=0.2299, ctc_loss=0.1557, cr_loss=0.371, over 3354117.10 frames. ], batch size: 56, lr: 8.04e-03, grad_scale: 16.0 2024-09-23 12:59:35,425 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=263638.6666666667, ans=0.125 2024-09-23 12:59:46,591 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=263685.3333333333, ans=0.125 2024-09-23 13:00:01,742 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=263685.3333333333, ans=0.125 2024-09-23 13:00:05,621 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=263732.0, ans=0.125 2024-09-23 13:00:07,294 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=263732.0, ans=0.025 2024-09-23 13:00:22,080 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=3.40 vs. limit=6.0 2024-09-23 13:00:37,854 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.46 vs. limit=10.0 2024-09-23 13:00:54,760 INFO [train.py:1198] (2/4) Epoch 15, batch 2000, loss[loss=0.2307, ctc_loss=0.1549, cr_loss=0.3791, over 17172.00 frames. ], tot_loss[loss=0.229, ctc_loss=0.1551, cr_loss=0.3695, over 3360349.70 frames. ], batch size: 45, lr: 8.04e-03, grad_scale: 32.0 2024-09-23 13:00:55,145 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=263872.0, ans=0.125 2024-09-23 13:01:00,418 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=512, metric=2.42 vs. limit=15.0 2024-09-23 13:01:12,778 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=263918.6666666667, ans=0.1 2024-09-23 13:01:14,002 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.286e+02 1.410e+02 1.590e+02 2.174e+02, threshold=2.819e+02, percent-clipped=0.0 2024-09-23 13:01:31,668 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=263965.3333333333, ans=0.0 2024-09-23 13:01:33,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=263965.3333333333, ans=0.125 2024-09-23 13:01:38,149 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=263965.3333333333, ans=0.125 2024-09-23 13:01:38,459 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=10.60 vs. limit=15.0 2024-09-23 13:01:42,845 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=264012.0, ans=0.125 2024-09-23 13:01:50,783 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=264012.0, ans=0.125 2024-09-23 13:01:58,749 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=264058.6666666667, ans=0.125 2024-09-23 13:02:03,543 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=264058.6666666667, ans=0.125 2024-09-23 13:02:16,687 INFO [train.py:1198] (2/4) Epoch 15, batch 2050, loss[loss=0.2468, ctc_loss=0.1683, cr_loss=0.3922, over 16930.00 frames. ], tot_loss[loss=0.2287, ctc_loss=0.1548, cr_loss=0.3693, over 3366190.75 frames. ], batch size: 58, lr: 8.03e-03, grad_scale: 32.0 2024-09-23 13:02:20,148 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=264105.3333333333, ans=0.2 2024-09-23 13:02:31,789 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=264152.0, ans=0.125 2024-09-23 13:02:33,347 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=264152.0, ans=0.125 2024-09-23 13:02:44,274 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=264152.0, ans=0.125 2024-09-23 13:02:44,309 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=264152.0, ans=0.1 2024-09-23 13:02:44,849 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.61 vs. limit=15.0 2024-09-23 13:02:45,836 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=264152.0, ans=0.0 2024-09-23 13:03:00,203 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=264198.6666666667, ans=0.0 2024-09-23 13:03:06,425 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=264245.3333333333, ans=0.125 2024-09-23 13:03:09,642 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer_ff3.min_abs, batch_count=264245.3333333333, ans=0.2 2024-09-23 13:03:36,265 INFO [train.py:1198] (2/4) Epoch 15, batch 2100, loss[loss=0.2814, ctc_loss=0.2016, cr_loss=0.3993, over 11890.00 frames. ], tot_loss[loss=0.2299, ctc_loss=0.1558, cr_loss=0.3705, over 3351390.29 frames. ], batch size: 123, lr: 8.03e-03, grad_scale: 32.0 2024-09-23 13:03:55,211 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.095e+02 1.282e+02 1.387e+02 1.573e+02 2.128e+02, threshold=2.775e+02, percent-clipped=0.0 2024-09-23 13:04:08,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten.whitening_limit, batch_count=264432.0, ans=15.0 2024-09-23 13:04:25,673 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=264478.6666666667, ans=0.025 2024-09-23 13:04:49,363 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=264525.3333333333, ans=0.125 2024-09-23 13:04:54,918 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=264525.3333333333, ans=0.0 2024-09-23 13:05:03,753 INFO [train.py:1198] (2/4) Epoch 15, batch 2150, loss[loss=0.2028, ctc_loss=0.1352, cr_loss=0.3378, over 16965.00 frames. ], tot_loss[loss=0.2296, ctc_loss=0.1555, cr_loss=0.3702, over 3345125.36 frames. ], batch size: 42, lr: 8.03e-03, grad_scale: 32.0 2024-09-23 13:05:21,591 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=264618.6666666667, ans=0.125 2024-09-23 13:05:28,478 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.02 vs. limit=22.5 2024-09-23 13:05:32,780 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=264618.6666666667, ans=0.2 2024-09-23 13:05:32,864 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=264618.6666666667, ans=0.0 2024-09-23 13:05:42,371 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=264665.3333333333, ans=0.07 2024-09-23 13:05:44,392 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.93 vs. limit=12.0 2024-09-23 13:06:01,509 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=264712.0, ans=0.125 2024-09-23 13:06:03,196 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=264712.0, ans=0.125 2024-09-23 13:06:17,822 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=264758.6666666667, ans=0.2 2024-09-23 13:06:20,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=264758.6666666667, ans=0.2 2024-09-23 13:06:22,959 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=14.79 vs. limit=22.5 2024-09-23 13:06:23,715 INFO [train.py:1198] (2/4) Epoch 15, batch 2200, loss[loss=0.1834, ctc_loss=0.1214, cr_loss=0.3101, over 17254.00 frames. ], tot_loss[loss=0.2292, ctc_loss=0.1552, cr_loss=0.3698, over 3347077.05 frames. ], batch size: 44, lr: 8.02e-03, grad_scale: 32.0 2024-09-23 13:06:42,813 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.110e+02 1.291e+02 1.443e+02 1.638e+02 2.419e+02, threshold=2.885e+02, percent-clipped=0.0 2024-09-23 13:06:51,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=264852.0, ans=0.2 2024-09-23 13:07:37,256 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=264992.0, ans=0.0 2024-09-23 13:07:46,502 INFO [train.py:1198] (2/4) Epoch 15, batch 2250, loss[loss=0.2425, ctc_loss=0.1642, cr_loss=0.3916, over 16921.00 frames. ], tot_loss[loss=0.2288, ctc_loss=0.1549, cr_loss=0.3697, over 3354692.80 frames. ], batch size: 58, lr: 8.02e-03, grad_scale: 32.0 2024-09-23 13:08:04,701 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=265085.3333333333, ans=0.0 2024-09-23 13:08:17,807 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=265132.0, ans=0.2 2024-09-23 13:08:37,007 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=265178.6666666667, ans=0.0 2024-09-23 13:08:52,786 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=265225.3333333333, ans=0.1 2024-09-23 13:09:00,604 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=265225.3333333333, ans=0.1 2024-09-23 13:09:05,480 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-23 13:09:06,702 INFO [train.py:1198] (2/4) Epoch 15, batch 2300, loss[loss=0.1968, ctc_loss=0.1286, cr_loss=0.3409, over 17094.00 frames. ], tot_loss[loss=0.2301, ctc_loss=0.1559, cr_loss=0.371, over 3346814.56 frames. ], batch size: 43, lr: 8.02e-03, grad_scale: 16.0 2024-09-23 13:09:10,261 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=265272.0, ans=0.1 2024-09-23 13:09:27,202 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.234e+02 1.309e+02 1.490e+02 3.155e+02, threshold=2.619e+02, percent-clipped=1.0 2024-09-23 13:09:55,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=265365.3333333333, ans=0.0 2024-09-23 13:09:55,306 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=265365.3333333333, ans=0.125 2024-09-23 13:10:06,295 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.18 vs. limit=15.0 2024-09-23 13:10:33,244 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=265505.3333333333, ans=0.125 2024-09-23 13:10:34,408 INFO [train.py:1198] (2/4) Epoch 15, batch 2350, loss[loss=0.2874, ctc_loss=0.2056, cr_loss=0.4091, over 12244.00 frames. ], tot_loss[loss=0.2308, ctc_loss=0.1565, cr_loss=0.3718, over 3348147.28 frames. ], batch size: 123, lr: 8.01e-03, grad_scale: 16.0 2024-09-23 13:10:36,311 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=265505.3333333333, ans=0.125 2024-09-23 13:11:31,844 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=10.55 vs. limit=15.0 2024-09-23 13:11:40,360 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=6.76 vs. limit=15.0 2024-09-23 13:11:53,674 INFO [train.py:1198] (2/4) Epoch 15, batch 2400, loss[loss=0.2353, ctc_loss=0.1593, cr_loss=0.3798, over 17267.00 frames. ], tot_loss[loss=0.2289, ctc_loss=0.1549, cr_loss=0.37, over 3359151.83 frames. ], batch size: 44, lr: 8.01e-03, grad_scale: 16.0 2024-09-23 13:12:10,971 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=265785.3333333333, ans=0.125 2024-09-23 13:12:17,741 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=265785.3333333333, ans=0.125 2024-09-23 13:12:18,951 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.338e+02 1.450e+02 1.574e+02 3.453e+02, threshold=2.900e+02, percent-clipped=1.0 2024-09-23 13:12:23,089 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=4.00 vs. limit=15.0 2024-09-23 13:12:35,877 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=7.42 vs. limit=15.0 2024-09-23 13:12:41,711 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=265832.0, ans=0.0 2024-09-23 13:13:03,266 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.58 vs. limit=6.0 2024-09-23 13:13:16,576 INFO [train.py:1198] (2/4) Epoch 15, batch 2450, loss[loss=0.1879, ctc_loss=0.1244, cr_loss=0.3176, over 17078.00 frames. ], tot_loss[loss=0.2297, ctc_loss=0.1555, cr_loss=0.3711, over 3360306.85 frames. ], batch size: 43, lr: 8.00e-03, grad_scale: 16.0 2024-09-23 13:13:44,792 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.45 vs. limit=15.0 2024-09-23 13:13:58,672 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=266065.3333333333, ans=0.1 2024-09-23 13:14:12,937 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=266112.0, ans=0.125 2024-09-23 13:14:22,578 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=266158.6666666667, ans=0.125 2024-09-23 13:14:27,272 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=266158.6666666667, ans=0.0 2024-09-23 13:14:31,787 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=266158.6666666667, ans=0.125 2024-09-23 13:14:38,832 INFO [train.py:1198] (2/4) Epoch 15, batch 2500, loss[loss=0.2614, ctc_loss=0.1777, cr_loss=0.4185, over 17216.00 frames. ], tot_loss[loss=0.229, ctc_loss=0.1549, cr_loss=0.3702, over 3371408.76 frames. ], batch size: 55, lr: 8.00e-03, grad_scale: 16.0 2024-09-23 13:15:02,934 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=266252.0, ans=0.5 2024-09-23 13:15:05,874 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.066e+02 1.240e+02 1.331e+02 1.479e+02 2.623e+02, threshold=2.663e+02, percent-clipped=0.0 2024-09-23 13:15:15,867 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=266298.6666666667, ans=0.125 2024-09-23 13:15:52,455 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=266392.0, ans=0.2 2024-09-23 13:16:03,517 INFO [train.py:1198] (2/4) Epoch 15, batch 2550, loss[loss=0.2382, ctc_loss=0.1628, cr_loss=0.377, over 17239.00 frames. ], tot_loss[loss=0.2312, ctc_loss=0.1566, cr_loss=0.3727, over 3361784.95 frames. ], batch size: 55, lr: 8.00e-03, grad_scale: 16.0 2024-09-23 13:16:29,137 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=266485.3333333333, ans=0.2 2024-09-23 13:16:39,057 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.02 vs. limit=10.0 2024-09-23 13:16:58,967 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=266578.6666666667, ans=0.0 2024-09-23 13:17:04,790 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=266578.6666666667, ans=0.2 2024-09-23 13:17:05,191 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.99 vs. limit=15.0 2024-09-23 13:17:23,800 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=266672.0, ans=0.2 2024-09-23 13:17:25,062 INFO [train.py:1198] (2/4) Epoch 15, batch 2600, loss[loss=0.2452, ctc_loss=0.1699, cr_loss=0.3762, over 17081.00 frames. ], tot_loss[loss=0.2299, ctc_loss=0.1556, cr_loss=0.3716, over 3363074.10 frames. ], batch size: 49, lr: 7.99e-03, grad_scale: 16.0 2024-09-23 13:17:38,079 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=266672.0, ans=0.2 2024-09-23 13:17:47,325 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.078e+02 1.291e+02 1.396e+02 1.519e+02 2.239e+02, threshold=2.791e+02, percent-clipped=0.0 2024-09-23 13:17:49,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=266718.6666666667, ans=0.125 2024-09-23 13:18:00,175 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=266765.3333333333, ans=0.1 2024-09-23 13:18:19,259 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=266812.0, ans=0.125 2024-09-23 13:18:25,742 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=266812.0, ans=0.05 2024-09-23 13:18:30,960 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=512, metric=4.35 vs. limit=15.0 2024-09-23 13:18:44,566 INFO [train.py:1198] (2/4) Epoch 15, batch 2650, loss[loss=0.2766, ctc_loss=0.187, cr_loss=0.4479, over 17320.00 frames. ], tot_loss[loss=0.2292, ctc_loss=0.1551, cr_loss=0.3708, over 3375140.59 frames. ], batch size: 54, lr: 7.99e-03, grad_scale: 16.0 2024-09-23 13:19:15,584 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=266998.6666666667, ans=0.025 2024-09-23 13:19:32,148 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=266998.6666666667, ans=0.09899494936611666 2024-09-23 13:20:07,513 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=267092.0, ans=0.1 2024-09-23 13:20:12,066 INFO [train.py:1198] (2/4) Epoch 15, batch 2700, loss[loss=0.2127, ctc_loss=0.1397, cr_loss=0.3652, over 17169.00 frames. ], tot_loss[loss=0.2282, ctc_loss=0.1542, cr_loss=0.3699, over 3377981.14 frames. ], batch size: 45, lr: 7.99e-03, grad_scale: 16.0 2024-09-23 13:20:15,975 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.87 vs. limit=15.0 2024-09-23 13:20:34,472 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.035e+02 1.269e+02 1.346e+02 1.479e+02 2.065e+02, threshold=2.692e+02, percent-clipped=0.0 2024-09-23 13:20:36,394 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=267185.3333333333, ans=0.125 2024-09-23 13:20:58,694 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=267278.6666666667, ans=0.2 2024-09-23 13:21:09,634 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=267278.6666666667, ans=0.0 2024-09-23 13:21:31,860 INFO [train.py:1198] (2/4) Epoch 15, batch 2750, loss[loss=0.1823, ctc_loss=0.1266, cr_loss=0.2789, over 17045.00 frames. ], tot_loss[loss=0.2281, ctc_loss=0.1544, cr_loss=0.3688, over 3371021.86 frames. ], batch size: 46, lr: 7.98e-03, grad_scale: 16.0 2024-09-23 13:21:32,244 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=267372.0, ans=0.1 2024-09-23 13:22:00,307 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=267418.6666666667, ans=0.0 2024-09-23 13:22:01,846 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=267418.6666666667, ans=0.125 2024-09-23 13:22:53,941 INFO [train.py:1198] (2/4) Epoch 15, batch 2800, loss[loss=0.2179, ctc_loss=0.145, cr_loss=0.3649, over 17136.00 frames. ], tot_loss[loss=0.2288, ctc_loss=0.155, cr_loss=0.3692, over 3361636.70 frames. ], batch size: 45, lr: 7.98e-03, grad_scale: 32.0 2024-09-23 13:23:03,900 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=267605.3333333333, ans=0.1 2024-09-23 13:23:11,067 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=3.56 vs. limit=15.0 2024-09-23 13:23:11,749 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=267652.0, ans=0.2 2024-09-23 13:23:17,756 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.059e+02 1.278e+02 1.476e+02 1.758e+02 2.429e+02, threshold=2.952e+02, percent-clipped=0.0 2024-09-23 13:23:42,500 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=14.09 vs. limit=15.0 2024-09-23 13:24:08,991 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=267792.0, ans=0.125 2024-09-23 13:24:13,372 INFO [train.py:1198] (2/4) Epoch 15, batch 2850, loss[loss=0.2239, ctc_loss=0.1478, cr_loss=0.3806, over 17036.00 frames. ], tot_loss[loss=0.2295, ctc_loss=0.1556, cr_loss=0.3696, over 3346640.62 frames. ], batch size: 39, lr: 7.98e-03, grad_scale: 16.0 2024-09-23 13:24:25,422 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=267838.6666666667, ans=0.125 2024-09-23 13:24:31,333 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=267885.3333333333, ans=0.125 2024-09-23 13:24:54,028 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=267932.0, ans=0.025 2024-09-23 13:25:02,463 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=12.82 vs. limit=15.0 2024-09-23 13:25:25,809 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.min_abs, batch_count=268025.3333333333, ans=0.5 2024-09-23 13:25:25,890 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=268025.3333333333, ans=0.125 2024-09-23 13:25:34,319 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=16.90 vs. limit=22.5 2024-09-23 13:25:35,434 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=268025.3333333333, ans=0.0 2024-09-23 13:25:41,408 INFO [train.py:1198] (2/4) Epoch 15, batch 2900, loss[loss=0.239, ctc_loss=0.1607, cr_loss=0.3914, over 17321.00 frames. ], tot_loss[loss=0.2294, ctc_loss=0.1555, cr_loss=0.3694, over 3341988.47 frames. ], batch size: 51, lr: 7.97e-03, grad_scale: 16.0 2024-09-23 13:25:59,709 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=268118.6666666667, ans=0.0 2024-09-23 13:26:05,969 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.264e+02 1.372e+02 1.552e+02 2.806e+02, threshold=2.744e+02, percent-clipped=0.0 2024-09-23 13:26:07,900 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=268118.6666666667, ans=0.0 2024-09-23 13:26:12,800 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=268165.3333333333, ans=0.0 2024-09-23 13:26:14,431 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=268165.3333333333, ans=0.0 2024-09-23 13:26:19,210 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=268165.3333333333, ans=0.125 2024-09-23 13:26:20,691 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=268165.3333333333, ans=0.125 2024-09-23 13:26:28,974 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.84 vs. limit=22.5 2024-09-23 13:26:49,309 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=268258.6666666667, ans=0.0 2024-09-23 13:26:51,205 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=11.46 vs. limit=22.5 2024-09-23 13:27:03,974 INFO [train.py:1198] (2/4) Epoch 15, batch 2950, loss[loss=0.2399, ctc_loss=0.1653, cr_loss=0.3727, over 14930.00 frames. ], tot_loss[loss=0.2286, ctc_loss=0.1548, cr_loss=0.3689, over 3354004.44 frames. ], batch size: 89, lr: 7.97e-03, grad_scale: 16.0 2024-09-23 13:27:04,351 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 13:27:11,193 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=5.82 vs. limit=12.0 2024-09-23 13:27:21,801 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=268352.0, ans=0.0 2024-09-23 13:27:37,012 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=6.91 vs. limit=15.0 2024-09-23 13:27:44,075 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=268398.6666666667, ans=0.0 2024-09-23 13:27:52,229 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer_ff3.min_abs, batch_count=268445.3333333333, ans=0.2 2024-09-23 13:28:06,635 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.64 vs. limit=15.0 2024-09-23 13:28:20,804 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=4.86 vs. limit=12.0 2024-09-23 13:28:22,713 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=13.31 vs. limit=22.5 2024-09-23 13:28:23,347 INFO [train.py:1198] (2/4) Epoch 15, batch 3000, loss[loss=0.2238, ctc_loss=0.1494, cr_loss=0.3724, over 17148.00 frames. ], tot_loss[loss=0.2282, ctc_loss=0.1544, cr_loss=0.369, over 3358766.12 frames. ], batch size: 48, lr: 7.97e-03, grad_scale: 16.0 2024-09-23 13:28:23,347 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 13:28:32,005 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.1.encoder.layers.1.self_attn_weights, attn_weights_entropy = tensor([4.9854, 4.7832, 3.9892, 4.6887], device='cuda:2') 2024-09-23 13:28:36,676 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.4.encoder.layers.2.self_attn_weights, attn_weights_entropy = tensor([3.5502, 2.9644, 3.1292, 3.2567], device='cuda:2') 2024-09-23 13:28:38,968 INFO [train.py:1230] (2/4) Epoch 15, validation: loss=0.04166, ctc_loss=0.04166, cr_loss=7.464e-15, over 944034.00 frames. 2024-09-23 13:28:38,969 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 13:29:02,495 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.093e+02 1.290e+02 1.376e+02 1.476e+02 2.234e+02, threshold=2.753e+02, percent-clipped=0.0 2024-09-23 13:29:13,976 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=4.36 vs. limit=12.0 2024-09-23 13:29:27,823 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=268678.6666666667, ans=0.0 2024-09-23 13:29:39,246 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=8.87 vs. limit=15.0 2024-09-23 13:29:52,419 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=268725.3333333333, ans=0.125 2024-09-23 13:29:53,879 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=268725.3333333333, ans=0.125 2024-09-23 13:29:56,909 INFO [train.py:1198] (2/4) Epoch 15, batch 3050, loss[loss=0.2263, ctc_loss=0.1541, cr_loss=0.361, over 17229.00 frames. ], tot_loss[loss=0.2288, ctc_loss=0.1549, cr_loss=0.3695, over 3356790.88 frames. ], batch size: 50, lr: 7.96e-03, grad_scale: 16.0 2024-09-23 13:30:16,343 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 13:30:17,780 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=268818.6666666667, ans=0.125 2024-09-23 13:30:20,885 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=268818.6666666667, ans=0.2 2024-09-23 13:30:35,155 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=268865.3333333333, ans=0.0 2024-09-23 13:31:22,605 INFO [train.py:1198] (2/4) Epoch 15, batch 3100, loss[loss=0.2196, ctc_loss=0.1474, cr_loss=0.3611, over 17032.00 frames. ], tot_loss[loss=0.2283, ctc_loss=0.1545, cr_loss=0.3691, over 3359617.99 frames. ], batch size: 53, lr: 7.96e-03, grad_scale: 16.0 2024-09-23 13:31:34,092 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.17 vs. limit=6.0 2024-09-23 13:31:34,435 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=5.00 vs. limit=10.0 2024-09-23 13:31:36,952 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=269052.0, ans=0.125 2024-09-23 13:31:43,308 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=269052.0, ans=0.025 2024-09-23 13:31:46,076 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.323e+02 1.397e+02 1.534e+02 2.167e+02, threshold=2.795e+02, percent-clipped=0.0 2024-09-23 13:31:51,224 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 13:32:00,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=269098.6666666667, ans=0.125 2024-09-23 13:32:38,185 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=269192.0, ans=0.0 2024-09-23 13:32:41,148 INFO [train.py:1198] (2/4) Epoch 15, batch 3150, loss[loss=0.1841, ctc_loss=0.1217, cr_loss=0.3122, over 17227.00 frames. ], tot_loss[loss=0.2274, ctc_loss=0.1538, cr_loss=0.368, over 3363779.90 frames. ], batch size: 41, lr: 7.96e-03, grad_scale: 16.0 2024-09-23 13:32:56,901 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=269285.3333333333, ans=0.125 2024-09-23 13:33:00,300 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.00 vs. limit=10.0 2024-09-23 13:33:18,828 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=269332.0, ans=0.025 2024-09-23 13:33:59,101 INFO [train.py:1198] (2/4) Epoch 15, batch 3200, loss[loss=0.2416, ctc_loss=0.1622, cr_loss=0.3969, over 17018.00 frames. ], tot_loss[loss=0.2279, ctc_loss=0.1542, cr_loss=0.3684, over 3363379.85 frames. ], batch size: 53, lr: 7.95e-03, grad_scale: 32.0 2024-09-23 13:34:07,073 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=269472.0, ans=0.0 2024-09-23 13:34:22,377 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.268e+02 1.359e+02 1.495e+02 2.696e+02, threshold=2.719e+02, percent-clipped=0.0 2024-09-23 13:34:35,242 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=269565.3333333333, ans=0.0 2024-09-23 13:34:50,883 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=269612.0, ans=0.125 2024-09-23 13:34:53,972 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=269612.0, ans=0.0 2024-09-23 13:35:17,024 INFO [train.py:1198] (2/4) Epoch 15, batch 3250, loss[loss=0.2356, ctc_loss=0.1568, cr_loss=0.3943, over 17222.00 frames. ], tot_loss[loss=0.2287, ctc_loss=0.1548, cr_loss=0.3697, over 3364923.80 frames. ], batch size: 50, lr: 7.95e-03, grad_scale: 32.0 2024-09-23 13:35:30,391 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=269705.3333333333, ans=0.1 2024-09-23 13:35:43,142 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.whiten.whitening_limit, batch_count=269752.0, ans=12.0 2024-09-23 13:35:51,413 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=269798.6666666667, ans=0.2 2024-09-23 13:35:58,934 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=269798.6666666667, ans=0.125 2024-09-23 13:36:01,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=269798.6666666667, ans=0.0 2024-09-23 13:36:03,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.max_abs, batch_count=269798.6666666667, ans=10.0 2024-09-23 13:36:16,187 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=269845.3333333333, ans=0.125 2024-09-23 13:36:37,978 INFO [train.py:1198] (2/4) Epoch 15, batch 3300, loss[loss=0.2063, ctc_loss=0.1386, cr_loss=0.3387, over 17197.00 frames. ], tot_loss[loss=0.2273, ctc_loss=0.1536, cr_loss=0.368, over 3374295.99 frames. ], batch size: 41, lr: 7.95e-03, grad_scale: 32.0 2024-09-23 13:36:38,345 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=269938.6666666667, ans=0.125 2024-09-23 13:36:51,439 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=11.26 vs. limit=15.0 2024-09-23 13:37:01,532 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.092e+02 1.277e+02 1.362e+02 1.536e+02 4.994e+02, threshold=2.723e+02, percent-clipped=1.0 2024-09-23 13:37:14,776 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=11.69 vs. limit=15.0 2024-09-23 13:37:15,672 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=270032.0, ans=0.125 2024-09-23 13:37:41,025 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=7.04 vs. limit=15.0 2024-09-23 13:37:55,869 INFO [train.py:1198] (2/4) Epoch 15, batch 3350, loss[loss=0.2947, ctc_loss=0.2077, cr_loss=0.435, over 16176.00 frames. ], tot_loss[loss=0.2279, ctc_loss=0.1543, cr_loss=0.3684, over 3369827.90 frames. ], batch size: 74, lr: 7.94e-03, grad_scale: 32.0 2024-09-23 13:38:03,791 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=270172.0, ans=0.0 2024-09-23 13:38:03,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=270172.0, ans=0.07 2024-09-23 13:38:28,912 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=270265.3333333333, ans=0.125 2024-09-23 13:38:30,832 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=7.77 vs. limit=15.0 2024-09-23 13:38:39,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=270265.3333333333, ans=0.125 2024-09-23 13:38:41,194 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=270312.0, ans=0.0 2024-09-23 13:38:47,451 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=270312.0, ans=0.035 2024-09-23 13:39:14,019 INFO [train.py:1198] (2/4) Epoch 15, batch 3400, loss[loss=0.259, ctc_loss=0.1779, cr_loss=0.4057, over 16572.00 frames. ], tot_loss[loss=0.2297, ctc_loss=0.1558, cr_loss=0.3694, over 3347373.96 frames. ], batch size: 66, lr: 7.94e-03, grad_scale: 32.0 2024-09-23 13:39:23,637 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=270405.3333333333, ans=0.0 2024-09-23 13:39:37,359 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.143e+02 1.273e+02 1.371e+02 1.521e+02 2.083e+02, threshold=2.741e+02, percent-clipped=0.0 2024-09-23 13:39:39,290 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=270452.0, ans=0.0 2024-09-23 13:40:11,548 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.90 vs. limit=6.0 2024-09-23 13:40:26,283 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=270592.0, ans=0.125 2024-09-23 13:40:32,268 INFO [train.py:1198] (2/4) Epoch 15, batch 3450, loss[loss=0.2092, ctc_loss=0.1412, cr_loss=0.34, over 17013.00 frames. ], tot_loss[loss=0.2298, ctc_loss=0.1559, cr_loss=0.3694, over 3338328.04 frames. ], batch size: 44, lr: 7.94e-03, grad_scale: 32.0 2024-09-23 13:40:59,368 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=270685.3333333333, ans=0.0 2024-09-23 13:41:07,497 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=270732.0, ans=0.015 2024-09-23 13:41:09,694 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.88 vs. limit=6.0 2024-09-23 13:41:09,815 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=17.84 vs. limit=22.5 2024-09-23 13:41:13,354 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten.whitening_limit, batch_count=270732.0, ans=15.0 2024-09-23 13:41:16,163 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=270732.0, ans=0.5 2024-09-23 13:41:26,095 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=270778.6666666667, ans=0.1 2024-09-23 13:41:35,574 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=12.25 vs. limit=12.0 2024-09-23 13:41:36,905 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=270778.6666666667, ans=0.125 2024-09-23 13:41:56,575 INFO [train.py:1198] (2/4) Epoch 15, batch 3500, loss[loss=0.2317, ctc_loss=0.1577, cr_loss=0.3704, over 17318.00 frames. ], tot_loss[loss=0.229, ctc_loss=0.1552, cr_loss=0.3688, over 3340350.28 frames. ], batch size: 51, lr: 7.93e-03, grad_scale: 32.0 2024-09-23 13:42:12,432 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=270918.6666666667, ans=0.125 2024-09-23 13:42:19,966 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.027e+02 1.268e+02 1.406e+02 1.575e+02 2.426e+02, threshold=2.811e+02, percent-clipped=0.0 2024-09-23 13:42:26,864 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten.whitening_limit, batch_count=270965.3333333333, ans=15.0 2024-09-23 13:42:52,903 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=5.85 vs. limit=12.0 2024-09-23 13:43:11,483 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=8.63 vs. limit=15.0 2024-09-23 13:43:15,425 INFO [train.py:1198] (2/4) Epoch 15, batch 3550, loss[loss=0.2132, ctc_loss=0.1398, cr_loss=0.3673, over 17300.00 frames. ], tot_loss[loss=0.2293, ctc_loss=0.1554, cr_loss=0.3693, over 3349700.91 frames. ], batch size: 46, lr: 7.93e-03, grad_scale: 32.0 2024-09-23 13:43:43,242 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.71 vs. limit=5.0 2024-09-23 13:43:57,004 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=7.50 vs. limit=15.0 2024-09-23 13:44:07,127 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=271245.3333333333, ans=0.2 2024-09-23 13:44:13,877 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.17 vs. limit=15.0 2024-09-23 13:44:21,484 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=271292.0, ans=0.5 2024-09-23 13:44:31,843 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=9.08 vs. limit=15.0 2024-09-23 13:44:33,800 INFO [train.py:1198] (2/4) Epoch 15, batch 3600, loss[loss=0.2379, ctc_loss=0.1643, cr_loss=0.368, over 16878.00 frames. ], tot_loss[loss=0.2293, ctc_loss=0.1554, cr_loss=0.3695, over 3351068.16 frames. ], batch size: 58, lr: 7.93e-03, grad_scale: 32.0 2024-09-23 13:44:36,069 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.70 vs. limit=15.0 2024-09-23 13:44:57,245 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.121e+02 1.288e+02 1.401e+02 1.561e+02 2.194e+02, threshold=2.802e+02, percent-clipped=0.0 2024-09-23 13:45:28,929 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=271478.6666666667, ans=0.125 2024-09-23 13:45:53,864 INFO [train.py:1198] (2/4) Epoch 15, batch 3650, loss[loss=0.2342, ctc_loss=0.1576, cr_loss=0.3831, over 17014.00 frames. ], tot_loss[loss=0.2298, ctc_loss=0.1556, cr_loss=0.3709, over 3358409.74 frames. ], batch size: 53, lr: 7.92e-03, grad_scale: 16.0 2024-09-23 13:46:17,771 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=11.41 vs. limit=12.0 2024-09-23 13:46:29,861 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=271665.3333333333, ans=0.0 2024-09-23 13:46:31,330 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=271665.3333333333, ans=0.125 2024-09-23 13:46:33,800 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=6.71 vs. limit=15.0 2024-09-23 13:46:51,784 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=271712.0, ans=0.125 2024-09-23 13:47:12,692 INFO [train.py:1198] (2/4) Epoch 15, batch 3700, loss[loss=0.2153, ctc_loss=0.1463, cr_loss=0.345, over 17342.00 frames. ], tot_loss[loss=0.2292, ctc_loss=0.1551, cr_loss=0.3703, over 3352871.89 frames. ], batch size: 52, lr: 7.92e-03, grad_scale: 16.0 2024-09-23 13:47:37,539 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.112e+02 1.268e+02 1.357e+02 1.522e+02 2.745e+02, threshold=2.715e+02, percent-clipped=0.0 2024-09-23 13:48:27,138 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=271992.0, ans=0.0 2024-09-23 13:48:31,495 INFO [train.py:1198] (2/4) Epoch 15, batch 3750, loss[loss=0.238, ctc_loss=0.1597, cr_loss=0.3915, over 16985.00 frames. ], tot_loss[loss=0.2295, ctc_loss=0.1556, cr_loss=0.3695, over 3329745.10 frames. ], batch size: 56, lr: 7.92e-03, grad_scale: 16.0 2024-09-23 13:48:34,959 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=272038.6666666667, ans=0.1 2024-09-23 13:48:44,283 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=272038.6666666667, ans=0.125 2024-09-23 13:49:00,246 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=272085.3333333333, ans=0.2 2024-09-23 13:49:04,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=272132.0, ans=0.0 2024-09-23 13:49:04,912 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=272132.0, ans=0.0 2024-09-23 13:49:07,335 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=4.81 vs. limit=10.0 2024-09-23 13:49:22,407 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=272178.6666666667, ans=0.125 2024-09-23 13:49:33,437 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=272225.3333333333, ans=0.2 2024-09-23 13:49:38,684 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.36 vs. limit=10.0 2024-09-23 13:49:41,392 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=272225.3333333333, ans=0.125 2024-09-23 13:49:42,833 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=272225.3333333333, ans=0.0 2024-09-23 13:49:50,617 INFO [train.py:1198] (2/4) Epoch 15, batch 3800, loss[loss=0.1975, ctc_loss=0.1318, cr_loss=0.3285, over 17185.00 frames. ], tot_loss[loss=0.2296, ctc_loss=0.1558, cr_loss=0.3693, over 3319726.60 frames. ], batch size: 41, lr: 7.91e-03, grad_scale: 16.0 2024-09-23 13:49:53,932 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=272272.0, ans=0.0 2024-09-23 13:50:16,159 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.102e+02 1.263e+02 1.390e+02 1.534e+02 1.887e+02, threshold=2.779e+02, percent-clipped=0.0 2024-09-23 13:50:19,638 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=272318.6666666667, ans=0.025 2024-09-23 13:51:08,737 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=272458.6666666667, ans=0.04949747468305833 2024-09-23 13:51:11,675 INFO [train.py:1198] (2/4) Epoch 15, batch 3850, loss[loss=0.276, ctc_loss=0.1907, cr_loss=0.4267, over 12094.00 frames. ], tot_loss[loss=0.2312, ctc_loss=0.1573, cr_loss=0.3693, over 3265597.64 frames. ], batch size: 123, lr: 7.91e-03, grad_scale: 16.0 2024-09-23 13:51:12,691 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=15.46 vs. limit=22.5 2024-09-23 13:52:08,873 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=7.34 vs. limit=15.0 2024-09-23 13:52:11,256 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.max_abs, batch_count=272692.0, ans=10.0 2024-09-23 13:52:17,234 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=272692.0, ans=0.0 2024-09-23 13:53:14,133 INFO [train.py:1198] (2/4) Epoch 16, batch 0, loss[loss=0.2426, ctc_loss=0.1672, cr_loss=0.3769, over 17346.00 frames. ], tot_loss[loss=0.2426, ctc_loss=0.1672, cr_loss=0.3769, over 17346.00 frames. ], batch size: 48, lr: 7.65e-03, grad_scale: 32.0 2024-09-23 13:53:14,133 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 13:53:30,198 INFO [train.py:1230] (2/4) Epoch 16, validation: loss=0.04222, ctc_loss=0.04222, cr_loss=7.738e-15, over 944034.00 frames. 2024-09-23 13:53:30,198 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 13:53:30,309 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=272720.0, ans=0.125 2024-09-23 13:53:37,208 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.46 vs. limit=22.5 2024-09-23 13:53:44,723 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=272766.6666666667, ans=0.2 2024-09-23 13:54:02,002 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.157e+02 1.456e+02 1.611e+02 1.770e+02 2.340e+02, threshold=3.223e+02, percent-clipped=0.0 2024-09-23 13:54:15,834 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=5.06 vs. limit=15.0 2024-09-23 13:54:50,284 INFO [train.py:1198] (2/4) Epoch 16, batch 50, loss[loss=0.248, ctc_loss=0.1689, cr_loss=0.3956, over 17223.00 frames. ], tot_loss[loss=0.2284, ctc_loss=0.1545, cr_loss=0.3693, over 757670.66 frames. ], batch size: 50, lr: 7.65e-03, grad_scale: 32.0 2024-09-23 13:55:05,167 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 13:55:32,112 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.27 vs. limit=6.0 2024-09-23 13:55:33,341 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=273046.6666666667, ans=0.125 2024-09-23 13:56:09,765 INFO [train.py:1198] (2/4) Epoch 16, batch 100, loss[loss=0.2119, ctc_loss=0.1395, cr_loss=0.362, over 16945.00 frames. ], tot_loss[loss=0.2277, ctc_loss=0.1536, cr_loss=0.3707, over 1337982.78 frames. ], batch size: 42, lr: 7.65e-03, grad_scale: 32.0 2024-09-23 13:56:11,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=273186.6666666667, ans=0.125 2024-09-23 13:56:13,119 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=273186.6666666667, ans=0.1 2024-09-23 13:56:51,568 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.239e+02 1.320e+02 1.510e+02 1.777e+02, threshold=2.639e+02, percent-clipped=0.0 2024-09-23 13:57:00,175 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=12.13 vs. limit=15.0 2024-09-23 13:57:20,323 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=273326.6666666667, ans=0.09899494936611666 2024-09-23 13:57:29,778 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=273373.3333333333, ans=0.125 2024-09-23 13:57:38,811 INFO [train.py:1198] (2/4) Epoch 16, batch 150, loss[loss=0.1646, ctc_loss=0.1086, cr_loss=0.2802, over 17267.00 frames. ], tot_loss[loss=0.2252, ctc_loss=0.1519, cr_loss=0.3663, over 1792360.86 frames. ], batch size: 42, lr: 7.64e-03, grad_scale: 32.0 2024-09-23 13:58:44,016 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.26 vs. limit=15.0 2024-09-23 13:58:51,493 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=273606.6666666667, ans=0.125 2024-09-23 13:58:56,785 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.09 vs. limit=10.0 2024-09-23 13:58:59,125 INFO [train.py:1198] (2/4) Epoch 16, batch 200, loss[loss=0.1939, ctc_loss=0.1275, cr_loss=0.3317, over 16966.00 frames. ], tot_loss[loss=0.2258, ctc_loss=0.1523, cr_loss=0.3672, over 2143254.05 frames. ], batch size: 42, lr: 7.64e-03, grad_scale: 32.0 2024-09-23 13:59:10,434 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=273653.3333333333, ans=0.1 2024-09-23 13:59:16,667 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=273700.0, ans=0.04949747468305833 2024-09-23 13:59:21,447 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=273700.0, ans=0.125 2024-09-23 13:59:31,012 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.259e+02 1.374e+02 1.473e+02 2.348e+02, threshold=2.748e+02, percent-clipped=0.0 2024-09-23 13:59:31,413 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=273746.6666666667, ans=0.07 2024-09-23 13:59:42,530 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 13:59:44,035 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=273746.6666666667, ans=0.125 2024-09-23 13:59:52,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=273793.3333333333, ans=0.125 2024-09-23 14:00:14,485 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=273840.0, ans=0.1 2024-09-23 14:00:18,769 INFO [train.py:1198] (2/4) Epoch 16, batch 250, loss[loss=0.2104, ctc_loss=0.1387, cr_loss=0.3586, over 17046.00 frames. ], tot_loss[loss=0.225, ctc_loss=0.1518, cr_loss=0.3659, over 2415721.18 frames. ], batch size: 39, lr: 7.64e-03, grad_scale: 32.0 2024-09-23 14:00:20,715 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=273886.6666666667, ans=0.125 2024-09-23 14:00:54,279 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=9.71 vs. limit=15.0 2024-09-23 14:01:46,158 INFO [train.py:1198] (2/4) Epoch 16, batch 300, loss[loss=0.2466, ctc_loss=0.1651, cr_loss=0.4074, over 17064.00 frames. ], tot_loss[loss=0.2261, ctc_loss=0.1527, cr_loss=0.3671, over 2622132.95 frames. ], batch size: 46, lr: 7.63e-03, grad_scale: 32.0 2024-09-23 14:02:03,934 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=274166.6666666667, ans=0.0 2024-09-23 14:02:22,714 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.271e+02 1.353e+02 1.525e+02 2.781e+02, threshold=2.705e+02, percent-clipped=1.0 2024-09-23 14:02:59,790 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=274306.6666666667, ans=0.125 2024-09-23 14:03:08,654 INFO [train.py:1198] (2/4) Epoch 16, batch 350, loss[loss=0.2305, ctc_loss=0.1534, cr_loss=0.3857, over 17082.00 frames. ], tot_loss[loss=0.2271, ctc_loss=0.1536, cr_loss=0.3678, over 2766124.65 frames. ], batch size: 49, lr: 7.63e-03, grad_scale: 16.0 2024-09-23 14:03:55,061 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=274493.3333333333, ans=0.1 2024-09-23 14:04:08,803 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=7.01 vs. limit=15.0 2024-09-23 14:04:28,535 INFO [train.py:1198] (2/4) Epoch 16, batch 400, loss[loss=0.1817, ctc_loss=0.1186, cr_loss=0.3153, over 17074.00 frames. ], tot_loss[loss=0.2276, ctc_loss=0.1538, cr_loss=0.3688, over 2896866.90 frames. ], batch size: 43, lr: 7.63e-03, grad_scale: 32.0 2024-09-23 14:04:43,027 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=274633.3333333333, ans=0.0 2024-09-23 14:05:01,857 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.049e+02 1.271e+02 1.344e+02 1.500e+02 2.841e+02, threshold=2.689e+02, percent-clipped=1.0 2024-09-23 14:05:38,930 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=274773.3333333333, ans=0.125 2024-09-23 14:05:47,970 INFO [train.py:1198] (2/4) Epoch 16, batch 450, loss[loss=0.213, ctc_loss=0.1448, cr_loss=0.3411, over 17146.00 frames. ], tot_loss[loss=0.2273, ctc_loss=0.1538, cr_loss=0.3675, over 2993032.17 frames. ], batch size: 48, lr: 7.62e-03, grad_scale: 32.0 2024-09-23 14:06:12,085 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=274866.6666666667, ans=0.125 2024-09-23 14:06:12,087 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=274866.6666666667, ans=0.0 2024-09-23 14:06:31,694 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=274913.3333333333, ans=0.125 2024-09-23 14:06:38,185 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=274913.3333333333, ans=0.125 2024-09-23 14:06:53,881 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=274960.0, ans=0.125 2024-09-23 14:06:58,400 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=275006.6666666667, ans=0.125 2024-09-23 14:07:03,749 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_whiten.whitening_limit, batch_count=275006.6666666667, ans=15.0 2024-09-23 14:07:15,703 INFO [train.py:1198] (2/4) Epoch 16, batch 500, loss[loss=0.2216, ctc_loss=0.149, cr_loss=0.3626, over 17226.00 frames. ], tot_loss[loss=0.2268, ctc_loss=0.1534, cr_loss=0.3671, over 3075403.36 frames. ], batch size: 50, lr: 7.62e-03, grad_scale: 16.0 2024-09-23 14:07:15,978 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=275053.3333333333, ans=0.125 2024-09-23 14:07:26,130 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=11.38 vs. limit=15.0 2024-09-23 14:07:33,399 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=275100.0, ans=0.125 2024-09-23 14:07:46,345 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=275146.6666666667, ans=0.0 2024-09-23 14:07:52,532 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.254e+02 1.336e+02 1.469e+02 8.615e+02, threshold=2.672e+02, percent-clipped=1.0 2024-09-23 14:08:09,051 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=275193.3333333333, ans=0.125 2024-09-23 14:08:16,305 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.63 vs. limit=6.0 2024-09-23 14:08:35,387 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.86 vs. limit=15.0 2024-09-23 14:08:36,359 INFO [train.py:1198] (2/4) Epoch 16, batch 550, loss[loss=0.2469, ctc_loss=0.1682, cr_loss=0.3931, over 17101.00 frames. ], tot_loss[loss=0.2284, ctc_loss=0.1546, cr_loss=0.3693, over 3130786.12 frames. ], batch size: 49, lr: 7.62e-03, grad_scale: 8.0 2024-09-23 14:08:47,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=275286.6666666667, ans=0.0 2024-09-23 14:08:59,015 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=275333.3333333333, ans=0.0 2024-09-23 14:09:56,354 INFO [train.py:1198] (2/4) Epoch 16, batch 600, loss[loss=0.2301, ctc_loss=0.1522, cr_loss=0.3895, over 17234.00 frames. ], tot_loss[loss=0.2288, ctc_loss=0.1548, cr_loss=0.3698, over 3184053.99 frames. ], batch size: 50, lr: 7.61e-03, grad_scale: 8.0 2024-09-23 14:10:03,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=275520.0, ans=0.125 2024-09-23 14:10:32,682 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.280e+02 1.400e+02 1.549e+02 2.106e+02, threshold=2.799e+02, percent-clipped=0.0 2024-09-23 14:10:50,291 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=275660.0, ans=0.07 2024-09-23 14:11:21,416 INFO [train.py:1198] (2/4) Epoch 16, batch 650, loss[loss=0.2577, ctc_loss=0.1764, cr_loss=0.4063, over 15410.00 frames. ], tot_loss[loss=0.2276, ctc_loss=0.1539, cr_loss=0.3685, over 3225620.08 frames. ], batch size: 89, lr: 7.61e-03, grad_scale: 8.0 2024-09-23 14:11:24,975 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=275753.3333333333, ans=0.0 2024-09-23 14:11:45,114 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=275800.0, ans=0.0 2024-09-23 14:12:28,676 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=8.45 vs. limit=15.0 2024-09-23 14:12:39,242 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=275940.0, ans=0.125 2024-09-23 14:12:43,638 INFO [train.py:1198] (2/4) Epoch 16, batch 700, loss[loss=0.2312, ctc_loss=0.1573, cr_loss=0.3695, over 17346.00 frames. ], tot_loss[loss=0.2266, ctc_loss=0.1531, cr_loss=0.3676, over 3258779.64 frames. ], batch size: 48, lr: 7.61e-03, grad_scale: 8.0 2024-09-23 14:12:56,828 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=275986.6666666667, ans=0.125 2024-09-23 14:13:20,906 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.058e+02 1.263e+02 1.367e+02 1.500e+02 2.228e+02, threshold=2.734e+02, percent-clipped=0.0 2024-09-23 14:13:31,097 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 14:13:49,944 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=276173.3333333333, ans=0.1 2024-09-23 14:14:03,947 INFO [train.py:1198] (2/4) Epoch 16, batch 750, loss[loss=0.2206, ctc_loss=0.1469, cr_loss=0.3683, over 17227.00 frames. ], tot_loss[loss=0.2273, ctc_loss=0.1536, cr_loss=0.3685, over 3289698.85 frames. ], batch size: 50, lr: 7.61e-03, grad_scale: 8.0 2024-09-23 14:14:31,544 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=276266.6666666667, ans=0.2 2024-09-23 14:14:36,767 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.55 vs. limit=15.0 2024-09-23 14:15:13,269 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=276406.6666666667, ans=0.2 2024-09-23 14:15:14,766 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=276406.6666666667, ans=0.0 2024-09-23 14:15:17,089 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=6.00 vs. limit=12.0 2024-09-23 14:15:23,830 INFO [train.py:1198] (2/4) Epoch 16, batch 800, loss[loss=0.2491, ctc_loss=0.1715, cr_loss=0.3879, over 14965.00 frames. ], tot_loss[loss=0.2275, ctc_loss=0.1539, cr_loss=0.368, over 3290921.02 frames. ], batch size: 89, lr: 7.60e-03, grad_scale: 16.0 2024-09-23 14:15:43,378 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=276500.0, ans=0.125 2024-09-23 14:15:52,805 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=276500.0, ans=0.2 2024-09-23 14:15:55,893 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=276546.6666666667, ans=0.0 2024-09-23 14:16:00,467 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.161e+02 1.250e+02 1.343e+02 1.464e+02 3.040e+02, threshold=2.687e+02, percent-clipped=1.0 2024-09-23 14:16:14,095 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 14:16:51,108 INFO [train.py:1198] (2/4) Epoch 16, batch 850, loss[loss=0.2441, ctc_loss=0.1646, cr_loss=0.3974, over 16540.00 frames. ], tot_loss[loss=0.2268, ctc_loss=0.1532, cr_loss=0.3679, over 3316055.35 frames. ], batch size: 66, lr: 7.60e-03, grad_scale: 16.0 2024-09-23 14:16:52,942 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=276686.6666666667, ans=0.2 2024-09-23 14:16:56,872 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.63 vs. limit=15.0 2024-09-23 14:17:08,848 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=276733.3333333333, ans=0.0 2024-09-23 14:17:27,995 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=276780.0, ans=0.1 2024-09-23 14:17:48,493 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=12.36 vs. limit=22.5 2024-09-23 14:18:02,268 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=12.17 vs. limit=15.0 2024-09-23 14:18:11,440 INFO [train.py:1198] (2/4) Epoch 16, batch 900, loss[loss=0.2192, ctc_loss=0.1465, cr_loss=0.3634, over 17276.00 frames. ], tot_loss[loss=0.227, ctc_loss=0.1533, cr_loss=0.3682, over 3312609.40 frames. ], batch size: 44, lr: 7.60e-03, grad_scale: 16.0 2024-09-23 14:18:21,396 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=276920.0, ans=0.2 2024-09-23 14:18:22,842 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=276920.0, ans=0.125 2024-09-23 14:18:22,970 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=276920.0, ans=0.0 2024-09-23 14:18:24,926 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=17.73 vs. limit=22.5 2024-09-23 14:18:38,759 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=276966.6666666667, ans=0.125 2024-09-23 14:18:47,947 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.062e+02 1.278e+02 1.412e+02 1.648e+02 4.971e+02, threshold=2.824e+02, percent-clipped=1.0 2024-09-23 14:18:56,462 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.61 vs. limit=22.5 2024-09-23 14:19:02,671 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=277060.0, ans=0.125 2024-09-23 14:19:09,091 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys.whitening_limit, batch_count=277060.0, ans=6.0 2024-09-23 14:19:20,004 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=277106.6666666667, ans=0.1 2024-09-23 14:19:23,059 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=277106.6666666667, ans=0.125 2024-09-23 14:19:30,870 INFO [train.py:1198] (2/4) Epoch 16, batch 950, loss[loss=0.2371, ctc_loss=0.1601, cr_loss=0.3852, over 17027.00 frames. ], tot_loss[loss=0.2281, ctc_loss=0.1542, cr_loss=0.3692, over 3317287.31 frames. ], batch size: 56, lr: 7.59e-03, grad_scale: 16.0 2024-09-23 14:20:03,021 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=277246.6666666667, ans=0.2 2024-09-23 14:20:50,273 INFO [train.py:1198] (2/4) Epoch 16, batch 1000, loss[loss=0.2035, ctc_loss=0.1323, cr_loss=0.356, over 17182.00 frames. ], tot_loss[loss=0.2267, ctc_loss=0.1531, cr_loss=0.3679, over 3322115.03 frames. ], batch size: 41, lr: 7.59e-03, grad_scale: 16.0 2024-09-23 14:21:02,216 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=277386.6666666667, ans=0.0 2024-09-23 14:21:11,132 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=277386.6666666667, ans=0.04949747468305833 2024-09-23 14:21:17,413 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=277433.3333333333, ans=0.125 2024-09-23 14:21:25,566 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=277433.3333333333, ans=0.09899494936611666 2024-09-23 14:21:28,895 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=277480.0, ans=0.0 2024-09-23 14:21:32,803 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=3.63 vs. limit=15.0 2024-09-23 14:21:37,363 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.289e+02 1.371e+02 1.510e+02 2.522e+02, threshold=2.742e+02, percent-clipped=0.0 2024-09-23 14:21:39,316 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=277480.0, ans=0.025 2024-09-23 14:21:40,758 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=277480.0, ans=0.125 2024-09-23 14:21:45,456 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=277480.0, ans=0.0 2024-09-23 14:21:47,022 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=277526.6666666667, ans=0.025 2024-09-23 14:21:50,273 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=277526.6666666667, ans=0.1 2024-09-23 14:21:51,755 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=277526.6666666667, ans=0.04949747468305833 2024-09-23 14:21:53,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=277526.6666666667, ans=0.04949747468305833 2024-09-23 14:22:20,155 INFO [train.py:1198] (2/4) Epoch 16, batch 1050, loss[loss=0.2549, ctc_loss=0.1737, cr_loss=0.406, over 16489.00 frames. ], tot_loss[loss=0.227, ctc_loss=0.1532, cr_loss=0.3686, over 3331565.77 frames. ], batch size: 66, lr: 7.59e-03, grad_scale: 16.0 2024-09-23 14:22:22,181 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=277620.0, ans=0.125 2024-09-23 14:22:51,065 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten.whitening_limit, batch_count=277713.3333333333, ans=22.5 2024-09-23 14:23:39,992 INFO [train.py:1198] (2/4) Epoch 16, batch 1100, loss[loss=0.1977, ctc_loss=0.135, cr_loss=0.3135, over 17259.00 frames. ], tot_loss[loss=0.2277, ctc_loss=0.1539, cr_loss=0.369, over 3331627.02 frames. ], batch size: 42, lr: 7.58e-03, grad_scale: 16.0 2024-09-23 14:23:44,971 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=277853.3333333333, ans=0.125 2024-09-23 14:23:48,632 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=11.63 vs. limit=15.0 2024-09-23 14:23:49,861 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=277853.3333333333, ans=0.1 2024-09-23 14:23:53,049 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=277853.3333333333, ans=0.0 2024-09-23 14:24:16,322 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.065e+02 1.265e+02 1.345e+02 1.521e+02 2.827e+02, threshold=2.691e+02, percent-clipped=1.0 2024-09-23 14:24:24,616 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=277946.6666666667, ans=0.125 2024-09-23 14:24:46,253 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=6.85 vs. limit=15.0 2024-09-23 14:24:53,400 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=278040.0, ans=0.2 2024-09-23 14:24:56,638 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=278040.0, ans=0.2 2024-09-23 14:24:59,427 INFO [train.py:1198] (2/4) Epoch 16, batch 1150, loss[loss=0.2316, ctc_loss=0.1578, cr_loss=0.3689, over 17368.00 frames. ], tot_loss[loss=0.2269, ctc_loss=0.1532, cr_loss=0.3685, over 3345161.54 frames. ], batch size: 48, lr: 7.58e-03, grad_scale: 16.0 2024-09-23 14:25:24,028 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=278133.3333333333, ans=0.0 2024-09-23 14:25:46,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=278226.6666666667, ans=0.025 2024-09-23 14:26:13,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=278273.3333333333, ans=0.125 2024-09-23 14:26:15,308 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=11.02 vs. limit=15.0 2024-09-23 14:26:27,404 INFO [train.py:1198] (2/4) Epoch 16, batch 1200, loss[loss=0.2422, ctc_loss=0.1675, cr_loss=0.3736, over 16052.00 frames. ], tot_loss[loss=0.2258, ctc_loss=0.1524, cr_loss=0.367, over 3359570.14 frames. ], batch size: 74, lr: 7.58e-03, grad_scale: 16.0 2024-09-23 14:26:27,775 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=278320.0, ans=0.125 2024-09-23 14:26:45,234 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2.whitening_limit, batch_count=278366.6666666667, ans=15.0 2024-09-23 14:26:46,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=278366.6666666667, ans=0.1 2024-09-23 14:27:08,387 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.263e+02 1.375e+02 1.495e+02 2.178e+02, threshold=2.750e+02, percent-clipped=0.0 2024-09-23 14:27:08,691 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=278413.3333333333, ans=0.07 2024-09-23 14:27:09,268 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.28 vs. limit=12.0 2024-09-23 14:27:49,956 INFO [train.py:1198] (2/4) Epoch 16, batch 1250, loss[loss=0.2224, ctc_loss=0.1503, cr_loss=0.3607, over 17154.00 frames. ], tot_loss[loss=0.2252, ctc_loss=0.1518, cr_loss=0.3669, over 3369179.85 frames. ], batch size: 45, lr: 7.57e-03, grad_scale: 16.0 2024-09-23 14:28:11,574 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.80 vs. limit=10.0 2024-09-23 14:28:24,258 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=512, metric=4.86 vs. limit=15.0 2024-09-23 14:28:28,534 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=278646.6666666667, ans=0.125 2024-09-23 14:28:31,880 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=278646.6666666667, ans=0.0 2024-09-23 14:29:10,195 INFO [train.py:1198] (2/4) Epoch 16, batch 1300, loss[loss=0.2428, ctc_loss=0.1659, cr_loss=0.3845, over 17302.00 frames. ], tot_loss[loss=0.2254, ctc_loss=0.152, cr_loss=0.3671, over 3366673.26 frames. ], batch size: 49, lr: 7.57e-03, grad_scale: 16.0 2024-09-23 14:29:19,831 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=278786.6666666667, ans=0.5 2024-09-23 14:29:48,132 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.053e+02 1.264e+02 1.359e+02 1.551e+02 2.304e+02, threshold=2.718e+02, percent-clipped=0.0 2024-09-23 14:30:12,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=278973.3333333333, ans=0.0 2024-09-23 14:30:29,687 INFO [train.py:1198] (2/4) Epoch 16, batch 1350, loss[loss=0.223, ctc_loss=0.1503, cr_loss=0.3631, over 16944.00 frames. ], tot_loss[loss=0.227, ctc_loss=0.1532, cr_loss=0.369, over 3367680.18 frames. ], batch size: 42, lr: 7.57e-03, grad_scale: 16.0 2024-09-23 14:31:02,378 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=279066.6666666667, ans=0.125 2024-09-23 14:31:38,862 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=12.91 vs. limit=15.0 2024-09-23 14:31:39,986 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=279160.0, ans=0.0 2024-09-23 14:32:00,029 INFO [train.py:1198] (2/4) Epoch 16, batch 1400, loss[loss=0.2026, ctc_loss=0.1323, cr_loss=0.3514, over 17107.00 frames. ], tot_loss[loss=0.227, ctc_loss=0.1532, cr_loss=0.3691, over 3366075.74 frames. ], batch size: 43, lr: 7.56e-03, grad_scale: 16.0 2024-09-23 14:32:16,818 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.47 vs. limit=15.0 2024-09-23 14:32:24,318 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.42 vs. limit=15.0 2024-09-23 14:32:38,350 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.056e+02 1.268e+02 1.363e+02 1.512e+02 2.184e+02, threshold=2.726e+02, percent-clipped=0.0 2024-09-23 14:32:41,760 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=279346.6666666667, ans=0.0 2024-09-23 14:32:51,651 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=279393.3333333333, ans=0.1 2024-09-23 14:33:05,071 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=13.23 vs. limit=15.0 2024-09-23 14:33:05,956 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=279440.0, ans=0.1 2024-09-23 14:33:09,151 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=279440.0, ans=0.125 2024-09-23 14:33:20,075 INFO [train.py:1198] (2/4) Epoch 16, batch 1450, loss[loss=0.2278, ctc_loss=0.1543, cr_loss=0.3678, over 17216.00 frames. ], tot_loss[loss=0.2257, ctc_loss=0.1522, cr_loss=0.3676, over 3360976.06 frames. ], batch size: 47, lr: 7.56e-03, grad_scale: 16.0 2024-09-23 14:33:21,066 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=5.22 vs. limit=10.0 2024-09-23 14:33:31,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=279486.6666666667, ans=0.125 2024-09-23 14:33:37,496 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.94 vs. limit=15.0 2024-09-23 14:33:41,347 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=279533.3333333333, ans=0.1 2024-09-23 14:34:10,130 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.96 vs. limit=15.0 2024-09-23 14:34:36,948 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=279673.3333333333, ans=0.125 2024-09-23 14:34:39,814 INFO [train.py:1198] (2/4) Epoch 16, batch 1500, loss[loss=0.243, ctc_loss=0.1635, cr_loss=0.3974, over 17359.00 frames. ], tot_loss[loss=0.2266, ctc_loss=0.1529, cr_loss=0.3685, over 3356084.32 frames. ], batch size: 48, lr: 7.56e-03, grad_scale: 16.0 2024-09-23 14:34:40,218 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=279720.0, ans=0.07 2024-09-23 14:34:43,265 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=279720.0, ans=0.2 2024-09-23 14:34:56,101 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=279766.6666666667, ans=0.025 2024-09-23 14:34:59,420 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=279766.6666666667, ans=0.125 2024-09-23 14:35:18,097 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.064e+02 1.287e+02 1.379e+02 1.521e+02 2.599e+02, threshold=2.758e+02, percent-clipped=0.0 2024-09-23 14:35:18,386 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=279813.3333333333, ans=0.125 2024-09-23 14:35:29,708 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=279860.0, ans=0.125 2024-09-23 14:35:37,989 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=18.68 vs. limit=22.5 2024-09-23 14:36:06,618 INFO [train.py:1198] (2/4) Epoch 16, batch 1550, loss[loss=0.2294, ctc_loss=0.1528, cr_loss=0.3831, over 17079.00 frames. ], tot_loss[loss=0.2274, ctc_loss=0.1536, cr_loss=0.3692, over 3358649.32 frames. ], batch size: 46, lr: 7.56e-03, grad_scale: 16.0 2024-09-23 14:36:52,958 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=280046.6666666667, ans=0.0 2024-09-23 14:36:59,356 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=280093.3333333333, ans=0.125 2024-09-23 14:37:07,224 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=280093.3333333333, ans=0.125 2024-09-23 14:37:23,238 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=280140.0, ans=0.0 2024-09-23 14:37:30,855 INFO [train.py:1198] (2/4) Epoch 16, batch 1600, loss[loss=0.2421, ctc_loss=0.1629, cr_loss=0.396, over 17020.00 frames. ], tot_loss[loss=0.2266, ctc_loss=0.1529, cr_loss=0.3684, over 3361420.41 frames. ], batch size: 51, lr: 7.55e-03, grad_scale: 32.0 2024-09-23 14:37:35,931 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=280186.6666666667, ans=0.2 2024-09-23 14:38:10,325 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.068e+02 1.239e+02 1.329e+02 1.436e+02 2.606e+02, threshold=2.657e+02, percent-clipped=0.0 2024-09-23 14:38:21,778 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=280326.6666666667, ans=0.125 2024-09-23 14:38:31,276 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.min_positive, batch_count=280326.6666666667, ans=0.05 2024-09-23 14:38:32,872 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=280373.3333333333, ans=0.125 2024-09-23 14:38:50,311 INFO [train.py:1198] (2/4) Epoch 16, batch 1650, loss[loss=0.2483, ctc_loss=0.1684, cr_loss=0.3994, over 17222.00 frames. ], tot_loss[loss=0.2265, ctc_loss=0.1529, cr_loss=0.368, over 3368428.80 frames. ], batch size: 55, lr: 7.55e-03, grad_scale: 16.0 2024-09-23 14:38:56,949 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=280420.0, ans=0.025 2024-09-23 14:39:03,383 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=280420.0, ans=0.125 2024-09-23 14:39:11,078 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=280466.6666666667, ans=0.0 2024-09-23 14:39:33,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer_ff3.min_abs, batch_count=280513.3333333333, ans=0.2 2024-09-23 14:39:49,415 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=280560.0, ans=0.125 2024-09-23 14:40:09,864 INFO [train.py:1198] (2/4) Epoch 16, batch 1700, loss[loss=0.2223, ctc_loss=0.1494, cr_loss=0.3645, over 16912.00 frames. ], tot_loss[loss=0.2262, ctc_loss=0.1528, cr_loss=0.3673, over 3367237.47 frames. ], batch size: 58, lr: 7.55e-03, grad_scale: 16.0 2024-09-23 14:40:11,809 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 14:40:14,894 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=280653.3333333333, ans=0.125 2024-09-23 14:40:18,264 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=280653.3333333333, ans=0.0 2024-09-23 14:40:24,662 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=280700.0, ans=0.125 2024-09-23 14:40:26,222 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=280700.0, ans=0.125 2024-09-23 14:40:27,845 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=280700.0, ans=0.125 2024-09-23 14:40:29,801 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.59 vs. limit=15.0 2024-09-23 14:40:32,562 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=280700.0, ans=0.125 2024-09-23 14:40:49,782 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=11.44 vs. limit=22.5 2024-09-23 14:40:52,361 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.241e+02 1.314e+02 1.415e+02 2.347e+02, threshold=2.628e+02, percent-clipped=0.0 2024-09-23 14:41:25,172 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=280840.0, ans=0.1 2024-09-23 14:41:30,336 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=280840.0, ans=0.2 2024-09-23 14:41:36,297 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.07 vs. limit=22.5 2024-09-23 14:41:40,567 INFO [train.py:1198] (2/4) Epoch 16, batch 1750, loss[loss=0.2275, ctc_loss=0.1593, cr_loss=0.341, over 16705.00 frames. ], tot_loss[loss=0.2256, ctc_loss=0.1525, cr_loss=0.3655, over 3354768.17 frames. ], batch size: 61, lr: 7.54e-03, grad_scale: 16.0 2024-09-23 14:42:04,004 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.38 vs. limit=12.0 2024-09-23 14:42:22,543 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=280980.0, ans=0.0 2024-09-23 14:42:27,434 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.max_positive, batch_count=281026.6666666667, ans=0.95 2024-09-23 14:43:00,732 INFO [train.py:1198] (2/4) Epoch 16, batch 1800, loss[loss=0.2456, ctc_loss=0.1659, cr_loss=0.3986, over 17295.00 frames. ], tot_loss[loss=0.2258, ctc_loss=0.1526, cr_loss=0.3659, over 3361358.13 frames. ], batch size: 46, lr: 7.54e-03, grad_scale: 16.0 2024-09-23 14:43:20,040 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=281166.6666666667, ans=0.05 2024-09-23 14:43:27,925 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.min_positive, batch_count=281166.6666666667, ans=0.025 2024-09-23 14:43:37,491 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=281213.3333333333, ans=0.125 2024-09-23 14:43:40,372 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.120e+02 1.265e+02 1.365e+02 1.521e+02 2.085e+02, threshold=2.729e+02, percent-clipped=0.0 2024-09-23 14:43:54,996 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 14:44:20,023 INFO [train.py:1198] (2/4) Epoch 16, batch 1850, loss[loss=0.2346, ctc_loss=0.158, cr_loss=0.383, over 17009.00 frames. ], tot_loss[loss=0.2258, ctc_loss=0.1527, cr_loss=0.3658, over 3356459.54 frames. ], batch size: 44, lr: 7.54e-03, grad_scale: 16.0 2024-09-23 14:44:49,217 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.96 vs. limit=15.0 2024-09-23 14:44:50,120 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=281446.6666666667, ans=0.1 2024-09-23 14:45:19,133 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=281493.3333333333, ans=0.1 2024-09-23 14:45:42,170 INFO [train.py:1198] (2/4) Epoch 16, batch 1900, loss[loss=0.2153, ctc_loss=0.1436, cr_loss=0.3581, over 17066.00 frames. ], tot_loss[loss=0.2263, ctc_loss=0.153, cr_loss=0.3665, over 3354595.79 frames. ], batch size: 46, lr: 7.53e-03, grad_scale: 16.0 2024-09-23 14:45:45,568 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=281586.6666666667, ans=0.0 2024-09-23 14:45:56,751 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=281586.6666666667, ans=0.125 2024-09-23 14:46:17,675 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=281680.0, ans=0.0 2024-09-23 14:46:23,999 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=281680.0, ans=0.2 2024-09-23 14:46:27,054 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.113e+02 1.270e+02 1.346e+02 1.420e+02 2.030e+02, threshold=2.692e+02, percent-clipped=0.0 2024-09-23 14:46:30,592 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=281680.0, ans=0.0 2024-09-23 14:47:03,429 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=281773.3333333333, ans=0.0 2024-09-23 14:47:09,366 INFO [train.py:1198] (2/4) Epoch 16, batch 1950, loss[loss=0.2136, ctc_loss=0.1462, cr_loss=0.3369, over 17166.00 frames. ], tot_loss[loss=0.2265, ctc_loss=0.1531, cr_loss=0.3673, over 3356377.60 frames. ], batch size: 45, lr: 7.53e-03, grad_scale: 16.0 2024-09-23 14:47:21,109 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=281820.0, ans=0.125 2024-09-23 14:47:27,733 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=281866.6666666667, ans=0.1 2024-09-23 14:47:34,096 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.min_positive, batch_count=281866.6666666667, ans=0.05 2024-09-23 14:47:35,587 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=281866.6666666667, ans=0.125 2024-09-23 14:48:18,707 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=282006.6666666667, ans=0.125 2024-09-23 14:48:26,684 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=282006.6666666667, ans=0.0 2024-09-23 14:48:29,654 INFO [train.py:1198] (2/4) Epoch 16, batch 2000, loss[loss=0.2385, ctc_loss=0.161, cr_loss=0.3872, over 17210.00 frames. ], tot_loss[loss=0.2256, ctc_loss=0.1523, cr_loss=0.3661, over 3354041.20 frames. ], batch size: 47, lr: 7.53e-03, grad_scale: 32.0 2024-09-23 14:48:30,062 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=282053.3333333333, ans=0.0 2024-09-23 14:48:39,600 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=282053.3333333333, ans=0.0 2024-09-23 14:49:07,881 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=282146.6666666667, ans=0.1 2024-09-23 14:49:09,155 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.302e+02 1.409e+02 1.610e+02 3.619e+02, threshold=2.818e+02, percent-clipped=1.0 2024-09-23 14:49:15,960 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 14:49:26,073 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=5.35 vs. limit=10.0 2024-09-23 14:49:37,019 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=282240.0, ans=0.0 2024-09-23 14:49:41,856 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=282240.0, ans=0.025 2024-09-23 14:49:49,170 INFO [train.py:1198] (2/4) Epoch 16, batch 2050, loss[loss=0.2702, ctc_loss=0.1835, cr_loss=0.4335, over 17215.00 frames. ], tot_loss[loss=0.2268, ctc_loss=0.1532, cr_loss=0.3682, over 3363521.92 frames. ], batch size: 50, lr: 7.52e-03, grad_scale: 32.0 2024-09-23 14:50:09,106 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.97 vs. limit=10.0 2024-09-23 14:50:16,901 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.99 vs. limit=15.0 2024-09-23 14:50:27,613 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 14:50:56,218 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=282426.6666666667, ans=0.125 2024-09-23 14:51:16,373 INFO [train.py:1198] (2/4) Epoch 16, batch 2100, loss[loss=0.2207, ctc_loss=0.153, cr_loss=0.3384, over 16880.00 frames. ], tot_loss[loss=0.2278, ctc_loss=0.154, cr_loss=0.3689, over 3345229.25 frames. ], batch size: 58, lr: 7.52e-03, grad_scale: 32.0 2024-09-23 14:51:33,490 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.min_positive, batch_count=282566.6666666667, ans=0.025 2024-09-23 14:51:40,125 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.46 vs. limit=15.0 2024-09-23 14:51:50,981 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=282613.3333333333, ans=0.2 2024-09-23 14:51:58,360 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.067e+02 1.277e+02 1.347e+02 1.477e+02 3.259e+02, threshold=2.693e+02, percent-clipped=1.0 2024-09-23 14:52:28,877 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=282706.6666666667, ans=0.125 2024-09-23 14:52:38,056 INFO [train.py:1198] (2/4) Epoch 16, batch 2150, loss[loss=0.2094, ctc_loss=0.1408, cr_loss=0.3426, over 17083.00 frames. ], tot_loss[loss=0.2271, ctc_loss=0.1535, cr_loss=0.3684, over 3353387.00 frames. ], batch size: 43, lr: 7.52e-03, grad_scale: 32.0 2024-09-23 14:52:52,761 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=282800.0, ans=0.2 2024-09-23 14:53:04,360 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=282800.0, ans=0.2 2024-09-23 14:53:05,953 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=282800.0, ans=0.0 2024-09-23 14:53:15,522 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=282846.6666666667, ans=0.0 2024-09-23 14:53:20,115 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.min_abs, batch_count=282846.6666666667, ans=0.5 2024-09-23 14:53:21,775 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=282846.6666666667, ans=0.025 2024-09-23 14:53:42,622 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=282940.0, ans=0.125 2024-09-23 14:53:47,477 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.84 vs. limit=10.0 2024-09-23 14:53:52,145 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 14:53:58,103 INFO [train.py:1198] (2/4) Epoch 16, batch 2200, loss[loss=0.2166, ctc_loss=0.1482, cr_loss=0.3421, over 17360.00 frames. ], tot_loss[loss=0.2253, ctc_loss=0.1521, cr_loss=0.3662, over 3361945.18 frames. ], batch size: 48, lr: 7.52e-03, grad_scale: 32.0 2024-09-23 14:54:00,454 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=11.98 vs. limit=15.0 2024-09-23 14:54:17,884 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=283033.3333333333, ans=0.125 2024-09-23 14:54:19,631 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=12.03 vs. limit=12.0 2024-09-23 14:54:21,401 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.89 vs. limit=6.0 2024-09-23 14:54:36,853 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=283080.0, ans=0.0 2024-09-23 14:54:38,169 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.245e+02 1.360e+02 1.489e+02 2.276e+02, threshold=2.720e+02, percent-clipped=0.0 2024-09-23 14:54:38,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer_na.min_abs, batch_count=283080.0, ans=0.02 2024-09-23 14:54:41,764 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=283080.0, ans=0.2 2024-09-23 14:55:18,497 INFO [train.py:1198] (2/4) Epoch 16, batch 2250, loss[loss=0.2476, ctc_loss=0.1689, cr_loss=0.3934, over 15945.00 frames. ], tot_loss[loss=0.2246, ctc_loss=0.1514, cr_loss=0.3661, over 3368413.18 frames. ], batch size: 74, lr: 7.51e-03, grad_scale: 32.0 2024-09-23 14:55:27,126 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.64 vs. limit=22.5 2024-09-23 14:55:37,762 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 14:56:02,769 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=283313.3333333333, ans=0.2 2024-09-23 14:56:12,528 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=283360.0, ans=0.1 2024-09-23 14:56:46,617 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=283453.3333333333, ans=0.125 2024-09-23 14:56:47,891 INFO [train.py:1198] (2/4) Epoch 16, batch 2300, loss[loss=0.1745, ctc_loss=0.1175, cr_loss=0.285, over 17154.00 frames. ], tot_loss[loss=0.224, ctc_loss=0.151, cr_loss=0.3649, over 3361813.98 frames. ], batch size: 41, lr: 7.51e-03, grad_scale: 32.0 2024-09-23 14:57:01,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten.whitening_limit, batch_count=283453.3333333333, ans=22.5 2024-09-23 14:57:23,896 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=18.84 vs. limit=22.5 2024-09-23 14:57:27,769 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.135e+02 1.282e+02 1.381e+02 1.559e+02 2.607e+02, threshold=2.763e+02, percent-clipped=0.0 2024-09-23 14:58:07,689 INFO [train.py:1198] (2/4) Epoch 16, batch 2350, loss[loss=0.2534, ctc_loss=0.1736, cr_loss=0.399, over 17045.00 frames. ], tot_loss[loss=0.2244, ctc_loss=0.1516, cr_loss=0.3643, over 3342505.57 frames. ], batch size: 52, lr: 7.51e-03, grad_scale: 32.0 2024-09-23 14:58:14,325 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=283686.6666666667, ans=0.1 2024-09-23 14:58:48,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=283780.0, ans=0.0 2024-09-23 14:58:59,349 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=283826.6666666667, ans=0.125 2024-09-23 14:59:06,526 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module1.whiten, num_groups=1, num_channels=192, metric=8.04 vs. limit=15.0 2024-09-23 14:59:27,621 INFO [train.py:1198] (2/4) Epoch 16, batch 2400, loss[loss=0.2553, ctc_loss=0.1739, cr_loss=0.4069, over 15888.00 frames. ], tot_loss[loss=0.2246, ctc_loss=0.1517, cr_loss=0.3645, over 3336668.49 frames. ], batch size: 74, lr: 7.50e-03, grad_scale: 32.0 2024-09-23 14:59:56,621 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=283966.6666666667, ans=0.0 2024-09-23 15:00:07,588 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.106e+02 1.274e+02 1.411e+02 1.575e+02 2.045e+02, threshold=2.822e+02, percent-clipped=0.0 2024-09-23 15:00:09,590 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=284013.3333333333, ans=0.125 2024-09-23 15:00:26,967 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=284060.0, ans=0.0 2024-09-23 15:00:42,956 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=284106.6666666667, ans=0.125 2024-09-23 15:00:44,710 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer_ff2.min_abs, batch_count=284106.6666666667, ans=0.1 2024-09-23 15:00:51,919 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=284106.6666666667, ans=0.2 2024-09-23 15:00:54,749 INFO [train.py:1198] (2/4) Epoch 16, batch 2450, loss[loss=0.2204, ctc_loss=0.1498, cr_loss=0.3533, over 17025.00 frames. ], tot_loss[loss=0.2267, ctc_loss=0.1533, cr_loss=0.3669, over 3326464.19 frames. ], batch size: 51, lr: 7.50e-03, grad_scale: 32.0 2024-09-23 15:01:07,442 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=284153.3333333333, ans=0.125 2024-09-23 15:01:43,777 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=284293.3333333333, ans=0.2 2024-09-23 15:02:15,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=284386.6666666667, ans=0.125 2024-09-23 15:02:15,673 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=284386.6666666667, ans=0.125 2024-09-23 15:02:17,031 INFO [train.py:1198] (2/4) Epoch 16, batch 2500, loss[loss=0.2327, ctc_loss=0.1611, cr_loss=0.3577, over 17360.00 frames. ], tot_loss[loss=0.2257, ctc_loss=0.1524, cr_loss=0.3664, over 3341469.52 frames. ], batch size: 48, lr: 7.50e-03, grad_scale: 32.0 2024-09-23 15:02:55,474 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=284480.0, ans=0.125 2024-09-23 15:02:56,744 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.046e+02 1.256e+02 1.352e+02 1.487e+02 2.472e+02, threshold=2.703e+02, percent-clipped=0.0 2024-09-23 15:03:09,030 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn1.whiten, num_groups=1, num_channels=192, metric=9.01 vs. limit=22.5 2024-09-23 15:03:17,785 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 15:03:32,004 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=284573.3333333333, ans=0.125 2024-09-23 15:03:36,420 INFO [train.py:1198] (2/4) Epoch 16, batch 2550, loss[loss=0.2325, ctc_loss=0.1565, cr_loss=0.3802, over 17067.00 frames. ], tot_loss[loss=0.2252, ctc_loss=0.152, cr_loss=0.3658, over 3348069.47 frames. ], batch size: 46, lr: 7.49e-03, grad_scale: 32.0 2024-09-23 15:04:04,472 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=284666.6666666667, ans=0.125 2024-09-23 15:04:20,837 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.84 vs. limit=12.0 2024-09-23 15:04:51,284 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.36 vs. limit=12.0 2024-09-23 15:04:56,919 INFO [train.py:1198] (2/4) Epoch 16, batch 2600, loss[loss=0.2161, ctc_loss=0.1464, cr_loss=0.3485, over 17295.00 frames. ], tot_loss[loss=0.2243, ctc_loss=0.1513, cr_loss=0.3648, over 3361682.70 frames. ], batch size: 46, lr: 7.49e-03, grad_scale: 32.0 2024-09-23 15:05:21,213 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=284900.0, ans=0.0 2024-09-23 15:05:41,680 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.034e+02 1.252e+02 1.364e+02 1.589e+02 2.408e+02, threshold=2.728e+02, percent-clipped=0.0 2024-09-23 15:05:50,884 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=284993.3333333333, ans=0.125 2024-09-23 15:05:51,483 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.96 vs. limit=10.0 2024-09-23 15:06:24,111 INFO [train.py:1198] (2/4) Epoch 16, batch 2650, loss[loss=0.2204, ctc_loss=0.1497, cr_loss=0.3535, over 16851.00 frames. ], tot_loss[loss=0.2239, ctc_loss=0.1509, cr_loss=0.3648, over 3359208.94 frames. ], batch size: 58, lr: 7.49e-03, grad_scale: 32.0 2024-09-23 15:06:47,927 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=285133.3333333333, ans=0.0 2024-09-23 15:07:16,925 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=285226.6666666667, ans=0.125 2024-09-23 15:07:20,518 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=10.59 vs. limit=15.0 2024-09-23 15:07:35,948 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=285273.3333333333, ans=0.125 2024-09-23 15:07:45,570 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=285320.0, ans=0.0 2024-09-23 15:07:46,885 INFO [train.py:1198] (2/4) Epoch 16, batch 2700, loss[loss=0.1997, ctc_loss=0.1286, cr_loss=0.3556, over 16966.00 frames. ], tot_loss[loss=0.2242, ctc_loss=0.1512, cr_loss=0.3649, over 3356402.35 frames. ], batch size: 42, lr: 7.48e-03, grad_scale: 32.0 2024-09-23 15:07:48,733 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=285320.0, ans=0.125 2024-09-23 15:08:04,508 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=285366.6666666667, ans=0.07 2024-09-23 15:08:09,207 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=285366.6666666667, ans=0.05 2024-09-23 15:08:26,735 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.106e+02 1.250e+02 1.328e+02 1.437e+02 2.275e+02, threshold=2.655e+02, percent-clipped=0.0 2024-09-23 15:08:46,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=285460.0, ans=0.125 2024-09-23 15:08:46,292 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=285460.0, ans=0.0 2024-09-23 15:08:54,030 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=285506.6666666667, ans=0.2 2024-09-23 15:09:06,362 INFO [train.py:1198] (2/4) Epoch 16, batch 2750, loss[loss=0.188, ctc_loss=0.1236, cr_loss=0.3221, over 17089.00 frames. ], tot_loss[loss=0.225, ctc_loss=0.1519, cr_loss=0.3656, over 3352935.49 frames. ], batch size: 43, lr: 7.48e-03, grad_scale: 32.0 2024-09-23 15:09:10,210 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=9.26 vs. limit=15.0 2024-09-23 15:09:16,343 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.max_positive, batch_count=285553.3333333333, ans=0.95 2024-09-23 15:09:18,055 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=285553.3333333333, ans=0.125 2024-09-23 15:09:30,651 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=285600.0, ans=0.1 2024-09-23 15:09:30,795 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=285600.0, ans=0.2 2024-09-23 15:09:32,563 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.55 vs. limit=15.0 2024-09-23 15:10:05,478 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=285693.3333333333, ans=0.1 2024-09-23 15:10:18,136 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=285740.0, ans=0.125 2024-09-23 15:10:26,019 INFO [train.py:1198] (2/4) Epoch 16, batch 2800, loss[loss=0.2268, ctc_loss=0.151, cr_loss=0.3794, over 17076.00 frames. ], tot_loss[loss=0.2243, ctc_loss=0.1512, cr_loss=0.3652, over 3359798.62 frames. ], batch size: 46, lr: 7.48e-03, grad_scale: 32.0 2024-09-23 15:11:11,238 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.075e+02 1.268e+02 1.378e+02 1.547e+02 2.101e+02, threshold=2.756e+02, percent-clipped=0.0 2024-09-23 15:11:21,396 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.19 vs. limit=6.0 2024-09-23 15:11:36,470 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=285973.3333333333, ans=0.125 2024-09-23 15:11:53,454 INFO [train.py:1198] (2/4) Epoch 16, batch 2850, loss[loss=0.2475, ctc_loss=0.1696, cr_loss=0.3895, over 17136.00 frames. ], tot_loss[loss=0.2249, ctc_loss=0.1519, cr_loss=0.3652, over 3345908.42 frames. ], batch size: 48, lr: 7.48e-03, grad_scale: 32.0 2024-09-23 15:12:14,043 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=286066.6666666667, ans=0.125 2024-09-23 15:12:21,823 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=286066.6666666667, ans=0.1 2024-09-23 15:12:28,328 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=286113.3333333333, ans=0.125 2024-09-23 15:12:33,277 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=286113.3333333333, ans=0.125 2024-09-23 15:13:00,470 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=286206.6666666667, ans=0.2 2024-09-23 15:13:02,533 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.44 vs. limit=10.0 2024-09-23 15:13:13,161 INFO [train.py:1198] (2/4) Epoch 16, batch 2900, loss[loss=0.2369, ctc_loss=0.1634, cr_loss=0.3677, over 17298.00 frames. ], tot_loss[loss=0.2253, ctc_loss=0.1522, cr_loss=0.366, over 3347222.91 frames. ], batch size: 51, lr: 7.47e-03, grad_scale: 32.0 2024-09-23 15:13:35,824 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=286300.0, ans=0.0 2024-09-23 15:13:39,308 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=286300.0, ans=0.0 2024-09-23 15:13:53,318 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.106e+02 1.305e+02 1.434e+02 1.582e+02 2.466e+02, threshold=2.868e+02, percent-clipped=0.0 2024-09-23 15:13:54,377 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=9.62 vs. limit=15.0 2024-09-23 15:14:33,434 INFO [train.py:1198] (2/4) Epoch 16, batch 2950, loss[loss=0.2626, ctc_loss=0.1811, cr_loss=0.4078, over 16662.00 frames. ], tot_loss[loss=0.2253, ctc_loss=0.152, cr_loss=0.3661, over 3358742.02 frames. ], batch size: 61, lr: 7.47e-03, grad_scale: 32.0 2024-09-23 15:14:35,710 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=7.28 vs. limit=15.0 2024-09-23 15:14:41,809 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=286486.6666666667, ans=0.125 2024-09-23 15:15:10,635 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=286580.0, ans=0.125 2024-09-23 15:15:23,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=286626.6666666667, ans=0.1 2024-09-23 15:15:41,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=286673.3333333333, ans=0.0 2024-09-23 15:15:43,264 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=286673.3333333333, ans=0.025 2024-09-23 15:15:51,185 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=286673.3333333333, ans=0.0 2024-09-23 15:15:58,476 INFO [train.py:1198] (2/4) Epoch 16, batch 3000, loss[loss=0.2543, ctc_loss=0.1757, cr_loss=0.3931, over 16528.00 frames. ], tot_loss[loss=0.2256, ctc_loss=0.1524, cr_loss=0.3661, over 3361012.20 frames. ], batch size: 66, lr: 7.47e-03, grad_scale: 32.0 2024-09-23 15:15:58,477 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 15:16:14,044 INFO [train.py:1230] (2/4) Epoch 16, validation: loss=0.04215, ctc_loss=0.04215, cr_loss=7.551e-15, over 944034.00 frames. 2024-09-23 15:16:14,044 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 15:16:19,115 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=286720.0, ans=0.125 2024-09-23 15:16:36,409 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=286766.6666666667, ans=0.125 2024-09-23 15:16:53,330 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.253e+02 1.384e+02 1.509e+02 2.501e+02, threshold=2.769e+02, percent-clipped=0.0 2024-09-23 15:16:53,685 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=286813.3333333333, ans=0.125 2024-09-23 15:17:04,783 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=286860.0, ans=0.0 2024-09-23 15:17:31,987 INFO [train.py:1198] (2/4) Epoch 16, batch 3050, loss[loss=0.2094, ctc_loss=0.1395, cr_loss=0.3494, over 17252.00 frames. ], tot_loss[loss=0.2265, ctc_loss=0.1531, cr_loss=0.3674, over 3364963.01 frames. ], batch size: 44, lr: 7.46e-03, grad_scale: 32.0 2024-09-23 15:18:08,489 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 15:18:50,317 INFO [train.py:1198] (2/4) Epoch 16, batch 3100, loss[loss=0.287, ctc_loss=0.2047, cr_loss=0.4117, over 12378.00 frames. ], tot_loss[loss=0.2265, ctc_loss=0.1531, cr_loss=0.3668, over 3350907.78 frames. ], batch size: 123, lr: 7.46e-03, grad_scale: 32.0 2024-09-23 15:18:54,126 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=4.36 vs. limit=15.0 2024-09-23 15:19:21,774 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=287280.0, ans=0.1 2024-09-23 15:19:24,876 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=287280.0, ans=0.09899494936611666 2024-09-23 15:19:29,118 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.115e+02 1.248e+02 1.355e+02 1.474e+02 3.116e+02, threshold=2.710e+02, percent-clipped=1.0 2024-09-23 15:19:35,774 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 15:19:54,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=287373.3333333333, ans=0.0 2024-09-23 15:19:54,919 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=287373.3333333333, ans=0.025 2024-09-23 15:20:08,479 INFO [train.py:1198] (2/4) Epoch 16, batch 3150, loss[loss=0.1961, ctc_loss=0.1292, cr_loss=0.3347, over 16922.00 frames. ], tot_loss[loss=0.2256, ctc_loss=0.1524, cr_loss=0.3661, over 3352551.80 frames. ], batch size: 42, lr: 7.46e-03, grad_scale: 16.0 2024-09-23 15:20:19,515 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=287420.0, ans=0.125 2024-09-23 15:20:23,079 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.11 vs. limit=15.0 2024-09-23 15:20:26,087 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=287466.6666666667, ans=0.125 2024-09-23 15:20:40,050 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=287513.3333333333, ans=0.1 2024-09-23 15:20:49,414 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=287513.3333333333, ans=0.0 2024-09-23 15:21:09,367 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=287606.6666666667, ans=0.0 2024-09-23 15:21:26,302 INFO [train.py:1198] (2/4) Epoch 16, batch 3200, loss[loss=0.23, ctc_loss=0.1559, cr_loss=0.3707, over 17072.00 frames. ], tot_loss[loss=0.2261, ctc_loss=0.1528, cr_loss=0.3666, over 3345517.18 frames. ], batch size: 46, lr: 7.45e-03, grad_scale: 32.0 2024-09-23 15:21:37,535 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=287653.3333333333, ans=0.025 2024-09-23 15:21:51,516 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=287700.0, ans=0.0 2024-09-23 15:21:59,315 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 15:22:06,510 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.274e+02 1.402e+02 1.535e+02 1.896e+02, threshold=2.804e+02, percent-clipped=0.0 2024-09-23 15:22:17,632 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=287793.3333333333, ans=0.0 2024-09-23 15:22:22,354 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=287793.3333333333, ans=0.125 2024-09-23 15:22:28,949 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=11.95 vs. limit=15.0 2024-09-23 15:22:43,940 INFO [train.py:1198] (2/4) Epoch 16, batch 3250, loss[loss=0.2391, ctc_loss=0.1603, cr_loss=0.394, over 17143.00 frames. ], tot_loss[loss=0.2272, ctc_loss=0.1535, cr_loss=0.3686, over 3350716.79 frames. ], batch size: 48, lr: 7.45e-03, grad_scale: 32.0 2024-09-23 15:23:19,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=287980.0, ans=0.125 2024-09-23 15:23:24,743 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=287980.0, ans=0.07 2024-09-23 15:23:57,844 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=8.97 vs. limit=15.0 2024-09-23 15:24:01,923 INFO [train.py:1198] (2/4) Epoch 16, batch 3300, loss[loss=0.2268, ctc_loss=0.1533, cr_loss=0.3676, over 17216.00 frames. ], tot_loss[loss=0.2268, ctc_loss=0.1533, cr_loss=0.3677, over 3341724.89 frames. ], batch size: 47, lr: 7.45e-03, grad_scale: 32.0 2024-09-23 15:24:05,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=288120.0, ans=0.1 2024-09-23 15:24:11,554 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=288120.0, ans=0.0 2024-09-23 15:24:13,019 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=288120.0, ans=0.125 2024-09-23 15:24:13,022 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=288120.0, ans=0.125 2024-09-23 15:24:46,430 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.267e+02 1.362e+02 1.524e+02 2.882e+02, threshold=2.723e+02, percent-clipped=1.0 2024-09-23 15:24:53,207 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=288260.0, ans=0.0 2024-09-23 15:24:54,708 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=288260.0, ans=0.125 2024-09-23 15:25:24,524 INFO [train.py:1198] (2/4) Epoch 16, batch 3350, loss[loss=0.191, ctc_loss=0.1246, cr_loss=0.3321, over 16639.00 frames. ], tot_loss[loss=0.2265, ctc_loss=0.1531, cr_loss=0.367, over 3339563.13 frames. ], batch size: 37, lr: 7.45e-03, grad_scale: 32.0 2024-09-23 15:25:28,182 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.55 vs. limit=15.0 2024-09-23 15:25:32,621 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=288353.3333333333, ans=0.2 2024-09-23 15:25:32,736 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=288353.3333333333, ans=0.2 2024-09-23 15:25:48,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=288400.0, ans=0.1 2024-09-23 15:25:49,539 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=18.07 vs. limit=22.5 2024-09-23 15:26:00,134 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.min_positive, batch_count=288446.6666666667, ans=0.05 2024-09-23 15:26:02,198 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=8.88 vs. limit=15.0 2024-09-23 15:26:14,237 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=288493.3333333333, ans=0.07 2024-09-23 15:26:19,472 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=19.60 vs. limit=22.5 2024-09-23 15:26:20,552 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=288493.3333333333, ans=0.125 2024-09-23 15:26:40,626 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=288540.0, ans=0.125 2024-09-23 15:26:45,093 INFO [train.py:1198] (2/4) Epoch 16, batch 3400, loss[loss=0.2081, ctc_loss=0.1396, cr_loss=0.3428, over 17196.00 frames. ], tot_loss[loss=0.2255, ctc_loss=0.1522, cr_loss=0.3666, over 3350802.83 frames. ], batch size: 47, lr: 7.44e-03, grad_scale: 32.0 2024-09-23 15:26:48,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=288586.6666666667, ans=0.0 2024-09-23 15:26:57,889 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=288586.6666666667, ans=0.5 2024-09-23 15:27:01,183 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=288633.3333333333, ans=0.1 2024-09-23 15:27:07,508 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=288633.3333333333, ans=0.1 2024-09-23 15:27:24,128 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.69 vs. limit=15.0 2024-09-23 15:27:27,928 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.080e+02 1.234e+02 1.362e+02 1.557e+02 3.833e+02, threshold=2.724e+02, percent-clipped=1.0 2024-09-23 15:27:44,036 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=288726.6666666667, ans=0.0 2024-09-23 15:28:05,748 INFO [train.py:1198] (2/4) Epoch 16, batch 3450, loss[loss=0.2108, ctc_loss=0.1414, cr_loss=0.3473, over 17061.00 frames. ], tot_loss[loss=0.2267, ctc_loss=0.153, cr_loss=0.3681, over 3355120.28 frames. ], batch size: 39, lr: 7.44e-03, grad_scale: 32.0 2024-09-23 15:28:07,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=288820.0, ans=0.125 2024-09-23 15:28:16,985 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=288820.0, ans=0.1 2024-09-23 15:28:54,339 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=288960.0, ans=0.2 2024-09-23 15:28:56,255 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=5.45 vs. limit=12.0 2024-09-23 15:29:06,725 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=289006.6666666667, ans=0.125 2024-09-23 15:29:18,027 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=5.91 vs. limit=12.0 2024-09-23 15:29:23,286 INFO [train.py:1198] (2/4) Epoch 16, batch 3500, loss[loss=0.2152, ctc_loss=0.1444, cr_loss=0.354, over 17099.00 frames. ], tot_loss[loss=0.2264, ctc_loss=0.1528, cr_loss=0.368, over 3361635.27 frames. ], batch size: 43, lr: 7.44e-03, grad_scale: 32.0 2024-09-23 15:29:23,927 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.37 vs. limit=15.0 2024-09-23 15:29:39,013 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=289100.0, ans=0.0 2024-09-23 15:30:03,920 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.282e+02 1.368e+02 1.525e+02 3.473e+02, threshold=2.736e+02, percent-clipped=1.0 2024-09-23 15:30:09,040 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 15:30:29,339 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=289240.0, ans=0.0 2024-09-23 15:30:33,884 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=289240.0, ans=0.0 2024-09-23 15:30:41,433 INFO [train.py:1198] (2/4) Epoch 16, batch 3550, loss[loss=0.2183, ctc_loss=0.1439, cr_loss=0.3724, over 17032.00 frames. ], tot_loss[loss=0.2252, ctc_loss=0.1518, cr_loss=0.3671, over 3368366.93 frames. ], batch size: 44, lr: 7.43e-03, grad_scale: 32.0 2024-09-23 15:30:44,762 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=289286.6666666667, ans=0.0 2024-09-23 15:30:44,882 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=289286.6666666667, ans=0.125 2024-09-23 15:31:05,644 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=289333.3333333333, ans=0.025 2024-09-23 15:31:05,754 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=289333.3333333333, ans=0.125 2024-09-23 15:31:16,537 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=289380.0, ans=0.125 2024-09-23 15:31:34,136 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.51 vs. limit=10.0 2024-09-23 15:32:00,000 INFO [train.py:1198] (2/4) Epoch 16, batch 3600, loss[loss=0.1986, ctc_loss=0.133, cr_loss=0.3282, over 16296.00 frames. ], tot_loss[loss=0.2247, ctc_loss=0.1515, cr_loss=0.3663, over 3369331.16 frames. ], batch size: 36, lr: 7.43e-03, grad_scale: 32.0 2024-09-23 15:32:12,369 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=289520.0, ans=0.0 2024-09-23 15:32:18,654 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=289566.6666666667, ans=0.1 2024-09-23 15:32:36,022 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=289613.3333333333, ans=0.1 2024-09-23 15:32:40,498 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.033e+02 1.232e+02 1.305e+02 1.381e+02 1.906e+02, threshold=2.610e+02, percent-clipped=0.0 2024-09-23 15:32:44,980 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=8.07 vs. limit=15.0 2024-09-23 15:32:58,170 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 15:33:18,230 INFO [train.py:1198] (2/4) Epoch 16, batch 3650, loss[loss=0.2605, ctc_loss=0.174, cr_loss=0.4328, over 17051.00 frames. ], tot_loss[loss=0.2244, ctc_loss=0.1513, cr_loss=0.3654, over 3357239.66 frames. ], batch size: 52, lr: 7.43e-03, grad_scale: 16.0 2024-09-23 15:33:28,397 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.38 vs. limit=15.0 2024-09-23 15:33:32,897 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=14.21 vs. limit=22.5 2024-09-23 15:33:34,104 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=289800.0, ans=0.025 2024-09-23 15:34:02,615 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=289846.6666666667, ans=0.125 2024-09-23 15:34:06,215 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.33 vs. limit=6.0 2024-09-23 15:34:13,332 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=289893.3333333333, ans=0.1 2024-09-23 15:34:18,818 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=6.57 vs. limit=15.0 2024-09-23 15:34:40,766 INFO [train.py:1198] (2/4) Epoch 16, batch 3700, loss[loss=0.2204, ctc_loss=0.1486, cr_loss=0.359, over 17371.00 frames. ], tot_loss[loss=0.224, ctc_loss=0.151, cr_loss=0.3651, over 3360419.45 frames. ], batch size: 48, lr: 7.43e-03, grad_scale: 16.0 2024-09-23 15:35:11,746 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=290080.0, ans=0.0 2024-09-23 15:35:14,668 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=290080.0, ans=0.125 2024-09-23 15:35:23,810 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.296e+02 1.389e+02 1.546e+02 2.430e+02, threshold=2.778e+02, percent-clipped=0.0 2024-09-23 15:35:59,677 INFO [train.py:1198] (2/4) Epoch 16, batch 3750, loss[loss=0.2415, ctc_loss=0.1642, cr_loss=0.3868, over 16930.00 frames. ], tot_loss[loss=0.2258, ctc_loss=0.1523, cr_loss=0.3676, over 3359690.68 frames. ], batch size: 58, lr: 7.42e-03, grad_scale: 16.0 2024-09-23 15:36:07,846 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=290220.0, ans=0.125 2024-09-23 15:36:16,002 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=290266.6666666667, ans=0.0 2024-09-23 15:36:23,603 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=290266.6666666667, ans=0.0 2024-09-23 15:36:45,536 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=290360.0, ans=0.0 2024-09-23 15:36:46,951 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=290360.0, ans=0.125 2024-09-23 15:37:02,047 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.min_positive, batch_count=290406.6666666667, ans=0.05 2024-09-23 15:37:18,710 INFO [train.py:1198] (2/4) Epoch 16, batch 3800, loss[loss=0.2082, ctc_loss=0.1397, cr_loss=0.3426, over 17152.00 frames. ], tot_loss[loss=0.2267, ctc_loss=0.1532, cr_loss=0.3677, over 3326950.45 frames. ], batch size: 45, lr: 7.42e-03, grad_scale: 16.0 2024-09-23 15:37:26,799 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=290453.3333333333, ans=0.125 2024-09-23 15:37:45,244 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=290500.0, ans=0.2 2024-09-23 15:37:52,873 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=290546.6666666667, ans=0.125 2024-09-23 15:38:00,416 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.074e+02 1.261e+02 1.370e+02 1.506e+02 3.479e+02, threshold=2.739e+02, percent-clipped=1.0 2024-09-23 15:38:00,687 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=290546.6666666667, ans=0.07 2024-09-23 15:38:21,079 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=290640.0, ans=0.0 2024-09-23 15:38:29,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=290640.0, ans=0.0 2024-09-23 15:38:32,913 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=8.94 vs. limit=15.0 2024-09-23 15:38:36,619 INFO [train.py:1198] (2/4) Epoch 16, batch 3850, loss[loss=0.2504, ctc_loss=0.1694, cr_loss=0.405, over 17311.00 frames. ], tot_loss[loss=0.2292, ctc_loss=0.1555, cr_loss=0.3688, over 3266640.86 frames. ], batch size: 51, lr: 7.42e-03, grad_scale: 16.0 2024-09-23 15:38:44,910 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=290686.6666666667, ans=0.0 2024-09-23 15:39:01,669 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 15:39:06,221 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=290780.0, ans=0.2 2024-09-23 15:39:11,129 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.83 vs. limit=15.0 2024-09-23 15:39:18,976 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.86 vs. limit=6.0 2024-09-23 15:39:35,355 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=290826.6666666667, ans=0.125 2024-09-23 15:39:40,210 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=290873.3333333333, ans=0.125 2024-09-23 15:40:38,385 INFO [train.py:1198] (2/4) Epoch 17, batch 0, loss[loss=0.2491, ctc_loss=0.1684, cr_loss=0.4031, over 17149.00 frames. ], tot_loss[loss=0.2491, ctc_loss=0.1684, cr_loss=0.4031, over 17149.00 frames. ], batch size: 48, lr: 7.19e-03, grad_scale: 32.0 2024-09-23 15:40:38,386 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 15:40:45,997 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.4.encoder.layers.2.self_attn_weights, attn_weights_entropy = tensor([4.2099, 3.4019, 3.8471, 3.7204], device='cuda:2') 2024-09-23 15:40:53,762 INFO [train.py:1230] (2/4) Epoch 17, validation: loss=0.04104, ctc_loss=0.04104, cr_loss=7.589e-15, over 944034.00 frames. 2024-09-23 15:40:53,763 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 15:41:10,281 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=290948.0, ans=0.0 2024-09-23 15:41:21,068 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=290948.0, ans=0.125 2024-09-23 15:41:22,683 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.max_abs, batch_count=290948.0, ans=10.0 2024-09-23 15:41:24,272 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=290948.0, ans=0.2 2024-09-23 15:41:46,227 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.183e+02 1.415e+02 1.552e+02 1.651e+02 2.695e+02, threshold=3.103e+02, percent-clipped=0.0 2024-09-23 15:42:11,205 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=10.20 vs. limit=15.0 2024-09-23 15:42:18,130 INFO [train.py:1198] (2/4) Epoch 17, batch 50, loss[loss=0.2443, ctc_loss=0.1656, cr_loss=0.3936, over 17188.00 frames. ], tot_loss[loss=0.2306, ctc_loss=0.156, cr_loss=0.3727, over 748286.53 frames. ], batch size: 55, lr: 7.19e-03, grad_scale: 16.0 2024-09-23 15:42:32,681 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=291181.3333333333, ans=0.1 2024-09-23 15:42:38,856 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=291181.3333333333, ans=0.125 2024-09-23 15:42:39,704 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=5.51 vs. limit=15.0 2024-09-23 15:42:53,466 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=291228.0, ans=0.0 2024-09-23 15:43:03,163 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=291228.0, ans=0.125 2024-09-23 15:43:26,956 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=291321.3333333333, ans=0.125 2024-09-23 15:43:30,422 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=12.86 vs. limit=15.0 2024-09-23 15:43:39,789 INFO [train.py:1198] (2/4) Epoch 17, batch 100, loss[loss=0.2063, ctc_loss=0.1367, cr_loss=0.3479, over 17037.00 frames. ], tot_loss[loss=0.2269, ctc_loss=0.1529, cr_loss=0.3696, over 1328066.47 frames. ], batch size: 56, lr: 7.18e-03, grad_scale: 16.0 2024-09-23 15:44:21,891 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 15:44:31,114 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.243e+02 1.327e+02 1.449e+02 2.389e+02, threshold=2.654e+02, percent-clipped=0.0 2024-09-23 15:44:44,231 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 15:44:55,502 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=291554.6666666667, ans=0.0 2024-09-23 15:44:59,988 INFO [train.py:1198] (2/4) Epoch 17, batch 150, loss[loss=0.2247, ctc_loss=0.1469, cr_loss=0.3889, over 17093.00 frames. ], tot_loss[loss=0.2269, ctc_loss=0.1532, cr_loss=0.3686, over 1764839.51 frames. ], batch size: 49, lr: 7.18e-03, grad_scale: 16.0 2024-09-23 15:45:09,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=291601.3333333333, ans=0.0 2024-09-23 15:45:14,852 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=291648.0, ans=0.125 2024-09-23 15:45:22,687 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=291648.0, ans=0.0 2024-09-23 15:45:25,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=291648.0, ans=0.0 2024-09-23 15:45:27,437 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=291648.0, ans=0.125 2024-09-23 15:45:30,650 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=291694.6666666667, ans=0.125 2024-09-23 15:45:41,762 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=291694.6666666667, ans=0.125 2024-09-23 15:45:43,175 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=291694.6666666667, ans=0.0 2024-09-23 15:45:46,471 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=291694.6666666667, ans=0.025 2024-09-23 15:45:54,414 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=291741.3333333333, ans=0.125 2024-09-23 15:45:54,833 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.01 vs. limit=15.0 2024-09-23 15:46:16,486 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=291788.0, ans=0.025 2024-09-23 15:46:25,721 INFO [train.py:1198] (2/4) Epoch 17, batch 200, loss[loss=0.2404, ctc_loss=0.1654, cr_loss=0.3751, over 16712.00 frames. ], tot_loss[loss=0.2267, ctc_loss=0.153, cr_loss=0.3684, over 2119413.87 frames. ], batch size: 61, lr: 7.18e-03, grad_scale: 16.0 2024-09-23 15:47:12,185 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer_ff2.min_abs, batch_count=291928.0, ans=0.1 2024-09-23 15:47:12,752 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.77 vs. limit=6.0 2024-09-23 15:47:18,310 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.110e+02 1.306e+02 1.386e+02 1.614e+02 2.877e+02, threshold=2.773e+02, percent-clipped=1.0 2024-09-23 15:47:31,442 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=292021.3333333333, ans=0.125 2024-09-23 15:47:36,434 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 15:47:49,025 INFO [train.py:1198] (2/4) Epoch 17, batch 250, loss[loss=0.2293, ctc_loss=0.1541, cr_loss=0.3756, over 17349.00 frames. ], tot_loss[loss=0.2266, ctc_loss=0.1529, cr_loss=0.3685, over 2391721.78 frames. ], batch size: 48, lr: 7.18e-03, grad_scale: 16.0 2024-09-23 15:48:41,093 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.58 vs. limit=6.0 2024-09-23 15:49:05,871 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=292254.6666666667, ans=0.0 2024-09-23 15:49:07,880 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=4.41 vs. limit=12.0 2024-09-23 15:49:08,854 INFO [train.py:1198] (2/4) Epoch 17, batch 300, loss[loss=0.2945, ctc_loss=0.2032, cr_loss=0.4566, over 16505.00 frames. ], tot_loss[loss=0.2271, ctc_loss=0.1532, cr_loss=0.3693, over 2613893.53 frames. ], batch size: 66, lr: 7.17e-03, grad_scale: 16.0 2024-09-23 15:50:00,584 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.112e+02 1.265e+02 1.346e+02 1.452e+02 2.269e+02, threshold=2.693e+02, percent-clipped=0.0 2024-09-23 15:50:04,349 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=292441.3333333333, ans=0.1 2024-09-23 15:50:17,127 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer_na.min_abs, batch_count=292488.0, ans=0.02 2024-09-23 15:50:32,872 INFO [train.py:1198] (2/4) Epoch 17, batch 350, loss[loss=0.2387, ctc_loss=0.1597, cr_loss=0.3949, over 17005.00 frames. ], tot_loss[loss=0.2266, ctc_loss=0.1527, cr_loss=0.3691, over 2777422.33 frames. ], batch size: 51, lr: 7.17e-03, grad_scale: 16.0 2024-09-23 15:51:28,580 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=292674.6666666667, ans=0.125 2024-09-23 15:51:42,694 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=292721.3333333333, ans=0.0 2024-09-23 15:51:57,137 INFO [train.py:1198] (2/4) Epoch 17, batch 400, loss[loss=0.2105, ctc_loss=0.1402, cr_loss=0.3515, over 17144.00 frames. ], tot_loss[loss=0.2259, ctc_loss=0.1524, cr_loss=0.3675, over 2903046.74 frames. ], batch size: 45, lr: 7.17e-03, grad_scale: 32.0 2024-09-23 15:52:21,104 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=292814.6666666667, ans=0.0 2024-09-23 15:52:21,433 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=14.15 vs. limit=22.5 2024-09-23 15:52:24,243 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=292814.6666666667, ans=0.125 2024-09-23 15:52:25,909 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=292814.6666666667, ans=0.025 2024-09-23 15:52:50,103 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.274e+02 1.356e+02 1.494e+02 2.535e+02, threshold=2.713e+02, percent-clipped=0.0 2024-09-23 15:53:04,826 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=292954.6666666667, ans=0.07 2024-09-23 15:53:18,847 INFO [train.py:1198] (2/4) Epoch 17, batch 450, loss[loss=0.2253, ctc_loss=0.1548, cr_loss=0.3525, over 17165.00 frames. ], tot_loss[loss=0.2262, ctc_loss=0.1526, cr_loss=0.3678, over 3005019.71 frames. ], batch size: 45, lr: 7.16e-03, grad_scale: 32.0 2024-09-23 15:53:53,240 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=293094.6666666667, ans=0.0 2024-09-23 15:53:57,843 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=293094.6666666667, ans=0.125 2024-09-23 15:54:17,363 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=293141.3333333333, ans=0.125 2024-09-23 15:54:31,472 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=293188.0, ans=0.125 2024-09-23 15:54:39,067 INFO [train.py:1198] (2/4) Epoch 17, batch 500, loss[loss=0.2181, ctc_loss=0.1444, cr_loss=0.3686, over 17307.00 frames. ], tot_loss[loss=0.2247, ctc_loss=0.1514, cr_loss=0.3662, over 3090905.37 frames. ], batch size: 49, lr: 7.16e-03, grad_scale: 32.0 2024-09-23 15:54:48,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=293234.6666666667, ans=0.1 2024-09-23 15:55:06,745 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=293281.3333333333, ans=0.0 2024-09-23 15:55:32,746 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.054e+02 1.258e+02 1.318e+02 1.426e+02 2.177e+02, threshold=2.636e+02, percent-clipped=0.0 2024-09-23 15:55:42,563 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=293374.6666666667, ans=0.025 2024-09-23 15:55:47,371 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=293421.3333333333, ans=0.125 2024-09-23 15:55:55,464 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=293421.3333333333, ans=0.0 2024-09-23 15:56:03,960 INFO [train.py:1198] (2/4) Epoch 17, batch 550, loss[loss=0.2246, ctc_loss=0.1528, cr_loss=0.3592, over 17042.00 frames. ], tot_loss[loss=0.2251, ctc_loss=0.1516, cr_loss=0.3673, over 3153150.66 frames. ], batch size: 53, lr: 7.16e-03, grad_scale: 32.0 2024-09-23 15:56:14,181 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=5.59 vs. limit=12.0 2024-09-23 15:56:38,437 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=3.77 vs. limit=15.0 2024-09-23 15:56:45,804 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 15:57:07,023 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=11.87 vs. limit=15.0 2024-09-23 15:57:27,129 INFO [train.py:1198] (2/4) Epoch 17, batch 600, loss[loss=0.1983, ctc_loss=0.1347, cr_loss=0.318, over 17040.00 frames. ], tot_loss[loss=0.2254, ctc_loss=0.1518, cr_loss=0.3676, over 3191084.71 frames. ], batch size: 39, lr: 7.16e-03, grad_scale: 32.0 2024-09-23 15:57:49,213 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=293748.0, ans=0.0 2024-09-23 15:58:20,482 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.130e+02 1.285e+02 1.384e+02 1.494e+02 2.358e+02, threshold=2.768e+02, percent-clipped=0.0 2024-09-23 15:58:24,272 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 15:58:28,274 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module1.whiten, num_groups=1, num_channels=192, metric=7.39 vs. limit=15.0 2024-09-23 15:58:30,594 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=293841.3333333333, ans=0.125 2024-09-23 15:58:49,412 INFO [train.py:1198] (2/4) Epoch 17, batch 650, loss[loss=0.1842, ctc_loss=0.1249, cr_loss=0.2965, over 17027.00 frames. ], tot_loss[loss=0.2242, ctc_loss=0.151, cr_loss=0.3658, over 3224725.04 frames. ], batch size: 39, lr: 7.15e-03, grad_scale: 32.0 2024-09-23 15:59:16,842 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=293981.3333333333, ans=0.125 2024-09-23 15:59:49,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=294074.6666666667, ans=0.0 2024-09-23 15:59:54,285 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=294121.3333333333, ans=0.1 2024-09-23 16:00:09,840 INFO [train.py:1198] (2/4) Epoch 17, batch 700, loss[loss=0.204, ctc_loss=0.1351, cr_loss=0.3445, over 17029.00 frames. ], tot_loss[loss=0.2235, ctc_loss=0.1504, cr_loss=0.3658, over 3252428.57 frames. ], batch size: 39, lr: 7.15e-03, grad_scale: 32.0 2024-09-23 16:00:39,815 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=294214.6666666667, ans=0.0 2024-09-23 16:00:45,656 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=10.95 vs. limit=22.5 2024-09-23 16:00:52,720 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=294261.3333333333, ans=0.125 2024-09-23 16:01:06,301 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.243e+02 1.328e+02 1.463e+02 2.107e+02, threshold=2.656e+02, percent-clipped=0.0 2024-09-23 16:01:22,445 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=294354.6666666667, ans=0.125 2024-09-23 16:01:28,768 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=294354.6666666667, ans=0.035 2024-09-23 16:01:28,940 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=294354.6666666667, ans=0.1 2024-09-23 16:01:34,890 INFO [train.py:1198] (2/4) Epoch 17, batch 750, loss[loss=0.1855, ctc_loss=0.1229, cr_loss=0.3131, over 17107.00 frames. ], tot_loss[loss=0.2236, ctc_loss=0.1505, cr_loss=0.3653, over 3281407.67 frames. ], batch size: 40, lr: 7.15e-03, grad_scale: 32.0 2024-09-23 16:01:35,550 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.22 vs. limit=15.0 2024-09-23 16:01:36,883 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=294401.3333333333, ans=0.125 2024-09-23 16:01:41,451 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=294401.3333333333, ans=0.125 2024-09-23 16:02:00,401 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=294448.0, ans=0.2 2024-09-23 16:02:15,009 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=294494.6666666667, ans=0.125 2024-09-23 16:02:16,411 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=294494.6666666667, ans=0.0 2024-09-23 16:02:22,778 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=294494.6666666667, ans=0.0 2024-09-23 16:02:32,464 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=294541.3333333333, ans=0.1 2024-09-23 16:02:35,443 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=294541.3333333333, ans=0.2 2024-09-23 16:02:37,798 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=4.83 vs. limit=10.0 2024-09-23 16:02:54,357 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=294588.0, ans=0.1 2024-09-23 16:02:57,399 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=294588.0, ans=0.125 2024-09-23 16:03:00,185 INFO [train.py:1198] (2/4) Epoch 17, batch 800, loss[loss=0.2161, ctc_loss=0.1461, cr_loss=0.3501, over 17347.00 frames. ], tot_loss[loss=0.225, ctc_loss=0.1516, cr_loss=0.3667, over 3296053.10 frames. ], batch size: 48, lr: 7.14e-03, grad_scale: 32.0 2024-09-23 16:03:24,892 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=512, metric=14.24 vs. limit=22.5 2024-09-23 16:03:46,859 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=294774.6666666667, ans=0.125 2024-09-23 16:03:51,161 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.056e+02 1.255e+02 1.363e+02 1.453e+02 2.011e+02, threshold=2.725e+02, percent-clipped=0.0 2024-09-23 16:04:00,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.max_abs, batch_count=294774.6666666667, ans=10.0 2024-09-23 16:04:08,925 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=294821.3333333333, ans=0.2 2024-09-23 16:04:15,281 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=294821.3333333333, ans=0.0 2024-09-23 16:04:19,627 INFO [train.py:1198] (2/4) Epoch 17, batch 850, loss[loss=0.2369, ctc_loss=0.1619, cr_loss=0.3751, over 16894.00 frames. ], tot_loss[loss=0.2243, ctc_loss=0.1512, cr_loss=0.366, over 3312144.60 frames. ], batch size: 58, lr: 7.14e-03, grad_scale: 32.0 2024-09-23 16:04:40,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=294914.6666666667, ans=0.0 2024-09-23 16:05:00,482 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=5.08 vs. limit=12.0 2024-09-23 16:05:08,514 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=6.22 vs. limit=15.0 2024-09-23 16:05:20,083 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=295008.0, ans=0.125 2024-09-23 16:05:41,832 INFO [train.py:1198] (2/4) Epoch 17, batch 900, loss[loss=0.2223, ctc_loss=0.1487, cr_loss=0.3678, over 17292.00 frames. ], tot_loss[loss=0.2251, ctc_loss=0.1518, cr_loss=0.3666, over 3325435.75 frames. ], batch size: 46, lr: 7.14e-03, grad_scale: 32.0 2024-09-23 16:05:43,710 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=295101.3333333333, ans=0.2 2024-09-23 16:06:01,103 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.77 vs. limit=15.0 2024-09-23 16:06:02,312 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=295148.0, ans=0.0 2024-09-23 16:06:31,025 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=295241.3333333333, ans=0.0 2024-09-23 16:06:34,304 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=295241.3333333333, ans=0.2 2024-09-23 16:06:34,377 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=295241.3333333333, ans=0.0 2024-09-23 16:06:35,667 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.082e+02 1.252e+02 1.345e+02 1.506e+02 2.061e+02, threshold=2.689e+02, percent-clipped=0.0 2024-09-23 16:06:38,952 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=295241.3333333333, ans=0.0 2024-09-23 16:06:53,193 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 16:07:07,053 INFO [train.py:1198] (2/4) Epoch 17, batch 950, loss[loss=0.2111, ctc_loss=0.1372, cr_loss=0.3694, over 16985.00 frames. ], tot_loss[loss=0.225, ctc_loss=0.1517, cr_loss=0.3663, over 3333733.90 frames. ], batch size: 39, lr: 7.14e-03, grad_scale: 32.0 2024-09-23 16:07:08,401 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.97 vs. limit=6.0 2024-09-23 16:07:18,328 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=295334.6666666667, ans=0.125 2024-09-23 16:07:18,382 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=295334.6666666667, ans=0.125 2024-09-23 16:07:57,763 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 16:08:27,868 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=295568.0, ans=0.125 2024-09-23 16:08:29,164 INFO [train.py:1198] (2/4) Epoch 17, batch 1000, loss[loss=0.2588, ctc_loss=0.1731, cr_loss=0.4286, over 17240.00 frames. ], tot_loss[loss=0.2249, ctc_loss=0.1516, cr_loss=0.3667, over 3343327.85 frames. ], batch size: 55, lr: 7.13e-03, grad_scale: 32.0 2024-09-23 16:08:37,243 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=295568.0, ans=0.125 2024-09-23 16:08:40,455 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 16:08:45,280 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=6.19 vs. limit=15.0 2024-09-23 16:08:58,236 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.41 vs. limit=15.0 2024-09-23 16:09:01,049 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=295661.3333333333, ans=0.0 2024-09-23 16:09:19,593 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.278e+02 1.364e+02 1.541e+02 2.196e+02, threshold=2.728e+02, percent-clipped=0.0 2024-09-23 16:09:47,780 INFO [train.py:1198] (2/4) Epoch 17, batch 1050, loss[loss=0.2147, ctc_loss=0.1414, cr_loss=0.3662, over 17119.00 frames. ], tot_loss[loss=0.2247, ctc_loss=0.1515, cr_loss=0.3657, over 3331496.27 frames. ], batch size: 40, lr: 7.13e-03, grad_scale: 32.0 2024-09-23 16:10:20,758 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=295894.6666666667, ans=0.125 2024-09-23 16:10:33,850 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=295894.6666666667, ans=0.0 2024-09-23 16:10:35,481 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=295894.6666666667, ans=0.125 2024-09-23 16:11:11,620 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 16:11:12,826 INFO [train.py:1198] (2/4) Epoch 17, batch 1100, loss[loss=0.1789, ctc_loss=0.1201, cr_loss=0.294, over 17123.00 frames. ], tot_loss[loss=0.2235, ctc_loss=0.1505, cr_loss=0.3646, over 3336201.94 frames. ], batch size: 40, lr: 7.13e-03, grad_scale: 32.0 2024-09-23 16:11:23,006 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=13.31 vs. limit=15.0 2024-09-23 16:11:38,182 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 16:11:39,615 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=296081.3333333333, ans=0.0 2024-09-23 16:12:00,532 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 16:12:03,605 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=296174.6666666667, ans=0.125 2024-09-23 16:12:06,482 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.092e+02 1.273e+02 1.407e+02 1.625e+02 2.289e+02, threshold=2.813e+02, percent-clipped=0.0 2024-09-23 16:12:19,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=296221.3333333333, ans=0.125 2024-09-23 16:12:35,106 INFO [train.py:1198] (2/4) Epoch 17, batch 1150, loss[loss=0.2239, ctc_loss=0.1488, cr_loss=0.3759, over 17184.00 frames. ], tot_loss[loss=0.224, ctc_loss=0.1509, cr_loss=0.3654, over 3339530.41 frames. ], batch size: 45, lr: 7.13e-03, grad_scale: 32.0 2024-09-23 16:12:44,567 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=296268.0, ans=0.0 2024-09-23 16:12:54,267 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=296314.6666666667, ans=0.125 2024-09-23 16:13:28,208 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=23.38 vs. limit=22.5 2024-09-23 16:13:35,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=296408.0, ans=0.0 2024-09-23 16:13:45,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.max_abs, batch_count=296454.6666666667, ans=10.0 2024-09-23 16:13:57,590 INFO [train.py:1198] (2/4) Epoch 17, batch 1200, loss[loss=0.2024, ctc_loss=0.1319, cr_loss=0.3525, over 17165.00 frames. ], tot_loss[loss=0.2233, ctc_loss=0.1504, cr_loss=0.3646, over 3352445.98 frames. ], batch size: 45, lr: 7.12e-03, grad_scale: 32.0 2024-09-23 16:14:32,578 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=296594.6666666667, ans=0.125 2024-09-23 16:14:49,695 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.095e+02 1.250e+02 1.332e+02 1.452e+02 2.634e+02, threshold=2.664e+02, percent-clipped=0.0 2024-09-23 16:15:19,279 INFO [train.py:1198] (2/4) Epoch 17, batch 1250, loss[loss=0.2374, ctc_loss=0.1597, cr_loss=0.3885, over 17128.00 frames. ], tot_loss[loss=0.2221, ctc_loss=0.1494, cr_loss=0.3635, over 3359816.04 frames. ], batch size: 48, lr: 7.12e-03, grad_scale: 32.0 2024-09-23 16:15:32,332 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=296734.6666666667, ans=0.125 2024-09-23 16:15:35,537 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=296781.3333333333, ans=0.025 2024-09-23 16:15:47,727 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=296781.3333333333, ans=0.125 2024-09-23 16:16:27,989 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=296921.3333333333, ans=0.125 2024-09-23 16:16:44,492 INFO [train.py:1198] (2/4) Epoch 17, batch 1300, loss[loss=0.2097, ctc_loss=0.1414, cr_loss=0.3412, over 17272.00 frames. ], tot_loss[loss=0.2221, ctc_loss=0.1492, cr_loss=0.3644, over 3371804.73 frames. ], batch size: 44, lr: 7.12e-03, grad_scale: 32.0 2024-09-23 16:16:44,876 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=296968.0, ans=0.1 2024-09-23 16:16:57,433 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=296968.0, ans=0.09899494936611666 2024-09-23 16:17:00,670 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer_ff3.min_abs, batch_count=297014.6666666667, ans=0.2 2024-09-23 16:17:16,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=297061.3333333333, ans=0.125 2024-09-23 16:17:19,757 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=297061.3333333333, ans=0.125 2024-09-23 16:17:36,178 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.32 vs. limit=15.0 2024-09-23 16:17:36,627 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.91 vs. limit=5.0 2024-09-23 16:17:36,805 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.242e+02 1.322e+02 1.445e+02 3.373e+02, threshold=2.644e+02, percent-clipped=1.0 2024-09-23 16:17:54,243 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=297154.6666666667, ans=0.125 2024-09-23 16:17:57,456 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=297154.6666666667, ans=0.07 2024-09-23 16:18:06,583 INFO [train.py:1198] (2/4) Epoch 17, batch 1350, loss[loss=0.216, ctc_loss=0.1442, cr_loss=0.3591, over 17293.00 frames. ], tot_loss[loss=0.222, ctc_loss=0.149, cr_loss=0.3646, over 3368116.12 frames. ], batch size: 46, lr: 7.11e-03, grad_scale: 32.0 2024-09-23 16:18:19,858 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=297201.3333333333, ans=0.1 2024-09-23 16:18:40,189 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=297294.6666666667, ans=0.0 2024-09-23 16:18:45,126 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=297294.6666666667, ans=0.125 2024-09-23 16:18:45,541 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.52 vs. limit=6.0 2024-09-23 16:18:49,799 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=297294.6666666667, ans=0.0 2024-09-23 16:19:26,084 INFO [train.py:1198] (2/4) Epoch 17, batch 1400, loss[loss=0.2044, ctc_loss=0.1349, cr_loss=0.3472, over 17141.00 frames. ], tot_loss[loss=0.2221, ctc_loss=0.1492, cr_loss=0.3648, over 3379507.31 frames. ], batch size: 48, lr: 7.11e-03, grad_scale: 32.0 2024-09-23 16:19:33,389 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.72 vs. limit=6.0 2024-09-23 16:20:07,644 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=297528.0, ans=0.015 2024-09-23 16:20:20,766 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.246e+02 1.345e+02 1.562e+02 2.130e+02, threshold=2.691e+02, percent-clipped=0.0 2024-09-23 16:20:21,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=297574.6666666667, ans=0.0 2024-09-23 16:20:21,582 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=17.92 vs. limit=22.5 2024-09-23 16:20:29,201 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=297574.6666666667, ans=0.2 2024-09-23 16:20:50,354 INFO [train.py:1198] (2/4) Epoch 17, batch 1450, loss[loss=0.2264, ctc_loss=0.1518, cr_loss=0.3727, over 17212.00 frames. ], tot_loss[loss=0.2221, ctc_loss=0.1492, cr_loss=0.3646, over 3373781.36 frames. ], batch size: 47, lr: 7.11e-03, grad_scale: 32.0 2024-09-23 16:21:00,608 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.23 vs. limit=12.0 2024-09-23 16:21:16,458 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=297714.6666666667, ans=0.0 2024-09-23 16:21:39,582 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=297808.0, ans=0.2 2024-09-23 16:21:39,583 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=297808.0, ans=0.0 2024-09-23 16:21:47,656 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=297808.0, ans=0.125 2024-09-23 16:22:12,530 INFO [train.py:1198] (2/4) Epoch 17, batch 1500, loss[loss=0.2555, ctc_loss=0.1728, cr_loss=0.4137, over 15963.00 frames. ], tot_loss[loss=0.222, ctc_loss=0.1491, cr_loss=0.3648, over 3368226.73 frames. ], batch size: 74, lr: 7.11e-03, grad_scale: 32.0 2024-09-23 16:22:20,763 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=297901.3333333333, ans=0.125 2024-09-23 16:22:27,027 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=297948.0, ans=0.0 2024-09-23 16:22:50,436 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=297994.6666666667, ans=0.0 2024-09-23 16:22:59,703 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=297994.6666666667, ans=0.0 2024-09-23 16:23:04,501 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=298041.3333333333, ans=0.2 2024-09-23 16:23:07,241 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.004e+02 1.245e+02 1.341e+02 1.437e+02 3.249e+02, threshold=2.682e+02, percent-clipped=1.0 2024-09-23 16:23:34,455 INFO [train.py:1198] (2/4) Epoch 17, batch 1550, loss[loss=0.2096, ctc_loss=0.1374, cr_loss=0.3608, over 17063.00 frames. ], tot_loss[loss=0.2224, ctc_loss=0.1493, cr_loss=0.365, over 3370335.15 frames. ], batch size: 46, lr: 7.10e-03, grad_scale: 32.0 2024-09-23 16:23:52,350 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=298181.3333333333, ans=0.07 2024-09-23 16:24:34,119 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 16:24:54,217 INFO [train.py:1198] (2/4) Epoch 17, batch 1600, loss[loss=0.2637, ctc_loss=0.1876, cr_loss=0.3809, over 11684.00 frames. ], tot_loss[loss=0.2225, ctc_loss=0.1495, cr_loss=0.3648, over 3371181.26 frames. ], batch size: 123, lr: 7.10e-03, grad_scale: 32.0 2024-09-23 16:25:48,226 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=7.09 vs. limit=8.0 2024-09-23 16:25:51,724 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.003e+02 1.268e+02 1.375e+02 1.538e+02 2.240e+02, threshold=2.750e+02, percent-clipped=0.0 2024-09-23 16:25:58,687 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=298508.0, ans=0.125 2024-09-23 16:25:59,071 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=11.71 vs. limit=22.5 2024-09-23 16:26:00,217 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=298508.0, ans=0.125 2024-09-23 16:26:15,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=298554.6666666667, ans=0.125 2024-09-23 16:26:18,870 INFO [train.py:1198] (2/4) Epoch 17, batch 1650, loss[loss=0.1986, ctc_loss=0.1265, cr_loss=0.3602, over 17057.00 frames. ], tot_loss[loss=0.223, ctc_loss=0.1499, cr_loss=0.3657, over 3370342.45 frames. ], batch size: 46, lr: 7.10e-03, grad_scale: 32.0 2024-09-23 16:26:39,309 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=298648.0, ans=0.0 2024-09-23 16:26:51,724 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=9.70 vs. limit=15.0 2024-09-23 16:26:54,202 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=298694.6666666667, ans=0.125 2024-09-23 16:27:00,550 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=298694.6666666667, ans=0.125 2024-09-23 16:27:25,384 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=11.35 vs. limit=15.0 2024-09-23 16:27:45,762 INFO [train.py:1198] (2/4) Epoch 17, batch 1700, loss[loss=0.233, ctc_loss=0.1554, cr_loss=0.3878, over 17229.00 frames. ], tot_loss[loss=0.2238, ctc_loss=0.1504, cr_loss=0.3667, over 3369239.90 frames. ], batch size: 55, lr: 7.09e-03, grad_scale: 32.0 2024-09-23 16:28:02,752 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=9.09 vs. limit=15.0 2024-09-23 16:28:19,964 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=298928.0, ans=0.0 2024-09-23 16:28:27,766 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=298928.0, ans=0.125 2024-09-23 16:28:33,028 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=19.85 vs. limit=22.5 2024-09-23 16:28:38,521 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.087e+02 1.237e+02 1.323e+02 1.443e+02 1.876e+02, threshold=2.646e+02, percent-clipped=0.0 2024-09-23 16:29:05,468 INFO [train.py:1198] (2/4) Epoch 17, batch 1750, loss[loss=0.2156, ctc_loss=0.1457, cr_loss=0.349, over 17135.00 frames. ], tot_loss[loss=0.2213, ctc_loss=0.1486, cr_loss=0.3639, over 3377026.93 frames. ], batch size: 48, lr: 7.09e-03, grad_scale: 32.0 2024-09-23 16:29:05,853 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=299068.0, ans=0.125 2024-09-23 16:29:20,722 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn1.whiten, num_groups=1, num_channels=192, metric=10.40 vs. limit=22.5 2024-09-23 16:29:24,753 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=299114.6666666667, ans=0.2 2024-09-23 16:29:28,009 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=299114.6666666667, ans=0.0 2024-09-23 16:30:06,575 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=14.18 vs. limit=15.0 2024-09-23 16:30:27,807 INFO [train.py:1198] (2/4) Epoch 17, batch 1800, loss[loss=0.2638, ctc_loss=0.1839, cr_loss=0.3993, over 16446.00 frames. ], tot_loss[loss=0.2215, ctc_loss=0.1487, cr_loss=0.3638, over 3377422.81 frames. ], batch size: 66, lr: 7.09e-03, grad_scale: 32.0 2024-09-23 16:30:56,290 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 16:31:12,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=299394.6666666667, ans=0.1 2024-09-23 16:31:17,084 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=299441.3333333333, ans=0.125 2024-09-23 16:31:23,056 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.041e+02 1.259e+02 1.337e+02 1.488e+02 2.205e+02, threshold=2.673e+02, percent-clipped=0.0 2024-09-23 16:31:23,318 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.max_abs, batch_count=299441.3333333333, ans=10.0 2024-09-23 16:31:52,511 INFO [train.py:1198] (2/4) Epoch 17, batch 1850, loss[loss=0.2382, ctc_loss=0.1607, cr_loss=0.3878, over 16928.00 frames. ], tot_loss[loss=0.2222, ctc_loss=0.1494, cr_loss=0.3642, over 3367811.58 frames. ], batch size: 58, lr: 7.09e-03, grad_scale: 32.0 2024-09-23 16:32:14,067 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.69 vs. limit=15.0 2024-09-23 16:32:15,845 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.17 vs. limit=22.5 2024-09-23 16:33:07,334 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=299721.3333333333, ans=0.125 2024-09-23 16:33:14,614 INFO [train.py:1198] (2/4) Epoch 17, batch 1900, loss[loss=0.1872, ctc_loss=0.1224, cr_loss=0.3243, over 17269.00 frames. ], tot_loss[loss=0.2213, ctc_loss=0.1487, cr_loss=0.3635, over 3372129.24 frames. ], batch size: 42, lr: 7.08e-03, grad_scale: 32.0 2024-09-23 16:33:16,652 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=299768.0, ans=0.0 2024-09-23 16:34:02,282 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=299908.0, ans=0.5 2024-09-23 16:34:06,709 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.256e+02 1.309e+02 1.429e+02 1.873e+02, threshold=2.618e+02, percent-clipped=0.0 2024-09-23 16:34:24,391 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=299954.6666666667, ans=0.125 2024-09-23 16:34:33,599 INFO [train.py:1198] (2/4) Epoch 17, batch 1950, loss[loss=0.2369, ctc_loss=0.1583, cr_loss=0.3931, over 17007.00 frames. ], tot_loss[loss=0.2211, ctc_loss=0.1484, cr_loss=0.3631, over 3363733.34 frames. ], batch size: 53, lr: 7.08e-03, grad_scale: 16.0 2024-09-23 16:34:40,154 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=300001.3333333333, ans=0.125 2024-09-23 16:34:40,423 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.21 vs. limit=15.0 2024-09-23 16:34:41,697 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=300001.3333333333, ans=0.0 2024-09-23 16:34:59,473 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=300048.0, ans=0.125 2024-09-23 16:35:29,398 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=8.99 vs. limit=15.0 2024-09-23 16:35:58,734 INFO [train.py:1198] (2/4) Epoch 17, batch 2000, loss[loss=0.2173, ctc_loss=0.1454, cr_loss=0.3592, over 17027.00 frames. ], tot_loss[loss=0.2209, ctc_loss=0.1485, cr_loss=0.3621, over 3358447.86 frames. ], batch size: 51, lr: 7.08e-03, grad_scale: 32.0 2024-09-23 16:36:03,809 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=300234.6666666667, ans=0.125 2024-09-23 16:36:05,424 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=300234.6666666667, ans=0.0 2024-09-23 16:36:24,492 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.06 vs. limit=15.0 2024-09-23 16:36:27,314 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=300281.3333333333, ans=0.125 2024-09-23 16:36:28,944 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=300281.3333333333, ans=0.05 2024-09-23 16:36:28,963 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.min_positive, batch_count=300281.3333333333, ans=0.05 2024-09-23 16:36:32,066 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=300328.0, ans=0.2 2024-09-23 16:36:55,626 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.280e+02 1.363e+02 1.514e+02 2.601e+02, threshold=2.726e+02, percent-clipped=0.0 2024-09-23 16:37:21,170 INFO [train.py:1198] (2/4) Epoch 17, batch 2050, loss[loss=0.2319, ctc_loss=0.1585, cr_loss=0.3669, over 17049.00 frames. ], tot_loss[loss=0.2225, ctc_loss=0.1498, cr_loss=0.3639, over 3345091.53 frames. ], batch size: 56, lr: 7.08e-03, grad_scale: 32.0 2024-09-23 16:37:49,265 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=300514.6666666667, ans=0.1 2024-09-23 16:37:50,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=300514.6666666667, ans=0.5 2024-09-23 16:37:55,585 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=300561.3333333333, ans=0.125 2024-09-23 16:38:31,932 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=300654.6666666667, ans=0.0 2024-09-23 16:38:36,884 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=300654.6666666667, ans=0.1 2024-09-23 16:38:38,960 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=13.07 vs. limit=22.5 2024-09-23 16:38:43,057 INFO [train.py:1198] (2/4) Epoch 17, batch 2100, loss[loss=0.2637, ctc_loss=0.1793, cr_loss=0.4218, over 17210.00 frames. ], tot_loss[loss=0.2236, ctc_loss=0.1506, cr_loss=0.3653, over 3344482.18 frames. ], batch size: 55, lr: 7.07e-03, grad_scale: 32.0 2024-09-23 16:39:28,538 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=7.22 vs. limit=15.0 2024-09-23 16:39:32,940 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer_na.min_abs, batch_count=300841.3333333333, ans=0.02 2024-09-23 16:39:37,336 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.012e+02 1.282e+02 1.378e+02 1.629e+02 2.500e+02, threshold=2.755e+02, percent-clipped=0.0 2024-09-23 16:40:05,318 INFO [train.py:1198] (2/4) Epoch 17, batch 2150, loss[loss=0.1723, ctc_loss=0.1111, cr_loss=0.306, over 17181.00 frames. ], tot_loss[loss=0.2234, ctc_loss=0.1503, cr_loss=0.3653, over 3350374.83 frames. ], batch size: 41, lr: 7.07e-03, grad_scale: 32.0 2024-09-23 16:40:08,907 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=300934.6666666667, ans=0.0 2024-09-23 16:40:08,912 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=300934.6666666667, ans=0.0 2024-09-23 16:40:21,868 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=300981.3333333333, ans=0.125 2024-09-23 16:40:39,888 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=301028.0, ans=0.125 2024-09-23 16:40:44,601 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=301028.0, ans=0.125 2024-09-23 16:40:55,763 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=301074.6666666667, ans=0.2 2024-09-23 16:41:00,423 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=301074.6666666667, ans=0.0 2024-09-23 16:41:17,969 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.69 vs. limit=12.0 2024-09-23 16:41:25,114 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=301121.3333333333, ans=0.125 2024-09-23 16:41:29,568 INFO [train.py:1198] (2/4) Epoch 17, batch 2200, loss[loss=0.192, ctc_loss=0.1249, cr_loss=0.3352, over 16988.00 frames. ], tot_loss[loss=0.2234, ctc_loss=0.1502, cr_loss=0.3658, over 3358849.16 frames. ], batch size: 44, lr: 7.07e-03, grad_scale: 32.0 2024-09-23 16:41:33,004 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=301168.0, ans=0.125 2024-09-23 16:42:11,119 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=16.53 vs. limit=22.5 2024-09-23 16:42:23,222 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.050e+02 1.227e+02 1.315e+02 1.424e+02 2.310e+02, threshold=2.629e+02, percent-clipped=0.0 2024-09-23 16:42:23,561 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=301308.0, ans=0.0 2024-09-23 16:42:40,447 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=301354.6666666667, ans=0.025 2024-09-23 16:42:45,167 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=301354.6666666667, ans=0.125 2024-09-23 16:42:47,963 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=9.89 vs. limit=15.0 2024-09-23 16:42:51,584 INFO [train.py:1198] (2/4) Epoch 17, batch 2250, loss[loss=0.2237, ctc_loss=0.1517, cr_loss=0.3599, over 17106.00 frames. ], tot_loss[loss=0.2235, ctc_loss=0.1503, cr_loss=0.3659, over 3353909.86 frames. ], batch size: 49, lr: 7.07e-03, grad_scale: 32.0 2024-09-23 16:42:52,010 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=301401.3333333333, ans=0.125 2024-09-23 16:42:53,361 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=301401.3333333333, ans=0.0 2024-09-23 16:42:59,798 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=301401.3333333333, ans=0.125 2024-09-23 16:43:12,312 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=301448.0, ans=0.0 2024-09-23 16:43:18,826 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=301448.0, ans=0.125 2024-09-23 16:44:11,216 INFO [train.py:1198] (2/4) Epoch 17, batch 2300, loss[loss=0.2185, ctc_loss=0.1457, cr_loss=0.3639, over 17002.00 frames. ], tot_loss[loss=0.2242, ctc_loss=0.1509, cr_loss=0.3668, over 3356142.19 frames. ], batch size: 53, lr: 7.06e-03, grad_scale: 32.0 2024-09-23 16:44:27,905 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=512, metric=14.84 vs. limit=22.5 2024-09-23 16:44:37,129 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 16:44:42,116 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=4.76 vs. limit=12.0 2024-09-23 16:44:56,107 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=301728.0, ans=0.025 2024-09-23 16:45:02,456 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.01 vs. limit=22.5 2024-09-23 16:45:05,240 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=301774.6666666667, ans=0.0 2024-09-23 16:45:08,137 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.084e+02 1.278e+02 1.376e+02 1.551e+02 2.468e+02, threshold=2.752e+02, percent-clipped=0.0 2024-09-23 16:45:28,470 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=301821.3333333333, ans=0.125 2024-09-23 16:45:35,976 INFO [train.py:1198] (2/4) Epoch 17, batch 2350, loss[loss=0.278, ctc_loss=0.2005, cr_loss=0.3875, over 12375.00 frames. ], tot_loss[loss=0.2238, ctc_loss=0.1504, cr_loss=0.367, over 3355316.92 frames. ], batch size: 123, lr: 7.06e-03, grad_scale: 32.0 2024-09-23 16:45:42,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=301868.0, ans=0.2 2024-09-23 16:45:52,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=301914.6666666667, ans=0.125 2024-09-23 16:46:02,948 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=301914.6666666667, ans=0.1 2024-09-23 16:46:09,348 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 16:46:42,268 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=302054.6666666667, ans=0.125 2024-09-23 16:46:57,671 INFO [train.py:1198] (2/4) Epoch 17, batch 2400, loss[loss=0.2225, ctc_loss=0.1508, cr_loss=0.3585, over 17296.00 frames. ], tot_loss[loss=0.2249, ctc_loss=0.1513, cr_loss=0.3676, over 3343394.89 frames. ], batch size: 49, lr: 7.06e-03, grad_scale: 32.0 2024-09-23 16:47:54,481 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.150e+02 1.300e+02 1.437e+02 1.590e+02 2.245e+02, threshold=2.874e+02, percent-clipped=0.0 2024-09-23 16:48:04,379 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=302288.0, ans=0.125 2024-09-23 16:48:14,409 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.33 vs. limit=15.0 2024-09-23 16:48:15,413 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=302288.0, ans=0.0 2024-09-23 16:48:19,880 INFO [train.py:1198] (2/4) Epoch 17, batch 2450, loss[loss=0.2624, ctc_loss=0.1784, cr_loss=0.4199, over 17042.00 frames. ], tot_loss[loss=0.2253, ctc_loss=0.1518, cr_loss=0.3679, over 3337692.18 frames. ], batch size: 52, lr: 7.05e-03, grad_scale: 32.0 2024-09-23 16:48:23,298 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.min_positive, batch_count=302334.6666666667, ans=0.05 2024-09-23 16:48:28,702 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.17 vs. limit=15.0 2024-09-23 16:48:29,780 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=302334.6666666667, ans=0.2 2024-09-23 16:48:34,644 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=302381.3333333333, ans=0.125 2024-09-23 16:48:54,251 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=302428.0, ans=0.0 2024-09-23 16:48:57,291 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=302428.0, ans=0.2 2024-09-23 16:49:04,542 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=7.10 vs. limit=8.0 2024-09-23 16:49:17,622 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=302474.6666666667, ans=0.1 2024-09-23 16:49:17,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=302474.6666666667, ans=0.07 2024-09-23 16:49:29,103 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=302521.3333333333, ans=0.0 2024-09-23 16:49:39,967 INFO [train.py:1198] (2/4) Epoch 17, batch 2500, loss[loss=0.205, ctc_loss=0.1381, cr_loss=0.3341, over 17157.00 frames. ], tot_loss[loss=0.225, ctc_loss=0.1515, cr_loss=0.3679, over 3338265.63 frames. ], batch size: 48, lr: 7.05e-03, grad_scale: 32.0 2024-09-23 16:49:40,609 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.26 vs. limit=15.0 2024-09-23 16:49:58,693 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=302614.6666666667, ans=0.125 2024-09-23 16:50:27,978 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=5.18 vs. limit=15.0 2024-09-23 16:50:35,713 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=17.74 vs. limit=22.5 2024-09-23 16:50:39,595 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.051e+02 1.277e+02 1.416e+02 1.598e+02 3.065e+02, threshold=2.832e+02, percent-clipped=1.0 2024-09-23 16:50:43,172 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 16:50:43,198 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=302708.0, ans=0.05 2024-09-23 16:50:55,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=302754.6666666667, ans=0.125 2024-09-23 16:50:56,356 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=3.31 vs. limit=6.0 2024-09-23 16:51:04,638 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=302754.6666666667, ans=0.125 2024-09-23 16:51:07,612 INFO [train.py:1198] (2/4) Epoch 17, batch 2550, loss[loss=0.226, ctc_loss=0.1508, cr_loss=0.376, over 17288.00 frames. ], tot_loss[loss=0.2239, ctc_loss=0.1506, cr_loss=0.3668, over 3350452.53 frames. ], batch size: 46, lr: 7.05e-03, grad_scale: 32.0 2024-09-23 16:51:43,167 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=302894.6666666667, ans=0.125 2024-09-23 16:51:44,062 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=11.71 vs. limit=22.5 2024-09-23 16:52:08,427 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=302941.3333333333, ans=0.1 2024-09-23 16:52:13,603 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=8.06 vs. limit=15.0 2024-09-23 16:52:29,691 INFO [train.py:1198] (2/4) Epoch 17, batch 2600, loss[loss=0.2078, ctc_loss=0.1404, cr_loss=0.337, over 17226.00 frames. ], tot_loss[loss=0.2224, ctc_loss=0.1493, cr_loss=0.3651, over 3350336.86 frames. ], batch size: 50, lr: 7.05e-03, grad_scale: 32.0 2024-09-23 16:52:49,134 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 16:52:50,739 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=303081.3333333333, ans=0.2 2024-09-23 16:52:56,901 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=303081.3333333333, ans=0.2 2024-09-23 16:53:05,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=303128.0, ans=0.125 2024-09-23 16:53:09,673 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=303128.0, ans=0.0 2024-09-23 16:53:23,708 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.065e+02 1.309e+02 1.432e+02 1.509e+02 2.078e+02, threshold=2.863e+02, percent-clipped=0.0 2024-09-23 16:53:43,658 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.87 vs. limit=15.0 2024-09-23 16:53:49,176 INFO [train.py:1198] (2/4) Epoch 17, batch 2650, loss[loss=0.232, ctc_loss=0.1541, cr_loss=0.3897, over 15985.00 frames. ], tot_loss[loss=0.2217, ctc_loss=0.1488, cr_loss=0.3645, over 3352833.26 frames. ], batch size: 74, lr: 7.04e-03, grad_scale: 32.0 2024-09-23 16:53:49,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=303268.0, ans=0.1 2024-09-23 16:53:52,625 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-23 16:53:52,643 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=303268.0, ans=0.125 2024-09-23 16:54:05,317 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=303314.6666666667, ans=0.125 2024-09-23 16:54:24,519 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=303361.3333333333, ans=0.2 2024-09-23 16:55:13,734 INFO [train.py:1198] (2/4) Epoch 17, batch 2700, loss[loss=0.2497, ctc_loss=0.1697, cr_loss=0.4001, over 16472.00 frames. ], tot_loss[loss=0.2228, ctc_loss=0.1497, cr_loss=0.3655, over 3352318.08 frames. ], batch size: 66, lr: 7.04e-03, grad_scale: 32.0 2024-09-23 16:55:26,502 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=303501.3333333333, ans=0.0 2024-09-23 16:56:01,233 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.33 vs. limit=15.0 2024-09-23 16:56:10,277 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.142e+02 1.299e+02 1.395e+02 1.578e+02 3.213e+02, threshold=2.790e+02, percent-clipped=1.0 2024-09-23 16:56:35,876 INFO [train.py:1198] (2/4) Epoch 17, batch 2750, loss[loss=0.2349, ctc_loss=0.1612, cr_loss=0.3685, over 16489.00 frames. ], tot_loss[loss=0.2224, ctc_loss=0.1495, cr_loss=0.3647, over 3352331.97 frames. ], batch size: 66, lr: 7.04e-03, grad_scale: 32.0 2024-09-23 16:56:48,954 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.max_abs, batch_count=303734.6666666667, ans=10.0 2024-09-23 16:56:57,122 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=303781.3333333333, ans=0.0 2024-09-23 16:57:00,282 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=303781.3333333333, ans=0.0 2024-09-23 16:57:20,498 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.63 vs. limit=15.0 2024-09-23 16:57:58,357 INFO [train.py:1198] (2/4) Epoch 17, batch 2800, loss[loss=0.2114, ctc_loss=0.1386, cr_loss=0.3644, over 17050.00 frames. ], tot_loss[loss=0.222, ctc_loss=0.1492, cr_loss=0.3639, over 3351458.09 frames. ], batch size: 46, lr: 7.04e-03, grad_scale: 32.0 2024-09-23 16:58:35,062 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=304061.3333333333, ans=0.125 2024-09-23 16:58:39,711 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=304061.3333333333, ans=0.125 2024-09-23 16:58:47,688 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.max_abs, batch_count=304108.0, ans=10.0 2024-09-23 16:58:51,517 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=5.09 vs. limit=10.0 2024-09-23 16:58:53,780 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.311e+02 1.385e+02 1.466e+02 2.534e+02, threshold=2.770e+02, percent-clipped=0.0 2024-09-23 16:58:54,118 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 16:58:54,788 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.96 vs. limit=15.0 2024-09-23 16:59:07,236 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=14.18 vs. limit=15.0 2024-09-23 16:59:17,881 INFO [train.py:1198] (2/4) Epoch 17, batch 2850, loss[loss=0.2625, ctc_loss=0.1838, cr_loss=0.3933, over 15128.00 frames. ], tot_loss[loss=0.2217, ctc_loss=0.1491, cr_loss=0.3629, over 3348214.04 frames. ], batch size: 89, lr: 7.03e-03, grad_scale: 16.0 2024-09-23 16:59:22,745 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=304201.3333333333, ans=0.125 2024-09-23 16:59:37,232 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=304248.0, ans=0.0 2024-09-23 16:59:40,687 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=304248.0, ans=0.0 2024-09-23 16:59:41,351 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.84 vs. limit=10.0 2024-09-23 17:00:05,943 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=7.49 vs. limit=15.0 2024-09-23 17:00:19,177 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=304341.3333333333, ans=0.1 2024-09-23 17:00:24,084 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=304341.3333333333, ans=0.125 2024-09-23 17:00:28,827 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=304388.0, ans=0.125 2024-09-23 17:00:33,477 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=304388.0, ans=0.2 2024-09-23 17:00:42,927 INFO [train.py:1198] (2/4) Epoch 17, batch 2900, loss[loss=0.2506, ctc_loss=0.1688, cr_loss=0.4089, over 16903.00 frames. ], tot_loss[loss=0.2223, ctc_loss=0.1495, cr_loss=0.3643, over 3353714.81 frames. ], batch size: 58, lr: 7.03e-03, grad_scale: 16.0 2024-09-23 17:01:01,638 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=304481.3333333333, ans=0.1 2024-09-23 17:01:01,661 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=304481.3333333333, ans=0.04949747468305833 2024-09-23 17:01:16,153 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=304528.0, ans=0.2 2024-09-23 17:01:20,032 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.95 vs. limit=15.0 2024-09-23 17:01:25,698 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=304528.0, ans=0.1 2024-09-23 17:01:42,986 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.082e+02 1.263e+02 1.350e+02 1.442e+02 2.620e+02, threshold=2.699e+02, percent-clipped=0.0 2024-09-23 17:01:43,358 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=304574.6666666667, ans=0.025 2024-09-23 17:01:57,732 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=304621.3333333333, ans=0.1 2024-09-23 17:01:59,234 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.min_positive, batch_count=304621.3333333333, ans=0.05 2024-09-23 17:01:59,333 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=304621.3333333333, ans=0.0 2024-09-23 17:02:04,212 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=304668.0, ans=0.2 2024-09-23 17:02:05,521 INFO [train.py:1198] (2/4) Epoch 17, batch 2950, loss[loss=0.2334, ctc_loss=0.1606, cr_loss=0.3643, over 17042.00 frames. ], tot_loss[loss=0.2229, ctc_loss=0.1499, cr_loss=0.3648, over 3353514.40 frames. ], batch size: 56, lr: 7.03e-03, grad_scale: 8.0 2024-09-23 17:02:21,635 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=304714.6666666667, ans=0.125 2024-09-23 17:03:26,627 INFO [train.py:1198] (2/4) Epoch 17, batch 3000, loss[loss=0.2156, ctc_loss=0.1413, cr_loss=0.3715, over 17113.00 frames. ], tot_loss[loss=0.2222, ctc_loss=0.1493, cr_loss=0.3644, over 3351073.12 frames. ], batch size: 49, lr: 7.02e-03, grad_scale: 8.0 2024-09-23 17:03:26,628 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 17:03:42,424 INFO [train.py:1230] (2/4) Epoch 17, validation: loss=0.0409, ctc_loss=0.0409, cr_loss=7.678e-15, over 944034.00 frames. 2024-09-23 17:03:42,425 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 17:04:10,574 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=304948.0, ans=0.0 2024-09-23 17:04:38,120 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.276e+02 1.374e+02 1.513e+02 2.906e+02, threshold=2.749e+02, percent-clipped=1.0 2024-09-23 17:04:59,730 INFO [train.py:1198] (2/4) Epoch 17, batch 3050, loss[loss=0.1905, ctc_loss=0.1284, cr_loss=0.3102, over 17166.00 frames. ], tot_loss[loss=0.2219, ctc_loss=0.1492, cr_loss=0.3635, over 3335771.12 frames. ], batch size: 45, lr: 7.02e-03, grad_scale: 8.0 2024-09-23 17:05:01,655 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=305134.6666666667, ans=0.125 2024-09-23 17:05:37,750 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=305228.0, ans=0.1 2024-09-23 17:06:04,224 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=3.96 vs. limit=10.0 2024-09-23 17:06:20,561 INFO [train.py:1198] (2/4) Epoch 17, batch 3100, loss[loss=0.2509, ctc_loss=0.1724, cr_loss=0.3922, over 17012.00 frames. ], tot_loss[loss=0.2222, ctc_loss=0.1495, cr_loss=0.3636, over 3341546.46 frames. ], batch size: 56, lr: 7.02e-03, grad_scale: 8.0 2024-09-23 17:06:51,591 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn2.whiten, num_groups=1, num_channels=192, metric=8.64 vs. limit=22.5 2024-09-23 17:07:00,869 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=305461.3333333333, ans=0.0 2024-09-23 17:07:19,080 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.257e+02 1.347e+02 1.447e+02 2.070e+02, threshold=2.694e+02, percent-clipped=0.0 2024-09-23 17:07:19,363 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=305508.0, ans=0.0 2024-09-23 17:07:31,948 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=305554.6666666667, ans=0.0 2024-09-23 17:07:41,116 INFO [train.py:1198] (2/4) Epoch 17, batch 3150, loss[loss=0.2414, ctc_loss=0.1641, cr_loss=0.3866, over 17006.00 frames. ], tot_loss[loss=0.2213, ctc_loss=0.1489, cr_loss=0.3619, over 3344860.95 frames. ], batch size: 53, lr: 7.02e-03, grad_scale: 8.0 2024-09-23 17:08:18,619 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=305694.6666666667, ans=0.0 2024-09-23 17:08:28,534 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=20.36 vs. limit=22.5 2024-09-23 17:08:38,617 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.72 vs. limit=15.0 2024-09-23 17:08:45,230 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=305788.0, ans=0.025 2024-09-23 17:09:00,534 INFO [train.py:1198] (2/4) Epoch 17, batch 3200, loss[loss=0.2104, ctc_loss=0.1407, cr_loss=0.3488, over 17319.00 frames. ], tot_loss[loss=0.2219, ctc_loss=0.1493, cr_loss=0.3632, over 3348596.90 frames. ], batch size: 51, lr: 7.01e-03, grad_scale: 16.0 2024-09-23 17:09:38,441 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.05 vs. limit=22.5 2024-09-23 17:09:40,108 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.31 vs. limit=15.0 2024-09-23 17:09:52,057 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=305974.6666666667, ans=0.125 2024-09-23 17:09:56,502 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.224e+02 1.300e+02 1.391e+02 2.057e+02, threshold=2.599e+02, percent-clipped=0.0 2024-09-23 17:10:01,556 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 17:10:03,471 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.52 vs. limit=10.0 2024-09-23 17:10:18,312 INFO [train.py:1198] (2/4) Epoch 17, batch 3250, loss[loss=0.2474, ctc_loss=0.1678, cr_loss=0.3979, over 17229.00 frames. ], tot_loss[loss=0.2221, ctc_loss=0.1495, cr_loss=0.363, over 3348543.00 frames. ], batch size: 50, lr: 7.01e-03, grad_scale: 16.0 2024-09-23 17:10:25,259 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=5.75 vs. limit=12.0 2024-09-23 17:10:31,082 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=306068.0, ans=0.125 2024-09-23 17:10:40,702 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=306114.6666666667, ans=0.2 2024-09-23 17:11:30,979 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=306254.6666666667, ans=0.0 2024-09-23 17:11:32,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=306254.6666666667, ans=0.0 2024-09-23 17:11:36,836 INFO [train.py:1198] (2/4) Epoch 17, batch 3300, loss[loss=0.2238, ctc_loss=0.1499, cr_loss=0.3697, over 16758.00 frames. ], tot_loss[loss=0.2212, ctc_loss=0.1488, cr_loss=0.3621, over 3359800.18 frames. ], batch size: 61, lr: 7.01e-03, grad_scale: 16.0 2024-09-23 17:11:45,685 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=306301.3333333333, ans=0.0 2024-09-23 17:11:56,776 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=4.41 vs. limit=12.0 2024-09-23 17:12:05,756 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=306348.0, ans=0.0 2024-09-23 17:12:08,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=306394.6666666667, ans=0.0 2024-09-23 17:12:22,708 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=306394.6666666667, ans=0.2 2024-09-23 17:12:34,842 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.074e+02 1.271e+02 1.388e+02 1.557e+02 2.598e+02, threshold=2.776e+02, percent-clipped=0.0 2024-09-23 17:12:37,191 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=9.72 vs. limit=15.0 2024-09-23 17:12:38,601 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=6.38 vs. limit=15.0 2024-09-23 17:12:49,337 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=306488.0, ans=0.125 2024-09-23 17:12:56,557 INFO [train.py:1198] (2/4) Epoch 17, batch 3350, loss[loss=0.2621, ctc_loss=0.1799, cr_loss=0.4108, over 15032.00 frames. ], tot_loss[loss=0.2231, ctc_loss=0.1501, cr_loss=0.3648, over 3350802.16 frames. ], batch size: 89, lr: 7.01e-03, grad_scale: 16.0 2024-09-23 17:13:29,841 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 17:13:34,160 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=306628.0, ans=0.0 2024-09-23 17:13:38,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=306628.0, ans=0.125 2024-09-23 17:13:44,122 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.04 vs. limit=15.0 2024-09-23 17:13:51,189 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=306674.6666666667, ans=0.2 2024-09-23 17:13:57,414 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=306721.3333333333, ans=0.1 2024-09-23 17:14:14,499 INFO [train.py:1198] (2/4) Epoch 17, batch 3400, loss[loss=0.198, ctc_loss=0.1352, cr_loss=0.3138, over 17030.00 frames. ], tot_loss[loss=0.223, ctc_loss=0.1501, cr_loss=0.3644, over 3352617.33 frames. ], batch size: 39, lr: 7.00e-03, grad_scale: 16.0 2024-09-23 17:14:20,808 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=306768.0, ans=0.125 2024-09-23 17:14:35,157 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.05 vs. limit=12.0 2024-09-23 17:14:48,511 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=306861.3333333333, ans=0.125 2024-09-23 17:14:54,641 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=306861.3333333333, ans=0.1 2024-09-23 17:15:09,980 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.095e+02 1.230e+02 1.305e+02 1.432e+02 2.019e+02, threshold=2.610e+02, percent-clipped=0.0 2024-09-23 17:15:14,843 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=306954.6666666667, ans=0.125 2024-09-23 17:15:26,075 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.max_positive, batch_count=306954.6666666667, ans=0.95 2024-09-23 17:15:32,031 INFO [train.py:1198] (2/4) Epoch 17, batch 3450, loss[loss=0.2341, ctc_loss=0.1589, cr_loss=0.3759, over 17219.00 frames. ], tot_loss[loss=0.2231, ctc_loss=0.1502, cr_loss=0.3644, over 3357764.36 frames. ], batch size: 55, lr: 7.00e-03, grad_scale: 16.0 2024-09-23 17:15:40,141 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=307001.3333333333, ans=0.125 2024-09-23 17:16:25,955 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=307141.3333333333, ans=0.2 2024-09-23 17:16:36,844 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=307188.0, ans=0.0 2024-09-23 17:16:36,895 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=307188.0, ans=0.125 2024-09-23 17:16:39,944 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=307188.0, ans=0.0 2024-09-23 17:16:52,833 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=307234.6666666667, ans=0.125 2024-09-23 17:16:53,994 INFO [train.py:1198] (2/4) Epoch 17, batch 3500, loss[loss=0.2494, ctc_loss=0.1676, cr_loss=0.4093, over 17026.00 frames. ], tot_loss[loss=0.2235, ctc_loss=0.1505, cr_loss=0.3649, over 3354762.55 frames. ], batch size: 51, lr: 7.00e-03, grad_scale: 16.0 2024-09-23 17:17:00,602 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=307234.6666666667, ans=0.125 2024-09-23 17:17:25,511 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=307328.0, ans=0.125 2024-09-23 17:17:50,323 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.079e+02 1.294e+02 1.368e+02 1.478e+02 3.708e+02, threshold=2.737e+02, percent-clipped=1.0 2024-09-23 17:18:01,419 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=307421.3333333333, ans=0.025 2024-09-23 17:18:09,319 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=307421.3333333333, ans=0.0 2024-09-23 17:18:10,861 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=307468.0, ans=0.025 2024-09-23 17:18:12,054 INFO [train.py:1198] (2/4) Epoch 17, batch 3550, loss[loss=0.2153, ctc_loss=0.1408, cr_loss=0.3722, over 16971.00 frames. ], tot_loss[loss=0.2235, ctc_loss=0.1505, cr_loss=0.3653, over 3358674.52 frames. ], batch size: 42, lr: 7.00e-03, grad_scale: 16.0 2024-09-23 17:18:18,522 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=307468.0, ans=0.2 2024-09-23 17:18:18,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=307468.0, ans=0.1 2024-09-23 17:18:25,115 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=307468.0, ans=0.0 2024-09-23 17:18:28,461 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=307514.6666666667, ans=0.0 2024-09-23 17:18:42,418 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=307514.6666666667, ans=0.025 2024-09-23 17:19:13,219 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=307608.0, ans=0.0 2024-09-23 17:19:20,835 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=307654.6666666667, ans=0.0 2024-09-23 17:19:31,530 INFO [train.py:1198] (2/4) Epoch 17, batch 3600, loss[loss=0.2412, ctc_loss=0.163, cr_loss=0.3912, over 17211.00 frames. ], tot_loss[loss=0.2236, ctc_loss=0.1505, cr_loss=0.3658, over 3356739.26 frames. ], batch size: 50, lr: 6.99e-03, grad_scale: 16.0 2024-09-23 17:19:42,843 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=307701.3333333333, ans=0.1 2024-09-23 17:19:45,759 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=307748.0, ans=0.125 2024-09-23 17:20:05,917 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=307794.6666666667, ans=0.125 2024-09-23 17:20:07,967 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.26 vs. limit=22.5 2024-09-23 17:20:09,082 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=307794.6666666667, ans=0.125 2024-09-23 17:20:20,149 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=307841.3333333333, ans=0.2 2024-09-23 17:20:21,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=307841.3333333333, ans=0.125 2024-09-23 17:20:29,251 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.066e+02 1.229e+02 1.313e+02 1.436e+02 1.870e+02, threshold=2.625e+02, percent-clipped=0.0 2024-09-23 17:20:31,932 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn1.whiten, num_groups=1, num_channels=192, metric=9.02 vs. limit=22.5 2024-09-23 17:20:48,141 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=307934.6666666667, ans=0.0 2024-09-23 17:20:49,540 INFO [train.py:1198] (2/4) Epoch 17, batch 3650, loss[loss=0.2361, ctc_loss=0.1569, cr_loss=0.3962, over 17097.00 frames. ], tot_loss[loss=0.2244, ctc_loss=0.151, cr_loss=0.367, over 3356521.10 frames. ], batch size: 49, lr: 6.99e-03, grad_scale: 16.0 2024-09-23 17:21:08,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=307981.3333333333, ans=0.125 2024-09-23 17:21:16,823 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.max_abs, batch_count=307981.3333333333, ans=10.0 2024-09-23 17:21:39,093 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.35 vs. limit=10.0 2024-09-23 17:21:41,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=308074.6666666667, ans=0.125 2024-09-23 17:21:44,384 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=308074.6666666667, ans=0.1 2024-09-23 17:21:52,464 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=308074.6666666667, ans=0.2 2024-09-23 17:21:52,480 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=308074.6666666667, ans=0.0 2024-09-23 17:21:52,481 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=308074.6666666667, ans=0.1 2024-09-23 17:21:53,835 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=308121.3333333333, ans=0.0 2024-09-23 17:22:00,032 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=308121.3333333333, ans=0.0 2024-09-23 17:22:10,729 INFO [train.py:1198] (2/4) Epoch 17, batch 3700, loss[loss=0.1958, ctc_loss=0.1296, cr_loss=0.3307, over 15844.00 frames. ], tot_loss[loss=0.2248, ctc_loss=0.1514, cr_loss=0.3672, over 3346986.18 frames. ], batch size: 35, lr: 6.99e-03, grad_scale: 16.0 2024-09-23 17:22:22,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=308168.0, ans=0.125 2024-09-23 17:22:22,141 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=308168.0, ans=0.125 2024-09-23 17:22:23,781 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=308168.0, ans=0.0 2024-09-23 17:22:24,520 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.57 vs. limit=15.0 2024-09-23 17:22:54,921 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=308261.3333333333, ans=0.125 2024-09-23 17:23:08,758 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.143e+02 1.332e+02 1.441e+02 1.620e+02 2.318e+02, threshold=2.882e+02, percent-clipped=0.0 2024-09-23 17:23:28,818 INFO [train.py:1198] (2/4) Epoch 17, batch 3750, loss[loss=0.2942, ctc_loss=0.2113, cr_loss=0.4147, over 11580.00 frames. ], tot_loss[loss=0.2249, ctc_loss=0.1515, cr_loss=0.367, over 3330539.89 frames. ], batch size: 123, lr: 6.99e-03, grad_scale: 16.0 2024-09-23 17:23:40,248 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=308401.3333333333, ans=0.2 2024-09-23 17:23:49,563 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=308448.0, ans=0.1 2024-09-23 17:24:00,621 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=308494.6666666667, ans=0.07 2024-09-23 17:24:02,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=308494.6666666667, ans=0.1 2024-09-23 17:24:16,351 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=308541.3333333333, ans=0.2 2024-09-23 17:24:28,779 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=308541.3333333333, ans=0.125 2024-09-23 17:24:34,944 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=308588.0, ans=0.0 2024-09-23 17:24:47,306 INFO [train.py:1198] (2/4) Epoch 17, batch 3800, loss[loss=0.2135, ctc_loss=0.1421, cr_loss=0.357, over 17359.00 frames. ], tot_loss[loss=0.2245, ctc_loss=0.1512, cr_loss=0.3668, over 3330195.08 frames. ], batch size: 48, lr: 6.98e-03, grad_scale: 16.0 2024-09-23 17:24:52,323 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=308634.6666666667, ans=0.0 2024-09-23 17:24:55,589 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=308634.6666666667, ans=0.125 2024-09-23 17:25:07,173 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.46 vs. limit=6.0 2024-09-23 17:25:11,195 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=308681.3333333333, ans=0.1 2024-09-23 17:25:16,187 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=308681.3333333333, ans=0.125 2024-09-23 17:25:19,340 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=308728.0, ans=0.0 2024-09-23 17:25:45,745 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.085e+02 1.280e+02 1.358e+02 1.516e+02 2.710e+02, threshold=2.715e+02, percent-clipped=0.0 2024-09-23 17:25:53,065 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.42 vs. limit=15.0 2024-09-23 17:25:53,944 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=308821.3333333333, ans=0.125 2024-09-23 17:25:57,055 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=308821.3333333333, ans=0.0 2024-09-23 17:26:05,959 INFO [train.py:1198] (2/4) Epoch 17, batch 3850, loss[loss=0.198, ctc_loss=0.1307, cr_loss=0.3366, over 16285.00 frames. ], tot_loss[loss=0.2257, ctc_loss=0.1522, cr_loss=0.3672, over 3298410.73 frames. ], batch size: 36, lr: 6.98e-03, grad_scale: 16.0 2024-09-23 17:26:15,262 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=308868.0, ans=0.1 2024-09-23 17:26:18,977 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.47 vs. limit=15.0 2024-09-23 17:26:32,890 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=308914.6666666667, ans=0.0 2024-09-23 17:27:01,349 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=309008.0, ans=0.125 2024-09-23 17:27:01,439 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=309008.0, ans=0.2 2024-09-23 17:27:01,673 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=13.95 vs. limit=15.0 2024-09-23 17:28:06,788 INFO [train.py:1198] (2/4) Epoch 18, batch 0, loss[loss=0.2321, ctc_loss=0.1525, cr_loss=0.3981, over 17289.00 frames. ], tot_loss[loss=0.2321, ctc_loss=0.1525, cr_loss=0.3981, over 17289.00 frames. ], batch size: 42, lr: 6.78e-03, grad_scale: 32.0 2024-09-23 17:28:06,788 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 17:28:21,931 INFO [train.py:1230] (2/4) Epoch 18, validation: loss=0.03994, ctc_loss=0.03994, cr_loss=8.27e-15, over 944034.00 frames. 2024-09-23 17:28:21,932 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 17:28:26,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=309082.6666666667, ans=0.04949747468305833 2024-09-23 17:28:27,065 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=309082.6666666667, ans=0.1 2024-09-23 17:28:43,330 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.97 vs. limit=10.0 2024-09-23 17:28:49,409 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=309129.3333333333, ans=0.0 2024-09-23 17:28:58,740 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=309176.0, ans=0.2 2024-09-23 17:29:09,236 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=3.98 vs. limit=10.0 2024-09-23 17:29:15,505 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=10.96 vs. limit=22.5 2024-09-23 17:29:25,661 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=309222.6666666667, ans=0.125 2024-09-23 17:29:30,230 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.283e+02 1.487e+02 1.642e+02 2.774e+02, threshold=2.974e+02, percent-clipped=1.0 2024-09-23 17:29:44,550 INFO [train.py:1198] (2/4) Epoch 18, batch 50, loss[loss=0.2276, ctc_loss=0.1527, cr_loss=0.3746, over 17134.00 frames. ], tot_loss[loss=0.224, ctc_loss=0.1505, cr_loss=0.3673, over 760342.81 frames. ], batch size: 48, lr: 6.78e-03, grad_scale: 32.0 2024-09-23 17:30:15,930 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=16.22 vs. limit=22.5 2024-09-23 17:30:29,404 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=309409.3333333333, ans=0.125 2024-09-23 17:30:40,612 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=309456.0, ans=0.5 2024-09-23 17:30:56,953 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.98 vs. limit=22.5 2024-09-23 17:31:06,896 INFO [train.py:1198] (2/4) Epoch 18, batch 100, loss[loss=0.2383, ctc_loss=0.1626, cr_loss=0.3786, over 17147.00 frames. ], tot_loss[loss=0.223, ctc_loss=0.1498, cr_loss=0.3663, over 1329220.56 frames. ], batch size: 48, lr: 6.77e-03, grad_scale: 16.0 2024-09-23 17:31:25,616 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.42 vs. limit=15.0 2024-09-23 17:31:28,555 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=18.76 vs. limit=22.5 2024-09-23 17:31:48,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=309642.6666666667, ans=0.0 2024-09-23 17:31:50,205 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=309642.6666666667, ans=0.2 2024-09-23 17:32:09,414 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=309736.0, ans=0.125 2024-09-23 17:32:11,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=309736.0, ans=0.1 2024-09-23 17:32:13,773 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.062e+02 1.252e+02 1.337e+02 1.407e+02 3.310e+02, threshold=2.674e+02, percent-clipped=1.0 2024-09-23 17:32:28,257 INFO [train.py:1198] (2/4) Epoch 18, batch 150, loss[loss=0.1851, ctc_loss=0.1188, cr_loss=0.3316, over 17038.00 frames. ], tot_loss[loss=0.2224, ctc_loss=0.1491, cr_loss=0.3667, over 1786031.79 frames. ], batch size: 39, lr: 6.77e-03, grad_scale: 16.0 2024-09-23 17:32:50,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=309829.3333333333, ans=0.125 2024-09-23 17:32:51,425 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=9.10 vs. limit=15.0 2024-09-23 17:33:19,491 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.20 vs. limit=6.0 2024-09-23 17:33:38,181 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=309969.3333333333, ans=0.0 2024-09-23 17:33:42,910 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=309969.3333333333, ans=0.0 2024-09-23 17:33:50,777 INFO [train.py:1198] (2/4) Epoch 18, batch 200, loss[loss=0.2473, ctc_loss=0.1679, cr_loss=0.3969, over 14807.00 frames. ], tot_loss[loss=0.2211, ctc_loss=0.1483, cr_loss=0.3643, over 2132556.88 frames. ], batch size: 89, lr: 6.77e-03, grad_scale: 16.0 2024-09-23 17:33:54,371 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=310016.0, ans=0.125 2024-09-23 17:33:54,777 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=11.60 vs. limit=22.5 2024-09-23 17:34:27,452 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=310109.3333333333, ans=0.1 2024-09-23 17:34:33,935 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=310109.3333333333, ans=0.0 2024-09-23 17:34:35,673 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=310109.3333333333, ans=0.0 2024-09-23 17:34:46,586 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=310156.0, ans=0.125 2024-09-23 17:35:00,728 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.233e+02 1.340e+02 1.521e+02 2.141e+02, threshold=2.680e+02, percent-clipped=0.0 2024-09-23 17:35:04,367 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 17:35:04,379 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=310202.6666666667, ans=0.125 2024-09-23 17:35:13,460 INFO [train.py:1198] (2/4) Epoch 18, batch 250, loss[loss=0.2089, ctc_loss=0.137, cr_loss=0.3594, over 17075.00 frames. ], tot_loss[loss=0.2226, ctc_loss=0.1495, cr_loss=0.3656, over 2393355.76 frames. ], batch size: 46, lr: 6.77e-03, grad_scale: 16.0 2024-09-23 17:35:41,175 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.80 vs. limit=15.0 2024-09-23 17:35:44,748 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.whiten, num_groups=1, num_channels=192, metric=4.99 vs. limit=12.0 2024-09-23 17:36:02,707 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 17:36:07,929 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.82 vs. limit=10.0 2024-09-23 17:36:15,545 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=310389.3333333333, ans=0.125 2024-09-23 17:36:33,453 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=310436.0, ans=0.125 2024-09-23 17:36:36,379 INFO [train.py:1198] (2/4) Epoch 18, batch 300, loss[loss=0.1845, ctc_loss=0.1243, cr_loss=0.301, over 17017.00 frames. ], tot_loss[loss=0.2227, ctc_loss=0.1495, cr_loss=0.3662, over 2610791.33 frames. ], batch size: 39, lr: 6.76e-03, grad_scale: 16.0 2024-09-23 17:36:50,898 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=310529.3333333333, ans=0.0 2024-09-23 17:37:13,343 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=310576.0, ans=0.2 2024-09-23 17:37:15,383 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.24 vs. limit=12.0 2024-09-23 17:37:28,963 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=310622.6666666667, ans=0.125 2024-09-23 17:37:34,026 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.83 vs. limit=10.0 2024-09-23 17:37:36,843 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=310622.6666666667, ans=0.0 2024-09-23 17:37:46,270 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.058e+02 1.225e+02 1.352e+02 1.557e+02 2.949e+02, threshold=2.705e+02, percent-clipped=1.0 2024-09-23 17:38:00,131 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=310716.0, ans=0.125 2024-09-23 17:38:01,322 INFO [train.py:1198] (2/4) Epoch 18, batch 350, loss[loss=0.2335, ctc_loss=0.1583, cr_loss=0.376, over 17087.00 frames. ], tot_loss[loss=0.2213, ctc_loss=0.1484, cr_loss=0.3642, over 2776630.80 frames. ], batch size: 49, lr: 6.76e-03, grad_scale: 16.0 2024-09-23 17:38:56,264 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=310856.0, ans=0.1 2024-09-23 17:39:09,492 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.57 vs. limit=6.0 2024-09-23 17:39:24,258 INFO [train.py:1198] (2/4) Epoch 18, batch 400, loss[loss=0.2262, ctc_loss=0.1497, cr_loss=0.3827, over 17353.00 frames. ], tot_loss[loss=0.2221, ctc_loss=0.149, cr_loss=0.3655, over 2904331.83 frames. ], batch size: 48, lr: 6.76e-03, grad_scale: 32.0 2024-09-23 17:39:24,632 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=310949.3333333333, ans=0.1 2024-09-23 17:39:26,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=310949.3333333333, ans=0.0 2024-09-23 17:39:45,185 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=310996.0, ans=0.1 2024-09-23 17:39:58,596 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=13.82 vs. limit=15.0 2024-09-23 17:40:13,936 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=311089.3333333333, ans=0.125 2024-09-23 17:40:25,026 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=311089.3333333333, ans=0.1 2024-09-23 17:40:26,718 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=311136.0, ans=0.1 2024-09-23 17:40:31,100 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.269e+02 1.393e+02 1.575e+02 2.470e+02, threshold=2.786e+02, percent-clipped=0.0 2024-09-23 17:40:43,671 INFO [train.py:1198] (2/4) Epoch 18, batch 450, loss[loss=0.2327, ctc_loss=0.1591, cr_loss=0.3682, over 17146.00 frames. ], tot_loss[loss=0.2223, ctc_loss=0.1492, cr_loss=0.3659, over 3008841.95 frames. ], batch size: 48, lr: 6.76e-03, grad_scale: 32.0 2024-09-23 17:41:10,811 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=7.03 vs. limit=15.0 2024-09-23 17:41:14,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=311229.3333333333, ans=0.0 2024-09-23 17:41:29,413 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=311276.0, ans=0.0 2024-09-23 17:41:48,365 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 17:41:56,082 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 17:42:05,309 INFO [train.py:1198] (2/4) Epoch 18, batch 500, loss[loss=0.2098, ctc_loss=0.1361, cr_loss=0.3683, over 17262.00 frames. ], tot_loss[loss=0.2222, ctc_loss=0.1491, cr_loss=0.3654, over 3084307.55 frames. ], batch size: 44, lr: 6.75e-03, grad_scale: 32.0 2024-09-23 17:42:05,999 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=5.25 vs. limit=12.0 2024-09-23 17:42:13,671 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.35 vs. limit=15.0 2024-09-23 17:42:25,866 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=311462.6666666667, ans=0.125 2024-09-23 17:42:49,856 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=19.37 vs. limit=22.5 2024-09-23 17:43:17,645 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.265e+02 1.370e+02 1.572e+02 2.414e+02, threshold=2.740e+02, percent-clipped=0.0 2024-09-23 17:43:24,212 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=311602.6666666667, ans=0.5 2024-09-23 17:43:30,320 INFO [train.py:1198] (2/4) Epoch 18, batch 550, loss[loss=0.2473, ctc_loss=0.1687, cr_loss=0.3933, over 17308.00 frames. ], tot_loss[loss=0.222, ctc_loss=0.149, cr_loss=0.365, over 3134902.23 frames. ], batch size: 49, lr: 6.75e-03, grad_scale: 32.0 2024-09-23 17:44:01,275 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=16.57 vs. limit=22.5 2024-09-23 17:44:18,203 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=311742.6666666667, ans=0.125 2024-09-23 17:44:26,788 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=5.29 vs. limit=12.0 2024-09-23 17:44:35,690 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=311836.0, ans=0.0 2024-09-23 17:44:53,370 INFO [train.py:1198] (2/4) Epoch 18, batch 600, loss[loss=0.2023, ctc_loss=0.134, cr_loss=0.3417, over 17026.00 frames. ], tot_loss[loss=0.2216, ctc_loss=0.1488, cr_loss=0.3642, over 3183566.28 frames. ], batch size: 44, lr: 6.75e-03, grad_scale: 32.0 2024-09-23 17:44:57,086 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=311882.6666666667, ans=0.0 2024-09-23 17:45:03,327 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.max_abs, batch_count=311882.6666666667, ans=10.0 2024-09-23 17:45:13,764 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=5.00 vs. limit=15.0 2024-09-23 17:45:43,270 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=312022.6666666667, ans=0.0 2024-09-23 17:46:00,294 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=312069.3333333333, ans=0.0 2024-09-23 17:46:04,639 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.142e+02 1.282e+02 1.384e+02 1.538e+02 2.458e+02, threshold=2.768e+02, percent-clipped=0.0 2024-09-23 17:46:12,849 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=312069.3333333333, ans=0.2 2024-09-23 17:46:16,003 INFO [train.py:1198] (2/4) Epoch 18, batch 650, loss[loss=0.2156, ctc_loss=0.1434, cr_loss=0.3607, over 17081.00 frames. ], tot_loss[loss=0.222, ctc_loss=0.1491, cr_loss=0.3646, over 3221778.09 frames. ], batch size: 46, lr: 6.75e-03, grad_scale: 16.0 2024-09-23 17:46:26,686 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=8.23 vs. limit=15.0 2024-09-23 17:47:22,335 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.76 vs. limit=15.0 2024-09-23 17:47:38,725 INFO [train.py:1198] (2/4) Epoch 18, batch 700, loss[loss=0.2542, ctc_loss=0.1733, cr_loss=0.4047, over 17227.00 frames. ], tot_loss[loss=0.2219, ctc_loss=0.1491, cr_loss=0.3641, over 3244942.29 frames. ], batch size: 50, lr: 6.74e-03, grad_scale: 16.0 2024-09-23 17:48:49,978 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.087e+02 1.252e+02 1.369e+02 1.593e+02 2.409e+02, threshold=2.738e+02, percent-clipped=0.0 2024-09-23 17:49:00,254 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=5.69 vs. limit=12.0 2024-09-23 17:49:00,933 INFO [train.py:1198] (2/4) Epoch 18, batch 750, loss[loss=0.205, ctc_loss=0.1365, cr_loss=0.3428, over 17161.00 frames. ], tot_loss[loss=0.2215, ctc_loss=0.1488, cr_loss=0.3638, over 3275043.17 frames. ], batch size: 45, lr: 6.74e-03, grad_scale: 16.0 2024-09-23 17:49:04,268 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=312582.6666666667, ans=0.125 2024-09-23 17:49:13,232 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=312582.6666666667, ans=0.125 2024-09-23 17:49:14,901 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=312582.6666666667, ans=0.1 2024-09-23 17:49:30,688 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=312629.3333333333, ans=0.125 2024-09-23 17:49:34,958 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=10.66 vs. limit=15.0 2024-09-23 17:50:09,936 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.73 vs. limit=15.0 2024-09-23 17:50:10,950 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=312769.3333333333, ans=0.2 2024-09-23 17:50:23,307 INFO [train.py:1198] (2/4) Epoch 18, batch 800, loss[loss=0.2335, ctc_loss=0.1566, cr_loss=0.3843, over 17180.00 frames. ], tot_loss[loss=0.221, ctc_loss=0.1484, cr_loss=0.3629, over 3291242.60 frames. ], batch size: 45, lr: 6.74e-03, grad_scale: 32.0 2024-09-23 17:51:17,750 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.62 vs. limit=6.0 2024-09-23 17:51:26,698 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=312956.0, ans=0.025 2024-09-23 17:51:34,556 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.074e+02 1.268e+02 1.368e+02 1.483e+02 2.318e+02, threshold=2.737e+02, percent-clipped=0.0 2024-09-23 17:51:45,655 INFO [train.py:1198] (2/4) Epoch 18, batch 850, loss[loss=0.209, ctc_loss=0.138, cr_loss=0.355, over 16997.00 frames. ], tot_loss[loss=0.2212, ctc_loss=0.1485, cr_loss=0.3635, over 3310859.50 frames. ], batch size: 51, lr: 6.74e-03, grad_scale: 32.0 2024-09-23 17:52:16,888 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=313096.0, ans=0.0 2024-09-23 17:52:29,730 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=313142.6666666667, ans=0.025 2024-09-23 17:53:10,236 INFO [train.py:1198] (2/4) Epoch 18, batch 900, loss[loss=0.2378, ctc_loss=0.1699, cr_loss=0.3397, over 11628.00 frames. ], tot_loss[loss=0.2213, ctc_loss=0.1486, cr_loss=0.3634, over 3320156.85 frames. ], batch size: 123, lr: 6.73e-03, grad_scale: 32.0 2024-09-23 17:53:43,871 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=313376.0, ans=0.0 2024-09-23 17:54:09,491 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.min_positive, batch_count=313422.6666666667, ans=0.05 2024-09-23 17:54:09,546 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=313422.6666666667, ans=0.125 2024-09-23 17:54:21,425 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.286e+02 1.405e+02 1.600e+02 2.669e+02, threshold=2.810e+02, percent-clipped=0.0 2024-09-23 17:54:22,633 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module1.whiten, num_groups=1, num_channels=192, metric=7.97 vs. limit=15.0 2024-09-23 17:54:32,762 INFO [train.py:1198] (2/4) Epoch 18, batch 950, loss[loss=0.1736, ctc_loss=0.1125, cr_loss=0.3058, over 17068.00 frames. ], tot_loss[loss=0.2221, ctc_loss=0.1492, cr_loss=0.3647, over 3331607.69 frames. ], batch size: 39, lr: 6.73e-03, grad_scale: 32.0 2024-09-23 17:54:53,808 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=313562.6666666667, ans=0.125 2024-09-23 17:55:00,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=313562.6666666667, ans=0.0 2024-09-23 17:55:00,789 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module1.whiten, num_groups=1, num_channels=512, metric=9.82 vs. limit=15.0 2024-09-23 17:55:42,317 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=313702.6666666667, ans=0.2 2024-09-23 17:55:51,281 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=313702.6666666667, ans=0.0 2024-09-23 17:55:55,788 INFO [train.py:1198] (2/4) Epoch 18, batch 1000, loss[loss=0.2299, ctc_loss=0.1546, cr_loss=0.3765, over 16720.00 frames. ], tot_loss[loss=0.2207, ctc_loss=0.1481, cr_loss=0.363, over 3339126.36 frames. ], batch size: 61, lr: 6.73e-03, grad_scale: 32.0 2024-09-23 17:55:56,039 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=313749.3333333333, ans=0.125 2024-09-23 17:56:47,396 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.32 vs. limit=15.0 2024-09-23 17:56:53,374 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=313889.3333333333, ans=0.125 2024-09-23 17:57:06,775 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.066e+02 1.230e+02 1.328e+02 1.431e+02 2.241e+02, threshold=2.657e+02, percent-clipped=0.0 2024-09-23 17:57:13,312 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=313936.0, ans=0.0 2024-09-23 17:57:14,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=313936.0, ans=0.0 2024-09-23 17:57:17,824 INFO [train.py:1198] (2/4) Epoch 18, batch 1050, loss[loss=0.2375, ctc_loss=0.1603, cr_loss=0.3861, over 16812.00 frames. ], tot_loss[loss=0.2196, ctc_loss=0.1472, cr_loss=0.3619, over 3352576.05 frames. ], batch size: 58, lr: 6.73e-03, grad_scale: 32.0 2024-09-23 17:57:30,858 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=313982.6666666667, ans=0.1 2024-09-23 17:57:30,864 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=313982.6666666667, ans=0.2 2024-09-23 17:58:19,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=314122.6666666667, ans=0.125 2024-09-23 17:58:37,795 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.80 vs. limit=15.0 2024-09-23 17:58:40,062 INFO [train.py:1198] (2/4) Epoch 18, batch 1100, loss[loss=0.2196, ctc_loss=0.1459, cr_loss=0.3686, over 17064.00 frames. ], tot_loss[loss=0.2199, ctc_loss=0.1474, cr_loss=0.3623, over 3352500.21 frames. ], batch size: 46, lr: 6.72e-03, grad_scale: 32.0 2024-09-23 17:58:48,213 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=314216.0, ans=0.1 2024-09-23 17:58:49,912 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=314216.0, ans=0.1 2024-09-23 17:58:55,004 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.49 vs. limit=15.0 2024-09-23 17:59:33,936 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=314356.0, ans=0.0 2024-09-23 17:59:49,889 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=314402.6666666667, ans=0.125 2024-09-23 17:59:51,144 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.280e+02 1.383e+02 1.491e+02 2.284e+02, threshold=2.766e+02, percent-clipped=0.0 2024-09-23 18:00:02,377 INFO [train.py:1198] (2/4) Epoch 18, batch 1150, loss[loss=0.1974, ctc_loss=0.1308, cr_loss=0.3331, over 16269.00 frames. ], tot_loss[loss=0.2197, ctc_loss=0.1473, cr_loss=0.3619, over 3358232.33 frames. ], batch size: 36, lr: 6.72e-03, grad_scale: 32.0 2024-09-23 18:00:23,842 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.min_abs, batch_count=314496.0, ans=0.5 2024-09-23 18:00:32,219 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.26 vs. limit=15.0 2024-09-23 18:01:25,141 INFO [train.py:1198] (2/4) Epoch 18, batch 1200, loss[loss=0.2342, ctc_loss=0.1579, cr_loss=0.3816, over 17109.00 frames. ], tot_loss[loss=0.2197, ctc_loss=0.1475, cr_loss=0.3614, over 3348211.36 frames. ], batch size: 49, lr: 6.72e-03, grad_scale: 32.0 2024-09-23 18:02:05,393 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=12.12 vs. limit=22.5 2024-09-23 18:02:09,374 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=314776.0, ans=0.125 2024-09-23 18:02:38,964 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.109e+02 1.291e+02 1.389e+02 1.552e+02 2.070e+02, threshold=2.777e+02, percent-clipped=0.0 2024-09-23 18:02:49,967 INFO [train.py:1198] (2/4) Epoch 18, batch 1250, loss[loss=0.227, ctc_loss=0.1553, cr_loss=0.3587, over 16925.00 frames. ], tot_loss[loss=0.2193, ctc_loss=0.1472, cr_loss=0.3602, over 3346857.54 frames. ], batch size: 58, lr: 6.72e-03, grad_scale: 32.0 2024-09-23 18:02:51,828 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=314916.0, ans=0.2 2024-09-23 18:03:06,362 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=314962.6666666667, ans=0.125 2024-09-23 18:03:08,399 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.29 vs. limit=15.0 2024-09-23 18:03:27,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=315009.3333333333, ans=0.0 2024-09-23 18:03:31,145 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.16 vs. limit=22.5 2024-09-23 18:03:43,151 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=315056.0, ans=0.125 2024-09-23 18:03:47,891 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=315056.0, ans=0.0 2024-09-23 18:03:48,057 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 18:03:57,484 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=315102.6666666667, ans=0.0 2024-09-23 18:03:59,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=315102.6666666667, ans=0.2 2024-09-23 18:04:10,278 INFO [train.py:1198] (2/4) Epoch 18, batch 1300, loss[loss=0.2421, ctc_loss=0.1609, cr_loss=0.4061, over 17353.00 frames. ], tot_loss[loss=0.2196, ctc_loss=0.1473, cr_loss=0.3613, over 3357572.63 frames. ], batch size: 48, lr: 6.71e-03, grad_scale: 32.0 2024-09-23 18:04:17,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=315149.3333333333, ans=0.125 2024-09-23 18:04:48,141 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=315242.6666666667, ans=0.0 2024-09-23 18:05:02,706 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=315289.3333333333, ans=0.125 2024-09-23 18:05:04,164 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=315289.3333333333, ans=0.2 2024-09-23 18:05:09,095 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.max_abs, batch_count=315289.3333333333, ans=10.0 2024-09-23 18:05:10,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=315289.3333333333, ans=0.0 2024-09-23 18:05:10,756 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=315289.3333333333, ans=0.025 2024-09-23 18:05:21,639 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.126e+02 1.312e+02 1.416e+02 1.661e+02 2.501e+02, threshold=2.832e+02, percent-clipped=0.0 2024-09-23 18:05:32,770 INFO [train.py:1198] (2/4) Epoch 18, batch 1350, loss[loss=0.2347, ctc_loss=0.1569, cr_loss=0.389, over 17144.00 frames. ], tot_loss[loss=0.2191, ctc_loss=0.1469, cr_loss=0.3612, over 3369967.71 frames. ], batch size: 48, lr: 6.71e-03, grad_scale: 32.0 2024-09-23 18:05:57,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=315429.3333333333, ans=0.0 2024-09-23 18:05:57,717 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=315429.3333333333, ans=0.125 2024-09-23 18:06:05,921 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=315476.0, ans=0.1 2024-09-23 18:06:18,521 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.min_positive, batch_count=315476.0, ans=0.025 2024-09-23 18:06:28,196 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=315522.6666666667, ans=0.1 2024-09-23 18:06:29,769 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=315522.6666666667, ans=0.0 2024-09-23 18:06:45,621 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=315569.3333333333, ans=0.1 2024-09-23 18:06:54,861 INFO [train.py:1198] (2/4) Epoch 18, batch 1400, loss[loss=0.2374, ctc_loss=0.1567, cr_loss=0.4035, over 16991.00 frames. ], tot_loss[loss=0.2201, ctc_loss=0.1476, cr_loss=0.3626, over 3371164.65 frames. ], batch size: 53, lr: 6.71e-03, grad_scale: 32.0 2024-09-23 18:06:55,139 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=315616.0, ans=0.1 2024-09-23 18:06:58,478 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=315616.0, ans=0.05 2024-09-23 18:07:04,370 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=315616.0, ans=0.025 2024-09-23 18:07:43,368 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=315709.3333333333, ans=0.125 2024-09-23 18:08:04,755 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=4.50 vs. limit=12.0 2024-09-23 18:08:08,561 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.069e+02 1.236e+02 1.313e+02 1.450e+02 2.376e+02, threshold=2.626e+02, percent-clipped=0.0 2024-09-23 18:08:12,386 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.65 vs. limit=15.0 2024-09-23 18:08:19,845 INFO [train.py:1198] (2/4) Epoch 18, batch 1450, loss[loss=0.2302, ctc_loss=0.1572, cr_loss=0.3649, over 17258.00 frames. ], tot_loss[loss=0.2195, ctc_loss=0.1471, cr_loss=0.3623, over 3368983.31 frames. ], batch size: 44, lr: 6.71e-03, grad_scale: 32.0 2024-09-23 18:09:21,921 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=315989.3333333333, ans=0.05 2024-09-23 18:09:33,260 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=13.62 vs. limit=15.0 2024-09-23 18:09:40,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=316082.6666666667, ans=0.125 2024-09-23 18:09:42,228 INFO [train.py:1198] (2/4) Epoch 18, batch 1500, loss[loss=0.244, ctc_loss=0.1646, cr_loss=0.397, over 16031.00 frames. ], tot_loss[loss=0.2201, ctc_loss=0.1474, cr_loss=0.3632, over 3369094.42 frames. ], batch size: 74, lr: 6.70e-03, grad_scale: 32.0 2024-09-23 18:09:50,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=316082.6666666667, ans=0.2 2024-09-23 18:09:53,950 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=316082.6666666667, ans=0.0 2024-09-23 18:10:22,599 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=316176.0, ans=0.125 2024-09-23 18:10:40,380 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=316222.6666666667, ans=0.125 2024-09-23 18:10:53,734 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.112e+02 1.296e+02 1.405e+02 1.565e+02 2.957e+02, threshold=2.810e+02, percent-clipped=1.0 2024-09-23 18:10:54,034 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=316269.3333333333, ans=0.025 2024-09-23 18:10:54,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=316269.3333333333, ans=0.125 2024-09-23 18:10:59,151 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.min_positive, batch_count=316269.3333333333, ans=0.025 2024-09-23 18:11:05,019 INFO [train.py:1198] (2/4) Epoch 18, batch 1550, loss[loss=0.2192, ctc_loss=0.1477, cr_loss=0.3571, over 17296.00 frames. ], tot_loss[loss=0.2195, ctc_loss=0.147, cr_loss=0.3624, over 3369482.14 frames. ], batch size: 46, lr: 6.70e-03, grad_scale: 32.0 2024-09-23 18:11:18,027 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=316316.0, ans=0.1 2024-09-23 18:11:35,592 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=316409.3333333333, ans=0.1 2024-09-23 18:12:02,520 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=316456.0, ans=0.125 2024-09-23 18:12:06,574 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=19.35 vs. limit=22.5 2024-09-23 18:12:17,191 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=316502.6666666667, ans=0.125 2024-09-23 18:12:28,051 INFO [train.py:1198] (2/4) Epoch 18, batch 1600, loss[loss=0.2128, ctc_loss=0.1427, cr_loss=0.3505, over 16983.00 frames. ], tot_loss[loss=0.2188, ctc_loss=0.1465, cr_loss=0.3614, over 3376558.95 frames. ], batch size: 53, lr: 6.70e-03, grad_scale: 32.0 2024-09-23 18:12:45,421 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=316596.0, ans=0.125 2024-09-23 18:12:47,428 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=6.13 vs. limit=12.0 2024-09-23 18:12:51,554 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=316596.0, ans=0.1 2024-09-23 18:13:07,749 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=316642.6666666667, ans=0.09899494936611666 2024-09-23 18:13:10,962 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=316642.6666666667, ans=0.0 2024-09-23 18:13:38,090 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.min_abs, batch_count=316736.0, ans=0.5 2024-09-23 18:13:39,405 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.138e+02 1.280e+02 1.386e+02 1.539e+02 3.392e+02, threshold=2.772e+02, percent-clipped=1.0 2024-09-23 18:13:50,542 INFO [train.py:1198] (2/4) Epoch 18, batch 1650, loss[loss=0.2313, ctc_loss=0.1558, cr_loss=0.3776, over 17299.00 frames. ], tot_loss[loss=0.2202, ctc_loss=0.1477, cr_loss=0.3624, over 3370163.60 frames. ], batch size: 46, lr: 6.70e-03, grad_scale: 32.0 2024-09-23 18:14:01,989 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=316782.6666666667, ans=0.1 2024-09-23 18:14:14,276 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 18:14:17,366 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=316829.3333333333, ans=0.125 2024-09-23 18:14:36,493 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=316876.0, ans=0.125 2024-09-23 18:14:36,662 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=316876.0, ans=0.04949747468305833 2024-09-23 18:14:39,906 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=316922.6666666667, ans=0.1 2024-09-23 18:14:54,145 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=316922.6666666667, ans=0.0 2024-09-23 18:15:12,112 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=317016.0, ans=0.0 2024-09-23 18:15:13,362 INFO [train.py:1198] (2/4) Epoch 18, batch 1700, loss[loss=0.23, ctc_loss=0.1552, cr_loss=0.3739, over 17303.00 frames. ], tot_loss[loss=0.2201, ctc_loss=0.1477, cr_loss=0.3623, over 3364645.73 frames. ], batch size: 46, lr: 6.69e-03, grad_scale: 32.0 2024-09-23 18:15:34,207 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=317062.6666666667, ans=0.125 2024-09-23 18:15:45,253 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.84 vs. limit=6.0 2024-09-23 18:15:52,567 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=317109.3333333333, ans=0.125 2024-09-23 18:15:55,809 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=317109.3333333333, ans=0.025 2024-09-23 18:15:55,836 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=317109.3333333333, ans=0.1 2024-09-23 18:16:10,343 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=317156.0, ans=0.2 2024-09-23 18:16:11,931 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=317156.0, ans=0.0 2024-09-23 18:16:24,388 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.272e+02 1.364e+02 1.549e+02 1.950e+02, threshold=2.728e+02, percent-clipped=0.0 2024-09-23 18:16:35,522 INFO [train.py:1198] (2/4) Epoch 18, batch 1750, loss[loss=0.2604, ctc_loss=0.1764, cr_loss=0.4199, over 16607.00 frames. ], tot_loss[loss=0.2205, ctc_loss=0.1479, cr_loss=0.3629, over 3368845.65 frames. ], batch size: 66, lr: 6.69e-03, grad_scale: 32.0 2024-09-23 18:17:44,219 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=317389.3333333333, ans=0.125 2024-09-23 18:17:44,324 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=317389.3333333333, ans=0.125 2024-09-23 18:17:52,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=317436.0, ans=0.0 2024-09-23 18:17:55,419 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=317436.0, ans=0.125 2024-09-23 18:18:02,969 INFO [train.py:1198] (2/4) Epoch 18, batch 1800, loss[loss=0.2193, ctc_loss=0.1506, cr_loss=0.3436, over 16064.00 frames. ], tot_loss[loss=0.22, ctc_loss=0.1475, cr_loss=0.3624, over 3368551.09 frames. ], batch size: 74, lr: 6.69e-03, grad_scale: 32.0 2024-09-23 18:18:37,394 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=317576.0, ans=0.1 2024-09-23 18:18:50,298 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=317622.6666666667, ans=0.125 2024-09-23 18:18:51,950 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=317622.6666666667, ans=0.1 2024-09-23 18:19:07,936 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=317669.3333333333, ans=0.0 2024-09-23 18:19:14,905 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.089e+02 1.243e+02 1.320e+02 1.455e+02 2.029e+02, threshold=2.640e+02, percent-clipped=0.0 2024-09-23 18:19:26,146 INFO [train.py:1198] (2/4) Epoch 18, batch 1850, loss[loss=0.19, ctc_loss=0.1233, cr_loss=0.3338, over 17019.00 frames. ], tot_loss[loss=0.2199, ctc_loss=0.1475, cr_loss=0.3622, over 3361517.64 frames. ], batch size: 44, lr: 6.69e-03, grad_scale: 32.0 2024-09-23 18:19:47,649 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=317762.6666666667, ans=0.1 2024-09-23 18:20:05,060 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=317809.3333333333, ans=0.1 2024-09-23 18:20:12,204 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=18.85 vs. limit=22.5 2024-09-23 18:20:49,117 INFO [train.py:1198] (2/4) Epoch 18, batch 1900, loss[loss=0.2263, ctc_loss=0.1537, cr_loss=0.3626, over 17295.00 frames. ], tot_loss[loss=0.2212, ctc_loss=0.1484, cr_loss=0.3641, over 3357037.88 frames. ], batch size: 49, lr: 6.68e-03, grad_scale: 32.0 2024-09-23 18:21:18,850 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=317996.0, ans=0.125 2024-09-23 18:21:29,871 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=318042.6666666667, ans=0.025 2024-09-23 18:21:58,144 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.235e+02 1.336e+02 1.423e+02 1.924e+02, threshold=2.671e+02, percent-clipped=0.0 2024-09-23 18:22:05,671 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=318136.0, ans=0.125 2024-09-23 18:22:11,954 INFO [train.py:1198] (2/4) Epoch 18, batch 1950, loss[loss=0.2164, ctc_loss=0.1446, cr_loss=0.3589, over 17010.00 frames. ], tot_loss[loss=0.2198, ctc_loss=0.1474, cr_loss=0.3622, over 3356598.95 frames. ], batch size: 53, lr: 6.68e-03, grad_scale: 32.0 2024-09-23 18:22:26,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=318229.3333333333, ans=0.125 2024-09-23 18:22:28,788 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=11.59 vs. limit=15.0 2024-09-23 18:22:37,814 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=318229.3333333333, ans=0.0 2024-09-23 18:22:43,952 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=318229.3333333333, ans=0.0 2024-09-23 18:23:17,060 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=318369.3333333333, ans=0.125 2024-09-23 18:23:22,367 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=10.93 vs. limit=15.0 2024-09-23 18:23:34,523 INFO [train.py:1198] (2/4) Epoch 18, batch 2000, loss[loss=0.2006, ctc_loss=0.1321, cr_loss=0.3426, over 17099.00 frames. ], tot_loss[loss=0.2192, ctc_loss=0.1468, cr_loss=0.3619, over 3366743.58 frames. ], batch size: 49, lr: 6.68e-03, grad_scale: 32.0 2024-09-23 18:24:05,906 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.19 vs. limit=6.0 2024-09-23 18:24:13,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=318509.3333333333, ans=0.125 2024-09-23 18:24:32,234 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=318556.0, ans=0.125 2024-09-23 18:24:46,013 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.121e+02 1.249e+02 1.342e+02 1.462e+02 2.296e+02, threshold=2.684e+02, percent-clipped=0.0 2024-09-23 18:24:46,278 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=318602.6666666667, ans=0.1 2024-09-23 18:24:52,656 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=318602.6666666667, ans=0.1 2024-09-23 18:24:57,327 INFO [train.py:1198] (2/4) Epoch 18, batch 2050, loss[loss=0.1908, ctc_loss=0.1273, cr_loss=0.3173, over 17289.00 frames. ], tot_loss[loss=0.2197, ctc_loss=0.1472, cr_loss=0.3626, over 3366966.33 frames. ], batch size: 42, lr: 6.68e-03, grad_scale: 32.0 2024-09-23 18:26:19,581 INFO [train.py:1198] (2/4) Epoch 18, batch 2100, loss[loss=0.2003, ctc_loss=0.1347, cr_loss=0.3281, over 17018.00 frames. ], tot_loss[loss=0.2191, ctc_loss=0.1468, cr_loss=0.3617, over 3366209.54 frames. ], batch size: 44, lr: 6.67e-03, grad_scale: 32.0 2024-09-23 18:26:32,806 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=318882.6666666667, ans=0.125 2024-09-23 18:27:11,996 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=319022.6666666667, ans=0.0 2024-09-23 18:27:31,218 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.263e+02 1.339e+02 1.437e+02 2.082e+02, threshold=2.677e+02, percent-clipped=0.0 2024-09-23 18:27:34,556 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 18:27:44,921 INFO [train.py:1198] (2/4) Epoch 18, batch 2150, loss[loss=0.2458, ctc_loss=0.1651, cr_loss=0.4034, over 17252.00 frames. ], tot_loss[loss=0.2185, ctc_loss=0.1462, cr_loss=0.3615, over 3372669.85 frames. ], batch size: 55, lr: 6.67e-03, grad_scale: 32.0 2024-09-23 18:27:54,936 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=319116.0, ans=0.125 2024-09-23 18:28:19,166 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer_na.min_abs, batch_count=319209.3333333333, ans=0.02 2024-09-23 18:28:25,596 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=319209.3333333333, ans=0.125 2024-09-23 18:28:48,119 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.03 vs. limit=15.0 2024-09-23 18:29:02,481 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=319302.6666666667, ans=0.125 2024-09-23 18:29:05,237 INFO [train.py:1198] (2/4) Epoch 18, batch 2200, loss[loss=0.2607, ctc_loss=0.176, cr_loss=0.4235, over 16531.00 frames. ], tot_loss[loss=0.2187, ctc_loss=0.1463, cr_loss=0.362, over 3379939.00 frames. ], batch size: 66, lr: 6.67e-03, grad_scale: 16.0 2024-09-23 18:29:14,033 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=3.75 vs. limit=15.0 2024-09-23 18:29:36,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=319396.0, ans=0.2 2024-09-23 18:29:41,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=319442.6666666667, ans=0.0 2024-09-23 18:29:53,601 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.93 vs. limit=6.0 2024-09-23 18:30:00,999 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=3.16 vs. limit=6.0 2024-09-23 18:30:02,189 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=319489.3333333333, ans=0.0 2024-09-23 18:30:08,800 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=11.43 vs. limit=22.5 2024-09-23 18:30:17,667 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.053e+02 1.244e+02 1.317e+02 1.469e+02 2.029e+02, threshold=2.633e+02, percent-clipped=0.0 2024-09-23 18:30:27,417 INFO [train.py:1198] (2/4) Epoch 18, batch 2250, loss[loss=0.2704, ctc_loss=0.1906, cr_loss=0.3988, over 11828.00 frames. ], tot_loss[loss=0.2197, ctc_loss=0.1471, cr_loss=0.3631, over 3368426.22 frames. ], batch size: 123, lr: 6.67e-03, grad_scale: 16.0 2024-09-23 18:30:34,217 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=319582.6666666667, ans=0.125 2024-09-23 18:31:45,693 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=319769.3333333333, ans=0.125 2024-09-23 18:31:50,360 INFO [train.py:1198] (2/4) Epoch 18, batch 2300, loss[loss=0.2391, ctc_loss=0.1582, cr_loss=0.4046, over 17007.00 frames. ], tot_loss[loss=0.2192, ctc_loss=0.1467, cr_loss=0.3625, over 3365214.56 frames. ], batch size: 53, lr: 6.67e-03, grad_scale: 16.0 2024-09-23 18:31:53,916 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 18:32:26,956 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=319909.3333333333, ans=0.0 2024-09-23 18:32:33,336 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=319909.3333333333, ans=0.1 2024-09-23 18:32:51,647 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=319956.0, ans=0.125 2024-09-23 18:33:05,707 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.162e+02 1.272e+02 1.367e+02 1.553e+02 2.225e+02, threshold=2.733e+02, percent-clipped=0.0 2024-09-23 18:33:15,256 INFO [train.py:1198] (2/4) Epoch 18, batch 2350, loss[loss=0.2214, ctc_loss=0.1509, cr_loss=0.3527, over 17005.00 frames. ], tot_loss[loss=0.2202, ctc_loss=0.1474, cr_loss=0.364, over 3363510.14 frames. ], batch size: 51, lr: 6.66e-03, grad_scale: 16.0 2024-09-23 18:33:18,675 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=320049.3333333333, ans=0.1 2024-09-23 18:33:36,358 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=320096.0, ans=0.125 2024-09-23 18:33:42,716 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=320096.0, ans=0.1 2024-09-23 18:33:48,046 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 18:34:00,683 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=320142.6666666667, ans=0.1 2024-09-23 18:34:17,890 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=7.01 vs. limit=15.0 2024-09-23 18:34:27,445 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=11.11 vs. limit=15.0 2024-09-23 18:34:37,842 INFO [train.py:1198] (2/4) Epoch 18, batch 2400, loss[loss=0.2027, ctc_loss=0.1328, cr_loss=0.3493, over 16342.00 frames. ], tot_loss[loss=0.2214, ctc_loss=0.1484, cr_loss=0.3649, over 3355210.48 frames. ], batch size: 36, lr: 6.66e-03, grad_scale: 32.0 2024-09-23 18:34:43,015 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=320282.6666666667, ans=0.04949747468305833 2024-09-23 18:34:51,318 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=9.67 vs. limit=15.0 2024-09-23 18:35:32,698 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=320422.6666666667, ans=0.07 2024-09-23 18:35:41,475 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn2.whiten, num_groups=1, num_channels=192, metric=8.99 vs. limit=22.5 2024-09-23 18:35:50,936 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.059e+02 1.273e+02 1.349e+02 1.474e+02 2.344e+02, threshold=2.697e+02, percent-clipped=0.0 2024-09-23 18:35:59,175 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=320516.0, ans=0.1 2024-09-23 18:36:00,511 INFO [train.py:1198] (2/4) Epoch 18, batch 2450, loss[loss=0.2464, ctc_loss=0.1684, cr_loss=0.3896, over 16902.00 frames. ], tot_loss[loss=0.2212, ctc_loss=0.1483, cr_loss=0.3646, over 3350172.58 frames. ], batch size: 58, lr: 6.66e-03, grad_scale: 32.0 2024-09-23 18:36:28,048 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=320562.6666666667, ans=0.0 2024-09-23 18:36:47,409 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=320656.0, ans=0.07 2024-09-23 18:36:57,709 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=320656.0, ans=0.125 2024-09-23 18:37:05,603 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=320702.6666666667, ans=0.125 2024-09-23 18:37:11,822 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=320702.6666666667, ans=0.0 2024-09-23 18:37:22,752 INFO [train.py:1198] (2/4) Epoch 18, batch 2500, loss[loss=0.2377, ctc_loss=0.1633, cr_loss=0.3721, over 16999.00 frames. ], tot_loss[loss=0.2198, ctc_loss=0.1472, cr_loss=0.3627, over 3361327.75 frames. ], batch size: 53, lr: 6.66e-03, grad_scale: 32.0 2024-09-23 18:37:49,646 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=320796.0, ans=0.125 2024-09-23 18:37:56,181 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=320842.6666666667, ans=0.1 2024-09-23 18:38:08,790 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=320842.6666666667, ans=0.0 2024-09-23 18:38:12,094 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=320889.3333333333, ans=0.0 2024-09-23 18:38:16,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=320889.3333333333, ans=0.125 2024-09-23 18:38:26,718 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 18:38:37,301 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.301e+02 1.381e+02 1.501e+02 2.403e+02, threshold=2.762e+02, percent-clipped=0.0 2024-09-23 18:38:42,919 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=512, metric=2.65 vs. limit=15.0 2024-09-23 18:38:45,213 INFO [train.py:1198] (2/4) Epoch 18, batch 2550, loss[loss=0.225, ctc_loss=0.1519, cr_loss=0.3655, over 17019.00 frames. ], tot_loss[loss=0.2197, ctc_loss=0.1471, cr_loss=0.3629, over 3366683.62 frames. ], batch size: 44, lr: 6.65e-03, grad_scale: 16.0 2024-09-23 18:39:26,891 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=321076.0, ans=0.0 2024-09-23 18:39:27,550 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.97 vs. limit=10.0 2024-09-23 18:39:49,201 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=321122.6666666667, ans=0.125 2024-09-23 18:39:58,971 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.07 vs. limit=15.0 2024-09-23 18:40:07,838 INFO [train.py:1198] (2/4) Epoch 18, batch 2600, loss[loss=0.1908, ctc_loss=0.1286, cr_loss=0.3111, over 17119.00 frames. ], tot_loss[loss=0.2193, ctc_loss=0.1467, cr_loss=0.3627, over 3373006.79 frames. ], batch size: 40, lr: 6.65e-03, grad_scale: 16.0 2024-09-23 18:40:10,532 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=6.69 vs. limit=15.0 2024-09-23 18:40:12,249 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.whiten, num_groups=1, num_channels=192, metric=4.85 vs. limit=12.0 2024-09-23 18:40:35,484 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.75 vs. limit=5.0 2024-09-23 18:41:22,320 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.312e+02 1.421e+02 1.594e+02 2.386e+02, threshold=2.841e+02, percent-clipped=0.0 2024-09-23 18:41:30,434 INFO [train.py:1198] (2/4) Epoch 18, batch 2650, loss[loss=0.1948, ctc_loss=0.1297, cr_loss=0.3255, over 17292.00 frames. ], tot_loss[loss=0.2203, ctc_loss=0.1476, cr_loss=0.3636, over 3375557.32 frames. ], batch size: 46, lr: 6.65e-03, grad_scale: 16.0 2024-09-23 18:41:53,152 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=13.48 vs. limit=22.5 2024-09-23 18:42:33,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=321589.3333333333, ans=0.125 2024-09-23 18:42:36,521 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=13.39 vs. limit=15.0 2024-09-23 18:42:49,826 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=11.59 vs. limit=15.0 2024-09-23 18:42:55,424 INFO [train.py:1198] (2/4) Epoch 18, batch 2700, loss[loss=0.1985, ctc_loss=0.1328, cr_loss=0.3282, over 17091.00 frames. ], tot_loss[loss=0.2211, ctc_loss=0.1482, cr_loss=0.3643, over 3359516.97 frames. ], batch size: 43, lr: 6.65e-03, grad_scale: 16.0 2024-09-23 18:42:58,262 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=5.03 vs. limit=5.0 2024-09-23 18:43:30,670 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=321776.0, ans=0.0 2024-09-23 18:44:02,344 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=321869.3333333333, ans=0.05 2024-09-23 18:44:09,951 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.074e+02 1.267e+02 1.364e+02 1.525e+02 2.538e+02, threshold=2.728e+02, percent-clipped=0.0 2024-09-23 18:44:17,772 INFO [train.py:1198] (2/4) Epoch 18, batch 2750, loss[loss=0.2307, ctc_loss=0.1592, cr_loss=0.3576, over 16751.00 frames. ], tot_loss[loss=0.2223, ctc_loss=0.1493, cr_loss=0.3652, over 3348776.74 frames. ], batch size: 61, lr: 6.64e-03, grad_scale: 16.0 2024-09-23 18:44:24,733 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=321916.0, ans=0.0 2024-09-23 18:44:37,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=321962.6666666667, ans=0.125 2024-09-23 18:44:53,421 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=322009.3333333333, ans=0.125 2024-09-23 18:44:54,998 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=322009.3333333333, ans=0.2 2024-09-23 18:45:11,090 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=322056.0, ans=0.125 2024-09-23 18:45:11,535 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.28 vs. limit=12.0 2024-09-23 18:45:29,902 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=14.18 vs. limit=22.5 2024-09-23 18:45:40,662 INFO [train.py:1198] (2/4) Epoch 18, batch 2800, loss[loss=0.2579, ctc_loss=0.1769, cr_loss=0.4049, over 16886.00 frames. ], tot_loss[loss=0.2231, ctc_loss=0.1498, cr_loss=0.3665, over 3346225.48 frames. ], batch size: 58, lr: 6.64e-03, grad_scale: 32.0 2024-09-23 18:45:47,342 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=322149.3333333333, ans=0.2 2024-09-23 18:46:01,702 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=322196.0, ans=0.1 2024-09-23 18:46:04,715 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=322196.0, ans=0.125 2024-09-23 18:46:10,216 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=11.72 vs. limit=15.0 2024-09-23 18:46:49,560 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=15.69 vs. limit=22.5 2024-09-23 18:46:55,162 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.235e+02 1.345e+02 1.484e+02 2.490e+02, threshold=2.689e+02, percent-clipped=0.0 2024-09-23 18:47:03,239 INFO [train.py:1198] (2/4) Epoch 18, batch 2850, loss[loss=0.2776, ctc_loss=0.2001, cr_loss=0.3875, over 11978.00 frames. ], tot_loss[loss=0.222, ctc_loss=0.149, cr_loss=0.3653, over 3350701.99 frames. ], batch size: 123, lr: 6.64e-03, grad_scale: 32.0 2024-09-23 18:47:30,406 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=322429.3333333333, ans=0.125 2024-09-23 18:47:46,780 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.58 vs. limit=15.0 2024-09-23 18:47:53,654 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=11.34 vs. limit=15.0 2024-09-23 18:48:10,271 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=322569.3333333333, ans=0.125 2024-09-23 18:48:15,015 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=322569.3333333333, ans=0.1 2024-09-23 18:48:26,237 INFO [train.py:1198] (2/4) Epoch 18, batch 2900, loss[loss=0.2362, ctc_loss=0.16, cr_loss=0.3814, over 16723.00 frames. ], tot_loss[loss=0.2218, ctc_loss=0.1488, cr_loss=0.3654, over 3354979.87 frames. ], batch size: 61, lr: 6.64e-03, grad_scale: 32.0 2024-09-23 18:48:31,345 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 18:48:48,985 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=322662.6666666667, ans=0.025 2024-09-23 18:49:40,536 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.288e+02 1.376e+02 1.532e+02 2.384e+02, threshold=2.751e+02, percent-clipped=0.0 2024-09-23 18:49:48,607 INFO [train.py:1198] (2/4) Epoch 18, batch 2950, loss[loss=0.2005, ctc_loss=0.1316, cr_loss=0.3443, over 17300.00 frames. ], tot_loss[loss=0.2216, ctc_loss=0.1486, cr_loss=0.3651, over 3349738.04 frames. ], batch size: 42, lr: 6.63e-03, grad_scale: 32.0 2024-09-23 18:50:01,807 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=322849.3333333333, ans=0.125 2024-09-23 18:50:10,489 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=17.93 vs. limit=22.5 2024-09-23 18:50:27,295 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=322942.6666666667, ans=0.025 2024-09-23 18:50:43,195 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=322989.3333333333, ans=0.0 2024-09-23 18:50:44,675 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=322989.3333333333, ans=0.1 2024-09-23 18:50:49,677 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=322989.3333333333, ans=0.07 2024-09-23 18:50:51,387 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=322989.3333333333, ans=0.125 2024-09-23 18:50:54,441 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.min_positive, batch_count=323036.0, ans=0.025 2024-09-23 18:51:00,824 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=323036.0, ans=0.125 2024-09-23 18:51:08,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=323036.0, ans=0.2 2024-09-23 18:51:11,407 INFO [train.py:1198] (2/4) Epoch 18, batch 3000, loss[loss=0.2052, ctc_loss=0.1366, cr_loss=0.3431, over 17025.00 frames. ], tot_loss[loss=0.2215, ctc_loss=0.1485, cr_loss=0.3649, over 3357831.26 frames. ], batch size: 56, lr: 6.63e-03, grad_scale: 32.0 2024-09-23 18:51:11,408 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 18:51:26,831 INFO [train.py:1230] (2/4) Epoch 18, validation: loss=0.04062, ctc_loss=0.04062, cr_loss=7.511e-15, over 944034.00 frames. 2024-09-23 18:51:26,832 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 18:51:28,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=323082.6666666667, ans=0.125 2024-09-23 18:52:39,991 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.237e+02 1.313e+02 1.391e+02 1.785e+02, threshold=2.627e+02, percent-clipped=0.0 2024-09-23 18:52:47,793 INFO [train.py:1198] (2/4) Epoch 18, batch 3050, loss[loss=0.2505, ctc_loss=0.1758, cr_loss=0.3737, over 12415.00 frames. ], tot_loss[loss=0.2212, ctc_loss=0.1483, cr_loss=0.3646, over 3354292.57 frames. ], batch size: 123, lr: 6.63e-03, grad_scale: 32.0 2024-09-23 18:52:54,409 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=323316.0, ans=0.5 2024-09-23 18:53:36,787 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn2.whiten, num_groups=1, num_channels=192, metric=8.95 vs. limit=22.5 2024-09-23 18:54:08,393 INFO [train.py:1198] (2/4) Epoch 18, batch 3100, loss[loss=0.1897, ctc_loss=0.1212, cr_loss=0.343, over 17035.00 frames. ], tot_loss[loss=0.2209, ctc_loss=0.1481, cr_loss=0.3641, over 3339388.36 frames. ], batch size: 44, lr: 6.63e-03, grad_scale: 32.0 2024-09-23 18:54:19,625 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.min_abs, batch_count=323549.3333333333, ans=0.5 2024-09-23 18:54:49,094 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=323642.6666666667, ans=0.1 2024-09-23 18:55:19,122 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.021e+02 1.256e+02 1.342e+02 1.439e+02 3.475e+02, threshold=2.684e+02, percent-clipped=1.0 2024-09-23 18:55:26,830 INFO [train.py:1198] (2/4) Epoch 18, batch 3150, loss[loss=0.2371, ctc_loss=0.1602, cr_loss=0.3843, over 17020.00 frames. ], tot_loss[loss=0.221, ctc_loss=0.1482, cr_loss=0.3644, over 3339987.30 frames. ], batch size: 51, lr: 6.62e-03, grad_scale: 32.0 2024-09-23 18:55:36,046 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.53 vs. limit=15.0 2024-09-23 18:55:36,633 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=323782.6666666667, ans=0.1 2024-09-23 18:55:39,667 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=323782.6666666667, ans=0.125 2024-09-23 18:55:52,213 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=323829.3333333333, ans=0.125 2024-09-23 18:56:11,912 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.49 vs. limit=10.0 2024-09-23 18:56:37,496 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=12.48 vs. limit=15.0 2024-09-23 18:56:45,842 INFO [train.py:1198] (2/4) Epoch 18, batch 3200, loss[loss=0.2425, ctc_loss=0.1647, cr_loss=0.3893, over 15167.00 frames. ], tot_loss[loss=0.22, ctc_loss=0.1473, cr_loss=0.3635, over 3348665.25 frames. ], batch size: 89, lr: 6.62e-03, grad_scale: 32.0 2024-09-23 18:56:46,057 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=324016.0, ans=0.0 2024-09-23 18:56:48,382 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=4.50 vs. limit=15.0 2024-09-23 18:56:58,606 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=324016.0, ans=0.125 2024-09-23 18:57:01,776 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 18:57:12,593 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=18.74 vs. limit=22.5 2024-09-23 18:57:49,503 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=324202.6666666667, ans=0.0 2024-09-23 18:57:55,597 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 18:57:58,436 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.249e+02 1.354e+02 1.459e+02 2.306e+02, threshold=2.708e+02, percent-clipped=0.0 2024-09-23 18:58:06,266 INFO [train.py:1198] (2/4) Epoch 18, batch 3250, loss[loss=0.191, ctc_loss=0.1304, cr_loss=0.3027, over 17051.00 frames. ], tot_loss[loss=0.22, ctc_loss=0.1473, cr_loss=0.3632, over 3344800.27 frames. ], batch size: 46, lr: 6.62e-03, grad_scale: 32.0 2024-09-23 18:58:06,651 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=324249.3333333333, ans=0.125 2024-09-23 18:58:13,054 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=324249.3333333333, ans=0.0 2024-09-23 18:58:17,595 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=324249.3333333333, ans=0.125 2024-09-23 18:58:37,933 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=324342.6666666667, ans=0.1 2024-09-23 18:58:48,778 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=324342.6666666667, ans=0.125 2024-09-23 18:58:51,937 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=324389.3333333333, ans=0.0 2024-09-23 18:59:18,532 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=324436.0, ans=0.125 2024-09-23 18:59:24,401 INFO [train.py:1198] (2/4) Epoch 18, batch 3300, loss[loss=0.2678, ctc_loss=0.1864, cr_loss=0.4069, over 14971.00 frames. ], tot_loss[loss=0.2209, ctc_loss=0.148, cr_loss=0.3644, over 3342967.42 frames. ], batch size: 89, lr: 6.62e-03, grad_scale: 32.0 2024-09-23 18:59:25,265 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.46 vs. limit=15.0 2024-09-23 19:00:03,559 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=324576.0, ans=0.125 2024-09-23 19:00:16,680 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=11.68 vs. limit=22.5 2024-09-23 19:00:22,590 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=324622.6666666667, ans=0.125 2024-09-23 19:00:26,296 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.83 vs. limit=10.0 2024-09-23 19:00:30,896 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=16.14 vs. limit=22.5 2024-09-23 19:00:36,180 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.318e+02 1.406e+02 1.543e+02 2.189e+02, threshold=2.811e+02, percent-clipped=0.0 2024-09-23 19:00:41,661 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.71 vs. limit=6.0 2024-09-23 19:00:44,001 INFO [train.py:1198] (2/4) Epoch 18, batch 3350, loss[loss=0.2432, ctc_loss=0.1637, cr_loss=0.3973, over 17038.00 frames. ], tot_loss[loss=0.2202, ctc_loss=0.1474, cr_loss=0.364, over 3357692.20 frames. ], batch size: 56, lr: 6.62e-03, grad_scale: 32.0 2024-09-23 19:01:03,555 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=5.64 vs. limit=12.0 2024-09-23 19:01:12,755 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=324762.6666666667, ans=0.125 2024-09-23 19:01:32,516 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=6.37 vs. limit=15.0 2024-09-23 19:01:42,498 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=324856.0, ans=0.0 2024-09-23 19:01:44,183 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=324856.0, ans=0.125 2024-09-23 19:02:04,708 INFO [train.py:1198] (2/4) Epoch 18, batch 3400, loss[loss=0.2204, ctc_loss=0.1495, cr_loss=0.3548, over 17322.00 frames. ], tot_loss[loss=0.221, ctc_loss=0.148, cr_loss=0.3649, over 3356042.30 frames. ], batch size: 51, lr: 6.61e-03, grad_scale: 32.0 2024-09-23 19:02:05,046 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 19:02:51,674 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.min_abs, batch_count=325089.3333333333, ans=0.5 2024-09-23 19:02:58,043 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=325089.3333333333, ans=0.0 2024-09-23 19:03:03,073 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.09 vs. limit=15.0 2024-09-23 19:03:14,633 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.106e+02 1.235e+02 1.326e+02 1.505e+02 2.372e+02, threshold=2.653e+02, percent-clipped=0.0 2024-09-23 19:03:22,389 INFO [train.py:1198] (2/4) Epoch 18, batch 3450, loss[loss=0.2499, ctc_loss=0.1758, cr_loss=0.3706, over 12129.00 frames. ], tot_loss[loss=0.2196, ctc_loss=0.147, cr_loss=0.3628, over 3354376.78 frames. ], batch size: 124, lr: 6.61e-03, grad_scale: 32.0 2024-09-23 19:03:29,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=325182.6666666667, ans=0.125 2024-09-23 19:03:30,890 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=325182.6666666667, ans=0.125 2024-09-23 19:03:32,589 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.min_positive, batch_count=325182.6666666667, ans=0.025 2024-09-23 19:03:34,020 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=325182.6666666667, ans=0.0 2024-09-23 19:03:35,681 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=325182.6666666667, ans=0.125 2024-09-23 19:03:49,456 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=325229.3333333333, ans=0.0 2024-09-23 19:04:00,902 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=325276.0, ans=0.1 2024-09-23 19:04:19,983 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.67 vs. limit=10.0 2024-09-23 19:04:36,015 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.01 vs. limit=6.0 2024-09-23 19:04:39,707 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=325369.3333333333, ans=0.125 2024-09-23 19:04:42,580 INFO [train.py:1198] (2/4) Epoch 18, batch 3500, loss[loss=0.2064, ctc_loss=0.1361, cr_loss=0.3517, over 17208.00 frames. ], tot_loss[loss=0.2195, ctc_loss=0.147, cr_loss=0.3627, over 3356375.37 frames. ], batch size: 47, lr: 6.61e-03, grad_scale: 32.0 2024-09-23 19:04:44,479 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=325416.0, ans=0.07 2024-09-23 19:05:24,730 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=10.40 vs. limit=15.0 2024-09-23 19:05:30,908 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module1.whiten, num_groups=1, num_channels=192, metric=7.98 vs. limit=15.0 2024-09-23 19:05:31,619 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=325556.0, ans=0.0 2024-09-23 19:05:39,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=325556.0, ans=0.0 2024-09-23 19:05:39,301 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=325556.0, ans=0.2 2024-09-23 19:05:43,891 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=325602.6666666667, ans=0.125 2024-09-23 19:05:51,172 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.whiten, num_groups=1, num_channels=192, metric=4.80 vs. limit=12.0 2024-09-23 19:05:53,068 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.293e+02 1.389e+02 1.509e+02 2.092e+02, threshold=2.779e+02, percent-clipped=0.0 2024-09-23 19:05:53,413 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=325602.6666666667, ans=0.125 2024-09-23 19:06:00,905 INFO [train.py:1198] (2/4) Epoch 18, batch 3550, loss[loss=0.213, ctc_loss=0.1438, cr_loss=0.3462, over 17009.00 frames. ], tot_loss[loss=0.2199, ctc_loss=0.1473, cr_loss=0.363, over 3347052.90 frames. ], batch size: 51, lr: 6.61e-03, grad_scale: 32.0 2024-09-23 19:06:02,779 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=325649.3333333333, ans=0.125 2024-09-23 19:06:15,652 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=325696.0, ans=0.05 2024-09-23 19:06:18,724 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=325696.0, ans=0.125 2024-09-23 19:07:21,663 INFO [train.py:1198] (2/4) Epoch 18, batch 3600, loss[loss=0.2365, ctc_loss=0.1626, cr_loss=0.3699, over 17192.00 frames. ], tot_loss[loss=0.2197, ctc_loss=0.1472, cr_loss=0.3624, over 3343806.15 frames. ], batch size: 55, lr: 6.60e-03, grad_scale: 32.0 2024-09-23 19:07:24,959 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=325882.6666666667, ans=0.125 2024-09-23 19:07:42,332 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=325929.3333333333, ans=0.0 2024-09-23 19:07:44,493 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.50 vs. limit=6.0 2024-09-23 19:08:04,215 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.15 vs. limit=22.5 2024-09-23 19:08:16,211 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=326022.6666666667, ans=0.0 2024-09-23 19:08:21,724 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=3.49 vs. limit=15.0 2024-09-23 19:08:31,458 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.271e+02 1.404e+02 1.560e+02 2.337e+02, threshold=2.809e+02, percent-clipped=0.0 2024-09-23 19:08:31,843 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 19:08:39,250 INFO [train.py:1198] (2/4) Epoch 18, batch 3650, loss[loss=0.2099, ctc_loss=0.1401, cr_loss=0.349, over 16973.00 frames. ], tot_loss[loss=0.2194, ctc_loss=0.147, cr_loss=0.362, over 3346287.94 frames. ], batch size: 42, lr: 6.60e-03, grad_scale: 32.0 2024-09-23 19:08:49,160 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.min_positive, batch_count=326116.0, ans=0.05 2024-09-23 19:08:49,637 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=11.40 vs. limit=15.0 2024-09-23 19:09:15,845 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=326209.3333333333, ans=0.1 2024-09-23 19:09:32,242 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=326256.0, ans=0.125 2024-09-23 19:10:00,632 INFO [train.py:1198] (2/4) Epoch 18, batch 3700, loss[loss=0.2085, ctc_loss=0.1386, cr_loss=0.3493, over 17354.00 frames. ], tot_loss[loss=0.2186, ctc_loss=0.1465, cr_loss=0.3606, over 3347502.44 frames. ], batch size: 48, lr: 6.60e-03, grad_scale: 32.0 2024-09-23 19:10:10,119 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=326349.3333333333, ans=0.125 2024-09-23 19:10:13,225 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=326349.3333333333, ans=0.125 2024-09-23 19:10:16,438 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=326396.0, ans=0.0 2024-09-23 19:10:40,087 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=326442.6666666667, ans=0.125 2024-09-23 19:10:41,482 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer_ff2.min_abs, batch_count=326442.6666666667, ans=0.1 2024-09-23 19:10:50,647 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=326489.3333333333, ans=0.025 2024-09-23 19:10:54,361 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=17.58 vs. limit=22.5 2024-09-23 19:11:06,357 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=326536.0, ans=0.125 2024-09-23 19:11:08,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=326536.0, ans=0.125 2024-09-23 19:11:10,730 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.270e+02 1.376e+02 1.485e+02 2.172e+02, threshold=2.752e+02, percent-clipped=0.0 2024-09-23 19:11:18,535 INFO [train.py:1198] (2/4) Epoch 18, batch 3750, loss[loss=0.2221, ctc_loss=0.1473, cr_loss=0.3743, over 17021.00 frames. ], tot_loss[loss=0.2196, ctc_loss=0.1473, cr_loss=0.3615, over 3340340.34 frames. ], batch size: 52, lr: 6.60e-03, grad_scale: 32.0 2024-09-23 19:12:03,784 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=326676.0, ans=0.0 2024-09-23 19:12:31,917 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=326769.3333333333, ans=0.1 2024-09-23 19:12:33,637 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=326769.3333333333, ans=0.0 2024-09-23 19:12:37,968 INFO [train.py:1198] (2/4) Epoch 18, batch 3800, loss[loss=0.2069, ctc_loss=0.1361, cr_loss=0.3538, over 17027.00 frames. ], tot_loss[loss=0.2199, ctc_loss=0.1476, cr_loss=0.3619, over 3327219.29 frames. ], batch size: 44, lr: 6.59e-03, grad_scale: 32.0 2024-09-23 19:12:43,130 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=10.46 vs. limit=22.5 2024-09-23 19:12:55,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=326862.6666666667, ans=0.0 2024-09-23 19:12:58,968 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=326862.6666666667, ans=0.0 2024-09-23 19:13:01,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=326862.6666666667, ans=0.0 2024-09-23 19:13:03,673 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=326862.6666666667, ans=0.0 2024-09-23 19:13:16,324 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=326909.3333333333, ans=0.125 2024-09-23 19:13:28,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=326956.0, ans=0.125 2024-09-23 19:13:44,642 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=327002.6666666667, ans=0.0 2024-09-23 19:13:49,015 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.113e+02 1.286e+02 1.423e+02 1.561e+02 2.075e+02, threshold=2.847e+02, percent-clipped=0.0 2024-09-23 19:13:56,833 INFO [train.py:1198] (2/4) Epoch 18, batch 3850, loss[loss=0.1846, ctc_loss=0.1214, cr_loss=0.3158, over 16946.00 frames. ], tot_loss[loss=0.2202, ctc_loss=0.1479, cr_loss=0.3616, over 3310251.92 frames. ], batch size: 42, lr: 6.59e-03, grad_scale: 32.0 2024-09-23 19:14:06,878 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=11.08 vs. limit=22.5 2024-09-23 19:14:14,493 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=327096.0, ans=0.1 2024-09-23 19:14:33,998 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=327142.6666666667, ans=0.125 2024-09-23 19:14:47,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=327189.3333333333, ans=0.125 2024-09-23 19:14:48,035 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.83 vs. limit=15.0 2024-09-23 19:14:53,377 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=327189.3333333333, ans=0.125 2024-09-23 19:15:59,140 INFO [train.py:1198] (2/4) Epoch 19, batch 0, loss[loss=0.2173, ctc_loss=0.1417, cr_loss=0.3778, over 17023.00 frames. ], tot_loss[loss=0.2173, ctc_loss=0.1417, cr_loss=0.3778, over 17023.00 frames. ], batch size: 44, lr: 6.41e-03, grad_scale: 32.0 2024-09-23 19:15:59,140 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 19:16:14,368 INFO [train.py:1230] (2/4) Epoch 19, validation: loss=0.03972, ctc_loss=0.03972, cr_loss=8.025e-15, over 944034.00 frames. 2024-09-23 19:16:14,368 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 19:16:19,328 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=327264.0, ans=0.125 2024-09-23 19:16:27,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=327264.0, ans=0.2 2024-09-23 19:16:42,338 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=327310.6666666667, ans=0.035 2024-09-23 19:17:06,745 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=327404.0, ans=0.125 2024-09-23 19:17:39,622 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.290e+02 1.473e+02 1.789e+02 2.583e+02, threshold=2.945e+02, percent-clipped=0.0 2024-09-23 19:17:41,330 INFO [train.py:1198] (2/4) Epoch 19, batch 50, loss[loss=0.1879, ctc_loss=0.1218, cr_loss=0.3305, over 17174.00 frames. ], tot_loss[loss=0.214, ctc_loss=0.1421, cr_loss=0.3597, over 769433.05 frames. ], batch size: 41, lr: 6.41e-03, grad_scale: 32.0 2024-09-23 19:17:41,574 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=327497.3333333333, ans=0.125 2024-09-23 19:17:41,579 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=327497.3333333333, ans=0.2 2024-09-23 19:18:07,479 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=327544.0, ans=0.125 2024-09-23 19:18:15,449 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=327590.6666666667, ans=0.125 2024-09-23 19:18:49,496 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 19:18:52,627 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=327684.0, ans=0.0 2024-09-23 19:18:55,814 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=327684.0, ans=0.125 2024-09-23 19:19:00,666 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=327684.0, ans=0.1 2024-09-23 19:19:03,588 INFO [train.py:1198] (2/4) Epoch 19, batch 100, loss[loss=0.2208, ctc_loss=0.1462, cr_loss=0.3732, over 17310.00 frames. ], tot_loss[loss=0.2145, ctc_loss=0.1428, cr_loss=0.3588, over 1338208.14 frames. ], batch size: 46, lr: 6.41e-03, grad_scale: 32.0 2024-09-23 19:19:06,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=327730.6666666667, ans=0.2 2024-09-23 19:19:08,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=327730.6666666667, ans=0.125 2024-09-23 19:19:15,299 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.57 vs. limit=15.0 2024-09-23 19:19:22,942 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=327777.3333333333, ans=0.125 2024-09-23 19:19:48,343 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=327824.0, ans=0.0 2024-09-23 19:20:06,133 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=512, metric=13.45 vs. limit=22.5 2024-09-23 19:20:15,086 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=327917.3333333333, ans=0.125 2024-09-23 19:20:21,266 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.064e+02 1.272e+02 1.373e+02 1.513e+02 2.086e+02, threshold=2.746e+02, percent-clipped=0.0 2024-09-23 19:20:22,822 INFO [train.py:1198] (2/4) Epoch 19, batch 150, loss[loss=0.191, ctc_loss=0.1275, cr_loss=0.3178, over 16994.00 frames. ], tot_loss[loss=0.2161, ctc_loss=0.1441, cr_loss=0.36, over 1788985.90 frames. ], batch size: 39, lr: 6.40e-03, grad_scale: 32.0 2024-09-23 19:20:23,201 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=327964.0, ans=0.125 2024-09-23 19:20:31,125 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=327964.0, ans=0.125 2024-09-23 19:20:48,547 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=328010.6666666667, ans=0.025 2024-09-23 19:20:57,842 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=328057.3333333333, ans=0.125 2024-09-23 19:21:04,351 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=328057.3333333333, ans=0.125 2024-09-23 19:21:33,336 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=328150.6666666667, ans=0.5 2024-09-23 19:21:34,856 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=328150.6666666667, ans=0.125 2024-09-23 19:21:48,158 INFO [train.py:1198] (2/4) Epoch 19, batch 200, loss[loss=0.2567, ctc_loss=0.1715, cr_loss=0.4259, over 16996.00 frames. ], tot_loss[loss=0.2182, ctc_loss=0.1457, cr_loss=0.3625, over 2129114.11 frames. ], batch size: 53, lr: 6.40e-03, grad_scale: 32.0 2024-09-23 19:22:07,746 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=328244.0, ans=0.07 2024-09-23 19:22:10,930 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=328244.0, ans=0.04949747468305833 2024-09-23 19:22:36,578 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=16.35 vs. limit=22.5 2024-09-23 19:22:38,169 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.97 vs. limit=6.0 2024-09-23 19:22:52,034 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=328337.3333333333, ans=0.125 2024-09-23 19:23:00,058 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=328384.0, ans=0.07 2024-09-23 19:23:01,584 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=328384.0, ans=0.2 2024-09-23 19:23:01,730 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=328384.0, ans=0.125 2024-09-23 19:23:06,451 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=328384.0, ans=0.125 2024-09-23 19:23:08,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=328384.0, ans=0.2 2024-09-23 19:23:09,156 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.059e+02 1.256e+02 1.366e+02 1.552e+02 2.193e+02, threshold=2.732e+02, percent-clipped=0.0 2024-09-23 19:23:09,521 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=328430.6666666667, ans=0.125 2024-09-23 19:23:10,719 INFO [train.py:1198] (2/4) Epoch 19, batch 250, loss[loss=0.2058, ctc_loss=0.1359, cr_loss=0.3492, over 17100.00 frames. ], tot_loss[loss=0.2198, ctc_loss=0.1469, cr_loss=0.3647, over 2410644.64 frames. ], batch size: 40, lr: 6.40e-03, grad_scale: 32.0 2024-09-23 19:23:20,608 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.min_positive, batch_count=328430.6666666667, ans=0.025 2024-09-23 19:23:32,318 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=328477.3333333333, ans=0.125 2024-09-23 19:24:00,785 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.72 vs. limit=5.0 2024-09-23 19:24:07,764 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=328570.6666666667, ans=0.125 2024-09-23 19:24:09,354 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=328570.6666666667, ans=0.125 2024-09-23 19:24:31,665 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=6.93 vs. limit=15.0 2024-09-23 19:24:32,609 INFO [train.py:1198] (2/4) Epoch 19, batch 300, loss[loss=0.2613, ctc_loss=0.1794, cr_loss=0.4092, over 16048.00 frames. ], tot_loss[loss=0.2189, ctc_loss=0.1464, cr_loss=0.3626, over 2611668.91 frames. ], batch size: 74, lr: 6.40e-03, grad_scale: 32.0 2024-09-23 19:24:52,146 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=328710.6666666667, ans=0.2 2024-09-23 19:25:00,126 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=328710.6666666667, ans=0.0 2024-09-23 19:25:13,018 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=328757.3333333333, ans=0.125 2024-09-23 19:25:43,543 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=328850.6666666667, ans=0.0 2024-09-23 19:25:51,167 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.306e+02 1.376e+02 1.560e+02 2.317e+02, threshold=2.752e+02, percent-clipped=0.0 2024-09-23 19:25:52,831 INFO [train.py:1198] (2/4) Epoch 19, batch 350, loss[loss=0.2187, ctc_loss=0.1461, cr_loss=0.3632, over 16978.00 frames. ], tot_loss[loss=0.2206, ctc_loss=0.1477, cr_loss=0.3645, over 2773382.79 frames. ], batch size: 56, lr: 6.40e-03, grad_scale: 32.0 2024-09-23 19:26:04,384 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=328897.3333333333, ans=0.125 2024-09-23 19:26:05,913 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=328897.3333333333, ans=0.125 2024-09-23 19:26:21,573 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=328944.0, ans=0.1 2024-09-23 19:27:01,798 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=329084.0, ans=0.125 2024-09-23 19:27:10,150 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.26 vs. limit=12.0 2024-09-23 19:27:11,323 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=329084.0, ans=0.0 2024-09-23 19:27:12,780 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=329084.0, ans=0.1 2024-09-23 19:27:12,895 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=329084.0, ans=0.0 2024-09-23 19:27:17,481 INFO [train.py:1198] (2/4) Epoch 19, batch 400, loss[loss=0.1803, ctc_loss=0.1174, cr_loss=0.3147, over 17038.00 frames. ], tot_loss[loss=0.2196, ctc_loss=0.1468, cr_loss=0.3637, over 2915623.81 frames. ], batch size: 39, lr: 6.39e-03, grad_scale: 32.0 2024-09-23 19:27:17,832 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=329130.6666666667, ans=0.125 2024-09-23 19:27:25,229 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=329130.6666666667, ans=0.125 2024-09-23 19:27:56,244 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=10.96 vs. limit=22.5 2024-09-23 19:28:13,507 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=329270.6666666667, ans=0.125 2024-09-23 19:28:29,135 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=329317.3333333333, ans=0.07 2024-09-23 19:28:35,404 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=329317.3333333333, ans=0.125 2024-09-23 19:28:38,629 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=329317.3333333333, ans=0.125 2024-09-23 19:28:41,423 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.241e+02 1.311e+02 1.431e+02 1.671e+02, threshold=2.622e+02, percent-clipped=0.0 2024-09-23 19:28:43,031 INFO [train.py:1198] (2/4) Epoch 19, batch 450, loss[loss=0.2318, ctc_loss=0.1588, cr_loss=0.365, over 17001.00 frames. ], tot_loss[loss=0.2208, ctc_loss=0.1478, cr_loss=0.3649, over 3012630.57 frames. ], batch size: 53, lr: 6.39e-03, grad_scale: 32.0 2024-09-23 19:28:49,534 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=329364.0, ans=0.125 2024-09-23 19:28:58,186 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.85 vs. limit=10.0 2024-09-23 19:28:59,004 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=329410.6666666667, ans=0.125 2024-09-23 19:29:00,723 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=329410.6666666667, ans=0.125 2024-09-23 19:29:02,095 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=329410.6666666667, ans=0.0 2024-09-23 19:29:10,359 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=329410.6666666667, ans=0.2 2024-09-23 19:29:33,367 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=11.28 vs. limit=15.0 2024-09-23 19:29:44,230 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=329504.0, ans=0.025 2024-09-23 19:29:59,269 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=10.73 vs. limit=15.0 2024-09-23 19:30:03,311 INFO [train.py:1198] (2/4) Epoch 19, batch 500, loss[loss=0.2734, ctc_loss=0.1853, cr_loss=0.4408, over 16981.00 frames. ], tot_loss[loss=0.2191, ctc_loss=0.1466, cr_loss=0.3626, over 3096994.82 frames. ], batch size: 53, lr: 6.39e-03, grad_scale: 32.0 2024-09-23 19:30:34,258 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=329690.6666666667, ans=0.125 2024-09-23 19:30:36,483 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=3.38 vs. limit=15.0 2024-09-23 19:30:43,555 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=329690.6666666667, ans=0.1 2024-09-23 19:31:02,765 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.min_positive, batch_count=329737.3333333333, ans=0.025 2024-09-23 19:31:12,505 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=329784.0, ans=0.125 2024-09-23 19:31:21,603 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.126e+02 1.237e+02 1.337e+02 1.500e+02 1.948e+02, threshold=2.674e+02, percent-clipped=0.0 2024-09-23 19:31:22,068 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=329830.6666666667, ans=0.1 2024-09-23 19:31:23,261 INFO [train.py:1198] (2/4) Epoch 19, batch 550, loss[loss=0.2355, ctc_loss=0.1592, cr_loss=0.3814, over 17315.00 frames. ], tot_loss[loss=0.2181, ctc_loss=0.1458, cr_loss=0.3616, over 3163058.72 frames. ], batch size: 51, lr: 6.39e-03, grad_scale: 32.0 2024-09-23 19:32:04,298 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.min_positive, batch_count=329924.0, ans=0.025 2024-09-23 19:32:27,629 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=329970.6666666667, ans=0.025 2024-09-23 19:32:29,314 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=329970.6666666667, ans=0.125 2024-09-23 19:32:34,599 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.08 vs. limit=15.0 2024-09-23 19:32:37,152 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=330017.3333333333, ans=0.2 2024-09-23 19:32:40,341 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=330017.3333333333, ans=0.2 2024-09-23 19:32:45,535 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.28 vs. limit=6.0 2024-09-23 19:32:48,241 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer_na.min_abs, batch_count=330017.3333333333, ans=0.02 2024-09-23 19:32:51,188 INFO [train.py:1198] (2/4) Epoch 19, batch 600, loss[loss=0.2444, ctc_loss=0.1639, cr_loss=0.4027, over 17025.00 frames. ], tot_loss[loss=0.2187, ctc_loss=0.1462, cr_loss=0.3623, over 3206474.60 frames. ], batch size: 52, lr: 6.38e-03, grad_scale: 32.0 2024-09-23 19:33:00,851 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=330064.0, ans=0.1 2024-09-23 19:33:05,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=330110.6666666667, ans=0.025 2024-09-23 19:33:22,991 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=330110.6666666667, ans=0.125 2024-09-23 19:33:26,203 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=330157.3333333333, ans=0.0 2024-09-23 19:33:31,079 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=330157.3333333333, ans=0.2 2024-09-23 19:33:39,202 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=330157.3333333333, ans=0.125 2024-09-23 19:33:52,214 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=330204.0, ans=0.2 2024-09-23 19:34:05,102 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=330250.6666666667, ans=0.0 2024-09-23 19:34:12,589 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.274e+02 1.356e+02 1.568e+02 2.511e+02, threshold=2.713e+02, percent-clipped=0.0 2024-09-23 19:34:14,266 INFO [train.py:1198] (2/4) Epoch 19, batch 650, loss[loss=0.173, ctc_loss=0.1117, cr_loss=0.3063, over 16696.00 frames. ], tot_loss[loss=0.219, ctc_loss=0.1464, cr_loss=0.3632, over 3242720.55 frames. ], batch size: 37, lr: 6.38e-03, grad_scale: 64.0 2024-09-23 19:34:45,705 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.16 vs. limit=10.0 2024-09-23 19:35:14,238 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=330437.3333333333, ans=0.2 2024-09-23 19:35:34,528 INFO [train.py:1198] (2/4) Epoch 19, batch 700, loss[loss=0.2315, ctc_loss=0.1551, cr_loss=0.3821, over 16763.00 frames. ], tot_loss[loss=0.2184, ctc_loss=0.146, cr_loss=0.3619, over 3266951.16 frames. ], batch size: 61, lr: 6.38e-03, grad_scale: 64.0 2024-09-23 19:36:10,382 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=330624.0, ans=0.2 2024-09-23 19:36:36,829 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=330670.6666666667, ans=0.125 2024-09-23 19:36:55,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=330717.3333333333, ans=0.125 2024-09-23 19:36:58,032 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.050e+02 1.250e+02 1.371e+02 1.487e+02 1.794e+02, threshold=2.742e+02, percent-clipped=0.0 2024-09-23 19:36:59,667 INFO [train.py:1198] (2/4) Epoch 19, batch 750, loss[loss=0.2352, ctc_loss=0.155, cr_loss=0.4011, over 16526.00 frames. ], tot_loss[loss=0.2176, ctc_loss=0.1455, cr_loss=0.3605, over 3280193.18 frames. ], batch size: 66, lr: 6.38e-03, grad_scale: 64.0 2024-09-23 19:37:16,079 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=330810.6666666667, ans=0.0 2024-09-23 19:38:25,055 INFO [train.py:1198] (2/4) Epoch 19, batch 800, loss[loss=0.2605, ctc_loss=0.1755, cr_loss=0.4245, over 17186.00 frames. ], tot_loss[loss=0.2184, ctc_loss=0.1462, cr_loss=0.3609, over 3288699.34 frames. ], batch size: 55, lr: 6.38e-03, grad_scale: 32.0 2024-09-23 19:38:39,547 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=331044.0, ans=0.125 2024-09-23 19:38:48,246 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=5.54 vs. limit=15.0 2024-09-23 19:38:58,743 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=331090.6666666667, ans=0.0 2024-09-23 19:39:30,734 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=331184.0, ans=0.0 2024-09-23 19:39:44,567 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.257e+02 1.307e+02 1.431e+02 3.261e+02, threshold=2.613e+02, percent-clipped=1.0 2024-09-23 19:39:44,591 INFO [train.py:1198] (2/4) Epoch 19, batch 850, loss[loss=0.1786, ctc_loss=0.1175, cr_loss=0.3055, over 17207.00 frames. ], tot_loss[loss=0.2189, ctc_loss=0.1466, cr_loss=0.3617, over 3297408.23 frames. ], batch size: 41, lr: 6.37e-03, grad_scale: 32.0 2024-09-23 19:39:50,022 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module1.whiten, num_groups=1, num_channels=512, metric=6.14 vs. limit=15.0 2024-09-23 19:40:05,282 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=331277.3333333333, ans=0.1 2024-09-23 19:40:05,448 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=331277.3333333333, ans=0.125 2024-09-23 19:40:52,105 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=331417.3333333333, ans=0.0 2024-09-23 19:41:03,177 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=331464.0, ans=0.5 2024-09-23 19:41:04,316 INFO [train.py:1198] (2/4) Epoch 19, batch 900, loss[loss=0.2164, ctc_loss=0.1436, cr_loss=0.3637, over 17351.00 frames. ], tot_loss[loss=0.2198, ctc_loss=0.1471, cr_loss=0.3633, over 3309582.58 frames. ], batch size: 48, lr: 6.37e-03, grad_scale: 16.0 2024-09-23 19:41:31,177 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 19:41:36,547 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn2.whiten, num_groups=1, num_channels=192, metric=8.17 vs. limit=22.5 2024-09-23 19:41:49,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=331557.3333333333, ans=0.125 2024-09-23 19:42:21,518 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=7.52 vs. limit=15.0 2024-09-23 19:42:26,452 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.15 vs. limit=15.0 2024-09-23 19:42:30,989 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.10 vs. limit=6.0 2024-09-23 19:42:31,884 INFO [train.py:1198] (2/4) Epoch 19, batch 950, loss[loss=0.2117, ctc_loss=0.1416, cr_loss=0.3505, over 17022.00 frames. ], tot_loss[loss=0.2188, ctc_loss=0.1464, cr_loss=0.3619, over 3321080.59 frames. ], batch size: 44, lr: 6.37e-03, grad_scale: 16.0 2024-09-23 19:42:33,536 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.067e+02 1.264e+02 1.377e+02 1.501e+02 1.833e+02, threshold=2.753e+02, percent-clipped=0.0 2024-09-23 19:42:48,455 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=331744.0, ans=0.0 2024-09-23 19:42:59,725 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=331744.0, ans=0.025 2024-09-23 19:43:13,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=331790.6666666667, ans=0.125 2024-09-23 19:43:24,638 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=331837.3333333333, ans=0.125 2024-09-23 19:43:54,919 INFO [train.py:1198] (2/4) Epoch 19, batch 1000, loss[loss=0.2276, ctc_loss=0.153, cr_loss=0.3733, over 17315.00 frames. ], tot_loss[loss=0.2191, ctc_loss=0.1466, cr_loss=0.3624, over 3327641.43 frames. ], batch size: 51, lr: 6.37e-03, grad_scale: 16.0 2024-09-23 19:44:52,782 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.51 vs. limit=15.0 2024-09-23 19:45:01,749 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=332117.3333333333, ans=0.125 2024-09-23 19:45:12,998 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=332164.0, ans=0.125 2024-09-23 19:45:14,335 INFO [train.py:1198] (2/4) Epoch 19, batch 1050, loss[loss=0.2428, ctc_loss=0.1626, cr_loss=0.401, over 16515.00 frames. ], tot_loss[loss=0.22, ctc_loss=0.1473, cr_loss=0.3632, over 3326349.25 frames. ], batch size: 66, lr: 6.36e-03, grad_scale: 16.0 2024-09-23 19:45:14,771 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=332164.0, ans=0.125 2024-09-23 19:45:15,987 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.099e+02 1.292e+02 1.398e+02 1.505e+02 1.868e+02, threshold=2.795e+02, percent-clipped=0.0 2024-09-23 19:45:36,020 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.53 vs. limit=6.0 2024-09-23 19:46:06,275 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=332304.0, ans=0.125 2024-09-23 19:46:11,752 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=18.87 vs. limit=22.5 2024-09-23 19:46:24,879 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=332350.6666666667, ans=0.05 2024-09-23 19:46:39,718 INFO [train.py:1198] (2/4) Epoch 19, batch 1100, loss[loss=0.2502, ctc_loss=0.1688, cr_loss=0.4071, over 16944.00 frames. ], tot_loss[loss=0.2203, ctc_loss=0.1476, cr_loss=0.3637, over 3323981.46 frames. ], batch size: 58, lr: 6.36e-03, grad_scale: 16.0 2024-09-23 19:46:52,716 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=332397.3333333333, ans=0.04949747468305833 2024-09-23 19:46:59,197 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=332444.0, ans=0.125 2024-09-23 19:47:14,390 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=332490.6666666667, ans=0.1 2024-09-23 19:47:35,568 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=11.81 vs. limit=22.5 2024-09-23 19:47:39,825 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=332537.3333333333, ans=0.125 2024-09-23 19:47:55,826 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=332584.0, ans=0.0 2024-09-23 19:48:00,472 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=332630.6666666667, ans=0.125 2024-09-23 19:48:01,784 INFO [train.py:1198] (2/4) Epoch 19, batch 1150, loss[loss=0.1875, ctc_loss=0.1265, cr_loss=0.3048, over 17256.00 frames. ], tot_loss[loss=0.2178, ctc_loss=0.1456, cr_loss=0.3609, over 3336266.84 frames. ], batch size: 44, lr: 6.36e-03, grad_scale: 16.0 2024-09-23 19:48:06,032 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.131e+02 1.288e+02 1.429e+02 1.556e+02 3.211e+02, threshold=2.859e+02, percent-clipped=1.0 2024-09-23 19:48:14,376 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=332630.6666666667, ans=0.125 2024-09-23 19:48:22,469 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=332677.3333333333, ans=0.125 2024-09-23 19:49:05,888 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=332770.6666666667, ans=0.1 2024-09-23 19:49:24,949 INFO [train.py:1198] (2/4) Epoch 19, batch 1200, loss[loss=0.2552, ctc_loss=0.176, cr_loss=0.396, over 17197.00 frames. ], tot_loss[loss=0.218, ctc_loss=0.1458, cr_loss=0.3611, over 3345808.85 frames. ], batch size: 55, lr: 6.36e-03, grad_scale: 32.0 2024-09-23 19:49:37,058 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=4.97 vs. limit=12.0 2024-09-23 19:49:44,689 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=332910.6666666667, ans=0.125 2024-09-23 19:49:54,432 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=332910.6666666667, ans=0.125 2024-09-23 19:50:07,254 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=332957.3333333333, ans=0.125 2024-09-23 19:50:07,366 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=332957.3333333333, ans=0.125 2024-09-23 19:50:08,780 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=332957.3333333333, ans=0.0 2024-09-23 19:50:13,710 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=333004.0, ans=0.0 2024-09-23 19:50:20,561 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=19.93 vs. limit=22.5 2024-09-23 19:50:26,583 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.87 vs. limit=10.0 2024-09-23 19:50:45,364 INFO [train.py:1198] (2/4) Epoch 19, batch 1250, loss[loss=0.234, ctc_loss=0.1568, cr_loss=0.3863, over 17005.00 frames. ], tot_loss[loss=0.2188, ctc_loss=0.1463, cr_loss=0.3621, over 3347902.81 frames. ], batch size: 56, lr: 6.36e-03, grad_scale: 32.0 2024-09-23 19:50:46,878 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.324e+02 1.446e+02 1.574e+02 2.086e+02, threshold=2.891e+02, percent-clipped=0.0 2024-09-23 19:51:07,897 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=333144.0, ans=0.0 2024-09-23 19:51:11,207 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=333144.0, ans=0.2 2024-09-23 19:51:14,241 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=333144.0, ans=0.0 2024-09-23 19:52:01,178 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=333284.0, ans=0.125 2024-09-23 19:52:12,987 INFO [train.py:1198] (2/4) Epoch 19, batch 1300, loss[loss=0.2077, ctc_loss=0.1381, cr_loss=0.3478, over 17195.00 frames. ], tot_loss[loss=0.2187, ctc_loss=0.1463, cr_loss=0.3617, over 3349068.72 frames. ], batch size: 41, lr: 6.35e-03, grad_scale: 16.0 2024-09-23 19:52:19,504 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=333330.6666666667, ans=0.2 2024-09-23 19:52:42,756 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=7.31 vs. limit=15.0 2024-09-23 19:52:43,774 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=333424.0, ans=0.125 2024-09-23 19:52:56,745 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=333424.0, ans=0.125 2024-09-23 19:53:29,921 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=333517.3333333333, ans=0.1 2024-09-23 19:53:35,915 INFO [train.py:1198] (2/4) Epoch 19, batch 1350, loss[loss=0.2196, ctc_loss=0.1487, cr_loss=0.3547, over 15857.00 frames. ], tot_loss[loss=0.218, ctc_loss=0.1458, cr_loss=0.3609, over 3358725.99 frames. ], batch size: 74, lr: 6.35e-03, grad_scale: 16.0 2024-09-23 19:53:39,093 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.112e+02 1.258e+02 1.337e+02 1.461e+02 2.024e+02, threshold=2.674e+02, percent-clipped=0.0 2024-09-23 19:53:43,212 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=7.20 vs. limit=15.0 2024-09-23 19:53:54,202 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=12.90 vs. limit=22.5 2024-09-23 19:53:57,306 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=333610.6666666667, ans=0.0 2024-09-23 19:54:07,798 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=8.28 vs. limit=15.0 2024-09-23 19:54:18,399 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=333657.3333333333, ans=0.0 2024-09-23 19:54:35,724 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=333704.0, ans=0.2 2024-09-23 19:54:45,460 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=333750.6666666667, ans=0.125 2024-09-23 19:54:56,452 INFO [train.py:1198] (2/4) Epoch 19, batch 1400, loss[loss=0.195, ctc_loss=0.1309, cr_loss=0.3201, over 17082.00 frames. ], tot_loss[loss=0.2187, ctc_loss=0.1463, cr_loss=0.3621, over 3355147.45 frames. ], batch size: 43, lr: 6.35e-03, grad_scale: 16.0 2024-09-23 19:54:57,409 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.22 vs. limit=15.0 2024-09-23 19:55:11,300 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=333844.0, ans=0.125 2024-09-23 19:55:43,267 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=333937.3333333333, ans=0.125 2024-09-23 19:56:00,503 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=333984.0, ans=0.125 2024-09-23 19:56:02,012 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=333984.0, ans=0.125 2024-09-23 19:56:15,964 INFO [train.py:1198] (2/4) Epoch 19, batch 1450, loss[loss=0.187, ctc_loss=0.124, cr_loss=0.3147, over 16363.00 frames. ], tot_loss[loss=0.2186, ctc_loss=0.1462, cr_loss=0.3617, over 3359020.69 frames. ], batch size: 36, lr: 6.35e-03, grad_scale: 16.0 2024-09-23 19:56:21,632 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.283e+02 1.375e+02 1.499e+02 2.283e+02, threshold=2.750e+02, percent-clipped=0.0 2024-09-23 19:56:46,801 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=334077.3333333333, ans=0.125 2024-09-23 19:56:46,863 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=334077.3333333333, ans=0.0 2024-09-23 19:57:10,536 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=334170.6666666667, ans=0.125 2024-09-23 19:57:15,203 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=334170.6666666667, ans=0.125 2024-09-23 19:57:43,769 INFO [train.py:1198] (2/4) Epoch 19, batch 1500, loss[loss=0.2134, ctc_loss=0.1427, cr_loss=0.3535, over 16830.00 frames. ], tot_loss[loss=0.2185, ctc_loss=0.1461, cr_loss=0.3621, over 3359871.13 frames. ], batch size: 61, lr: 6.34e-03, grad_scale: 16.0 2024-09-23 19:58:22,041 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=334357.3333333333, ans=0.125 2024-09-23 19:58:33,105 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=334404.0, ans=0.0 2024-09-23 19:59:06,199 INFO [train.py:1198] (2/4) Epoch 19, batch 1550, loss[loss=0.1778, ctc_loss=0.1157, cr_loss=0.3108, over 17271.00 frames. ], tot_loss[loss=0.2193, ctc_loss=0.1467, cr_loss=0.3629, over 3363189.84 frames. ], batch size: 42, lr: 6.34e-03, grad_scale: 16.0 2024-09-23 19:59:09,456 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.039e+02 1.278e+02 1.387e+02 1.518e+02 2.007e+02, threshold=2.775e+02, percent-clipped=0.0 2024-09-23 19:59:32,137 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=334544.0, ans=0.0 2024-09-23 20:00:05,747 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=334637.3333333333, ans=0.125 2024-09-23 20:00:20,372 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 20:00:26,299 INFO [train.py:1198] (2/4) Epoch 19, batch 1600, loss[loss=0.2595, ctc_loss=0.172, cr_loss=0.4376, over 17024.00 frames. ], tot_loss[loss=0.2177, ctc_loss=0.1455, cr_loss=0.3613, over 3368375.79 frames. ], batch size: 52, lr: 6.34e-03, grad_scale: 32.0 2024-09-23 20:00:40,912 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=334777.3333333333, ans=0.125 2024-09-23 20:00:53,771 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.66 vs. limit=10.0 2024-09-23 20:00:55,047 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=334777.3333333333, ans=0.125 2024-09-23 20:00:55,619 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module1.whiten, num_groups=1, num_channels=512, metric=5.86 vs. limit=15.0 2024-09-23 20:01:33,661 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=334917.3333333333, ans=0.125 2024-09-23 20:01:47,925 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.min_positive, batch_count=334917.3333333333, ans=0.025 2024-09-23 20:01:50,761 INFO [train.py:1198] (2/4) Epoch 19, batch 1650, loss[loss=0.1911, ctc_loss=0.1276, cr_loss=0.3172, over 17216.00 frames. ], tot_loss[loss=0.2175, ctc_loss=0.1453, cr_loss=0.361, over 3370922.87 frames. ], batch size: 47, lr: 6.34e-03, grad_scale: 32.0 2024-09-23 20:01:53,946 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.297e+02 1.356e+02 1.480e+02 2.099e+02, threshold=2.713e+02, percent-clipped=0.0 2024-09-23 20:02:07,999 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=335010.6666666667, ans=0.125 2024-09-23 20:02:40,040 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=335104.0, ans=0.0 2024-09-23 20:03:16,143 INFO [train.py:1198] (2/4) Epoch 19, batch 1700, loss[loss=0.221, ctc_loss=0.1455, cr_loss=0.3773, over 17247.00 frames. ], tot_loss[loss=0.2175, ctc_loss=0.1453, cr_loss=0.3609, over 3367852.91 frames. ], batch size: 55, lr: 6.34e-03, grad_scale: 32.0 2024-09-23 20:03:40,236 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=335244.0, ans=0.125 2024-09-23 20:03:51,573 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=335290.6666666667, ans=0.125 2024-09-23 20:03:51,699 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=335290.6666666667, ans=0.1 2024-09-23 20:03:55,236 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=7.56 vs. limit=15.0 2024-09-23 20:03:57,829 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=335290.6666666667, ans=0.125 2024-09-23 20:04:01,164 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=335290.6666666667, ans=0.125 2024-09-23 20:04:12,413 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=335337.3333333333, ans=0.2 2024-09-23 20:04:31,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=335384.0, ans=0.125 2024-09-23 20:04:31,536 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 20:04:35,134 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=14.84 vs. limit=15.0 2024-09-23 20:04:36,068 INFO [train.py:1198] (2/4) Epoch 19, batch 1750, loss[loss=0.2757, ctc_loss=0.1916, cr_loss=0.4205, over 12264.00 frames. ], tot_loss[loss=0.2176, ctc_loss=0.1455, cr_loss=0.3603, over 3361455.18 frames. ], batch size: 123, lr: 6.33e-03, grad_scale: 32.0 2024-09-23 20:04:38,118 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 20:04:39,273 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.109e+02 1.243e+02 1.345e+02 1.441e+02 1.973e+02, threshold=2.691e+02, percent-clipped=0.0 2024-09-23 20:04:57,007 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=335477.3333333333, ans=0.2 2024-09-23 20:05:01,657 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=335477.3333333333, ans=0.125 2024-09-23 20:05:08,542 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.81 vs. limit=15.0 2024-09-23 20:05:18,322 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=4.79 vs. limit=15.0 2024-09-23 20:05:28,964 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=335570.6666666667, ans=0.1 2024-09-23 20:05:34,046 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=6.17 vs. limit=15.0 2024-09-23 20:05:39,877 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=335617.3333333333, ans=0.0 2024-09-23 20:05:55,468 INFO [train.py:1198] (2/4) Epoch 19, batch 1800, loss[loss=0.2434, ctc_loss=0.1636, cr_loss=0.399, over 17227.00 frames. ], tot_loss[loss=0.2168, ctc_loss=0.1449, cr_loss=0.3596, over 3363275.69 frames. ], batch size: 55, lr: 6.33e-03, grad_scale: 32.0 2024-09-23 20:06:02,311 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=335664.0, ans=0.0 2024-09-23 20:06:10,395 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=335710.6666666667, ans=0.0 2024-09-23 20:06:13,424 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=335710.6666666667, ans=0.0 2024-09-23 20:06:22,486 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=335710.6666666667, ans=0.125 2024-09-23 20:06:39,391 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 20:06:50,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=335804.0, ans=0.0 2024-09-23 20:07:05,809 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=335850.6666666667, ans=0.125 2024-09-23 20:07:22,992 INFO [train.py:1198] (2/4) Epoch 19, batch 1850, loss[loss=0.2301, ctc_loss=0.1556, cr_loss=0.3727, over 17001.00 frames. ], tot_loss[loss=0.2165, ctc_loss=0.1447, cr_loss=0.359, over 3360158.40 frames. ], batch size: 51, lr: 6.33e-03, grad_scale: 32.0 2024-09-23 20:07:23,337 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=335897.3333333333, ans=0.125 2024-09-23 20:07:26,227 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.069e+02 1.276e+02 1.381e+02 1.510e+02 2.241e+02, threshold=2.761e+02, percent-clipped=0.0 2024-09-23 20:07:28,455 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=11.23 vs. limit=12.0 2024-09-23 20:07:34,477 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=335897.3333333333, ans=0.95 2024-09-23 20:08:16,987 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=10.59 vs. limit=15.0 2024-09-23 20:08:21,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=336037.3333333333, ans=0.125 2024-09-23 20:08:28,218 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.37 vs. limit=15.0 2024-09-23 20:08:36,187 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.65 vs. limit=15.0 2024-09-23 20:08:48,088 INFO [train.py:1198] (2/4) Epoch 19, batch 1900, loss[loss=0.2363, ctc_loss=0.1581, cr_loss=0.391, over 17139.00 frames. ], tot_loss[loss=0.2157, ctc_loss=0.1441, cr_loss=0.3581, over 3362980.47 frames. ], batch size: 48, lr: 6.33e-03, grad_scale: 32.0 2024-09-23 20:09:29,627 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=336224.0, ans=0.0 2024-09-23 20:09:58,362 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=336317.3333333333, ans=0.0 2024-09-23 20:10:07,580 INFO [train.py:1198] (2/4) Epoch 19, batch 1950, loss[loss=0.1693, ctc_loss=0.1108, cr_loss=0.2927, over 17289.00 frames. ], tot_loss[loss=0.2163, ctc_loss=0.1445, cr_loss=0.3587, over 3360109.09 frames. ], batch size: 42, lr: 6.32e-03, grad_scale: 32.0 2024-09-23 20:10:10,844 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.098e+02 1.232e+02 1.325e+02 1.435e+02 2.417e+02, threshold=2.650e+02, percent-clipped=0.0 2024-09-23 20:11:14,086 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=336550.6666666667, ans=0.125 2024-09-23 20:11:15,532 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=336550.6666666667, ans=0.1 2024-09-23 20:11:33,080 INFO [train.py:1198] (2/4) Epoch 19, batch 2000, loss[loss=0.1735, ctc_loss=0.1098, cr_loss=0.3184, over 17251.00 frames. ], tot_loss[loss=0.2162, ctc_loss=0.1446, cr_loss=0.3583, over 3354228.41 frames. ], batch size: 42, lr: 6.32e-03, grad_scale: 32.0 2024-09-23 20:11:42,364 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=7.42 vs. limit=8.0 2024-09-23 20:12:04,627 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=336644.0, ans=0.125 2024-09-23 20:12:27,536 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=3.88 vs. limit=10.0 2024-09-23 20:12:28,247 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=336737.3333333333, ans=0.125 2024-09-23 20:12:44,744 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.56 vs. limit=15.0 2024-09-23 20:12:55,370 INFO [train.py:1198] (2/4) Epoch 19, batch 2050, loss[loss=0.2452, ctc_loss=0.1644, cr_loss=0.4038, over 16478.00 frames. ], tot_loss[loss=0.2161, ctc_loss=0.1444, cr_loss=0.3588, over 3365101.79 frames. ], batch size: 66, lr: 6.32e-03, grad_scale: 32.0 2024-09-23 20:12:58,517 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.061e+02 1.236e+02 1.316e+02 1.436e+02 2.046e+02, threshold=2.631e+02, percent-clipped=0.0 2024-09-23 20:13:23,712 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 20:14:00,882 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=337017.3333333333, ans=0.2 2024-09-23 20:14:18,066 INFO [train.py:1198] (2/4) Epoch 19, batch 2100, loss[loss=0.221, ctc_loss=0.15, cr_loss=0.3551, over 17138.00 frames. ], tot_loss[loss=0.2169, ctc_loss=0.1449, cr_loss=0.3596, over 3359708.91 frames. ], batch size: 48, lr: 6.32e-03, grad_scale: 16.0 2024-09-23 20:14:42,311 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=337110.6666666667, ans=0.1 2024-09-23 20:14:45,520 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=337110.6666666667, ans=0.0 2024-09-23 20:14:56,841 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=337157.3333333333, ans=0.125 2024-09-23 20:15:37,917 INFO [train.py:1198] (2/4) Epoch 19, batch 2150, loss[loss=0.2158, ctc_loss=0.1467, cr_loss=0.3453, over 17239.00 frames. ], tot_loss[loss=0.2173, ctc_loss=0.1454, cr_loss=0.3597, over 3357363.55 frames. ], batch size: 55, lr: 6.32e-03, grad_scale: 8.0 2024-09-23 20:15:44,256 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.099e+02 1.267e+02 1.340e+02 1.517e+02 2.263e+02, threshold=2.680e+02, percent-clipped=0.0 2024-09-23 20:16:20,654 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=337390.6666666667, ans=0.0 2024-09-23 20:16:20,659 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=337390.6666666667, ans=0.125 2024-09-23 20:16:26,709 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=337390.6666666667, ans=0.125 2024-09-23 20:17:05,771 INFO [train.py:1198] (2/4) Epoch 19, batch 2200, loss[loss=0.2237, ctc_loss=0.1523, cr_loss=0.3571, over 17316.00 frames. ], tot_loss[loss=0.2172, ctc_loss=0.1452, cr_loss=0.3597, over 3363134.39 frames. ], batch size: 46, lr: 6.31e-03, grad_scale: 8.0 2024-09-23 20:17:09,274 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=337530.6666666667, ans=0.1 2024-09-23 20:17:22,119 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=337577.3333333333, ans=0.2 2024-09-23 20:18:01,236 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=337670.6666666667, ans=0.125 2024-09-23 20:18:28,142 INFO [train.py:1198] (2/4) Epoch 19, batch 2250, loss[loss=0.2226, ctc_loss=0.1448, cr_loss=0.3893, over 17243.00 frames. ], tot_loss[loss=0.2173, ctc_loss=0.1452, cr_loss=0.3607, over 3366233.35 frames. ], batch size: 44, lr: 6.31e-03, grad_scale: 8.0 2024-09-23 20:18:30,135 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=337764.0, ans=0.1 2024-09-23 20:18:34,522 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.290e+02 1.402e+02 1.494e+02 5.352e+02, threshold=2.803e+02, percent-clipped=1.0 2024-09-23 20:18:38,580 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.11 vs. limit=15.0 2024-09-23 20:18:42,734 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=337810.6666666667, ans=0.125 2024-09-23 20:18:45,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=337810.6666666667, ans=0.125 2024-09-23 20:19:02,269 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.17 vs. limit=6.0 2024-09-23 20:19:37,299 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=337950.6666666667, ans=0.0 2024-09-23 20:19:40,782 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=337950.6666666667, ans=0.0 2024-09-23 20:19:48,403 INFO [train.py:1198] (2/4) Epoch 19, batch 2300, loss[loss=0.2867, ctc_loss=0.207, cr_loss=0.3983, over 11778.00 frames. ], tot_loss[loss=0.2172, ctc_loss=0.1452, cr_loss=0.3599, over 3362123.50 frames. ], batch size: 123, lr: 6.31e-03, grad_scale: 8.0 2024-09-23 20:19:51,190 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=12.92 vs. limit=22.5 2024-09-23 20:20:03,938 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=5.09 vs. limit=10.0 2024-09-23 20:20:14,721 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=10.93 vs. limit=15.0 2024-09-23 20:21:09,226 INFO [train.py:1198] (2/4) Epoch 19, batch 2350, loss[loss=0.2113, ctc_loss=0.1433, cr_loss=0.3399, over 17167.00 frames. ], tot_loss[loss=0.2168, ctc_loss=0.1449, cr_loss=0.3593, over 3358302.60 frames. ], batch size: 45, lr: 6.31e-03, grad_scale: 8.0 2024-09-23 20:21:18,094 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.113e+02 1.269e+02 1.353e+02 1.497e+02 2.252e+02, threshold=2.706e+02, percent-clipped=0.0 2024-09-23 20:21:19,279 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.71 vs. limit=15.0 2024-09-23 20:22:06,576 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=338370.6666666667, ans=0.0 2024-09-23 20:22:16,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=338370.6666666667, ans=0.1 2024-09-23 20:22:17,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=338370.6666666667, ans=0.125 2024-09-23 20:22:25,888 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=338417.3333333333, ans=0.125 2024-09-23 20:22:36,778 INFO [train.py:1198] (2/4) Epoch 19, batch 2400, loss[loss=0.2426, ctc_loss=0.1633, cr_loss=0.3962, over 17042.00 frames. ], tot_loss[loss=0.216, ctc_loss=0.1443, cr_loss=0.3586, over 3361440.24 frames. ], batch size: 52, lr: 6.31e-03, grad_scale: 16.0 2024-09-23 20:22:38,697 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=338464.0, ans=0.0 2024-09-23 20:23:01,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=338510.6666666667, ans=0.125 2024-09-23 20:23:08,024 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=338510.6666666667, ans=0.0 2024-09-23 20:23:16,585 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.38 vs. limit=6.0 2024-09-23 20:23:43,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=338650.6666666667, ans=0.1 2024-09-23 20:23:50,095 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=338650.6666666667, ans=0.125 2024-09-23 20:23:53,275 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=338650.6666666667, ans=0.07 2024-09-23 20:23:59,488 INFO [train.py:1198] (2/4) Epoch 19, batch 2450, loss[loss=0.2007, ctc_loss=0.1346, cr_loss=0.3306, over 17256.00 frames. ], tot_loss[loss=0.2159, ctc_loss=0.1442, cr_loss=0.3585, over 3365206.68 frames. ], batch size: 42, lr: 6.30e-03, grad_scale: 16.0 2024-09-23 20:24:05,811 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.149e+02 1.266e+02 1.363e+02 1.497e+02 1.978e+02, threshold=2.725e+02, percent-clipped=0.0 2024-09-23 20:24:30,386 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.55 vs. limit=15.0 2024-09-23 20:24:49,857 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=13.17 vs. limit=22.5 2024-09-23 20:25:19,141 INFO [train.py:1198] (2/4) Epoch 19, batch 2500, loss[loss=0.2216, ctc_loss=0.1491, cr_loss=0.3625, over 17086.00 frames. ], tot_loss[loss=0.217, ctc_loss=0.145, cr_loss=0.3599, over 3354714.03 frames. ], batch size: 46, lr: 6.30e-03, grad_scale: 16.0 2024-09-23 20:25:19,757 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=8.40 vs. limit=12.0 2024-09-23 20:25:42,294 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=10.83 vs. limit=15.0 2024-09-23 20:25:47,049 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.96 vs. limit=15.0 2024-09-23 20:25:57,835 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=339024.0, ans=0.125 2024-09-23 20:26:39,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=339117.3333333333, ans=0.125 2024-09-23 20:26:39,313 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=339117.3333333333, ans=0.2 2024-09-23 20:26:43,705 INFO [train.py:1198] (2/4) Epoch 19, batch 2550, loss[loss=0.2059, ctc_loss=0.1359, cr_loss=0.3502, over 17070.00 frames. ], tot_loss[loss=0.2182, ctc_loss=0.1458, cr_loss=0.362, over 3360642.01 frames. ], batch size: 46, lr: 6.30e-03, grad_scale: 16.0 2024-09-23 20:26:44,081 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=339164.0, ans=0.125 2024-09-23 20:26:50,570 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=7.64 vs. limit=15.0 2024-09-23 20:26:52,585 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.116e+02 1.236e+02 1.315e+02 1.425e+02 2.139e+02, threshold=2.630e+02, percent-clipped=0.0 2024-09-23 20:27:15,150 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=339210.6666666667, ans=0.125 2024-09-23 20:27:18,344 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=339257.3333333333, ans=0.1 2024-09-23 20:27:35,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=339304.0, ans=0.1 2024-09-23 20:27:40,605 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=339304.0, ans=0.125 2024-09-23 20:27:45,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=339304.0, ans=0.0 2024-09-23 20:27:53,261 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten.whitening_limit, batch_count=339350.6666666667, ans=15.0 2024-09-23 20:28:08,357 INFO [train.py:1198] (2/4) Epoch 19, batch 2600, loss[loss=0.2325, ctc_loss=0.1559, cr_loss=0.3832, over 17216.00 frames. ], tot_loss[loss=0.2184, ctc_loss=0.1459, cr_loss=0.3624, over 3369526.60 frames. ], batch size: 55, lr: 6.30e-03, grad_scale: 16.0 2024-09-23 20:28:14,902 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.max_abs, batch_count=339397.3333333333, ans=10.0 2024-09-23 20:28:37,333 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=339444.0, ans=0.035 2024-09-23 20:28:43,847 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=339490.6666666667, ans=0.2 2024-09-23 20:29:05,984 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=339537.3333333333, ans=0.125 2024-09-23 20:29:20,441 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=339584.0, ans=0.1 2024-09-23 20:29:22,118 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=339584.0, ans=0.125 2024-09-23 20:29:28,190 INFO [train.py:1198] (2/4) Epoch 19, batch 2650, loss[loss=0.2165, ctc_loss=0.1473, cr_loss=0.3458, over 16963.00 frames. ], tot_loss[loss=0.2187, ctc_loss=0.1462, cr_loss=0.3623, over 3363067.84 frames. ], batch size: 42, lr: 6.29e-03, grad_scale: 16.0 2024-09-23 20:29:31,678 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=339630.6666666667, ans=0.125 2024-09-23 20:29:33,647 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.whiten, num_groups=1, num_channels=512, metric=4.87 vs. limit=12.0 2024-09-23 20:29:34,367 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.106e+02 1.315e+02 1.464e+02 1.614e+02 2.211e+02, threshold=2.927e+02, percent-clipped=0.0 2024-09-23 20:29:41,114 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=339630.6666666667, ans=0.125 2024-09-23 20:29:50,743 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=339677.3333333333, ans=0.0 2024-09-23 20:30:10,777 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.33 vs. limit=15.0 2024-09-23 20:30:13,445 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=339724.0, ans=0.0 2024-09-23 20:30:48,089 INFO [train.py:1198] (2/4) Epoch 19, batch 2700, loss[loss=0.2567, ctc_loss=0.1738, cr_loss=0.4146, over 14956.00 frames. ], tot_loss[loss=0.2174, ctc_loss=0.1453, cr_loss=0.3604, over 3364485.63 frames. ], batch size: 89, lr: 6.29e-03, grad_scale: 16.0 2024-09-23 20:31:10,295 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=339910.6666666667, ans=0.0 2024-09-23 20:31:17,934 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=339910.6666666667, ans=0.2 2024-09-23 20:31:24,181 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=339957.3333333333, ans=0.0 2024-09-23 20:31:30,455 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=339957.3333333333, ans=0.125 2024-09-23 20:32:11,626 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=340050.6666666667, ans=0.1 2024-09-23 20:32:13,160 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=340050.6666666667, ans=0.1 2024-09-23 20:32:16,073 INFO [train.py:1198] (2/4) Epoch 19, batch 2750, loss[loss=0.2107, ctc_loss=0.1424, cr_loss=0.3415, over 17230.00 frames. ], tot_loss[loss=0.2169, ctc_loss=0.145, cr_loss=0.3599, over 3364724.23 frames. ], batch size: 50, lr: 6.29e-03, grad_scale: 16.0 2024-09-23 20:32:19,885 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=16.98 vs. limit=22.5 2024-09-23 20:32:22,222 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.266e+02 1.341e+02 1.484e+02 3.814e+02, threshold=2.681e+02, percent-clipped=1.0 2024-09-23 20:32:27,309 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=340097.3333333333, ans=0.2 2024-09-23 20:32:33,782 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=340144.0, ans=0.1 2024-09-23 20:32:58,683 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=340190.6666666667, ans=0.125 2024-09-23 20:33:00,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=340190.6666666667, ans=0.0 2024-09-23 20:33:16,338 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=340237.3333333333, ans=0.125 2024-09-23 20:33:38,449 INFO [train.py:1198] (2/4) Epoch 19, batch 2800, loss[loss=0.1827, ctc_loss=0.1211, cr_loss=0.3079, over 17095.00 frames. ], tot_loss[loss=0.2168, ctc_loss=0.1449, cr_loss=0.3596, over 3367386.58 frames. ], batch size: 40, lr: 6.29e-03, grad_scale: 32.0 2024-09-23 20:33:45,026 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=340330.6666666667, ans=0.025 2024-09-23 20:33:51,866 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.17 vs. limit=6.0 2024-09-23 20:34:07,643 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.41 vs. limit=6.0 2024-09-23 20:34:36,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=340470.6666666667, ans=0.95 2024-09-23 20:34:47,186 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=340517.3333333333, ans=0.0 2024-09-23 20:34:58,053 INFO [train.py:1198] (2/4) Epoch 19, batch 2850, loss[loss=0.2522, ctc_loss=0.1755, cr_loss=0.3831, over 16626.00 frames. ], tot_loss[loss=0.2168, ctc_loss=0.1449, cr_loss=0.3593, over 3364362.65 frames. ], batch size: 61, lr: 6.29e-03, grad_scale: 32.0 2024-09-23 20:35:04,575 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.252e+02 1.382e+02 1.524e+02 2.298e+02, threshold=2.764e+02, percent-clipped=0.0 2024-09-23 20:35:22,798 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.28 vs. limit=15.0 2024-09-23 20:35:39,863 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=340657.3333333333, ans=0.2 2024-09-23 20:35:59,641 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=340704.0, ans=0.0 2024-09-23 20:36:16,667 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=340750.6666666667, ans=0.125 2024-09-23 20:36:22,620 INFO [train.py:1198] (2/4) Epoch 19, batch 2900, loss[loss=0.2189, ctc_loss=0.1462, cr_loss=0.3635, over 17015.00 frames. ], tot_loss[loss=0.2169, ctc_loss=0.1451, cr_loss=0.3592, over 3362404.00 frames. ], batch size: 44, lr: 6.28e-03, grad_scale: 16.0 2024-09-23 20:37:38,967 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=340984.0, ans=0.5 2024-09-23 20:37:47,690 INFO [train.py:1198] (2/4) Epoch 19, batch 2950, loss[loss=0.2413, ctc_loss=0.1615, cr_loss=0.399, over 17227.00 frames. ], tot_loss[loss=0.2181, ctc_loss=0.1458, cr_loss=0.3616, over 3363505.25 frames. ], batch size: 55, lr: 6.28e-03, grad_scale: 16.0 2024-09-23 20:37:55,446 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.075e+02 1.260e+02 1.402e+02 1.500e+02 2.241e+02, threshold=2.804e+02, percent-clipped=0.0 2024-09-23 20:38:07,072 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=341077.3333333333, ans=0.125 2024-09-23 20:38:49,702 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=341217.3333333333, ans=0.125 2024-09-23 20:38:52,838 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer_ff2.min_abs, batch_count=341217.3333333333, ans=0.1 2024-09-23 20:39:00,641 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=341217.3333333333, ans=0.04949747468305833 2024-09-23 20:39:06,540 INFO [train.py:1198] (2/4) Epoch 19, batch 3000, loss[loss=0.2039, ctc_loss=0.1375, cr_loss=0.3321, over 17011.00 frames. ], tot_loss[loss=0.218, ctc_loss=0.1457, cr_loss=0.3616, over 3361026.92 frames. ], batch size: 53, lr: 6.28e-03, grad_scale: 16.0 2024-09-23 20:39:06,540 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 20:39:18,436 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.1.encoder.layers.1.self_attn_weights, attn_weights_entropy = tensor([5.4499, 5.2350, 4.4245, 5.1170], device='cuda:2') 2024-09-23 20:39:21,850 INFO [train.py:1230] (2/4) Epoch 19, validation: loss=0.03984, ctc_loss=0.03984, cr_loss=8.01e-15, over 944034.00 frames. 2024-09-23 20:39:21,851 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 20:40:27,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=341450.6666666667, ans=0.0 2024-09-23 20:40:35,937 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=341450.6666666667, ans=0.125 2024-09-23 20:40:37,363 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=341450.6666666667, ans=0.125 2024-09-23 20:40:40,304 INFO [train.py:1198] (2/4) Epoch 19, batch 3050, loss[loss=0.2056, ctc_loss=0.1348, cr_loss=0.3536, over 17306.00 frames. ], tot_loss[loss=0.2182, ctc_loss=0.1458, cr_loss=0.362, over 3355668.99 frames. ], batch size: 49, lr: 6.28e-03, grad_scale: 16.0 2024-09-23 20:40:45,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=341497.3333333333, ans=0.0 2024-09-23 20:40:48,106 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.094e+02 1.265e+02 1.361e+02 1.501e+02 2.045e+02, threshold=2.721e+02, percent-clipped=0.0 2024-09-23 20:41:20,152 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=341590.6666666667, ans=0.5 2024-09-23 20:41:26,539 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=341637.3333333333, ans=0.125 2024-09-23 20:41:56,656 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=341684.0, ans=0.125 2024-09-23 20:41:59,534 INFO [train.py:1198] (2/4) Epoch 19, batch 3100, loss[loss=0.2165, ctc_loss=0.1441, cr_loss=0.362, over 17259.00 frames. ], tot_loss[loss=0.2187, ctc_loss=0.1462, cr_loss=0.3625, over 3367215.24 frames. ], batch size: 44, lr: 6.28e-03, grad_scale: 16.0 2024-09-23 20:42:18,687 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=341777.3333333333, ans=0.0 2024-09-23 20:42:23,321 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=341777.3333333333, ans=0.2 2024-09-23 20:42:38,312 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.whiten.whitening_limit, batch_count=341824.0, ans=15.0 2024-09-23 20:42:47,017 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=341870.6666666667, ans=0.0 2024-09-23 20:42:52,179 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.85 vs. limit=15.0 2024-09-23 20:43:05,314 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=5.37 vs. limit=12.0 2024-09-23 20:43:06,430 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=341917.3333333333, ans=0.1 2024-09-23 20:43:14,162 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=341917.3333333333, ans=0.1 2024-09-23 20:43:15,623 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=341917.3333333333, ans=0.1 2024-09-23 20:43:15,684 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=341917.3333333333, ans=0.125 2024-09-23 20:43:20,310 INFO [train.py:1198] (2/4) Epoch 19, batch 3150, loss[loss=0.1758, ctc_loss=0.1146, cr_loss=0.3056, over 16728.00 frames. ], tot_loss[loss=0.2196, ctc_loss=0.1469, cr_loss=0.3634, over 3362209.57 frames. ], batch size: 37, lr: 6.27e-03, grad_scale: 16.0 2024-09-23 20:43:26,810 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=341964.0, ans=0.125 2024-09-23 20:43:30,466 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.288e+02 1.385e+02 1.538e+02 2.696e+02, threshold=2.771e+02, percent-clipped=0.0 2024-09-23 20:43:52,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=342057.3333333333, ans=0.1 2024-09-23 20:43:59,162 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.min_positive, batch_count=342057.3333333333, ans=0.05 2024-09-23 20:44:10,626 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=342104.0, ans=0.0 2024-09-23 20:44:12,140 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=342104.0, ans=0.0 2024-09-23 20:44:26,196 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=342150.6666666667, ans=0.0 2024-09-23 20:44:36,511 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=342150.6666666667, ans=0.125 2024-09-23 20:44:44,043 INFO [train.py:1198] (2/4) Epoch 19, batch 3200, loss[loss=0.2106, ctc_loss=0.1404, cr_loss=0.3512, over 17020.00 frames. ], tot_loss[loss=0.2185, ctc_loss=0.146, cr_loss=0.3628, over 3360486.37 frames. ], batch size: 44, lr: 6.27e-03, grad_scale: 32.0 2024-09-23 20:44:55,048 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=342197.3333333333, ans=0.125 2024-09-23 20:45:06,369 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=342244.0, ans=0.0 2024-09-23 20:45:13,927 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=342290.6666666667, ans=0.125 2024-09-23 20:45:58,944 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=342384.0, ans=0.125 2024-09-23 20:46:01,866 INFO [train.py:1198] (2/4) Epoch 19, batch 3250, loss[loss=0.2424, ctc_loss=0.163, cr_loss=0.3973, over 17064.00 frames. ], tot_loss[loss=0.2192, ctc_loss=0.1464, cr_loss=0.3638, over 3351865.79 frames. ], batch size: 56, lr: 6.27e-03, grad_scale: 16.0 2024-09-23 20:46:11,249 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.085e+02 1.296e+02 1.376e+02 1.501e+02 2.697e+02, threshold=2.752e+02, percent-clipped=0.0 2024-09-23 20:46:19,979 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=7.48 vs. limit=15.0 2024-09-23 20:46:25,516 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=342477.3333333333, ans=0.0 2024-09-23 20:46:40,420 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=342524.0, ans=0.0 2024-09-23 20:46:54,314 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=342570.6666666667, ans=0.1 2024-09-23 20:47:00,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=342570.6666666667, ans=0.5 2024-09-23 20:47:10,145 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=342617.3333333333, ans=0.2 2024-09-23 20:47:22,259 INFO [train.py:1198] (2/4) Epoch 19, batch 3300, loss[loss=0.222, ctc_loss=0.1452, cr_loss=0.3838, over 16565.00 frames. ], tot_loss[loss=0.2175, ctc_loss=0.1451, cr_loss=0.3618, over 3366675.63 frames. ], batch size: 66, lr: 6.27e-03, grad_scale: 16.0 2024-09-23 20:47:51,462 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.35 vs. limit=12.0 2024-09-23 20:47:58,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=342757.3333333333, ans=0.035 2024-09-23 20:48:00,516 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=342757.3333333333, ans=0.125 2024-09-23 20:48:19,420 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=342804.0, ans=0.0 2024-09-23 20:48:25,660 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=342850.6666666667, ans=0.125 2024-09-23 20:48:41,090 INFO [train.py:1198] (2/4) Epoch 19, batch 3350, loss[loss=0.2126, ctc_loss=0.1404, cr_loss=0.3607, over 17161.00 frames. ], tot_loss[loss=0.2175, ctc_loss=0.1452, cr_loss=0.3616, over 3351895.93 frames. ], batch size: 45, lr: 6.26e-03, grad_scale: 16.0 2024-09-23 20:48:42,889 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=342897.3333333333, ans=0.125 2024-09-23 20:48:44,573 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 20:48:47,626 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=342897.3333333333, ans=0.05 2024-09-23 20:48:50,525 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.035e+02 1.298e+02 1.381e+02 1.515e+02 2.027e+02, threshold=2.761e+02, percent-clipped=0.0 2024-09-23 20:49:10,224 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.73 vs. limit=12.0 2024-09-23 20:49:39,587 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=343037.3333333333, ans=0.0 2024-09-23 20:49:59,660 INFO [train.py:1198] (2/4) Epoch 19, batch 3400, loss[loss=0.1663, ctc_loss=0.1071, cr_loss=0.2958, over 17045.00 frames. ], tot_loss[loss=0.2181, ctc_loss=0.1458, cr_loss=0.3616, over 3345705.61 frames. ], batch size: 39, lr: 6.26e-03, grad_scale: 16.0 2024-09-23 20:50:10,914 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=343130.6666666667, ans=0.125 2024-09-23 20:50:48,941 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.29 vs. limit=15.0 2024-09-23 20:51:01,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=343317.3333333333, ans=0.125 2024-09-23 20:51:04,262 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=343317.3333333333, ans=0.0 2024-09-23 20:51:17,913 INFO [train.py:1198] (2/4) Epoch 19, batch 3450, loss[loss=0.2333, ctc_loss=0.1573, cr_loss=0.3803, over 16996.00 frames. ], tot_loss[loss=0.2184, ctc_loss=0.146, cr_loss=0.362, over 3344734.90 frames. ], batch size: 56, lr: 6.26e-03, grad_scale: 16.0 2024-09-23 20:51:24,504 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=343364.0, ans=0.125 2024-09-23 20:51:26,141 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=343364.0, ans=0.0 2024-09-23 20:51:27,572 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.110e+02 1.246e+02 1.361e+02 1.474e+02 2.072e+02, threshold=2.723e+02, percent-clipped=0.0 2024-09-23 20:51:38,751 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=343410.6666666667, ans=0.0 2024-09-23 20:51:54,530 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=343457.3333333333, ans=0.05 2024-09-23 20:52:36,230 INFO [train.py:1198] (2/4) Epoch 19, batch 3500, loss[loss=0.2124, ctc_loss=0.1419, cr_loss=0.3528, over 16766.00 frames. ], tot_loss[loss=0.218, ctc_loss=0.1457, cr_loss=0.3614, over 3347045.47 frames. ], batch size: 61, lr: 6.26e-03, grad_scale: 16.0 2024-09-23 20:52:42,832 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=343597.3333333333, ans=0.125 2024-09-23 20:52:53,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=343644.0, ans=0.1 2024-09-23 20:52:59,504 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=343644.0, ans=0.125 2024-09-23 20:53:02,757 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=343644.0, ans=0.125 2024-09-23 20:53:17,259 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=343690.6666666667, ans=0.1 2024-09-23 20:53:19,043 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=343690.6666666667, ans=0.0 2024-09-23 20:53:57,647 INFO [train.py:1198] (2/4) Epoch 19, batch 3550, loss[loss=0.2703, ctc_loss=0.1876, cr_loss=0.4134, over 17005.00 frames. ], tot_loss[loss=0.2178, ctc_loss=0.1456, cr_loss=0.3612, over 3356984.03 frames. ], batch size: 53, lr: 6.26e-03, grad_scale: 16.0 2024-09-23 20:54:07,040 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.247e+02 1.326e+02 1.445e+02 2.020e+02, threshold=2.653e+02, percent-clipped=0.0 2024-09-23 20:54:12,019 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=343877.3333333333, ans=0.2 2024-09-23 20:54:20,061 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=343877.3333333333, ans=0.125 2024-09-23 20:54:32,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=343924.0, ans=0.04949747468305833 2024-09-23 20:55:07,250 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.19 vs. limit=15.0 2024-09-23 20:55:17,372 INFO [train.py:1198] (2/4) Epoch 19, batch 3600, loss[loss=0.2591, ctc_loss=0.1769, cr_loss=0.4115, over 16571.00 frames. ], tot_loss[loss=0.2171, ctc_loss=0.1452, cr_loss=0.3598, over 3360766.86 frames. ], batch size: 66, lr: 6.25e-03, grad_scale: 32.0 2024-09-23 20:55:31,862 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=8.65 vs. limit=15.0 2024-09-23 20:55:33,263 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 20:55:34,682 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=344110.6666666667, ans=0.125 2024-09-23 20:55:47,975 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.92 vs. limit=15.0 2024-09-23 20:56:03,984 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=344204.0, ans=0.2 2024-09-23 20:56:04,003 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=344204.0, ans=0.0 2024-09-23 20:56:07,058 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=344204.0, ans=0.0 2024-09-23 20:56:36,797 INFO [train.py:1198] (2/4) Epoch 19, batch 3650, loss[loss=0.2495, ctc_loss=0.1662, cr_loss=0.4165, over 17020.00 frames. ], tot_loss[loss=0.2158, ctc_loss=0.1442, cr_loss=0.3583, over 3365480.07 frames. ], batch size: 53, lr: 6.25e-03, grad_scale: 32.0 2024-09-23 20:56:44,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=344297.3333333333, ans=0.125 2024-09-23 20:56:47,598 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.165e+02 1.302e+02 1.379e+02 1.507e+02 2.534e+02, threshold=2.758e+02, percent-clipped=0.0 2024-09-23 20:56:58,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=344344.0, ans=0.0 2024-09-23 20:57:06,778 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=344390.6666666667, ans=0.125 2024-09-23 20:57:36,037 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=344437.3333333333, ans=0.125 2024-09-23 20:57:51,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=344484.0, ans=0.125 2024-09-23 20:57:54,992 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=344530.6666666667, ans=0.2 2024-09-23 20:57:56,273 INFO [train.py:1198] (2/4) Epoch 19, batch 3700, loss[loss=0.2026, ctc_loss=0.1334, cr_loss=0.3458, over 17291.00 frames. ], tot_loss[loss=0.2166, ctc_loss=0.1446, cr_loss=0.3598, over 3358462.90 frames. ], batch size: 46, lr: 6.25e-03, grad_scale: 16.0 2024-09-23 20:58:12,282 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=344577.3333333333, ans=0.2 2024-09-23 20:58:20,101 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=344577.3333333333, ans=0.1 2024-09-23 20:58:43,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.min_positive, batch_count=344670.6666666667, ans=0.025 2024-09-23 20:58:44,192 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.24 vs. limit=15.0 2024-09-23 20:58:58,514 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.80 vs. limit=12.0 2024-09-23 20:59:14,535 INFO [train.py:1198] (2/4) Epoch 19, batch 3750, loss[loss=0.2865, ctc_loss=0.2045, cr_loss=0.4098, over 12349.00 frames. ], tot_loss[loss=0.2164, ctc_loss=0.1444, cr_loss=0.3597, over 3357431.27 frames. ], batch size: 124, lr: 6.25e-03, grad_scale: 16.0 2024-09-23 20:59:25,491 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.059e+02 1.321e+02 1.412e+02 1.562e+02 2.185e+02, threshold=2.824e+02, percent-clipped=0.0 2024-09-23 20:59:27,902 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=512, metric=3.85 vs. limit=15.0 2024-09-23 20:59:46,293 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.82 vs. limit=15.0 2024-09-23 21:00:01,513 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 21:00:32,421 INFO [train.py:1198] (2/4) Epoch 19, batch 3800, loss[loss=0.1716, ctc_loss=0.1103, cr_loss=0.3067, over 17046.00 frames. ], tot_loss[loss=0.2164, ctc_loss=0.1446, cr_loss=0.3589, over 3323750.37 frames. ], batch size: 39, lr: 6.25e-03, grad_scale: 16.0 2024-09-23 21:00:34,955 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.44 vs. limit=6.0 2024-09-23 21:01:00,786 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=345044.0, ans=0.0 2024-09-23 21:01:03,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=345090.6666666667, ans=0.0 2024-09-23 21:01:22,821 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=345137.3333333333, ans=0.1 2024-09-23 21:01:41,522 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=345184.0, ans=0.1 2024-09-23 21:01:50,702 INFO [train.py:1198] (2/4) Epoch 19, batch 3850, loss[loss=0.2465, ctc_loss=0.1671, cr_loss=0.3969, over 14977.00 frames. ], tot_loss[loss=0.2179, ctc_loss=0.1459, cr_loss=0.3598, over 3270719.97 frames. ], batch size: 89, lr: 6.24e-03, grad_scale: 16.0 2024-09-23 21:01:50,922 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=345230.6666666667, ans=0.125 2024-09-23 21:02:01,516 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.310e+02 1.460e+02 1.598e+02 2.355e+02, threshold=2.920e+02, percent-clipped=0.0 2024-09-23 21:02:09,967 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=345277.3333333333, ans=0.125 2024-09-23 21:02:43,603 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=345370.6666666667, ans=0.1 2024-09-23 21:02:56,700 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.91 vs. limit=10.0 2024-09-23 21:03:53,556 INFO [train.py:1198] (2/4) Epoch 20, batch 0, loss[loss=0.2064, ctc_loss=0.139, cr_loss=0.3366, over 17274.00 frames. ], tot_loss[loss=0.2064, ctc_loss=0.139, cr_loss=0.3366, over 17274.00 frames. ], batch size: 42, lr: 6.08e-03, grad_scale: 32.0 2024-09-23 21:03:53,557 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 21:04:08,657 INFO [train.py:1230] (2/4) Epoch 20, validation: loss=0.03935, ctc_loss=0.03935, cr_loss=7.664e-15, over 944034.00 frames. 2024-09-23 21:04:08,658 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 21:04:27,732 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=345492.0, ans=0.0 2024-09-23 21:04:30,713 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=345492.0, ans=0.0 2024-09-23 21:05:10,464 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=345585.3333333333, ans=0.125 2024-09-23 21:05:15,140 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=345585.3333333333, ans=0.125 2024-09-23 21:05:17,467 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=13.04 vs. limit=15.0 2024-09-23 21:05:33,923 INFO [train.py:1198] (2/4) Epoch 20, batch 50, loss[loss=0.2261, ctc_loss=0.1505, cr_loss=0.378, over 17341.00 frames. ], tot_loss[loss=0.2219, ctc_loss=0.1482, cr_loss=0.3684, over 757284.90 frames. ], batch size: 48, lr: 6.08e-03, grad_scale: 32.0 2024-09-23 21:05:45,637 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.02 vs. limit=15.0 2024-09-23 21:05:51,260 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.032e+02 1.262e+02 1.457e+02 1.636e+02 2.185e+02, threshold=2.915e+02, percent-clipped=0.0 2024-09-23 21:06:31,637 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=345818.6666666667, ans=0.1 2024-09-23 21:06:53,841 INFO [train.py:1198] (2/4) Epoch 20, batch 100, loss[loss=0.1922, ctc_loss=0.1292, cr_loss=0.3149, over 17076.00 frames. ], tot_loss[loss=0.2214, ctc_loss=0.1477, cr_loss=0.3683, over 1342072.52 frames. ], batch size: 43, lr: 6.08e-03, grad_scale: 32.0 2024-09-23 21:06:57,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=345912.0, ans=0.0 2024-09-23 21:07:02,125 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=345912.0, ans=0.0 2024-09-23 21:07:19,818 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=345958.6666666667, ans=0.125 2024-09-23 21:07:20,324 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.37 vs. limit=6.0 2024-09-23 21:07:26,156 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=346005.3333333333, ans=0.125 2024-09-23 21:07:33,913 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=346005.3333333333, ans=0.125 2024-09-23 21:07:43,083 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=346052.0, ans=0.1 2024-09-23 21:07:46,967 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=6.30 vs. limit=15.0 2024-09-23 21:08:18,775 INFO [train.py:1198] (2/4) Epoch 20, batch 150, loss[loss=0.2027, ctc_loss=0.1309, cr_loss=0.3591, over 17121.00 frames. ], tot_loss[loss=0.2193, ctc_loss=0.1462, cr_loss=0.3655, over 1782555.31 frames. ], batch size: 40, lr: 6.07e-03, grad_scale: 16.0 2024-09-23 21:08:24,378 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=7.04 vs. limit=15.0 2024-09-23 21:08:36,814 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=346192.0, ans=0.0 2024-09-23 21:08:38,025 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.021e+02 1.321e+02 1.429e+02 1.602e+02 2.448e+02, threshold=2.858e+02, percent-clipped=0.0 2024-09-23 21:08:59,403 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.70 vs. limit=15.0 2024-09-23 21:09:05,763 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=346285.3333333333, ans=0.125 2024-09-23 21:09:38,947 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=346332.0, ans=0.125 2024-09-23 21:09:45,099 INFO [train.py:1198] (2/4) Epoch 20, batch 200, loss[loss=0.2047, ctc_loss=0.1346, cr_loss=0.3505, over 17251.00 frames. ], tot_loss[loss=0.2187, ctc_loss=0.1457, cr_loss=0.3653, over 2137858.71 frames. ], batch size: 42, lr: 6.07e-03, grad_scale: 16.0 2024-09-23 21:09:54,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=346378.6666666667, ans=0.1 2024-09-23 21:10:07,592 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=346425.3333333333, ans=0.2 2024-09-23 21:10:22,325 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.98 vs. limit=15.0 2024-09-23 21:10:34,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=346518.6666666667, ans=10.0 2024-09-23 21:10:49,124 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=346565.3333333333, ans=0.125 2024-09-23 21:11:04,972 INFO [train.py:1198] (2/4) Epoch 20, batch 250, loss[loss=0.2, ctc_loss=0.1312, cr_loss=0.3441, over 17260.00 frames. ], tot_loss[loss=0.217, ctc_loss=0.1444, cr_loss=0.3626, over 2413553.43 frames. ], batch size: 42, lr: 6.07e-03, grad_scale: 16.0 2024-09-23 21:11:06,256 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.62 vs. limit=5.0 2024-09-23 21:11:23,903 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.297e+02 1.428e+02 1.617e+02 1.854e+02, threshold=2.857e+02, percent-clipped=0.0 2024-09-23 21:11:33,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=346658.6666666667, ans=0.125 2024-09-23 21:12:02,241 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=346752.0, ans=0.1 2024-09-23 21:12:23,149 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=346845.3333333333, ans=0.07 2024-09-23 21:12:24,538 INFO [train.py:1198] (2/4) Epoch 20, batch 300, loss[loss=0.2112, ctc_loss=0.1421, cr_loss=0.3457, over 17008.00 frames. ], tot_loss[loss=0.2163, ctc_loss=0.144, cr_loss=0.3614, over 2626764.15 frames. ], batch size: 51, lr: 6.07e-03, grad_scale: 16.0 2024-09-23 21:12:31,164 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.min_positive, batch_count=346845.3333333333, ans=0.025 2024-09-23 21:12:56,242 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=346892.0, ans=0.125 2024-09-23 21:13:02,553 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=346938.6666666667, ans=0.0 2024-09-23 21:13:05,797 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=346938.6666666667, ans=0.025 2024-09-23 21:13:06,143 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.18 vs. limit=15.0 2024-09-23 21:13:32,377 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=347032.0, ans=0.125 2024-09-23 21:13:49,681 INFO [train.py:1198] (2/4) Epoch 20, batch 350, loss[loss=0.2082, ctc_loss=0.1408, cr_loss=0.3371, over 16689.00 frames. ], tot_loss[loss=0.2158, ctc_loss=0.1437, cr_loss=0.3606, over 2801602.16 frames. ], batch size: 61, lr: 6.07e-03, grad_scale: 16.0 2024-09-23 21:14:08,665 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.051e+02 1.225e+02 1.312e+02 1.421e+02 1.795e+02, threshold=2.625e+02, percent-clipped=0.0 2024-09-23 21:14:56,746 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.min_positive, batch_count=347265.3333333333, ans=0.05 2024-09-23 21:14:58,945 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=17.30 vs. limit=22.5 2024-09-23 21:15:03,116 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=347265.3333333333, ans=0.0 2024-09-23 21:15:03,137 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=347265.3333333333, ans=0.125 2024-09-23 21:15:12,641 INFO [train.py:1198] (2/4) Epoch 20, batch 400, loss[loss=0.2306, ctc_loss=0.1554, cr_loss=0.3759, over 17130.00 frames. ], tot_loss[loss=0.217, ctc_loss=0.1445, cr_loss=0.3622, over 2924630.59 frames. ], batch size: 48, lr: 6.06e-03, grad_scale: 32.0 2024-09-23 21:15:25,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=347312.0, ans=0.125 2024-09-23 21:15:40,315 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.19 vs. limit=15.0 2024-09-23 21:15:41,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=347358.6666666667, ans=0.2 2024-09-23 21:16:32,400 INFO [train.py:1198] (2/4) Epoch 20, batch 450, loss[loss=0.2289, ctc_loss=0.1544, cr_loss=0.3722, over 14822.00 frames. ], tot_loss[loss=0.2171, ctc_loss=0.1446, cr_loss=0.3626, over 3030663.16 frames. ], batch size: 89, lr: 6.06e-03, grad_scale: 32.0 2024-09-23 21:16:32,683 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.min_abs, batch_count=347545.3333333333, ans=0.5 2024-09-23 21:16:52,900 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.150e+02 1.293e+02 1.421e+02 1.585e+02 2.168e+02, threshold=2.842e+02, percent-clipped=0.0 2024-09-23 21:17:30,748 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=11.57 vs. limit=15.0 2024-09-23 21:17:47,295 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=347732.0, ans=0.0 2024-09-23 21:17:51,996 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=347732.0, ans=0.0 2024-09-23 21:17:54,852 INFO [train.py:1198] (2/4) Epoch 20, batch 500, loss[loss=0.2106, ctc_loss=0.1429, cr_loss=0.3386, over 16764.00 frames. ], tot_loss[loss=0.216, ctc_loss=0.1439, cr_loss=0.3607, over 3110019.09 frames. ], batch size: 61, lr: 6.06e-03, grad_scale: 16.0 2024-09-23 21:17:55,101 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=347778.6666666667, ans=0.125 2024-09-23 21:17:55,216 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=347778.6666666667, ans=0.125 2024-09-23 21:18:25,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=347825.3333333333, ans=0.125 2024-09-23 21:19:11,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=347965.3333333333, ans=0.1 2024-09-23 21:19:14,281 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=347965.3333333333, ans=0.1 2024-09-23 21:19:17,191 INFO [train.py:1198] (2/4) Epoch 20, batch 550, loss[loss=0.2526, ctc_loss=0.1788, cr_loss=0.369, over 11500.00 frames. ], tot_loss[loss=0.2161, ctc_loss=0.1439, cr_loss=0.3606, over 3160443.48 frames. ], batch size: 123, lr: 6.06e-03, grad_scale: 16.0 2024-09-23 21:19:31,091 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.48 vs. limit=22.5 2024-09-23 21:19:36,823 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 21:19:41,694 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.58 vs. limit=15.0 2024-09-23 21:19:42,496 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.238e+02 1.328e+02 1.432e+02 2.472e+02, threshold=2.656e+02, percent-clipped=0.0 2024-09-23 21:20:41,715 INFO [train.py:1198] (2/4) Epoch 20, batch 600, loss[loss=0.1992, ctc_loss=0.1336, cr_loss=0.3277, over 16114.00 frames. ], tot_loss[loss=0.2151, ctc_loss=0.1433, cr_loss=0.359, over 3206550.51 frames. ], batch size: 74, lr: 6.06e-03, grad_scale: 16.0 2024-09-23 21:20:56,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=348292.0, ans=0.0 2024-09-23 21:20:56,175 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=348292.0, ans=0.125 2024-09-23 21:21:15,927 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.75 vs. limit=15.0 2024-09-23 21:21:23,580 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.94 vs. limit=6.0 2024-09-23 21:21:26,231 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=348338.6666666667, ans=0.0 2024-09-23 21:21:32,623 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=348385.3333333333, ans=0.1 2024-09-23 21:21:46,938 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=348432.0, ans=0.1 2024-09-23 21:21:54,846 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=348432.0, ans=0.125 2024-09-23 21:22:01,010 INFO [train.py:1198] (2/4) Epoch 20, batch 650, loss[loss=0.2167, ctc_loss=0.1426, cr_loss=0.3702, over 17157.00 frames. ], tot_loss[loss=0.215, ctc_loss=0.1432, cr_loss=0.3593, over 3250208.72 frames. ], batch size: 48, lr: 6.05e-03, grad_scale: 16.0 2024-09-23 21:22:21,604 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.109e+02 1.264e+02 1.348e+02 1.477e+02 2.156e+02, threshold=2.696e+02, percent-clipped=0.0 2024-09-23 21:22:28,291 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=348525.3333333333, ans=0.0 2024-09-23 21:22:36,371 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=348572.0, ans=0.2 2024-09-23 21:23:26,231 INFO [train.py:1198] (2/4) Epoch 20, batch 700, loss[loss=0.2228, ctc_loss=0.1504, cr_loss=0.362, over 17004.00 frames. ], tot_loss[loss=0.2159, ctc_loss=0.1437, cr_loss=0.3609, over 3278574.03 frames. ], batch size: 56, lr: 6.05e-03, grad_scale: 16.0 2024-09-23 21:23:30,478 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.48 vs. limit=12.0 2024-09-23 21:23:32,989 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=348712.0, ans=0.125 2024-09-23 21:23:45,722 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=348758.6666666667, ans=0.125 2024-09-23 21:23:48,918 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=348758.6666666667, ans=0.125 2024-09-23 21:23:51,587 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=12.29 vs. limit=15.0 2024-09-23 21:24:08,340 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=348805.3333333333, ans=0.125 2024-09-23 21:24:14,683 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=348852.0, ans=0.125 2024-09-23 21:24:23,775 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=348852.0, ans=0.125 2024-09-23 21:24:25,260 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.min_positive, batch_count=348852.0, ans=0.025 2024-09-23 21:24:26,945 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=348852.0, ans=0.1 2024-09-23 21:24:36,056 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=9.76 vs. limit=15.0 2024-09-23 21:24:51,549 INFO [train.py:1198] (2/4) Epoch 20, batch 750, loss[loss=0.18, ctc_loss=0.1175, cr_loss=0.3121, over 17192.00 frames. ], tot_loss[loss=0.2153, ctc_loss=0.1433, cr_loss=0.36, over 3302438.89 frames. ], batch size: 41, lr: 6.05e-03, grad_scale: 16.0 2024-09-23 21:25:12,208 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.238e+02 1.344e+02 1.442e+02 2.138e+02, threshold=2.688e+02, percent-clipped=0.0 2024-09-23 21:25:14,574 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.02 vs. limit=6.0 2024-09-23 21:25:18,959 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=348992.0, ans=0.125 2024-09-23 21:25:38,046 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=349085.3333333333, ans=0.125 2024-09-23 21:26:00,959 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=23.36 vs. limit=22.5 2024-09-23 21:26:11,141 INFO [train.py:1198] (2/4) Epoch 20, batch 800, loss[loss=0.2145, ctc_loss=0.1385, cr_loss=0.3801, over 17150.00 frames. ], tot_loss[loss=0.2152, ctc_loss=0.1432, cr_loss=0.3601, over 3324336.51 frames. ], batch size: 45, lr: 6.05e-03, grad_scale: 32.0 2024-09-23 21:26:49,588 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=349272.0, ans=0.125 2024-09-23 21:26:56,230 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=349272.0, ans=0.0 2024-09-23 21:27:08,999 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=349318.6666666667, ans=0.0 2024-09-23 21:27:15,471 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=349365.3333333333, ans=0.2 2024-09-23 21:27:18,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=349365.3333333333, ans=0.125 2024-09-23 21:27:20,452 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=5.87 vs. limit=12.0 2024-09-23 21:27:21,759 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=349365.3333333333, ans=0.0 2024-09-23 21:27:31,180 INFO [train.py:1198] (2/4) Epoch 20, batch 850, loss[loss=0.1981, ctc_loss=0.1313, cr_loss=0.3338, over 17106.00 frames. ], tot_loss[loss=0.2143, ctc_loss=0.1425, cr_loss=0.3591, over 3339408.04 frames. ], batch size: 40, lr: 6.05e-03, grad_scale: 32.0 2024-09-23 21:27:43,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=349412.0, ans=0.125 2024-09-23 21:27:46,521 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=349412.0, ans=0.125 2024-09-23 21:27:54,462 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.268e+02 1.363e+02 1.494e+02 2.147e+02, threshold=2.726e+02, percent-clipped=0.0 2024-09-23 21:28:03,312 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=8.66 vs. limit=15.0 2024-09-23 21:28:45,234 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=349598.6666666667, ans=0.07 2024-09-23 21:28:56,003 INFO [train.py:1198] (2/4) Epoch 20, batch 900, loss[loss=0.1973, ctc_loss=0.1291, cr_loss=0.3406, over 17247.00 frames. ], tot_loss[loss=0.2146, ctc_loss=0.1427, cr_loss=0.3595, over 3351781.21 frames. ], batch size: 44, lr: 6.04e-03, grad_scale: 32.0 2024-09-23 21:29:10,787 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=349692.0, ans=0.0 2024-09-23 21:29:17,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=349692.0, ans=0.1 2024-09-23 21:29:33,355 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=349738.6666666667, ans=0.2 2024-09-23 21:29:43,445 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.41 vs. limit=10.0 2024-09-23 21:30:12,282 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=6.14 vs. limit=15.0 2024-09-23 21:30:21,167 INFO [train.py:1198] (2/4) Epoch 20, batch 950, loss[loss=0.2395, ctc_loss=0.158, cr_loss=0.4075, over 17302.00 frames. ], tot_loss[loss=0.2153, ctc_loss=0.1432, cr_loss=0.3603, over 3361097.18 frames. ], batch size: 49, lr: 6.04e-03, grad_scale: 32.0 2024-09-23 21:30:42,123 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.058e+02 1.279e+02 1.392e+02 1.519e+02 1.959e+02, threshold=2.784e+02, percent-clipped=0.0 2024-09-23 21:31:35,332 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=350065.3333333333, ans=0.1 2024-09-23 21:31:41,386 INFO [train.py:1198] (2/4) Epoch 20, batch 1000, loss[loss=0.227, ctc_loss=0.1478, cr_loss=0.396, over 17352.00 frames. ], tot_loss[loss=0.2159, ctc_loss=0.1437, cr_loss=0.3605, over 3345445.36 frames. ], batch size: 48, lr: 6.04e-03, grad_scale: 32.0 2024-09-23 21:31:51,190 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=350112.0, ans=0.125 2024-09-23 21:32:17,098 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=350205.3333333333, ans=0.125 2024-09-23 21:33:02,595 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.92 vs. limit=15.0 2024-09-23 21:33:04,840 INFO [train.py:1198] (2/4) Epoch 20, batch 1050, loss[loss=0.1822, ctc_loss=0.1225, cr_loss=0.2987, over 17187.00 frames. ], tot_loss[loss=0.2158, ctc_loss=0.1438, cr_loss=0.3601, over 3342902.10 frames. ], batch size: 41, lr: 6.04e-03, grad_scale: 32.0 2024-09-23 21:33:27,842 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.295e+02 1.374e+02 1.534e+02 2.501e+02, threshold=2.748e+02, percent-clipped=0.0 2024-09-23 21:33:50,447 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=350438.6666666667, ans=0.0 2024-09-23 21:33:53,527 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=350485.3333333333, ans=0.0 2024-09-23 21:33:56,850 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=350485.3333333333, ans=0.0 2024-09-23 21:34:03,691 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.63 vs. limit=22.5 2024-09-23 21:34:32,260 INFO [train.py:1198] (2/4) Epoch 20, batch 1100, loss[loss=0.1844, ctc_loss=0.118, cr_loss=0.332, over 16407.00 frames. ], tot_loss[loss=0.2143, ctc_loss=0.1427, cr_loss=0.3582, over 3350140.00 frames. ], batch size: 36, lr: 6.04e-03, grad_scale: 16.0 2024-09-23 21:34:32,600 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=350578.6666666667, ans=0.125 2024-09-23 21:34:48,632 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=350625.3333333333, ans=0.125 2024-09-23 21:35:16,173 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=3.04 vs. limit=6.0 2024-09-23 21:35:20,889 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.88 vs. limit=12.0 2024-09-23 21:35:27,080 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.28 vs. limit=6.0 2024-09-23 21:35:33,920 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=6.86 vs. limit=15.0 2024-09-23 21:35:39,482 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=350765.3333333333, ans=0.0 2024-09-23 21:35:52,149 INFO [train.py:1198] (2/4) Epoch 20, batch 1150, loss[loss=0.2018, ctc_loss=0.1317, cr_loss=0.3505, over 17022.00 frames. ], tot_loss[loss=0.2133, ctc_loss=0.1419, cr_loss=0.3573, over 3363969.06 frames. ], batch size: 44, lr: 6.03e-03, grad_scale: 16.0 2024-09-23 21:35:57,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=350812.0, ans=0.1 2024-09-23 21:36:06,900 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=350858.6666666667, ans=0.125 2024-09-23 21:36:14,527 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.089e+02 1.209e+02 1.298e+02 1.420e+02 2.385e+02, threshold=2.595e+02, percent-clipped=0.0 2024-09-23 21:36:38,972 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=350952.0, ans=0.0 2024-09-23 21:36:42,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=350952.0, ans=0.0 2024-09-23 21:36:42,225 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=350952.0, ans=0.025 2024-09-23 21:37:02,869 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=350998.6666666667, ans=0.125 2024-09-23 21:37:07,647 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=350998.6666666667, ans=0.125 2024-09-23 21:37:10,934 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=351045.3333333333, ans=0.2 2024-09-23 21:37:12,176 INFO [train.py:1198] (2/4) Epoch 20, batch 1200, loss[loss=0.1598, ctc_loss=0.1029, cr_loss=0.2848, over 17052.00 frames. ], tot_loss[loss=0.2137, ctc_loss=0.1421, cr_loss=0.3581, over 3365794.28 frames. ], batch size: 39, lr: 6.03e-03, grad_scale: 32.0 2024-09-23 21:37:16,021 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.33 vs. limit=6.0 2024-09-23 21:37:32,580 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=351092.0, ans=0.125 2024-09-23 21:37:32,736 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=351092.0, ans=0.125 2024-09-23 21:37:47,095 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=351138.6666666667, ans=0.035 2024-09-23 21:38:23,083 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=351232.0, ans=0.125 2024-09-23 21:38:35,086 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=16.98 vs. limit=22.5 2024-09-23 21:38:36,677 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=6.70 vs. limit=15.0 2024-09-23 21:38:37,551 INFO [train.py:1198] (2/4) Epoch 20, batch 1250, loss[loss=0.2151, ctc_loss=0.1451, cr_loss=0.35, over 17006.00 frames. ], tot_loss[loss=0.2153, ctc_loss=0.1432, cr_loss=0.3602, over 3370325.99 frames. ], batch size: 56, lr: 6.03e-03, grad_scale: 32.0 2024-09-23 21:38:39,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=351278.6666666667, ans=0.125 2024-09-23 21:38:47,332 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=351278.6666666667, ans=0.125 2024-09-23 21:38:48,959 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=351278.6666666667, ans=0.1 2024-09-23 21:38:52,862 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.90 vs. limit=15.0 2024-09-23 21:38:53,657 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=351325.3333333333, ans=0.125 2024-09-23 21:38:59,140 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.62 vs. limit=15.0 2024-09-23 21:38:59,801 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.265e+02 1.387e+02 1.557e+02 1.898e+02, threshold=2.775e+02, percent-clipped=0.0 2024-09-23 21:39:03,301 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=351325.3333333333, ans=0.0 2024-09-23 21:39:13,836 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=351372.0, ans=0.125 2024-09-23 21:39:43,298 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=351418.6666666667, ans=0.0 2024-09-23 21:39:51,615 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=15.88 vs. limit=22.5 2024-09-23 21:40:02,002 INFO [train.py:1198] (2/4) Epoch 20, batch 1300, loss[loss=0.2559, ctc_loss=0.1795, cr_loss=0.3818, over 11827.00 frames. ], tot_loss[loss=0.2153, ctc_loss=0.1433, cr_loss=0.3597, over 3364788.24 frames. ], batch size: 123, lr: 6.03e-03, grad_scale: 32.0 2024-09-23 21:40:23,922 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=13.67 vs. limit=22.5 2024-09-23 21:40:57,993 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=351652.0, ans=0.0 2024-09-23 21:41:01,255 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 21:41:10,810 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=351698.6666666667, ans=0.125 2024-09-23 21:41:21,502 INFO [train.py:1198] (2/4) Epoch 20, batch 1350, loss[loss=0.1861, ctc_loss=0.1223, cr_loss=0.3194, over 17271.00 frames. ], tot_loss[loss=0.2167, ctc_loss=0.1446, cr_loss=0.3606, over 3350435.60 frames. ], batch size: 42, lr: 6.03e-03, grad_scale: 32.0 2024-09-23 21:41:42,375 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=351792.0, ans=0.0 2024-09-23 21:41:43,586 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.053e+02 1.285e+02 1.356e+02 1.494e+02 3.061e+02, threshold=2.711e+02, percent-clipped=1.0 2024-09-23 21:41:44,142 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=19.10 vs. limit=22.5 2024-09-23 21:41:49,246 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.whiten, num_groups=1, num_channels=512, metric=4.95 vs. limit=12.0 2024-09-23 21:41:55,234 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=351838.6666666667, ans=0.125 2024-09-23 21:41:55,271 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=351838.6666666667, ans=0.125 2024-09-23 21:42:04,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=351838.6666666667, ans=0.125 2024-09-23 21:42:08,116 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=351885.3333333333, ans=0.125 2024-09-23 21:42:26,331 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=351932.0, ans=0.125 2024-09-23 21:42:34,205 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 21:42:43,485 INFO [train.py:1198] (2/4) Epoch 20, batch 1400, loss[loss=0.2283, ctc_loss=0.153, cr_loss=0.3765, over 17158.00 frames. ], tot_loss[loss=0.2175, ctc_loss=0.1451, cr_loss=0.3621, over 3352239.51 frames. ], batch size: 45, lr: 6.02e-03, grad_scale: 32.0 2024-09-23 21:42:53,672 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.26 vs. limit=15.0 2024-09-23 21:42:56,300 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=351978.6666666667, ans=0.125 2024-09-23 21:43:00,766 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.min_positive, batch_count=352025.3333333333, ans=0.025 2024-09-23 21:43:13,533 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=12.37 vs. limit=22.5 2024-09-23 21:43:18,178 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 21:43:56,827 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.78 vs. limit=15.0 2024-09-23 21:44:05,191 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=352165.3333333333, ans=0.125 2024-09-23 21:44:08,212 INFO [train.py:1198] (2/4) Epoch 20, batch 1450, loss[loss=0.22, ctc_loss=0.1429, cr_loss=0.3855, over 16687.00 frames. ], tot_loss[loss=0.2165, ctc_loss=0.1443, cr_loss=0.361, over 3361919.38 frames. ], batch size: 37, lr: 6.02e-03, grad_scale: 32.0 2024-09-23 21:44:14,242 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=352212.0, ans=0.04949747468305833 2024-09-23 21:44:22,286 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.min_positive, batch_count=352212.0, ans=0.025 2024-09-23 21:44:32,962 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.276e+02 1.364e+02 1.479e+02 2.768e+02, threshold=2.727e+02, percent-clipped=1.0 2024-09-23 21:44:37,917 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=352258.6666666667, ans=0.125 2024-09-23 21:44:42,763 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=352305.3333333333, ans=0.1 2024-09-23 21:45:07,002 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=352352.0, ans=0.125 2024-09-23 21:45:11,585 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=352352.0, ans=0.1 2024-09-23 21:45:14,592 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer_ff3.min_abs, batch_count=352398.6666666667, ans=0.2 2024-09-23 21:45:20,782 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=352398.6666666667, ans=0.0 2024-09-23 21:45:30,175 INFO [train.py:1198] (2/4) Epoch 20, batch 1500, loss[loss=0.1941, ctc_loss=0.1264, cr_loss=0.3384, over 17232.00 frames. ], tot_loss[loss=0.2169, ctc_loss=0.1446, cr_loss=0.3616, over 3359919.22 frames. ], batch size: 50, lr: 6.02e-03, grad_scale: 16.0 2024-09-23 21:45:40,474 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.60 vs. limit=22.5 2024-09-23 21:46:12,425 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 21:46:18,772 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=352585.3333333333, ans=0.125 2024-09-23 21:46:51,135 INFO [train.py:1198] (2/4) Epoch 20, batch 1550, loss[loss=0.1979, ctc_loss=0.1298, cr_loss=0.3404, over 16931.00 frames. ], tot_loss[loss=0.2171, ctc_loss=0.1447, cr_loss=0.3622, over 3369161.75 frames. ], batch size: 42, lr: 6.02e-03, grad_scale: 16.0 2024-09-23 21:46:51,350 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=352678.6666666667, ans=0.0 2024-09-23 21:46:51,956 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=9.39 vs. limit=15.0 2024-09-23 21:46:57,823 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=352678.6666666667, ans=0.2 2024-09-23 21:47:17,719 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.019e+02 1.276e+02 1.387e+02 1.513e+02 2.213e+02, threshold=2.774e+02, percent-clipped=0.0 2024-09-23 21:47:20,298 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=19.46 vs. limit=22.5 2024-09-23 21:47:24,349 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=352772.0, ans=0.07 2024-09-23 21:47:29,139 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=352772.0, ans=0.2 2024-09-23 21:47:35,674 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=352772.0, ans=0.2 2024-09-23 21:48:16,279 INFO [train.py:1198] (2/4) Epoch 20, batch 1600, loss[loss=0.1785, ctc_loss=0.118, cr_loss=0.3027, over 16980.00 frames. ], tot_loss[loss=0.2163, ctc_loss=0.1441, cr_loss=0.3611, over 3368583.70 frames. ], batch size: 42, lr: 6.02e-03, grad_scale: 32.0 2024-09-23 21:48:49,998 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.min_abs, batch_count=353005.3333333333, ans=0.5 2024-09-23 21:49:22,045 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=353052.0, ans=0.0 2024-09-23 21:49:25,567 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=17.46 vs. limit=22.5 2024-09-23 21:49:41,002 INFO [train.py:1198] (2/4) Epoch 20, batch 1650, loss[loss=0.2383, ctc_loss=0.1632, cr_loss=0.3753, over 17005.00 frames. ], tot_loss[loss=0.2152, ctc_loss=0.1433, cr_loss=0.3597, over 3376050.60 frames. ], batch size: 56, lr: 6.02e-03, grad_scale: 32.0 2024-09-23 21:50:03,972 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=353192.0, ans=0.125 2024-09-23 21:50:05,308 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.132e+02 1.288e+02 1.374e+02 1.505e+02 2.586e+02, threshold=2.748e+02, percent-clipped=0.0 2024-09-23 21:50:17,445 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.34 vs. limit=6.0 2024-09-23 21:50:20,092 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=353238.6666666667, ans=0.125 2024-09-23 21:50:52,330 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=13.27 vs. limit=22.5 2024-09-23 21:50:53,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=353332.0, ans=0.125 2024-09-23 21:50:59,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=353378.6666666667, ans=0.125 2024-09-23 21:51:01,006 INFO [train.py:1198] (2/4) Epoch 20, batch 1700, loss[loss=0.2137, ctc_loss=0.1385, cr_loss=0.3756, over 17308.00 frames. ], tot_loss[loss=0.2144, ctc_loss=0.1427, cr_loss=0.3585, over 3385623.95 frames. ], batch size: 46, lr: 6.01e-03, grad_scale: 32.0 2024-09-23 21:51:10,962 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.min_abs, batch_count=353378.6666666667, ans=0.5 2024-09-23 21:51:14,386 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=353378.6666666667, ans=0.09899494936611666 2024-09-23 21:51:17,678 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=353425.3333333333, ans=0.05 2024-09-23 21:52:05,737 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=353565.3333333333, ans=0.125 2024-09-23 21:52:12,067 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=10.60 vs. limit=15.0 2024-09-23 21:52:22,962 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=353612.0, ans=0.125 2024-09-23 21:52:24,263 INFO [train.py:1198] (2/4) Epoch 20, batch 1750, loss[loss=0.1736, ctc_loss=0.113, cr_loss=0.3032, over 17096.00 frames. ], tot_loss[loss=0.2141, ctc_loss=0.1424, cr_loss=0.3583, over 3380143.33 frames. ], batch size: 43, lr: 6.01e-03, grad_scale: 16.0 2024-09-23 21:52:49,735 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.235e+02 1.321e+02 1.434e+02 2.366e+02, threshold=2.642e+02, percent-clipped=0.0 2024-09-23 21:53:02,697 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=4.53 vs. limit=12.0 2024-09-23 21:53:13,106 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=5.11 vs. limit=12.0 2024-09-23 21:53:25,264 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=353752.0, ans=0.0 2024-09-23 21:53:32,602 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.64 vs. limit=15.0 2024-09-23 21:53:39,607 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=353798.6666666667, ans=0.2 2024-09-23 21:53:44,707 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=8.93 vs. limit=12.0 2024-09-23 21:53:47,292 INFO [train.py:1198] (2/4) Epoch 20, batch 1800, loss[loss=0.2224, ctc_loss=0.146, cr_loss=0.3818, over 17017.00 frames. ], tot_loss[loss=0.2151, ctc_loss=0.1433, cr_loss=0.3594, over 3370124.62 frames. ], batch size: 51, lr: 6.01e-03, grad_scale: 16.0 2024-09-23 21:54:18,894 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.61 vs. limit=15.0 2024-09-23 21:54:32,707 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=353938.6666666667, ans=0.025 2024-09-23 21:54:45,456 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=353985.3333333333, ans=0.025 2024-09-23 21:54:59,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=354032.0, ans=0.1 2024-09-23 21:55:12,165 INFO [train.py:1198] (2/4) Epoch 20, batch 1850, loss[loss=0.2276, ctc_loss=0.1534, cr_loss=0.3708, over 16039.00 frames. ], tot_loss[loss=0.2155, ctc_loss=0.1435, cr_loss=0.3597, over 3370993.57 frames. ], batch size: 74, lr: 6.01e-03, grad_scale: 16.0 2024-09-23 21:55:26,346 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=14.64 vs. limit=15.0 2024-09-23 21:55:37,927 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.093e+02 1.237e+02 1.317e+02 1.410e+02 2.955e+02, threshold=2.633e+02, percent-clipped=1.0 2024-09-23 21:55:39,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=354125.3333333333, ans=0.125 2024-09-23 21:56:05,401 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=354218.6666666667, ans=0.125 2024-09-23 21:56:16,426 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=354265.3333333333, ans=0.125 2024-09-23 21:56:32,457 INFO [train.py:1198] (2/4) Epoch 20, batch 1900, loss[loss=0.2397, ctc_loss=0.1587, cr_loss=0.4049, over 16818.00 frames. ], tot_loss[loss=0.2158, ctc_loss=0.1438, cr_loss=0.3602, over 3370718.23 frames. ], batch size: 58, lr: 6.01e-03, grad_scale: 16.0 2024-09-23 21:56:41,260 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.20 vs. limit=15.0 2024-09-23 21:57:47,724 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=354498.6666666667, ans=0.0 2024-09-23 21:57:50,734 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=354498.6666666667, ans=0.0 2024-09-23 21:57:55,485 INFO [train.py:1198] (2/4) Epoch 20, batch 1950, loss[loss=0.2445, ctc_loss=0.1635, cr_loss=0.4048, over 17043.00 frames. ], tot_loss[loss=0.2161, ctc_loss=0.1441, cr_loss=0.3602, over 3364712.85 frames. ], batch size: 52, lr: 6.00e-03, grad_scale: 16.0 2024-09-23 21:58:00,065 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 21:58:05,163 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=11.10 vs. limit=15.0 2024-09-23 21:58:23,509 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.283e+02 1.409e+02 1.574e+02 2.318e+02, threshold=2.818e+02, percent-clipped=0.0 2024-09-23 21:58:42,548 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=354638.6666666667, ans=0.125 2024-09-23 21:58:52,365 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=354685.3333333333, ans=0.0 2024-09-23 21:59:10,861 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=354732.0, ans=0.025 2024-09-23 21:59:25,720 INFO [train.py:1198] (2/4) Epoch 20, batch 2000, loss[loss=0.2113, ctc_loss=0.139, cr_loss=0.3616, over 17304.00 frames. ], tot_loss[loss=0.2149, ctc_loss=0.1432, cr_loss=0.3587, over 3374413.96 frames. ], batch size: 49, lr: 6.00e-03, grad_scale: 32.0 2024-09-23 21:59:41,782 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=354825.3333333333, ans=0.0 2024-09-23 22:00:00,278 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=3.91 vs. limit=15.0 2024-09-23 22:00:17,232 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=354918.6666666667, ans=0.0 2024-09-23 22:00:18,924 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=354918.6666666667, ans=0.125 2024-09-23 22:00:29,953 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=354965.3333333333, ans=0.125 2024-09-23 22:00:45,847 INFO [train.py:1198] (2/4) Epoch 20, batch 2050, loss[loss=0.2359, ctc_loss=0.1606, cr_loss=0.3766, over 16999.00 frames. ], tot_loss[loss=0.2145, ctc_loss=0.143, cr_loss=0.3575, over 3371083.69 frames. ], batch size: 53, lr: 6.00e-03, grad_scale: 32.0 2024-09-23 22:01:05,694 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=355058.6666666667, ans=0.1 2024-09-23 22:01:08,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=355058.6666666667, ans=0.125 2024-09-23 22:01:11,577 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.282e+02 1.359e+02 1.463e+02 2.527e+02, threshold=2.718e+02, percent-clipped=0.0 2024-09-23 22:01:18,963 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=20.69 vs. limit=22.5 2024-09-23 22:01:28,039 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=11.34 vs. limit=22.5 2024-09-23 22:01:37,284 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=355152.0, ans=0.125 2024-09-23 22:01:41,861 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=355152.0, ans=0.125 2024-09-23 22:02:05,991 INFO [train.py:1198] (2/4) Epoch 20, batch 2100, loss[loss=0.2157, ctc_loss=0.143, cr_loss=0.3637, over 17215.00 frames. ], tot_loss[loss=0.2154, ctc_loss=0.1438, cr_loss=0.3583, over 3358900.52 frames. ], batch size: 47, lr: 6.00e-03, grad_scale: 32.0 2024-09-23 22:02:34,078 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=355292.0, ans=0.2 2024-09-23 22:02:35,617 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=355292.0, ans=0.125 2024-09-23 22:02:56,187 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=355385.3333333333, ans=0.0 2024-09-23 22:03:30,232 INFO [train.py:1198] (2/4) Epoch 20, batch 2150, loss[loss=0.1904, ctc_loss=0.1241, cr_loss=0.3318, over 17009.00 frames. ], tot_loss[loss=0.2157, ctc_loss=0.1439, cr_loss=0.359, over 3361635.74 frames. ], batch size: 39, lr: 6.00e-03, grad_scale: 32.0 2024-09-23 22:03:37,043 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=355478.6666666667, ans=0.0 2024-09-23 22:03:44,881 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=355525.3333333333, ans=0.0 2024-09-23 22:03:49,717 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=355525.3333333333, ans=0.125 2024-09-23 22:03:58,368 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.113e+02 1.263e+02 1.377e+02 1.523e+02 2.016e+02, threshold=2.755e+02, percent-clipped=0.0 2024-09-23 22:04:02,033 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=355525.3333333333, ans=0.1 2024-09-23 22:04:26,023 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=355618.6666666667, ans=0.0 2024-09-23 22:04:28,983 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=355618.6666666667, ans=0.1 2024-09-23 22:04:32,368 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=355618.6666666667, ans=0.1 2024-09-23 22:04:36,999 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 22:04:55,854 INFO [train.py:1198] (2/4) Epoch 20, batch 2200, loss[loss=0.1788, ctc_loss=0.1157, cr_loss=0.3158, over 17214.00 frames. ], tot_loss[loss=0.2144, ctc_loss=0.143, cr_loss=0.3568, over 3355431.62 frames. ], batch size: 50, lr: 5.99e-03, grad_scale: 32.0 2024-09-23 22:05:12,326 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 22:05:15,368 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=355758.6666666667, ans=0.125 2024-09-23 22:05:37,896 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=355805.3333333333, ans=0.2 2024-09-23 22:05:57,321 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=14.00 vs. limit=15.0 2024-09-23 22:06:09,205 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.68 vs. limit=6.0 2024-09-23 22:06:16,040 INFO [train.py:1198] (2/4) Epoch 20, batch 2250, loss[loss=0.1768, ctc_loss=0.1145, cr_loss=0.3115, over 17094.00 frames. ], tot_loss[loss=0.2148, ctc_loss=0.1433, cr_loss=0.3576, over 3355761.37 frames. ], batch size: 40, lr: 5.99e-03, grad_scale: 32.0 2024-09-23 22:06:27,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=355945.3333333333, ans=0.2 2024-09-23 22:06:38,964 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=355992.0, ans=0.2 2024-09-23 22:06:41,714 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.041e+02 1.265e+02 1.369e+02 1.505e+02 1.904e+02, threshold=2.738e+02, percent-clipped=0.0 2024-09-23 22:06:43,775 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=355992.0, ans=0.07 2024-09-23 22:06:43,779 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=355992.0, ans=0.125 2024-09-23 22:07:32,246 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=356132.0, ans=0.125 2024-09-23 22:07:34,007 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=356132.0, ans=0.2 2024-09-23 22:07:37,874 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.23 vs. limit=15.0 2024-09-23 22:07:38,555 INFO [train.py:1198] (2/4) Epoch 20, batch 2300, loss[loss=0.1875, ctc_loss=0.1177, cr_loss=0.3489, over 17051.00 frames. ], tot_loss[loss=0.2142, ctc_loss=0.1426, cr_loss=0.3577, over 3363362.55 frames. ], batch size: 39, lr: 5.99e-03, grad_scale: 32.0 2024-09-23 22:07:43,650 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=356178.6666666667, ans=0.2 2024-09-23 22:07:44,184 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=12.40 vs. limit=15.0 2024-09-23 22:08:01,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=356225.3333333333, ans=0.125 2024-09-23 22:08:01,991 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=356225.3333333333, ans=0.07 2024-09-23 22:08:21,708 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=6.14 vs. limit=15.0 2024-09-23 22:08:22,378 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=356272.0, ans=0.0 2024-09-23 22:08:22,521 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=356272.0, ans=0.1 2024-09-23 22:08:26,025 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=18.42 vs. limit=22.5 2024-09-23 22:08:39,722 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=356318.6666666667, ans=0.1 2024-09-23 22:08:58,194 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=356365.3333333333, ans=0.05 2024-09-23 22:09:02,582 INFO [train.py:1198] (2/4) Epoch 20, batch 2350, loss[loss=0.2173, ctc_loss=0.1427, cr_loss=0.3733, over 17323.00 frames. ], tot_loss[loss=0.2144, ctc_loss=0.1427, cr_loss=0.3583, over 3365646.17 frames. ], batch size: 51, lr: 5.99e-03, grad_scale: 32.0 2024-09-23 22:09:30,700 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.083e+02 1.236e+02 1.335e+02 1.500e+02 2.118e+02, threshold=2.671e+02, percent-clipped=0.0 2024-09-23 22:09:32,568 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=356458.6666666667, ans=0.125 2024-09-23 22:09:42,514 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.58 vs. limit=6.0 2024-09-23 22:09:43,574 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=356505.3333333333, ans=0.2 2024-09-23 22:09:43,727 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=356505.3333333333, ans=0.2 2024-09-23 22:10:04,702 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=356552.0, ans=0.05 2024-09-23 22:10:07,757 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=356598.6666666667, ans=0.1 2024-09-23 22:10:22,043 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=356598.6666666667, ans=0.0 2024-09-23 22:10:24,947 INFO [train.py:1198] (2/4) Epoch 20, batch 2400, loss[loss=0.2252, ctc_loss=0.1505, cr_loss=0.3736, over 17213.00 frames. ], tot_loss[loss=0.2145, ctc_loss=0.1428, cr_loss=0.3586, over 3363222.63 frames. ], batch size: 47, lr: 5.99e-03, grad_scale: 32.0 2024-09-23 22:10:51,956 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.96 vs. limit=6.0 2024-09-23 22:10:54,218 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=356692.0, ans=0.125 2024-09-23 22:11:07,011 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=356738.6666666667, ans=0.125 2024-09-23 22:11:28,820 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.out_whiten.whitening_limit, batch_count=356832.0, ans=8.0 2024-09-23 22:11:29,549 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=356832.0, ans=0.0 2024-09-23 22:11:45,254 INFO [train.py:1198] (2/4) Epoch 20, batch 2450, loss[loss=0.2768, ctc_loss=0.1935, cr_loss=0.4168, over 15056.00 frames. ], tot_loss[loss=0.2146, ctc_loss=0.143, cr_loss=0.3584, over 3366536.84 frames. ], batch size: 89, lr: 5.98e-03, grad_scale: 32.0 2024-09-23 22:11:57,086 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.31 vs. limit=6.0 2024-09-23 22:12:10,613 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=356925.3333333333, ans=0.0 2024-09-23 22:12:13,315 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.093e+02 1.240e+02 1.349e+02 1.469e+02 2.826e+02, threshold=2.697e+02, percent-clipped=1.0 2024-09-23 22:12:15,238 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=356925.3333333333, ans=0.1 2024-09-23 22:12:45,490 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=357018.6666666667, ans=0.1 2024-09-23 22:13:01,027 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=357065.3333333333, ans=0.125 2024-09-23 22:13:10,054 INFO [train.py:1198] (2/4) Epoch 20, batch 2500, loss[loss=0.2067, ctc_loss=0.1369, cr_loss=0.3489, over 17298.00 frames. ], tot_loss[loss=0.2148, ctc_loss=0.143, cr_loss=0.3589, over 3368036.54 frames. ], batch size: 51, lr: 5.98e-03, grad_scale: 32.0 2024-09-23 22:13:10,390 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=357112.0, ans=0.0 2024-09-23 22:13:18,933 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.18 vs. limit=15.0 2024-09-23 22:13:32,466 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=357158.6666666667, ans=0.125 2024-09-23 22:13:37,857 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.41 vs. limit=15.0 2024-09-23 22:13:38,868 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=357158.6666666667, ans=0.0 2024-09-23 22:13:41,102 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=9.07 vs. limit=15.0 2024-09-23 22:13:50,076 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=10.03 vs. limit=15.0 2024-09-23 22:14:20,621 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=357298.6666666667, ans=0.05 2024-09-23 22:14:30,205 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=357298.6666666667, ans=0.125 2024-09-23 22:14:34,664 INFO [train.py:1198] (2/4) Epoch 20, batch 2550, loss[loss=0.1818, ctc_loss=0.118, cr_loss=0.3187, over 17073.00 frames. ], tot_loss[loss=0.2161, ctc_loss=0.1439, cr_loss=0.361, over 3363376.61 frames. ], batch size: 43, lr: 5.98e-03, grad_scale: 32.0 2024-09-23 22:14:35,969 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=11.76 vs. limit=22.5 2024-09-23 22:14:47,934 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 22:14:56,620 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn1.whiten, num_groups=1, num_channels=192, metric=9.22 vs. limit=22.5 2024-09-23 22:15:00,261 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.269e+02 1.372e+02 1.542e+02 2.100e+02, threshold=2.744e+02, percent-clipped=0.0 2024-09-23 22:15:18,501 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=357438.6666666667, ans=0.0 2024-09-23 22:15:31,462 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.76 vs. limit=15.0 2024-09-23 22:15:32,749 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=357485.3333333333, ans=0.09899494936611666 2024-09-23 22:15:34,268 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=357485.3333333333, ans=0.125 2024-09-23 22:15:43,868 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=357532.0, ans=0.1 2024-09-23 22:15:50,324 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=357532.0, ans=0.125 2024-09-23 22:15:54,810 INFO [train.py:1198] (2/4) Epoch 20, batch 2600, loss[loss=0.1826, ctc_loss=0.1175, cr_loss=0.3258, over 17273.00 frames. ], tot_loss[loss=0.2152, ctc_loss=0.1433, cr_loss=0.3594, over 3360570.86 frames. ], batch size: 42, lr: 5.98e-03, grad_scale: 32.0 2024-09-23 22:15:58,343 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=357578.6666666667, ans=0.04949747468305833 2024-09-23 22:16:00,068 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=357578.6666666667, ans=0.04949747468305833 2024-09-23 22:16:05,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=357578.6666666667, ans=10.0 2024-09-23 22:16:11,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=357625.3333333333, ans=0.025 2024-09-23 22:16:14,756 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=357625.3333333333, ans=0.1 2024-09-23 22:16:16,442 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=357625.3333333333, ans=0.0 2024-09-23 22:16:29,312 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=357672.0, ans=0.125 2024-09-23 22:16:43,457 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=357718.6666666667, ans=0.125 2024-09-23 22:16:48,186 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=357718.6666666667, ans=0.2 2024-09-23 22:17:17,931 INFO [train.py:1198] (2/4) Epoch 20, batch 2650, loss[loss=0.2185, ctc_loss=0.1473, cr_loss=0.356, over 17281.00 frames. ], tot_loss[loss=0.2162, ctc_loss=0.1441, cr_loss=0.3606, over 3359117.00 frames. ], batch size: 49, lr: 5.98e-03, grad_scale: 32.0 2024-09-23 22:17:19,992 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=357812.0, ans=0.5 2024-09-23 22:17:30,081 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=9.91 vs. limit=15.0 2024-09-23 22:17:43,557 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.094e+02 1.263e+02 1.370e+02 1.503e+02 2.130e+02, threshold=2.739e+02, percent-clipped=0.0 2024-09-23 22:17:56,281 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=357905.3333333333, ans=0.125 2024-09-23 22:18:04,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=357905.3333333333, ans=0.0 2024-09-23 22:18:24,968 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=357998.6666666667, ans=0.1 2024-09-23 22:18:43,128 INFO [train.py:1198] (2/4) Epoch 20, batch 2700, loss[loss=0.1906, ctc_loss=0.123, cr_loss=0.3381, over 17105.00 frames. ], tot_loss[loss=0.2168, ctc_loss=0.1445, cr_loss=0.3611, over 3352126.90 frames. ], batch size: 43, lr: 5.97e-03, grad_scale: 32.0 2024-09-23 22:19:44,148 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=10.65 vs. limit=22.5 2024-09-23 22:19:53,208 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 22:20:05,465 INFO [train.py:1198] (2/4) Epoch 20, batch 2750, loss[loss=0.2064, ctc_loss=0.1371, cr_loss=0.3464, over 17223.00 frames. ], tot_loss[loss=0.2149, ctc_loss=0.1432, cr_loss=0.3583, over 3350276.48 frames. ], batch size: 50, lr: 5.97e-03, grad_scale: 32.0 2024-09-23 22:20:09,182 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=12.59 vs. limit=15.0 2024-09-23 22:20:16,905 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=358278.6666666667, ans=0.2 2024-09-23 22:20:31,065 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.059e+02 1.266e+02 1.360e+02 1.482e+02 2.958e+02, threshold=2.720e+02, percent-clipped=1.0 2024-09-23 22:20:44,234 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=358372.0, ans=0.1 2024-09-23 22:20:52,433 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.max_abs, batch_count=358418.6666666667, ans=10.0 2024-09-23 22:21:00,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=358418.6666666667, ans=0.125 2024-09-23 22:21:05,708 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=12.89 vs. limit=15.0 2024-09-23 22:21:22,730 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=358465.3333333333, ans=0.0 2024-09-23 22:21:25,605 INFO [train.py:1198] (2/4) Epoch 20, batch 2800, loss[loss=0.2122, ctc_loss=0.1407, cr_loss=0.3571, over 17095.00 frames. ], tot_loss[loss=0.214, ctc_loss=0.1425, cr_loss=0.3577, over 3352055.59 frames. ], batch size: 49, lr: 5.97e-03, grad_scale: 32.0 2024-09-23 22:21:33,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=358512.0, ans=0.125 2024-09-23 22:21:55,375 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=358558.6666666667, ans=0.0 2024-09-23 22:22:09,735 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=12.53 vs. limit=22.5 2024-09-23 22:22:12,727 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=358605.3333333333, ans=0.0 2024-09-23 22:22:24,712 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=9.03 vs. limit=10.0 2024-09-23 22:22:35,033 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=358698.6666666667, ans=0.0 2024-09-23 22:22:45,877 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=358698.6666666667, ans=0.125 2024-09-23 22:22:50,492 INFO [train.py:1198] (2/4) Epoch 20, batch 2850, loss[loss=0.2737, ctc_loss=0.1957, cr_loss=0.3902, over 12152.00 frames. ], tot_loss[loss=0.2151, ctc_loss=0.1434, cr_loss=0.3589, over 3346187.28 frames. ], batch size: 123, lr: 5.97e-03, grad_scale: 32.0 2024-09-23 22:23:09,757 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=358792.0, ans=0.1 2024-09-23 22:23:15,984 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.093e+02 1.246e+02 1.349e+02 1.409e+02 2.188e+02, threshold=2.698e+02, percent-clipped=0.0 2024-09-23 22:23:18,089 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=358792.0, ans=0.2 2024-09-23 22:23:50,587 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=358885.3333333333, ans=0.0 2024-09-23 22:24:15,311 INFO [train.py:1198] (2/4) Epoch 20, batch 2900, loss[loss=0.2347, ctc_loss=0.1592, cr_loss=0.3773, over 16828.00 frames. ], tot_loss[loss=0.215, ctc_loss=0.1432, cr_loss=0.359, over 3354434.05 frames. ], batch size: 61, lr: 5.97e-03, grad_scale: 32.0 2024-09-23 22:24:25,335 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=358978.6666666667, ans=0.0 2024-09-23 22:24:30,138 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=359025.3333333333, ans=0.125 2024-09-23 22:25:07,502 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=12.06 vs. limit=15.0 2024-09-23 22:25:10,085 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=359118.6666666667, ans=0.0 2024-09-23 22:25:13,364 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=359118.6666666667, ans=0.125 2024-09-23 22:25:35,526 INFO [train.py:1198] (2/4) Epoch 20, batch 2950, loss[loss=0.1692, ctc_loss=0.109, cr_loss=0.3009, over 17039.00 frames. ], tot_loss[loss=0.214, ctc_loss=0.1424, cr_loss=0.358, over 3365567.91 frames. ], batch size: 39, lr: 5.96e-03, grad_scale: 32.0 2024-09-23 22:26:00,732 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.080e+02 1.258e+02 1.349e+02 1.512e+02 2.191e+02, threshold=2.699e+02, percent-clipped=0.0 2024-09-23 22:26:37,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=359398.6666666667, ans=0.0 2024-09-23 22:26:40,938 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=359398.6666666667, ans=0.125 2024-09-23 22:26:54,936 INFO [train.py:1198] (2/4) Epoch 20, batch 3000, loss[loss=0.1762, ctc_loss=0.1144, cr_loss=0.3092, over 17050.00 frames. ], tot_loss[loss=0.2134, ctc_loss=0.142, cr_loss=0.3575, over 3356195.70 frames. ], batch size: 39, lr: 5.96e-03, grad_scale: 32.0 2024-09-23 22:26:54,936 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 22:27:10,605 INFO [train.py:1230] (2/4) Epoch 20, validation: loss=0.03912, ctc_loss=0.03912, cr_loss=8.309e-15, over 944034.00 frames. 2024-09-23 22:27:10,606 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 22:27:12,585 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=359445.3333333333, ans=0.125 2024-09-23 22:27:26,696 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=359492.0, ans=0.125 2024-09-23 22:27:37,668 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=359492.0, ans=0.2 2024-09-23 22:27:54,640 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=359538.6666666667, ans=0.0 2024-09-23 22:28:05,838 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=359585.3333333333, ans=0.125 2024-09-23 22:28:15,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=359632.0, ans=0.2 2024-09-23 22:28:22,856 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=359632.0, ans=0.0 2024-09-23 22:28:31,108 INFO [train.py:1198] (2/4) Epoch 20, batch 3050, loss[loss=0.213, ctc_loss=0.1382, cr_loss=0.374, over 17070.00 frames. ], tot_loss[loss=0.2124, ctc_loss=0.1412, cr_loss=0.3559, over 3360406.13 frames. ], batch size: 46, lr: 5.96e-03, grad_scale: 32.0 2024-09-23 22:28:55,539 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.26 vs. limit=15.0 2024-09-23 22:28:56,199 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.297e+02 1.389e+02 1.525e+02 3.485e+02, threshold=2.778e+02, percent-clipped=1.0 2024-09-23 22:29:22,883 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=359818.6666666667, ans=0.125 2024-09-23 22:29:49,413 INFO [train.py:1198] (2/4) Epoch 20, batch 3100, loss[loss=0.2106, ctc_loss=0.1389, cr_loss=0.3587, over 17133.00 frames. ], tot_loss[loss=0.2138, ctc_loss=0.1422, cr_loss=0.3582, over 3358688.85 frames. ], batch size: 40, lr: 5.96e-03, grad_scale: 32.0 2024-09-23 22:30:12,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=359958.6666666667, ans=0.125 2024-09-23 22:30:15,579 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 22:30:20,184 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=359958.6666666667, ans=0.1 2024-09-23 22:30:37,850 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=2.94 vs. limit=15.0 2024-09-23 22:30:40,479 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=360052.0, ans=0.0 2024-09-23 22:30:48,289 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=360052.0, ans=0.0 2024-09-23 22:30:48,419 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=360052.0, ans=0.125 2024-09-23 22:30:55,495 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=3.97 vs. limit=15.0 2024-09-23 22:31:07,506 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=360098.6666666667, ans=0.125 2024-09-23 22:31:12,474 INFO [train.py:1198] (2/4) Epoch 20, batch 3150, loss[loss=0.249, ctc_loss=0.1675, cr_loss=0.4072, over 15788.00 frames. ], tot_loss[loss=0.214, ctc_loss=0.1423, cr_loss=0.3583, over 3354243.41 frames. ], batch size: 74, lr: 5.96e-03, grad_scale: 32.0 2024-09-23 22:31:37,610 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.113e+02 1.291e+02 1.400e+02 1.624e+02 2.307e+02, threshold=2.800e+02, percent-clipped=0.0 2024-09-23 22:31:53,031 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module1.whiten, num_groups=1, num_channels=192, metric=7.55 vs. limit=15.0 2024-09-23 22:32:17,946 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.03 vs. limit=15.0 2024-09-23 22:32:31,335 INFO [train.py:1198] (2/4) Epoch 20, batch 3200, loss[loss=0.2303, ctc_loss=0.1538, cr_loss=0.3828, over 17190.00 frames. ], tot_loss[loss=0.2133, ctc_loss=0.1419, cr_loss=0.3573, over 3355760.12 frames. ], batch size: 55, lr: 5.95e-03, grad_scale: 32.0 2024-09-23 22:32:36,390 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=360378.6666666667, ans=0.0 2024-09-23 22:32:58,784 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=8.44 vs. limit=15.0 2024-09-23 22:33:34,772 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=7.55 vs. limit=15.0 2024-09-23 22:33:49,550 INFO [train.py:1198] (2/4) Epoch 20, batch 3250, loss[loss=0.2244, ctc_loss=0.1525, cr_loss=0.3597, over 17299.00 frames. ], tot_loss[loss=0.214, ctc_loss=0.1424, cr_loss=0.3579, over 3359203.96 frames. ], batch size: 46, lr: 5.95e-03, grad_scale: 32.0 2024-09-23 22:33:56,189 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=360612.0, ans=0.1 2024-09-23 22:33:57,606 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=360612.0, ans=0.1 2024-09-23 22:34:16,116 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.062e+02 1.218e+02 1.307e+02 1.462e+02 2.065e+02, threshold=2.615e+02, percent-clipped=0.0 2024-09-23 22:34:58,714 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=360798.6666666667, ans=0.125 2024-09-23 22:35:08,070 INFO [train.py:1198] (2/4) Epoch 20, batch 3300, loss[loss=0.2531, ctc_loss=0.1827, cr_loss=0.3522, over 12017.00 frames. ], tot_loss[loss=0.2156, ctc_loss=0.1435, cr_loss=0.3604, over 3358657.67 frames. ], batch size: 123, lr: 5.95e-03, grad_scale: 32.0 2024-09-23 22:35:47,438 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=360938.6666666667, ans=0.125 2024-09-23 22:36:26,306 INFO [train.py:1198] (2/4) Epoch 20, batch 3350, loss[loss=0.2049, ctc_loss=0.135, cr_loss=0.3495, over 17094.00 frames. ], tot_loss[loss=0.216, ctc_loss=0.1438, cr_loss=0.3606, over 3351231.33 frames. ], batch size: 43, lr: 5.95e-03, grad_scale: 16.0 2024-09-23 22:36:42,693 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module1.whiten, num_groups=1, num_channels=512, metric=5.45 vs. limit=15.0 2024-09-23 22:36:45,827 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=16.31 vs. limit=22.5 2024-09-23 22:36:56,430 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.040e+02 1.260e+02 1.354e+02 1.461e+02 2.333e+02, threshold=2.708e+02, percent-clipped=0.0 2024-09-23 22:36:58,375 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-23 22:36:58,381 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=361172.0, ans=0.2 2024-09-23 22:37:02,724 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=361172.0, ans=0.1 2024-09-23 22:37:14,065 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=361218.6666666667, ans=0.025 2024-09-23 22:37:17,181 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=361218.6666666667, ans=0.025 2024-09-23 22:37:45,432 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=361312.0, ans=0.025 2024-09-23 22:37:46,671 INFO [train.py:1198] (2/4) Epoch 20, batch 3400, loss[loss=0.2635, ctc_loss=0.1887, cr_loss=0.374, over 11419.00 frames. ], tot_loss[loss=0.2146, ctc_loss=0.1429, cr_loss=0.3584, over 3356004.28 frames. ], batch size: 123, lr: 5.95e-03, grad_scale: 16.0 2024-09-23 22:38:39,480 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=361452.0, ans=0.125 2024-09-23 22:38:40,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=361452.0, ans=0.2 2024-09-23 22:39:00,117 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.max_abs, batch_count=361498.6666666667, ans=10.0 2024-09-23 22:39:06,007 INFO [train.py:1198] (2/4) Epoch 20, batch 3450, loss[loss=0.261, ctc_loss=0.1797, cr_loss=0.4063, over 15058.00 frames. ], tot_loss[loss=0.2154, ctc_loss=0.1434, cr_loss=0.3599, over 3350037.65 frames. ], batch size: 89, lr: 5.95e-03, grad_scale: 16.0 2024-09-23 22:39:34,087 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.266e+02 1.362e+02 1.520e+02 1.983e+02, threshold=2.723e+02, percent-clipped=0.0 2024-09-23 22:39:40,716 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=361638.6666666667, ans=0.04949747468305833 2024-09-23 22:39:43,688 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=361638.6666666667, ans=0.125 2024-09-23 22:40:24,483 INFO [train.py:1198] (2/4) Epoch 20, batch 3500, loss[loss=0.2327, ctc_loss=0.1588, cr_loss=0.3697, over 16020.00 frames. ], tot_loss[loss=0.2141, ctc_loss=0.1425, cr_loss=0.358, over 3351935.83 frames. ], batch size: 74, lr: 5.94e-03, grad_scale: 16.0 2024-09-23 22:40:31,036 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=361778.6666666667, ans=0.125 2024-09-23 22:40:34,044 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=361778.6666666667, ans=0.2 2024-09-23 22:40:50,931 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=11.24 vs. limit=15.0 2024-09-23 22:40:51,965 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=361825.3333333333, ans=0.0 2024-09-23 22:40:59,762 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=361872.0, ans=0.125 2024-09-23 22:41:10,810 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=361872.0, ans=0.1 2024-09-23 22:41:12,490 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=361918.6666666667, ans=0.125 2024-09-23 22:41:21,918 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=361918.6666666667, ans=0.125 2024-09-23 22:41:26,500 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 22:41:47,245 INFO [train.py:1198] (2/4) Epoch 20, batch 3550, loss[loss=0.2218, ctc_loss=0.1458, cr_loss=0.3801, over 16748.00 frames. ], tot_loss[loss=0.2151, ctc_loss=0.1433, cr_loss=0.3592, over 3350915.63 frames. ], batch size: 61, lr: 5.94e-03, grad_scale: 16.0 2024-09-23 22:41:49,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=362012.0, ans=0.0 2024-09-23 22:42:04,845 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=362058.6666666667, ans=0.09899494936611666 2024-09-23 22:42:15,394 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.065e+02 1.278e+02 1.363e+02 1.489e+02 4.233e+02, threshold=2.726e+02, percent-clipped=1.0 2024-09-23 22:43:05,644 INFO [train.py:1198] (2/4) Epoch 20, batch 3600, loss[loss=0.2035, ctc_loss=0.136, cr_loss=0.3375, over 17098.00 frames. ], tot_loss[loss=0.2155, ctc_loss=0.1436, cr_loss=0.3594, over 3351194.76 frames. ], batch size: 49, lr: 5.94e-03, grad_scale: 32.0 2024-09-23 22:43:21,556 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=362292.0, ans=0.1 2024-09-23 22:43:21,561 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=362292.0, ans=0.125 2024-09-23 22:43:24,616 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=362292.0, ans=0.07 2024-09-23 22:43:32,608 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=362292.0, ans=0.125 2024-09-23 22:44:11,521 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 22:44:23,798 INFO [train.py:1198] (2/4) Epoch 20, batch 3650, loss[loss=0.2082, ctc_loss=0.1396, cr_loss=0.3432, over 17098.00 frames. ], tot_loss[loss=0.2164, ctc_loss=0.1442, cr_loss=0.361, over 3351953.54 frames. ], batch size: 49, lr: 5.94e-03, grad_scale: 32.0 2024-09-23 22:44:25,677 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=362478.6666666667, ans=0.125 2024-09-23 22:44:51,968 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.102e+02 1.248e+02 1.353e+02 1.463e+02 2.228e+02, threshold=2.707e+02, percent-clipped=0.0 2024-09-23 22:44:53,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=362572.0, ans=0.125 2024-09-23 22:44:59,080 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.24 vs. limit=15.0 2024-09-23 22:45:12,524 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer_na.min_abs, batch_count=362618.6666666667, ans=0.02 2024-09-23 22:45:23,343 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=362618.6666666667, ans=0.0 2024-09-23 22:45:28,872 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=362665.3333333333, ans=0.025 2024-09-23 22:45:42,631 INFO [train.py:1198] (2/4) Epoch 20, batch 3700, loss[loss=0.2073, ctc_loss=0.1357, cr_loss=0.3582, over 17008.00 frames. ], tot_loss[loss=0.2162, ctc_loss=0.1441, cr_loss=0.3608, over 3355828.69 frames. ], batch size: 44, lr: 5.94e-03, grad_scale: 32.0 2024-09-23 22:45:48,557 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.79 vs. limit=6.0 2024-09-23 22:46:01,659 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=362758.6666666667, ans=0.1 2024-09-23 22:46:07,952 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=362758.6666666667, ans=0.0 2024-09-23 22:46:16,065 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=362805.3333333333, ans=0.2 2024-09-23 22:46:17,544 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=362805.3333333333, ans=0.025 2024-09-23 22:46:22,394 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 22:46:32,074 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.53 vs. limit=6.0 2024-09-23 22:46:32,845 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=362852.0, ans=0.1 2024-09-23 22:46:39,358 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=7.55 vs. limit=15.0 2024-09-23 22:46:53,090 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=6.94 vs. limit=12.0 2024-09-23 22:47:01,265 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=3.88 vs. limit=10.0 2024-09-23 22:47:01,721 INFO [train.py:1198] (2/4) Epoch 20, batch 3750, loss[loss=0.2367, ctc_loss=0.1566, cr_loss=0.4005, over 16927.00 frames. ], tot_loss[loss=0.2152, ctc_loss=0.1434, cr_loss=0.359, over 3352320.99 frames. ], batch size: 58, lr: 5.93e-03, grad_scale: 32.0 2024-09-23 22:47:27,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=362992.0, ans=0.125 2024-09-23 22:47:30,494 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.110e+02 1.286e+02 1.371e+02 1.493e+02 3.473e+02, threshold=2.742e+02, percent-clipped=1.0 2024-09-23 22:47:59,486 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=363085.3333333333, ans=0.2 2024-09-23 22:48:07,765 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.42 vs. limit=15.0 2024-09-23 22:48:07,883 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.48 vs. limit=6.0 2024-09-23 22:48:15,120 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=363132.0, ans=0.2 2024-09-23 22:48:18,283 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=363132.0, ans=0.0 2024-09-23 22:48:22,096 INFO [train.py:1198] (2/4) Epoch 20, batch 3800, loss[loss=0.2188, ctc_loss=0.1454, cr_loss=0.3674, over 17203.00 frames. ], tot_loss[loss=0.215, ctc_loss=0.1434, cr_loss=0.3577, over 3318917.27 frames. ], batch size: 41, lr: 5.93e-03, grad_scale: 32.0 2024-09-23 22:48:22,369 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=363178.6666666667, ans=0.09899494936611666 2024-09-23 22:48:23,933 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=363178.6666666667, ans=0.125 2024-09-23 22:48:41,155 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=363225.3333333333, ans=0.2 2024-09-23 22:48:42,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer_na.min_abs, batch_count=363225.3333333333, ans=0.02 2024-09-23 22:48:46,038 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=363225.3333333333, ans=0.2 2024-09-23 22:48:49,020 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=363225.3333333333, ans=0.95 2024-09-23 22:48:50,649 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=363225.3333333333, ans=0.125 2024-09-23 22:48:58,536 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=363272.0, ans=0.1 2024-09-23 22:49:03,152 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=363272.0, ans=0.125 2024-09-23 22:49:08,054 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=363318.6666666667, ans=0.125 2024-09-23 22:49:14,591 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=6.23 vs. limit=15.0 2024-09-23 22:49:15,852 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=363318.6666666667, ans=0.1 2024-09-23 22:49:15,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=363318.6666666667, ans=0.125 2024-09-23 22:49:20,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=363318.6666666667, ans=0.1 2024-09-23 22:49:34,397 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=363365.3333333333, ans=0.1 2024-09-23 22:49:40,056 INFO [train.py:1198] (2/4) Epoch 20, batch 3850, loss[loss=0.2405, ctc_loss=0.1625, cr_loss=0.3903, over 15195.00 frames. ], tot_loss[loss=0.218, ctc_loss=0.146, cr_loss=0.36, over 3245214.57 frames. ], batch size: 89, lr: 5.93e-03, grad_scale: 32.0 2024-09-23 22:49:56,494 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=363458.6666666667, ans=0.2 2024-09-23 22:49:57,904 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=363458.6666666667, ans=0.0 2024-09-23 22:49:57,910 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=363458.6666666667, ans=0.1 2024-09-23 22:50:08,337 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.115e+02 1.324e+02 1.457e+02 1.626e+02 2.908e+02, threshold=2.914e+02, percent-clipped=1.0 2024-09-23 22:50:10,131 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=363505.3333333333, ans=0.2 2024-09-23 22:51:41,132 INFO [train.py:1198] (2/4) Epoch 21, batch 0, loss[loss=0.1998, ctc_loss=0.1287, cr_loss=0.3552, over 17031.00 frames. ], tot_loss[loss=0.1998, ctc_loss=0.1287, cr_loss=0.3552, over 17031.00 frames. ], batch size: 44, lr: 5.78e-03, grad_scale: 32.0 2024-09-23 22:51:41,132 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-23 22:51:57,011 INFO [train.py:1230] (2/4) Epoch 21, validation: loss=0.03907, ctc_loss=0.03907, cr_loss=7.91e-15, over 944034.00 frames. 2024-09-23 22:51:57,011 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-23 22:52:23,510 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-23 22:52:41,077 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=363720.0, ans=0.125 2024-09-23 22:52:49,414 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=10.55 vs. limit=15.0 2024-09-23 22:53:09,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=363813.3333333333, ans=0.2 2024-09-23 22:53:11,243 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=363813.3333333333, ans=0.04949747468305833 2024-09-23 22:53:15,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=363813.3333333333, ans=0.125 2024-09-23 22:53:18,865 INFO [train.py:1198] (2/4) Epoch 21, batch 50, loss[loss=0.2185, ctc_loss=0.1435, cr_loss=0.3746, over 17143.00 frames. ], tot_loss[loss=0.2138, ctc_loss=0.1422, cr_loss=0.3578, over 753754.75 frames. ], batch size: 48, lr: 5.78e-03, grad_scale: 32.0 2024-09-23 22:53:28,391 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=363860.0, ans=0.1 2024-09-23 22:53:56,543 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.319e+02 1.470e+02 1.661e+02 2.685e+02, threshold=2.941e+02, percent-clipped=0.0 2024-09-23 22:54:41,603 INFO [train.py:1198] (2/4) Epoch 21, batch 100, loss[loss=0.2009, ctc_loss=0.1319, cr_loss=0.3449, over 17160.00 frames. ], tot_loss[loss=0.2165, ctc_loss=0.1442, cr_loss=0.3612, over 1331128.46 frames. ], batch size: 41, lr: 5.78e-03, grad_scale: 32.0 2024-09-23 22:54:57,708 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=364140.0, ans=0.125 2024-09-23 22:54:57,741 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=364140.0, ans=0.125 2024-09-23 22:55:00,859 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=364140.0, ans=0.0 2024-09-23 22:55:03,143 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.85 vs. limit=12.0 2024-09-23 22:55:24,682 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.44 vs. limit=12.0 2024-09-23 22:55:51,438 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=364280.0, ans=0.125 2024-09-23 22:56:00,942 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=364280.0, ans=0.025 2024-09-23 22:56:03,804 INFO [train.py:1198] (2/4) Epoch 21, batch 150, loss[loss=0.2123, ctc_loss=0.1421, cr_loss=0.3513, over 17221.00 frames. ], tot_loss[loss=0.2153, ctc_loss=0.1431, cr_loss=0.3612, over 1791133.09 frames. ], batch size: 50, lr: 5.78e-03, grad_scale: 32.0 2024-09-23 22:56:13,724 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=364326.6666666667, ans=0.0 2024-09-23 22:56:15,988 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn2.whiten, num_groups=1, num_channels=192, metric=9.45 vs. limit=22.5 2024-09-23 22:56:23,979 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.69 vs. limit=15.0 2024-09-23 22:56:24,975 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-23 22:56:36,034 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=364420.0, ans=0.04949747468305833 2024-09-23 22:56:38,962 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.055e+02 1.242e+02 1.342e+02 1.442e+02 2.042e+02, threshold=2.684e+02, percent-clipped=0.0 2024-09-23 22:57:12,335 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=364513.3333333333, ans=0.07 2024-09-23 22:57:19,024 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 22:57:29,599 INFO [train.py:1198] (2/4) Epoch 21, batch 200, loss[loss=0.1866, ctc_loss=0.1236, cr_loss=0.315, over 17099.00 frames. ], tot_loss[loss=0.2147, ctc_loss=0.1429, cr_loss=0.3589, over 2132939.85 frames. ], batch size: 43, lr: 5.78e-03, grad_scale: 32.0 2024-09-23 22:57:38,159 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=364560.0, ans=0.125 2024-09-23 22:57:52,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=364606.6666666667, ans=0.04949747468305833 2024-09-23 22:58:17,545 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=364653.3333333333, ans=0.1 2024-09-23 22:58:33,502 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=364700.0, ans=0.1 2024-09-23 22:58:52,293 INFO [train.py:1198] (2/4) Epoch 21, batch 250, loss[loss=0.2236, ctc_loss=0.1478, cr_loss=0.379, over 17090.00 frames. ], tot_loss[loss=0.2159, ctc_loss=0.1437, cr_loss=0.3607, over 2401690.20 frames. ], batch size: 49, lr: 5.77e-03, grad_scale: 32.0 2024-09-23 22:59:15,047 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=364840.0, ans=0.125 2024-09-23 22:59:27,564 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.024e+02 1.297e+02 1.364e+02 1.577e+02 2.161e+02, threshold=2.728e+02, percent-clipped=0.0 2024-09-23 22:59:29,450 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=364886.6666666667, ans=0.125 2024-09-23 22:59:51,144 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.47 vs. limit=15.0 2024-09-23 23:00:03,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=364980.0, ans=0.1 2024-09-23 23:00:15,862 INFO [train.py:1198] (2/4) Epoch 21, batch 300, loss[loss=0.1941, ctc_loss=0.1292, cr_loss=0.3248, over 16619.00 frames. ], tot_loss[loss=0.2167, ctc_loss=0.1444, cr_loss=0.3616, over 2610823.75 frames. ], batch size: 37, lr: 5.77e-03, grad_scale: 32.0 2024-09-23 23:00:49,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=365120.0, ans=0.125 2024-09-23 23:00:59,322 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=365120.0, ans=0.0 2024-09-23 23:01:02,624 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=365166.6666666667, ans=0.025 2024-09-23 23:01:15,228 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=365166.6666666667, ans=0.125 2024-09-23 23:01:26,533 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.85 vs. limit=15.0 2024-09-23 23:01:35,405 INFO [train.py:1198] (2/4) Epoch 21, batch 350, loss[loss=0.2146, ctc_loss=0.1466, cr_loss=0.3398, over 16749.00 frames. ], tot_loss[loss=0.2151, ctc_loss=0.1431, cr_loss=0.3599, over 2781716.08 frames. ], batch size: 61, lr: 5.77e-03, grad_scale: 32.0 2024-09-23 23:01:36,313 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=11.31 vs. limit=15.0 2024-09-23 23:01:57,795 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=365306.6666666667, ans=0.5 2024-09-23 23:02:15,505 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=365353.3333333333, ans=0.125 2024-09-23 23:02:16,055 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=5.82 vs. limit=12.0 2024-09-23 23:02:16,762 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.263e+02 1.348e+02 1.461e+02 2.184e+02, threshold=2.696e+02, percent-clipped=0.0 2024-09-23 23:02:34,773 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=365400.0, ans=0.125 2024-09-23 23:02:41,172 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=365400.0, ans=0.125 2024-09-23 23:02:58,568 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=365446.6666666667, ans=0.125 2024-09-23 23:03:01,497 INFO [train.py:1198] (2/4) Epoch 21, batch 400, loss[loss=0.2142, ctc_loss=0.1438, cr_loss=0.3518, over 17317.00 frames. ], tot_loss[loss=0.2155, ctc_loss=0.1436, cr_loss=0.3594, over 2896352.75 frames. ], batch size: 51, lr: 5.77e-03, grad_scale: 32.0 2024-09-23 23:03:06,677 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.91 vs. limit=15.0 2024-09-23 23:03:15,512 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=365493.3333333333, ans=0.125 2024-09-23 23:03:34,634 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=365586.6666666667, ans=0.125 2024-09-23 23:03:43,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=365586.6666666667, ans=0.125 2024-09-23 23:03:48,863 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=365586.6666666667, ans=0.1 2024-09-23 23:03:53,330 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=365633.3333333333, ans=0.125 2024-09-23 23:03:55,644 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.82 vs. limit=6.0 2024-09-23 23:04:20,666 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=365680.0, ans=0.125 2024-09-23 23:04:23,584 INFO [train.py:1198] (2/4) Epoch 21, batch 450, loss[loss=0.2523, ctc_loss=0.1748, cr_loss=0.3872, over 12322.00 frames. ], tot_loss[loss=0.2157, ctc_loss=0.1437, cr_loss=0.36, over 2997149.07 frames. ], batch size: 123, lr: 5.77e-03, grad_scale: 32.0 2024-09-23 23:04:36,863 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=365726.6666666667, ans=0.0 2024-09-23 23:04:43,437 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 23:04:51,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=365773.3333333333, ans=0.0 2024-09-23 23:04:51,470 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=365773.3333333333, ans=0.1 2024-09-23 23:04:53,177 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=365773.3333333333, ans=0.0 2024-09-23 23:04:59,071 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.052e+02 1.246e+02 1.320e+02 1.437e+02 2.140e+02, threshold=2.640e+02, percent-clipped=0.0 2024-09-23 23:04:59,438 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=365820.0, ans=0.025 2024-09-23 23:05:03,230 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.95 vs. limit=6.0 2024-09-23 23:05:04,092 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=365820.0, ans=0.025 2024-09-23 23:05:33,771 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=365913.3333333333, ans=0.2 2024-09-23 23:05:46,206 INFO [train.py:1198] (2/4) Epoch 21, batch 500, loss[loss=0.223, ctc_loss=0.1491, cr_loss=0.3696, over 17360.00 frames. ], tot_loss[loss=0.2153, ctc_loss=0.1434, cr_loss=0.3595, over 3081657.90 frames. ], batch size: 48, lr: 5.76e-03, grad_scale: 32.0 2024-09-23 23:05:48,888 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.36 vs. limit=6.0 2024-09-23 23:06:05,917 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=366006.6666666667, ans=0.0 2024-09-23 23:06:07,399 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=366006.6666666667, ans=0.125 2024-09-23 23:07:11,259 INFO [train.py:1198] (2/4) Epoch 21, batch 550, loss[loss=0.2127, ctc_loss=0.1367, cr_loss=0.3801, over 16947.00 frames. ], tot_loss[loss=0.2151, ctc_loss=0.1433, cr_loss=0.3593, over 3142576.62 frames. ], batch size: 42, lr: 5.76e-03, grad_scale: 32.0 2024-09-23 23:07:20,977 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=366193.3333333333, ans=0.125 2024-09-23 23:07:21,662 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=6.98 vs. limit=15.0 2024-09-23 23:07:45,228 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=366286.6666666667, ans=0.125 2024-09-23 23:07:46,482 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.227e+02 1.339e+02 1.480e+02 1.923e+02, threshold=2.679e+02, percent-clipped=0.0 2024-09-23 23:07:46,783 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=366286.6666666667, ans=0.125 2024-09-23 23:07:48,316 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=366286.6666666667, ans=0.125 2024-09-23 23:08:02,923 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=366333.3333333333, ans=0.025 2024-09-23 23:08:04,372 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=366333.3333333333, ans=0.125 2024-09-23 23:08:25,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=366380.0, ans=0.0 2024-09-23 23:08:27,500 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=366380.0, ans=0.0 2024-09-23 23:08:33,667 INFO [train.py:1198] (2/4) Epoch 21, batch 600, loss[loss=0.1752, ctc_loss=0.1147, cr_loss=0.3021, over 17273.00 frames. ], tot_loss[loss=0.2147, ctc_loss=0.143, cr_loss=0.3588, over 3198063.59 frames. ], batch size: 44, lr: 5.76e-03, grad_scale: 32.0 2024-09-23 23:08:41,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=366426.6666666667, ans=0.0 2024-09-23 23:09:16,401 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.60 vs. limit=15.0 2024-09-23 23:09:24,162 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.93 vs. limit=15.0 2024-09-23 23:09:28,312 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=366566.6666666667, ans=0.0 2024-09-23 23:09:28,886 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.27 vs. limit=15.0 2024-09-23 23:09:39,932 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.11 vs. limit=15.0 2024-09-23 23:09:53,532 INFO [train.py:1198] (2/4) Epoch 21, batch 650, loss[loss=0.1926, ctc_loss=0.1254, cr_loss=0.3361, over 17194.00 frames. ], tot_loss[loss=0.2158, ctc_loss=0.1438, cr_loss=0.3603, over 3214223.06 frames. ], batch size: 41, lr: 5.76e-03, grad_scale: 16.0 2024-09-23 23:10:32,741 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.251e+02 1.353e+02 1.422e+02 2.083e+02, threshold=2.707e+02, percent-clipped=0.0 2024-09-23 23:10:33,109 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=366753.3333333333, ans=0.0 2024-09-23 23:11:13,028 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=366846.6666666667, ans=0.125 2024-09-23 23:11:15,971 INFO [train.py:1198] (2/4) Epoch 21, batch 700, loss[loss=0.1803, ctc_loss=0.1174, cr_loss=0.3143, over 17225.00 frames. ], tot_loss[loss=0.2154, ctc_loss=0.1434, cr_loss=0.36, over 3245540.45 frames. ], batch size: 41, lr: 5.76e-03, grad_scale: 16.0 2024-09-23 23:11:41,733 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 23:11:51,286 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=366986.6666666667, ans=10.0 2024-09-23 23:12:21,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=367033.3333333333, ans=0.125 2024-09-23 23:12:24,178 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=367080.0, ans=0.0 2024-09-23 23:12:38,481 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=367080.0, ans=0.125 2024-09-23 23:12:41,459 INFO [train.py:1198] (2/4) Epoch 21, batch 750, loss[loss=0.2409, ctc_loss=0.1641, cr_loss=0.3839, over 17216.00 frames. ], tot_loss[loss=0.2162, ctc_loss=0.144, cr_loss=0.3607, over 3268857.92 frames. ], batch size: 55, lr: 5.76e-03, grad_scale: 16.0 2024-09-23 23:12:51,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=367126.6666666667, ans=0.5 2024-09-23 23:13:13,698 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=7.46 vs. limit=15.0 2024-09-23 23:13:21,100 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.308e+02 1.415e+02 1.587e+02 2.063e+02, threshold=2.831e+02, percent-clipped=0.0 2024-09-23 23:13:24,549 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=367220.0, ans=0.125 2024-09-23 23:13:26,271 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=367220.0, ans=0.09899494936611666 2024-09-23 23:13:27,890 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=367220.0, ans=0.025 2024-09-23 23:13:42,441 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=367266.6666666667, ans=0.0 2024-09-23 23:14:03,770 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=367360.0, ans=0.125 2024-09-23 23:14:04,986 INFO [train.py:1198] (2/4) Epoch 21, batch 800, loss[loss=0.2226, ctc_loss=0.1502, cr_loss=0.3616, over 17291.00 frames. ], tot_loss[loss=0.2163, ctc_loss=0.1441, cr_loss=0.361, over 3280414.58 frames. ], batch size: 46, lr: 5.75e-03, grad_scale: 32.0 2024-09-23 23:14:13,935 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=15.66 vs. limit=22.5 2024-09-23 23:14:14,923 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=367360.0, ans=0.125 2024-09-23 23:14:18,260 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=367360.0, ans=0.125 2024-09-23 23:15:07,494 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=18.01 vs. limit=22.5 2024-09-23 23:15:11,714 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=367546.6666666667, ans=0.1 2024-09-23 23:15:14,926 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=367546.6666666667, ans=0.125 2024-09-23 23:15:19,729 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=367546.6666666667, ans=0.125 2024-09-23 23:15:24,930 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.75 vs. limit=22.5 2024-09-23 23:15:27,328 INFO [train.py:1198] (2/4) Epoch 21, batch 850, loss[loss=0.2326, ctc_loss=0.152, cr_loss=0.4029, over 16988.00 frames. ], tot_loss[loss=0.2165, ctc_loss=0.1443, cr_loss=0.361, over 3291768.05 frames. ], batch size: 53, lr: 5.75e-03, grad_scale: 32.0 2024-09-23 23:16:05,379 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.026e+02 1.235e+02 1.359e+02 1.551e+02 2.196e+02, threshold=2.717e+02, percent-clipped=0.0 2024-09-23 23:16:41,153 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=367780.0, ans=0.1 2024-09-23 23:16:53,127 INFO [train.py:1198] (2/4) Epoch 21, batch 900, loss[loss=0.1866, ctc_loss=0.1245, cr_loss=0.3108, over 17273.00 frames. ], tot_loss[loss=0.216, ctc_loss=0.144, cr_loss=0.36, over 3289835.80 frames. ], batch size: 42, lr: 5.75e-03, grad_scale: 16.0 2024-09-23 23:17:06,278 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=367826.6666666667, ans=0.0 2024-09-23 23:17:17,748 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.43 vs. limit=15.0 2024-09-23 23:17:31,602 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=367920.0, ans=0.125 2024-09-23 23:18:15,415 INFO [train.py:1198] (2/4) Epoch 21, batch 950, loss[loss=0.1936, ctc_loss=0.129, cr_loss=0.3233, over 17054.00 frames. ], tot_loss[loss=0.2156, ctc_loss=0.1437, cr_loss=0.3596, over 3306544.23 frames. ], batch size: 46, lr: 5.75e-03, grad_scale: 16.0 2024-09-23 23:18:22,125 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=368060.0, ans=0.2 2024-09-23 23:18:33,104 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=368106.6666666667, ans=0.125 2024-09-23 23:18:33,211 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=368106.6666666667, ans=0.125 2024-09-23 23:18:41,053 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=368106.6666666667, ans=0.2 2024-09-23 23:18:53,167 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.253e+02 1.356e+02 1.481e+02 2.989e+02, threshold=2.713e+02, percent-clipped=1.0 2024-09-23 23:19:09,467 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=368200.0, ans=0.125 2024-09-23 23:19:17,605 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=368246.6666666667, ans=0.125 2024-09-23 23:19:17,708 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=368246.6666666667, ans=0.1 2024-09-23 23:19:28,849 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=368246.6666666667, ans=0.125 2024-09-23 23:19:34,877 INFO [train.py:1198] (2/4) Epoch 21, batch 1000, loss[loss=0.1717, ctc_loss=0.1129, cr_loss=0.2943, over 16965.00 frames. ], tot_loss[loss=0.2157, ctc_loss=0.1437, cr_loss=0.3602, over 3320767.28 frames. ], batch size: 42, lr: 5.75e-03, grad_scale: 16.0 2024-09-23 23:19:48,031 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=368293.3333333333, ans=0.0 2024-09-23 23:19:52,284 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=368340.0, ans=0.0 2024-09-23 23:19:55,559 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=368340.0, ans=0.0 2024-09-23 23:20:58,191 INFO [train.py:1198] (2/4) Epoch 21, batch 1050, loss[loss=0.2018, ctc_loss=0.1354, cr_loss=0.3321, over 17085.00 frames. ], tot_loss[loss=0.2163, ctc_loss=0.1443, cr_loss=0.3604, over 3326085.86 frames. ], batch size: 49, lr: 5.74e-03, grad_scale: 16.0 2024-09-23 23:21:30,506 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=368573.3333333333, ans=0.125 2024-09-23 23:21:39,785 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.060e+02 1.259e+02 1.337e+02 1.472e+02 2.048e+02, threshold=2.674e+02, percent-clipped=0.0 2024-09-23 23:21:54,255 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=368666.6666666667, ans=0.2 2024-09-23 23:22:02,778 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=13.09 vs. limit=22.5 2024-09-23 23:22:24,100 INFO [train.py:1198] (2/4) Epoch 21, batch 1100, loss[loss=0.2712, ctc_loss=0.1921, cr_loss=0.3958, over 11635.00 frames. ], tot_loss[loss=0.2164, ctc_loss=0.1442, cr_loss=0.3609, over 3336081.86 frames. ], batch size: 123, lr: 5.74e-03, grad_scale: 16.0 2024-09-23 23:23:00,721 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.45 vs. limit=10.0 2024-09-23 23:23:19,713 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=11.96 vs. limit=15.0 2024-09-23 23:23:36,688 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=368946.6666666667, ans=0.125 2024-09-23 23:23:46,068 INFO [train.py:1198] (2/4) Epoch 21, batch 1150, loss[loss=0.2536, ctc_loss=0.1721, cr_loss=0.4075, over 15936.00 frames. ], tot_loss[loss=0.2152, ctc_loss=0.1433, cr_loss=0.3596, over 3347132.12 frames. ], batch size: 74, lr: 5.74e-03, grad_scale: 16.0 2024-09-23 23:24:24,272 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.242e+02 1.332e+02 1.452e+02 2.606e+02, threshold=2.664e+02, percent-clipped=0.0 2024-09-23 23:24:29,338 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=369086.6666666667, ans=0.125 2024-09-23 23:24:47,905 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=369133.3333333333, ans=0.1 2024-09-23 23:24:51,122 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=369180.0, ans=0.025 2024-09-23 23:25:08,569 INFO [train.py:1198] (2/4) Epoch 21, batch 1200, loss[loss=0.244, ctc_loss=0.1718, cr_loss=0.3608, over 15068.00 frames. ], tot_loss[loss=0.2144, ctc_loss=0.1427, cr_loss=0.3585, over 3350498.72 frames. ], batch size: 89, lr: 5.74e-03, grad_scale: 32.0 2024-09-23 23:25:15,312 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=369226.6666666667, ans=0.125 2024-09-23 23:25:24,973 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-23 23:26:01,323 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=369366.6666666667, ans=0.025 2024-09-23 23:26:24,448 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=369413.3333333333, ans=0.0 2024-09-23 23:26:26,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=369413.3333333333, ans=0.2 2024-09-23 23:26:26,663 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.71 vs. limit=15.0 2024-09-23 23:26:27,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=369413.3333333333, ans=0.125 2024-09-23 23:26:30,657 INFO [train.py:1198] (2/4) Epoch 21, batch 1250, loss[loss=0.1899, ctc_loss=0.1248, cr_loss=0.3256, over 16775.00 frames. ], tot_loss[loss=0.2141, ctc_loss=0.1425, cr_loss=0.3581, over 3346615.15 frames. ], batch size: 37, lr: 5.74e-03, grad_scale: 16.0 2024-09-23 23:26:50,194 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=9.33 vs. limit=15.0 2024-09-23 23:27:09,041 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=369553.3333333333, ans=0.125 2024-09-23 23:27:13,542 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 9.996e+01 1.239e+02 1.348e+02 1.443e+02 2.209e+02, threshold=2.696e+02, percent-clipped=0.0 2024-09-23 23:27:27,500 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.13 vs. limit=10.0 2024-09-23 23:27:44,032 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=369646.6666666667, ans=0.125 2024-09-23 23:27:44,178 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten.whitening_limit, batch_count=369646.6666666667, ans=22.5 2024-09-23 23:27:56,568 INFO [train.py:1198] (2/4) Epoch 21, batch 1300, loss[loss=0.1658, ctc_loss=0.1087, cr_loss=0.2855, over 16711.00 frames. ], tot_loss[loss=0.2146, ctc_loss=0.1429, cr_loss=0.3586, over 3346701.10 frames. ], batch size: 37, lr: 5.74e-03, grad_scale: 16.0 2024-09-23 23:28:03,540 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=369693.3333333333, ans=0.125 2024-09-23 23:28:05,174 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=369693.3333333333, ans=0.125 2024-09-23 23:28:22,541 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=369740.0, ans=0.0 2024-09-23 23:28:22,963 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.21 vs. limit=6.0 2024-09-23 23:28:29,079 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=369786.6666666667, ans=0.2 2024-09-23 23:28:29,087 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=369786.6666666667, ans=0.125 2024-09-23 23:28:33,897 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=369786.6666666667, ans=0.2 2024-09-23 23:29:16,750 INFO [train.py:1198] (2/4) Epoch 21, batch 1350, loss[loss=0.2479, ctc_loss=0.1677, cr_loss=0.4008, over 17020.00 frames. ], tot_loss[loss=0.2145, ctc_loss=0.1428, cr_loss=0.3586, over 3355098.16 frames. ], batch size: 52, lr: 5.73e-03, grad_scale: 16.0 2024-09-23 23:29:21,831 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=369926.6666666667, ans=0.125 2024-09-23 23:29:22,045 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=16.71 vs. limit=22.5 2024-09-23 23:29:57,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=370020.0, ans=0.0 2024-09-23 23:29:58,707 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.126e+02 1.263e+02 1.337e+02 1.513e+02 2.330e+02, threshold=2.674e+02, percent-clipped=0.0 2024-09-23 23:30:07,595 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.59 vs. limit=10.0 2024-09-23 23:30:31,760 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.84 vs. limit=12.0 2024-09-23 23:30:38,706 INFO [train.py:1198] (2/4) Epoch 21, batch 1400, loss[loss=0.2648, ctc_loss=0.1862, cr_loss=0.3928, over 11700.00 frames. ], tot_loss[loss=0.2156, ctc_loss=0.1437, cr_loss=0.3598, over 3325221.98 frames. ], batch size: 123, lr: 5.73e-03, grad_scale: 16.0 2024-09-23 23:30:56,305 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=370206.6666666667, ans=0.125 2024-09-23 23:31:24,293 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=370253.3333333333, ans=0.125 2024-09-23 23:31:27,569 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=370300.0, ans=0.0 2024-09-23 23:31:36,959 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=370300.0, ans=0.1 2024-09-23 23:32:03,658 INFO [train.py:1198] (2/4) Epoch 21, batch 1450, loss[loss=0.2111, ctc_loss=0.1414, cr_loss=0.3483, over 17304.00 frames. ], tot_loss[loss=0.2144, ctc_loss=0.1428, cr_loss=0.3583, over 3329260.57 frames. ], batch size: 49, lr: 5.73e-03, grad_scale: 16.0 2024-09-23 23:32:08,752 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=370393.3333333333, ans=0.0 2024-09-23 23:32:12,620 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=13.96 vs. limit=15.0 2024-09-23 23:32:33,326 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=370440.0, ans=0.125 2024-09-23 23:32:39,698 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=12.44 vs. limit=22.5 2024-09-23 23:32:43,861 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=370486.6666666667, ans=0.125 2024-09-23 23:32:46,581 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.064e+02 1.255e+02 1.316e+02 1.424e+02 2.117e+02, threshold=2.631e+02, percent-clipped=0.0 2024-09-23 23:33:08,411 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=6.01 vs. limit=15.0 2024-09-23 23:33:14,183 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=370580.0, ans=0.125 2024-09-23 23:33:26,656 INFO [train.py:1198] (2/4) Epoch 21, batch 1500, loss[loss=0.2717, ctc_loss=0.1867, cr_loss=0.4252, over 17022.00 frames. ], tot_loss[loss=0.2149, ctc_loss=0.1431, cr_loss=0.3589, over 3324194.72 frames. ], batch size: 52, lr: 5.73e-03, grad_scale: 16.0 2024-09-23 23:33:39,908 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.min_positive, batch_count=370626.6666666667, ans=0.05 2024-09-23 23:33:44,912 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 23:34:11,961 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=370720.0, ans=0.125 2024-09-23 23:34:12,481 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=8.14 vs. limit=15.0 2024-09-23 23:34:23,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=370766.6666666667, ans=0.125 2024-09-23 23:34:23,202 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=370766.6666666667, ans=0.0 2024-09-23 23:34:31,076 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=370813.3333333333, ans=0.125 2024-09-23 23:34:32,792 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=370813.3333333333, ans=0.0 2024-09-23 23:34:40,729 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.41 vs. limit=12.0 2024-09-23 23:34:41,601 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=370813.3333333333, ans=0.125 2024-09-23 23:34:44,817 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=370813.3333333333, ans=0.125 2024-09-23 23:34:45,004 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-23 23:34:49,370 INFO [train.py:1198] (2/4) Epoch 21, batch 1550, loss[loss=0.2098, ctc_loss=0.1352, cr_loss=0.3734, over 17219.00 frames. ], tot_loss[loss=0.2159, ctc_loss=0.1438, cr_loss=0.3603, over 3325872.12 frames. ], batch size: 47, lr: 5.73e-03, grad_scale: 16.0 2024-09-23 23:34:49,754 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=370860.0, ans=0.125 2024-09-23 23:34:52,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=370860.0, ans=0.125 2024-09-23 23:35:20,108 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=370953.3333333333, ans=0.0 2024-09-23 23:35:29,477 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.107e+02 1.240e+02 1.330e+02 1.451e+02 2.029e+02, threshold=2.659e+02, percent-clipped=0.0 2024-09-23 23:35:39,407 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=371000.0, ans=0.2 2024-09-23 23:35:39,499 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=371000.0, ans=0.1 2024-09-23 23:35:48,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=371000.0, ans=0.125 2024-09-23 23:36:12,045 INFO [train.py:1198] (2/4) Epoch 21, batch 1600, loss[loss=0.2133, ctc_loss=0.1426, cr_loss=0.3534, over 17011.00 frames. ], tot_loss[loss=0.2146, ctc_loss=0.1428, cr_loss=0.3588, over 3335608.47 frames. ], batch size: 44, lr: 5.73e-03, grad_scale: 32.0 2024-09-23 23:36:29,734 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=371140.0, ans=0.125 2024-09-23 23:36:48,893 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.33 vs. limit=15.0 2024-09-23 23:37:37,163 INFO [train.py:1198] (2/4) Epoch 21, batch 1650, loss[loss=0.2307, ctc_loss=0.1498, cr_loss=0.4044, over 16943.00 frames. ], tot_loss[loss=0.2161, ctc_loss=0.1439, cr_loss=0.3608, over 3327152.92 frames. ], batch size: 42, lr: 5.72e-03, grad_scale: 32.0 2024-09-23 23:38:02,912 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.min_positive, batch_count=371373.3333333333, ans=0.05 2024-09-23 23:38:16,970 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.130e+02 1.286e+02 1.382e+02 1.548e+02 2.796e+02, threshold=2.764e+02, percent-clipped=1.0 2024-09-23 23:38:37,892 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=371466.6666666667, ans=0.125 2024-09-23 23:38:47,501 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=371513.3333333333, ans=0.09899494936611666 2024-09-23 23:38:52,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=371513.3333333333, ans=0.04949747468305833 2024-09-23 23:38:56,658 INFO [train.py:1198] (2/4) Epoch 21, batch 1700, loss[loss=0.2159, ctc_loss=0.144, cr_loss=0.3598, over 17029.00 frames. ], tot_loss[loss=0.2149, ctc_loss=0.143, cr_loss=0.3594, over 3342611.33 frames. ], batch size: 44, lr: 5.72e-03, grad_scale: 32.0 2024-09-23 23:39:03,283 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=371560.0, ans=0.04949747468305833 2024-09-23 23:39:17,552 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=371606.6666666667, ans=0.2 2024-09-23 23:39:20,716 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=371606.6666666667, ans=0.2 2024-09-23 23:39:39,018 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=371653.3333333333, ans=0.125 2024-09-23 23:40:15,606 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=371746.6666666667, ans=0.125 2024-09-23 23:40:18,410 INFO [train.py:1198] (2/4) Epoch 21, batch 1750, loss[loss=0.2058, ctc_loss=0.1345, cr_loss=0.3564, over 17184.00 frames. ], tot_loss[loss=0.2145, ctc_loss=0.1425, cr_loss=0.3597, over 3357186.88 frames. ], batch size: 41, lr: 5.72e-03, grad_scale: 32.0 2024-09-23 23:40:45,921 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=371840.0, ans=0.025 2024-09-23 23:40:47,300 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=371840.0, ans=0.1 2024-09-23 23:40:58,126 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.082e+02 1.218e+02 1.302e+02 1.379e+02 1.745e+02, threshold=2.603e+02, percent-clipped=0.0 2024-09-23 23:41:18,392 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=371933.3333333333, ans=0.0 2024-09-23 23:41:40,627 INFO [train.py:1198] (2/4) Epoch 21, batch 1800, loss[loss=0.2457, ctc_loss=0.1608, cr_loss=0.4244, over 17218.00 frames. ], tot_loss[loss=0.2138, ctc_loss=0.1421, cr_loss=0.3588, over 3362628.17 frames. ], batch size: 50, lr: 5.72e-03, grad_scale: 32.0 2024-09-23 23:41:46,716 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=372026.6666666667, ans=0.1 2024-09-23 23:41:51,931 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=6.99 vs. limit=15.0 2024-09-23 23:42:50,281 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=372213.3333333333, ans=0.125 2024-09-23 23:42:58,086 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=372213.3333333333, ans=0.1 2024-09-23 23:43:05,767 INFO [train.py:1198] (2/4) Epoch 21, batch 1850, loss[loss=0.2, ctc_loss=0.1292, cr_loss=0.3541, over 17084.00 frames. ], tot_loss[loss=0.2142, ctc_loss=0.1424, cr_loss=0.3589, over 3365181.24 frames. ], batch size: 46, lr: 5.72e-03, grad_scale: 32.0 2024-09-23 23:43:12,413 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=372260.0, ans=0.2 2024-09-23 23:43:25,013 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=372306.6666666667, ans=0.0 2024-09-23 23:43:39,436 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=372353.3333333333, ans=0.025 2024-09-23 23:43:44,324 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=372353.3333333333, ans=0.125 2024-09-23 23:43:45,460 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.067e+02 1.299e+02 1.413e+02 1.617e+02 3.494e+02, threshold=2.826e+02, percent-clipped=2.0 2024-09-23 23:43:49,055 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=372353.3333333333, ans=0.0 2024-09-23 23:44:28,064 INFO [train.py:1198] (2/4) Epoch 21, batch 1900, loss[loss=0.2077, ctc_loss=0.1381, cr_loss=0.3483, over 17301.00 frames. ], tot_loss[loss=0.2136, ctc_loss=0.1421, cr_loss=0.3577, over 3369910.97 frames. ], batch size: 46, lr: 5.71e-03, grad_scale: 32.0 2024-09-23 23:44:28,894 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.43 vs. limit=15.0 2024-09-23 23:44:58,268 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=372586.6666666667, ans=0.2 2024-09-23 23:45:03,163 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=372586.6666666667, ans=0.125 2024-09-23 23:45:13,442 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.68 vs. limit=15.0 2024-09-23 23:45:47,537 INFO [train.py:1198] (2/4) Epoch 21, batch 1950, loss[loss=0.1948, ctc_loss=0.1273, cr_loss=0.3375, over 17048.00 frames. ], tot_loss[loss=0.2121, ctc_loss=0.1408, cr_loss=0.3561, over 3379160.27 frames. ], batch size: 39, lr: 5.71e-03, grad_scale: 32.0 2024-09-23 23:45:47,823 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=372726.6666666667, ans=0.0 2024-09-23 23:46:17,494 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=372773.3333333333, ans=0.025 2024-09-23 23:46:22,261 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=372820.0, ans=0.125 2024-09-23 23:46:22,414 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=372820.0, ans=0.07 2024-09-23 23:46:26,963 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=372820.0, ans=0.125 2024-09-23 23:46:29,780 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.064e+02 1.272e+02 1.378e+02 1.476e+02 2.541e+02, threshold=2.755e+02, percent-clipped=0.0 2024-09-23 23:46:38,244 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=372866.6666666667, ans=0.2 2024-09-23 23:46:42,296 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=372866.6666666667, ans=0.2 2024-09-23 23:46:56,711 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=372913.3333333333, ans=0.125 2024-09-23 23:47:12,283 INFO [train.py:1198] (2/4) Epoch 21, batch 2000, loss[loss=0.2308, ctc_loss=0.153, cr_loss=0.389, over 17107.00 frames. ], tot_loss[loss=0.2135, ctc_loss=0.1419, cr_loss=0.358, over 3380872.96 frames. ], batch size: 49, lr: 5.71e-03, grad_scale: 32.0 2024-09-23 23:47:31,036 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=373006.6666666667, ans=0.5 2024-09-23 23:47:37,862 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=14.84 vs. limit=15.0 2024-09-23 23:47:50,384 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=373053.3333333333, ans=0.0 2024-09-23 23:47:50,535 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=373053.3333333333, ans=0.125 2024-09-23 23:47:52,448 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.81 vs. limit=15.0 2024-09-23 23:48:06,705 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.87 vs. limit=15.0 2024-09-23 23:48:34,486 INFO [train.py:1198] (2/4) Epoch 21, batch 2050, loss[loss=0.1702, ctc_loss=0.1116, cr_loss=0.2934, over 16969.00 frames. ], tot_loss[loss=0.2131, ctc_loss=0.1415, cr_loss=0.3579, over 3380941.58 frames. ], batch size: 42, lr: 5.71e-03, grad_scale: 32.0 2024-09-23 23:48:36,285 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=373193.3333333333, ans=0.2 2024-09-23 23:49:05,949 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module1.whiten, num_groups=1, num_channels=192, metric=7.83 vs. limit=15.0 2024-09-23 23:49:14,303 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.094e+02 1.256e+02 1.387e+02 1.531e+02 2.398e+02, threshold=2.775e+02, percent-clipped=0.0 2024-09-23 23:49:40,197 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=373333.3333333333, ans=0.125 2024-09-23 23:49:41,969 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=373380.0, ans=0.125 2024-09-23 23:49:50,081 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=373380.0, ans=0.125 2024-09-23 23:49:50,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=373380.0, ans=0.1 2024-09-23 23:49:59,550 INFO [train.py:1198] (2/4) Epoch 21, batch 2100, loss[loss=0.2291, ctc_loss=0.1503, cr_loss=0.3942, over 17265.00 frames. ], tot_loss[loss=0.2129, ctc_loss=0.1413, cr_loss=0.3579, over 3380154.17 frames. ], batch size: 46, lr: 5.71e-03, grad_scale: 32.0 2024-09-23 23:50:11,041 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=373426.6666666667, ans=0.125 2024-09-23 23:50:13,078 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.54 vs. limit=15.0 2024-09-23 23:50:15,929 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=373473.3333333333, ans=0.2 2024-09-23 23:51:05,032 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=373613.3333333333, ans=0.125 2024-09-23 23:51:22,279 INFO [train.py:1198] (2/4) Epoch 21, batch 2150, loss[loss=0.2205, ctc_loss=0.1454, cr_loss=0.3752, over 17292.00 frames. ], tot_loss[loss=0.2122, ctc_loss=0.1407, cr_loss=0.3571, over 3378631.83 frames. ], batch size: 49, lr: 5.71e-03, grad_scale: 32.0 2024-09-23 23:51:25,764 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=373660.0, ans=0.125 2024-09-23 23:51:46,517 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.25 vs. limit=15.0 2024-09-23 23:51:55,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=373753.3333333333, ans=0.0 2024-09-23 23:52:03,262 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=373753.3333333333, ans=0.1 2024-09-23 23:52:04,668 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.068e+02 1.257e+02 1.382e+02 1.555e+02 2.014e+02, threshold=2.764e+02, percent-clipped=0.0 2024-09-23 23:52:32,278 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=6.02 vs. limit=15.0 2024-09-23 23:52:34,566 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=373846.6666666667, ans=0.0 2024-09-23 23:52:47,116 INFO [train.py:1198] (2/4) Epoch 21, batch 2200, loss[loss=0.2125, ctc_loss=0.1414, cr_loss=0.3554, over 17066.00 frames. ], tot_loss[loss=0.2118, ctc_loss=0.1406, cr_loss=0.3561, over 3378312.03 frames. ], batch size: 46, lr: 5.70e-03, grad_scale: 32.0 2024-09-23 23:53:28,025 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.85 vs. limit=22.5 2024-09-23 23:53:29,096 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=373986.6666666667, ans=0.125 2024-09-23 23:53:32,228 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=373986.6666666667, ans=0.125 2024-09-23 23:53:34,019 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=374033.3333333333, ans=0.1 2024-09-23 23:53:35,635 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=374033.3333333333, ans=0.125 2024-09-23 23:53:38,906 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=9.90 vs. limit=15.0 2024-09-23 23:53:40,061 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=374033.3333333333, ans=0.0 2024-09-23 23:53:45,087 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.59 vs. limit=15.0 2024-09-23 23:53:47,952 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=374033.3333333333, ans=0.025 2024-09-23 23:53:55,895 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=374080.0, ans=0.125 2024-09-23 23:53:57,528 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=374080.0, ans=0.125 2024-09-23 23:54:06,785 INFO [train.py:1198] (2/4) Epoch 21, batch 2250, loss[loss=0.2001, ctc_loss=0.132, cr_loss=0.3408, over 16266.00 frames. ], tot_loss[loss=0.2123, ctc_loss=0.1409, cr_loss=0.3573, over 3379014.34 frames. ], batch size: 36, lr: 5.70e-03, grad_scale: 32.0 2024-09-23 23:54:43,207 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=374220.0, ans=0.035 2024-09-23 23:54:49,164 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.146e+02 1.304e+02 1.416e+02 1.549e+02 2.189e+02, threshold=2.832e+02, percent-clipped=0.0 2024-09-23 23:55:29,454 INFO [train.py:1198] (2/4) Epoch 21, batch 2300, loss[loss=0.1795, ctc_loss=0.1174, cr_loss=0.3105, over 17005.00 frames. ], tot_loss[loss=0.2105, ctc_loss=0.1396, cr_loss=0.3545, over 3379096.66 frames. ], batch size: 44, lr: 5.70e-03, grad_scale: 32.0 2024-09-23 23:55:47,191 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=374406.6666666667, ans=0.09899494936611666 2024-09-23 23:55:52,183 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=374406.6666666667, ans=0.125 2024-09-23 23:56:54,567 INFO [train.py:1198] (2/4) Epoch 21, batch 2350, loss[loss=0.2273, ctc_loss=0.1488, cr_loss=0.3924, over 17107.00 frames. ], tot_loss[loss=0.2106, ctc_loss=0.1396, cr_loss=0.3548, over 3373666.69 frames. ], batch size: 49, lr: 5.70e-03, grad_scale: 32.0 2024-09-23 23:57:16,352 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=16.70 vs. limit=22.5 2024-09-23 23:57:20,667 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=374640.0, ans=0.1 2024-09-23 23:57:38,981 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.011e+02 1.274e+02 1.389e+02 1.537e+02 2.355e+02, threshold=2.778e+02, percent-clipped=0.0 2024-09-23 23:57:52,148 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=374733.3333333333, ans=0.1 2024-09-23 23:58:06,407 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=374780.0, ans=0.125 2024-09-23 23:58:06,643 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=6.71 vs. limit=15.0 2024-09-23 23:58:12,725 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=374780.0, ans=0.0 2024-09-23 23:58:17,351 INFO [train.py:1198] (2/4) Epoch 21, batch 2400, loss[loss=0.2331, ctc_loss=0.1566, cr_loss=0.3826, over 16530.00 frames. ], tot_loss[loss=0.2112, ctc_loss=0.1402, cr_loss=0.3552, over 3369382.13 frames. ], batch size: 66, lr: 5.70e-03, grad_scale: 32.0 2024-09-23 23:58:27,737 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.46 vs. limit=15.0 2024-09-23 23:58:30,504 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=374826.6666666667, ans=0.125 2024-09-23 23:58:32,053 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=374873.3333333333, ans=0.125 2024-09-23 23:58:58,968 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-23 23:59:08,608 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=374966.6666666667, ans=0.0 2024-09-23 23:59:27,092 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=375013.3333333333, ans=0.125 2024-09-23 23:59:39,561 INFO [train.py:1198] (2/4) Epoch 21, batch 2450, loss[loss=0.1874, ctc_loss=0.1216, cr_loss=0.3286, over 17232.00 frames. ], tot_loss[loss=0.2112, ctc_loss=0.1403, cr_loss=0.3549, over 3357686.66 frames. ], batch size: 47, lr: 5.70e-03, grad_scale: 32.0 2024-09-23 23:59:58,892 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=375106.6666666667, ans=0.2 2024-09-24 00:00:20,974 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.089e+02 1.266e+02 1.364e+02 1.496e+02 2.065e+02, threshold=2.727e+02, percent-clipped=0.0 2024-09-24 00:00:26,123 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=375200.0, ans=0.125 2024-09-24 00:00:47,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=375246.6666666667, ans=0.025 2024-09-24 00:00:50,234 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=375246.6666666667, ans=0.125 2024-09-24 00:01:02,063 INFO [train.py:1198] (2/4) Epoch 21, batch 2500, loss[loss=0.2487, ctc_loss=0.1691, cr_loss=0.398, over 15812.00 frames. ], tot_loss[loss=0.2121, ctc_loss=0.141, cr_loss=0.3557, over 3345897.44 frames. ], batch size: 74, lr: 5.69e-03, grad_scale: 32.0 2024-09-24 00:01:10,200 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.min_positive, batch_count=375293.3333333333, ans=0.025 2024-09-24 00:01:18,544 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.22 vs. limit=10.0 2024-09-24 00:01:47,728 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.46 vs. limit=10.0 2024-09-24 00:02:01,150 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=6.17 vs. limit=15.0 2024-09-24 00:02:26,987 INFO [train.py:1198] (2/4) Epoch 21, batch 2550, loss[loss=0.1839, ctc_loss=0.1185, cr_loss=0.327, over 17090.00 frames. ], tot_loss[loss=0.2122, ctc_loss=0.141, cr_loss=0.356, over 3337453.26 frames. ], batch size: 43, lr: 5.69e-03, grad_scale: 32.0 2024-09-24 00:02:44,844 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=375573.3333333333, ans=0.2 2024-09-24 00:03:08,316 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.019e+02 1.246e+02 1.350e+02 1.498e+02 2.836e+02, threshold=2.701e+02, percent-clipped=1.0 2024-09-24 00:03:11,843 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=375620.0, ans=0.2 2024-09-24 00:03:46,557 INFO [train.py:1198] (2/4) Epoch 21, batch 2600, loss[loss=0.2201, ctc_loss=0.147, cr_loss=0.3655, over 17302.00 frames. ], tot_loss[loss=0.2124, ctc_loss=0.1411, cr_loss=0.3567, over 3340464.42 frames. ], batch size: 46, lr: 5.69e-03, grad_scale: 32.0 2024-09-24 00:03:49,190 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=3.94 vs. limit=15.0 2024-09-24 00:04:02,931 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=375806.6666666667, ans=0.125 2024-09-24 00:04:43,871 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=8.90 vs. limit=15.0 2024-09-24 00:04:56,693 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.64 vs. limit=15.0 2024-09-24 00:05:08,529 INFO [train.py:1198] (2/4) Epoch 21, batch 2650, loss[loss=0.2511, ctc_loss=0.1696, cr_loss=0.4076, over 17017.00 frames. ], tot_loss[loss=0.2134, ctc_loss=0.1418, cr_loss=0.3577, over 3337608.36 frames. ], batch size: 56, lr: 5.69e-03, grad_scale: 32.0 2024-09-24 00:05:15,085 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=375993.3333333333, ans=0.05 2024-09-24 00:05:15,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=375993.3333333333, ans=0.0 2024-09-24 00:05:43,007 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.65 vs. limit=22.5 2024-09-24 00:05:52,572 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.095e+02 1.243e+02 1.351e+02 1.464e+02 2.576e+02, threshold=2.701e+02, percent-clipped=0.0 2024-09-24 00:06:32,297 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 00:06:33,562 INFO [train.py:1198] (2/4) Epoch 21, batch 2700, loss[loss=0.1927, ctc_loss=0.1256, cr_loss=0.3357, over 17013.00 frames. ], tot_loss[loss=0.213, ctc_loss=0.1414, cr_loss=0.3577, over 3343500.73 frames. ], batch size: 51, lr: 5.69e-03, grad_scale: 32.0 2024-09-24 00:06:49,639 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=376273.3333333333, ans=0.1 2024-09-24 00:07:23,811 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=376366.6666666667, ans=0.125 2024-09-24 00:07:41,417 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=376413.3333333333, ans=0.125 2024-09-24 00:07:55,724 INFO [train.py:1198] (2/4) Epoch 21, batch 2750, loss[loss=0.2597, ctc_loss=0.1727, cr_loss=0.4349, over 17037.00 frames. ], tot_loss[loss=0.2135, ctc_loss=0.1418, cr_loss=0.3586, over 3341418.70 frames. ], batch size: 52, lr: 5.68e-03, grad_scale: 32.0 2024-09-24 00:08:29,854 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.min_positive, batch_count=376553.3333333333, ans=0.05 2024-09-24 00:08:31,363 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=376553.3333333333, ans=0.125 2024-09-24 00:08:37,708 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.067e+02 1.226e+02 1.342e+02 1.468e+02 2.196e+02, threshold=2.683e+02, percent-clipped=0.0 2024-09-24 00:08:50,593 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=376600.0, ans=0.1 2024-09-24 00:09:13,678 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=376646.6666666667, ans=0.125 2024-09-24 00:09:18,226 INFO [train.py:1198] (2/4) Epoch 21, batch 2800, loss[loss=0.2225, ctc_loss=0.1471, cr_loss=0.3769, over 17285.00 frames. ], tot_loss[loss=0.2116, ctc_loss=0.1406, cr_loss=0.3552, over 3349758.55 frames. ], batch size: 51, lr: 5.68e-03, grad_scale: 32.0 2024-09-24 00:10:38,125 INFO [train.py:1198] (2/4) Epoch 21, batch 2850, loss[loss=0.2221, ctc_loss=0.1458, cr_loss=0.3815, over 17200.00 frames. ], tot_loss[loss=0.2117, ctc_loss=0.1405, cr_loss=0.3559, over 3352684.24 frames. ], batch size: 55, lr: 5.68e-03, grad_scale: 32.0 2024-09-24 00:10:43,139 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=376926.6666666667, ans=0.125 2024-09-24 00:11:17,912 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=377020.0, ans=0.125 2024-09-24 00:11:22,307 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.032e+02 1.291e+02 1.370e+02 1.483e+02 1.856e+02, threshold=2.740e+02, percent-clipped=0.0 2024-09-24 00:11:35,105 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=377066.6666666667, ans=0.0 2024-09-24 00:11:41,537 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=377066.6666666667, ans=0.0 2024-09-24 00:11:44,576 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=377066.6666666667, ans=0.0 2024-09-24 00:11:51,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=377113.3333333333, ans=0.125 2024-09-24 00:12:03,409 INFO [train.py:1198] (2/4) Epoch 21, batch 2900, loss[loss=0.2571, ctc_loss=0.1763, cr_loss=0.4043, over 16636.00 frames. ], tot_loss[loss=0.2112, ctc_loss=0.1403, cr_loss=0.3549, over 3351273.00 frames. ], batch size: 66, lr: 5.68e-03, grad_scale: 32.0 2024-09-24 00:12:49,626 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 00:13:02,412 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=377300.0, ans=0.0 2024-09-24 00:13:23,380 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=377346.6666666667, ans=0.5 2024-09-24 00:13:26,192 INFO [train.py:1198] (2/4) Epoch 21, batch 2950, loss[loss=0.1649, ctc_loss=0.1075, cr_loss=0.2872, over 16274.00 frames. ], tot_loss[loss=0.2114, ctc_loss=0.1405, cr_loss=0.3547, over 3337262.60 frames. ], batch size: 36, lr: 5.68e-03, grad_scale: 16.0 2024-09-24 00:13:31,914 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=21.54 vs. limit=22.5 2024-09-24 00:13:39,482 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn2.whiten.whitening_limit, batch_count=377393.3333333333, ans=22.5 2024-09-24 00:13:47,446 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.01 vs. limit=15.0 2024-09-24 00:14:11,336 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.234e+02 1.337e+02 1.460e+02 2.034e+02, threshold=2.674e+02, percent-clipped=0.0 2024-09-24 00:14:13,510 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=11.70 vs. limit=15.0 2024-09-24 00:14:19,356 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=377533.3333333333, ans=0.2 2024-09-24 00:14:33,559 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=377580.0, ans=0.125 2024-09-24 00:14:47,420 INFO [train.py:1198] (2/4) Epoch 21, batch 3000, loss[loss=0.1823, ctc_loss=0.1154, cr_loss=0.3344, over 17107.00 frames. ], tot_loss[loss=0.2107, ctc_loss=0.1399, cr_loss=0.354, over 3344005.25 frames. ], batch size: 40, lr: 5.68e-03, grad_scale: 16.0 2024-09-24 00:14:47,421 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 00:15:01,271 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.0.layers.0.self_attn_weights, attn_weights_entropy = tensor([5.3249, 5.0781, 4.6465, 4.8893], device='cuda:2') 2024-09-24 00:15:02,880 INFO [train.py:1230] (2/4) Epoch 21, validation: loss=0.03893, ctc_loss=0.03893, cr_loss=7.803e-15, over 944034.00 frames. 2024-09-24 00:15:02,880 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 00:15:32,784 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.96 vs. limit=6.0 2024-09-24 00:15:38,047 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=377720.0, ans=0.0 2024-09-24 00:15:45,837 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=377720.0, ans=0.125 2024-09-24 00:16:21,772 INFO [train.py:1198] (2/4) Epoch 21, batch 3050, loss[loss=0.2013, ctc_loss=0.1317, cr_loss=0.348, over 17289.00 frames. ], tot_loss[loss=0.2122, ctc_loss=0.141, cr_loss=0.3561, over 3343790.60 frames. ], batch size: 42, lr: 5.67e-03, grad_scale: 16.0 2024-09-24 00:16:42,531 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=377906.6666666667, ans=0.0 2024-09-24 00:17:04,477 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.136e+02 1.246e+02 1.316e+02 1.406e+02 3.085e+02, threshold=2.632e+02, percent-clipped=1.0 2024-09-24 00:17:06,304 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=377953.3333333333, ans=0.125 2024-09-24 00:17:43,200 INFO [train.py:1198] (2/4) Epoch 21, batch 3100, loss[loss=0.2206, ctc_loss=0.1485, cr_loss=0.3602, over 15953.00 frames. ], tot_loss[loss=0.2123, ctc_loss=0.1411, cr_loss=0.3562, over 3332137.81 frames. ], batch size: 74, lr: 5.67e-03, grad_scale: 16.0 2024-09-24 00:18:04,151 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=378140.0, ans=0.2 2024-09-24 00:18:15,553 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=9.01 vs. limit=12.0 2024-09-24 00:18:46,905 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.max_abs, batch_count=378280.0, ans=10.0 2024-09-24 00:18:55,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten.whitening_limit, batch_count=378280.0, ans=15.0 2024-09-24 00:19:00,053 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.15 vs. limit=22.5 2024-09-24 00:19:04,078 INFO [train.py:1198] (2/4) Epoch 21, batch 3150, loss[loss=0.1862, ctc_loss=0.1234, cr_loss=0.314, over 16674.00 frames. ], tot_loss[loss=0.2128, ctc_loss=0.1414, cr_loss=0.3572, over 3329489.60 frames. ], batch size: 37, lr: 5.67e-03, grad_scale: 16.0 2024-09-24 00:19:19,984 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=378373.3333333333, ans=0.0 2024-09-24 00:19:46,189 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 9.987e+01 1.255e+02 1.337e+02 1.461e+02 2.046e+02, threshold=2.675e+02, percent-clipped=0.0 2024-09-24 00:19:48,077 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=378420.0, ans=0.0 2024-09-24 00:19:51,029 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=378466.6666666667, ans=0.0 2024-09-24 00:20:21,949 INFO [train.py:1198] (2/4) Epoch 21, batch 3200, loss[loss=0.2617, ctc_loss=0.181, cr_loss=0.4031, over 16516.00 frames. ], tot_loss[loss=0.2134, ctc_loss=0.1418, cr_loss=0.3581, over 3334758.87 frames. ], batch size: 66, lr: 5.67e-03, grad_scale: 32.0 2024-09-24 00:20:37,137 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=378560.0, ans=0.125 2024-09-24 00:20:47,887 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=378606.6666666667, ans=0.0 2024-09-24 00:20:47,973 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=378606.6666666667, ans=0.0 2024-09-24 00:21:01,804 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 00:21:06,400 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=378653.3333333333, ans=0.2 2024-09-24 00:21:42,247 INFO [train.py:1198] (2/4) Epoch 21, batch 3250, loss[loss=0.1822, ctc_loss=0.1156, cr_loss=0.3327, over 17116.00 frames. ], tot_loss[loss=0.2129, ctc_loss=0.1413, cr_loss=0.3579, over 3349132.95 frames. ], batch size: 40, lr: 5.67e-03, grad_scale: 32.0 2024-09-24 00:22:08,074 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.61 vs. limit=15.0 2024-09-24 00:22:12,409 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=378886.6666666667, ans=0.0 2024-09-24 00:22:15,277 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=378886.6666666667, ans=0.125 2024-09-24 00:22:24,295 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.107e+02 1.280e+02 1.348e+02 1.474e+02 1.911e+02, threshold=2.696e+02, percent-clipped=0.0 2024-09-24 00:22:29,517 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer_na.min_abs, batch_count=378933.3333333333, ans=0.02 2024-09-24 00:22:32,589 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=378933.3333333333, ans=0.125 2024-09-24 00:22:41,727 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=378933.3333333333, ans=0.125 2024-09-24 00:22:43,214 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=378980.0, ans=0.125 2024-09-24 00:22:52,699 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=378980.0, ans=0.1 2024-09-24 00:22:58,918 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 00:22:59,367 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.99 vs. limit=15.0 2024-09-24 00:23:00,167 INFO [train.py:1198] (2/4) Epoch 21, batch 3300, loss[loss=0.2301, ctc_loss=0.1543, cr_loss=0.3786, over 16636.00 frames. ], tot_loss[loss=0.213, ctc_loss=0.1414, cr_loss=0.358, over 3355702.32 frames. ], batch size: 61, lr: 5.67e-03, grad_scale: 32.0 2024-09-24 00:23:03,618 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=379026.6666666667, ans=0.125 2024-09-24 00:23:36,741 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=10.65 vs. limit=22.5 2024-09-24 00:23:50,319 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=379166.6666666667, ans=0.025 2024-09-24 00:23:58,130 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=379166.6666666667, ans=0.125 2024-09-24 00:24:07,681 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=379213.3333333333, ans=0.1 2024-09-24 00:24:15,941 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.51 vs. limit=15.0 2024-09-24 00:24:18,313 INFO [train.py:1198] (2/4) Epoch 21, batch 3350, loss[loss=0.1619, ctc_loss=0.1046, cr_loss=0.2865, over 17115.00 frames. ], tot_loss[loss=0.2122, ctc_loss=0.1409, cr_loss=0.3568, over 3357097.40 frames. ], batch size: 40, lr: 5.66e-03, grad_scale: 32.0 2024-09-24 00:24:48,782 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.min_abs, batch_count=379306.6666666667, ans=0.5 2024-09-24 00:24:50,458 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=379353.3333333333, ans=0.0 2024-09-24 00:25:01,139 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=379353.3333333333, ans=0.0 2024-09-24 00:25:02,489 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.292e+02 1.373e+02 1.509e+02 2.435e+02, threshold=2.745e+02, percent-clipped=0.0 2024-09-24 00:25:12,356 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=379400.0, ans=0.125 2024-09-24 00:25:18,575 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=379400.0, ans=0.0 2024-09-24 00:25:38,757 INFO [train.py:1198] (2/4) Epoch 21, batch 3400, loss[loss=0.2328, ctc_loss=0.1602, cr_loss=0.3628, over 16887.00 frames. ], tot_loss[loss=0.2129, ctc_loss=0.1414, cr_loss=0.3576, over 3365013.13 frames. ], batch size: 58, lr: 5.66e-03, grad_scale: 32.0 2024-09-24 00:25:51,507 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=379493.3333333333, ans=0.2 2024-09-24 00:26:05,602 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=379540.0, ans=0.125 2024-09-24 00:26:56,713 INFO [train.py:1198] (2/4) Epoch 21, batch 3450, loss[loss=0.1893, ctc_loss=0.1236, cr_loss=0.3289, over 17082.00 frames. ], tot_loss[loss=0.2125, ctc_loss=0.1412, cr_loss=0.3565, over 3363348.47 frames. ], batch size: 43, lr: 5.66e-03, grad_scale: 32.0 2024-09-24 00:27:23,983 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=17.10 vs. limit=22.5 2024-09-24 00:27:38,510 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.144e+02 1.292e+02 1.385e+02 1.500e+02 2.362e+02, threshold=2.770e+02, percent-clipped=0.0 2024-09-24 00:27:41,968 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=379866.6666666667, ans=0.2 2024-09-24 00:28:00,078 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=10.26 vs. limit=15.0 2024-09-24 00:28:16,467 INFO [train.py:1198] (2/4) Epoch 21, batch 3500, loss[loss=0.1833, ctc_loss=0.1191, cr_loss=0.3208, over 16967.00 frames. ], tot_loss[loss=0.2124, ctc_loss=0.1411, cr_loss=0.3561, over 3359423.53 frames. ], batch size: 42, lr: 5.66e-03, grad_scale: 32.0 2024-09-24 00:28:40,638 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=380006.6666666667, ans=0.2 2024-09-24 00:28:55,014 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=380053.3333333333, ans=0.125 2024-09-24 00:29:08,774 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=380100.0, ans=0.0 2024-09-24 00:29:18,588 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=6.45 vs. limit=15.0 2024-09-24 00:29:32,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=380146.6666666667, ans=0.0 2024-09-24 00:29:36,654 INFO [train.py:1198] (2/4) Epoch 21, batch 3550, loss[loss=0.225, ctc_loss=0.1514, cr_loss=0.3677, over 17077.00 frames. ], tot_loss[loss=0.2119, ctc_loss=0.1407, cr_loss=0.3558, over 3357344.58 frames. ], batch size: 46, lr: 5.66e-03, grad_scale: 32.0 2024-09-24 00:29:52,583 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=380240.0, ans=0.025 2024-09-24 00:30:00,848 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.88 vs. limit=6.0 2024-09-24 00:30:20,797 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.084e+02 1.243e+02 1.324e+02 1.443e+02 2.322e+02, threshold=2.648e+02, percent-clipped=0.0 2024-09-24 00:30:33,530 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=380333.3333333333, ans=0.125 2024-09-24 00:30:39,797 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=380380.0, ans=0.125 2024-09-24 00:30:49,865 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=10.96 vs. limit=15.0 2024-09-24 00:30:56,886 INFO [train.py:1198] (2/4) Epoch 21, batch 3600, loss[loss=0.2057, ctc_loss=0.1361, cr_loss=0.3479, over 17001.00 frames. ], tot_loss[loss=0.2115, ctc_loss=0.1403, cr_loss=0.3558, over 3358995.25 frames. ], batch size: 52, lr: 5.66e-03, grad_scale: 32.0 2024-09-24 00:30:58,603 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=380426.6666666667, ans=0.125 2024-09-24 00:31:26,614 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=380520.0, ans=0.0 2024-09-24 00:31:51,659 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=380566.6666666667, ans=0.2 2024-09-24 00:32:03,968 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=380613.3333333333, ans=0.125 2024-09-24 00:32:05,810 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=380613.3333333333, ans=0.2 2024-09-24 00:32:06,109 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=14.06 vs. limit=22.5 2024-09-24 00:32:09,096 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=380613.3333333333, ans=0.1 2024-09-24 00:32:13,546 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=380660.0, ans=0.125 2024-09-24 00:32:14,557 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=9.45 vs. limit=15.0 2024-09-24 00:32:14,911 INFO [train.py:1198] (2/4) Epoch 21, batch 3650, loss[loss=0.1908, ctc_loss=0.1234, cr_loss=0.3367, over 17229.00 frames. ], tot_loss[loss=0.2118, ctc_loss=0.1405, cr_loss=0.3564, over 3357810.70 frames. ], batch size: 50, lr: 5.65e-03, grad_scale: 32.0 2024-09-24 00:32:21,860 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=11.98 vs. limit=22.5 2024-09-24 00:32:35,330 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=380706.6666666667, ans=0.125 2024-09-24 00:32:35,670 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=5.56 vs. limit=12.0 2024-09-24 00:32:45,321 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=380753.3333333333, ans=0.125 2024-09-24 00:32:53,243 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=380753.3333333333, ans=0.0 2024-09-24 00:32:57,524 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.078e+02 1.280e+02 1.369e+02 1.514e+02 2.640e+02, threshold=2.739e+02, percent-clipped=0.0 2024-09-24 00:33:02,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=380800.0, ans=0.025 2024-09-24 00:33:16,984 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=380846.6666666667, ans=0.125 2024-09-24 00:33:35,127 INFO [train.py:1198] (2/4) Epoch 21, batch 3700, loss[loss=0.1729, ctc_loss=0.1122, cr_loss=0.3034, over 17309.00 frames. ], tot_loss[loss=0.2139, ctc_loss=0.142, cr_loss=0.3595, over 3355165.74 frames. ], batch size: 46, lr: 5.65e-03, grad_scale: 32.0 2024-09-24 00:33:35,537 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 00:33:41,554 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=380893.3333333333, ans=0.1 2024-09-24 00:34:00,546 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=380940.0, ans=0.0 2024-09-24 00:34:20,947 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=381033.3333333333, ans=0.125 2024-09-24 00:34:40,232 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=381080.0, ans=0.125 2024-09-24 00:34:53,989 INFO [train.py:1198] (2/4) Epoch 21, batch 3750, loss[loss=0.2429, ctc_loss=0.1647, cr_loss=0.391, over 14910.00 frames. ], tot_loss[loss=0.2135, ctc_loss=0.1418, cr_loss=0.3586, over 3347906.01 frames. ], batch size: 89, lr: 5.65e-03, grad_scale: 32.0 2024-09-24 00:34:58,931 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=381126.6666666667, ans=0.125 2024-09-24 00:35:16,470 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=381173.3333333333, ans=0.1 2024-09-24 00:35:17,805 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=381173.3333333333, ans=0.125 2024-09-24 00:35:25,681 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=381220.0, ans=0.125 2024-09-24 00:35:25,729 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=381220.0, ans=0.2 2024-09-24 00:35:36,265 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.058e+02 1.271e+02 1.353e+02 1.516e+02 2.351e+02, threshold=2.707e+02, percent-clipped=0.0 2024-09-24 00:35:55,476 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=381313.3333333333, ans=0.125 2024-09-24 00:36:08,085 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=381313.3333333333, ans=0.0 2024-09-24 00:36:11,308 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 00:36:12,540 INFO [train.py:1198] (2/4) Epoch 21, batch 3800, loss[loss=0.2365, ctc_loss=0.1624, cr_loss=0.3705, over 15062.00 frames. ], tot_loss[loss=0.2136, ctc_loss=0.142, cr_loss=0.3578, over 3329973.16 frames. ], batch size: 89, lr: 5.65e-03, grad_scale: 32.0 2024-09-24 00:36:25,332 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=381360.0, ans=0.125 2024-09-24 00:36:34,604 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=381406.6666666667, ans=0.125 2024-09-24 00:36:39,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=381406.6666666667, ans=0.04949747468305833 2024-09-24 00:36:40,933 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=381406.6666666667, ans=0.1 2024-09-24 00:36:54,484 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=11.93 vs. limit=15.0 2024-09-24 00:37:17,465 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=381546.6666666667, ans=0.2 2024-09-24 00:37:31,011 INFO [train.py:1198] (2/4) Epoch 21, batch 3850, loss[loss=0.2091, ctc_loss=0.1389, cr_loss=0.351, over 17138.00 frames. ], tot_loss[loss=0.2153, ctc_loss=0.1434, cr_loss=0.3592, over 3301067.68 frames. ], batch size: 48, lr: 5.65e-03, grad_scale: 32.0 2024-09-24 00:37:31,201 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=381593.3333333333, ans=0.0 2024-09-24 00:37:37,284 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=381593.3333333333, ans=0.0 2024-09-24 00:37:38,207 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.33 vs. limit=15.0 2024-09-24 00:37:39,137 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=381593.3333333333, ans=0.125 2024-09-24 00:38:13,072 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=381686.6666666667, ans=0.0 2024-09-24 00:38:14,455 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.170e+02 1.386e+02 1.498e+02 1.641e+02 2.451e+02, threshold=2.996e+02, percent-clipped=0.0 2024-09-24 00:38:24,331 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=381733.3333333333, ans=0.125 2024-09-24 00:38:30,761 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=381733.3333333333, ans=0.0 2024-09-24 00:39:39,373 INFO [train.py:1198] (2/4) Epoch 22, batch 0, loss[loss=0.2158, ctc_loss=0.1419, cr_loss=0.3695, over 17227.00 frames. ], tot_loss[loss=0.2158, ctc_loss=0.1419, cr_loss=0.3695, over 17227.00 frames. ], batch size: 44, lr: 5.51e-03, grad_scale: 32.0 2024-09-24 00:39:39,374 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 00:39:54,638 INFO [train.py:1230] (2/4) Epoch 22, validation: loss=0.03827, ctc_loss=0.03827, cr_loss=8.092e-15, over 944034.00 frames. 2024-09-24 00:39:54,639 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 00:39:54,928 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.max_abs, batch_count=381808.0, ans=10.0 2024-09-24 00:40:02,876 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=381808.0, ans=0.025 2024-09-24 00:40:07,581 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=381808.0, ans=0.125 2024-09-24 00:40:42,807 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.22 vs. limit=15.0 2024-09-24 00:41:01,948 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=16.20 vs. limit=22.5 2024-09-24 00:41:12,953 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=381994.6666666667, ans=0.025 2024-09-24 00:41:16,631 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.42 vs. limit=10.0 2024-09-24 00:41:17,589 INFO [train.py:1198] (2/4) Epoch 22, batch 50, loss[loss=0.2167, ctc_loss=0.1439, cr_loss=0.3642, over 17043.00 frames. ], tot_loss[loss=0.2087, ctc_loss=0.1382, cr_loss=0.3526, over 760823.39 frames. ], batch size: 39, lr: 5.51e-03, grad_scale: 32.0 2024-09-24 00:42:06,803 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.269e+02 1.377e+02 1.581e+02 4.736e+02, threshold=2.753e+02, percent-clipped=1.0 2024-09-24 00:42:15,030 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=382181.3333333333, ans=0.125 2024-09-24 00:42:28,006 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=382228.0, ans=0.0 2024-09-24 00:42:29,632 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=382228.0, ans=0.125 2024-09-24 00:42:40,201 INFO [train.py:1198] (2/4) Epoch 22, batch 100, loss[loss=0.2354, ctc_loss=0.157, cr_loss=0.3921, over 17083.00 frames. ], tot_loss[loss=0.2084, ctc_loss=0.138, cr_loss=0.3523, over 1337178.50 frames. ], batch size: 49, lr: 5.51e-03, grad_scale: 32.0 2024-09-24 00:42:42,450 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.36 vs. limit=15.0 2024-09-24 00:42:43,537 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer_ff2.min_abs, batch_count=382274.6666666667, ans=0.1 2024-09-24 00:43:08,811 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=382321.3333333333, ans=0.1 2024-09-24 00:43:10,326 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=382368.0, ans=0.5 2024-09-24 00:43:16,744 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=382368.0, ans=0.125 2024-09-24 00:43:28,958 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.whiten, num_groups=1, num_channels=192, metric=5.55 vs. limit=12.0 2024-09-24 00:43:39,822 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=5.64 vs. limit=12.0 2024-09-24 00:43:53,811 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-24 00:43:58,490 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=382508.0, ans=0.125 2024-09-24 00:43:59,805 INFO [train.py:1198] (2/4) Epoch 22, batch 150, loss[loss=0.2492, ctc_loss=0.166, cr_loss=0.4155, over 17024.00 frames. ], tot_loss[loss=0.211, ctc_loss=0.1399, cr_loss=0.3553, over 1778719.68 frames. ], batch size: 53, lr: 5.51e-03, grad_scale: 32.0 2024-09-24 00:44:19,099 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 00:44:55,447 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.262e+02 1.352e+02 1.515e+02 2.166e+02, threshold=2.703e+02, percent-clipped=0.0 2024-09-24 00:44:55,727 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=382648.0, ans=0.125 2024-09-24 00:45:01,769 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=13.35 vs. limit=22.5 2024-09-24 00:45:04,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.max_abs, batch_count=382648.0, ans=10.0 2024-09-24 00:45:09,004 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=382694.6666666667, ans=0.0 2024-09-24 00:45:10,496 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=382694.6666666667, ans=0.1 2024-09-24 00:45:16,793 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=382694.6666666667, ans=0.125 2024-09-24 00:45:29,258 INFO [train.py:1198] (2/4) Epoch 22, batch 200, loss[loss=0.212, ctc_loss=0.1419, cr_loss=0.3505, over 16781.00 frames. ], tot_loss[loss=0.2112, ctc_loss=0.1401, cr_loss=0.3553, over 2128935.18 frames. ], batch size: 61, lr: 5.51e-03, grad_scale: 32.0 2024-09-24 00:46:21,968 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=382881.3333333333, ans=0.0 2024-09-24 00:46:48,687 INFO [train.py:1198] (2/4) Epoch 22, batch 250, loss[loss=0.2395, ctc_loss=0.1584, cr_loss=0.4056, over 16496.00 frames. ], tot_loss[loss=0.2116, ctc_loss=0.1403, cr_loss=0.3565, over 2413277.01 frames. ], batch size: 66, lr: 5.50e-03, grad_scale: 32.0 2024-09-24 00:46:55,604 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 00:47:11,699 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=383021.3333333333, ans=0.09899494936611666 2024-09-24 00:47:13,327 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 00:47:41,217 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.081e+02 1.260e+02 1.348e+02 1.458e+02 2.828e+02, threshold=2.696e+02, percent-clipped=1.0 2024-09-24 00:47:49,555 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=383114.6666666667, ans=0.125 2024-09-24 00:47:55,846 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=383161.3333333333, ans=0.1 2024-09-24 00:47:57,594 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=383161.3333333333, ans=0.125 2024-09-24 00:48:11,508 INFO [train.py:1198] (2/4) Epoch 22, batch 300, loss[loss=0.1973, ctc_loss=0.1318, cr_loss=0.3275, over 17057.00 frames. ], tot_loss[loss=0.2115, ctc_loss=0.1403, cr_loss=0.3564, over 2628070.47 frames. ], batch size: 39, lr: 5.50e-03, grad_scale: 32.0 2024-09-24 00:48:18,051 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=383208.0, ans=0.125 2024-09-24 00:48:42,367 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=383301.3333333333, ans=0.1 2024-09-24 00:48:45,579 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=383301.3333333333, ans=0.0 2024-09-24 00:49:04,425 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=383348.0, ans=0.2 2024-09-24 00:49:23,469 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=383394.6666666667, ans=0.125 2024-09-24 00:49:33,029 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module1.whiten, num_groups=1, num_channels=512, metric=6.05 vs. limit=15.0 2024-09-24 00:49:37,075 INFO [train.py:1198] (2/4) Epoch 22, batch 350, loss[loss=0.1804, ctc_loss=0.1148, cr_loss=0.328, over 16928.00 frames. ], tot_loss[loss=0.2107, ctc_loss=0.1396, cr_loss=0.3555, over 2790410.92 frames. ], batch size: 42, lr: 5.50e-03, grad_scale: 32.0 2024-09-24 00:49:37,458 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=383441.3333333333, ans=0.2 2024-09-24 00:49:47,102 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=383441.3333333333, ans=0.125 2024-09-24 00:49:50,192 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=383441.3333333333, ans=0.0 2024-09-24 00:49:57,967 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=383488.0, ans=0.09899494936611666 2024-09-24 00:50:09,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=383534.6666666667, ans=0.1 2024-09-24 00:50:12,374 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=383534.6666666667, ans=0.1 2024-09-24 00:50:19,733 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=383534.6666666667, ans=0.0 2024-09-24 00:50:28,763 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.255e+02 1.333e+02 1.486e+02 2.174e+02, threshold=2.666e+02, percent-clipped=0.0 2024-09-24 00:50:59,529 INFO [train.py:1198] (2/4) Epoch 22, batch 400, loss[loss=0.2037, ctc_loss=0.1359, cr_loss=0.3388, over 17232.00 frames. ], tot_loss[loss=0.2111, ctc_loss=0.1399, cr_loss=0.3558, over 2923264.71 frames. ], batch size: 50, lr: 5.50e-03, grad_scale: 32.0 2024-09-24 00:51:23,606 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=383721.3333333333, ans=0.125 2024-09-24 00:51:33,246 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=383768.0, ans=0.125 2024-09-24 00:52:19,399 INFO [train.py:1198] (2/4) Epoch 22, batch 450, loss[loss=0.192, ctc_loss=0.1252, cr_loss=0.3338, over 16316.00 frames. ], tot_loss[loss=0.2118, ctc_loss=0.1404, cr_loss=0.357, over 3016315.63 frames. ], batch size: 36, lr: 5.50e-03, grad_scale: 32.0 2024-09-24 00:52:21,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=383908.0, ans=0.0 2024-09-24 00:52:48,249 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=383954.6666666667, ans=0.125 2024-09-24 00:52:59,197 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=384001.3333333333, ans=0.025 2024-09-24 00:53:00,983 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=384001.3333333333, ans=0.0 2024-09-24 00:53:11,769 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.087e+02 1.289e+02 1.376e+02 1.526e+02 3.562e+02, threshold=2.753e+02, percent-clipped=1.0 2024-09-24 00:53:12,023 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=384048.0, ans=0.0 2024-09-24 00:53:38,172 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.51 vs. limit=15.0 2024-09-24 00:53:41,925 INFO [train.py:1198] (2/4) Epoch 22, batch 500, loss[loss=0.2384, ctc_loss=0.1612, cr_loss=0.3863, over 17009.00 frames. ], tot_loss[loss=0.2124, ctc_loss=0.1408, cr_loss=0.358, over 3089456.85 frames. ], batch size: 51, lr: 5.50e-03, grad_scale: 32.0 2024-09-24 00:54:41,981 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=384281.3333333333, ans=0.125 2024-09-24 00:54:43,520 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=384281.3333333333, ans=0.1 2024-09-24 00:54:54,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=384328.0, ans=0.05 2024-09-24 00:54:54,687 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=384328.0, ans=0.125 2024-09-24 00:54:56,114 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=384328.0, ans=0.0 2024-09-24 00:55:00,983 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=384328.0, ans=0.125 2024-09-24 00:55:02,525 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=384328.0, ans=0.1 2024-09-24 00:55:07,116 INFO [train.py:1198] (2/4) Epoch 22, batch 550, loss[loss=0.1804, ctc_loss=0.1145, cr_loss=0.3296, over 17251.00 frames. ], tot_loss[loss=0.212, ctc_loss=0.1405, cr_loss=0.3576, over 3155034.92 frames. ], batch size: 44, lr: 5.49e-03, grad_scale: 32.0 2024-09-24 00:55:41,059 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=11.48 vs. limit=15.0 2024-09-24 00:55:59,357 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.087e+02 1.258e+02 1.357e+02 1.512e+02 2.238e+02, threshold=2.715e+02, percent-clipped=0.0 2024-09-24 00:56:24,147 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=384561.3333333333, ans=0.1 2024-09-24 00:56:30,183 INFO [train.py:1198] (2/4) Epoch 22, batch 600, loss[loss=0.204, ctc_loss=0.1342, cr_loss=0.3493, over 16846.00 frames. ], tot_loss[loss=0.2116, ctc_loss=0.1403, cr_loss=0.3568, over 3198111.52 frames. ], batch size: 61, lr: 5.49e-03, grad_scale: 32.0 2024-09-24 00:56:44,926 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=384654.6666666667, ans=0.125 2024-09-24 00:56:58,403 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.25 vs. limit=22.5 2024-09-24 00:57:02,480 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=384701.3333333333, ans=0.0 2024-09-24 00:57:46,393 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=384794.6666666667, ans=0.1 2024-09-24 00:57:47,090 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.38 vs. limit=15.0 2024-09-24 00:57:52,481 INFO [train.py:1198] (2/4) Epoch 22, batch 650, loss[loss=0.2178, ctc_loss=0.1423, cr_loss=0.3776, over 16819.00 frames. ], tot_loss[loss=0.2122, ctc_loss=0.1407, cr_loss=0.3572, over 3223881.78 frames. ], batch size: 61, lr: 5.49e-03, grad_scale: 16.0 2024-09-24 00:57:58,818 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=384841.3333333333, ans=0.2 2024-09-24 00:58:31,012 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=384934.6666666667, ans=0.0 2024-09-24 00:58:43,614 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.116e+02 1.285e+02 1.401e+02 1.551e+02 2.497e+02, threshold=2.802e+02, percent-clipped=0.0 2024-09-24 00:58:48,829 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=384981.3333333333, ans=0.035 2024-09-24 00:59:06,364 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=385028.0, ans=0.04949747468305833 2024-09-24 00:59:15,116 INFO [train.py:1198] (2/4) Epoch 22, batch 700, loss[loss=0.2317, ctc_loss=0.1575, cr_loss=0.371, over 17019.00 frames. ], tot_loss[loss=0.2122, ctc_loss=0.1408, cr_loss=0.3568, over 3260534.12 frames. ], batch size: 51, lr: 5.49e-03, grad_scale: 16.0 2024-09-24 00:59:17,048 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=385074.6666666667, ans=0.025 2024-09-24 00:59:19,285 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=16.96 vs. limit=22.5 2024-09-24 00:59:33,013 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.87 vs. limit=10.0 2024-09-24 00:59:51,697 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=385168.0, ans=0.125 2024-09-24 00:59:53,315 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=385168.0, ans=0.0 2024-09-24 00:59:59,520 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=385168.0, ans=0.1 2024-09-24 01:00:34,586 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=8.21 vs. limit=15.0 2024-09-24 01:00:40,658 INFO [train.py:1198] (2/4) Epoch 22, batch 750, loss[loss=0.2056, ctc_loss=0.1353, cr_loss=0.3512, over 17104.00 frames. ], tot_loss[loss=0.2113, ctc_loss=0.1402, cr_loss=0.3558, over 3275126.20 frames. ], batch size: 43, lr: 5.49e-03, grad_scale: 8.0 2024-09-24 01:00:47,334 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=385308.0, ans=0.125 2024-09-24 01:00:48,784 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=385308.0, ans=0.0 2024-09-24 01:00:55,245 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=385354.6666666667, ans=0.125 2024-09-24 01:01:01,651 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=385354.6666666667, ans=0.0 2024-09-24 01:01:31,779 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=385448.0, ans=0.125 2024-09-24 01:01:33,050 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.232e+02 1.335e+02 1.516e+02 2.427e+02, threshold=2.671e+02, percent-clipped=0.0 2024-09-24 01:02:00,313 INFO [train.py:1198] (2/4) Epoch 22, batch 800, loss[loss=0.1912, ctc_loss=0.1273, cr_loss=0.3193, over 16850.00 frames. ], tot_loss[loss=0.2094, ctc_loss=0.1388, cr_loss=0.3534, over 3303026.94 frames. ], batch size: 58, lr: 5.49e-03, grad_scale: 16.0 2024-09-24 01:02:15,013 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=385588.0, ans=0.2 2024-09-24 01:02:34,216 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=385634.6666666667, ans=0.1 2024-09-24 01:02:39,974 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=385634.6666666667, ans=0.125 2024-09-24 01:02:50,357 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=4.85 vs. limit=15.0 2024-09-24 01:02:54,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=385681.3333333333, ans=0.125 2024-09-24 01:02:57,745 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=385681.3333333333, ans=0.125 2024-09-24 01:03:01,489 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.50 vs. limit=15.0 2024-09-24 01:03:23,165 INFO [train.py:1198] (2/4) Epoch 22, batch 850, loss[loss=0.1946, ctc_loss=0.129, cr_loss=0.3276, over 17316.00 frames. ], tot_loss[loss=0.2112, ctc_loss=0.1401, cr_loss=0.3555, over 3306607.04 frames. ], batch size: 46, lr: 5.48e-03, grad_scale: 16.0 2024-09-24 01:03:37,920 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=385821.3333333333, ans=0.125 2024-09-24 01:03:42,750 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=385821.3333333333, ans=0.125 2024-09-24 01:03:49,221 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=385821.3333333333, ans=0.1 2024-09-24 01:03:55,731 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=385868.0, ans=0.125 2024-09-24 01:03:58,902 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=385868.0, ans=0.125 2024-09-24 01:04:18,868 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.098e+02 1.270e+02 1.386e+02 1.514e+02 2.172e+02, threshold=2.772e+02, percent-clipped=0.0 2024-09-24 01:04:34,731 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=385961.3333333333, ans=0.025 2024-09-24 01:04:38,450 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.57 vs. limit=15.0 2024-09-24 01:04:48,968 INFO [train.py:1198] (2/4) Epoch 22, batch 900, loss[loss=0.2165, ctc_loss=0.1446, cr_loss=0.3598, over 15863.00 frames. ], tot_loss[loss=0.211, ctc_loss=0.1399, cr_loss=0.3555, over 3320457.86 frames. ], batch size: 74, lr: 5.48e-03, grad_scale: 16.0 2024-09-24 01:04:57,198 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=386008.0, ans=0.125 2024-09-24 01:05:06,836 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=386054.6666666667, ans=0.125 2024-09-24 01:05:11,595 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=386054.6666666667, ans=0.5 2024-09-24 01:05:20,980 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.46 vs. limit=6.0 2024-09-24 01:05:23,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=386101.3333333333, ans=0.0 2024-09-24 01:05:23,989 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=386101.3333333333, ans=0.0 2024-09-24 01:05:53,163 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=386148.0, ans=0.125 2024-09-24 01:06:11,837 INFO [train.py:1198] (2/4) Epoch 22, batch 950, loss[loss=0.2097, ctc_loss=0.1393, cr_loss=0.3516, over 15997.00 frames. ], tot_loss[loss=0.21, ctc_loss=0.1391, cr_loss=0.3544, over 3332048.66 frames. ], batch size: 74, lr: 5.48e-03, grad_scale: 16.0 2024-09-24 01:06:18,456 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=386241.3333333333, ans=0.125 2024-09-24 01:06:38,271 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=7.91 vs. limit=15.0 2024-09-24 01:06:48,180 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=386334.6666666667, ans=0.125 2024-09-24 01:06:58,071 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=6.92 vs. limit=12.0 2024-09-24 01:07:03,756 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.224e+02 1.313e+02 1.395e+02 1.890e+02, threshold=2.625e+02, percent-clipped=0.0 2024-09-24 01:07:19,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten.whitening_limit, batch_count=386428.0, ans=22.5 2024-09-24 01:07:21,649 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 01:07:31,055 INFO [train.py:1198] (2/4) Epoch 22, batch 1000, loss[loss=0.2074, ctc_loss=0.1357, cr_loss=0.3585, over 17304.00 frames. ], tot_loss[loss=0.2113, ctc_loss=0.1401, cr_loss=0.3559, over 3334446.56 frames. ], batch size: 49, lr: 5.48e-03, grad_scale: 16.0 2024-09-24 01:07:48,332 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=386521.3333333333, ans=0.5 2024-09-24 01:08:21,579 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=386614.6666666667, ans=0.125 2024-09-24 01:08:24,723 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=386614.6666666667, ans=0.0 2024-09-24 01:08:36,133 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=386661.3333333333, ans=0.125 2024-09-24 01:08:39,126 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=386661.3333333333, ans=10.0 2024-09-24 01:08:45,593 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.max_positive, batch_count=386661.3333333333, ans=0.95 2024-09-24 01:08:45,687 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=386661.3333333333, ans=0.125 2024-09-24 01:08:53,198 INFO [train.py:1198] (2/4) Epoch 22, batch 1050, loss[loss=0.2099, ctc_loss=0.1396, cr_loss=0.3519, over 17286.00 frames. ], tot_loss[loss=0.2116, ctc_loss=0.1404, cr_loss=0.3562, over 3337742.78 frames. ], batch size: 46, lr: 5.48e-03, grad_scale: 16.0 2024-09-24 01:09:07,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=386754.6666666667, ans=0.0 2024-09-24 01:09:35,356 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=386801.3333333333, ans=0.025 2024-09-24 01:09:45,611 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=15.35 vs. limit=22.5 2024-09-24 01:09:46,468 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=386848.0, ans=0.1 2024-09-24 01:09:50,909 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.102e+02 1.253e+02 1.357e+02 1.506e+02 3.378e+02, threshold=2.715e+02, percent-clipped=1.0 2024-09-24 01:10:04,046 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=386894.6666666667, ans=0.025 2024-09-24 01:10:08,951 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=386894.6666666667, ans=0.2 2024-09-24 01:10:20,568 INFO [train.py:1198] (2/4) Epoch 22, batch 1100, loss[loss=0.1894, ctc_loss=0.1259, cr_loss=0.3176, over 17297.00 frames. ], tot_loss[loss=0.2136, ctc_loss=0.1419, cr_loss=0.3583, over 3326290.48 frames. ], batch size: 42, lr: 5.48e-03, grad_scale: 16.0 2024-09-24 01:10:28,920 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=386941.3333333333, ans=0.0 2024-09-24 01:10:49,694 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=386988.0, ans=0.125 2024-09-24 01:10:52,235 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.95 vs. limit=15.0 2024-09-24 01:11:02,611 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=387034.6666666667, ans=0.0 2024-09-24 01:11:40,357 INFO [train.py:1198] (2/4) Epoch 22, batch 1150, loss[loss=0.1862, ctc_loss=0.1222, cr_loss=0.3198, over 16958.00 frames. ], tot_loss[loss=0.2136, ctc_loss=0.1419, cr_loss=0.3583, over 3335106.04 frames. ], batch size: 42, lr: 5.47e-03, grad_scale: 16.0 2024-09-24 01:11:50,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=387174.6666666667, ans=0.125 2024-09-24 01:12:27,545 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=387314.6666666667, ans=0.025 2024-09-24 01:12:36,141 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.050e+02 1.223e+02 1.330e+02 1.443e+02 2.592e+02, threshold=2.661e+02, percent-clipped=0.0 2024-09-24 01:12:47,985 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=387361.3333333333, ans=0.125 2024-09-24 01:12:54,529 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 01:13:03,659 INFO [train.py:1198] (2/4) Epoch 22, batch 1200, loss[loss=0.2096, ctc_loss=0.1389, cr_loss=0.3538, over 17010.00 frames. ], tot_loss[loss=0.213, ctc_loss=0.1415, cr_loss=0.3575, over 3334636.39 frames. ], batch size: 51, lr: 5.47e-03, grad_scale: 32.0 2024-09-24 01:13:15,577 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.23 vs. limit=15.0 2024-09-24 01:13:23,026 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=387454.6666666667, ans=0.125 2024-09-24 01:14:19,957 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=387594.6666666667, ans=0.0 2024-09-24 01:14:25,965 INFO [train.py:1198] (2/4) Epoch 22, batch 1250, loss[loss=0.245, ctc_loss=0.1698, cr_loss=0.376, over 17009.00 frames. ], tot_loss[loss=0.2128, ctc_loss=0.1413, cr_loss=0.3574, over 3333112.39 frames. ], batch size: 56, lr: 5.47e-03, grad_scale: 32.0 2024-09-24 01:14:30,456 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=387641.3333333333, ans=0.0 2024-09-24 01:15:12,274 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=11.75 vs. limit=22.5 2024-09-24 01:15:23,793 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.120e+02 1.292e+02 1.416e+02 1.548e+02 3.016e+02, threshold=2.831e+02, percent-clipped=1.0 2024-09-24 01:15:25,647 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=387781.3333333333, ans=0.125 2024-09-24 01:15:50,985 INFO [train.py:1198] (2/4) Epoch 22, batch 1300, loss[loss=0.2126, ctc_loss=0.1403, cr_loss=0.3616, over 17140.00 frames. ], tot_loss[loss=0.2118, ctc_loss=0.1405, cr_loss=0.3563, over 3346458.04 frames. ], batch size: 48, lr: 5.47e-03, grad_scale: 32.0 2024-09-24 01:15:57,883 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=387874.6666666667, ans=0.125 2024-09-24 01:16:01,885 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=3.87 vs. limit=10.0 2024-09-24 01:16:39,245 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 01:16:51,353 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=388014.6666666667, ans=0.95 2024-09-24 01:17:08,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=388108.0, ans=0.2 2024-09-24 01:17:10,181 INFO [train.py:1198] (2/4) Epoch 22, batch 1350, loss[loss=0.2574, ctc_loss=0.1784, cr_loss=0.3949, over 11816.00 frames. ], tot_loss[loss=0.2116, ctc_loss=0.1404, cr_loss=0.3562, over 3352454.26 frames. ], batch size: 123, lr: 5.47e-03, grad_scale: 32.0 2024-09-24 01:17:10,519 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=388108.0, ans=0.0 2024-09-24 01:17:21,661 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=388108.0, ans=0.2 2024-09-24 01:17:35,587 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=388154.6666666667, ans=0.125 2024-09-24 01:18:05,574 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.078e+02 1.287e+02 1.390e+02 1.521e+02 2.749e+02, threshold=2.781e+02, percent-clipped=0.0 2024-09-24 01:18:13,101 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=5.98 vs. limit=12.0 2024-09-24 01:18:32,815 INFO [train.py:1198] (2/4) Epoch 22, batch 1400, loss[loss=0.2162, ctc_loss=0.1452, cr_loss=0.355, over 17263.00 frames. ], tot_loss[loss=0.2101, ctc_loss=0.1392, cr_loss=0.3546, over 3355825.89 frames. ], batch size: 55, lr: 5.47e-03, grad_scale: 32.0 2024-09-24 01:18:33,198 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=388341.3333333333, ans=0.2 2024-09-24 01:18:34,837 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 01:18:49,151 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=388388.0, ans=10.0 2024-09-24 01:18:58,693 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=388388.0, ans=0.1 2024-09-24 01:19:08,985 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=388434.6666666667, ans=0.09899494936611666 2024-09-24 01:19:15,328 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=388434.6666666667, ans=0.0 2024-09-24 01:19:33,173 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=13.13 vs. limit=15.0 2024-09-24 01:19:56,439 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=388574.6666666667, ans=0.0 2024-09-24 01:19:56,547 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=388574.6666666667, ans=0.125 2024-09-24 01:19:57,781 INFO [train.py:1198] (2/4) Epoch 22, batch 1450, loss[loss=0.219, ctc_loss=0.1434, cr_loss=0.3779, over 16879.00 frames. ], tot_loss[loss=0.2101, ctc_loss=0.1391, cr_loss=0.3548, over 3363106.29 frames. ], batch size: 58, lr: 5.46e-03, grad_scale: 32.0 2024-09-24 01:20:35,443 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=388668.0, ans=0.1 2024-09-24 01:20:52,837 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.075e+02 1.236e+02 1.340e+02 1.484e+02 2.143e+02, threshold=2.680e+02, percent-clipped=0.0 2024-09-24 01:21:14,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=388761.3333333333, ans=0.0 2024-09-24 01:21:20,283 INFO [train.py:1198] (2/4) Epoch 22, batch 1500, loss[loss=0.2155, ctc_loss=0.139, cr_loss=0.3822, over 17155.00 frames. ], tot_loss[loss=0.2095, ctc_loss=0.1387, cr_loss=0.354, over 3355716.77 frames. ], batch size: 45, lr: 5.46e-03, grad_scale: 32.0 2024-09-24 01:22:14,668 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=388948.0, ans=0.0 2024-09-24 01:22:42,363 INFO [train.py:1198] (2/4) Epoch 22, batch 1550, loss[loss=0.2386, ctc_loss=0.1637, cr_loss=0.3744, over 15008.00 frames. ], tot_loss[loss=0.2097, ctc_loss=0.1388, cr_loss=0.3547, over 3356864.95 frames. ], batch size: 89, lr: 5.46e-03, grad_scale: 16.0 2024-09-24 01:22:49,090 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=389041.3333333333, ans=0.125 2024-09-24 01:23:09,835 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=389088.0, ans=0.025 2024-09-24 01:23:20,996 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=389134.6666666667, ans=0.0 2024-09-24 01:23:25,927 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=389134.6666666667, ans=0.0 2024-09-24 01:23:29,827 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=6.18 vs. limit=15.0 2024-09-24 01:23:37,028 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.256e+02 1.352e+02 1.468e+02 3.649e+02, threshold=2.704e+02, percent-clipped=1.0 2024-09-24 01:23:45,507 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=389228.0, ans=0.0 2024-09-24 01:23:50,976 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.48 vs. limit=22.5 2024-09-24 01:24:02,774 INFO [train.py:1198] (2/4) Epoch 22, batch 1600, loss[loss=0.2486, ctc_loss=0.1699, cr_loss=0.3935, over 15100.00 frames. ], tot_loss[loss=0.2112, ctc_loss=0.1399, cr_loss=0.3563, over 3353204.00 frames. ], batch size: 89, lr: 5.46e-03, grad_scale: 32.0 2024-09-24 01:24:21,514 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=389321.3333333333, ans=0.1 2024-09-24 01:24:26,489 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=389321.3333333333, ans=0.0 2024-09-24 01:24:53,440 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.88 vs. limit=15.0 2024-09-24 01:24:57,578 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=389414.6666666667, ans=0.0 2024-09-24 01:25:17,339 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=389461.3333333333, ans=0.1 2024-09-24 01:25:29,442 INFO [train.py:1198] (2/4) Epoch 22, batch 1650, loss[loss=0.1972, ctc_loss=0.1293, cr_loss=0.3397, over 17099.00 frames. ], tot_loss[loss=0.2105, ctc_loss=0.1394, cr_loss=0.3554, over 3358313.40 frames. ], batch size: 49, lr: 5.46e-03, grad_scale: 32.0 2024-09-24 01:25:36,254 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=389508.0, ans=0.125 2024-09-24 01:25:39,590 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.06 vs. limit=15.0 2024-09-24 01:25:41,001 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=389508.0, ans=0.125 2024-09-24 01:25:44,866 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.37 vs. limit=10.0 2024-09-24 01:25:59,381 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=5.05 vs. limit=15.0 2024-09-24 01:26:18,940 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=389648.0, ans=0.0 2024-09-24 01:26:23,522 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.218e+02 1.295e+02 1.408e+02 1.987e+02, threshold=2.589e+02, percent-clipped=0.0 2024-09-24 01:26:27,061 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=389648.0, ans=0.2 2024-09-24 01:26:28,732 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=389648.0, ans=0.0 2024-09-24 01:26:49,450 INFO [train.py:1198] (2/4) Epoch 22, batch 1700, loss[loss=0.1822, ctc_loss=0.1174, cr_loss=0.3239, over 17278.00 frames. ], tot_loss[loss=0.2117, ctc_loss=0.1403, cr_loss=0.3567, over 3357995.59 frames. ], batch size: 42, lr: 5.46e-03, grad_scale: 32.0 2024-09-24 01:27:00,254 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.97 vs. limit=6.0 2024-09-24 01:27:02,534 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=389741.3333333333, ans=0.0 2024-09-24 01:27:21,563 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=389788.0, ans=0.125 2024-09-24 01:27:53,241 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=389881.3333333333, ans=0.0 2024-09-24 01:28:11,028 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=389974.6666666667, ans=0.125 2024-09-24 01:28:12,341 INFO [train.py:1198] (2/4) Epoch 22, batch 1750, loss[loss=0.2196, ctc_loss=0.1484, cr_loss=0.3557, over 16697.00 frames. ], tot_loss[loss=0.2125, ctc_loss=0.141, cr_loss=0.3576, over 3348011.86 frames. ], batch size: 61, lr: 5.46e-03, grad_scale: 32.0 2024-09-24 01:28:18,888 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=389974.6666666667, ans=0.2 2024-09-24 01:28:23,928 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=389974.6666666667, ans=0.0 2024-09-24 01:28:36,695 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=390021.3333333333, ans=0.125 2024-09-24 01:29:09,553 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.027e+02 1.289e+02 1.384e+02 1.529e+02 2.458e+02, threshold=2.768e+02, percent-clipped=0.0 2024-09-24 01:29:37,592 INFO [train.py:1198] (2/4) Epoch 22, batch 1800, loss[loss=0.1916, ctc_loss=0.1271, cr_loss=0.3221, over 17065.00 frames. ], tot_loss[loss=0.2118, ctc_loss=0.1406, cr_loss=0.356, over 3346836.16 frames. ], batch size: 46, lr: 5.45e-03, grad_scale: 32.0 2024-09-24 01:29:53,754 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=390254.6666666667, ans=0.125 2024-09-24 01:29:58,637 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=390254.6666666667, ans=0.125 2024-09-24 01:30:20,371 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=390301.3333333333, ans=0.125 2024-09-24 01:30:34,766 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=390348.0, ans=0.2 2024-09-24 01:31:00,215 INFO [train.py:1198] (2/4) Epoch 22, batch 1850, loss[loss=0.2015, ctc_loss=0.1322, cr_loss=0.3464, over 17257.00 frames. ], tot_loss[loss=0.2131, ctc_loss=0.1414, cr_loss=0.3582, over 3348957.00 frames. ], batch size: 44, lr: 5.45e-03, grad_scale: 32.0 2024-09-24 01:31:22,755 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=390488.0, ans=0.125 2024-09-24 01:31:46,285 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=390581.3333333333, ans=0.1 2024-09-24 01:31:47,750 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=390581.3333333333, ans=0.0 2024-09-24 01:31:54,069 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.256e+02 1.338e+02 1.460e+02 2.391e+02, threshold=2.675e+02, percent-clipped=0.0 2024-09-24 01:31:54,346 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=390581.3333333333, ans=0.0 2024-09-24 01:32:21,841 INFO [train.py:1198] (2/4) Epoch 22, batch 1900, loss[loss=0.2198, ctc_loss=0.1488, cr_loss=0.3549, over 15978.00 frames. ], tot_loss[loss=0.2126, ctc_loss=0.141, cr_loss=0.3579, over 3350270.42 frames. ], batch size: 74, lr: 5.45e-03, grad_scale: 32.0 2024-09-24 01:32:52,777 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=390768.0, ans=0.0 2024-09-24 01:33:03,956 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=390768.0, ans=0.0 2024-09-24 01:33:29,306 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=390861.3333333333, ans=0.125 2024-09-24 01:33:29,350 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=390861.3333333333, ans=0.09899494936611666 2024-09-24 01:33:37,177 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=390861.3333333333, ans=0.0 2024-09-24 01:33:41,708 INFO [train.py:1198] (2/4) Epoch 22, batch 1950, loss[loss=0.223, ctc_loss=0.1493, cr_loss=0.3683, over 17201.00 frames. ], tot_loss[loss=0.2116, ctc_loss=0.1403, cr_loss=0.3565, over 3352876.67 frames. ], batch size: 47, lr: 5.45e-03, grad_scale: 32.0 2024-09-24 01:33:47,168 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=13.49 vs. limit=22.5 2024-09-24 01:33:50,526 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.30 vs. limit=15.0 2024-09-24 01:34:00,450 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=390954.6666666667, ans=0.1 2024-09-24 01:34:41,242 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.293e+02 1.356e+02 1.523e+02 3.316e+02, threshold=2.712e+02, percent-clipped=1.0 2024-09-24 01:35:02,883 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=391094.6666666667, ans=0.5 2024-09-24 01:35:09,056 INFO [train.py:1198] (2/4) Epoch 22, batch 2000, loss[loss=0.18, ctc_loss=0.1121, cr_loss=0.3394, over 17072.00 frames. ], tot_loss[loss=0.2118, ctc_loss=0.1404, cr_loss=0.357, over 3350250.42 frames. ], batch size: 40, lr: 5.45e-03, grad_scale: 32.0 2024-09-24 01:35:09,464 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=391141.3333333333, ans=0.125 2024-09-24 01:35:15,626 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=391141.3333333333, ans=0.04949747468305833 2024-09-24 01:35:20,514 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=391141.3333333333, ans=0.0 2024-09-24 01:35:45,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=391234.6666666667, ans=0.1 2024-09-24 01:35:47,330 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 01:35:50,615 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=391234.6666666667, ans=0.1 2024-09-24 01:35:55,611 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=391281.3333333333, ans=0.125 2024-09-24 01:36:08,930 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=16.48 vs. limit=22.5 2024-09-24 01:36:13,022 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=391328.0, ans=0.1 2024-09-24 01:36:17,715 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=391328.0, ans=0.1 2024-09-24 01:36:28,618 INFO [train.py:1198] (2/4) Epoch 22, batch 2050, loss[loss=0.2281, ctc_loss=0.1529, cr_loss=0.3757, over 17317.00 frames. ], tot_loss[loss=0.2114, ctc_loss=0.14, cr_loss=0.3566, over 3352036.97 frames. ], batch size: 51, lr: 5.45e-03, grad_scale: 16.0 2024-09-24 01:36:43,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=391421.3333333333, ans=0.1 2024-09-24 01:36:46,616 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=391421.3333333333, ans=0.0 2024-09-24 01:37:03,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=391468.0, ans=0.1 2024-09-24 01:37:07,620 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=6.37 vs. limit=15.0 2024-09-24 01:37:14,294 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=7.05 vs. limit=8.0 2024-09-24 01:37:19,608 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=391514.6666666667, ans=0.1 2024-09-24 01:37:24,198 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=391514.6666666667, ans=0.0 2024-09-24 01:37:27,042 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.055e+02 1.255e+02 1.364e+02 1.443e+02 3.272e+02, threshold=2.728e+02, percent-clipped=1.0 2024-09-24 01:37:44,977 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=391561.3333333333, ans=0.125 2024-09-24 01:37:51,061 INFO [train.py:1198] (2/4) Epoch 22, batch 2100, loss[loss=0.2294, ctc_loss=0.1559, cr_loss=0.3674, over 16995.00 frames. ], tot_loss[loss=0.2104, ctc_loss=0.1392, cr_loss=0.3558, over 3365572.61 frames. ], batch size: 53, lr: 5.44e-03, grad_scale: 16.0 2024-09-24 01:37:54,625 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=391608.0, ans=0.0 2024-09-24 01:38:07,113 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=391654.6666666667, ans=0.125 2024-09-24 01:38:25,468 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.95 vs. limit=15.0 2024-09-24 01:39:04,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=391794.6666666667, ans=0.1 2024-09-24 01:39:16,044 INFO [train.py:1198] (2/4) Epoch 22, batch 2150, loss[loss=0.2144, ctc_loss=0.1443, cr_loss=0.3505, over 16613.00 frames. ], tot_loss[loss=0.2109, ctc_loss=0.1396, cr_loss=0.3563, over 3359993.67 frames. ], batch size: 66, lr: 5.44e-03, grad_scale: 16.0 2024-09-24 01:39:21,076 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=391841.3333333333, ans=0.0 2024-09-24 01:39:33,711 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=391888.0, ans=0.2 2024-09-24 01:40:00,430 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=391934.6666666667, ans=0.125 2024-09-24 01:40:16,908 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.098e+02 1.257e+02 1.348e+02 1.506e+02 2.274e+02, threshold=2.696e+02, percent-clipped=0.0 2024-09-24 01:40:39,313 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=392074.6666666667, ans=10.0 2024-09-24 01:40:40,616 INFO [train.py:1198] (2/4) Epoch 22, batch 2200, loss[loss=0.1998, ctc_loss=0.1282, cr_loss=0.3579, over 17150.00 frames. ], tot_loss[loss=0.2106, ctc_loss=0.1393, cr_loss=0.3563, over 3368237.24 frames. ], batch size: 48, lr: 5.44e-03, grad_scale: 16.0 2024-09-24 01:41:01,518 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=392121.3333333333, ans=0.125 2024-09-24 01:41:09,538 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=392121.3333333333, ans=0.125 2024-09-24 01:41:18,197 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=6.36 vs. limit=15.0 2024-09-24 01:41:42,258 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=15.09 vs. limit=15.0 2024-09-24 01:42:03,260 INFO [train.py:1198] (2/4) Epoch 22, batch 2250, loss[loss=0.1774, ctc_loss=0.1162, cr_loss=0.3057, over 17098.00 frames. ], tot_loss[loss=0.2114, ctc_loss=0.1399, cr_loss=0.3575, over 3361551.72 frames. ], batch size: 43, lr: 5.44e-03, grad_scale: 16.0 2024-09-24 01:42:11,503 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=392308.0, ans=0.0 2024-09-24 01:42:59,349 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.083e+02 1.230e+02 1.321e+02 1.423e+02 2.235e+02, threshold=2.642e+02, percent-clipped=0.0 2024-09-24 01:43:12,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=392494.6666666667, ans=0.025 2024-09-24 01:43:23,762 INFO [train.py:1198] (2/4) Epoch 22, batch 2300, loss[loss=0.2354, ctc_loss=0.159, cr_loss=0.3818, over 16443.00 frames. ], tot_loss[loss=0.2113, ctc_loss=0.1399, cr_loss=0.3572, over 3362811.50 frames. ], batch size: 66, lr: 5.44e-03, grad_scale: 16.0 2024-09-24 01:43:32,341 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=392541.3333333333, ans=0.0 2024-09-24 01:43:46,975 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=9.27 vs. limit=10.0 2024-09-24 01:43:57,712 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.78 vs. limit=15.0 2024-09-24 01:44:51,568 INFO [train.py:1198] (2/4) Epoch 22, batch 2350, loss[loss=0.2138, ctc_loss=0.1381, cr_loss=0.3784, over 17157.00 frames. ], tot_loss[loss=0.212, ctc_loss=0.1405, cr_loss=0.3573, over 3364126.40 frames. ], batch size: 45, lr: 5.44e-03, grad_scale: 16.0 2024-09-24 01:44:55,032 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=392774.6666666667, ans=0.0 2024-09-24 01:45:01,379 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=392774.6666666667, ans=0.1 2024-09-24 01:45:11,048 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=392821.3333333333, ans=0.0 2024-09-24 01:45:12,509 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=392821.3333333333, ans=0.0 2024-09-24 01:45:12,678 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=392821.3333333333, ans=0.1 2024-09-24 01:45:14,202 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=392821.3333333333, ans=0.025 2024-09-24 01:45:23,481 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 01:45:28,215 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=392868.0, ans=0.0 2024-09-24 01:45:44,233 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=392914.6666666667, ans=0.1 2024-09-24 01:45:47,292 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.099e+02 1.237e+02 1.319e+02 1.405e+02 2.078e+02, threshold=2.638e+02, percent-clipped=0.0 2024-09-24 01:46:11,757 INFO [train.py:1198] (2/4) Epoch 22, batch 2400, loss[loss=0.2187, ctc_loss=0.1459, cr_loss=0.364, over 17107.00 frames. ], tot_loss[loss=0.2116, ctc_loss=0.1403, cr_loss=0.3568, over 3369214.20 frames. ], batch size: 49, lr: 5.43e-03, grad_scale: 32.0 2024-09-24 01:46:28,573 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=9.03 vs. limit=15.0 2024-09-24 01:47:04,875 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=11.91 vs. limit=15.0 2024-09-24 01:47:22,178 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.min_positive, batch_count=393194.6666666667, ans=0.05 2024-09-24 01:47:34,554 INFO [train.py:1198] (2/4) Epoch 22, batch 2450, loss[loss=0.2312, ctc_loss=0.1534, cr_loss=0.3888, over 16770.00 frames. ], tot_loss[loss=0.2117, ctc_loss=0.1404, cr_loss=0.3566, over 3369996.76 frames. ], batch size: 61, lr: 5.43e-03, grad_scale: 32.0 2024-09-24 01:47:47,048 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=6.71 vs. limit=15.0 2024-09-24 01:47:49,447 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=393288.0, ans=0.0 2024-09-24 01:48:08,514 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=393334.6666666667, ans=0.0 2024-09-24 01:48:30,720 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.273e+02 1.360e+02 1.558e+02 2.301e+02, threshold=2.721e+02, percent-clipped=0.0 2024-09-24 01:48:37,371 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=393428.0, ans=0.125 2024-09-24 01:48:57,400 INFO [train.py:1198] (2/4) Epoch 22, batch 2500, loss[loss=0.1894, ctc_loss=0.1242, cr_loss=0.326, over 17305.00 frames. ], tot_loss[loss=0.2104, ctc_loss=0.1394, cr_loss=0.3551, over 3375545.78 frames. ], batch size: 49, lr: 5.43e-03, grad_scale: 32.0 2024-09-24 01:49:55,094 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=393614.6666666667, ans=0.125 2024-09-24 01:50:22,182 INFO [train.py:1198] (2/4) Epoch 22, batch 2550, loss[loss=0.2173, ctc_loss=0.1443, cr_loss=0.365, over 17306.00 frames. ], tot_loss[loss=0.2117, ctc_loss=0.1405, cr_loss=0.3562, over 3369474.45 frames. ], batch size: 51, lr: 5.43e-03, grad_scale: 32.0 2024-09-24 01:50:49,636 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=393754.6666666667, ans=0.125 2024-09-24 01:51:09,289 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=393848.0, ans=0.5 2024-09-24 01:51:18,854 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.258e+02 1.341e+02 1.475e+02 2.313e+02, threshold=2.682e+02, percent-clipped=0.0 2024-09-24 01:51:36,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=393894.6666666667, ans=0.125 2024-09-24 01:51:43,205 INFO [train.py:1198] (2/4) Epoch 22, batch 2600, loss[loss=0.2091, ctc_loss=0.1374, cr_loss=0.3586, over 17180.00 frames. ], tot_loss[loss=0.2107, ctc_loss=0.1396, cr_loss=0.3556, over 3370004.63 frames. ], batch size: 55, lr: 5.43e-03, grad_scale: 32.0 2024-09-24 01:52:13,835 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn2.whiten.whitening_limit, batch_count=393988.0, ans=22.5 2024-09-24 01:52:18,534 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=394034.6666666667, ans=0.125 2024-09-24 01:52:31,303 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=394034.6666666667, ans=0.0 2024-09-24 01:52:39,292 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=394081.3333333333, ans=0.0 2024-09-24 01:52:48,817 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 01:52:53,597 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=394128.0, ans=0.2 2024-09-24 01:53:06,118 INFO [train.py:1198] (2/4) Epoch 22, batch 2650, loss[loss=0.2146, ctc_loss=0.1403, cr_loss=0.3716, over 17282.00 frames. ], tot_loss[loss=0.2097, ctc_loss=0.1388, cr_loss=0.3543, over 3369588.20 frames. ], batch size: 46, lr: 5.43e-03, grad_scale: 32.0 2024-09-24 01:53:14,797 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=5.59 vs. limit=12.0 2024-09-24 01:53:46,443 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.38 vs. limit=10.0 2024-09-24 01:54:01,962 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=394314.6666666667, ans=0.125 2024-09-24 01:54:07,322 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.255e+02 1.320e+02 1.434e+02 1.908e+02, threshold=2.640e+02, percent-clipped=0.0 2024-09-24 01:54:31,190 INFO [train.py:1198] (2/4) Epoch 22, batch 2700, loss[loss=0.1749, ctc_loss=0.1141, cr_loss=0.3038, over 17117.00 frames. ], tot_loss[loss=0.2107, ctc_loss=0.1395, cr_loss=0.3558, over 3368138.80 frames. ], batch size: 40, lr: 5.42e-03, grad_scale: 32.0 2024-09-24 01:54:57,899 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=394454.6666666667, ans=0.1 2024-09-24 01:55:18,656 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=394501.3333333333, ans=0.125 2024-09-24 01:55:21,804 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=394548.0, ans=0.125 2024-09-24 01:55:23,331 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=394548.0, ans=0.0 2024-09-24 01:55:42,568 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=394594.6666666667, ans=0.125 2024-09-24 01:55:47,046 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=394594.6666666667, ans=0.125 2024-09-24 01:55:53,379 INFO [train.py:1198] (2/4) Epoch 22, batch 2750, loss[loss=0.2161, ctc_loss=0.1435, cr_loss=0.3631, over 17305.00 frames. ], tot_loss[loss=0.2112, ctc_loss=0.14, cr_loss=0.3559, over 3367087.92 frames. ], batch size: 51, lr: 5.42e-03, grad_scale: 32.0 2024-09-24 01:56:00,088 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.min_positive, batch_count=394641.3333333333, ans=0.05 2024-09-24 01:56:11,214 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=394688.0, ans=0.125 2024-09-24 01:56:17,683 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=394688.0, ans=0.0 2024-09-24 01:56:41,640 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=394781.3333333333, ans=0.0 2024-09-24 01:56:50,882 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=394781.3333333333, ans=0.125 2024-09-24 01:56:52,096 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.037e+02 1.297e+02 1.428e+02 1.577e+02 2.482e+02, threshold=2.855e+02, percent-clipped=0.0 2024-09-24 01:57:06,111 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.81 vs. limit=6.0 2024-09-24 01:57:07,247 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=394828.0, ans=0.2 2024-09-24 01:57:08,939 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=394828.0, ans=0.2 2024-09-24 01:57:16,458 INFO [train.py:1198] (2/4) Epoch 22, batch 2800, loss[loss=0.2719, ctc_loss=0.1836, cr_loss=0.4413, over 17058.00 frames. ], tot_loss[loss=0.2109, ctc_loss=0.1397, cr_loss=0.3559, over 3370447.65 frames. ], batch size: 56, lr: 5.42e-03, grad_scale: 32.0 2024-09-24 01:58:01,443 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.89 vs. limit=22.5 2024-09-24 01:58:08,806 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=395014.6666666667, ans=0.1 2024-09-24 01:58:13,536 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=395014.6666666667, ans=0.125 2024-09-24 01:58:18,336 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=395061.3333333333, ans=0.025 2024-09-24 01:58:21,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=395061.3333333333, ans=0.125 2024-09-24 01:58:38,324 INFO [train.py:1198] (2/4) Epoch 22, batch 2850, loss[loss=0.2085, ctc_loss=0.1379, cr_loss=0.3529, over 17009.00 frames. ], tot_loss[loss=0.2113, ctc_loss=0.1401, cr_loss=0.3562, over 3361692.40 frames. ], batch size: 44, lr: 5.42e-03, grad_scale: 32.0 2024-09-24 01:58:46,355 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=395108.0, ans=0.09899494936611666 2024-09-24 01:59:01,228 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=13.23 vs. limit=15.0 2024-09-24 01:59:36,607 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.264e+02 1.351e+02 1.437e+02 2.135e+02, threshold=2.702e+02, percent-clipped=0.0 2024-09-24 02:00:03,215 INFO [train.py:1198] (2/4) Epoch 22, batch 2900, loss[loss=0.207, ctc_loss=0.1387, cr_loss=0.3417, over 17160.00 frames. ], tot_loss[loss=0.2117, ctc_loss=0.1405, cr_loss=0.3564, over 3358267.73 frames. ], batch size: 45, lr: 5.42e-03, grad_scale: 32.0 2024-09-24 02:00:17,636 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=395388.0, ans=0.2 2024-09-24 02:00:52,820 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=395481.3333333333, ans=0.125 2024-09-24 02:01:23,078 INFO [train.py:1198] (2/4) Epoch 22, batch 2950, loss[loss=0.2163, ctc_loss=0.1445, cr_loss=0.359, over 17254.00 frames. ], tot_loss[loss=0.2121, ctc_loss=0.1407, cr_loss=0.3569, over 3354843.99 frames. ], batch size: 44, lr: 5.42e-03, grad_scale: 32.0 2024-09-24 02:01:36,192 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=395574.6666666667, ans=0.1 2024-09-24 02:01:45,378 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=395621.3333333333, ans=0.1 2024-09-24 02:02:21,884 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.109e+02 1.266e+02 1.346e+02 1.444e+02 1.756e+02, threshold=2.692e+02, percent-clipped=0.0 2024-09-24 02:02:22,117 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=395714.6666666667, ans=0.0 2024-09-24 02:02:44,717 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=4.92 vs. limit=15.0 2024-09-24 02:02:45,309 INFO [train.py:1198] (2/4) Epoch 22, batch 3000, loss[loss=0.212, ctc_loss=0.1422, cr_loss=0.349, over 16873.00 frames. ], tot_loss[loss=0.2101, ctc_loss=0.1392, cr_loss=0.3548, over 3365560.06 frames. ], batch size: 58, lr: 5.42e-03, grad_scale: 32.0 2024-09-24 02:02:45,309 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 02:02:52,300 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.2.encoder.layers.2.self_attn_weights, attn_weights_entropy = tensor([3.3708, 2.3596, 3.0891, 3.0985], device='cuda:2') 2024-09-24 02:03:00,736 INFO [train.py:1230] (2/4) Epoch 22, validation: loss=0.03869, ctc_loss=0.03869, cr_loss=8.188e-15, over 944034.00 frames. 2024-09-24 02:03:00,736 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 02:03:06,097 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.61 vs. limit=6.0 2024-09-24 02:03:27,077 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=395854.6666666667, ans=0.1 2024-09-24 02:03:28,797 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=395854.6666666667, ans=0.1 2024-09-24 02:04:02,008 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.01 vs. limit=15.0 2024-09-24 02:04:20,325 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=4.35 vs. limit=12.0 2024-09-24 02:04:21,130 INFO [train.py:1198] (2/4) Epoch 22, batch 3050, loss[loss=0.26, ctc_loss=0.1772, cr_loss=0.4136, over 17006.00 frames. ], tot_loss[loss=0.2118, ctc_loss=0.1404, cr_loss=0.357, over 3349733.61 frames. ], batch size: 53, lr: 5.41e-03, grad_scale: 32.0 2024-09-24 02:04:26,185 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=396041.3333333333, ans=0.0 2024-09-24 02:04:27,623 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=396041.3333333333, ans=0.1 2024-09-24 02:04:33,977 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=396041.3333333333, ans=0.1 2024-09-24 02:04:34,150 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=11.65 vs. limit=12.0 2024-09-24 02:04:43,303 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=396088.0, ans=0.125 2024-09-24 02:04:43,632 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.95 vs. limit=6.0 2024-09-24 02:04:44,742 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=396088.0, ans=0.0 2024-09-24 02:04:49,411 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=396088.0, ans=0.2 2024-09-24 02:04:54,166 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=396134.6666666667, ans=0.125 2024-09-24 02:05:15,411 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.106e+02 1.242e+02 1.330e+02 1.474e+02 2.506e+02, threshold=2.661e+02, percent-clipped=0.0 2024-09-24 02:05:19,501 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.20 vs. limit=15.0 2024-09-24 02:05:41,006 INFO [train.py:1198] (2/4) Epoch 22, batch 3100, loss[loss=0.2141, ctc_loss=0.1393, cr_loss=0.3741, over 17068.00 frames. ], tot_loss[loss=0.2119, ctc_loss=0.1405, cr_loss=0.3569, over 3353929.08 frames. ], batch size: 46, lr: 5.41e-03, grad_scale: 32.0 2024-09-24 02:05:50,678 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=396274.6666666667, ans=0.1 2024-09-24 02:06:00,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=396321.3333333333, ans=0.125 2024-09-24 02:06:09,424 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=396321.3333333333, ans=0.1 2024-09-24 02:06:12,495 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=396368.0, ans=0.125 2024-09-24 02:06:23,336 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=396368.0, ans=0.125 2024-09-24 02:06:26,452 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=396414.6666666667, ans=0.025 2024-09-24 02:06:35,733 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=396414.6666666667, ans=0.025 2024-09-24 02:07:01,221 INFO [train.py:1198] (2/4) Epoch 22, batch 3150, loss[loss=0.2233, ctc_loss=0.1479, cr_loss=0.3768, over 17035.00 frames. ], tot_loss[loss=0.2112, ctc_loss=0.14, cr_loss=0.3558, over 3343573.44 frames. ], batch size: 52, lr: 5.41e-03, grad_scale: 16.0 2024-09-24 02:07:26,327 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=396554.6666666667, ans=0.125 2024-09-24 02:07:44,153 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.57 vs. limit=10.0 2024-09-24 02:07:57,788 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.290e+02 1.399e+02 1.555e+02 2.844e+02, threshold=2.797e+02, percent-clipped=1.0 2024-09-24 02:08:08,837 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=396694.6666666667, ans=0.1 2024-09-24 02:08:19,605 INFO [train.py:1198] (2/4) Epoch 22, batch 3200, loss[loss=0.1999, ctc_loss=0.1318, cr_loss=0.3408, over 17169.00 frames. ], tot_loss[loss=0.2105, ctc_loss=0.1395, cr_loss=0.3549, over 3347965.67 frames. ], batch size: 41, lr: 5.41e-03, grad_scale: 32.0 2024-09-24 02:08:24,645 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=396741.3333333333, ans=0.0 2024-09-24 02:08:35,457 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=396788.0, ans=0.0 2024-09-24 02:08:54,884 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.30 vs. limit=15.0 2024-09-24 02:09:19,516 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=396881.3333333333, ans=0.125 2024-09-24 02:09:20,093 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.31 vs. limit=15.0 2024-09-24 02:09:37,797 INFO [train.py:1198] (2/4) Epoch 22, batch 3250, loss[loss=0.1766, ctc_loss=0.1123, cr_loss=0.3216, over 17078.00 frames. ], tot_loss[loss=0.2105, ctc_loss=0.1395, cr_loss=0.3552, over 3348155.90 frames. ], batch size: 43, lr: 5.41e-03, grad_scale: 32.0 2024-09-24 02:09:56,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=397021.3333333333, ans=0.1 2024-09-24 02:10:24,704 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=397114.6666666667, ans=0.125 2024-09-24 02:10:29,401 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=397114.6666666667, ans=0.0 2024-09-24 02:10:33,680 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.276e+02 1.353e+02 1.572e+02 3.957e+02, threshold=2.706e+02, percent-clipped=1.0 2024-09-24 02:10:42,924 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=397161.3333333333, ans=0.0 2024-09-24 02:10:50,925 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=397161.3333333333, ans=0.125 2024-09-24 02:10:55,243 INFO [train.py:1198] (2/4) Epoch 22, batch 3300, loss[loss=0.2452, ctc_loss=0.1654, cr_loss=0.3988, over 17035.00 frames. ], tot_loss[loss=0.2109, ctc_loss=0.1397, cr_loss=0.3563, over 3356927.89 frames. ], batch size: 52, lr: 5.41e-03, grad_scale: 32.0 2024-09-24 02:10:55,574 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 02:11:20,733 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer_na.min_abs, batch_count=397254.6666666667, ans=0.02 2024-09-24 02:11:39,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=397301.3333333333, ans=0.2 2024-09-24 02:11:50,366 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=397348.0, ans=0.025 2024-09-24 02:12:12,466 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=397394.6666666667, ans=0.1 2024-09-24 02:12:15,300 INFO [train.py:1198] (2/4) Epoch 22, batch 3350, loss[loss=0.2063, ctc_loss=0.1373, cr_loss=0.345, over 17212.00 frames. ], tot_loss[loss=0.2118, ctc_loss=0.1403, cr_loss=0.3575, over 3359392.67 frames. ], batch size: 55, lr: 5.40e-03, grad_scale: 32.0 2024-09-24 02:12:50,030 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=397534.6666666667, ans=0.1 2024-09-24 02:13:11,463 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.283e+02 1.436e+02 1.658e+02 2.229e+02, threshold=2.872e+02, percent-clipped=0.0 2024-09-24 02:13:18,132 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=397628.0, ans=0.0 2024-09-24 02:13:18,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=397628.0, ans=0.125 2024-09-24 02:13:33,127 INFO [train.py:1198] (2/4) Epoch 22, batch 3400, loss[loss=0.2033, ctc_loss=0.1327, cr_loss=0.3531, over 17091.00 frames. ], tot_loss[loss=0.2123, ctc_loss=0.1405, cr_loss=0.3588, over 3369382.23 frames. ], batch size: 43, lr: 5.40e-03, grad_scale: 32.0 2024-09-24 02:13:38,030 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=397674.6666666667, ans=0.125 2024-09-24 02:13:39,749 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=397674.6666666667, ans=0.0 2024-09-24 02:14:16,890 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=397768.0, ans=0.0 2024-09-24 02:14:53,442 INFO [train.py:1198] (2/4) Epoch 22, batch 3450, loss[loss=0.1851, ctc_loss=0.1161, cr_loss=0.3447, over 16955.00 frames. ], tot_loss[loss=0.2121, ctc_loss=0.1404, cr_loss=0.3585, over 3367081.08 frames. ], batch size: 42, lr: 5.40e-03, grad_scale: 32.0 2024-09-24 02:15:51,969 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.275e+02 1.378e+02 1.514e+02 2.011e+02, threshold=2.756e+02, percent-clipped=0.0 2024-09-24 02:16:01,756 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=398094.6666666667, ans=0.125 2024-09-24 02:16:13,997 INFO [train.py:1198] (2/4) Epoch 22, batch 3500, loss[loss=0.1696, ctc_loss=0.1085, cr_loss=0.3058, over 16945.00 frames. ], tot_loss[loss=0.2102, ctc_loss=0.139, cr_loss=0.3561, over 3371910.65 frames. ], batch size: 42, lr: 5.40e-03, grad_scale: 32.0 2024-09-24 02:16:15,900 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=398141.3333333333, ans=0.125 2024-09-24 02:16:31,496 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=398188.0, ans=0.125 2024-09-24 02:16:42,497 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=398188.0, ans=0.125 2024-09-24 02:16:54,836 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=398234.6666666667, ans=0.1 2024-09-24 02:17:04,607 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=398281.3333333333, ans=0.04949747468305833 2024-09-24 02:17:32,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=398374.6666666667, ans=0.0 2024-09-24 02:17:34,215 INFO [train.py:1198] (2/4) Epoch 22, batch 3550, loss[loss=0.2355, ctc_loss=0.1607, cr_loss=0.3737, over 15980.00 frames. ], tot_loss[loss=0.2105, ctc_loss=0.1393, cr_loss=0.3559, over 3358982.03 frames. ], batch size: 74, lr: 5.40e-03, grad_scale: 32.0 2024-09-24 02:17:57,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=398421.3333333333, ans=0.125 2024-09-24 02:18:01,233 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.51 vs. limit=15.0 2024-09-24 02:18:11,967 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=398468.0, ans=0.0 2024-09-24 02:18:32,111 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.257e+02 1.363e+02 1.494e+02 1.950e+02, threshold=2.726e+02, percent-clipped=0.0 2024-09-24 02:18:52,335 INFO [train.py:1198] (2/4) Epoch 22, batch 3600, loss[loss=0.2329, ctc_loss=0.1584, cr_loss=0.3724, over 16729.00 frames. ], tot_loss[loss=0.2113, ctc_loss=0.1399, cr_loss=0.3569, over 3357663.90 frames. ], batch size: 61, lr: 5.40e-03, grad_scale: 32.0 2024-09-24 02:19:46,928 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=398748.0, ans=0.1 2024-09-24 02:19:54,906 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=398794.6666666667, ans=0.0 2024-09-24 02:20:02,579 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=398794.6666666667, ans=0.125 2024-09-24 02:20:10,105 INFO [train.py:1198] (2/4) Epoch 22, batch 3650, loss[loss=0.2107, ctc_loss=0.139, cr_loss=0.3584, over 17218.00 frames. ], tot_loss[loss=0.2113, ctc_loss=0.14, cr_loss=0.3565, over 3349086.81 frames. ], batch size: 50, lr: 5.39e-03, grad_scale: 32.0 2024-09-24 02:20:15,109 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=398841.3333333333, ans=0.125 2024-09-24 02:20:32,096 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=398888.0, ans=0.125 2024-09-24 02:20:56,059 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=398934.6666666667, ans=0.0 2024-09-24 02:20:57,579 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=398981.3333333333, ans=0.125 2024-09-24 02:21:09,555 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.259e+02 1.359e+02 1.456e+02 2.035e+02, threshold=2.719e+02, percent-clipped=0.0 2024-09-24 02:21:11,345 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=398981.3333333333, ans=0.125 2024-09-24 02:21:30,491 INFO [train.py:1198] (2/4) Epoch 22, batch 3700, loss[loss=0.2444, ctc_loss=0.1627, cr_loss=0.4084, over 16567.00 frames. ], tot_loss[loss=0.2123, ctc_loss=0.1407, cr_loss=0.3581, over 3348600.23 frames. ], batch size: 66, lr: 5.39e-03, grad_scale: 32.0 2024-09-24 02:21:56,553 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=4.64 vs. limit=15.0 2024-09-24 02:22:04,833 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=399168.0, ans=0.0 2024-09-24 02:22:21,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=399214.6666666667, ans=0.0 2024-09-24 02:22:26,608 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=399214.6666666667, ans=0.125 2024-09-24 02:22:27,356 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.01 vs. limit=15.0 2024-09-24 02:22:32,972 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=399261.3333333333, ans=0.0 2024-09-24 02:22:48,075 INFO [train.py:1198] (2/4) Epoch 22, batch 3750, loss[loss=0.2111, ctc_loss=0.1356, cr_loss=0.3776, over 17275.00 frames. ], tot_loss[loss=0.2123, ctc_loss=0.1407, cr_loss=0.3581, over 3344385.30 frames. ], batch size: 46, lr: 5.39e-03, grad_scale: 32.0 2024-09-24 02:22:52,960 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=399308.0, ans=0.125 2024-09-24 02:23:09,322 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.69 vs. limit=15.0 2024-09-24 02:23:46,154 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.055e+02 1.265e+02 1.370e+02 1.488e+02 1.870e+02, threshold=2.739e+02, percent-clipped=0.0 2024-09-24 02:24:07,240 INFO [train.py:1198] (2/4) Epoch 22, batch 3800, loss[loss=0.2477, ctc_loss=0.175, cr_loss=0.3638, over 11525.00 frames. ], tot_loss[loss=0.2133, ctc_loss=0.1415, cr_loss=0.3593, over 3330048.67 frames. ], batch size: 123, lr: 5.39e-03, grad_scale: 32.0 2024-09-24 02:24:09,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=399541.3333333333, ans=0.125 2024-09-24 02:24:35,072 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=15.78 vs. limit=22.5 2024-09-24 02:24:46,713 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=399634.6666666667, ans=0.0 2024-09-24 02:24:48,112 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=399634.6666666667, ans=0.015 2024-09-24 02:24:51,181 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=399681.3333333333, ans=0.2 2024-09-24 02:25:00,400 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=399681.3333333333, ans=0.125 2024-09-24 02:25:23,718 INFO [train.py:1198] (2/4) Epoch 22, batch 3850, loss[loss=0.2471, ctc_loss=0.17, cr_loss=0.3856, over 12256.00 frames. ], tot_loss[loss=0.2155, ctc_loss=0.1434, cr_loss=0.3604, over 3272935.91 frames. ], batch size: 123, lr: 5.39e-03, grad_scale: 16.0 2024-09-24 02:25:23,914 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=399774.6666666667, ans=0.1 2024-09-24 02:25:53,641 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=399868.0, ans=0.125 2024-09-24 02:26:14,985 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=399914.6666666667, ans=0.0 2024-09-24 02:26:22,074 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.135e+02 1.394e+02 1.521e+02 1.653e+02 2.855e+02, threshold=3.042e+02, percent-clipped=1.0 2024-09-24 02:27:26,202 INFO [train.py:1198] (2/4) Epoch 23, batch 0, loss[loss=0.2334, ctc_loss=0.1585, cr_loss=0.3746, over 17215.00 frames. ], tot_loss[loss=0.2334, ctc_loss=0.1585, cr_loss=0.3746, over 17215.00 frames. ], batch size: 55, lr: 5.27e-03, grad_scale: 32.0 2024-09-24 02:27:26,202 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 02:27:40,426 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.3.encoder.layers.2.self_attn_weights, attn_weights_entropy = tensor([2.4207, 2.8935, 2.8161, 3.0662, 2.7055, 2.6207, 3.0988, 3.1955], device='cuda:2') 2024-09-24 02:27:41,785 INFO [train.py:1230] (2/4) Epoch 23, validation: loss=0.03754, ctc_loss=0.03754, cr_loss=8.311e-15, over 944034.00 frames. 2024-09-24 02:27:41,786 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 02:28:06,310 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=400040.6666666667, ans=0.0 2024-09-24 02:28:39,576 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.91 vs. limit=22.5 2024-09-24 02:28:55,301 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=400180.6666666667, ans=0.0 2024-09-24 02:28:55,747 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn2.whiten.whitening_limit, batch_count=400180.6666666667, ans=22.5 2024-09-24 02:29:03,259 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=400180.6666666667, ans=0.125 2024-09-24 02:29:04,930 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=400227.3333333333, ans=0.125 2024-09-24 02:29:06,169 INFO [train.py:1198] (2/4) Epoch 23, batch 50, loss[loss=0.2211, ctc_loss=0.1487, cr_loss=0.3622, over 16999.00 frames. ], tot_loss[loss=0.2138, ctc_loss=0.1418, cr_loss=0.3598, over 757866.71 frames. ], batch size: 56, lr: 5.26e-03, grad_scale: 32.0 2024-09-24 02:29:07,970 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=400227.3333333333, ans=0.1 2024-09-24 02:29:46,447 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=400320.6666666667, ans=0.2 2024-09-24 02:30:00,738 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=400367.3333333333, ans=0.125 2024-09-24 02:30:10,387 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=400414.0, ans=0.125 2024-09-24 02:30:11,720 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.083e+02 1.262e+02 1.335e+02 1.476e+02 2.366e+02, threshold=2.670e+02, percent-clipped=0.0 2024-09-24 02:30:26,016 INFO [train.py:1198] (2/4) Epoch 23, batch 100, loss[loss=0.2182, ctc_loss=0.1463, cr_loss=0.3593, over 17360.00 frames. ], tot_loss[loss=0.2127, ctc_loss=0.141, cr_loss=0.3588, over 1331956.02 frames. ], batch size: 48, lr: 5.26e-03, grad_scale: 32.0 2024-09-24 02:30:58,379 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=400554.0, ans=0.0 2024-09-24 02:31:00,323 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=7.09 vs. limit=15.0 2024-09-24 02:31:17,132 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=400600.6666666667, ans=0.125 2024-09-24 02:31:50,459 INFO [train.py:1198] (2/4) Epoch 23, batch 150, loss[loss=0.1932, ctc_loss=0.1273, cr_loss=0.3296, over 17155.00 frames. ], tot_loss[loss=0.2117, ctc_loss=0.1402, cr_loss=0.3576, over 1789822.15 frames. ], batch size: 45, lr: 5.26e-03, grad_scale: 32.0 2024-09-24 02:31:55,555 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=400694.0, ans=0.125 2024-09-24 02:32:14,719 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=400740.6666666667, ans=0.125 2024-09-24 02:32:55,956 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.034e+02 1.238e+02 1.330e+02 1.442e+02 1.852e+02, threshold=2.660e+02, percent-clipped=0.0 2024-09-24 02:33:00,669 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=400880.6666666667, ans=0.0 2024-09-24 02:33:13,117 INFO [train.py:1198] (2/4) Epoch 23, batch 200, loss[loss=0.2574, ctc_loss=0.1729, cr_loss=0.4229, over 17206.00 frames. ], tot_loss[loss=0.2132, ctc_loss=0.1414, cr_loss=0.359, over 2131871.25 frames. ], batch size: 55, lr: 5.26e-03, grad_scale: 32.0 2024-09-24 02:33:19,594 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=400927.3333333333, ans=0.125 2024-09-24 02:33:22,910 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 02:33:30,832 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=400974.0, ans=0.125 2024-09-24 02:33:34,069 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=400974.0, ans=0.125 2024-09-24 02:33:43,546 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=401020.6666666667, ans=0.0 2024-09-24 02:33:44,387 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=12.09 vs. limit=22.5 2024-09-24 02:33:45,097 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=401020.6666666667, ans=0.125 2024-09-24 02:34:04,060 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 02:34:04,068 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=401067.3333333333, ans=0.125 2024-09-24 02:34:05,585 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=401067.3333333333, ans=0.125 2024-09-24 02:34:35,729 INFO [train.py:1198] (2/4) Epoch 23, batch 250, loss[loss=0.1729, ctc_loss=0.1121, cr_loss=0.3039, over 16785.00 frames. ], tot_loss[loss=0.2124, ctc_loss=0.1407, cr_loss=0.3584, over 2405392.74 frames. ], batch size: 37, lr: 5.26e-03, grad_scale: 32.0 2024-09-24 02:35:12,646 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=401254.0, ans=0.125 2024-09-24 02:35:15,874 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=401254.0, ans=0.125 2024-09-24 02:35:28,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=401300.6666666667, ans=0.1 2024-09-24 02:35:30,190 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=401300.6666666667, ans=0.0 2024-09-24 02:35:40,336 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=6.10 vs. limit=15.0 2024-09-24 02:35:41,288 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.141e+02 1.253e+02 1.336e+02 1.468e+02 3.243e+02, threshold=2.673e+02, percent-clipped=1.0 2024-09-24 02:35:55,852 INFO [train.py:1198] (2/4) Epoch 23, batch 300, loss[loss=0.2548, ctc_loss=0.172, cr_loss=0.4141, over 16981.00 frames. ], tot_loss[loss=0.2109, ctc_loss=0.1397, cr_loss=0.356, over 2610134.28 frames. ], batch size: 53, lr: 5.26e-03, grad_scale: 32.0 2024-09-24 02:36:31,212 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=401487.3333333333, ans=0.05 2024-09-24 02:36:32,803 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.min_positive, batch_count=401487.3333333333, ans=0.025 2024-09-24 02:37:20,386 INFO [train.py:1198] (2/4) Epoch 23, batch 350, loss[loss=0.2148, ctc_loss=0.1427, cr_loss=0.3606, over 16701.00 frames. ], tot_loss[loss=0.2097, ctc_loss=0.1385, cr_loss=0.3555, over 2785413.27 frames. ], batch size: 61, lr: 5.26e-03, grad_scale: 32.0 2024-09-24 02:37:36,539 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-24 02:38:28,351 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.080e+02 1.260e+02 1.355e+02 1.507e+02 3.262e+02, threshold=2.709e+02, percent-clipped=1.0 2024-09-24 02:38:35,728 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=18.81 vs. limit=22.5 2024-09-24 02:38:42,870 INFO [train.py:1198] (2/4) Epoch 23, batch 400, loss[loss=0.2464, ctc_loss=0.1698, cr_loss=0.3834, over 16568.00 frames. ], tot_loss[loss=0.2114, ctc_loss=0.1402, cr_loss=0.3564, over 2887655.77 frames. ], batch size: 66, lr: 5.25e-03, grad_scale: 32.0 2024-09-24 02:38:57,254 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=401860.6666666667, ans=0.125 2024-09-24 02:39:47,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.min_positive, batch_count=402047.3333333333, ans=0.025 2024-09-24 02:39:58,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=402047.3333333333, ans=0.125 2024-09-24 02:40:04,967 INFO [train.py:1198] (2/4) Epoch 23, batch 450, loss[loss=0.2109, ctc_loss=0.1388, cr_loss=0.3606, over 17233.00 frames. ], tot_loss[loss=0.211, ctc_loss=0.1397, cr_loss=0.3565, over 2990709.77 frames. ], batch size: 50, lr: 5.25e-03, grad_scale: 32.0 2024-09-24 02:40:19,405 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=402140.6666666667, ans=0.0 2024-09-24 02:40:26,147 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=14.08 vs. limit=15.0 2024-09-24 02:40:33,826 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=402140.6666666667, ans=0.125 2024-09-24 02:40:41,808 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=402187.3333333333, ans=0.0 2024-09-24 02:40:49,870 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=402187.3333333333, ans=0.125 2024-09-24 02:40:57,557 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=402234.0, ans=0.2 2024-09-24 02:41:13,371 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.257e+02 1.349e+02 1.505e+02 2.195e+02, threshold=2.697e+02, percent-clipped=0.0 2024-09-24 02:41:18,564 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=402280.6666666667, ans=0.07 2024-09-24 02:41:27,647 INFO [train.py:1198] (2/4) Epoch 23, batch 500, loss[loss=0.2276, ctc_loss=0.1534, cr_loss=0.3713, over 17136.00 frames. ], tot_loss[loss=0.2116, ctc_loss=0.1401, cr_loss=0.3574, over 3075503.34 frames. ], batch size: 48, lr: 5.25e-03, grad_scale: 32.0 2024-09-24 02:41:27,846 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=402327.3333333333, ans=0.1 2024-09-24 02:41:32,749 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.max_abs, batch_count=402327.3333333333, ans=10.0 2024-09-24 02:41:45,570 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=402374.0, ans=0.125 2024-09-24 02:42:02,799 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 02:42:50,028 INFO [train.py:1198] (2/4) Epoch 23, batch 550, loss[loss=0.2144, ctc_loss=0.1418, cr_loss=0.3627, over 17344.00 frames. ], tot_loss[loss=0.2097, ctc_loss=0.1388, cr_loss=0.3547, over 3138181.92 frames. ], batch size: 48, lr: 5.25e-03, grad_scale: 32.0 2024-09-24 02:42:52,505 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.39 vs. limit=15.0 2024-09-24 02:43:05,171 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.58 vs. limit=12.0 2024-09-24 02:43:31,832 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=402654.0, ans=0.125 2024-09-24 02:43:57,895 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.095e+02 1.240e+02 1.347e+02 1.438e+02 1.839e+02, threshold=2.693e+02, percent-clipped=0.0 2024-09-24 02:44:07,892 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=402747.3333333333, ans=0.125 2024-09-24 02:44:12,300 INFO [train.py:1198] (2/4) Epoch 23, batch 600, loss[loss=0.2411, ctc_loss=0.1616, cr_loss=0.3973, over 16808.00 frames. ], tot_loss[loss=0.2108, ctc_loss=0.1396, cr_loss=0.3557, over 3182564.57 frames. ], batch size: 61, lr: 5.25e-03, grad_scale: 32.0 2024-09-24 02:44:12,647 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=402794.0, ans=0.125 2024-09-24 02:45:10,371 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=402934.0, ans=0.1 2024-09-24 02:45:15,115 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=402980.6666666667, ans=0.0 2024-09-24 02:45:25,218 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=10.49 vs. limit=15.0 2024-09-24 02:45:32,502 INFO [train.py:1198] (2/4) Epoch 23, batch 650, loss[loss=0.1828, ctc_loss=0.1177, cr_loss=0.3254, over 16265.00 frames. ], tot_loss[loss=0.2112, ctc_loss=0.1399, cr_loss=0.3569, over 3225500.88 frames. ], batch size: 36, lr: 5.25e-03, grad_scale: 32.0 2024-09-24 02:45:56,137 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=9.17 vs. limit=10.0 2024-09-24 02:46:10,569 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=403120.6666666667, ans=0.2 2024-09-24 02:46:35,803 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=403167.3333333333, ans=0.0 2024-09-24 02:46:43,831 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.037e+02 1.212e+02 1.314e+02 1.389e+02 1.753e+02, threshold=2.627e+02, percent-clipped=0.0 2024-09-24 02:46:57,956 INFO [train.py:1198] (2/4) Epoch 23, batch 700, loss[loss=0.2072, ctc_loss=0.134, cr_loss=0.3662, over 17123.00 frames. ], tot_loss[loss=0.2115, ctc_loss=0.1401, cr_loss=0.357, over 3240594.73 frames. ], batch size: 40, lr: 5.24e-03, grad_scale: 32.0 2024-09-24 02:47:19,184 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=403307.3333333333, ans=0.0 2024-09-24 02:47:54,129 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 02:48:05,482 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=403447.3333333333, ans=0.2 2024-09-24 02:48:16,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=403447.3333333333, ans=0.0 2024-09-24 02:48:20,956 INFO [train.py:1198] (2/4) Epoch 23, batch 750, loss[loss=0.1875, ctc_loss=0.1197, cr_loss=0.3391, over 17113.00 frames. ], tot_loss[loss=0.2108, ctc_loss=0.1395, cr_loss=0.3564, over 3275856.40 frames. ], batch size: 40, lr: 5.24e-03, grad_scale: 32.0 2024-09-24 02:48:21,349 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=403494.0, ans=0.125 2024-09-24 02:48:41,388 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=403540.6666666667, ans=0.0 2024-09-24 02:48:43,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=403540.6666666667, ans=0.1 2024-09-24 02:49:02,579 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=15.40 vs. limit=22.5 2024-09-24 02:49:05,208 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=403587.3333333333, ans=0.2 2024-09-24 02:49:06,700 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=403587.3333333333, ans=0.125 2024-09-24 02:49:30,262 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.065e+02 1.305e+02 1.385e+02 1.543e+02 2.308e+02, threshold=2.771e+02, percent-clipped=0.0 2024-09-24 02:49:32,246 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.max_positive, batch_count=403680.6666666667, ans=0.95 2024-09-24 02:49:43,265 INFO [train.py:1198] (2/4) Epoch 23, batch 800, loss[loss=0.2261, ctc_loss=0.1528, cr_loss=0.3663, over 17217.00 frames. ], tot_loss[loss=0.2096, ctc_loss=0.1386, cr_loss=0.3551, over 3306523.78 frames. ], batch size: 55, lr: 5.24e-03, grad_scale: 32.0 2024-09-24 02:50:20,887 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=403820.6666666667, ans=0.0 2024-09-24 02:50:20,914 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=403820.6666666667, ans=0.0 2024-09-24 02:50:39,938 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=403867.3333333333, ans=0.09899494936611666 2024-09-24 02:51:08,892 INFO [train.py:1198] (2/4) Epoch 23, batch 850, loss[loss=0.2081, ctc_loss=0.1381, cr_loss=0.3497, over 17300.00 frames. ], tot_loss[loss=0.2094, ctc_loss=0.1385, cr_loss=0.3544, over 3308632.72 frames. ], batch size: 46, lr: 5.24e-03, grad_scale: 16.0 2024-09-24 02:52:16,349 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=404147.3333333333, ans=0.0 2024-09-24 02:52:17,546 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.256e+02 1.343e+02 1.446e+02 2.174e+02, threshold=2.686e+02, percent-clipped=0.0 2024-09-24 02:52:28,723 INFO [train.py:1198] (2/4) Epoch 23, batch 900, loss[loss=0.1824, ctc_loss=0.1185, cr_loss=0.3195, over 17130.00 frames. ], tot_loss[loss=0.2082, ctc_loss=0.1376, cr_loss=0.3528, over 3329710.21 frames. ], batch size: 40, lr: 5.24e-03, grad_scale: 16.0 2024-09-24 02:53:00,226 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=404240.6666666667, ans=0.2 2024-09-24 02:53:04,187 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.78 vs. limit=15.0 2024-09-24 02:53:05,602 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=10.84 vs. limit=15.0 2024-09-24 02:53:25,838 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=404334.0, ans=0.0 2024-09-24 02:53:29,005 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=404334.0, ans=0.125 2024-09-24 02:53:36,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=404380.6666666667, ans=0.04949747468305833 2024-09-24 02:53:42,535 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=404380.6666666667, ans=0.025 2024-09-24 02:53:48,981 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=404380.6666666667, ans=0.0 2024-09-24 02:53:53,644 INFO [train.py:1198] (2/4) Epoch 23, batch 950, loss[loss=0.218, ctc_loss=0.1471, cr_loss=0.3547, over 17155.00 frames. ], tot_loss[loss=0.2079, ctc_loss=0.1374, cr_loss=0.3523, over 3334780.87 frames. ], batch size: 48, lr: 5.24e-03, grad_scale: 16.0 2024-09-24 02:54:24,621 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=404520.6666666667, ans=0.125 2024-09-24 02:54:31,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=404520.6666666667, ans=0.125 2024-09-24 02:54:40,022 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.39 vs. limit=15.0 2024-09-24 02:55:02,710 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.116e+02 1.281e+02 1.400e+02 1.572e+02 2.113e+02, threshold=2.800e+02, percent-clipped=0.0 2024-09-24 02:55:03,111 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=404614.0, ans=0.125 2024-09-24 02:55:13,696 INFO [train.py:1198] (2/4) Epoch 23, batch 1000, loss[loss=0.2558, ctc_loss=0.17, cr_loss=0.4289, over 16524.00 frames. ], tot_loss[loss=0.2084, ctc_loss=0.1378, cr_loss=0.3531, over 3346924.16 frames. ], batch size: 66, lr: 5.24e-03, grad_scale: 16.0 2024-09-24 02:55:56,129 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=15.53 vs. limit=22.5 2024-09-24 02:56:01,602 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=404754.0, ans=0.125 2024-09-24 02:56:05,078 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=9.10 vs. limit=15.0 2024-09-24 02:56:17,730 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=404800.6666666667, ans=0.2 2024-09-24 02:56:33,534 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=404847.3333333333, ans=0.0 2024-09-24 02:56:36,219 INFO [train.py:1198] (2/4) Epoch 23, batch 1050, loss[loss=0.1749, ctc_loss=0.1099, cr_loss=0.3252, over 17046.00 frames. ], tot_loss[loss=0.2079, ctc_loss=0.1374, cr_loss=0.3524, over 3345744.40 frames. ], batch size: 39, lr: 5.23e-03, grad_scale: 16.0 2024-09-24 02:56:52,544 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=404940.6666666667, ans=0.125 2024-09-24 02:56:58,627 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=404940.6666666667, ans=0.125 2024-09-24 02:57:12,983 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=404987.3333333333, ans=0.125 2024-09-24 02:57:46,811 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.138e+02 1.300e+02 1.378e+02 1.529e+02 2.270e+02, threshold=2.756e+02, percent-clipped=0.0 2024-09-24 02:57:57,825 INFO [train.py:1198] (2/4) Epoch 23, batch 1100, loss[loss=0.1922, ctc_loss=0.1279, cr_loss=0.3213, over 17295.00 frames. ], tot_loss[loss=0.2078, ctc_loss=0.1374, cr_loss=0.3521, over 3354766.06 frames. ], batch size: 49, lr: 5.23e-03, grad_scale: 16.0 2024-09-24 02:58:20,567 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.max_positive, batch_count=405174.0, ans=0.95 2024-09-24 02:58:39,495 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.92 vs. limit=10.0 2024-09-24 02:58:46,736 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=405267.3333333333, ans=0.2 2024-09-24 02:58:55,206 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=55.05 vs. limit=15.0 2024-09-24 02:58:57,117 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.82 vs. limit=10.0 2024-09-24 02:59:20,147 INFO [train.py:1198] (2/4) Epoch 23, batch 1150, loss[loss=0.1996, ctc_loss=0.1297, cr_loss=0.3497, over 17317.00 frames. ], tot_loss[loss=0.2085, ctc_loss=0.1378, cr_loss=0.3531, over 3356903.13 frames. ], batch size: 51, lr: 5.23e-03, grad_scale: 16.0 2024-09-24 02:59:43,083 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.max_abs, batch_count=405407.3333333333, ans=10.0 2024-09-24 02:59:52,659 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=405454.0, ans=0.2 2024-09-24 02:59:54,215 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=405454.0, ans=0.125 2024-09-24 03:00:03,876 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=405454.0, ans=0.125 2024-09-24 03:00:05,463 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 03:00:26,133 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=405547.3333333333, ans=0.1 2024-09-24 03:00:29,097 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.089e+02 1.259e+02 1.339e+02 1.439e+02 1.652e+02, threshold=2.679e+02, percent-clipped=0.0 2024-09-24 03:00:32,712 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 03:00:40,374 INFO [train.py:1198] (2/4) Epoch 23, batch 1200, loss[loss=0.2514, ctc_loss=0.1712, cr_loss=0.4013, over 16554.00 frames. ], tot_loss[loss=0.2076, ctc_loss=0.1372, cr_loss=0.3519, over 3362872.32 frames. ], batch size: 66, lr: 5.23e-03, grad_scale: 32.0 2024-09-24 03:00:50,175 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=405594.0, ans=0.2 2024-09-24 03:01:06,510 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=405640.6666666667, ans=0.125 2024-09-24 03:01:08,302 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=405640.6666666667, ans=0.0 2024-09-24 03:01:12,225 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=4.69 vs. limit=15.0 2024-09-24 03:01:17,713 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=405687.3333333333, ans=0.125 2024-09-24 03:01:20,971 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=405687.3333333333, ans=0.125 2024-09-24 03:01:59,467 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=405780.6666666667, ans=0.0 2024-09-24 03:02:05,556 INFO [train.py:1198] (2/4) Epoch 23, batch 1250, loss[loss=0.2545, ctc_loss=0.1748, cr_loss=0.3986, over 15153.00 frames. ], tot_loss[loss=0.2077, ctc_loss=0.1373, cr_loss=0.3521, over 3360095.13 frames. ], batch size: 89, lr: 5.23e-03, grad_scale: 32.0 2024-09-24 03:02:41,012 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=405920.6666666667, ans=0.0 2024-09-24 03:03:18,869 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.257e+02 1.354e+02 1.461e+02 2.818e+02, threshold=2.708e+02, percent-clipped=1.0 2024-09-24 03:03:30,974 INFO [train.py:1198] (2/4) Epoch 23, batch 1300, loss[loss=0.1947, ctc_loss=0.1264, cr_loss=0.3413, over 17090.00 frames. ], tot_loss[loss=0.2065, ctc_loss=0.1364, cr_loss=0.3506, over 3355959.14 frames. ], batch size: 40, lr: 5.23e-03, grad_scale: 16.0 2024-09-24 03:03:31,230 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=406060.6666666667, ans=0.0 2024-09-24 03:03:48,521 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=406107.3333333333, ans=0.04949747468305833 2024-09-24 03:03:51,713 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=406107.3333333333, ans=0.025 2024-09-24 03:03:54,833 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=406107.3333333333, ans=0.0 2024-09-24 03:04:23,679 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=406200.6666666667, ans=0.125 2024-09-24 03:04:30,215 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=406200.6666666667, ans=0.125 2024-09-24 03:04:36,592 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=406247.3333333333, ans=0.2 2024-09-24 03:04:50,622 INFO [train.py:1198] (2/4) Epoch 23, batch 1350, loss[loss=0.2055, ctc_loss=0.137, cr_loss=0.3423, over 17152.00 frames. ], tot_loss[loss=0.2083, ctc_loss=0.1377, cr_loss=0.3531, over 3352203.51 frames. ], batch size: 48, lr: 5.23e-03, grad_scale: 8.0 2024-09-24 03:05:06,751 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=406340.6666666667, ans=0.025 2024-09-24 03:05:22,723 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=406387.3333333333, ans=0.125 2024-09-24 03:05:51,178 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=406434.0, ans=0.0 2024-09-24 03:06:07,397 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.071e+02 1.244e+02 1.338e+02 1.449e+02 2.733e+02, threshold=2.676e+02, percent-clipped=1.0 2024-09-24 03:06:15,515 INFO [train.py:1198] (2/4) Epoch 23, batch 1400, loss[loss=0.2067, ctc_loss=0.1391, cr_loss=0.3381, over 16540.00 frames. ], tot_loss[loss=0.2097, ctc_loss=0.1389, cr_loss=0.3541, over 3342795.86 frames. ], batch size: 66, lr: 5.22e-03, grad_scale: 8.0 2024-09-24 03:06:22,663 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.50 vs. limit=15.0 2024-09-24 03:06:40,480 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=10.01 vs. limit=15.0 2024-09-24 03:06:50,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=406620.6666666667, ans=0.025 2024-09-24 03:06:54,224 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=13.66 vs. limit=22.5 2024-09-24 03:07:10,022 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=406667.3333333333, ans=0.125 2024-09-24 03:07:14,757 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=406667.3333333333, ans=0.125 2024-09-24 03:07:16,380 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=406667.3333333333, ans=0.125 2024-09-24 03:07:35,428 INFO [train.py:1198] (2/4) Epoch 23, batch 1450, loss[loss=0.2213, ctc_loss=0.1437, cr_loss=0.3882, over 17348.00 frames. ], tot_loss[loss=0.2089, ctc_loss=0.1382, cr_loss=0.3535, over 3353882.25 frames. ], batch size: 48, lr: 5.22e-03, grad_scale: 8.0 2024-09-24 03:07:40,555 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=406760.6666666667, ans=0.1 2024-09-24 03:07:57,205 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=406807.3333333333, ans=0.1 2024-09-24 03:08:02,304 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.13 vs. limit=15.0 2024-09-24 03:08:07,675 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=11.51 vs. limit=15.0 2024-09-24 03:08:29,019 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=406900.6666666667, ans=0.125 2024-09-24 03:08:32,196 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=406900.6666666667, ans=0.1 2024-09-24 03:08:52,574 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.081e+02 1.227e+02 1.307e+02 1.393e+02 1.809e+02, threshold=2.613e+02, percent-clipped=0.0 2024-09-24 03:08:56,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=406947.3333333333, ans=0.0 2024-09-24 03:09:00,495 INFO [train.py:1198] (2/4) Epoch 23, batch 1500, loss[loss=0.212, ctc_loss=0.1439, cr_loss=0.3402, over 16617.00 frames. ], tot_loss[loss=0.2091, ctc_loss=0.1384, cr_loss=0.3535, over 3349947.73 frames. ], batch size: 61, lr: 5.22e-03, grad_scale: 8.0 2024-09-24 03:09:36,011 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=407087.3333333333, ans=0.125 2024-09-24 03:10:06,323 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=407180.6666666667, ans=0.1 2024-09-24 03:10:09,312 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=407180.6666666667, ans=0.125 2024-09-24 03:10:12,576 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=407180.6666666667, ans=0.125 2024-09-24 03:10:20,385 INFO [train.py:1198] (2/4) Epoch 23, batch 1550, loss[loss=0.1796, ctc_loss=0.1152, cr_loss=0.322, over 16718.00 frames. ], tot_loss[loss=0.2101, ctc_loss=0.139, cr_loss=0.3552, over 3350320.10 frames. ], batch size: 37, lr: 5.22e-03, grad_scale: 8.0 2024-09-24 03:10:25,454 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=407227.3333333333, ans=0.125 2024-09-24 03:10:56,894 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=407320.6666666667, ans=0.1 2024-09-24 03:11:22,993 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=407367.3333333333, ans=0.125 2024-09-24 03:11:37,267 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.292e+02 1.390e+02 1.535e+02 2.050e+02, threshold=2.781e+02, percent-clipped=0.0 2024-09-24 03:11:45,284 INFO [train.py:1198] (2/4) Epoch 23, batch 1600, loss[loss=0.2705, ctc_loss=0.1836, cr_loss=0.4345, over 15084.00 frames. ], tot_loss[loss=0.2114, ctc_loss=0.14, cr_loss=0.3568, over 3337920.28 frames. ], batch size: 89, lr: 5.22e-03, grad_scale: 16.0 2024-09-24 03:12:01,061 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=10.70 vs. limit=15.0 2024-09-24 03:12:03,443 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.00 vs. limit=15.0 2024-09-24 03:12:10,269 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.41 vs. limit=15.0 2024-09-24 03:12:11,717 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten.whitening_limit, batch_count=407507.3333333333, ans=22.5 2024-09-24 03:12:17,543 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=407554.0, ans=0.125 2024-09-24 03:12:25,682 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=407554.0, ans=0.0 2024-09-24 03:12:32,492 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=4.82 vs. limit=12.0 2024-09-24 03:13:06,590 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=407694.0, ans=0.125 2024-09-24 03:13:08,009 INFO [train.py:1198] (2/4) Epoch 23, batch 1650, loss[loss=0.2207, ctc_loss=0.1442, cr_loss=0.3826, over 17299.00 frames. ], tot_loss[loss=0.2107, ctc_loss=0.1396, cr_loss=0.3555, over 3335857.48 frames. ], batch size: 51, lr: 5.22e-03, grad_scale: 16.0 2024-09-24 03:13:11,442 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=407694.0, ans=0.125 2024-09-24 03:13:19,300 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=407694.0, ans=0.0 2024-09-24 03:13:31,455 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer_ff3.min_abs, batch_count=407740.6666666667, ans=0.2 2024-09-24 03:14:21,924 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.102e+02 1.251e+02 1.320e+02 1.451e+02 2.604e+02, threshold=2.640e+02, percent-clipped=0.0 2024-09-24 03:14:29,902 INFO [train.py:1198] (2/4) Epoch 23, batch 1700, loss[loss=0.1908, ctc_loss=0.1231, cr_loss=0.3386, over 17108.00 frames. ], tot_loss[loss=0.2109, ctc_loss=0.1396, cr_loss=0.3562, over 3350774.68 frames. ], batch size: 40, lr: 5.22e-03, grad_scale: 16.0 2024-09-24 03:14:52,497 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=407974.0, ans=0.0 2024-09-24 03:15:21,171 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=408067.3333333333, ans=0.2 2024-09-24 03:15:21,184 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=408067.3333333333, ans=0.125 2024-09-24 03:15:22,716 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=408067.3333333333, ans=0.1 2024-09-24 03:15:38,824 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=408114.0, ans=0.125 2024-09-24 03:15:52,353 INFO [train.py:1198] (2/4) Epoch 23, batch 1750, loss[loss=0.2877, ctc_loss=0.2003, cr_loss=0.4369, over 11584.00 frames. ], tot_loss[loss=0.2116, ctc_loss=0.1401, cr_loss=0.3575, over 3350694.06 frames. ], batch size: 123, lr: 5.21e-03, grad_scale: 16.0 2024-09-24 03:16:08,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=408160.6666666667, ans=0.125 2024-09-24 03:16:14,914 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=10.40 vs. limit=22.5 2024-09-24 03:16:30,100 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=408254.0, ans=0.125 2024-09-24 03:16:57,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=408347.3333333333, ans=0.1 2024-09-24 03:17:06,658 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.263e+02 1.353e+02 1.472e+02 2.567e+02, threshold=2.706e+02, percent-clipped=0.0 2024-09-24 03:17:14,533 INFO [train.py:1198] (2/4) Epoch 23, batch 1800, loss[loss=0.2116, ctc_loss=0.1397, cr_loss=0.3593, over 17133.00 frames. ], tot_loss[loss=0.2117, ctc_loss=0.1402, cr_loss=0.3577, over 3358148.28 frames. ], batch size: 48, lr: 5.21e-03, grad_scale: 16.0 2024-09-24 03:17:21,234 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=408394.0, ans=0.0 2024-09-24 03:17:24,636 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=408394.0, ans=0.125 2024-09-24 03:17:32,987 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=15.34 vs. limit=22.5 2024-09-24 03:18:20,497 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=408534.0, ans=0.125 2024-09-24 03:18:32,072 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 03:18:39,948 INFO [train.py:1198] (2/4) Epoch 23, batch 1850, loss[loss=0.2096, ctc_loss=0.1382, cr_loss=0.3572, over 17156.00 frames. ], tot_loss[loss=0.2108, ctc_loss=0.1395, cr_loss=0.3567, over 3368826.34 frames. ], batch size: 45, lr: 5.21e-03, grad_scale: 16.0 2024-09-24 03:19:09,041 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=408674.0, ans=0.5 2024-09-24 03:19:16,156 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.78 vs. limit=6.0 2024-09-24 03:19:51,998 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.049e+02 1.253e+02 1.335e+02 1.430e+02 2.025e+02, threshold=2.670e+02, percent-clipped=0.0 2024-09-24 03:19:52,829 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=12.34 vs. limit=12.0 2024-09-24 03:19:59,998 INFO [train.py:1198] (2/4) Epoch 23, batch 1900, loss[loss=0.1961, ctc_loss=0.1273, cr_loss=0.3444, over 17270.00 frames. ], tot_loss[loss=0.2121, ctc_loss=0.1406, cr_loss=0.3576, over 3352794.13 frames. ], batch size: 42, lr: 5.21e-03, grad_scale: 16.0 2024-09-24 03:20:04,996 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=408860.6666666667, ans=0.1 2024-09-24 03:20:09,683 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=408860.6666666667, ans=0.0 2024-09-24 03:20:19,413 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=408907.3333333333, ans=0.125 2024-09-24 03:21:06,306 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=409000.6666666667, ans=0.0 2024-09-24 03:21:16,144 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=409047.3333333333, ans=0.2 2024-09-24 03:21:18,505 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn1.whiten, num_groups=1, num_channels=192, metric=10.22 vs. limit=22.5 2024-09-24 03:21:25,579 INFO [train.py:1198] (2/4) Epoch 23, batch 1950, loss[loss=0.2187, ctc_loss=0.1444, cr_loss=0.3715, over 17001.00 frames. ], tot_loss[loss=0.212, ctc_loss=0.1403, cr_loss=0.3582, over 3350533.45 frames. ], batch size: 51, lr: 5.21e-03, grad_scale: 16.0 2024-09-24 03:21:40,356 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=409140.6666666667, ans=0.025 2024-09-24 03:21:55,017 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=409140.6666666667, ans=0.125 2024-09-24 03:21:55,690 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.98 vs. limit=6.0 2024-09-24 03:22:40,198 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.272e+02 1.356e+02 1.502e+02 2.746e+02, threshold=2.712e+02, percent-clipped=1.0 2024-09-24 03:22:43,584 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=409280.6666666667, ans=0.125 2024-09-24 03:22:48,041 INFO [train.py:1198] (2/4) Epoch 23, batch 2000, loss[loss=0.179, ctc_loss=0.1188, cr_loss=0.3014, over 17287.00 frames. ], tot_loss[loss=0.2108, ctc_loss=0.1395, cr_loss=0.3567, over 3351231.88 frames. ], batch size: 42, lr: 5.21e-03, grad_scale: 32.0 2024-09-24 03:23:46,585 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=409467.3333333333, ans=0.0 2024-09-24 03:23:54,657 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=409514.0, ans=0.0 2024-09-24 03:23:57,159 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=9.47 vs. limit=15.0 2024-09-24 03:24:10,500 INFO [train.py:1198] (2/4) Epoch 23, batch 2050, loss[loss=0.2121, ctc_loss=0.1412, cr_loss=0.3543, over 17026.00 frames. ], tot_loss[loss=0.2105, ctc_loss=0.1392, cr_loss=0.3566, over 3354309.24 frames. ], batch size: 51, lr: 5.20e-03, grad_scale: 16.0 2024-09-24 03:24:22,667 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=8.20 vs. limit=15.0 2024-09-24 03:24:23,873 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=409560.6666666667, ans=0.2 2024-09-24 03:24:28,695 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=409607.3333333333, ans=0.0 2024-09-24 03:24:41,306 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=409654.0, ans=0.09899494936611666 2024-09-24 03:24:41,390 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=409654.0, ans=0.025 2024-09-24 03:25:09,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=409700.6666666667, ans=0.125 2024-09-24 03:25:12,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=409700.6666666667, ans=0.125 2024-09-24 03:25:12,648 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.50 vs. limit=15.0 2024-09-24 03:25:24,757 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.130e+02 1.284e+02 1.371e+02 1.476e+02 1.863e+02, threshold=2.742e+02, percent-clipped=0.0 2024-09-24 03:25:31,151 INFO [train.py:1198] (2/4) Epoch 23, batch 2100, loss[loss=0.2583, ctc_loss=0.1808, cr_loss=0.3873, over 11086.00 frames. ], tot_loss[loss=0.2105, ctc_loss=0.1392, cr_loss=0.3564, over 3340671.30 frames. ], batch size: 123, lr: 5.20e-03, grad_scale: 16.0 2024-09-24 03:25:43,470 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 03:26:00,397 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=409840.6666666667, ans=0.125 2024-09-24 03:26:38,608 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=409980.6666666667, ans=0.0 2024-09-24 03:26:51,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=409980.6666666667, ans=0.025 2024-09-24 03:26:56,009 INFO [train.py:1198] (2/4) Epoch 23, batch 2150, loss[loss=0.2008, ctc_loss=0.1344, cr_loss=0.332, over 16425.00 frames. ], tot_loss[loss=0.21, ctc_loss=0.1388, cr_loss=0.3558, over 3342231.14 frames. ], batch size: 66, lr: 5.20e-03, grad_scale: 16.0 2024-09-24 03:27:02,851 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.74 vs. limit=10.0 2024-09-24 03:27:15,512 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=410074.0, ans=0.125 2024-09-24 03:27:20,495 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=410074.0, ans=0.125 2024-09-24 03:27:21,911 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.max_abs, batch_count=410074.0, ans=10.0 2024-09-24 03:27:47,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=410167.3333333333, ans=0.2 2024-09-24 03:28:05,105 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=410214.0, ans=0.0 2024-09-24 03:28:15,214 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.089e+02 1.254e+02 1.321e+02 1.428e+02 2.805e+02, threshold=2.642e+02, percent-clipped=1.0 2024-09-24 03:28:21,743 INFO [train.py:1198] (2/4) Epoch 23, batch 2200, loss[loss=0.204, ctc_loss=0.1316, cr_loss=0.362, over 17206.00 frames. ], tot_loss[loss=0.2104, ctc_loss=0.1391, cr_loss=0.3566, over 3345243.23 frames. ], batch size: 47, lr: 5.20e-03, grad_scale: 16.0 2024-09-24 03:28:52,581 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.49 vs. limit=15.0 2024-09-24 03:28:53,056 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=5.18 vs. limit=15.0 2024-09-24 03:29:06,612 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=410354.0, ans=0.1 2024-09-24 03:29:29,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=410447.3333333333, ans=0.125 2024-09-24 03:29:37,178 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=410447.3333333333, ans=0.125 2024-09-24 03:29:41,637 INFO [train.py:1198] (2/4) Epoch 23, batch 2250, loss[loss=0.2401, ctc_loss=0.1641, cr_loss=0.3804, over 17032.00 frames. ], tot_loss[loss=0.2113, ctc_loss=0.1398, cr_loss=0.3576, over 3341651.88 frames. ], batch size: 52, lr: 5.20e-03, grad_scale: 16.0 2024-09-24 03:30:06,000 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=410540.6666666667, ans=0.09899494936611666 2024-09-24 03:30:20,262 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=410587.3333333333, ans=0.2 2024-09-24 03:30:23,409 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 03:30:36,240 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.48 vs. limit=6.0 2024-09-24 03:31:02,824 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.239e+02 1.305e+02 1.382e+02 2.270e+02, threshold=2.609e+02, percent-clipped=0.0 2024-09-24 03:31:08,208 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=410727.3333333333, ans=0.125 2024-09-24 03:31:09,443 INFO [train.py:1198] (2/4) Epoch 23, batch 2300, loss[loss=0.235, ctc_loss=0.1549, cr_loss=0.4005, over 15839.00 frames. ], tot_loss[loss=0.212, ctc_loss=0.1403, cr_loss=0.3581, over 3330995.40 frames. ], batch size: 74, lr: 5.20e-03, grad_scale: 16.0 2024-09-24 03:31:30,466 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=410774.0, ans=0.035 2024-09-24 03:31:41,823 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=410820.6666666667, ans=0.1 2024-09-24 03:31:46,606 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=410820.6666666667, ans=0.2 2024-09-24 03:31:51,400 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=410820.6666666667, ans=0.1 2024-09-24 03:32:31,673 INFO [train.py:1198] (2/4) Epoch 23, batch 2350, loss[loss=0.2047, ctc_loss=0.1333, cr_loss=0.3571, over 17005.00 frames. ], tot_loss[loss=0.2118, ctc_loss=0.1402, cr_loss=0.3582, over 3337095.55 frames. ], batch size: 51, lr: 5.20e-03, grad_scale: 16.0 2024-09-24 03:32:41,769 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=410960.6666666667, ans=0.2 2024-09-24 03:33:10,995 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=22.16 vs. limit=22.5 2024-09-24 03:33:47,828 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.079e+02 1.259e+02 1.351e+02 1.489e+02 2.005e+02, threshold=2.702e+02, percent-clipped=0.0 2024-09-24 03:33:48,145 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=411147.3333333333, ans=0.125 2024-09-24 03:33:54,326 INFO [train.py:1198] (2/4) Epoch 23, batch 2400, loss[loss=0.2094, ctc_loss=0.1405, cr_loss=0.3446, over 17224.00 frames. ], tot_loss[loss=0.2105, ctc_loss=0.1392, cr_loss=0.3567, over 3347128.97 frames. ], batch size: 50, lr: 5.19e-03, grad_scale: 32.0 2024-09-24 03:34:20,867 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=13.23 vs. limit=22.5 2024-09-24 03:34:42,806 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=411334.0, ans=0.125 2024-09-24 03:34:46,600 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=17.08 vs. limit=22.5 2024-09-24 03:35:14,350 INFO [train.py:1198] (2/4) Epoch 23, batch 2450, loss[loss=0.1672, ctc_loss=0.108, cr_loss=0.2956, over 17042.00 frames. ], tot_loss[loss=0.2116, ctc_loss=0.14, cr_loss=0.3579, over 3343769.31 frames. ], batch size: 39, lr: 5.19e-03, grad_scale: 16.0 2024-09-24 03:35:20,869 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=411427.3333333333, ans=0.125 2024-09-24 03:35:52,374 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=411520.6666666667, ans=0.1 2024-09-24 03:35:58,027 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=411520.6666666667, ans=0.125 2024-09-24 03:36:17,172 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=411567.3333333333, ans=0.2 2024-09-24 03:36:35,082 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.268e+02 1.337e+02 1.491e+02 2.179e+02, threshold=2.674e+02, percent-clipped=0.0 2024-09-24 03:36:38,585 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=411660.6666666667, ans=0.125 2024-09-24 03:36:39,829 INFO [train.py:1198] (2/4) Epoch 23, batch 2500, loss[loss=0.2793, ctc_loss=0.187, cr_loss=0.4612, over 16127.00 frames. ], tot_loss[loss=0.2124, ctc_loss=0.1405, cr_loss=0.3595, over 3345539.94 frames. ], batch size: 74, lr: 5.19e-03, grad_scale: 16.0 2024-09-24 03:36:40,125 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=411660.6666666667, ans=0.125 2024-09-24 03:37:09,833 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.32 vs. limit=15.0 2024-09-24 03:37:36,319 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=411800.6666666667, ans=0.09899494936611666 2024-09-24 03:37:50,221 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=411847.3333333333, ans=0.0 2024-09-24 03:38:02,886 INFO [train.py:1198] (2/4) Epoch 23, batch 2550, loss[loss=0.1788, ctc_loss=0.114, cr_loss=0.3241, over 17275.00 frames. ], tot_loss[loss=0.2115, ctc_loss=0.1399, cr_loss=0.3579, over 3348853.30 frames. ], batch size: 42, lr: 5.19e-03, grad_scale: 16.0 2024-09-24 03:38:03,727 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.51 vs. limit=10.0 2024-09-24 03:38:20,013 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=411940.6666666667, ans=0.125 2024-09-24 03:38:20,058 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=411940.6666666667, ans=0.1 2024-09-24 03:38:52,134 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=412034.0, ans=0.125 2024-09-24 03:38:59,008 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.42 vs. limit=15.0 2024-09-24 03:39:15,788 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=412080.6666666667, ans=0.125 2024-09-24 03:39:20,140 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.256e+02 1.367e+02 1.488e+02 2.148e+02, threshold=2.734e+02, percent-clipped=0.0 2024-09-24 03:39:24,808 INFO [train.py:1198] (2/4) Epoch 23, batch 2600, loss[loss=0.199, ctc_loss=0.1296, cr_loss=0.3472, over 17112.00 frames. ], tot_loss[loss=0.2121, ctc_loss=0.1405, cr_loss=0.3581, over 3346458.31 frames. ], batch size: 49, lr: 5.19e-03, grad_scale: 16.0 2024-09-24 03:39:37,704 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 03:39:56,808 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=412220.6666666667, ans=0.125 2024-09-24 03:40:07,144 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn2.whiten, num_groups=1, num_channels=192, metric=8.24 vs. limit=22.5 2024-09-24 03:40:07,912 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=412220.6666666667, ans=0.0 2024-09-24 03:40:20,399 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=14.93 vs. limit=22.5 2024-09-24 03:40:32,847 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=412314.0, ans=0.125 2024-09-24 03:40:34,388 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=412314.0, ans=0.0 2024-09-24 03:40:44,236 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=412314.0, ans=0.125 2024-09-24 03:40:49,868 INFO [train.py:1198] (2/4) Epoch 23, batch 2650, loss[loss=0.2276, ctc_loss=0.1539, cr_loss=0.3687, over 16808.00 frames. ], tot_loss[loss=0.2114, ctc_loss=0.1399, cr_loss=0.3575, over 3345078.78 frames. ], batch size: 61, lr: 5.19e-03, grad_scale: 16.0 2024-09-24 03:41:28,707 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=412454.0, ans=0.125 2024-09-24 03:41:30,365 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=412454.0, ans=0.0 2024-09-24 03:41:45,201 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=412500.6666666667, ans=0.125 2024-09-24 03:41:54,889 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=412547.3333333333, ans=0.125 2024-09-24 03:41:58,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=412547.3333333333, ans=0.0 2024-09-24 03:42:05,849 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.276e+02 1.370e+02 1.490e+02 2.063e+02, threshold=2.739e+02, percent-clipped=0.0 2024-09-24 03:42:10,850 INFO [train.py:1198] (2/4) Epoch 23, batch 2700, loss[loss=0.2342, ctc_loss=0.1558, cr_loss=0.3919, over 17003.00 frames. ], tot_loss[loss=0.2104, ctc_loss=0.1392, cr_loss=0.3559, over 3347454.83 frames. ], batch size: 53, lr: 5.19e-03, grad_scale: 16.0 2024-09-24 03:42:31,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=412640.6666666667, ans=0.125 2024-09-24 03:43:13,168 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.59 vs. limit=15.0 2024-09-24 03:43:34,026 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=412780.6666666667, ans=0.025 2024-09-24 03:43:37,010 INFO [train.py:1198] (2/4) Epoch 23, batch 2750, loss[loss=0.2161, ctc_loss=0.1447, cr_loss=0.3571, over 16731.00 frames. ], tot_loss[loss=0.2105, ctc_loss=0.1393, cr_loss=0.3561, over 3359948.06 frames. ], batch size: 61, lr: 5.18e-03, grad_scale: 16.0 2024-09-24 03:43:39,338 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=4.11 vs. limit=15.0 2024-09-24 03:43:53,198 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=412874.0, ans=0.0 2024-09-24 03:44:22,570 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=8.48 vs. limit=15.0 2024-09-24 03:44:52,047 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.288e+02 1.407e+02 1.537e+02 4.593e+02, threshold=2.814e+02, percent-clipped=2.0 2024-09-24 03:44:56,674 INFO [train.py:1198] (2/4) Epoch 23, batch 2800, loss[loss=0.2508, ctc_loss=0.1702, cr_loss=0.4028, over 17206.00 frames. ], tot_loss[loss=0.2107, ctc_loss=0.1395, cr_loss=0.356, over 3350008.22 frames. ], batch size: 55, lr: 5.18e-03, grad_scale: 32.0 2024-09-24 03:45:00,168 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=413060.6666666667, ans=0.1 2024-09-24 03:45:04,983 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=413060.6666666667, ans=0.025 2024-09-24 03:45:24,823 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=413107.3333333333, ans=0.125 2024-09-24 03:45:38,049 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=413154.0, ans=0.125 2024-09-24 03:45:39,792 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=7.91 vs. limit=15.0 2024-09-24 03:45:41,066 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=413154.0, ans=0.1 2024-09-24 03:46:21,928 INFO [train.py:1198] (2/4) Epoch 23, batch 2850, loss[loss=0.2065, ctc_loss=0.136, cr_loss=0.3525, over 17220.00 frames. ], tot_loss[loss=0.209, ctc_loss=0.1382, cr_loss=0.3537, over 3357485.36 frames. ], batch size: 47, lr: 5.18e-03, grad_scale: 32.0 2024-09-24 03:46:22,319 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=413294.0, ans=0.07 2024-09-24 03:46:48,672 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=6.80 vs. limit=15.0 2024-09-24 03:46:49,374 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=413340.6666666667, ans=0.1 2024-09-24 03:47:22,364 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=413434.0, ans=0.07 2024-09-24 03:47:25,723 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=413434.0, ans=0.125 2024-09-24 03:47:29,023 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=413480.6666666667, ans=0.125 2024-09-24 03:47:32,092 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=413480.6666666667, ans=0.0 2024-09-24 03:47:39,673 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.145e+02 1.282e+02 1.396e+02 1.534e+02 2.289e+02, threshold=2.792e+02, percent-clipped=0.0 2024-09-24 03:47:40,362 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=5.07 vs. limit=12.0 2024-09-24 03:47:41,636 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=413480.6666666667, ans=0.125 2024-09-24 03:47:44,651 INFO [train.py:1198] (2/4) Epoch 23, batch 2900, loss[loss=0.1713, ctc_loss=0.1109, cr_loss=0.3018, over 17109.00 frames. ], tot_loss[loss=0.2085, ctc_loss=0.1378, cr_loss=0.3535, over 3356251.99 frames. ], batch size: 40, lr: 5.18e-03, grad_scale: 32.0 2024-09-24 03:48:02,364 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 03:48:11,860 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=413574.0, ans=0.09899494936611666 2024-09-24 03:48:19,750 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=413620.6666666667, ans=0.1 2024-09-24 03:48:26,358 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 03:48:35,993 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.min_positive, batch_count=413667.3333333333, ans=0.025 2024-09-24 03:48:43,804 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=413667.3333333333, ans=0.1 2024-09-24 03:48:45,657 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=413667.3333333333, ans=0.125 2024-09-24 03:48:47,548 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.52 vs. limit=15.0 2024-09-24 03:48:59,665 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=413714.0, ans=0.1 2024-09-24 03:49:06,856 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys.whitening_limit, batch_count=413760.6666666667, ans=6.0 2024-09-24 03:49:07,438 INFO [train.py:1198] (2/4) Epoch 23, batch 2950, loss[loss=0.2068, ctc_loss=0.1367, cr_loss=0.3504, over 17195.00 frames. ], tot_loss[loss=0.2094, ctc_loss=0.1384, cr_loss=0.355, over 3355503.68 frames. ], batch size: 47, lr: 5.18e-03, grad_scale: 32.0 2024-09-24 03:49:30,680 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=413807.3333333333, ans=0.125 2024-09-24 03:49:47,099 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=16.21 vs. limit=22.5 2024-09-24 03:50:05,313 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 03:50:16,237 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=413947.3333333333, ans=0.1 2024-09-24 03:50:20,157 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.35 vs. limit=10.0 2024-09-24 03:50:22,355 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.031e+02 1.232e+02 1.296e+02 1.402e+02 3.065e+02, threshold=2.592e+02, percent-clipped=1.0 2024-09-24 03:50:27,182 INFO [train.py:1198] (2/4) Epoch 23, batch 3000, loss[loss=0.1529, ctc_loss=0.0973, cr_loss=0.2781, over 17150.00 frames. ], tot_loss[loss=0.2096, ctc_loss=0.1385, cr_loss=0.3552, over 3351214.61 frames. ], batch size: 41, lr: 5.18e-03, grad_scale: 32.0 2024-09-24 03:50:27,183 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 03:50:42,632 INFO [train.py:1230] (2/4) Epoch 23, validation: loss=0.03816, ctc_loss=0.03816, cr_loss=8.083e-15, over 944034.00 frames. 2024-09-24 03:50:42,633 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 03:50:43,619 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.24 vs. limit=6.0 2024-09-24 03:50:50,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=413994.0, ans=0.04949747468305833 2024-09-24 03:51:08,350 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.72 vs. limit=15.0 2024-09-24 03:51:25,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=414087.3333333333, ans=0.2 2024-09-24 03:51:40,315 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=414134.0, ans=0.125 2024-09-24 03:51:43,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=414134.0, ans=0.125 2024-09-24 03:52:03,781 INFO [train.py:1198] (2/4) Epoch 23, batch 3050, loss[loss=0.2064, ctc_loss=0.1359, cr_loss=0.3529, over 17330.00 frames. ], tot_loss[loss=0.2082, ctc_loss=0.1376, cr_loss=0.3533, over 3358931.00 frames. ], batch size: 52, lr: 5.18e-03, grad_scale: 16.0 2024-09-24 03:52:08,512 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=414227.3333333333, ans=0.125 2024-09-24 03:52:16,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=414227.3333333333, ans=0.125 2024-09-24 03:52:27,787 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=414274.0, ans=0.125 2024-09-24 03:52:38,646 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=414320.6666666667, ans=0.125 2024-09-24 03:52:39,421 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.66 vs. limit=15.0 2024-09-24 03:52:43,981 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.17 vs. limit=15.0 2024-09-24 03:52:58,802 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=414367.3333333333, ans=0.2 2024-09-24 03:53:02,543 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.69 vs. limit=12.0 2024-09-24 03:53:19,031 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.083e+02 1.284e+02 1.405e+02 1.521e+02 2.229e+02, threshold=2.810e+02, percent-clipped=0.0 2024-09-24 03:53:22,205 INFO [train.py:1198] (2/4) Epoch 23, batch 3100, loss[loss=0.235, ctc_loss=0.1531, cr_loss=0.4095, over 17139.00 frames. ], tot_loss[loss=0.2084, ctc_loss=0.1377, cr_loss=0.3533, over 3345829.01 frames. ], batch size: 48, lr: 5.17e-03, grad_scale: 16.0 2024-09-24 03:53:55,277 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=414554.0, ans=0.125 2024-09-24 03:54:38,451 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=414647.3333333333, ans=0.125 2024-09-24 03:54:38,798 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=6.50 vs. limit=12.0 2024-09-24 03:54:42,868 INFO [train.py:1198] (2/4) Epoch 23, batch 3150, loss[loss=0.2501, ctc_loss=0.1645, cr_loss=0.4283, over 17157.00 frames. ], tot_loss[loss=0.2093, ctc_loss=0.1384, cr_loss=0.3549, over 3350900.14 frames. ], batch size: 48, lr: 5.17e-03, grad_scale: 16.0 2024-09-24 03:54:57,095 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=414740.6666666667, ans=0.125 2024-09-24 03:55:00,291 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.min_positive, batch_count=414740.6666666667, ans=0.025 2024-09-24 03:55:19,060 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=414787.3333333333, ans=0.125 2024-09-24 03:55:20,746 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=414787.3333333333, ans=0.125 2024-09-24 03:55:27,012 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=414787.3333333333, ans=0.1 2024-09-24 03:55:38,655 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=13.31 vs. limit=15.0 2024-09-24 03:55:52,941 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.42 vs. limit=15.0 2024-09-24 03:55:55,657 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=414880.6666666667, ans=0.1 2024-09-24 03:55:59,845 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.084e+02 1.246e+02 1.369e+02 1.500e+02 3.916e+02, threshold=2.738e+02, percent-clipped=1.0 2024-09-24 03:56:03,016 INFO [train.py:1198] (2/4) Epoch 23, batch 3200, loss[loss=0.2257, ctc_loss=0.15, cr_loss=0.3786, over 17188.00 frames. ], tot_loss[loss=0.2094, ctc_loss=0.1385, cr_loss=0.3548, over 3347923.88 frames. ], batch size: 55, lr: 5.17e-03, grad_scale: 32.0 2024-09-24 03:56:15,085 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=6.83 vs. limit=15.0 2024-09-24 03:56:29,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=414974.0, ans=0.0 2024-09-24 03:56:43,123 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.whiten, num_groups=1, num_channels=192, metric=4.11 vs. limit=12.0 2024-09-24 03:56:44,285 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=6.63 vs. limit=12.0 2024-09-24 03:56:48,010 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.14 vs. limit=15.0 2024-09-24 03:56:50,392 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=415067.3333333333, ans=0.125 2024-09-24 03:57:04,517 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=415114.0, ans=0.125 2024-09-24 03:57:11,185 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.56 vs. limit=15.0 2024-09-24 03:57:21,131 INFO [train.py:1198] (2/4) Epoch 23, batch 3250, loss[loss=0.161, ctc_loss=0.1026, cr_loss=0.2917, over 16978.00 frames. ], tot_loss[loss=0.2099, ctc_loss=0.1389, cr_loss=0.3552, over 3334881.93 frames. ], batch size: 42, lr: 5.17e-03, grad_scale: 32.0 2024-09-24 03:57:25,165 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.61 vs. limit=15.0 2024-09-24 03:57:29,866 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.82 vs. limit=6.0 2024-09-24 03:57:33,801 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=415160.6666666667, ans=0.025 2024-09-24 03:57:39,999 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=415207.3333333333, ans=0.0 2024-09-24 03:57:46,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=415207.3333333333, ans=0.07 2024-09-24 03:58:15,573 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=415300.6666666667, ans=0.125 2024-09-24 03:58:19,020 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=415300.6666666667, ans=0.0 2024-09-24 03:58:27,590 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.77 vs. limit=15.0 2024-09-24 03:58:36,028 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.138e+02 1.250e+02 1.334e+02 1.506e+02 2.556e+02, threshold=2.668e+02, percent-clipped=0.0 2024-09-24 03:58:39,219 INFO [train.py:1198] (2/4) Epoch 23, batch 3300, loss[loss=0.2111, ctc_loss=0.1404, cr_loss=0.3538, over 16961.00 frames. ], tot_loss[loss=0.2092, ctc_loss=0.1384, cr_loss=0.3541, over 3350792.32 frames. ], batch size: 42, lr: 5.17e-03, grad_scale: 32.0 2024-09-24 03:58:39,420 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=415394.0, ans=0.125 2024-09-24 03:58:44,663 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=10.87 vs. limit=15.0 2024-09-24 03:58:47,460 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=415394.0, ans=0.1 2024-09-24 03:59:09,337 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=415487.3333333333, ans=0.125 2024-09-24 03:59:09,348 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=415487.3333333333, ans=0.1 2024-09-24 03:59:33,259 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.96 vs. limit=15.0 2024-09-24 03:59:37,414 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=415534.0, ans=0.125 2024-09-24 03:59:49,735 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=415580.6666666667, ans=0.125 2024-09-24 03:59:57,317 INFO [train.py:1198] (2/4) Epoch 23, batch 3350, loss[loss=0.1763, ctc_loss=0.1128, cr_loss=0.3177, over 17260.00 frames. ], tot_loss[loss=0.2097, ctc_loss=0.1387, cr_loss=0.3548, over 3355210.78 frames. ], batch size: 42, lr: 5.17e-03, grad_scale: 32.0 2024-09-24 04:00:11,817 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=415674.0, ans=0.0 2024-09-24 04:00:27,657 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=415720.6666666667, ans=0.0 2024-09-24 04:00:38,501 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=415720.6666666667, ans=0.025 2024-09-24 04:00:45,318 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=415767.3333333333, ans=0.0 2024-09-24 04:01:08,862 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=415814.0, ans=0.125 2024-09-24 04:01:14,621 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 1.296e+02 1.415e+02 1.543e+02 2.020e+02, threshold=2.829e+02, percent-clipped=0.0 2024-09-24 04:01:17,788 INFO [train.py:1198] (2/4) Epoch 23, batch 3400, loss[loss=0.23, ctc_loss=0.1552, cr_loss=0.3743, over 17247.00 frames. ], tot_loss[loss=0.2092, ctc_loss=0.1384, cr_loss=0.3541, over 3358154.47 frames. ], batch size: 55, lr: 5.17e-03, grad_scale: 32.0 2024-09-24 04:01:24,116 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=415860.6666666667, ans=0.025 2024-09-24 04:01:38,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=415907.3333333333, ans=0.2 2024-09-24 04:01:49,562 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=415954.0, ans=0.1 2024-09-24 04:02:30,499 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=416047.3333333333, ans=0.125 2024-09-24 04:02:38,113 INFO [train.py:1198] (2/4) Epoch 23, batch 3450, loss[loss=0.2082, ctc_loss=0.1359, cr_loss=0.3615, over 17155.00 frames. ], tot_loss[loss=0.2102, ctc_loss=0.1391, cr_loss=0.3554, over 3357366.36 frames. ], batch size: 48, lr: 5.16e-03, grad_scale: 32.0 2024-09-24 04:02:38,375 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=416094.0, ans=0.125 2024-09-24 04:02:46,959 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=8.09 vs. limit=15.0 2024-09-24 04:03:41,766 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=416280.6666666667, ans=0.125 2024-09-24 04:03:53,800 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.070e+02 1.237e+02 1.333e+02 1.427e+02 1.642e+02, threshold=2.667e+02, percent-clipped=0.0 2024-09-24 04:03:54,114 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=416280.6666666667, ans=0.0 2024-09-24 04:03:57,047 INFO [train.py:1198] (2/4) Epoch 23, batch 3500, loss[loss=0.1814, ctc_loss=0.1158, cr_loss=0.3279, over 17041.00 frames. ], tot_loss[loss=0.2103, ctc_loss=0.1392, cr_loss=0.3555, over 3352922.12 frames. ], batch size: 39, lr: 5.16e-03, grad_scale: 32.0 2024-09-24 04:04:02,036 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=416327.3333333333, ans=0.1 2024-09-24 04:05:17,346 INFO [train.py:1198] (2/4) Epoch 23, batch 3550, loss[loss=0.1796, ctc_loss=0.1168, cr_loss=0.3139, over 16772.00 frames. ], tot_loss[loss=0.2105, ctc_loss=0.1393, cr_loss=0.3561, over 3351482.11 frames. ], batch size: 37, lr: 5.16e-03, grad_scale: 32.0 2024-09-24 04:05:24,379 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=416560.6666666667, ans=0.1 2024-09-24 04:05:55,691 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-24 04:05:57,115 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=416654.0, ans=0.025 2024-09-24 04:05:58,681 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=416654.0, ans=0.125 2024-09-24 04:06:06,908 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=14.19 vs. limit=15.0 2024-09-24 04:06:09,642 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=416700.6666666667, ans=0.0 2024-09-24 04:06:12,639 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=416700.6666666667, ans=0.125 2024-09-24 04:06:34,282 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.083e+02 1.260e+02 1.366e+02 1.469e+02 1.975e+02, threshold=2.732e+02, percent-clipped=0.0 2024-09-24 04:06:37,347 INFO [train.py:1198] (2/4) Epoch 23, batch 3600, loss[loss=0.2042, ctc_loss=0.1347, cr_loss=0.3472, over 17222.00 frames. ], tot_loss[loss=0.2114, ctc_loss=0.14, cr_loss=0.357, over 3348033.90 frames. ], batch size: 50, lr: 5.16e-03, grad_scale: 32.0 2024-09-24 04:06:42,418 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=416794.0, ans=0.125 2024-09-24 04:06:49,971 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=416794.0, ans=0.04949747468305833 2024-09-24 04:06:56,102 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.min_abs, batch_count=416840.6666666667, ans=0.5 2024-09-24 04:07:55,449 INFO [train.py:1198] (2/4) Epoch 23, batch 3650, loss[loss=0.2062, ctc_loss=0.1341, cr_loss=0.3604, over 17008.00 frames. ], tot_loss[loss=0.211, ctc_loss=0.1396, cr_loss=0.3571, over 3356488.95 frames. ], batch size: 44, lr: 5.16e-03, grad_scale: 32.0 2024-09-24 04:08:14,511 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=417074.0, ans=0.0 2024-09-24 04:08:23,840 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=417074.0, ans=0.1 2024-09-24 04:08:57,845 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.min_abs, batch_count=417214.0, ans=0.5 2024-09-24 04:09:11,666 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.081e+02 1.268e+02 1.365e+02 1.526e+02 2.043e+02, threshold=2.731e+02, percent-clipped=0.0 2024-09-24 04:09:14,879 INFO [train.py:1198] (2/4) Epoch 23, batch 3700, loss[loss=0.205, ctc_loss=0.1348, cr_loss=0.3513, over 16988.00 frames. ], tot_loss[loss=0.2102, ctc_loss=0.139, cr_loss=0.3559, over 3350897.58 frames. ], batch size: 53, lr: 5.16e-03, grad_scale: 32.0 2024-09-24 04:09:26,119 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=417260.6666666667, ans=0.125 2024-09-24 04:09:26,136 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=417260.6666666667, ans=0.0 2024-09-24 04:09:26,554 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.62 vs. limit=15.0 2024-09-24 04:09:38,323 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=417307.3333333333, ans=0.025 2024-09-24 04:10:08,051 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=417400.6666666667, ans=0.1 2024-09-24 04:10:09,753 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=417400.6666666667, ans=0.0 2024-09-24 04:10:11,423 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=417400.6666666667, ans=0.125 2024-09-24 04:10:14,787 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.38 vs. limit=10.0 2024-09-24 04:10:27,718 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=417447.3333333333, ans=0.125 2024-09-24 04:10:33,639 INFO [train.py:1198] (2/4) Epoch 23, batch 3750, loss[loss=0.2395, ctc_loss=0.1615, cr_loss=0.3901, over 15870.00 frames. ], tot_loss[loss=0.211, ctc_loss=0.1397, cr_loss=0.3564, over 3341390.39 frames. ], batch size: 74, lr: 5.16e-03, grad_scale: 32.0 2024-09-24 04:10:33,879 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=417494.0, ans=0.0 2024-09-24 04:10:59,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=417540.6666666667, ans=0.05 2024-09-24 04:11:02,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=417540.6666666667, ans=0.125 2024-09-24 04:11:39,836 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=417680.6666666667, ans=0.125 2024-09-24 04:11:47,432 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=417680.6666666667, ans=0.0 2024-09-24 04:11:50,428 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.050e+02 1.304e+02 1.374e+02 1.461e+02 1.993e+02, threshold=2.748e+02, percent-clipped=0.0 2024-09-24 04:11:53,565 INFO [train.py:1198] (2/4) Epoch 23, batch 3800, loss[loss=0.2172, ctc_loss=0.1459, cr_loss=0.3565, over 16997.00 frames. ], tot_loss[loss=0.2118, ctc_loss=0.1404, cr_loss=0.3574, over 3338756.44 frames. ], batch size: 53, lr: 5.15e-03, grad_scale: 32.0 2024-09-24 04:11:53,877 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=417727.3333333333, ans=0.125 2024-09-24 04:12:07,999 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=417774.0, ans=0.0 2024-09-24 04:12:14,926 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.60 vs. limit=15.0 2024-09-24 04:12:23,937 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 04:12:26,922 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=417820.6666666667, ans=0.125 2024-09-24 04:12:34,841 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=417820.6666666667, ans=0.0 2024-09-24 04:12:38,538 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=13.14 vs. limit=15.0 2024-09-24 04:12:56,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=417914.0, ans=0.2 2024-09-24 04:13:12,142 INFO [train.py:1198] (2/4) Epoch 23, batch 3850, loss[loss=0.2016, ctc_loss=0.1305, cr_loss=0.3553, over 17024.00 frames. ], tot_loss[loss=0.2129, ctc_loss=0.1412, cr_loss=0.3582, over 3310675.59 frames. ], batch size: 51, lr: 5.15e-03, grad_scale: 32.0 2024-09-24 04:13:18,506 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=417960.6666666667, ans=0.125 2024-09-24 04:13:30,590 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=418007.3333333333, ans=0.07 2024-09-24 04:13:39,531 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=418007.3333333333, ans=0.0 2024-09-24 04:14:16,833 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=418147.3333333333, ans=0.125 2024-09-24 04:15:15,633 INFO [train.py:1198] (2/4) Epoch 24, batch 0, loss[loss=0.2425, ctc_loss=0.1628, cr_loss=0.3989, over 17044.00 frames. ], tot_loss[loss=0.2425, ctc_loss=0.1628, cr_loss=0.3989, over 17044.00 frames. ], batch size: 52, lr: 5.04e-03, grad_scale: 32.0 2024-09-24 04:15:15,634 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 04:15:26,919 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.4.encoder.layers.2.self_attn_weights, attn_weights_entropy = tensor([4.9031, 4.1626, 4.6043, 4.4801], device='cuda:2') 2024-09-24 04:15:33,340 INFO [train.py:1230] (2/4) Epoch 24, validation: loss=0.03789, ctc_loss=0.03789, cr_loss=8.011e-15, over 944034.00 frames. 2024-09-24 04:15:33,340 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 04:15:36,593 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.212e+02 1.389e+02 1.528e+02 1.642e+02 3.495e+02, threshold=3.056e+02, percent-clipped=0.0 2024-09-24 04:15:39,948 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=418175.3333333333, ans=0.125 2024-09-24 04:15:52,784 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=418222.0, ans=0.07 2024-09-24 04:15:55,894 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=418222.0, ans=0.0 2024-09-24 04:16:23,238 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=418315.3333333333, ans=0.0 2024-09-24 04:16:45,769 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=418362.0, ans=0.125 2024-09-24 04:16:50,615 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=418362.0, ans=0.125 2024-09-24 04:16:52,531 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.33 vs. limit=6.0 2024-09-24 04:16:53,367 INFO [train.py:1198] (2/4) Epoch 24, batch 50, loss[loss=0.1809, ctc_loss=0.1179, cr_loss=0.3148, over 17109.00 frames. ], tot_loss[loss=0.205, ctc_loss=0.1352, cr_loss=0.3492, over 764380.91 frames. ], batch size: 40, lr: 5.04e-03, grad_scale: 32.0 2024-09-24 04:16:55,721 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.64 vs. limit=6.0 2024-09-24 04:17:01,981 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=418408.6666666667, ans=0.125 2024-09-24 04:17:01,983 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=418408.6666666667, ans=0.125 2024-09-24 04:17:56,708 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=418548.6666666667, ans=0.0 2024-09-24 04:17:58,965 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=22.37 vs. limit=22.5 2024-09-24 04:18:15,655 INFO [train.py:1198] (2/4) Epoch 24, batch 100, loss[loss=0.2208, ctc_loss=0.1481, cr_loss=0.3634, over 17021.00 frames. ], tot_loss[loss=0.2064, ctc_loss=0.136, cr_loss=0.352, over 1331613.87 frames. ], batch size: 51, lr: 5.04e-03, grad_scale: 32.0 2024-09-24 04:18:18,829 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.073e+02 1.264e+02 1.318e+02 1.445e+02 2.140e+02, threshold=2.636e+02, percent-clipped=1.0 2024-09-24 04:18:25,532 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=418642.0, ans=0.125 2024-09-24 04:18:35,172 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=418688.6666666667, ans=0.125 2024-09-24 04:18:38,363 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=418688.6666666667, ans=0.1 2024-09-24 04:18:54,471 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=418735.3333333333, ans=0.125 2024-09-24 04:19:31,062 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=418828.6666666667, ans=0.1 2024-09-24 04:19:38,292 INFO [train.py:1198] (2/4) Epoch 24, batch 150, loss[loss=0.2093, ctc_loss=0.1385, cr_loss=0.3542, over 17025.00 frames. ], tot_loss[loss=0.2061, ctc_loss=0.1359, cr_loss=0.3507, over 1777712.87 frames. ], batch size: 44, lr: 5.04e-03, grad_scale: 32.0 2024-09-24 04:19:41,704 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=418875.3333333333, ans=0.125 2024-09-24 04:19:56,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=418922.0, ans=0.1 2024-09-24 04:20:00,867 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=418922.0, ans=0.0 2024-09-24 04:20:07,318 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=418922.0, ans=0.125 2024-09-24 04:21:03,916 INFO [train.py:1198] (2/4) Epoch 24, batch 200, loss[loss=0.2073, ctc_loss=0.1341, cr_loss=0.3658, over 17263.00 frames. ], tot_loss[loss=0.2066, ctc_loss=0.1363, cr_loss=0.3519, over 2127621.23 frames. ], batch size: 44, lr: 5.04e-03, grad_scale: 32.0 2024-09-24 04:21:07,012 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.102e+02 1.247e+02 1.339e+02 1.438e+02 1.930e+02, threshold=2.677e+02, percent-clipped=0.0 2024-09-24 04:21:10,843 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.42 vs. limit=15.0 2024-09-24 04:21:38,037 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=419202.0, ans=0.1 2024-09-24 04:21:49,734 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn1.whiten, num_groups=1, num_channels=192, metric=9.51 vs. limit=22.5 2024-09-24 04:22:26,694 INFO [train.py:1198] (2/4) Epoch 24, batch 250, loss[loss=0.2404, ctc_loss=0.1624, cr_loss=0.3896, over 17023.00 frames. ], tot_loss[loss=0.207, ctc_loss=0.1365, cr_loss=0.3523, over 2403952.18 frames. ], batch size: 56, lr: 5.03e-03, grad_scale: 32.0 2024-09-24 04:22:51,004 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=419388.6666666667, ans=0.1 2024-09-24 04:22:52,678 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=419388.6666666667, ans=0.1 2024-09-24 04:23:09,838 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=419435.3333333333, ans=0.0 2024-09-24 04:23:14,596 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=419482.0, ans=0.125 2024-09-24 04:23:17,921 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=419482.0, ans=0.2 2024-09-24 04:23:21,126 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=419482.0, ans=0.125 2024-09-24 04:23:32,236 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=419528.6666666667, ans=0.04949747468305833 2024-09-24 04:23:46,170 INFO [train.py:1198] (2/4) Epoch 24, batch 300, loss[loss=0.1901, ctc_loss=0.1236, cr_loss=0.3325, over 17339.00 frames. ], tot_loss[loss=0.2067, ctc_loss=0.1364, cr_loss=0.3518, over 2610935.79 frames. ], batch size: 48, lr: 5.03e-03, grad_scale: 32.0 2024-09-24 04:23:49,252 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.095e+02 1.252e+02 1.369e+02 1.490e+02 2.368e+02, threshold=2.737e+02, percent-clipped=0.0 2024-09-24 04:23:52,890 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=419575.3333333333, ans=0.1 2024-09-24 04:23:59,144 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=419575.3333333333, ans=0.125 2024-09-24 04:24:19,799 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=419668.6666666667, ans=0.125 2024-09-24 04:24:20,032 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=18.19 vs. limit=22.5 2024-09-24 04:24:27,588 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=419668.6666666667, ans=0.125 2024-09-24 04:24:45,553 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.95 vs. limit=22.5 2024-09-24 04:25:03,849 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.min_positive, batch_count=419762.0, ans=0.025 2024-09-24 04:25:08,435 INFO [train.py:1198] (2/4) Epoch 24, batch 350, loss[loss=0.2388, ctc_loss=0.1593, cr_loss=0.3977, over 17094.00 frames. ], tot_loss[loss=0.2067, ctc_loss=0.1364, cr_loss=0.3515, over 2777650.32 frames. ], batch size: 49, lr: 5.03e-03, grad_scale: 32.0 2024-09-24 04:25:15,881 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=4.57 vs. limit=15.0 2024-09-24 04:25:45,885 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=419902.0, ans=0.125 2024-09-24 04:25:55,317 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=419902.0, ans=0.125 2024-09-24 04:25:56,103 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=13.59 vs. limit=22.5 2024-09-24 04:26:03,477 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=419948.6666666667, ans=0.125 2024-09-24 04:26:05,488 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=15.90 vs. limit=22.5 2024-09-24 04:26:18,789 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.26 vs. limit=6.0 2024-09-24 04:26:34,308 INFO [train.py:1198] (2/4) Epoch 24, batch 400, loss[loss=0.2133, ctc_loss=0.1425, cr_loss=0.3543, over 17308.00 frames. ], tot_loss[loss=0.2081, ctc_loss=0.1375, cr_loss=0.3533, over 2896767.16 frames. ], batch size: 49, lr: 5.03e-03, grad_scale: 32.0 2024-09-24 04:26:37,588 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.274e+02 1.350e+02 1.482e+02 1.874e+02, threshold=2.700e+02, percent-clipped=0.0 2024-09-24 04:27:07,027 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.66 vs. limit=15.0 2024-09-24 04:27:52,928 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=420228.6666666667, ans=0.125 2024-09-24 04:27:56,960 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=11.75 vs. limit=15.0 2024-09-24 04:27:57,667 INFO [train.py:1198] (2/4) Epoch 24, batch 450, loss[loss=0.2139, ctc_loss=0.1413, cr_loss=0.3625, over 17373.00 frames. ], tot_loss[loss=0.2067, ctc_loss=0.1365, cr_loss=0.3513, over 3002024.93 frames. ], batch size: 48, lr: 5.03e-03, grad_scale: 32.0 2024-09-24 04:28:02,883 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer_na.min_abs, batch_count=420275.3333333333, ans=0.02 2024-09-24 04:28:07,496 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=420275.3333333333, ans=0.125 2024-09-24 04:28:20,743 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=420322.0, ans=0.125 2024-09-24 04:28:36,667 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=420368.6666666667, ans=0.125 2024-09-24 04:28:52,589 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=420415.3333333333, ans=0.125 2024-09-24 04:29:11,587 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.min_positive, batch_count=420462.0, ans=0.05 2024-09-24 04:29:17,789 INFO [train.py:1198] (2/4) Epoch 24, batch 500, loss[loss=0.1945, ctc_loss=0.1265, cr_loss=0.3403, over 17258.00 frames. ], tot_loss[loss=0.2087, ctc_loss=0.1381, cr_loss=0.3532, over 3074964.34 frames. ], batch size: 44, lr: 5.03e-03, grad_scale: 32.0 2024-09-24 04:29:21,053 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.026e+02 1.236e+02 1.308e+02 1.375e+02 2.594e+02, threshold=2.616e+02, percent-clipped=0.0 2024-09-24 04:29:30,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.max_abs, batch_count=420508.6666666667, ans=10.0 2024-09-24 04:29:52,131 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.83 vs. limit=15.0 2024-09-24 04:30:45,504 INFO [train.py:1198] (2/4) Epoch 24, batch 550, loss[loss=0.1838, ctc_loss=0.1198, cr_loss=0.3203, over 17133.00 frames. ], tot_loss[loss=0.209, ctc_loss=0.1382, cr_loss=0.3538, over 3125750.99 frames. ], batch size: 40, lr: 5.03e-03, grad_scale: 16.0 2024-09-24 04:31:09,862 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=420788.6666666667, ans=0.0 2024-09-24 04:31:17,929 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=420835.3333333333, ans=0.125 2024-09-24 04:31:46,449 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 04:32:08,521 INFO [train.py:1198] (2/4) Epoch 24, batch 600, loss[loss=0.1813, ctc_loss=0.1194, cr_loss=0.3091, over 17260.00 frames. ], tot_loss[loss=0.2084, ctc_loss=0.1378, cr_loss=0.3529, over 3172069.35 frames. ], batch size: 44, lr: 5.02e-03, grad_scale: 16.0 2024-09-24 04:32:13,197 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.029e+02 1.249e+02 1.314e+02 1.426e+02 3.030e+02, threshold=2.628e+02, percent-clipped=1.0 2024-09-24 04:32:18,260 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=420975.3333333333, ans=0.125 2024-09-24 04:32:28,187 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.61 vs. limit=22.5 2024-09-24 04:32:29,517 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=421022.0, ans=0.0 2024-09-24 04:33:01,539 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=421115.3333333333, ans=0.125 2024-09-24 04:33:04,662 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 04:33:20,771 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.min_positive, batch_count=421162.0, ans=0.025 2024-09-24 04:33:25,615 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=421162.0, ans=0.0 2024-09-24 04:33:28,573 INFO [train.py:1198] (2/4) Epoch 24, batch 650, loss[loss=0.2478, ctc_loss=0.1648, cr_loss=0.415, over 17049.00 frames. ], tot_loss[loss=0.2094, ctc_loss=0.1384, cr_loss=0.3549, over 3198743.03 frames. ], batch size: 52, lr: 5.02e-03, grad_scale: 16.0 2024-09-24 04:33:35,842 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.56 vs. limit=15.0 2024-09-24 04:33:41,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=421208.6666666667, ans=0.125 2024-09-24 04:33:51,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=421255.3333333333, ans=0.1 2024-09-24 04:34:02,570 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=421302.0, ans=0.2 2024-09-24 04:34:10,890 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=4.08 vs. limit=12.0 2024-09-24 04:34:29,001 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=421348.6666666667, ans=0.125 2024-09-24 04:34:29,078 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=421348.6666666667, ans=0.125 2024-09-24 04:34:48,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=421395.3333333333, ans=0.125 2024-09-24 04:34:50,328 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=421442.0, ans=0.1 2024-09-24 04:34:51,510 INFO [train.py:1198] (2/4) Epoch 24, batch 700, loss[loss=0.2588, ctc_loss=0.1765, cr_loss=0.4113, over 14853.00 frames. ], tot_loss[loss=0.2093, ctc_loss=0.1384, cr_loss=0.3548, over 3240937.86 frames. ], batch size: 89, lr: 5.02e-03, grad_scale: 16.0 2024-09-24 04:34:51,757 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=421442.0, ans=0.125 2024-09-24 04:34:53,366 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=421442.0, ans=0.125 2024-09-24 04:34:56,334 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.315e+02 1.415e+02 1.543e+02 2.275e+02, threshold=2.830e+02, percent-clipped=0.0 2024-09-24 04:35:41,048 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=421582.0, ans=0.2 2024-09-24 04:36:05,050 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer_ff2.min_abs, batch_count=421628.6666666667, ans=0.1 2024-09-24 04:36:14,458 INFO [train.py:1198] (2/4) Epoch 24, batch 750, loss[loss=0.2117, ctc_loss=0.1404, cr_loss=0.3565, over 17153.00 frames. ], tot_loss[loss=0.2086, ctc_loss=0.1378, cr_loss=0.354, over 3267526.26 frames. ], batch size: 48, lr: 5.02e-03, grad_scale: 16.0 2024-09-24 04:37:11,902 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=421815.3333333333, ans=0.125 2024-09-24 04:37:22,077 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=6.84 vs. limit=15.0 2024-09-24 04:37:37,162 INFO [train.py:1198] (2/4) Epoch 24, batch 800, loss[loss=0.1931, ctc_loss=0.1275, cr_loss=0.328, over 17311.00 frames. ], tot_loss[loss=0.2074, ctc_loss=0.1369, cr_loss=0.3521, over 3288186.50 frames. ], batch size: 51, lr: 5.02e-03, grad_scale: 32.0 2024-09-24 04:37:37,526 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=421908.6666666667, ans=0.0 2024-09-24 04:37:41,989 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.122e+02 1.296e+02 1.355e+02 1.500e+02 2.153e+02, threshold=2.711e+02, percent-clipped=0.0 2024-09-24 04:37:58,210 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=421955.3333333333, ans=0.125 2024-09-24 04:38:34,215 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.05 vs. limit=15.0 2024-09-24 04:38:38,922 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.70 vs. limit=15.0 2024-09-24 04:38:57,343 INFO [train.py:1198] (2/4) Epoch 24, batch 850, loss[loss=0.1816, ctc_loss=0.1164, cr_loss=0.3263, over 16683.00 frames. ], tot_loss[loss=0.2081, ctc_loss=0.1374, cr_loss=0.3537, over 3306304.73 frames. ], batch size: 37, lr: 5.02e-03, grad_scale: 32.0 2024-09-24 04:39:20,827 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=422188.6666666667, ans=0.025 2024-09-24 04:39:26,225 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=11.98 vs. limit=15.0 2024-09-24 04:39:38,988 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=19.83 vs. limit=22.5 2024-09-24 04:39:43,619 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.96 vs. limit=15.0 2024-09-24 04:39:47,999 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=422282.0, ans=0.125 2024-09-24 04:39:48,430 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=512, metric=13.37 vs. limit=22.5 2024-09-24 04:39:59,098 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=422282.0, ans=0.0 2024-09-24 04:40:25,045 INFO [train.py:1198] (2/4) Epoch 24, batch 900, loss[loss=0.2227, ctc_loss=0.1462, cr_loss=0.3825, over 16489.00 frames. ], tot_loss[loss=0.2082, ctc_loss=0.1375, cr_loss=0.3535, over 3320678.81 frames. ], batch size: 66, lr: 5.02e-03, grad_scale: 32.0 2024-09-24 04:40:29,775 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.142e+02 1.271e+02 1.382e+02 1.510e+02 2.333e+02, threshold=2.763e+02, percent-clipped=0.0 2024-09-24 04:40:49,127 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=422422.0, ans=0.125 2024-09-24 04:41:05,155 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=422468.6666666667, ans=0.2 2024-09-24 04:41:16,754 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.47 vs. limit=15.0 2024-09-24 04:41:19,477 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=422515.3333333333, ans=0.125 2024-09-24 04:41:22,732 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=422515.3333333333, ans=0.125 2024-09-24 04:41:29,389 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 04:41:36,462 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.36 vs. limit=15.0 2024-09-24 04:41:37,835 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.14 vs. limit=6.0 2024-09-24 04:41:42,721 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.47 vs. limit=10.0 2024-09-24 04:41:45,282 INFO [train.py:1198] (2/4) Epoch 24, batch 950, loss[loss=0.2415, ctc_loss=0.1621, cr_loss=0.3968, over 14888.00 frames. ], tot_loss[loss=0.2096, ctc_loss=0.1385, cr_loss=0.3554, over 3318885.86 frames. ], batch size: 89, lr: 5.01e-03, grad_scale: 32.0 2024-09-24 04:41:45,508 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=422608.6666666667, ans=0.035 2024-09-24 04:41:50,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=422608.6666666667, ans=0.125 2024-09-24 04:42:31,992 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=422702.0, ans=0.125 2024-09-24 04:43:00,914 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=422795.3333333333, ans=0.0 2024-09-24 04:43:08,570 INFO [train.py:1198] (2/4) Epoch 24, batch 1000, loss[loss=0.1867, ctc_loss=0.1214, cr_loss=0.3265, over 17094.00 frames. ], tot_loss[loss=0.2106, ctc_loss=0.1392, cr_loss=0.3569, over 3328139.85 frames. ], batch size: 40, lr: 5.01e-03, grad_scale: 32.0 2024-09-24 04:43:10,373 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.min_positive, batch_count=422842.0, ans=0.025 2024-09-24 04:43:13,212 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.084e+02 1.311e+02 1.412e+02 1.541e+02 1.926e+02, threshold=2.824e+02, percent-clipped=0.0 2024-09-24 04:43:40,613 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=422935.3333333333, ans=0.125 2024-09-24 04:43:59,695 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=422982.0, ans=0.125 2024-09-24 04:44:30,827 INFO [train.py:1198] (2/4) Epoch 24, batch 1050, loss[loss=0.1808, ctc_loss=0.1174, cr_loss=0.3171, over 17041.00 frames. ], tot_loss[loss=0.2096, ctc_loss=0.1386, cr_loss=0.3552, over 3331702.23 frames. ], batch size: 39, lr: 5.01e-03, grad_scale: 32.0 2024-09-24 04:45:09,075 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.53 vs. limit=15.0 2024-09-24 04:45:26,170 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=423215.3333333333, ans=0.125 2024-09-24 04:45:56,522 INFO [train.py:1198] (2/4) Epoch 24, batch 1100, loss[loss=0.2167, ctc_loss=0.1447, cr_loss=0.3602, over 16653.00 frames. ], tot_loss[loss=0.2098, ctc_loss=0.1385, cr_loss=0.3562, over 3338848.59 frames. ], batch size: 66, lr: 5.01e-03, grad_scale: 32.0 2024-09-24 04:46:01,273 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 9.919e+01 1.239e+02 1.343e+02 1.468e+02 1.769e+02, threshold=2.686e+02, percent-clipped=0.0 2024-09-24 04:46:14,476 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=423355.3333333333, ans=0.0 2024-09-24 04:46:19,050 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=423355.3333333333, ans=0.125 2024-09-24 04:46:19,167 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=423355.3333333333, ans=0.125 2024-09-24 04:46:25,557 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-24 04:46:35,146 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=423402.0, ans=0.125 2024-09-24 04:46:49,476 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 04:46:51,178 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=423448.6666666667, ans=0.0 2024-09-24 04:46:59,957 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=423448.6666666667, ans=0.125 2024-09-24 04:47:11,162 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=423495.3333333333, ans=0.125 2024-09-24 04:47:17,474 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=423542.0, ans=0.025 2024-09-24 04:47:18,698 INFO [train.py:1198] (2/4) Epoch 24, batch 1150, loss[loss=0.2192, ctc_loss=0.1439, cr_loss=0.3767, over 17000.00 frames. ], tot_loss[loss=0.21, ctc_loss=0.1386, cr_loss=0.357, over 3338470.63 frames. ], batch size: 56, lr: 5.01e-03, grad_scale: 32.0 2024-09-24 04:47:30,080 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=423542.0, ans=0.025 2024-09-24 04:47:36,632 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=423588.6666666667, ans=0.125 2024-09-24 04:47:54,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=423635.3333333333, ans=0.1 2024-09-24 04:48:14,077 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=11.28 vs. limit=15.0 2024-09-24 04:48:27,367 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.79 vs. limit=10.0 2024-09-24 04:48:39,272 INFO [train.py:1198] (2/4) Epoch 24, batch 1200, loss[loss=0.1898, ctc_loss=0.1238, cr_loss=0.33, over 17294.00 frames. ], tot_loss[loss=0.2107, ctc_loss=0.1392, cr_loss=0.3574, over 3336174.30 frames. ], batch size: 49, lr: 5.01e-03, grad_scale: 32.0 2024-09-24 04:48:44,027 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.099e+02 1.248e+02 1.325e+02 1.416e+02 2.562e+02, threshold=2.650e+02, percent-clipped=0.0 2024-09-24 04:48:46,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=423775.3333333333, ans=0.07 2024-09-24 04:49:00,455 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=423822.0, ans=0.0 2024-09-24 04:49:42,701 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=423915.3333333333, ans=0.2 2024-09-24 04:49:47,445 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=423962.0, ans=0.0 2024-09-24 04:49:48,920 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=423962.0, ans=0.125 2024-09-24 04:50:03,553 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=423962.0, ans=0.2 2024-09-24 04:50:06,367 INFO [train.py:1198] (2/4) Epoch 24, batch 1250, loss[loss=0.1953, ctc_loss=0.1248, cr_loss=0.3525, over 17195.00 frames. ], tot_loss[loss=0.2094, ctc_loss=0.1383, cr_loss=0.3555, over 3340539.39 frames. ], batch size: 41, lr: 5.01e-03, grad_scale: 32.0 2024-09-24 04:50:11,227 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=424008.6666666667, ans=0.1 2024-09-24 04:50:43,569 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=424102.0, ans=0.125 2024-09-24 04:50:53,308 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=424148.6666666667, ans=0.125 2024-09-24 04:51:17,320 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=424195.3333333333, ans=0.125 2024-09-24 04:51:26,792 INFO [train.py:1198] (2/4) Epoch 24, batch 1300, loss[loss=0.2366, ctc_loss=0.1568, cr_loss=0.3991, over 17231.00 frames. ], tot_loss[loss=0.2094, ctc_loss=0.1382, cr_loss=0.3556, over 3353060.49 frames. ], batch size: 50, lr: 5.00e-03, grad_scale: 32.0 2024-09-24 04:51:31,562 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.033e+02 1.254e+02 1.331e+02 1.449e+02 1.850e+02, threshold=2.662e+02, percent-clipped=0.0 2024-09-24 04:52:43,273 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=424428.6666666667, ans=0.125 2024-09-24 04:52:49,352 INFO [train.py:1198] (2/4) Epoch 24, batch 1350, loss[loss=0.1839, ctc_loss=0.1184, cr_loss=0.3271, over 17244.00 frames. ], tot_loss[loss=0.2094, ctc_loss=0.1382, cr_loss=0.3559, over 3355170.09 frames. ], batch size: 44, lr: 5.00e-03, grad_scale: 32.0 2024-09-24 04:52:50,061 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.95 vs. limit=6.0 2024-09-24 04:53:00,693 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=424475.3333333333, ans=0.125 2024-09-24 04:53:19,753 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=424568.6666666667, ans=0.0 2024-09-24 04:53:31,911 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.93 vs. limit=5.0 2024-09-24 04:53:32,498 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=424568.6666666667, ans=0.0 2024-09-24 04:53:48,696 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=424615.3333333333, ans=0.1 2024-09-24 04:54:11,954 INFO [train.py:1198] (2/4) Epoch 24, batch 1400, loss[loss=0.2274, ctc_loss=0.1506, cr_loss=0.3841, over 17203.00 frames. ], tot_loss[loss=0.2095, ctc_loss=0.1383, cr_loss=0.3562, over 3346143.65 frames. ], batch size: 55, lr: 5.00e-03, grad_scale: 32.0 2024-09-24 04:54:16,812 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.131e+02 1.257e+02 1.358e+02 1.495e+02 1.831e+02, threshold=2.716e+02, percent-clipped=0.0 2024-09-24 04:54:26,814 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=424755.3333333333, ans=0.0 2024-09-24 04:54:28,369 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=424755.3333333333, ans=0.125 2024-09-24 04:54:30,403 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=6.49 vs. limit=15.0 2024-09-24 04:54:56,713 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=424802.0, ans=0.125 2024-09-24 04:55:05,700 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=424848.6666666667, ans=0.125 2024-09-24 04:55:21,724 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=424895.3333333333, ans=0.0 2024-09-24 04:55:21,806 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=424895.3333333333, ans=0.125 2024-09-24 04:55:37,530 INFO [train.py:1198] (2/4) Epoch 24, batch 1450, loss[loss=0.2401, ctc_loss=0.1605, cr_loss=0.3983, over 16994.00 frames. ], tot_loss[loss=0.2109, ctc_loss=0.1393, cr_loss=0.3581, over 3337176.98 frames. ], batch size: 51, lr: 5.00e-03, grad_scale: 32.0 2024-09-24 04:55:49,137 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=424942.0, ans=0.2 2024-09-24 04:55:55,372 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=424988.6666666667, ans=0.125 2024-09-24 04:55:55,386 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=424988.6666666667, ans=0.125 2024-09-24 04:55:55,511 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 04:56:41,725 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 04:56:44,850 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=425128.6666666667, ans=0.125 2024-09-24 04:56:52,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=425128.6666666667, ans=0.0 2024-09-24 04:56:55,216 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=425128.6666666667, ans=0.0 2024-09-24 04:56:59,699 INFO [train.py:1198] (2/4) Epoch 24, batch 1500, loss[loss=0.2031, ctc_loss=0.1363, cr_loss=0.3341, over 17221.00 frames. ], tot_loss[loss=0.2096, ctc_loss=0.1383, cr_loss=0.3563, over 3349969.80 frames. ], batch size: 47, lr: 5.00e-03, grad_scale: 32.0 2024-09-24 04:57:01,967 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=9.48 vs. limit=15.0 2024-09-24 04:57:04,518 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.065e+02 1.246e+02 1.334e+02 1.449e+02 2.075e+02, threshold=2.667e+02, percent-clipped=0.0 2024-09-24 04:57:07,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=425175.3333333333, ans=0.0 2024-09-24 04:57:08,297 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=13.06 vs. limit=15.0 2024-09-24 04:57:17,651 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=425222.0, ans=0.125 2024-09-24 04:57:26,945 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.min_positive, batch_count=425222.0, ans=0.05 2024-09-24 04:57:28,735 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=425222.0, ans=0.2 2024-09-24 04:57:44,861 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=425268.6666666667, ans=0.1 2024-09-24 04:58:03,998 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=425362.0, ans=0.0 2024-09-24 04:58:19,423 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.88 vs. limit=6.0 2024-09-24 04:58:19,809 INFO [train.py:1198] (2/4) Epoch 24, batch 1550, loss[loss=0.2, ctc_loss=0.131, cr_loss=0.3447, over 17260.00 frames. ], tot_loss[loss=0.2096, ctc_loss=0.1384, cr_loss=0.3561, over 3349633.52 frames. ], batch size: 44, lr: 5.00e-03, grad_scale: 32.0 2024-09-24 04:58:43,591 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=7.41 vs. limit=8.0 2024-09-24 04:59:04,800 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=425502.0, ans=0.125 2024-09-24 04:59:13,732 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=425548.6666666667, ans=0.0 2024-09-24 04:59:42,275 INFO [train.py:1198] (2/4) Epoch 24, batch 1600, loss[loss=0.2642, ctc_loss=0.184, cr_loss=0.4008, over 11997.00 frames. ], tot_loss[loss=0.2095, ctc_loss=0.1384, cr_loss=0.3554, over 3337211.76 frames. ], batch size: 123, lr: 5.00e-03, grad_scale: 32.0 2024-09-24 04:59:42,643 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=425642.0, ans=0.125 2024-09-24 04:59:44,198 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=425642.0, ans=0.125 2024-09-24 04:59:47,003 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.230e+02 1.386e+02 1.499e+02 2.034e+02, threshold=2.773e+02, percent-clipped=0.0 2024-09-24 05:00:40,027 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.min_positive, batch_count=425782.0, ans=0.05 2024-09-24 05:00:46,530 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=425782.0, ans=0.125 2024-09-24 05:00:52,931 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=425828.6666666667, ans=0.05 2024-09-24 05:00:59,409 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=425828.6666666667, ans=0.2 2024-09-24 05:01:07,191 INFO [train.py:1198] (2/4) Epoch 24, batch 1650, loss[loss=0.1548, ctc_loss=0.09789, cr_loss=0.2845, over 17161.00 frames. ], tot_loss[loss=0.2087, ctc_loss=0.1378, cr_loss=0.3542, over 3349533.22 frames. ], batch size: 41, lr: 5.00e-03, grad_scale: 32.0 2024-09-24 05:01:07,578 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=425875.3333333333, ans=0.125 2024-09-24 05:01:19,261 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=16.85 vs. limit=22.5 2024-09-24 05:01:28,036 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=425922.0, ans=0.2 2024-09-24 05:01:42,407 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=425968.6666666667, ans=0.0 2024-09-24 05:01:43,140 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.32 vs. limit=22.5 2024-09-24 05:01:44,339 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=9.31 vs. limit=12.0 2024-09-24 05:02:02,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=426015.3333333333, ans=0.0 2024-09-24 05:02:10,800 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=426015.3333333333, ans=0.0 2024-09-24 05:02:13,823 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=426062.0, ans=0.125 2024-09-24 05:02:29,579 INFO [train.py:1198] (2/4) Epoch 24, batch 1700, loss[loss=0.1902, ctc_loss=0.1252, cr_loss=0.325, over 17221.00 frames. ], tot_loss[loss=0.2078, ctc_loss=0.1372, cr_loss=0.3529, over 3349259.98 frames. ], batch size: 47, lr: 4.99e-03, grad_scale: 32.0 2024-09-24 05:02:34,432 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.248e+02 1.319e+02 1.421e+02 3.276e+02, threshold=2.637e+02, percent-clipped=2.0 2024-09-24 05:03:05,223 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=426202.0, ans=0.125 2024-09-24 05:03:10,116 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=426202.0, ans=0.125 2024-09-24 05:03:31,354 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=426248.6666666667, ans=0.0 2024-09-24 05:03:47,739 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=426295.3333333333, ans=0.1 2024-09-24 05:03:50,634 INFO [train.py:1198] (2/4) Epoch 24, batch 1750, loss[loss=0.1885, ctc_loss=0.1235, cr_loss=0.3248, over 17093.00 frames. ], tot_loss[loss=0.2091, ctc_loss=0.1382, cr_loss=0.3543, over 3351083.73 frames. ], batch size: 40, lr: 4.99e-03, grad_scale: 32.0 2024-09-24 05:04:52,084 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=426482.0, ans=0.125 2024-09-24 05:05:10,327 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=426528.6666666667, ans=0.125 2024-09-24 05:05:17,716 INFO [train.py:1198] (2/4) Epoch 24, batch 1800, loss[loss=0.2067, ctc_loss=0.1351, cr_loss=0.3581, over 17170.00 frames. ], tot_loss[loss=0.2084, ctc_loss=0.1376, cr_loss=0.3542, over 3364208.29 frames. ], batch size: 45, lr: 4.99e-03, grad_scale: 32.0 2024-09-24 05:05:22,469 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.066e+02 1.235e+02 1.322e+02 1.422e+02 1.827e+02, threshold=2.643e+02, percent-clipped=0.0 2024-09-24 05:06:10,725 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=426715.3333333333, ans=0.1 2024-09-24 05:06:10,731 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=426715.3333333333, ans=0.0 2024-09-24 05:06:12,611 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=426715.3333333333, ans=0.2 2024-09-24 05:06:37,695 INFO [train.py:1198] (2/4) Epoch 24, batch 1850, loss[loss=0.2233, ctc_loss=0.1488, cr_loss=0.3725, over 16840.00 frames. ], tot_loss[loss=0.2098, ctc_loss=0.1386, cr_loss=0.3562, over 3364691.88 frames. ], batch size: 61, lr: 4.99e-03, grad_scale: 32.0 2024-09-24 05:06:38,167 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 05:06:44,366 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=426808.6666666667, ans=0.125 2024-09-24 05:07:02,978 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=426855.3333333333, ans=0.2 2024-09-24 05:07:16,163 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=3.03 vs. limit=6.0 2024-09-24 05:07:42,946 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=426995.3333333333, ans=0.2 2024-09-24 05:07:57,671 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.65 vs. limit=15.0 2024-09-24 05:08:00,025 INFO [train.py:1198] (2/4) Epoch 24, batch 1900, loss[loss=0.2123, ctc_loss=0.1392, cr_loss=0.3652, over 16896.00 frames. ], tot_loss[loss=0.2088, ctc_loss=0.1378, cr_loss=0.3549, over 3370409.57 frames. ], batch size: 58, lr: 4.99e-03, grad_scale: 16.0 2024-09-24 05:08:06,244 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.067e+02 1.237e+02 1.306e+02 1.387e+02 1.778e+02, threshold=2.611e+02, percent-clipped=0.0 2024-09-24 05:08:53,166 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=427182.0, ans=0.04949747468305833 2024-09-24 05:08:56,706 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=427182.0, ans=0.025 2024-09-24 05:09:21,414 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=427275.3333333333, ans=0.95 2024-09-24 05:09:22,692 INFO [train.py:1198] (2/4) Epoch 24, batch 1950, loss[loss=0.2281, ctc_loss=0.152, cr_loss=0.3808, over 17153.00 frames. ], tot_loss[loss=0.2079, ctc_loss=0.1371, cr_loss=0.3538, over 3371756.88 frames. ], batch size: 48, lr: 4.99e-03, grad_scale: 16.0 2024-09-24 05:09:31,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=427275.3333333333, ans=0.2 2024-09-24 05:10:08,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=427368.6666666667, ans=0.125 2024-09-24 05:10:20,784 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=427415.3333333333, ans=0.025 2024-09-24 05:10:47,725 INFO [train.py:1198] (2/4) Epoch 24, batch 2000, loss[loss=0.1977, ctc_loss=0.1298, cr_loss=0.3397, over 17314.00 frames. ], tot_loss[loss=0.2077, ctc_loss=0.1371, cr_loss=0.3532, over 3362867.04 frames. ], batch size: 51, lr: 4.99e-03, grad_scale: 16.0 2024-09-24 05:10:55,710 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.029e+02 1.252e+02 1.338e+02 1.434e+02 1.849e+02, threshold=2.676e+02, percent-clipped=0.0 2024-09-24 05:10:56,146 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=427508.6666666667, ans=0.2 2024-09-24 05:10:57,836 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=6.02 vs. limit=15.0 2024-09-24 05:11:01,190 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.28 vs. limit=15.0 2024-09-24 05:11:26,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=427602.0, ans=0.0 2024-09-24 05:12:09,510 INFO [train.py:1198] (2/4) Epoch 24, batch 2050, loss[loss=0.1864, ctc_loss=0.1232, cr_loss=0.3157, over 16795.00 frames. ], tot_loss[loss=0.2077, ctc_loss=0.137, cr_loss=0.3534, over 3370440.03 frames. ], batch size: 61, lr: 4.98e-03, grad_scale: 16.0 2024-09-24 05:12:17,764 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=427742.0, ans=0.0 2024-09-24 05:13:07,391 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=427882.0, ans=0.1 2024-09-24 05:13:15,673 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten.whitening_limit, batch_count=427928.6666666667, ans=15.0 2024-09-24 05:13:20,500 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=11.27 vs. limit=15.0 2024-09-24 05:13:29,442 INFO [train.py:1198] (2/4) Epoch 24, batch 2100, loss[loss=0.1946, ctc_loss=0.1239, cr_loss=0.3537, over 17297.00 frames. ], tot_loss[loss=0.2064, ctc_loss=0.136, cr_loss=0.352, over 3376935.98 frames. ], batch size: 46, lr: 4.98e-03, grad_scale: 16.0 2024-09-24 05:13:37,476 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.278e+02 1.335e+02 1.481e+02 2.167e+02, threshold=2.670e+02, percent-clipped=0.0 2024-09-24 05:13:49,162 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=428022.0, ans=0.0 2024-09-24 05:13:52,469 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=428022.0, ans=0.2 2024-09-24 05:14:03,597 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=428068.6666666667, ans=0.0 2024-09-24 05:14:03,648 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=428068.6666666667, ans=0.0 2024-09-24 05:14:30,596 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=18.43 vs. limit=22.5 2024-09-24 05:14:54,809 INFO [train.py:1198] (2/4) Epoch 24, batch 2150, loss[loss=0.2102, ctc_loss=0.1373, cr_loss=0.3645, over 17352.00 frames. ], tot_loss[loss=0.2062, ctc_loss=0.1358, cr_loss=0.3519, over 3385362.73 frames. ], batch size: 48, lr: 4.98e-03, grad_scale: 16.0 2024-09-24 05:14:57,908 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=428208.6666666667, ans=0.025 2024-09-24 05:15:21,881 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=8.16 vs. limit=15.0 2024-09-24 05:15:26,503 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=428255.3333333333, ans=0.04949747468305833 2024-09-24 05:16:05,207 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=428395.3333333333, ans=0.125 2024-09-24 05:16:14,709 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=428395.3333333333, ans=0.0 2024-09-24 05:16:17,769 INFO [train.py:1198] (2/4) Epoch 24, batch 2200, loss[loss=0.232, ctc_loss=0.1528, cr_loss=0.3958, over 16751.00 frames. ], tot_loss[loss=0.2068, ctc_loss=0.1361, cr_loss=0.3533, over 3386294.30 frames. ], batch size: 61, lr: 4.98e-03, grad_scale: 16.0 2024-09-24 05:16:21,584 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.63 vs. limit=15.0 2024-09-24 05:16:25,766 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.082e+02 1.245e+02 1.329e+02 1.411e+02 2.102e+02, threshold=2.657e+02, percent-clipped=0.0 2024-09-24 05:16:35,897 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=428488.6666666667, ans=0.04949747468305833 2024-09-24 05:17:24,336 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=6.44 vs. limit=15.0 2024-09-24 05:17:40,961 INFO [train.py:1198] (2/4) Epoch 24, batch 2250, loss[loss=0.1841, ctc_loss=0.1209, cr_loss=0.3159, over 17023.00 frames. ], tot_loss[loss=0.2063, ctc_loss=0.1358, cr_loss=0.3525, over 3379244.25 frames. ], batch size: 44, lr: 4.98e-03, grad_scale: 16.0 2024-09-24 05:17:42,905 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=428675.3333333333, ans=0.125 2024-09-24 05:18:14,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=428768.6666666667, ans=0.2 2024-09-24 05:18:29,411 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=428815.3333333333, ans=0.125 2024-09-24 05:19:01,240 INFO [train.py:1198] (2/4) Epoch 24, batch 2300, loss[loss=0.2077, ctc_loss=0.1379, cr_loss=0.349, over 17008.00 frames. ], tot_loss[loss=0.207, ctc_loss=0.1362, cr_loss=0.3537, over 3384940.68 frames. ], batch size: 44, lr: 4.98e-03, grad_scale: 16.0 2024-09-24 05:19:02,205 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=8.58 vs. limit=15.0 2024-09-24 05:19:11,767 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.244e+02 1.345e+02 1.429e+02 2.009e+02, threshold=2.690e+02, percent-clipped=0.0 2024-09-24 05:19:29,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=428955.3333333333, ans=0.0 2024-09-24 05:20:14,819 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=429095.3333333333, ans=0.125 2024-09-24 05:20:28,878 INFO [train.py:1198] (2/4) Epoch 24, batch 2350, loss[loss=0.2122, ctc_loss=0.1375, cr_loss=0.3737, over 17031.00 frames. ], tot_loss[loss=0.207, ctc_loss=0.1363, cr_loss=0.3537, over 3377046.01 frames. ], batch size: 52, lr: 4.98e-03, grad_scale: 16.0 2024-09-24 05:20:40,814 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=429142.0, ans=0.125 2024-09-24 05:20:53,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=429188.6666666667, ans=0.1 2024-09-24 05:21:03,320 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=429235.3333333333, ans=0.125 2024-09-24 05:21:53,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=429375.3333333333, ans=0.1 2024-09-24 05:21:54,430 INFO [train.py:1198] (2/4) Epoch 24, batch 2400, loss[loss=0.1992, ctc_loss=0.1333, cr_loss=0.3292, over 17209.00 frames. ], tot_loss[loss=0.2075, ctc_loss=0.1367, cr_loss=0.354, over 3374025.73 frames. ], batch size: 50, lr: 4.97e-03, grad_scale: 32.0 2024-09-24 05:22:02,373 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.270e+02 1.408e+02 1.564e+02 2.432e+02, threshold=2.816e+02, percent-clipped=0.0 2024-09-24 05:22:09,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=429422.0, ans=0.125 2024-09-24 05:22:17,012 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=429422.0, ans=0.1 2024-09-24 05:22:57,227 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=429562.0, ans=0.125 2024-09-24 05:23:13,377 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=429608.6666666667, ans=0.125 2024-09-24 05:23:14,640 INFO [train.py:1198] (2/4) Epoch 24, batch 2450, loss[loss=0.1819, ctc_loss=0.1201, cr_loss=0.309, over 17166.00 frames. ], tot_loss[loss=0.2074, ctc_loss=0.1365, cr_loss=0.3543, over 3373651.51 frames. ], batch size: 41, lr: 4.97e-03, grad_scale: 32.0 2024-09-24 05:23:27,937 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=429608.6666666667, ans=0.0 2024-09-24 05:23:29,416 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=429655.3333333333, ans=0.07 2024-09-24 05:23:40,048 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module1.whiten, num_groups=1, num_channels=192, metric=8.26 vs. limit=15.0 2024-09-24 05:23:55,666 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=5.09 vs. limit=10.0 2024-09-24 05:24:14,392 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=11.39 vs. limit=15.0 2024-09-24 05:24:34,551 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=429795.3333333333, ans=0.1 2024-09-24 05:24:37,351 INFO [train.py:1198] (2/4) Epoch 24, batch 2500, loss[loss=0.2187, ctc_loss=0.1424, cr_loss=0.3814, over 17065.00 frames. ], tot_loss[loss=0.2074, ctc_loss=0.1365, cr_loss=0.3545, over 3370072.67 frames. ], batch size: 46, lr: 4.97e-03, grad_scale: 32.0 2024-09-24 05:24:39,509 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 05:24:48,051 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.078e+02 1.250e+02 1.362e+02 1.457e+02 2.366e+02, threshold=2.724e+02, percent-clipped=0.0 2024-09-24 05:24:52,701 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=429842.0, ans=0.125 2024-09-24 05:25:11,250 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=5.10 vs. limit=10.0 2024-09-24 05:25:12,215 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.83 vs. limit=10.0 2024-09-24 05:25:21,902 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=429935.3333333333, ans=0.04949747468305833 2024-09-24 05:25:31,526 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 05:25:34,559 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=429982.0, ans=0.125 2024-09-24 05:26:02,839 INFO [train.py:1198] (2/4) Epoch 24, batch 2550, loss[loss=0.2311, ctc_loss=0.1562, cr_loss=0.3744, over 16272.00 frames. ], tot_loss[loss=0.2076, ctc_loss=0.1367, cr_loss=0.3543, over 3360169.50 frames. ], batch size: 74, lr: 4.97e-03, grad_scale: 32.0 2024-09-24 05:26:04,786 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=430075.3333333333, ans=0.0 2024-09-24 05:26:06,781 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.15 vs. limit=15.0 2024-09-24 05:26:15,819 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=430075.3333333333, ans=0.1 2024-09-24 05:26:33,310 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=430168.6666666667, ans=0.0 2024-09-24 05:26:55,642 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=430215.3333333333, ans=0.0 2024-09-24 05:27:05,163 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=430215.3333333333, ans=0.125 2024-09-24 05:27:22,967 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=430262.0, ans=0.125 2024-09-24 05:27:25,867 INFO [train.py:1198] (2/4) Epoch 24, batch 2600, loss[loss=0.1831, ctc_loss=0.1197, cr_loss=0.3172, over 17099.00 frames. ], tot_loss[loss=0.2072, ctc_loss=0.1364, cr_loss=0.3538, over 3360701.33 frames. ], batch size: 40, lr: 4.97e-03, grad_scale: 32.0 2024-09-24 05:27:33,811 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.253e+02 1.345e+02 1.495e+02 2.149e+02, threshold=2.690e+02, percent-clipped=0.0 2024-09-24 05:27:34,195 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=430308.6666666667, ans=0.025 2024-09-24 05:27:48,678 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=430355.3333333333, ans=0.1 2024-09-24 05:27:54,819 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=430355.3333333333, ans=0.2 2024-09-24 05:28:04,604 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.56 vs. limit=15.0 2024-09-24 05:28:05,808 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=430402.0, ans=0.125 2024-09-24 05:28:08,943 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=430402.0, ans=0.07 2024-09-24 05:28:31,154 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=430495.3333333333, ans=0.07 2024-09-24 05:28:34,207 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=430495.3333333333, ans=0.125 2024-09-24 05:28:36,558 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=3.66 vs. limit=15.0 2024-09-24 05:28:44,975 INFO [train.py:1198] (2/4) Epoch 24, batch 2650, loss[loss=0.1901, ctc_loss=0.1233, cr_loss=0.3339, over 17226.00 frames. ], tot_loss[loss=0.2081, ctc_loss=0.1373, cr_loss=0.3544, over 3348763.98 frames. ], batch size: 47, lr: 4.97e-03, grad_scale: 32.0 2024-09-24 05:29:03,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=430588.6666666667, ans=0.125 2024-09-24 05:29:18,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=430635.3333333333, ans=0.0 2024-09-24 05:29:29,378 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=430635.3333333333, ans=0.2 2024-09-24 05:29:52,156 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=430682.0, ans=0.0 2024-09-24 05:29:52,219 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=430682.0, ans=0.125 2024-09-24 05:30:09,758 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=430728.6666666667, ans=0.125 2024-09-24 05:30:12,482 INFO [train.py:1198] (2/4) Epoch 24, batch 2700, loss[loss=0.2268, ctc_loss=0.1497, cr_loss=0.3857, over 17032.00 frames. ], tot_loss[loss=0.2082, ctc_loss=0.1373, cr_loss=0.3542, over 3352343.63 frames. ], batch size: 56, lr: 4.97e-03, grad_scale: 32.0 2024-09-24 05:30:15,946 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=430775.3333333333, ans=0.125 2024-09-24 05:30:20,466 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.268e+02 1.356e+02 1.521e+02 3.128e+02, threshold=2.712e+02, percent-clipped=1.0 2024-09-24 05:30:22,425 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=430775.3333333333, ans=0.125 2024-09-24 05:31:00,343 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=430915.3333333333, ans=0.025 2024-09-24 05:31:06,823 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=430915.3333333333, ans=0.1 2024-09-24 05:31:14,977 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=430962.0, ans=0.125 2024-09-24 05:31:32,249 INFO [train.py:1198] (2/4) Epoch 24, batch 2750, loss[loss=0.2122, ctc_loss=0.1417, cr_loss=0.3524, over 17041.00 frames. ], tot_loss[loss=0.209, ctc_loss=0.138, cr_loss=0.3548, over 3349546.20 frames. ], batch size: 56, lr: 4.97e-03, grad_scale: 32.0 2024-09-24 05:32:00,272 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.38 vs. limit=15.0 2024-09-24 05:32:02,966 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=8.01 vs. limit=15.0 2024-09-24 05:32:22,545 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=8.22 vs. limit=15.0 2024-09-24 05:32:54,275 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=431242.0, ans=0.125 2024-09-24 05:32:55,668 INFO [train.py:1198] (2/4) Epoch 24, batch 2800, loss[loss=0.2141, ctc_loss=0.1448, cr_loss=0.3467, over 17226.00 frames. ], tot_loss[loss=0.2089, ctc_loss=0.1381, cr_loss=0.354, over 3353297.57 frames. ], batch size: 47, lr: 4.96e-03, grad_scale: 32.0 2024-09-24 05:33:03,460 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.094e+02 1.249e+02 1.392e+02 1.537e+02 2.011e+02, threshold=2.784e+02, percent-clipped=0.0 2024-09-24 05:33:06,344 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=13.74 vs. limit=22.5 2024-09-24 05:33:34,067 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=431335.3333333333, ans=0.125 2024-09-24 05:33:35,866 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=431335.3333333333, ans=0.125 2024-09-24 05:33:35,870 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=431335.3333333333, ans=0.0 2024-09-24 05:33:45,404 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=431382.0, ans=0.125 2024-09-24 05:33:50,239 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=431382.0, ans=0.1 2024-09-24 05:34:18,046 INFO [train.py:1198] (2/4) Epoch 24, batch 2850, loss[loss=0.2005, ctc_loss=0.1324, cr_loss=0.3406, over 17280.00 frames. ], tot_loss[loss=0.2078, ctc_loss=0.1372, cr_loss=0.3527, over 3356427.64 frames. ], batch size: 46, lr: 4.96e-03, grad_scale: 32.0 2024-09-24 05:34:26,991 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.87 vs. limit=6.0 2024-09-24 05:34:41,774 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.79 vs. limit=12.0 2024-09-24 05:34:55,561 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=431568.6666666667, ans=0.125 2024-09-24 05:35:13,388 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=431615.3333333333, ans=0.0 2024-09-24 05:35:25,195 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.30 vs. limit=15.0 2024-09-24 05:35:34,098 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=431662.0, ans=0.2 2024-09-24 05:35:43,353 INFO [train.py:1198] (2/4) Epoch 24, batch 2900, loss[loss=0.2202, ctc_loss=0.145, cr_loss=0.376, over 16973.00 frames. ], tot_loss[loss=0.2089, ctc_loss=0.1381, cr_loss=0.3541, over 3360010.08 frames. ], batch size: 58, lr: 4.96e-03, grad_scale: 16.0 2024-09-24 05:35:52,964 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.062e+02 1.236e+02 1.333e+02 1.474e+02 3.420e+02, threshold=2.666e+02, percent-clipped=1.0 2024-09-24 05:36:06,917 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.24 vs. limit=22.5 2024-09-24 05:36:11,035 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-24 05:36:12,486 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=431755.3333333333, ans=0.07 2024-09-24 05:36:14,148 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=431802.0, ans=0.125 2024-09-24 05:36:15,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=431802.0, ans=0.125 2024-09-24 05:36:29,213 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.65 vs. limit=15.0 2024-09-24 05:37:06,443 INFO [train.py:1198] (2/4) Epoch 24, batch 2950, loss[loss=0.2033, ctc_loss=0.136, cr_loss=0.3366, over 17022.00 frames. ], tot_loss[loss=0.2078, ctc_loss=0.1372, cr_loss=0.3529, over 3369307.10 frames. ], batch size: 51, lr: 4.96e-03, grad_scale: 16.0 2024-09-24 05:37:10,716 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=15.80 vs. limit=22.5 2024-09-24 05:37:11,636 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=431942.0, ans=0.125 2024-09-24 05:37:13,331 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=431942.0, ans=0.09899494936611666 2024-09-24 05:37:24,479 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=431988.6666666667, ans=0.0 2024-09-24 05:37:29,849 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten.whitening_limit, batch_count=431988.6666666667, ans=15.0 2024-09-24 05:37:55,084 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=432082.0, ans=0.125 2024-09-24 05:38:26,284 INFO [train.py:1198] (2/4) Epoch 24, batch 3000, loss[loss=0.2072, ctc_loss=0.1291, cr_loss=0.3902, over 17154.00 frames. ], tot_loss[loss=0.2083, ctc_loss=0.1375, cr_loss=0.3541, over 3369133.45 frames. ], batch size: 45, lr: 4.96e-03, grad_scale: 16.0 2024-09-24 05:38:26,285 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 05:38:42,487 INFO [train.py:1230] (2/4) Epoch 24, validation: loss=0.03786, ctc_loss=0.03786, cr_loss=8.617e-15, over 944034.00 frames. 2024-09-24 05:38:42,488 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 05:38:51,890 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.040e+02 1.243e+02 1.342e+02 1.455e+02 1.995e+02, threshold=2.684e+02, percent-clipped=0.0 2024-09-24 05:38:53,706 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=432175.3333333333, ans=0.125 2024-09-24 05:40:03,771 INFO [train.py:1198] (2/4) Epoch 24, batch 3050, loss[loss=0.2214, ctc_loss=0.1486, cr_loss=0.364, over 17043.00 frames. ], tot_loss[loss=0.2081, ctc_loss=0.1373, cr_loss=0.3539, over 3363957.80 frames. ], batch size: 52, lr: 4.96e-03, grad_scale: 16.0 2024-09-24 05:40:16,577 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 05:41:26,721 INFO [train.py:1198] (2/4) Epoch 24, batch 3100, loss[loss=0.1823, ctc_loss=0.1187, cr_loss=0.3183, over 17274.00 frames. ], tot_loss[loss=0.2069, ctc_loss=0.1365, cr_loss=0.3523, over 3360206.90 frames. ], batch size: 42, lr: 4.96e-03, grad_scale: 16.0 2024-09-24 05:41:26,981 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=432642.0, ans=0.0 2024-09-24 05:41:35,819 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.243e+02 1.328e+02 1.464e+02 2.073e+02, threshold=2.656e+02, percent-clipped=0.0 2024-09-24 05:41:45,613 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=432688.6666666667, ans=0.2 2024-09-24 05:41:49,576 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.73 vs. limit=6.0 2024-09-24 05:42:01,735 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=18.07 vs. limit=22.5 2024-09-24 05:42:44,646 INFO [train.py:1198] (2/4) Epoch 24, batch 3150, loss[loss=0.1835, ctc_loss=0.1181, cr_loss=0.3271, over 16272.00 frames. ], tot_loss[loss=0.2074, ctc_loss=0.1367, cr_loss=0.3534, over 3365336.28 frames. ], batch size: 36, lr: 4.95e-03, grad_scale: 16.0 2024-09-24 05:42:44,890 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=432875.3333333333, ans=0.07 2024-09-24 05:42:51,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=432875.3333333333, ans=0.125 2024-09-24 05:42:57,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=432875.3333333333, ans=0.0 2024-09-24 05:43:00,456 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=432922.0, ans=0.2 2024-09-24 05:43:06,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=432922.0, ans=0.125 2024-09-24 05:43:24,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=432968.6666666667, ans=0.125 2024-09-24 05:43:55,371 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=433062.0, ans=0.125 2024-09-24 05:44:02,936 INFO [train.py:1198] (2/4) Epoch 24, batch 3200, loss[loss=0.1925, ctc_loss=0.1292, cr_loss=0.3169, over 16979.00 frames. ], tot_loss[loss=0.2079, ctc_loss=0.1371, cr_loss=0.354, over 3354668.66 frames. ], batch size: 42, lr: 4.95e-03, grad_scale: 32.0 2024-09-24 05:44:12,034 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.252e+02 1.364e+02 1.478e+02 2.406e+02, threshold=2.728e+02, percent-clipped=0.0 2024-09-24 05:44:21,528 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=433155.3333333333, ans=0.0 2024-09-24 05:44:52,920 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=433248.6666666667, ans=0.0 2024-09-24 05:45:07,000 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=433295.3333333333, ans=0.2 2024-09-24 05:45:13,006 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=433295.3333333333, ans=0.125 2024-09-24 05:45:20,627 INFO [train.py:1198] (2/4) Epoch 24, batch 3250, loss[loss=0.1976, ctc_loss=0.1291, cr_loss=0.3425, over 17098.00 frames. ], tot_loss[loss=0.2069, ctc_loss=0.1364, cr_loss=0.3528, over 3348885.12 frames. ], batch size: 43, lr: 4.95e-03, grad_scale: 32.0 2024-09-24 05:45:30,286 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=433342.0, ans=0.0 2024-09-24 05:45:41,868 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.50 vs. limit=15.0 2024-09-24 05:45:52,175 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.85 vs. limit=10.0 2024-09-24 05:46:19,450 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.min_abs, batch_count=433482.0, ans=0.5 2024-09-24 05:46:19,726 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=10.73 vs. limit=15.0 2024-09-24 05:46:40,881 INFO [train.py:1198] (2/4) Epoch 24, batch 3300, loss[loss=0.2085, ctc_loss=0.1378, cr_loss=0.3537, over 17144.00 frames. ], tot_loss[loss=0.2065, ctc_loss=0.1361, cr_loss=0.352, over 3342059.17 frames. ], batch size: 48, lr: 4.95e-03, grad_scale: 32.0 2024-09-24 05:46:50,400 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.110e+02 1.267e+02 1.337e+02 1.528e+02 2.027e+02, threshold=2.674e+02, percent-clipped=0.0 2024-09-24 05:47:07,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=433622.0, ans=0.125 2024-09-24 05:47:07,622 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=433622.0, ans=0.0 2024-09-24 05:47:08,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=433622.0, ans=0.0 2024-09-24 05:47:10,705 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=433668.6666666667, ans=0.125 2024-09-24 05:47:29,238 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=433715.3333333333, ans=0.125 2024-09-24 05:47:39,110 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.10 vs. limit=15.0 2024-09-24 05:47:58,602 INFO [train.py:1198] (2/4) Epoch 24, batch 3350, loss[loss=0.1821, ctc_loss=0.119, cr_loss=0.3152, over 17231.00 frames. ], tot_loss[loss=0.2077, ctc_loss=0.137, cr_loss=0.3536, over 3357621.68 frames. ], batch size: 42, lr: 4.95e-03, grad_scale: 32.0 2024-09-24 05:47:59,446 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=5.95 vs. limit=10.0 2024-09-24 05:48:00,856 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.83 vs. limit=6.0 2024-09-24 05:48:17,790 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=433855.3333333333, ans=0.2 2024-09-24 05:48:28,654 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=433902.0, ans=0.125 2024-09-24 05:48:33,468 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=433902.0, ans=0.0 2024-09-24 05:48:34,961 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=433902.0, ans=0.125 2024-09-24 05:48:39,841 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=433902.0, ans=0.0 2024-09-24 05:48:46,321 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.41 vs. limit=15.0 2024-09-24 05:48:50,561 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=433948.6666666667, ans=0.125 2024-09-24 05:49:16,883 INFO [train.py:1198] (2/4) Epoch 24, batch 3400, loss[loss=0.1988, ctc_loss=0.1284, cr_loss=0.3523, over 17355.00 frames. ], tot_loss[loss=0.2082, ctc_loss=0.1374, cr_loss=0.3539, over 3334358.36 frames. ], batch size: 48, lr: 4.95e-03, grad_scale: 32.0 2024-09-24 05:49:26,482 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.071e+02 1.267e+02 1.361e+02 1.503e+02 2.338e+02, threshold=2.721e+02, percent-clipped=0.0 2024-09-24 05:49:34,678 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=434088.6666666667, ans=0.1 2024-09-24 05:50:18,688 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=9.36 vs. limit=12.0 2024-09-24 05:50:19,832 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=434228.6666666667, ans=0.125 2024-09-24 05:50:21,385 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=434228.6666666667, ans=0.125 2024-09-24 05:50:29,345 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=434228.6666666667, ans=0.0 2024-09-24 05:50:36,791 INFO [train.py:1198] (2/4) Epoch 24, batch 3450, loss[loss=0.2065, ctc_loss=0.1366, cr_loss=0.3495, over 17038.00 frames. ], tot_loss[loss=0.2092, ctc_loss=0.1381, cr_loss=0.3554, over 3342554.18 frames. ], batch size: 52, lr: 4.95e-03, grad_scale: 32.0 2024-09-24 05:50:40,644 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=13.06 vs. limit=15.0 2024-09-24 05:50:46,602 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=434275.3333333333, ans=0.2 2024-09-24 05:50:54,768 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=8.92 vs. limit=15.0 2024-09-24 05:51:25,278 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=434368.6666666667, ans=0.125 2024-09-24 05:51:58,559 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=434508.6666666667, ans=0.125 2024-09-24 05:51:58,561 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=434508.6666666667, ans=0.125 2024-09-24 05:51:59,811 INFO [train.py:1198] (2/4) Epoch 24, batch 3500, loss[loss=0.1773, ctc_loss=0.1156, cr_loss=0.3084, over 16969.00 frames. ], tot_loss[loss=0.21, ctc_loss=0.1388, cr_loss=0.3563, over 3342718.99 frames. ], batch size: 42, lr: 4.95e-03, grad_scale: 32.0 2024-09-24 05:52:06,425 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=434508.6666666667, ans=0.125 2024-09-24 05:52:10,825 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.042e+02 1.254e+02 1.358e+02 1.511e+02 3.142e+02, threshold=2.715e+02, percent-clipped=1.0 2024-09-24 05:52:28,203 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=434555.3333333333, ans=0.07 2024-09-24 05:52:48,777 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=434648.6666666667, ans=0.0 2024-09-24 05:52:55,190 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=434648.6666666667, ans=0.125 2024-09-24 05:53:07,003 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=5.53 vs. limit=15.0 2024-09-24 05:53:09,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=434695.3333333333, ans=0.0 2024-09-24 05:53:18,583 INFO [train.py:1198] (2/4) Epoch 24, batch 3550, loss[loss=0.2041, ctc_loss=0.1353, cr_loss=0.3439, over 17064.00 frames. ], tot_loss[loss=0.2094, ctc_loss=0.1382, cr_loss=0.3559, over 3345829.67 frames. ], batch size: 46, lr: 4.94e-03, grad_scale: 16.0 2024-09-24 05:53:28,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=434742.0, ans=0.2 2024-09-24 05:53:34,218 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=434788.6666666667, ans=0.125 2024-09-24 05:54:21,307 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=434928.6666666667, ans=0.125 2024-09-24 05:54:36,764 INFO [train.py:1198] (2/4) Epoch 24, batch 3600, loss[loss=0.1885, ctc_loss=0.1215, cr_loss=0.3352, over 17155.00 frames. ], tot_loss[loss=0.2079, ctc_loss=0.1371, cr_loss=0.3537, over 3351159.71 frames. ], batch size: 45, lr: 4.94e-03, grad_scale: 32.0 2024-09-24 05:54:40,110 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=434975.3333333333, ans=0.2 2024-09-24 05:54:47,631 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.041e+02 1.266e+02 1.361e+02 1.484e+02 1.804e+02, threshold=2.723e+02, percent-clipped=0.0 2024-09-24 05:54:54,019 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=435022.0, ans=0.125 2024-09-24 05:55:15,364 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten.whitening_limit, batch_count=435068.6666666667, ans=15.0 2024-09-24 05:55:19,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=435068.6666666667, ans=0.0 2024-09-24 05:55:57,298 INFO [train.py:1198] (2/4) Epoch 24, batch 3650, loss[loss=0.2164, ctc_loss=0.1413, cr_loss=0.3755, over 17290.00 frames. ], tot_loss[loss=0.2079, ctc_loss=0.1371, cr_loss=0.354, over 3354038.02 frames. ], batch size: 51, lr: 4.94e-03, grad_scale: 32.0 2024-09-24 05:56:21,319 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=435255.3333333333, ans=0.125 2024-09-24 05:56:58,255 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=435348.6666666667, ans=0.2 2024-09-24 05:57:13,255 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.68 vs. limit=6.0 2024-09-24 05:57:16,735 INFO [train.py:1198] (2/4) Epoch 24, batch 3700, loss[loss=0.2029, ctc_loss=0.1316, cr_loss=0.3562, over 17028.00 frames. ], tot_loss[loss=0.207, ctc_loss=0.1364, cr_loss=0.3529, over 3358270.32 frames. ], batch size: 39, lr: 4.94e-03, grad_scale: 16.0 2024-09-24 05:57:22,461 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=24.48 vs. limit=22.5 2024-09-24 05:57:29,292 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.270e+02 1.350e+02 1.462e+02 1.892e+02, threshold=2.701e+02, percent-clipped=0.0 2024-09-24 05:57:37,687 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=13.18 vs. limit=15.0 2024-09-24 05:58:11,953 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=435582.0, ans=0.04949747468305833 2024-09-24 05:58:35,097 INFO [train.py:1198] (2/4) Epoch 24, batch 3750, loss[loss=0.2217, ctc_loss=0.1472, cr_loss=0.3729, over 17264.00 frames. ], tot_loss[loss=0.2067, ctc_loss=0.1362, cr_loss=0.3529, over 3354258.18 frames. ], batch size: 44, lr: 4.94e-03, grad_scale: 16.0 2024-09-24 05:58:42,273 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.13 vs. limit=10.0 2024-09-24 05:58:57,084 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=6.74 vs. limit=15.0 2024-09-24 05:59:05,908 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=435768.6666666667, ans=0.0 2024-09-24 05:59:06,585 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.68 vs. limit=12.0 2024-09-24 05:59:18,618 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=435768.6666666667, ans=0.2 2024-09-24 05:59:37,645 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=6.70 vs. limit=12.0 2024-09-24 05:59:46,892 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=435862.0, ans=0.125 2024-09-24 05:59:54,384 INFO [train.py:1198] (2/4) Epoch 24, batch 3800, loss[loss=0.2357, ctc_loss=0.1634, cr_loss=0.3613, over 11563.00 frames. ], tot_loss[loss=0.2075, ctc_loss=0.1367, cr_loss=0.3538, over 3340941.75 frames. ], batch size: 123, lr: 4.94e-03, grad_scale: 16.0 2024-09-24 05:59:59,720 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=8.61 vs. limit=15.0 2024-09-24 06:00:07,111 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.258e+02 1.341e+02 1.479e+02 2.397e+02, threshold=2.682e+02, percent-clipped=0.0 2024-09-24 06:00:08,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=435955.3333333333, ans=0.125 2024-09-24 06:00:38,208 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=436002.0, ans=0.0 2024-09-24 06:00:39,873 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=436002.0, ans=0.125 2024-09-24 06:00:43,069 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=436048.6666666667, ans=0.0 2024-09-24 06:00:49,425 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=436048.6666666667, ans=0.125 2024-09-24 06:00:57,109 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=436095.3333333333, ans=0.125 2024-09-24 06:01:14,193 INFO [train.py:1198] (2/4) Epoch 24, batch 3850, loss[loss=0.2272, ctc_loss=0.1565, cr_loss=0.3538, over 11869.00 frames. ], tot_loss[loss=0.2087, ctc_loss=0.1377, cr_loss=0.3554, over 3308772.87 frames. ], batch size: 123, lr: 4.94e-03, grad_scale: 16.0 2024-09-24 06:01:14,889 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=15.19 vs. limit=15.0 2024-09-24 06:01:22,274 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=436142.0, ans=0.125 2024-09-24 06:01:28,666 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.66 vs. limit=15.0 2024-09-24 06:01:37,666 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=436188.6666666667, ans=0.0 2024-09-24 06:01:37,726 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=436188.6666666667, ans=0.125 2024-09-24 06:01:39,248 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=436188.6666666667, ans=0.025 2024-09-24 06:01:45,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=436235.3333333333, ans=0.1 2024-09-24 06:01:49,682 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=436235.3333333333, ans=0.125 2024-09-24 06:01:54,782 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=16.25 vs. limit=22.5 2024-09-24 06:02:21,564 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=436328.6666666667, ans=0.2 2024-09-24 06:03:16,543 INFO [train.py:1198] (2/4) Epoch 25, batch 0, loss[loss=0.1742, ctc_loss=0.113, cr_loss=0.3065, over 17185.00 frames. ], tot_loss[loss=0.1742, ctc_loss=0.113, cr_loss=0.3065, over 17185.00 frames. ], batch size: 41, lr: 4.83e-03, grad_scale: 32.0 2024-09-24 06:03:16,544 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 06:03:31,924 INFO [train.py:1230] (2/4) Epoch 25, validation: loss=0.03759, ctc_loss=0.03759, cr_loss=8.067e-15, over 944034.00 frames. 2024-09-24 06:03:31,925 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 06:03:51,061 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.051e+02 1.314e+02 1.430e+02 1.672e+02 2.033e+02, threshold=2.861e+02, percent-clipped=0.0 2024-09-24 06:04:14,938 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=436450.0, ans=0.1 2024-09-24 06:04:19,613 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=436450.0, ans=0.0 2024-09-24 06:04:21,457 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.58 vs. limit=22.5 2024-09-24 06:04:30,737 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=436496.6666666667, ans=0.125 2024-09-24 06:04:45,221 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=436543.3333333333, ans=0.0 2024-09-24 06:04:49,913 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=436543.3333333333, ans=0.1 2024-09-24 06:04:54,467 INFO [train.py:1198] (2/4) Epoch 25, batch 50, loss[loss=0.2023, ctc_loss=0.1331, cr_loss=0.3462, over 17142.00 frames. ], tot_loss[loss=0.2093, ctc_loss=0.1379, cr_loss=0.3571, over 759606.00 frames. ], batch size: 48, lr: 4.83e-03, grad_scale: 32.0 2024-09-24 06:04:59,542 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=436590.0, ans=0.1 2024-09-24 06:05:12,102 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=436636.6666666667, ans=0.025 2024-09-24 06:05:28,070 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=436683.3333333333, ans=0.125 2024-09-24 06:05:42,884 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=4.44 vs. limit=12.0 2024-09-24 06:06:19,638 INFO [train.py:1198] (2/4) Epoch 25, batch 100, loss[loss=0.2201, ctc_loss=0.1468, cr_loss=0.3668, over 17022.00 frames. ], tot_loss[loss=0.2101, ctc_loss=0.1384, cr_loss=0.3583, over 1346714.04 frames. ], batch size: 56, lr: 4.83e-03, grad_scale: 16.0 2024-09-24 06:06:23,068 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=436823.3333333333, ans=0.125 2024-09-24 06:06:30,966 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer_ff3.min_abs, batch_count=436823.3333333333, ans=0.2 2024-09-24 06:06:40,252 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.071e+02 1.292e+02 1.373e+02 1.497e+02 2.148e+02, threshold=2.747e+02, percent-clipped=0.0 2024-09-24 06:06:54,020 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.54 vs. limit=6.0 2024-09-24 06:07:00,471 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.10 vs. limit=15.0 2024-09-24 06:07:42,084 INFO [train.py:1198] (2/4) Epoch 25, batch 150, loss[loss=0.2179, ctc_loss=0.1445, cr_loss=0.3669, over 16840.00 frames. ], tot_loss[loss=0.2089, ctc_loss=0.1377, cr_loss=0.3563, over 1787580.41 frames. ], batch size: 58, lr: 4.83e-03, grad_scale: 16.0 2024-09-24 06:07:55,068 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=437056.6666666667, ans=0.1 2024-09-24 06:08:41,331 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=6.16 vs. limit=15.0 2024-09-24 06:08:47,196 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 06:08:51,919 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=437243.3333333333, ans=0.0 2024-09-24 06:08:55,247 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=437243.3333333333, ans=0.2 2024-09-24 06:09:04,124 INFO [train.py:1198] (2/4) Epoch 25, batch 200, loss[loss=0.192, ctc_loss=0.1232, cr_loss=0.3441, over 16953.00 frames. ], tot_loss[loss=0.2078, ctc_loss=0.1369, cr_loss=0.3541, over 2132929.92 frames. ], batch size: 42, lr: 4.83e-03, grad_scale: 8.0 2024-09-24 06:09:22,775 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.28 vs. limit=10.0 2024-09-24 06:09:26,583 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.083e+02 1.223e+02 1.322e+02 1.442e+02 1.903e+02, threshold=2.645e+02, percent-clipped=0.0 2024-09-24 06:10:02,126 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=437430.0, ans=0.1 2024-09-24 06:10:06,936 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=437476.6666666667, ans=0.1 2024-09-24 06:10:24,168 INFO [train.py:1198] (2/4) Epoch 25, batch 250, loss[loss=0.2238, ctc_loss=0.1451, cr_loss=0.3935, over 17024.00 frames. ], tot_loss[loss=0.2091, ctc_loss=0.1378, cr_loss=0.3562, over 2402653.27 frames. ], batch size: 44, lr: 4.83e-03, grad_scale: 8.0 2024-09-24 06:10:46,655 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=437570.0, ans=0.125 2024-09-24 06:11:12,856 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=437616.6666666667, ans=0.125 2024-09-24 06:11:16,354 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=437663.3333333333, ans=0.1 2024-09-24 06:11:20,130 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.32 vs. limit=6.0 2024-09-24 06:11:24,183 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=437663.3333333333, ans=0.1 2024-09-24 06:11:33,734 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=437710.0, ans=0.0 2024-09-24 06:11:42,221 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=437710.0, ans=0.125 2024-09-24 06:11:49,780 INFO [train.py:1198] (2/4) Epoch 25, batch 300, loss[loss=0.2404, ctc_loss=0.1586, cr_loss=0.409, over 17036.00 frames. ], tot_loss[loss=0.2091, ctc_loss=0.1381, cr_loss=0.3552, over 2607542.79 frames. ], batch size: 52, lr: 4.83e-03, grad_scale: 8.0 2024-09-24 06:11:59,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=437756.6666666667, ans=0.09899494936611666 2024-09-24 06:12:12,131 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.120e+02 1.257e+02 1.339e+02 1.438e+02 1.926e+02, threshold=2.678e+02, percent-clipped=0.0 2024-09-24 06:12:14,102 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=437803.3333333333, ans=0.025 2024-09-24 06:12:26,645 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=437850.0, ans=0.1 2024-09-24 06:12:26,656 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=437850.0, ans=0.125 2024-09-24 06:12:37,723 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 06:12:41,233 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=512, metric=2.65 vs. limit=15.0 2024-09-24 06:13:10,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=437990.0, ans=10.0 2024-09-24 06:13:12,245 INFO [train.py:1198] (2/4) Epoch 25, batch 350, loss[loss=0.1739, ctc_loss=0.1129, cr_loss=0.305, over 17259.00 frames. ], tot_loss[loss=0.2075, ctc_loss=0.1369, cr_loss=0.3531, over 2780451.61 frames. ], batch size: 42, lr: 4.82e-03, grad_scale: 8.0 2024-09-24 06:13:18,828 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=437990.0, ans=0.0 2024-09-24 06:13:21,956 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=437990.0, ans=0.0 2024-09-24 06:13:26,778 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=438036.6666666667, ans=0.125 2024-09-24 06:13:30,083 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=438036.6666666667, ans=0.125 2024-09-24 06:13:36,562 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=438036.6666666667, ans=0.0 2024-09-24 06:13:38,261 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=438036.6666666667, ans=0.09899494936611666 2024-09-24 06:13:55,704 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.17 vs. limit=15.0 2024-09-24 06:13:58,607 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=438083.3333333333, ans=0.0 2024-09-24 06:14:06,524 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=438130.0, ans=0.125 2024-09-24 06:14:08,626 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.44 vs. limit=6.0 2024-09-24 06:14:35,220 INFO [train.py:1198] (2/4) Epoch 25, batch 400, loss[loss=0.2544, ctc_loss=0.1765, cr_loss=0.3895, over 11635.00 frames. ], tot_loss[loss=0.2089, ctc_loss=0.138, cr_loss=0.355, over 2895978.31 frames. ], batch size: 123, lr: 4.82e-03, grad_scale: 16.0 2024-09-24 06:14:43,591 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=438223.3333333333, ans=0.125 2024-09-24 06:14:45,200 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=438223.3333333333, ans=0.0 2024-09-24 06:14:57,568 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.241e+02 1.339e+02 1.521e+02 2.224e+02, threshold=2.677e+02, percent-clipped=0.0 2024-09-24 06:15:01,167 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=438270.0, ans=0.125 2024-09-24 06:15:23,849 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.90 vs. limit=15.0 2024-09-24 06:15:24,835 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=438363.3333333333, ans=0.0 2024-09-24 06:15:41,293 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.42 vs. limit=22.5 2024-09-24 06:15:44,012 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=438410.0, ans=0.1 2024-09-24 06:15:57,604 INFO [train.py:1198] (2/4) Epoch 25, batch 450, loss[loss=0.1637, ctc_loss=0.1065, cr_loss=0.2859, over 17117.00 frames. ], tot_loss[loss=0.2077, ctc_loss=0.137, cr_loss=0.3533, over 2997944.05 frames. ], batch size: 40, lr: 4.82e-03, grad_scale: 16.0 2024-09-24 06:16:29,391 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=438503.3333333333, ans=0.125 2024-09-24 06:16:58,083 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=438596.6666666667, ans=0.2 2024-09-24 06:16:59,750 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=438596.6666666667, ans=0.0 2024-09-24 06:17:06,940 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.68 vs. limit=15.0 2024-09-24 06:17:11,086 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=438643.3333333333, ans=0.0 2024-09-24 06:17:20,718 INFO [train.py:1198] (2/4) Epoch 25, batch 500, loss[loss=0.1993, ctc_loss=0.128, cr_loss=0.3565, over 17173.00 frames. ], tot_loss[loss=0.2073, ctc_loss=0.1366, cr_loss=0.3538, over 3086244.05 frames. ], batch size: 41, lr: 4.82e-03, grad_scale: 16.0 2024-09-24 06:17:29,320 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=438690.0, ans=0.125 2024-09-24 06:17:37,487 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=12.06 vs. limit=15.0 2024-09-24 06:17:46,010 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module1.whiten, num_groups=1, num_channels=192, metric=8.23 vs. limit=15.0 2024-09-24 06:17:46,420 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.265e+02 1.364e+02 1.484e+02 2.816e+02, threshold=2.728e+02, percent-clipped=1.0 2024-09-24 06:17:53,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=438736.6666666667, ans=0.0 2024-09-24 06:18:06,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=438783.3333333333, ans=0.0 2024-09-24 06:18:40,824 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.96 vs. limit=6.0 2024-09-24 06:18:44,381 INFO [train.py:1198] (2/4) Epoch 25, batch 550, loss[loss=0.2047, ctc_loss=0.138, cr_loss=0.3332, over 16912.00 frames. ], tot_loss[loss=0.2073, ctc_loss=0.1366, cr_loss=0.3535, over 3145032.39 frames. ], batch size: 58, lr: 4.82e-03, grad_scale: 16.0 2024-09-24 06:18:58,842 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=438923.3333333333, ans=0.0 2024-09-24 06:19:12,797 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=438970.0, ans=0.025 2024-09-24 06:19:29,011 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=439016.6666666667, ans=0.07 2024-09-24 06:19:38,834 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=15.00 vs. limit=15.0 2024-09-24 06:20:04,082 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=439110.0, ans=0.1 2024-09-24 06:20:06,958 INFO [train.py:1198] (2/4) Epoch 25, batch 600, loss[loss=0.2197, ctc_loss=0.1452, cr_loss=0.3727, over 16923.00 frames. ], tot_loss[loss=0.2067, ctc_loss=0.1361, cr_loss=0.3529, over 3188273.03 frames. ], batch size: 58, lr: 4.82e-03, grad_scale: 16.0 2024-09-24 06:20:14,578 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module1.whiten, num_groups=1, num_channels=192, metric=8.38 vs. limit=15.0 2024-09-24 06:20:29,579 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.069e+02 1.242e+02 1.338e+02 1.453e+02 1.774e+02, threshold=2.676e+02, percent-clipped=0.0 2024-09-24 06:20:29,978 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=439203.3333333333, ans=0.125 2024-09-24 06:20:33,010 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=439203.3333333333, ans=0.0 2024-09-24 06:20:41,268 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=439250.0, ans=0.125 2024-09-24 06:20:42,818 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 06:21:12,256 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=439296.6666666667, ans=0.1 2024-09-24 06:21:18,715 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=439343.3333333333, ans=0.0 2024-09-24 06:21:32,780 INFO [train.py:1198] (2/4) Epoch 25, batch 650, loss[loss=0.1788, ctc_loss=0.1153, cr_loss=0.3176, over 16963.00 frames. ], tot_loss[loss=0.2072, ctc_loss=0.1365, cr_loss=0.3536, over 3221291.39 frames. ], batch size: 42, lr: 4.82e-03, grad_scale: 16.0 2024-09-24 06:21:33,842 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.00 vs. limit=15.0 2024-09-24 06:21:45,948 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=11.63 vs. limit=15.0 2024-09-24 06:21:50,417 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=439436.6666666667, ans=0.1 2024-09-24 06:21:55,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=439436.6666666667, ans=0.125 2024-09-24 06:22:28,136 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=439530.0, ans=0.125 2024-09-24 06:22:45,383 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=439576.6666666667, ans=0.0 2024-09-24 06:22:47,832 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.29 vs. limit=15.0 2024-09-24 06:22:54,950 INFO [train.py:1198] (2/4) Epoch 25, batch 700, loss[loss=0.1859, ctc_loss=0.1254, cr_loss=0.3029, over 17199.00 frames. ], tot_loss[loss=0.2058, ctc_loss=0.1355, cr_loss=0.3515, over 3262284.01 frames. ], batch size: 41, lr: 4.82e-03, grad_scale: 16.0 2024-09-24 06:23:17,410 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.239e+02 1.345e+02 1.493e+02 2.005e+02, threshold=2.689e+02, percent-clipped=0.0 2024-09-24 06:23:24,475 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.95 vs. limit=15.0 2024-09-24 06:23:25,584 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=439716.6666666667, ans=0.0 2024-09-24 06:23:25,601 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=439716.6666666667, ans=0.025 2024-09-24 06:23:29,688 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=9.30 vs. limit=10.0 2024-09-24 06:23:34,335 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=7.07 vs. limit=15.0 2024-09-24 06:23:35,768 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.12 vs. limit=15.0 2024-09-24 06:23:55,415 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.min_abs, batch_count=439763.3333333333, ans=0.5 2024-09-24 06:24:00,137 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=439810.0, ans=0.025 2024-09-24 06:24:01,675 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=439810.0, ans=0.04949747468305833 2024-09-24 06:24:17,399 INFO [train.py:1198] (2/4) Epoch 25, batch 750, loss[loss=0.177, ctc_loss=0.1172, cr_loss=0.2991, over 17155.00 frames. ], tot_loss[loss=0.2059, ctc_loss=0.1356, cr_loss=0.3514, over 3287562.60 frames. ], batch size: 45, lr: 4.81e-03, grad_scale: 16.0 2024-09-24 06:24:25,645 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=439856.6666666667, ans=0.05 2024-09-24 06:24:51,345 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=439950.0, ans=0.2 2024-09-24 06:24:59,269 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=439950.0, ans=0.125 2024-09-24 06:25:00,953 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=439950.0, ans=0.0 2024-09-24 06:25:28,155 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=440043.3333333333, ans=0.0 2024-09-24 06:25:28,744 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=10.07 vs. limit=15.0 2024-09-24 06:25:37,712 INFO [train.py:1198] (2/4) Epoch 25, batch 800, loss[loss=0.2136, ctc_loss=0.1396, cr_loss=0.3702, over 17056.00 frames. ], tot_loss[loss=0.2051, ctc_loss=0.135, cr_loss=0.3504, over 3306132.02 frames. ], batch size: 39, lr: 4.81e-03, grad_scale: 32.0 2024-09-24 06:25:49,203 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=440090.0, ans=0.07 2024-09-24 06:25:49,305 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=440090.0, ans=0.125 2024-09-24 06:26:02,614 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.255e+02 1.320e+02 1.414e+02 2.395e+02, threshold=2.641e+02, percent-clipped=0.0 2024-09-24 06:26:04,535 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=440136.6666666667, ans=0.125 2024-09-24 06:26:18,532 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=440183.3333333333, ans=0.0 2024-09-24 06:26:44,210 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=440230.0, ans=0.125 2024-09-24 06:26:46,247 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.34 vs. limit=15.0 2024-09-24 06:26:47,682 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=10.36 vs. limit=22.5 2024-09-24 06:26:53,700 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=440276.6666666667, ans=0.125 2024-09-24 06:27:03,008 INFO [train.py:1198] (2/4) Epoch 25, batch 850, loss[loss=0.1911, ctc_loss=0.1187, cr_loss=0.3619, over 17194.00 frames. ], tot_loss[loss=0.2057, ctc_loss=0.1355, cr_loss=0.3508, over 3305208.21 frames. ], batch size: 41, lr: 4.81e-03, grad_scale: 32.0 2024-09-24 06:27:28,799 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=440370.0, ans=0.125 2024-09-24 06:28:08,255 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=440510.0, ans=0.0 2024-09-24 06:28:10,266 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.68 vs. limit=10.0 2024-09-24 06:28:16,422 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=440510.0, ans=0.0 2024-09-24 06:28:19,865 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 06:28:26,004 INFO [train.py:1198] (2/4) Epoch 25, batch 900, loss[loss=0.1912, ctc_loss=0.1232, cr_loss=0.3403, over 17098.00 frames. ], tot_loss[loss=0.2061, ctc_loss=0.1358, cr_loss=0.3515, over 3321741.30 frames. ], batch size: 43, lr: 4.81e-03, grad_scale: 32.0 2024-09-24 06:28:41,741 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=12.50 vs. limit=22.5 2024-09-24 06:28:50,598 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=5.23 vs. limit=10.0 2024-09-24 06:28:50,994 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.308e+02 1.404e+02 1.529e+02 2.023e+02, threshold=2.807e+02, percent-clipped=0.0 2024-09-24 06:28:53,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=440603.3333333333, ans=0.0 2024-09-24 06:29:47,840 INFO [train.py:1198] (2/4) Epoch 25, batch 950, loss[loss=0.2466, ctc_loss=0.1626, cr_loss=0.42, over 16619.00 frames. ], tot_loss[loss=0.2059, ctc_loss=0.1357, cr_loss=0.3509, over 3323740.59 frames. ], batch size: 66, lr: 4.81e-03, grad_scale: 16.0 2024-09-24 06:30:38,850 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=440930.0, ans=0.125 2024-09-24 06:30:42,146 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=440930.0, ans=0.125 2024-09-24 06:31:06,639 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=440976.6666666667, ans=0.0 2024-09-24 06:31:12,904 INFO [train.py:1198] (2/4) Epoch 25, batch 1000, loss[loss=0.2147, ctc_loss=0.148, cr_loss=0.3338, over 16076.00 frames. ], tot_loss[loss=0.2065, ctc_loss=0.1362, cr_loss=0.3515, over 3328559.06 frames. ], batch size: 74, lr: 4.81e-03, grad_scale: 16.0 2024-09-24 06:31:27,504 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=441070.0, ans=0.0 2024-09-24 06:31:36,534 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.126e+02 1.287e+02 1.404e+02 1.496e+02 1.832e+02, threshold=2.807e+02, percent-clipped=0.0 2024-09-24 06:31:48,353 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=441116.6666666667, ans=0.0 2024-09-24 06:31:52,885 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=441116.6666666667, ans=0.125 2024-09-24 06:32:01,589 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.42 vs. limit=6.0 2024-09-24 06:32:04,312 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=441163.3333333333, ans=0.0 2024-09-24 06:32:18,610 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=441210.0, ans=0.125 2024-09-24 06:32:26,782 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=441210.0, ans=0.0 2024-09-24 06:32:32,883 INFO [train.py:1198] (2/4) Epoch 25, batch 1050, loss[loss=0.2315, ctc_loss=0.1573, cr_loss=0.3713, over 16349.00 frames. ], tot_loss[loss=0.206, ctc_loss=0.1359, cr_loss=0.3504, over 3324971.25 frames. ], batch size: 75, lr: 4.81e-03, grad_scale: 16.0 2024-09-24 06:33:35,250 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=441396.6666666667, ans=0.125 2024-09-24 06:33:55,510 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=441443.3333333333, ans=0.125 2024-09-24 06:33:58,445 INFO [train.py:1198] (2/4) Epoch 25, batch 1100, loss[loss=0.205, ctc_loss=0.1344, cr_loss=0.3529, over 17144.00 frames. ], tot_loss[loss=0.2076, ctc_loss=0.1369, cr_loss=0.3536, over 3330287.12 frames. ], batch size: 48, lr: 4.81e-03, grad_scale: 16.0 2024-09-24 06:34:11,988 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.08 vs. limit=10.0 2024-09-24 06:34:21,373 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=441536.6666666667, ans=0.125 2024-09-24 06:34:22,527 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.115e+02 1.251e+02 1.334e+02 1.435e+02 1.725e+02, threshold=2.668e+02, percent-clipped=0.0 2024-09-24 06:34:31,346 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.08 vs. limit=10.0 2024-09-24 06:34:35,367 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=441583.3333333333, ans=0.125 2024-09-24 06:34:51,390 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 06:35:12,498 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=441676.6666666667, ans=0.07 2024-09-24 06:35:12,817 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.54 vs. limit=15.0 2024-09-24 06:35:18,622 INFO [train.py:1198] (2/4) Epoch 25, batch 1150, loss[loss=0.1896, ctc_loss=0.1242, cr_loss=0.3269, over 17293.00 frames. ], tot_loss[loss=0.2072, ctc_loss=0.1365, cr_loss=0.3533, over 3332996.37 frames. ], batch size: 51, lr: 4.80e-03, grad_scale: 16.0 2024-09-24 06:35:20,522 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=441723.3333333333, ans=0.0 2024-09-24 06:35:33,127 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=441770.0, ans=0.125 2024-09-24 06:35:36,450 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=441770.0, ans=0.0 2024-09-24 06:35:45,992 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=441770.0, ans=0.125 2024-09-24 06:36:00,882 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=441816.6666666667, ans=0.125 2024-09-24 06:36:07,439 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=10.01 vs. limit=15.0 2024-09-24 06:36:08,564 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-24 06:36:42,700 INFO [train.py:1198] (2/4) Epoch 25, batch 1200, loss[loss=0.2046, ctc_loss=0.1352, cr_loss=0.3473, over 17216.00 frames. ], tot_loss[loss=0.2069, ctc_loss=0.1363, cr_loss=0.3531, over 3343799.01 frames. ], batch size: 50, lr: 4.80e-03, grad_scale: 32.0 2024-09-24 06:36:43,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=441956.6666666667, ans=0.1 2024-09-24 06:37:06,722 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.122e+02 1.253e+02 1.340e+02 1.429e+02 1.909e+02, threshold=2.679e+02, percent-clipped=0.0 2024-09-24 06:37:30,285 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=7.16 vs. limit=15.0 2024-09-24 06:37:36,945 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=442096.6666666667, ans=0.0 2024-09-24 06:37:37,059 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=442096.6666666667, ans=0.2 2024-09-24 06:38:05,401 INFO [train.py:1198] (2/4) Epoch 25, batch 1250, loss[loss=0.1863, ctc_loss=0.1209, cr_loss=0.3268, over 17256.00 frames. ], tot_loss[loss=0.2071, ctc_loss=0.1363, cr_loss=0.3536, over 3345727.54 frames. ], batch size: 44, lr: 4.80e-03, grad_scale: 32.0 2024-09-24 06:38:24,617 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=442236.6666666667, ans=0.1 2024-09-24 06:38:43,010 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=442283.3333333333, ans=0.0 2024-09-24 06:38:57,080 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=442330.0, ans=0.0 2024-09-24 06:39:02,076 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=442330.0, ans=0.125 2024-09-24 06:39:18,428 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=442376.6666666667, ans=0.1 2024-09-24 06:39:24,243 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=14.82 vs. limit=22.5 2024-09-24 06:39:27,932 INFO [train.py:1198] (2/4) Epoch 25, batch 1300, loss[loss=0.2011, ctc_loss=0.1316, cr_loss=0.3472, over 17023.00 frames. ], tot_loss[loss=0.2066, ctc_loss=0.136, cr_loss=0.353, over 3347037.72 frames. ], batch size: 44, lr: 4.80e-03, grad_scale: 32.0 2024-09-24 06:39:28,195 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=442423.3333333333, ans=0.125 2024-09-24 06:39:34,520 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 06:39:48,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=442470.0, ans=0.1 2024-09-24 06:39:53,305 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.059e+02 1.277e+02 1.396e+02 1.533e+02 2.196e+02, threshold=2.791e+02, percent-clipped=0.0 2024-09-24 06:40:15,391 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.59 vs. limit=5.0 2024-09-24 06:40:25,543 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=442563.3333333333, ans=0.1 2024-09-24 06:40:26,904 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=442563.3333333333, ans=0.125 2024-09-24 06:40:33,477 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=442610.0, ans=0.025 2024-09-24 06:40:42,938 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=442610.0, ans=0.125 2024-09-24 06:40:43,029 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=442610.0, ans=0.1 2024-09-24 06:40:47,418 INFO [train.py:1198] (2/4) Epoch 25, batch 1350, loss[loss=0.2, ctc_loss=0.1299, cr_loss=0.3503, over 17006.00 frames. ], tot_loss[loss=0.2063, ctc_loss=0.1358, cr_loss=0.3523, over 3351863.21 frames. ], batch size: 51, lr: 4.80e-03, grad_scale: 16.0 2024-09-24 06:41:02,349 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.min_abs, batch_count=442656.6666666667, ans=0.5 2024-09-24 06:41:19,899 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 06:41:31,152 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 06:41:38,969 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=442796.6666666667, ans=0.1 2024-09-24 06:41:50,224 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=442796.6666666667, ans=0.125 2024-09-24 06:42:06,132 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=442843.3333333333, ans=0.125 2024-09-24 06:42:07,693 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=442843.3333333333, ans=0.0 2024-09-24 06:42:12,268 INFO [train.py:1198] (2/4) Epoch 25, batch 1400, loss[loss=0.2367, ctc_loss=0.1599, cr_loss=0.384, over 15992.00 frames. ], tot_loss[loss=0.2061, ctc_loss=0.1357, cr_loss=0.3522, over 3355969.11 frames. ], batch size: 74, lr: 4.80e-03, grad_scale: 16.0 2024-09-24 06:42:30,072 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=442936.6666666667, ans=0.2 2024-09-24 06:42:33,208 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.max_positive, batch_count=442936.6666666667, ans=0.95 2024-09-24 06:42:33,239 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=442936.6666666667, ans=0.025 2024-09-24 06:42:37,797 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=11.67 vs. limit=15.0 2024-09-24 06:42:40,062 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.074e+02 1.265e+02 1.378e+02 1.497e+02 2.360e+02, threshold=2.756e+02, percent-clipped=0.0 2024-09-24 06:42:40,330 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=442936.6666666667, ans=0.0 2024-09-24 06:42:40,350 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=442936.6666666667, ans=0.0 2024-09-24 06:42:43,548 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=442936.6666666667, ans=0.125 2024-09-24 06:42:53,072 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=442983.3333333333, ans=0.1 2024-09-24 06:43:13,825 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=443030.0, ans=0.1 2024-09-24 06:43:29,539 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=443076.6666666667, ans=0.0 2024-09-24 06:43:36,560 INFO [train.py:1198] (2/4) Epoch 25, batch 1450, loss[loss=0.2174, ctc_loss=0.1451, cr_loss=0.3614, over 17031.00 frames. ], tot_loss[loss=0.2055, ctc_loss=0.1352, cr_loss=0.3512, over 3362268.18 frames. ], batch size: 56, lr: 4.80e-03, grad_scale: 16.0 2024-09-24 06:43:36,819 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=443123.3333333333, ans=0.125 2024-09-24 06:44:03,748 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=443170.0, ans=0.04949747468305833 2024-09-24 06:44:15,321 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.93 vs. limit=15.0 2024-09-24 06:44:20,861 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=443216.6666666667, ans=0.0 2024-09-24 06:44:32,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=443263.3333333333, ans=0.04949747468305833 2024-09-24 06:44:55,966 INFO [train.py:1198] (2/4) Epoch 25, batch 1500, loss[loss=0.2244, ctc_loss=0.1479, cr_loss=0.3826, over 17212.00 frames. ], tot_loss[loss=0.2059, ctc_loss=0.1355, cr_loss=0.3522, over 3372664.09 frames. ], batch size: 55, lr: 4.80e-03, grad_scale: 16.0 2024-09-24 06:44:56,318 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=443356.6666666667, ans=0.0 2024-09-24 06:45:21,556 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.070e+02 1.238e+02 1.333e+02 1.437e+02 1.693e+02, threshold=2.665e+02, percent-clipped=0.0 2024-09-24 06:45:37,696 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=443450.0, ans=0.125 2024-09-24 06:45:51,664 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 06:46:08,572 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=443543.3333333333, ans=0.0 2024-09-24 06:46:21,155 INFO [train.py:1198] (2/4) Epoch 25, batch 1550, loss[loss=0.2443, ctc_loss=0.1651, cr_loss=0.3959, over 15112.00 frames. ], tot_loss[loss=0.206, ctc_loss=0.1355, cr_loss=0.3525, over 3369608.47 frames. ], batch size: 89, lr: 4.79e-03, grad_scale: 16.0 2024-09-24 06:46:36,050 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=443636.6666666667, ans=0.0 2024-09-24 06:47:06,479 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=443683.3333333333, ans=0.125 2024-09-24 06:47:14,578 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 06:47:29,798 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=443776.6666666667, ans=0.2 2024-09-24 06:47:43,800 INFO [train.py:1198] (2/4) Epoch 25, batch 1600, loss[loss=0.1907, ctc_loss=0.1232, cr_loss=0.3374, over 17007.00 frames. ], tot_loss[loss=0.2052, ctc_loss=0.135, cr_loss=0.3509, over 3361905.21 frames. ], batch size: 44, lr: 4.79e-03, grad_scale: 32.0 2024-09-24 06:47:45,808 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=443823.3333333333, ans=0.1 2024-09-24 06:48:03,390 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=443870.0, ans=0.125 2024-09-24 06:48:03,513 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=443870.0, ans=0.125 2024-09-24 06:48:09,684 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.061e+02 1.242e+02 1.313e+02 1.472e+02 2.185e+02, threshold=2.626e+02, percent-clipped=0.0 2024-09-24 06:48:16,454 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 06:49:06,444 INFO [train.py:1198] (2/4) Epoch 25, batch 1650, loss[loss=0.2235, ctc_loss=0.1477, cr_loss=0.3787, over 17138.00 frames. ], tot_loss[loss=0.205, ctc_loss=0.1349, cr_loss=0.3507, over 3361137.06 frames. ], batch size: 48, lr: 4.79e-03, grad_scale: 32.0 2024-09-24 06:49:13,023 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=444056.6666666667, ans=0.0 2024-09-24 06:50:01,063 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.max_abs, batch_count=444196.6666666667, ans=10.0 2024-09-24 06:50:13,646 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=444243.3333333333, ans=0.0 2024-09-24 06:50:25,201 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=14.23 vs. limit=22.5 2024-09-24 06:50:26,221 INFO [train.py:1198] (2/4) Epoch 25, batch 1700, loss[loss=0.2014, ctc_loss=0.1348, cr_loss=0.3329, over 17000.00 frames. ], tot_loss[loss=0.2047, ctc_loss=0.1346, cr_loss=0.3504, over 3364316.93 frames. ], batch size: 53, lr: 4.79e-03, grad_scale: 32.0 2024-09-24 06:50:39,765 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=4.45 vs. limit=12.0 2024-09-24 06:50:45,514 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=444336.6666666667, ans=0.2 2024-09-24 06:50:49,404 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=11.34 vs. limit=22.5 2024-09-24 06:50:54,228 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.301e+02 1.388e+02 1.513e+02 1.890e+02, threshold=2.776e+02, percent-clipped=0.0 2024-09-24 06:50:54,513 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=444336.6666666667, ans=0.125 2024-09-24 06:51:27,535 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=444430.0, ans=0.025 2024-09-24 06:51:51,033 INFO [train.py:1198] (2/4) Epoch 25, batch 1750, loss[loss=0.2019, ctc_loss=0.1312, cr_loss=0.3539, over 17217.00 frames. ], tot_loss[loss=0.2051, ctc_loss=0.1349, cr_loss=0.3509, over 3360987.93 frames. ], batch size: 47, lr: 4.79e-03, grad_scale: 32.0 2024-09-24 06:52:20,197 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.43 vs. limit=15.0 2024-09-24 06:52:36,636 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=444616.6666666667, ans=0.125 2024-09-24 06:52:39,524 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=444663.3333333333, ans=0.2 2024-09-24 06:52:42,778 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=444663.3333333333, ans=0.125 2024-09-24 06:52:44,287 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=444663.3333333333, ans=0.025 2024-09-24 06:52:49,355 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=444663.3333333333, ans=0.125 2024-09-24 06:53:12,786 INFO [train.py:1198] (2/4) Epoch 25, batch 1800, loss[loss=0.1955, ctc_loss=0.1274, cr_loss=0.3406, over 17292.00 frames. ], tot_loss[loss=0.2055, ctc_loss=0.1351, cr_loss=0.3519, over 3368228.81 frames. ], batch size: 49, lr: 4.79e-03, grad_scale: 32.0 2024-09-24 06:53:26,854 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=444756.6666666667, ans=0.1 2024-09-24 06:53:40,996 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.233e+02 1.324e+02 1.428e+02 1.805e+02, threshold=2.648e+02, percent-clipped=0.0 2024-09-24 06:54:10,080 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=444896.6666666667, ans=0.0 2024-09-24 06:54:11,994 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.96 vs. limit=15.0 2024-09-24 06:54:19,525 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=444943.3333333333, ans=0.125 2024-09-24 06:54:19,526 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=444943.3333333333, ans=0.0 2024-09-24 06:54:30,975 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=444943.3333333333, ans=0.1 2024-09-24 06:54:32,524 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=444943.3333333333, ans=0.1 2024-09-24 06:54:35,385 INFO [train.py:1198] (2/4) Epoch 25, batch 1850, loss[loss=0.2203, ctc_loss=0.1457, cr_loss=0.3732, over 16637.00 frames. ], tot_loss[loss=0.2059, ctc_loss=0.1355, cr_loss=0.3524, over 3370553.09 frames. ], batch size: 66, lr: 4.79e-03, grad_scale: 32.0 2024-09-24 06:55:15,833 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=445083.3333333333, ans=0.1 2024-09-24 06:55:15,925 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.min_positive, batch_count=445083.3333333333, ans=0.05 2024-09-24 06:55:46,483 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.max_abs, batch_count=445176.6666666667, ans=10.0 2024-09-24 06:56:00,122 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=12.87 vs. limit=22.5 2024-09-24 06:56:01,115 INFO [train.py:1198] (2/4) Epoch 25, batch 1900, loss[loss=0.1852, ctc_loss=0.1198, cr_loss=0.3271, over 17264.00 frames. ], tot_loss[loss=0.2067, ctc_loss=0.136, cr_loss=0.3532, over 3366842.06 frames. ], batch size: 44, lr: 4.79e-03, grad_scale: 32.0 2024-09-24 06:56:06,139 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=445223.3333333333, ans=0.125 2024-09-24 06:56:26,825 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.079e+02 1.232e+02 1.312e+02 1.425e+02 2.956e+02, threshold=2.624e+02, percent-clipped=1.0 2024-09-24 06:56:27,442 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=8.70 vs. limit=12.0 2024-09-24 06:56:34,271 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.46 vs. limit=15.0 2024-09-24 06:56:34,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=445316.6666666667, ans=0.125 2024-09-24 06:56:54,065 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=445363.3333333333, ans=0.0 2024-09-24 06:56:55,704 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 06:57:00,372 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=445363.3333333333, ans=0.0 2024-09-24 06:57:16,414 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=445410.0, ans=0.1 2024-09-24 06:57:20,853 INFO [train.py:1198] (2/4) Epoch 25, batch 1950, loss[loss=0.2163, ctc_loss=0.1444, cr_loss=0.3592, over 17250.00 frames. ], tot_loss[loss=0.2074, ctc_loss=0.1365, cr_loss=0.3546, over 3369852.58 frames. ], batch size: 44, lr: 4.78e-03, grad_scale: 32.0 2024-09-24 06:57:24,415 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=445456.6666666667, ans=0.1 2024-09-24 06:57:29,323 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=445456.6666666667, ans=0.125 2024-09-24 06:57:52,707 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=445503.3333333333, ans=0.1 2024-09-24 06:58:00,487 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=445550.0, ans=0.0 2024-09-24 06:58:04,426 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=19.39 vs. limit=22.5 2024-09-24 06:58:28,808 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=445643.3333333333, ans=0.2 2024-09-24 06:58:38,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=445643.3333333333, ans=0.0 2024-09-24 06:58:46,269 INFO [train.py:1198] (2/4) Epoch 25, batch 2000, loss[loss=0.1791, ctc_loss=0.1139, cr_loss=0.3259, over 17265.00 frames. ], tot_loss[loss=0.2071, ctc_loss=0.1362, cr_loss=0.3546, over 3377205.17 frames. ], batch size: 42, lr: 4.78e-03, grad_scale: 32.0 2024-09-24 06:59:11,910 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.085e+02 1.272e+02 1.363e+02 1.482e+02 1.870e+02, threshold=2.725e+02, percent-clipped=0.0 2024-09-24 06:59:58,390 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=445876.6666666667, ans=0.2 2024-09-24 07:00:06,178 INFO [train.py:1198] (2/4) Epoch 25, batch 2050, loss[loss=0.2168, ctc_loss=0.1416, cr_loss=0.3763, over 17218.00 frames. ], tot_loss[loss=0.2077, ctc_loss=0.1367, cr_loss=0.3548, over 3366109.60 frames. ], batch size: 55, lr: 4.78e-03, grad_scale: 32.0 2024-09-24 07:00:11,284 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=445923.3333333333, ans=0.1 2024-09-24 07:00:19,275 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=445923.3333333333, ans=0.2 2024-09-24 07:00:21,583 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.60 vs. limit=15.0 2024-09-24 07:00:43,585 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=446016.6666666667, ans=10.0 2024-09-24 07:01:09,205 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=446063.3333333333, ans=0.0 2024-09-24 07:01:31,629 INFO [train.py:1198] (2/4) Epoch 25, batch 2100, loss[loss=0.2008, ctc_loss=0.1317, cr_loss=0.3457, over 17363.00 frames. ], tot_loss[loss=0.2073, ctc_loss=0.1364, cr_loss=0.3543, over 3367594.92 frames. ], batch size: 48, lr: 4.78e-03, grad_scale: 32.0 2024-09-24 07:01:56,975 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.073e+02 1.289e+02 1.372e+02 1.523e+02 2.426e+02, threshold=2.744e+02, percent-clipped=0.0 2024-09-24 07:01:58,935 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=446203.3333333333, ans=0.2 2024-09-24 07:02:03,481 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=446250.0, ans=0.125 2024-09-24 07:02:10,410 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=13.77 vs. limit=22.5 2024-09-24 07:02:16,492 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=446250.0, ans=0.1 2024-09-24 07:02:28,905 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=446296.6666666667, ans=0.2 2024-09-24 07:02:32,690 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.21 vs. limit=15.0 2024-09-24 07:02:45,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=446343.3333333333, ans=0.0 2024-09-24 07:02:54,490 INFO [train.py:1198] (2/4) Epoch 25, batch 2150, loss[loss=0.1923, ctc_loss=0.127, cr_loss=0.3264, over 17059.00 frames. ], tot_loss[loss=0.2073, ctc_loss=0.1365, cr_loss=0.3539, over 3362630.96 frames. ], batch size: 46, lr: 4.78e-03, grad_scale: 32.0 2024-09-24 07:02:59,516 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=446390.0, ans=0.0 2024-09-24 07:03:37,535 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=446483.3333333333, ans=0.0 2024-09-24 07:04:00,154 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=446576.6666666667, ans=0.0 2024-09-24 07:04:08,245 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=446576.6666666667, ans=0.125 2024-09-24 07:04:11,428 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=446576.6666666667, ans=0.0 2024-09-24 07:04:17,541 INFO [train.py:1198] (2/4) Epoch 25, batch 2200, loss[loss=0.1859, ctc_loss=0.1199, cr_loss=0.3304, over 17306.00 frames. ], tot_loss[loss=0.207, ctc_loss=0.1364, cr_loss=0.3533, over 3368402.68 frames. ], batch size: 51, lr: 4.78e-03, grad_scale: 16.0 2024-09-24 07:04:17,956 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=446623.3333333333, ans=0.125 2024-09-24 07:04:27,335 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-24 07:04:28,893 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=446623.3333333333, ans=0.125 2024-09-24 07:04:30,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=446623.3333333333, ans=0.025 2024-09-24 07:04:41,586 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=446670.0, ans=0.125 2024-09-24 07:04:44,641 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.080e+02 1.273e+02 1.353e+02 1.559e+02 2.204e+02, threshold=2.706e+02, percent-clipped=0.0 2024-09-24 07:05:06,107 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=446763.3333333333, ans=0.125 2024-09-24 07:05:15,854 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=446763.3333333333, ans=0.0 2024-09-24 07:05:36,719 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=446856.6666666667, ans=0.125 2024-09-24 07:05:37,840 INFO [train.py:1198] (2/4) Epoch 25, batch 2250, loss[loss=0.2183, ctc_loss=0.1476, cr_loss=0.3535, over 17032.00 frames. ], tot_loss[loss=0.2066, ctc_loss=0.1361, cr_loss=0.3522, over 3372975.87 frames. ], batch size: 56, lr: 4.78e-03, grad_scale: 16.0 2024-09-24 07:06:02,864 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.56 vs. limit=15.0 2024-09-24 07:06:58,667 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=447043.3333333333, ans=0.0 2024-09-24 07:07:03,314 INFO [train.py:1198] (2/4) Epoch 25, batch 2300, loss[loss=0.2531, ctc_loss=0.1739, cr_loss=0.3963, over 15109.00 frames. ], tot_loss[loss=0.2073, ctc_loss=0.1366, cr_loss=0.3532, over 3376044.43 frames. ], batch size: 89, lr: 4.78e-03, grad_scale: 16.0 2024-09-24 07:07:30,511 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.028e+02 1.259e+02 1.322e+02 1.446e+02 1.963e+02, threshold=2.643e+02, percent-clipped=0.0 2024-09-24 07:07:48,065 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1.whitening_limit, batch_count=447183.3333333333, ans=10.0 2024-09-24 07:08:03,733 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=447230.0, ans=0.1 2024-09-24 07:08:13,434 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=447276.6666666667, ans=0.125 2024-09-24 07:08:28,419 INFO [train.py:1198] (2/4) Epoch 25, batch 2350, loss[loss=0.1674, ctc_loss=0.1078, cr_loss=0.2979, over 17208.00 frames. ], tot_loss[loss=0.2078, ctc_loss=0.137, cr_loss=0.3543, over 3370977.95 frames. ], batch size: 47, lr: 4.77e-03, grad_scale: 16.0 2024-09-24 07:08:41,629 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.46 vs. limit=22.5 2024-09-24 07:08:44,730 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=12.34 vs. limit=12.0 2024-09-24 07:08:52,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=447370.0, ans=0.125 2024-09-24 07:08:55,450 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=447370.0, ans=0.125 2024-09-24 07:09:07,703 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=9.76 vs. limit=15.0 2024-09-24 07:09:13,644 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=4.17 vs. limit=15.0 2024-09-24 07:09:42,152 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=9.72 vs. limit=15.0 2024-09-24 07:09:47,656 INFO [train.py:1198] (2/4) Epoch 25, batch 2400, loss[loss=0.1937, ctc_loss=0.123, cr_loss=0.3536, over 17111.00 frames. ], tot_loss[loss=0.2077, ctc_loss=0.1369, cr_loss=0.354, over 3365965.17 frames. ], batch size: 40, lr: 4.77e-03, grad_scale: 32.0 2024-09-24 07:09:54,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=447556.6666666667, ans=0.1 2024-09-24 07:10:14,557 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.299e+02 1.395e+02 1.509e+02 2.801e+02, threshold=2.791e+02, percent-clipped=1.0 2024-09-24 07:10:22,882 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=447650.0, ans=0.1 2024-09-24 07:10:22,954 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=447650.0, ans=0.125 2024-09-24 07:11:06,551 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=447743.3333333333, ans=0.1 2024-09-24 07:11:12,576 INFO [train.py:1198] (2/4) Epoch 25, batch 2450, loss[loss=0.1871, ctc_loss=0.1245, cr_loss=0.3128, over 15983.00 frames. ], tot_loss[loss=0.2072, ctc_loss=0.1365, cr_loss=0.3533, over 3355069.67 frames. ], batch size: 74, lr: 4.77e-03, grad_scale: 32.0 2024-09-24 07:11:17,992 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten.whitening_limit, batch_count=447790.0, ans=15.0 2024-09-24 07:11:28,898 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=447836.6666666667, ans=0.1 2024-09-24 07:11:41,962 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=447836.6666666667, ans=0.125 2024-09-24 07:12:12,320 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=447930.0, ans=0.125 2024-09-24 07:12:13,848 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=447930.0, ans=0.1 2024-09-24 07:12:15,386 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=447976.6666666667, ans=0.0 2024-09-24 07:12:37,501 INFO [train.py:1198] (2/4) Epoch 25, batch 2500, loss[loss=0.2315, ctc_loss=0.1558, cr_loss=0.3788, over 15953.00 frames. ], tot_loss[loss=0.2074, ctc_loss=0.1366, cr_loss=0.3539, over 3358486.17 frames. ], batch size: 74, lr: 4.77e-03, grad_scale: 32.0 2024-09-24 07:13:00,763 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.84 vs. limit=22.5 2024-09-24 07:13:01,807 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=448070.0, ans=0.2 2024-09-24 07:13:04,742 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.089e+02 1.243e+02 1.319e+02 1.410e+02 2.410e+02, threshold=2.638e+02, percent-clipped=0.0 2024-09-24 07:13:11,431 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=448116.6666666667, ans=0.2 2024-09-24 07:13:28,068 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=448163.3333333333, ans=0.1 2024-09-24 07:13:31,263 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=448163.3333333333, ans=0.04949747468305833 2024-09-24 07:13:45,332 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.min_positive, batch_count=448210.0, ans=0.05 2024-09-24 07:13:59,333 INFO [train.py:1198] (2/4) Epoch 25, batch 2550, loss[loss=0.2226, ctc_loss=0.1456, cr_loss=0.3848, over 17086.00 frames. ], tot_loss[loss=0.207, ctc_loss=0.1364, cr_loss=0.3529, over 3354816.32 frames. ], batch size: 49, lr: 4.77e-03, grad_scale: 32.0 2024-09-24 07:14:23,551 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=448303.3333333333, ans=0.125 2024-09-24 07:14:45,958 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=448396.6666666667, ans=0.025 2024-09-24 07:15:07,531 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=9.26 vs. limit=15.0 2024-09-24 07:15:19,294 INFO [train.py:1198] (2/4) Epoch 25, batch 2600, loss[loss=0.2031, ctc_loss=0.1354, cr_loss=0.3385, over 16709.00 frames. ], tot_loss[loss=0.2075, ctc_loss=0.1367, cr_loss=0.354, over 3358268.73 frames. ], batch size: 61, lr: 4.77e-03, grad_scale: 32.0 2024-09-24 07:15:50,289 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=448536.6666666667, ans=0.125 2024-09-24 07:15:51,574 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.064e+02 1.258e+02 1.378e+02 1.492e+02 2.205e+02, threshold=2.755e+02, percent-clipped=0.0 2024-09-24 07:16:01,939 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=16.37 vs. limit=22.5 2024-09-24 07:16:03,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=448583.3333333333, ans=0.125 2024-09-24 07:16:04,878 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=448583.3333333333, ans=0.1 2024-09-24 07:16:41,543 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=448676.6666666667, ans=0.125 2024-09-24 07:16:44,543 INFO [train.py:1198] (2/4) Epoch 25, batch 2650, loss[loss=0.2106, ctc_loss=0.1399, cr_loss=0.3535, over 17214.00 frames. ], tot_loss[loss=0.2077, ctc_loss=0.1368, cr_loss=0.3544, over 3361595.82 frames. ], batch size: 55, lr: 4.77e-03, grad_scale: 32.0 2024-09-24 07:17:10,656 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=448770.0, ans=0.2 2024-09-24 07:18:06,261 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=448956.6666666667, ans=0.2 2024-09-24 07:18:10,103 INFO [train.py:1198] (2/4) Epoch 25, batch 2700, loss[loss=0.2414, ctc_loss=0.1608, cr_loss=0.4026, over 17033.00 frames. ], tot_loss[loss=0.2063, ctc_loss=0.1357, cr_loss=0.3526, over 3373474.20 frames. ], batch size: 52, lr: 4.77e-03, grad_scale: 32.0 2024-09-24 07:18:37,164 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.035e+02 1.249e+02 1.338e+02 1.409e+02 1.725e+02, threshold=2.676e+02, percent-clipped=0.0 2024-09-24 07:18:51,884 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=449050.0, ans=0.125 2024-09-24 07:19:09,477 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 07:19:29,777 INFO [train.py:1198] (2/4) Epoch 25, batch 2750, loss[loss=0.2406, ctc_loss=0.1644, cr_loss=0.381, over 14759.00 frames. ], tot_loss[loss=0.2061, ctc_loss=0.1358, cr_loss=0.3517, over 3367814.21 frames. ], batch size: 89, lr: 4.76e-03, grad_scale: 16.0 2024-09-24 07:19:42,629 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=449190.0, ans=0.125 2024-09-24 07:19:46,261 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.64 vs. limit=15.0 2024-09-24 07:19:54,019 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=449236.6666666667, ans=0.125 2024-09-24 07:19:55,516 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=449236.6666666667, ans=0.0 2024-09-24 07:20:55,050 INFO [train.py:1198] (2/4) Epoch 25, batch 2800, loss[loss=0.214, ctc_loss=0.1443, cr_loss=0.3486, over 16105.00 frames. ], tot_loss[loss=0.2059, ctc_loss=0.1356, cr_loss=0.3513, over 3366717.08 frames. ], batch size: 74, lr: 4.76e-03, grad_scale: 32.0 2024-09-24 07:21:08,783 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=4.54 vs. limit=15.0 2024-09-24 07:21:14,522 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=449470.0, ans=0.125 2024-09-24 07:21:23,915 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.107e+02 1.252e+02 1.365e+02 1.477e+02 3.816e+02, threshold=2.730e+02, percent-clipped=1.0 2024-09-24 07:22:06,166 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=449610.0, ans=0.0 2024-09-24 07:22:17,833 INFO [train.py:1198] (2/4) Epoch 25, batch 2850, loss[loss=0.2139, ctc_loss=0.1427, cr_loss=0.3561, over 15897.00 frames. ], tot_loss[loss=0.2065, ctc_loss=0.1361, cr_loss=0.3523, over 3365540.35 frames. ], batch size: 74, lr: 4.76e-03, grad_scale: 32.0 2024-09-24 07:22:22,948 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=449656.6666666667, ans=0.2 2024-09-24 07:22:30,944 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=449656.6666666667, ans=0.0 2024-09-24 07:22:32,681 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=449703.3333333333, ans=0.2 2024-09-24 07:23:08,124 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=12.72 vs. limit=22.5 2024-09-24 07:23:26,743 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=14.69 vs. limit=22.5 2024-09-24 07:23:37,886 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=449843.3333333333, ans=0.0 2024-09-24 07:23:40,781 INFO [train.py:1198] (2/4) Epoch 25, batch 2900, loss[loss=0.2253, ctc_loss=0.1491, cr_loss=0.3809, over 17014.00 frames. ], tot_loss[loss=0.2066, ctc_loss=0.1361, cr_loss=0.3522, over 3364027.39 frames. ], batch size: 53, lr: 4.76e-03, grad_scale: 32.0 2024-09-24 07:23:49,026 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=449890.0, ans=0.125 2024-09-24 07:23:49,063 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=449890.0, ans=0.125 2024-09-24 07:23:59,312 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=11.47 vs. limit=15.0 2024-09-24 07:24:08,433 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=449936.6666666667, ans=0.1 2024-09-24 07:24:09,619 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.121e+02 1.270e+02 1.360e+02 1.492e+02 4.410e+02, threshold=2.720e+02, percent-clipped=1.0 2024-09-24 07:24:09,895 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=449936.6666666667, ans=0.2 2024-09-24 07:24:11,611 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=449983.3333333333, ans=0.125 2024-09-24 07:24:29,641 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=3.87 vs. limit=10.0 2024-09-24 07:24:30,764 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=450030.0, ans=0.0 2024-09-24 07:24:35,895 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=15.29 vs. limit=22.5 2024-09-24 07:24:40,236 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=450030.0, ans=10.0 2024-09-24 07:25:00,611 INFO [train.py:1198] (2/4) Epoch 25, batch 2950, loss[loss=0.1879, ctc_loss=0.1195, cr_loss=0.3421, over 17263.00 frames. ], tot_loss[loss=0.2063, ctc_loss=0.1359, cr_loss=0.3517, over 3356128.61 frames. ], batch size: 42, lr: 4.76e-03, grad_scale: 32.0 2024-09-24 07:25:02,275 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=450123.3333333333, ans=0.125 2024-09-24 07:25:07,225 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=450123.3333333333, ans=0.125 2024-09-24 07:25:07,818 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.15 vs. limit=15.0 2024-09-24 07:26:06,771 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=450263.3333333333, ans=0.125 2024-09-24 07:26:13,018 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=450310.0, ans=0.07 2024-09-24 07:26:25,451 INFO [train.py:1198] (2/4) Epoch 25, batch 3000, loss[loss=0.216, ctc_loss=0.1436, cr_loss=0.3624, over 16904.00 frames. ], tot_loss[loss=0.2059, ctc_loss=0.1356, cr_loss=0.3516, over 3361627.35 frames. ], batch size: 58, lr: 4.76e-03, grad_scale: 32.0 2024-09-24 07:26:25,452 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 07:26:41,267 INFO [train.py:1230] (2/4) Epoch 25, validation: loss=0.03749, ctc_loss=0.03749, cr_loss=8.201e-15, over 944034.00 frames. 2024-09-24 07:26:41,268 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 07:26:43,320 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=450356.6666666667, ans=0.1 2024-09-24 07:26:47,253 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=6.01 vs. limit=15.0 2024-09-24 07:26:56,573 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.75 vs. limit=15.0 2024-09-24 07:26:57,476 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=450403.3333333333, ans=0.0 2024-09-24 07:27:09,643 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.102e+02 1.216e+02 1.332e+02 1.454e+02 2.331e+02, threshold=2.665e+02, percent-clipped=0.0 2024-09-24 07:27:22,287 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=450450.0, ans=0.0 2024-09-24 07:27:34,317 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.65 vs. limit=6.0 2024-09-24 07:27:50,557 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=450543.3333333333, ans=0.1 2024-09-24 07:28:02,383 INFO [train.py:1198] (2/4) Epoch 25, batch 3050, loss[loss=0.1626, ctc_loss=0.1001, cr_loss=0.3128, over 16967.00 frames. ], tot_loss[loss=0.2054, ctc_loss=0.1351, cr_loss=0.3512, over 3353360.00 frames. ], batch size: 42, lr: 4.76e-03, grad_scale: 32.0 2024-09-24 07:28:02,706 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=450590.0, ans=0.125 2024-09-24 07:28:11,913 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=450590.0, ans=0.05 2024-09-24 07:28:11,926 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=450590.0, ans=0.125 2024-09-24 07:28:19,617 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=450636.6666666667, ans=0.125 2024-09-24 07:28:21,115 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=450636.6666666667, ans=0.125 2024-09-24 07:28:55,756 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=450730.0, ans=0.125 2024-09-24 07:29:21,016 INFO [train.py:1198] (2/4) Epoch 25, batch 3100, loss[loss=0.2396, ctc_loss=0.1629, cr_loss=0.3836, over 14995.00 frames. ], tot_loss[loss=0.2059, ctc_loss=0.1355, cr_loss=0.3521, over 3348875.64 frames. ], batch size: 89, lr: 4.76e-03, grad_scale: 32.0 2024-09-24 07:29:24,484 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=450823.3333333333, ans=0.0 2024-09-24 07:29:42,373 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=450870.0, ans=0.2 2024-09-24 07:29:51,454 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.084e+02 1.263e+02 1.349e+02 1.464e+02 5.566e+02, threshold=2.698e+02, percent-clipped=1.0 2024-09-24 07:30:22,900 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=450963.3333333333, ans=0.125 2024-09-24 07:30:41,209 INFO [train.py:1198] (2/4) Epoch 25, batch 3150, loss[loss=0.2056, ctc_loss=0.1322, cr_loss=0.3668, over 16901.00 frames. ], tot_loss[loss=0.2047, ctc_loss=0.1346, cr_loss=0.3504, over 3352752.93 frames. ], batch size: 58, lr: 4.75e-03, grad_scale: 32.0 2024-09-24 07:31:00,896 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.03 vs. limit=15.0 2024-09-24 07:31:17,274 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=451150.0, ans=0.1 2024-09-24 07:31:21,005 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.26 vs. limit=15.0 2024-09-24 07:31:22,038 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=451150.0, ans=0.0 2024-09-24 07:31:50,362 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten.whitening_limit, batch_count=451243.3333333333, ans=15.0 2024-09-24 07:31:52,878 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.min_positive, batch_count=451243.3333333333, ans=0.025 2024-09-24 07:31:58,814 INFO [train.py:1198] (2/4) Epoch 25, batch 3200, loss[loss=0.2068, ctc_loss=0.1384, cr_loss=0.342, over 17217.00 frames. ], tot_loss[loss=0.2047, ctc_loss=0.1346, cr_loss=0.3503, over 3356526.01 frames. ], batch size: 47, lr: 4.75e-03, grad_scale: 32.0 2024-09-24 07:32:15,005 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=451336.6666666667, ans=0.07 2024-09-24 07:32:26,942 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.256e+02 1.371e+02 1.503e+02 1.985e+02, threshold=2.741e+02, percent-clipped=0.0 2024-09-24 07:32:28,854 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=451383.3333333333, ans=0.125 2024-09-24 07:32:30,377 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=451383.3333333333, ans=0.07 2024-09-24 07:32:34,960 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=451383.3333333333, ans=0.125 2024-09-24 07:32:42,791 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=451383.3333333333, ans=0.125 2024-09-24 07:33:02,395 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.46 vs. limit=22.5 2024-09-24 07:33:08,141 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=4.50 vs. limit=12.0 2024-09-24 07:33:12,340 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=451476.6666666667, ans=0.0 2024-09-24 07:33:16,639 INFO [train.py:1198] (2/4) Epoch 25, batch 3250, loss[loss=0.1849, ctc_loss=0.1217, cr_loss=0.3161, over 16291.00 frames. ], tot_loss[loss=0.205, ctc_loss=0.1348, cr_loss=0.3507, over 3356970.38 frames. ], batch size: 36, lr: 4.75e-03, grad_scale: 32.0 2024-09-24 07:33:33,136 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.30 vs. limit=15.0 2024-09-24 07:33:40,797 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=16.11 vs. limit=22.5 2024-09-24 07:34:03,778 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.05 vs. limit=15.0 2024-09-24 07:34:24,474 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.06 vs. limit=6.0 2024-09-24 07:34:34,603 INFO [train.py:1198] (2/4) Epoch 25, batch 3300, loss[loss=0.2371, ctc_loss=0.1575, cr_loss=0.398, over 15314.00 frames. ], tot_loss[loss=0.2061, ctc_loss=0.1356, cr_loss=0.3523, over 3351743.00 frames. ], batch size: 89, lr: 4.75e-03, grad_scale: 32.0 2024-09-24 07:34:45,692 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=451756.6666666667, ans=0.125 2024-09-24 07:35:03,224 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=451803.3333333333, ans=0.1 2024-09-24 07:35:04,442 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.271e+02 1.364e+02 1.475e+02 2.205e+02, threshold=2.729e+02, percent-clipped=0.0 2024-09-24 07:35:25,941 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=9.01 vs. limit=15.0 2024-09-24 07:35:33,291 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=451896.6666666667, ans=0.125 2024-09-24 07:35:56,273 INFO [train.py:1198] (2/4) Epoch 25, batch 3350, loss[loss=0.1916, ctc_loss=0.1271, cr_loss=0.3222, over 17149.00 frames. ], tot_loss[loss=0.2051, ctc_loss=0.1349, cr_loss=0.3512, over 3352974.44 frames. ], batch size: 41, lr: 4.75e-03, grad_scale: 32.0 2024-09-24 07:36:01,313 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=451990.0, ans=0.0 2024-09-24 07:36:07,418 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=451990.0, ans=0.2 2024-09-24 07:36:13,684 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=452036.6666666667, ans=0.0 2024-09-24 07:36:37,793 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=9.83 vs. limit=15.0 2024-09-24 07:37:10,780 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=452176.6666666667, ans=0.025 2024-09-24 07:37:15,193 INFO [train.py:1198] (2/4) Epoch 25, batch 3400, loss[loss=0.2051, ctc_loss=0.1365, cr_loss=0.3431, over 17036.00 frames. ], tot_loss[loss=0.2052, ctc_loss=0.1349, cr_loss=0.3515, over 3363264.55 frames. ], batch size: 56, lr: 4.75e-03, grad_scale: 32.0 2024-09-24 07:37:42,231 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=452270.0, ans=0.0 2024-09-24 07:37:43,413 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.052e+02 1.278e+02 1.359e+02 1.525e+02 3.338e+02, threshold=2.719e+02, percent-clipped=1.0 2024-09-24 07:37:48,202 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=452316.6666666667, ans=0.1 2024-09-24 07:38:18,351 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=4.83 vs. limit=15.0 2024-09-24 07:38:35,006 INFO [train.py:1198] (2/4) Epoch 25, batch 3450, loss[loss=0.2281, ctc_loss=0.1541, cr_loss=0.3703, over 15041.00 frames. ], tot_loss[loss=0.2059, ctc_loss=0.1355, cr_loss=0.3522, over 3359727.64 frames. ], batch size: 89, lr: 4.75e-03, grad_scale: 32.0 2024-09-24 07:38:43,098 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=452456.6666666667, ans=0.125 2024-09-24 07:38:48,001 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 07:39:06,767 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=452550.0, ans=0.125 2024-09-24 07:39:13,061 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=452550.0, ans=0.0 2024-09-24 07:39:55,459 INFO [train.py:1198] (2/4) Epoch 25, batch 3500, loss[loss=0.2022, ctc_loss=0.1319, cr_loss=0.3514, over 17221.00 frames. ], tot_loss[loss=0.2072, ctc_loss=0.1365, cr_loss=0.3537, over 3342154.36 frames. ], batch size: 50, lr: 4.75e-03, grad_scale: 32.0 2024-09-24 07:40:02,125 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=452690.0, ans=0.125 2024-09-24 07:40:20,783 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=452736.6666666667, ans=0.125 2024-09-24 07:40:23,617 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.059e+02 1.253e+02 1.349e+02 1.455e+02 2.168e+02, threshold=2.697e+02, percent-clipped=1.0 2024-09-24 07:40:34,369 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.30 vs. limit=15.0 2024-09-24 07:40:42,899 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=452830.0, ans=0.2 2024-09-24 07:40:44,457 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=452830.0, ans=0.1 2024-09-24 07:41:13,559 INFO [train.py:1198] (2/4) Epoch 25, batch 3550, loss[loss=0.2233, ctc_loss=0.1471, cr_loss=0.3807, over 16454.00 frames. ], tot_loss[loss=0.2062, ctc_loss=0.1358, cr_loss=0.3518, over 3346742.00 frames. ], batch size: 66, lr: 4.74e-03, grad_scale: 16.0 2024-09-24 07:41:50,900 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=453016.6666666667, ans=0.125 2024-09-24 07:42:02,163 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.39 vs. limit=6.0 2024-09-24 07:42:06,309 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=453063.3333333333, ans=0.125 2024-09-24 07:42:15,856 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=453110.0, ans=0.1 2024-09-24 07:42:23,978 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.22 vs. limit=6.0 2024-09-24 07:42:31,017 INFO [train.py:1198] (2/4) Epoch 25, batch 3600, loss[loss=0.1659, ctc_loss=0.1044, cr_loss=0.3072, over 17281.00 frames. ], tot_loss[loss=0.2057, ctc_loss=0.1355, cr_loss=0.3514, over 3350343.51 frames. ], batch size: 42, lr: 4.74e-03, grad_scale: 32.0 2024-09-24 07:42:57,563 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=453203.3333333333, ans=0.015 2024-09-24 07:43:00,635 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.246e+02 1.327e+02 1.467e+02 2.954e+02, threshold=2.655e+02, percent-clipped=1.0 2024-09-24 07:43:02,385 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=453250.0, ans=0.0 2024-09-24 07:43:05,601 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=453250.0, ans=0.025 2024-09-24 07:43:16,756 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=453296.6666666667, ans=0.125 2024-09-24 07:43:39,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=453343.3333333333, ans=0.125 2024-09-24 07:43:48,678 INFO [train.py:1198] (2/4) Epoch 25, batch 3650, loss[loss=0.2012, ctc_loss=0.1344, cr_loss=0.334, over 17301.00 frames. ], tot_loss[loss=0.2052, ctc_loss=0.135, cr_loss=0.3511, over 3361129.06 frames. ], batch size: 46, lr: 4.74e-03, grad_scale: 32.0 2024-09-24 07:44:29,918 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=453483.3333333333, ans=0.125 2024-09-24 07:44:30,701 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=4.76 vs. limit=15.0 2024-09-24 07:44:31,541 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=453483.3333333333, ans=0.1 2024-09-24 07:45:11,868 INFO [train.py:1198] (2/4) Epoch 25, batch 3700, loss[loss=0.1995, ctc_loss=0.1305, cr_loss=0.3452, over 17239.00 frames. ], tot_loss[loss=0.207, ctc_loss=0.1363, cr_loss=0.3532, over 3361065.12 frames. ], batch size: 50, lr: 4.74e-03, grad_scale: 32.0 2024-09-24 07:45:12,198 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=453623.3333333333, ans=0.125 2024-09-24 07:45:12,581 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.92 vs. limit=22.5 2024-09-24 07:45:37,175 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=453670.0, ans=0.125 2024-09-24 07:45:38,744 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=453670.0, ans=0.0 2024-09-24 07:45:41,667 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.073e+02 1.237e+02 1.315e+02 1.374e+02 1.764e+02, threshold=2.629e+02, percent-clipped=0.0 2024-09-24 07:45:42,515 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=10.65 vs. limit=15.0 2024-09-24 07:46:02,591 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=453763.3333333333, ans=0.1 2024-09-24 07:46:18,082 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=453810.0, ans=0.0 2024-09-24 07:46:27,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=453810.0, ans=0.0 2024-09-24 07:46:27,507 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=453810.0, ans=0.1 2024-09-24 07:46:27,617 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=453810.0, ans=0.0 2024-09-24 07:46:30,214 INFO [train.py:1198] (2/4) Epoch 25, batch 3750, loss[loss=0.2391, ctc_loss=0.1662, cr_loss=0.3648, over 11737.00 frames. ], tot_loss[loss=0.2072, ctc_loss=0.1366, cr_loss=0.3529, over 3339086.94 frames. ], batch size: 123, lr: 4.74e-03, grad_scale: 32.0 2024-09-24 07:46:30,399 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=453856.6666666667, ans=0.125 2024-09-24 07:46:50,901 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=453903.3333333333, ans=0.1 2024-09-24 07:46:52,392 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=453903.3333333333, ans=0.0 2024-09-24 07:46:52,438 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=453903.3333333333, ans=0.0 2024-09-24 07:46:55,460 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=453903.3333333333, ans=0.2 2024-09-24 07:47:07,892 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=453950.0, ans=0.05 2024-09-24 07:47:20,273 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=453996.6666666667, ans=0.0 2024-09-24 07:47:31,260 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=454043.3333333333, ans=0.0 2024-09-24 07:47:32,736 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=454043.3333333333, ans=0.0 2024-09-24 07:47:42,898 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=454043.3333333333, ans=0.1 2024-09-24 07:47:49,006 INFO [train.py:1198] (2/4) Epoch 25, batch 3800, loss[loss=0.1853, ctc_loss=0.1189, cr_loss=0.332, over 16330.00 frames. ], tot_loss[loss=0.2069, ctc_loss=0.1363, cr_loss=0.3526, over 3337221.82 frames. ], batch size: 36, lr: 4.74e-03, grad_scale: 32.0 2024-09-24 07:48:18,770 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.102e+02 1.251e+02 1.351e+02 1.487e+02 2.041e+02, threshold=2.702e+02, percent-clipped=0.0 2024-09-24 07:48:43,473 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=454230.0, ans=0.125 2024-09-24 07:48:44,949 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=454230.0, ans=0.0 2024-09-24 07:49:00,324 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=454276.6666666667, ans=0.125 2024-09-24 07:49:01,919 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=454276.6666666667, ans=0.0 2024-09-24 07:49:07,749 INFO [train.py:1198] (2/4) Epoch 25, batch 3850, loss[loss=0.2235, ctc_loss=0.1553, cr_loss=0.3412, over 11550.00 frames. ], tot_loss[loss=0.209, ctc_loss=0.1382, cr_loss=0.3544, over 3284628.55 frames. ], batch size: 123, lr: 4.74e-03, grad_scale: 16.0 2024-09-24 07:49:08,080 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=454323.3333333333, ans=0.125 2024-09-24 07:49:14,321 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=454323.3333333333, ans=0.0 2024-09-24 07:49:18,923 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=454323.3333333333, ans=0.125 2024-09-24 07:49:26,686 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 07:49:34,281 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=454370.0, ans=0.125 2024-09-24 07:49:37,433 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=454416.6666666667, ans=0.125 2024-09-24 07:49:38,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=454416.6666666667, ans=0.1 2024-09-24 07:50:01,313 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=454463.3333333333, ans=0.1 2024-09-24 07:50:09,337 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=10.53 vs. limit=22.5 2024-09-24 07:51:09,080 INFO [train.py:1198] (2/4) Epoch 26, batch 0, loss[loss=0.1717, ctc_loss=0.1123, cr_loss=0.297, over 16268.00 frames. ], tot_loss[loss=0.1717, ctc_loss=0.1123, cr_loss=0.297, over 16268.00 frames. ], batch size: 36, lr: 4.64e-03, grad_scale: 32.0 2024-09-24 07:51:09,081 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 07:51:25,119 INFO [train.py:1230] (2/4) Epoch 26, validation: loss=0.03743, ctc_loss=0.03743, cr_loss=8.662e-15, over 944034.00 frames. 2024-09-24 07:51:25,120 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 07:51:30,035 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=454538.0, ans=0.125 2024-09-24 07:51:36,414 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=454538.0, ans=0.0 2024-09-24 07:51:37,390 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module1.whiten, num_groups=1, num_channels=192, metric=8.72 vs. limit=15.0 2024-09-24 07:51:39,704 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=454584.6666666667, ans=0.0 2024-09-24 07:51:42,750 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=454584.6666666667, ans=0.025 2024-09-24 07:51:47,591 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=454584.6666666667, ans=0.0 2024-09-24 07:52:01,485 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=454631.3333333333, ans=0.0 2024-09-24 07:52:06,030 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.051e+02 1.313e+02 1.484e+02 1.654e+02 2.315e+02, threshold=2.969e+02, percent-clipped=0.0 2024-09-24 07:52:39,479 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=454724.6666666667, ans=0.0 2024-09-24 07:52:48,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=454771.3333333333, ans=0.125 2024-09-24 07:52:50,421 INFO [train.py:1198] (2/4) Epoch 26, batch 50, loss[loss=0.1992, ctc_loss=0.1304, cr_loss=0.344, over 17073.00 frames. ], tot_loss[loss=0.2039, ctc_loss=0.134, cr_loss=0.3494, over 767484.87 frames. ], batch size: 46, lr: 4.64e-03, grad_scale: 32.0 2024-09-24 07:53:29,626 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.88 vs. limit=15.0 2024-09-24 07:53:39,914 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=454911.3333333333, ans=0.0 2024-09-24 07:53:41,579 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=454911.3333333333, ans=0.2 2024-09-24 07:54:05,362 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=454958.0, ans=0.025 2024-09-24 07:54:11,279 INFO [train.py:1198] (2/4) Epoch 26, batch 100, loss[loss=0.2355, ctc_loss=0.1573, cr_loss=0.3912, over 17245.00 frames. ], tot_loss[loss=0.2063, ctc_loss=0.1358, cr_loss=0.3525, over 1340826.40 frames. ], batch size: 55, lr: 4.64e-03, grad_scale: 32.0 2024-09-24 07:54:16,498 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=455004.6666666667, ans=0.0 2024-09-24 07:54:51,770 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.228e+02 1.285e+02 1.398e+02 1.660e+02, threshold=2.570e+02, percent-clipped=0.0 2024-09-24 07:55:00,019 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=455144.6666666667, ans=0.125 2024-09-24 07:55:20,710 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=455191.3333333333, ans=0.125 2024-09-24 07:55:33,073 INFO [train.py:1198] (2/4) Epoch 26, batch 150, loss[loss=0.191, ctc_loss=0.1221, cr_loss=0.3445, over 17036.00 frames. ], tot_loss[loss=0.2061, ctc_loss=0.1357, cr_loss=0.352, over 1781402.08 frames. ], batch size: 44, lr: 4.64e-03, grad_scale: 32.0 2024-09-24 07:55:36,586 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=455238.0, ans=0.125 2024-09-24 07:55:42,836 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=455238.0, ans=0.0 2024-09-24 07:55:46,495 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten.whitening_limit, batch_count=455238.0, ans=15.0 2024-09-24 07:55:49,504 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=455284.6666666667, ans=0.2 2024-09-24 07:55:49,632 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=455284.6666666667, ans=0.2 2024-09-24 07:56:18,155 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=455331.3333333333, ans=0.025 2024-09-24 07:56:44,918 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=455424.6666666667, ans=0.125 2024-09-24 07:56:55,946 INFO [train.py:1198] (2/4) Epoch 26, batch 200, loss[loss=0.1914, ctc_loss=0.1213, cr_loss=0.3502, over 17152.00 frames. ], tot_loss[loss=0.2054, ctc_loss=0.135, cr_loss=0.352, over 2141237.96 frames. ], batch size: 48, lr: 4.64e-03, grad_scale: 32.0 2024-09-24 07:57:01,097 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=455471.3333333333, ans=0.0 2024-09-24 07:57:37,053 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.254e+02 1.327e+02 1.395e+02 2.472e+02, threshold=2.655e+02, percent-clipped=0.0 2024-09-24 07:57:40,505 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=455564.6666666667, ans=0.125 2024-09-24 07:58:10,832 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 07:58:17,095 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=455658.0, ans=0.1 2024-09-24 07:58:21,704 INFO [train.py:1198] (2/4) Epoch 26, batch 250, loss[loss=0.1981, ctc_loss=0.132, cr_loss=0.3304, over 17142.00 frames. ], tot_loss[loss=0.206, ctc_loss=0.1354, cr_loss=0.3527, over 2410180.26 frames. ], batch size: 48, lr: 4.64e-03, grad_scale: 32.0 2024-09-24 07:58:25,322 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=455704.6666666667, ans=0.125 2024-09-24 07:59:06,924 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=455798.0, ans=0.125 2024-09-24 07:59:11,704 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=455844.6666666667, ans=0.125 2024-09-24 07:59:25,654 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=455844.6666666667, ans=0.0 2024-09-24 07:59:41,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=455891.3333333333, ans=0.1 2024-09-24 07:59:44,774 INFO [train.py:1198] (2/4) Epoch 26, batch 300, loss[loss=0.2515, ctc_loss=0.1736, cr_loss=0.3892, over 11942.00 frames. ], tot_loss[loss=0.2059, ctc_loss=0.1355, cr_loss=0.3522, over 2608734.05 frames. ], batch size: 123, lr: 4.64e-03, grad_scale: 32.0 2024-09-24 07:59:55,111 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=5.84 vs. limit=15.0 2024-09-24 08:00:22,959 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.131e+02 1.258e+02 1.353e+02 1.431e+02 1.919e+02, threshold=2.706e+02, percent-clipped=0.0 2024-09-24 08:00:40,913 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 08:00:58,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=456124.6666666667, ans=0.125 2024-09-24 08:01:04,575 INFO [train.py:1198] (2/4) Epoch 26, batch 350, loss[loss=0.2292, ctc_loss=0.1479, cr_loss=0.4064, over 17311.00 frames. ], tot_loss[loss=0.2067, ctc_loss=0.1361, cr_loss=0.3531, over 2781541.58 frames. ], batch size: 46, lr: 4.63e-03, grad_scale: 32.0 2024-09-24 08:01:12,755 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=456171.3333333333, ans=0.125 2024-09-24 08:01:41,460 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=456264.6666666667, ans=0.0 2024-09-24 08:01:53,332 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module1.whiten, num_groups=1, num_channels=192, metric=9.26 vs. limit=15.0 2024-09-24 08:01:58,841 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=456311.3333333333, ans=0.0 2024-09-24 08:02:30,047 INFO [train.py:1198] (2/4) Epoch 26, batch 400, loss[loss=0.2278, ctc_loss=0.1511, cr_loss=0.3834, over 15875.00 frames. ], tot_loss[loss=0.2074, ctc_loss=0.1365, cr_loss=0.3544, over 2904216.92 frames. ], batch size: 74, lr: 4.63e-03, grad_scale: 32.0 2024-09-24 08:02:57,374 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=456451.3333333333, ans=0.125 2024-09-24 08:03:12,625 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.258e+02 1.321e+02 1.410e+02 2.001e+02, threshold=2.643e+02, percent-clipped=0.0 2024-09-24 08:03:19,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=456544.6666666667, ans=0.0 2024-09-24 08:03:29,443 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=456544.6666666667, ans=0.0 2024-09-24 08:03:52,934 INFO [train.py:1198] (2/4) Epoch 26, batch 450, loss[loss=0.204, ctc_loss=0.1372, cr_loss=0.3341, over 17308.00 frames. ], tot_loss[loss=0.206, ctc_loss=0.1355, cr_loss=0.3525, over 3010170.39 frames. ], batch size: 51, lr: 4.63e-03, grad_scale: 32.0 2024-09-24 08:04:37,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=456731.3333333333, ans=0.125 2024-09-24 08:05:15,525 INFO [train.py:1198] (2/4) Epoch 26, batch 500, loss[loss=0.1914, ctc_loss=0.123, cr_loss=0.3423, over 17073.00 frames. ], tot_loss[loss=0.206, ctc_loss=0.1354, cr_loss=0.3528, over 3079756.75 frames. ], batch size: 46, lr: 4.63e-03, grad_scale: 32.0 2024-09-24 08:05:17,952 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.81 vs. limit=10.0 2024-09-24 08:05:25,566 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=456871.3333333333, ans=0.0 2024-09-24 08:05:33,421 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=456918.0, ans=0.0 2024-09-24 08:05:41,366 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.min_positive, batch_count=456918.0, ans=0.025 2024-09-24 08:05:47,848 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=456964.6666666667, ans=0.0 2024-09-24 08:05:55,348 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.259e+02 1.357e+02 1.518e+02 2.705e+02, threshold=2.714e+02, percent-clipped=1.0 2024-09-24 08:06:17,640 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=457058.0, ans=0.125 2024-09-24 08:06:30,425 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=457058.0, ans=0.0 2024-09-24 08:06:30,907 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=9.78 vs. limit=15.0 2024-09-24 08:06:32,153 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=457058.0, ans=0.125 2024-09-24 08:06:37,535 INFO [train.py:1198] (2/4) Epoch 26, batch 550, loss[loss=0.2238, ctc_loss=0.1473, cr_loss=0.3825, over 17032.00 frames. ], tot_loss[loss=0.2071, ctc_loss=0.1362, cr_loss=0.3542, over 3139641.22 frames. ], batch size: 56, lr: 4.63e-03, grad_scale: 32.0 2024-09-24 08:06:47,539 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=457104.6666666667, ans=0.125 2024-09-24 08:07:02,073 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.max_abs, batch_count=457151.3333333333, ans=10.0 2024-09-24 08:07:02,646 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.49 vs. limit=15.0 2024-09-24 08:07:20,408 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.min_positive, batch_count=457198.0, ans=0.05 2024-09-24 08:07:31,448 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=457244.6666666667, ans=0.125 2024-09-24 08:08:00,421 INFO [train.py:1198] (2/4) Epoch 26, batch 600, loss[loss=0.2391, ctc_loss=0.1562, cr_loss=0.4142, over 17045.00 frames. ], tot_loss[loss=0.2055, ctc_loss=0.1351, cr_loss=0.3519, over 3200168.05 frames. ], batch size: 52, lr: 4.63e-03, grad_scale: 32.0 2024-09-24 08:08:22,512 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=457384.6666666667, ans=0.2 2024-09-24 08:08:24,110 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=457384.6666666667, ans=0.125 2024-09-24 08:08:37,054 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=457431.3333333333, ans=0.125 2024-09-24 08:08:40,104 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=457431.3333333333, ans=0.125 2024-09-24 08:08:42,969 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.060e+02 1.231e+02 1.314e+02 1.456e+02 1.940e+02, threshold=2.629e+02, percent-clipped=0.0 2024-09-24 08:08:49,697 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=457478.0, ans=0.125 2024-09-24 08:09:09,013 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer_na.min_abs, batch_count=457524.6666666667, ans=0.02 2024-09-24 08:09:21,486 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=457524.6666666667, ans=0.025 2024-09-24 08:09:21,540 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=457524.6666666667, ans=0.1 2024-09-24 08:09:26,035 INFO [train.py:1198] (2/4) Epoch 26, batch 650, loss[loss=0.2268, ctc_loss=0.1495, cr_loss=0.3867, over 15928.00 frames. ], tot_loss[loss=0.2044, ctc_loss=0.1344, cr_loss=0.35, over 3230140.23 frames. ], batch size: 74, lr: 4.63e-03, grad_scale: 32.0 2024-09-24 08:09:33,381 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.23 vs. limit=22.5 2024-09-24 08:10:27,587 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=457711.3333333333, ans=0.0 2024-09-24 08:10:46,655 INFO [train.py:1198] (2/4) Epoch 26, batch 700, loss[loss=0.1691, ctc_loss=0.1067, cr_loss=0.312, over 17033.00 frames. ], tot_loss[loss=0.2048, ctc_loss=0.1348, cr_loss=0.3498, over 3252329.59 frames. ], batch size: 39, lr: 4.63e-03, grad_scale: 32.0 2024-09-24 08:11:06,696 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.82 vs. limit=15.0 2024-09-24 08:11:28,400 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.068e+02 1.269e+02 1.380e+02 1.511e+02 2.346e+02, threshold=2.761e+02, percent-clipped=0.0 2024-09-24 08:11:30,336 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=457898.0, ans=0.0 2024-09-24 08:11:50,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=457944.6666666667, ans=0.0 2024-09-24 08:12:11,812 INFO [train.py:1198] (2/4) Epoch 26, batch 750, loss[loss=0.1986, ctc_loss=0.1311, cr_loss=0.3375, over 16094.00 frames. ], tot_loss[loss=0.204, ctc_loss=0.1342, cr_loss=0.3493, over 3282915.00 frames. ], batch size: 74, lr: 4.63e-03, grad_scale: 16.0 2024-09-24 08:12:18,885 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.65 vs. limit=15.0 2024-09-24 08:12:22,128 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=10.77 vs. limit=15.0 2024-09-24 08:12:42,583 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=458131.3333333333, ans=0.125 2024-09-24 08:12:45,840 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=458131.3333333333, ans=0.0 2024-09-24 08:12:47,389 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 08:13:12,513 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=458178.0, ans=0.0 2024-09-24 08:13:19,592 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.49 vs. limit=15.0 2024-09-24 08:13:22,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=458224.6666666667, ans=0.125 2024-09-24 08:13:34,684 INFO [train.py:1198] (2/4) Epoch 26, batch 800, loss[loss=0.2103, ctc_loss=0.1421, cr_loss=0.3412, over 17139.00 frames. ], tot_loss[loss=0.2039, ctc_loss=0.1341, cr_loss=0.349, over 3293750.08 frames. ], batch size: 45, lr: 4.62e-03, grad_scale: 32.0 2024-09-24 08:13:41,258 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=458271.3333333333, ans=0.0 2024-09-24 08:14:06,333 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=4.34 vs. limit=15.0 2024-09-24 08:14:16,067 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 08:14:18,900 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.021e+02 1.286e+02 1.386e+02 1.460e+02 2.197e+02, threshold=2.772e+02, percent-clipped=0.0 2024-09-24 08:14:20,821 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=458364.6666666667, ans=0.0 2024-09-24 08:14:27,218 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=458411.3333333333, ans=0.125 2024-09-24 08:14:33,909 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.36 vs. limit=22.5 2024-09-24 08:14:39,054 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.14 vs. limit=6.0 2024-09-24 08:14:57,079 INFO [train.py:1198] (2/4) Epoch 26, batch 850, loss[loss=0.1667, ctc_loss=0.1086, cr_loss=0.2905, over 16965.00 frames. ], tot_loss[loss=0.2042, ctc_loss=0.1343, cr_loss=0.3495, over 3307374.34 frames. ], batch size: 42, lr: 4.62e-03, grad_scale: 32.0 2024-09-24 08:15:34,394 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=458598.0, ans=0.125 2024-09-24 08:15:37,518 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=458598.0, ans=0.125 2024-09-24 08:15:44,087 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 08:16:03,552 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=10.67 vs. limit=22.5 2024-09-24 08:16:16,009 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=458738.0, ans=0.0 2024-09-24 08:16:17,229 INFO [train.py:1198] (2/4) Epoch 26, batch 900, loss[loss=0.19, ctc_loss=0.1224, cr_loss=0.338, over 17295.00 frames. ], tot_loss[loss=0.2042, ctc_loss=0.1342, cr_loss=0.3499, over 3322801.44 frames. ], batch size: 46, lr: 4.62e-03, grad_scale: 32.0 2024-09-24 08:16:19,544 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=5.13 vs. limit=10.0 2024-09-24 08:16:39,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=458784.6666666667, ans=0.125 2024-09-24 08:17:01,223 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.265e+02 1.347e+02 1.465e+02 1.812e+02, threshold=2.693e+02, percent-clipped=0.0 2024-09-24 08:17:01,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=458831.3333333333, ans=0.125 2024-09-24 08:17:11,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=458878.0, ans=0.0 2024-09-24 08:17:22,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=458878.0, ans=0.0 2024-09-24 08:17:24,433 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=458924.6666666667, ans=0.1 2024-09-24 08:17:26,239 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=458924.6666666667, ans=0.07 2024-09-24 08:17:36,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=458924.6666666667, ans=0.125 2024-09-24 08:17:41,524 INFO [train.py:1198] (2/4) Epoch 26, batch 950, loss[loss=0.2187, ctc_loss=0.1416, cr_loss=0.385, over 17213.00 frames. ], tot_loss[loss=0.2048, ctc_loss=0.1346, cr_loss=0.3507, over 3331473.07 frames. ], batch size: 47, lr: 4.62e-03, grad_scale: 32.0 2024-09-24 08:17:45,171 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=458971.3333333333, ans=0.125 2024-09-24 08:18:07,644 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.59 vs. limit=15.0 2024-09-24 08:18:19,128 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=12.08 vs. limit=15.0 2024-09-24 08:18:24,057 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.93 vs. limit=6.0 2024-09-24 08:18:26,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=459064.6666666667, ans=0.125 2024-09-24 08:18:28,083 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=459064.6666666667, ans=0.2 2024-09-24 08:18:31,220 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=459111.3333333333, ans=0.1 2024-09-24 08:18:41,510 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.55 vs. limit=15.0 2024-09-24 08:19:04,837 INFO [train.py:1198] (2/4) Epoch 26, batch 1000, loss[loss=0.1711, ctc_loss=0.1078, cr_loss=0.3165, over 17035.00 frames. ], tot_loss[loss=0.2047, ctc_loss=0.1346, cr_loss=0.3506, over 3332934.39 frames. ], batch size: 39, lr: 4.62e-03, grad_scale: 32.0 2024-09-24 08:19:17,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=459204.6666666667, ans=0.0 2024-09-24 08:19:29,304 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.59 vs. limit=15.0 2024-09-24 08:19:40,191 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.69 vs. limit=15.0 2024-09-24 08:19:48,775 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.098e+02 1.281e+02 1.338e+02 1.458e+02 2.157e+02, threshold=2.677e+02, percent-clipped=0.0 2024-09-24 08:20:27,260 INFO [train.py:1198] (2/4) Epoch 26, batch 1050, loss[loss=0.1904, ctc_loss=0.124, cr_loss=0.3316, over 17113.00 frames. ], tot_loss[loss=0.2056, ctc_loss=0.1353, cr_loss=0.3518, over 3334545.55 frames. ], batch size: 43, lr: 4.62e-03, grad_scale: 32.0 2024-09-24 08:20:40,303 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=459438.0, ans=0.125 2024-09-24 08:20:52,405 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=6.25 vs. limit=15.0 2024-09-24 08:21:17,914 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=18.69 vs. limit=22.5 2024-09-24 08:21:37,541 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=459624.6666666667, ans=0.1 2024-09-24 08:21:45,416 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=459624.6666666667, ans=0.07 2024-09-24 08:21:50,046 INFO [train.py:1198] (2/4) Epoch 26, batch 1100, loss[loss=0.1825, ctc_loss=0.1193, cr_loss=0.3159, over 17217.00 frames. ], tot_loss[loss=0.2053, ctc_loss=0.135, cr_loss=0.3513, over 3333485.70 frames. ], batch size: 47, lr: 4.62e-03, grad_scale: 32.0 2024-09-24 08:21:56,731 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=459671.3333333333, ans=0.125 2024-09-24 08:22:34,488 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.255e+02 1.355e+02 1.464e+02 2.621e+02, threshold=2.711e+02, percent-clipped=0.0 2024-09-24 08:22:34,877 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=459764.6666666667, ans=0.125 2024-09-24 08:22:55,262 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=459858.0, ans=0.125 2024-09-24 08:23:15,199 INFO [train.py:1198] (2/4) Epoch 26, batch 1150, loss[loss=0.1949, ctc_loss=0.1273, cr_loss=0.338, over 16961.00 frames. ], tot_loss[loss=0.2076, ctc_loss=0.1367, cr_loss=0.3547, over 3334871.91 frames. ], batch size: 42, lr: 4.62e-03, grad_scale: 32.0 2024-09-24 08:23:20,114 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=459904.6666666667, ans=0.125 2024-09-24 08:23:37,675 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=459951.3333333333, ans=0.1 2024-09-24 08:24:08,196 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=460044.6666666667, ans=0.2 2024-09-24 08:24:15,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=460044.6666666667, ans=0.0 2024-09-24 08:24:16,969 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=460044.6666666667, ans=0.1 2024-09-24 08:24:21,946 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=460091.3333333333, ans=0.09899494936611666 2024-09-24 08:24:37,840 INFO [train.py:1198] (2/4) Epoch 26, batch 1200, loss[loss=0.2138, ctc_loss=0.142, cr_loss=0.3588, over 17212.00 frames. ], tot_loss[loss=0.2061, ctc_loss=0.1356, cr_loss=0.3523, over 3335549.86 frames. ], batch size: 50, lr: 4.61e-03, grad_scale: 32.0 2024-09-24 08:25:12,437 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.30 vs. limit=15.0 2024-09-24 08:25:19,715 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.237e+02 1.345e+02 1.469e+02 1.862e+02, threshold=2.691e+02, percent-clipped=0.0 2024-09-24 08:25:20,006 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=460231.3333333333, ans=0.125 2024-09-24 08:25:58,189 INFO [train.py:1198] (2/4) Epoch 26, batch 1250, loss[loss=0.2063, ctc_loss=0.1367, cr_loss=0.3483, over 17299.00 frames. ], tot_loss[loss=0.2068, ctc_loss=0.1362, cr_loss=0.3529, over 3333191.01 frames. ], batch size: 51, lr: 4.61e-03, grad_scale: 32.0 2024-09-24 08:26:00,297 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 08:26:06,369 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=460371.3333333333, ans=0.1 2024-09-24 08:26:11,423 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=460371.3333333333, ans=0.0 2024-09-24 08:26:12,125 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.48 vs. limit=15.0 2024-09-24 08:26:30,056 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.51 vs. limit=15.0 2024-09-24 08:26:34,782 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=460464.6666666667, ans=0.125 2024-09-24 08:26:40,161 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=11.98 vs. limit=15.0 2024-09-24 08:26:41,753 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=7.67 vs. limit=15.0 2024-09-24 08:26:45,299 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.whiten, num_groups=1, num_channels=192, metric=4.96 vs. limit=12.0 2024-09-24 08:26:47,494 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=460511.3333333333, ans=0.125 2024-09-24 08:27:04,273 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=460511.3333333333, ans=0.0 2024-09-24 08:27:07,573 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=460558.0, ans=0.125 2024-09-24 08:27:20,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=460558.0, ans=0.125 2024-09-24 08:27:23,277 INFO [train.py:1198] (2/4) Epoch 26, batch 1300, loss[loss=0.2028, ctc_loss=0.133, cr_loss=0.3489, over 17032.00 frames. ], tot_loss[loss=0.2044, ctc_loss=0.1344, cr_loss=0.3498, over 3343703.68 frames. ], batch size: 44, lr: 4.61e-03, grad_scale: 32.0 2024-09-24 08:27:58,485 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.61 vs. limit=22.5 2024-09-24 08:28:02,695 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=460698.0, ans=0.0 2024-09-24 08:28:07,144 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.260e+02 1.346e+02 1.486e+02 2.192e+02, threshold=2.692e+02, percent-clipped=0.0 2024-09-24 08:28:10,809 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=460698.0, ans=0.125 2024-09-24 08:28:18,768 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=460744.6666666667, ans=0.125 2024-09-24 08:28:21,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=460744.6666666667, ans=0.0 2024-09-24 08:28:24,275 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=11.70 vs. limit=15.0 2024-09-24 08:28:45,545 INFO [train.py:1198] (2/4) Epoch 26, batch 1350, loss[loss=0.1675, ctc_loss=0.1063, cr_loss=0.3058, over 17032.00 frames. ], tot_loss[loss=0.2054, ctc_loss=0.1352, cr_loss=0.3512, over 3333008.43 frames. ], batch size: 39, lr: 4.61e-03, grad_scale: 32.0 2024-09-24 08:29:11,049 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 08:29:33,510 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=460931.3333333333, ans=0.125 2024-09-24 08:30:00,603 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=461024.6666666667, ans=0.125 2024-09-24 08:30:02,172 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=461024.6666666667, ans=0.0 2024-09-24 08:30:08,671 INFO [train.py:1198] (2/4) Epoch 26, batch 1400, loss[loss=0.1896, ctc_loss=0.1226, cr_loss=0.3351, over 17314.00 frames. ], tot_loss[loss=0.2057, ctc_loss=0.1354, cr_loss=0.3512, over 3334652.92 frames. ], batch size: 49, lr: 4.61e-03, grad_scale: 32.0 2024-09-24 08:30:38,201 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=461118.0, ans=0.125 2024-09-24 08:30:50,707 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.110e+02 1.251e+02 1.327e+02 1.445e+02 2.357e+02, threshold=2.654e+02, percent-clipped=0.0 2024-09-24 08:31:31,440 INFO [train.py:1198] (2/4) Epoch 26, batch 1450, loss[loss=0.2291, ctc_loss=0.1559, cr_loss=0.3661, over 11909.00 frames. ], tot_loss[loss=0.2047, ctc_loss=0.1347, cr_loss=0.35, over 3334134.44 frames. ], batch size: 123, lr: 4.61e-03, grad_scale: 32.0 2024-09-24 08:31:37,210 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=8.63 vs. limit=15.0 2024-09-24 08:31:47,507 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=461351.3333333333, ans=0.125 2024-09-24 08:32:02,532 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=461351.3333333333, ans=0.2 2024-09-24 08:32:26,756 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=461444.6666666667, ans=0.0 2024-09-24 08:32:32,902 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=461444.6666666667, ans=0.0 2024-09-24 08:32:55,982 INFO [train.py:1198] (2/4) Epoch 26, batch 1500, loss[loss=0.1999, ctc_loss=0.1306, cr_loss=0.3463, over 16755.00 frames. ], tot_loss[loss=0.2047, ctc_loss=0.1347, cr_loss=0.35, over 3333888.12 frames. ], batch size: 61, lr: 4.61e-03, grad_scale: 32.0 2024-09-24 08:33:16,856 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=461584.6666666667, ans=0.125 2024-09-24 08:33:37,703 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.270e+02 1.352e+02 1.469e+02 1.916e+02, threshold=2.703e+02, percent-clipped=0.0 2024-09-24 08:33:47,596 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=461678.0, ans=0.125 2024-09-24 08:33:49,186 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=461678.0, ans=0.1 2024-09-24 08:33:57,483 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=461678.0, ans=0.025 2024-09-24 08:34:16,300 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=461724.6666666667, ans=0.2 2024-09-24 08:34:19,043 INFO [train.py:1198] (2/4) Epoch 26, batch 1550, loss[loss=0.2319, ctc_loss=0.1557, cr_loss=0.3808, over 16991.00 frames. ], tot_loss[loss=0.2047, ctc_loss=0.1347, cr_loss=0.3499, over 3338712.50 frames. ], batch size: 53, lr: 4.61e-03, grad_scale: 32.0 2024-09-24 08:34:55,940 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=461864.6666666667, ans=0.2 2024-09-24 08:35:13,234 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=461911.3333333333, ans=0.125 2024-09-24 08:35:18,095 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=461911.3333333333, ans=0.0 2024-09-24 08:35:21,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=461958.0, ans=0.0 2024-09-24 08:35:22,007 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.00 vs. limit=15.0 2024-09-24 08:35:38,843 INFO [train.py:1198] (2/4) Epoch 26, batch 1600, loss[loss=0.1955, ctc_loss=0.1251, cr_loss=0.3521, over 17010.00 frames. ], tot_loss[loss=0.2048, ctc_loss=0.1348, cr_loss=0.3499, over 3347455.94 frames. ], batch size: 51, lr: 4.61e-03, grad_scale: 32.0 2024-09-24 08:35:47,315 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=462004.6666666667, ans=0.1 2024-09-24 08:36:15,136 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=462098.0, ans=0.125 2024-09-24 08:36:22,545 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.046e+02 1.247e+02 1.306e+02 1.406e+02 2.052e+02, threshold=2.612e+02, percent-clipped=0.0 2024-09-24 08:36:34,195 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=462144.6666666667, ans=0.1 2024-09-24 08:37:03,804 INFO [train.py:1198] (2/4) Epoch 26, batch 1650, loss[loss=0.2398, ctc_loss=0.1594, cr_loss=0.402, over 16114.00 frames. ], tot_loss[loss=0.2053, ctc_loss=0.135, cr_loss=0.3514, over 3353657.43 frames. ], batch size: 74, lr: 4.60e-03, grad_scale: 32.0 2024-09-24 08:37:16,838 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=462238.0, ans=0.2 2024-09-24 08:37:18,366 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=462284.6666666667, ans=0.125 2024-09-24 08:37:41,015 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=462331.3333333333, ans=0.0 2024-09-24 08:37:45,773 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=462331.3333333333, ans=0.0 2024-09-24 08:37:49,113 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=462331.3333333333, ans=0.2 2024-09-24 08:38:20,956 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.01 vs. limit=10.0 2024-09-24 08:38:26,354 INFO [train.py:1198] (2/4) Epoch 26, batch 1700, loss[loss=0.17, ctc_loss=0.1116, cr_loss=0.2918, over 16273.00 frames. ], tot_loss[loss=0.2048, ctc_loss=0.1346, cr_loss=0.3507, over 3352592.33 frames. ], batch size: 36, lr: 4.60e-03, grad_scale: 32.0 2024-09-24 08:38:44,241 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=462518.0, ans=0.125 2024-09-24 08:38:52,376 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=6.71 vs. limit=15.0 2024-09-24 08:39:10,601 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.092e+02 1.222e+02 1.318e+02 1.444e+02 2.333e+02, threshold=2.637e+02, percent-clipped=0.0 2024-09-24 08:39:12,468 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=462564.6666666667, ans=0.025 2024-09-24 08:39:41,377 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=462658.0, ans=0.0 2024-09-24 08:39:48,985 INFO [train.py:1198] (2/4) Epoch 26, batch 1750, loss[loss=0.1689, ctc_loss=0.1125, cr_loss=0.2819, over 17202.00 frames. ], tot_loss[loss=0.2037, ctc_loss=0.1339, cr_loss=0.3494, over 3356200.20 frames. ], batch size: 41, lr: 4.60e-03, grad_scale: 32.0 2024-09-24 08:39:49,417 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 08:40:01,926 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=462704.6666666667, ans=0.1 2024-09-24 08:40:21,500 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=4.80 vs. limit=15.0 2024-09-24 08:41:01,527 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=7.02 vs. limit=15.0 2024-09-24 08:41:08,744 INFO [train.py:1198] (2/4) Epoch 26, batch 1800, loss[loss=0.2078, ctc_loss=0.1343, cr_loss=0.3677, over 17007.00 frames. ], tot_loss[loss=0.2053, ctc_loss=0.1349, cr_loss=0.3517, over 3353127.19 frames. ], batch size: 58, lr: 4.60e-03, grad_scale: 32.0 2024-09-24 08:41:10,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=462938.0, ans=0.0 2024-09-24 08:41:26,099 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=462984.6666666667, ans=0.04949747468305833 2024-09-24 08:41:55,219 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.066e+02 1.293e+02 1.406e+02 1.567e+02 1.918e+02, threshold=2.813e+02, percent-clipped=0.0 2024-09-24 08:41:57,283 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=463031.3333333333, ans=0.125 2024-09-24 08:42:01,874 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=463078.0, ans=0.1 2024-09-24 08:42:03,893 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=5.22 vs. limit=12.0 2024-09-24 08:42:33,471 INFO [train.py:1198] (2/4) Epoch 26, batch 1850, loss[loss=0.2224, ctc_loss=0.1495, cr_loss=0.3644, over 16929.00 frames. ], tot_loss[loss=0.2064, ctc_loss=0.1356, cr_loss=0.3537, over 3359198.84 frames. ], batch size: 58, lr: 4.60e-03, grad_scale: 16.0 2024-09-24 08:42:56,843 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.min_positive, batch_count=463218.0, ans=0.05 2024-09-24 08:43:14,495 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.min_positive, batch_count=463264.6666666667, ans=0.05 2024-09-24 08:43:22,273 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=463311.3333333333, ans=0.125 2024-09-24 08:43:41,570 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=463358.0, ans=0.125 2024-09-24 08:43:55,584 INFO [train.py:1198] (2/4) Epoch 26, batch 1900, loss[loss=0.194, ctc_loss=0.1279, cr_loss=0.3301, over 17037.00 frames. ], tot_loss[loss=0.2052, ctc_loss=0.1347, cr_loss=0.3523, over 3367776.08 frames. ], batch size: 52, lr: 4.60e-03, grad_scale: 16.0 2024-09-24 08:44:28,590 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=463498.0, ans=0.0 2024-09-24 08:44:41,026 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.057e+02 1.265e+02 1.341e+02 1.479e+02 3.030e+02, threshold=2.683e+02, percent-clipped=1.0 2024-09-24 08:44:43,045 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=463498.0, ans=0.125 2024-09-24 08:45:17,988 INFO [train.py:1198] (2/4) Epoch 26, batch 1950, loss[loss=0.2016, ctc_loss=0.1313, cr_loss=0.3512, over 16656.00 frames. ], tot_loss[loss=0.2052, ctc_loss=0.1348, cr_loss=0.3524, over 3359047.79 frames. ], batch size: 61, lr: 4.60e-03, grad_scale: 16.0 2024-09-24 08:46:09,492 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=463778.0, ans=0.05 2024-09-24 08:46:40,692 INFO [train.py:1198] (2/4) Epoch 26, batch 2000, loss[loss=0.2127, ctc_loss=0.1417, cr_loss=0.3554, over 17302.00 frames. ], tot_loss[loss=0.2043, ctc_loss=0.134, cr_loss=0.3514, over 3368262.31 frames. ], batch size: 49, lr: 4.60e-03, grad_scale: 32.0 2024-09-24 08:46:42,659 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=463871.3333333333, ans=0.0 2024-09-24 08:47:06,133 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=463918.0, ans=0.1 2024-09-24 08:47:26,198 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=9.63 vs. limit=15.0 2024-09-24 08:47:26,725 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.293e+02 1.366e+02 1.489e+02 2.036e+02, threshold=2.732e+02, percent-clipped=0.0 2024-09-24 08:47:27,007 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=463964.6666666667, ans=0.125 2024-09-24 08:47:47,819 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=464058.0, ans=0.1 2024-09-24 08:47:50,027 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.37 vs. limit=22.5 2024-09-24 08:47:51,057 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=464058.0, ans=0.09899494936611666 2024-09-24 08:48:01,752 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=464058.0, ans=0.0 2024-09-24 08:48:01,805 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.max_abs, batch_count=464058.0, ans=10.0 2024-09-24 08:48:06,082 INFO [train.py:1198] (2/4) Epoch 26, batch 2050, loss[loss=0.2078, ctc_loss=0.1341, cr_loss=0.3685, over 17081.00 frames. ], tot_loss[loss=0.2049, ctc_loss=0.1345, cr_loss=0.3521, over 3372996.53 frames. ], batch size: 46, lr: 4.59e-03, grad_scale: 32.0 2024-09-24 08:48:19,308 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=464104.6666666667, ans=0.1 2024-09-24 08:48:19,695 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=11.72 vs. limit=15.0 2024-09-24 08:48:32,201 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=464151.3333333333, ans=0.0 2024-09-24 08:48:38,871 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 08:48:51,840 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=5.23 vs. limit=12.0 2024-09-24 08:48:53,125 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=464244.6666666667, ans=0.0 2024-09-24 08:49:09,828 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.min_abs, batch_count=464244.6666666667, ans=0.5 2024-09-24 08:49:14,725 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=464291.3333333333, ans=0.1 2024-09-24 08:49:28,855 INFO [train.py:1198] (2/4) Epoch 26, batch 2100, loss[loss=0.2218, ctc_loss=0.149, cr_loss=0.3644, over 16912.00 frames. ], tot_loss[loss=0.2056, ctc_loss=0.1351, cr_loss=0.3525, over 3368267.50 frames. ], batch size: 58, lr: 4.59e-03, grad_scale: 32.0 2024-09-24 08:49:30,775 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=464338.0, ans=0.125 2024-09-24 08:49:34,495 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.53 vs. limit=6.0 2024-09-24 08:49:44,061 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=5.19 vs. limit=12.0 2024-09-24 08:50:06,377 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=464431.3333333333, ans=0.125 2024-09-24 08:50:12,345 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.252e+02 1.369e+02 1.506e+02 2.347e+02, threshold=2.738e+02, percent-clipped=0.0 2024-09-24 08:50:19,149 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=464478.0, ans=0.0 2024-09-24 08:50:22,760 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.07 vs. limit=6.0 2024-09-24 08:50:48,960 INFO [train.py:1198] (2/4) Epoch 26, batch 2150, loss[loss=0.1952, ctc_loss=0.1275, cr_loss=0.3381, over 17104.00 frames. ], tot_loss[loss=0.2051, ctc_loss=0.1348, cr_loss=0.3513, over 3365469.08 frames. ], batch size: 40, lr: 4.59e-03, grad_scale: 32.0 2024-09-24 08:51:00,414 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=464571.3333333333, ans=0.125 2024-09-24 08:51:09,807 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=464618.0, ans=0.125 2024-09-24 08:51:26,476 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=464664.6666666667, ans=0.025 2024-09-24 08:51:36,160 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=464664.6666666667, ans=0.0 2024-09-24 08:52:05,817 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=464758.0, ans=0.125 2024-09-24 08:52:13,761 INFO [train.py:1198] (2/4) Epoch 26, batch 2200, loss[loss=0.2086, ctc_loss=0.1365, cr_loss=0.3602, over 17039.00 frames. ], tot_loss[loss=0.2042, ctc_loss=0.1341, cr_loss=0.3506, over 3368402.18 frames. ], batch size: 51, lr: 4.59e-03, grad_scale: 32.0 2024-09-24 08:52:56,972 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.150e+02 1.271e+02 1.361e+02 1.483e+02 2.576e+02, threshold=2.723e+02, percent-clipped=0.0 2024-09-24 08:53:05,077 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.48 vs. limit=10.0 2024-09-24 08:53:06,201 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=464944.6666666667, ans=0.0 2024-09-24 08:53:07,711 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=464944.6666666667, ans=0.125 2024-09-24 08:53:29,031 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=464991.3333333333, ans=0.125 2024-09-24 08:53:36,463 INFO [train.py:1198] (2/4) Epoch 26, batch 2250, loss[loss=0.236, ctc_loss=0.1575, cr_loss=0.3926, over 16814.00 frames. ], tot_loss[loss=0.2036, ctc_loss=0.1337, cr_loss=0.3495, over 3368299.37 frames. ], batch size: 61, lr: 4.59e-03, grad_scale: 32.0 2024-09-24 08:53:57,390 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer_ff2.min_abs, batch_count=465084.6666666667, ans=0.1 2024-09-24 08:54:19,609 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=9.88 vs. limit=22.5 2024-09-24 08:54:46,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer_na.min_abs, batch_count=465224.6666666667, ans=0.02 2024-09-24 08:54:51,426 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=465224.6666666667, ans=0.125 2024-09-24 08:54:52,812 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 08:54:59,001 INFO [train.py:1198] (2/4) Epoch 26, batch 2300, loss[loss=0.1889, ctc_loss=0.1194, cr_loss=0.3476, over 17040.00 frames. ], tot_loss[loss=0.2033, ctc_loss=0.1334, cr_loss=0.3495, over 3375330.45 frames. ], batch size: 39, lr: 4.59e-03, grad_scale: 32.0 2024-09-24 08:55:42,116 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=13.70 vs. limit=22.5 2024-09-24 08:55:42,650 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.066e+02 1.269e+02 1.366e+02 1.477e+02 2.036e+02, threshold=2.732e+02, percent-clipped=0.0 2024-09-24 08:55:45,090 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=8.94 vs. limit=15.0 2024-09-24 08:56:00,660 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=465411.3333333333, ans=0.2 2024-09-24 08:56:22,145 INFO [train.py:1198] (2/4) Epoch 26, batch 2350, loss[loss=0.2173, ctc_loss=0.141, cr_loss=0.3816, over 17301.00 frames. ], tot_loss[loss=0.2036, ctc_loss=0.1336, cr_loss=0.35, over 3374264.42 frames. ], batch size: 49, lr: 4.59e-03, grad_scale: 32.0 2024-09-24 08:56:27,424 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=465504.6666666667, ans=0.025 2024-09-24 08:56:45,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=465551.3333333333, ans=0.125 2024-09-24 08:56:52,505 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=465551.3333333333, ans=0.125 2024-09-24 08:56:58,895 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=465598.0, ans=0.0 2024-09-24 08:57:00,421 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=465598.0, ans=0.125 2024-09-24 08:57:00,923 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.70 vs. limit=15.0 2024-09-24 08:57:29,417 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 08:57:34,118 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=465691.3333333333, ans=0.025 2024-09-24 08:57:45,358 INFO [train.py:1198] (2/4) Epoch 26, batch 2400, loss[loss=0.2032, ctc_loss=0.1352, cr_loss=0.3403, over 17296.00 frames. ], tot_loss[loss=0.2046, ctc_loss=0.1344, cr_loss=0.3511, over 3359896.64 frames. ], batch size: 46, lr: 4.59e-03, grad_scale: 32.0 2024-09-24 08:58:23,336 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=465831.3333333333, ans=0.125 2024-09-24 08:58:28,225 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=465831.3333333333, ans=0.125 2024-09-24 08:58:32,543 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.070e+02 1.217e+02 1.274e+02 1.391e+02 1.998e+02, threshold=2.548e+02, percent-clipped=0.0 2024-09-24 08:58:39,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=465878.0, ans=0.025 2024-09-24 08:58:41,075 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.34 vs. limit=10.0 2024-09-24 08:58:56,800 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=465924.6666666667, ans=0.125 2024-09-24 08:59:10,330 INFO [train.py:1198] (2/4) Epoch 26, batch 2450, loss[loss=0.1958, ctc_loss=0.1302, cr_loss=0.3281, over 17219.00 frames. ], tot_loss[loss=0.205, ctc_loss=0.1347, cr_loss=0.3514, over 3364337.62 frames. ], batch size: 50, lr: 4.59e-03, grad_scale: 16.0 2024-09-24 08:59:19,047 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=8.46 vs. limit=15.0 2024-09-24 08:59:21,783 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=465971.3333333333, ans=0.0 2024-09-24 08:59:37,903 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=466018.0, ans=0.125 2024-09-24 09:00:11,292 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=466111.3333333333, ans=0.125 2024-09-24 09:00:25,933 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=466158.0, ans=0.1 2024-09-24 09:00:30,249 INFO [train.py:1198] (2/4) Epoch 26, batch 2500, loss[loss=0.2187, ctc_loss=0.1429, cr_loss=0.3791, over 17037.00 frames. ], tot_loss[loss=0.2048, ctc_loss=0.1345, cr_loss=0.3517, over 3366266.48 frames. ], batch size: 52, lr: 4.58e-03, grad_scale: 16.0 2024-09-24 09:00:32,075 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 09:00:56,280 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=466251.3333333333, ans=0.2 2024-09-24 09:01:05,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=466298.0, ans=0.125 2024-09-24 09:01:17,878 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.080e+02 1.260e+02 1.339e+02 1.444e+02 2.110e+02, threshold=2.678e+02, percent-clipped=0.0 2024-09-24 09:01:28,314 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.10 vs. limit=15.0 2024-09-24 09:01:50,110 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=466391.3333333333, ans=0.0 2024-09-24 09:01:56,196 INFO [train.py:1198] (2/4) Epoch 26, batch 2550, loss[loss=0.2231, ctc_loss=0.1483, cr_loss=0.374, over 17034.00 frames. ], tot_loss[loss=0.2048, ctc_loss=0.1346, cr_loss=0.3514, over 3354915.47 frames. ], batch size: 56, lr: 4.58e-03, grad_scale: 16.0 2024-09-24 09:02:07,702 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=466438.0, ans=0.0 2024-09-24 09:02:20,259 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=466484.6666666667, ans=0.1 2024-09-24 09:03:07,592 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=466624.6666666667, ans=0.0 2024-09-24 09:03:20,622 INFO [train.py:1198] (2/4) Epoch 26, batch 2600, loss[loss=0.2029, ctc_loss=0.1323, cr_loss=0.3532, over 17178.00 frames. ], tot_loss[loss=0.2061, ctc_loss=0.1355, cr_loss=0.3532, over 3354876.42 frames. ], batch size: 45, lr: 4.58e-03, grad_scale: 16.0 2024-09-24 09:03:38,892 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module1.whiten, num_groups=1, num_channels=512, metric=6.48 vs. limit=15.0 2024-09-24 09:03:44,938 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=466718.0, ans=0.125 2024-09-24 09:03:52,542 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten.whitening_limit, batch_count=466718.0, ans=15.0 2024-09-24 09:03:59,315 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=11.34 vs. limit=22.5 2024-09-24 09:04:07,742 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.094e+02 1.265e+02 1.355e+02 1.499e+02 2.594e+02, threshold=2.709e+02, percent-clipped=0.0 2024-09-24 09:04:42,886 INFO [train.py:1198] (2/4) Epoch 26, batch 2650, loss[loss=0.1638, ctc_loss=0.1052, cr_loss=0.293, over 17118.00 frames. ], tot_loss[loss=0.2063, ctc_loss=0.1357, cr_loss=0.3529, over 3335105.36 frames. ], batch size: 40, lr: 4.58e-03, grad_scale: 16.0 2024-09-24 09:05:33,603 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=467044.6666666667, ans=0.1 2024-09-24 09:06:04,510 INFO [train.py:1198] (2/4) Epoch 26, batch 2700, loss[loss=0.1991, ctc_loss=0.1315, cr_loss=0.3384, over 17212.00 frames. ], tot_loss[loss=0.2051, ctc_loss=0.1349, cr_loss=0.3508, over 3341151.91 frames. ], batch size: 50, lr: 4.58e-03, grad_scale: 16.0 2024-09-24 09:06:04,790 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=467138.0, ans=0.025 2024-09-24 09:06:11,106 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=467138.0, ans=0.125 2024-09-24 09:06:11,560 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.48 vs. limit=10.0 2024-09-24 09:06:31,875 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=467184.6666666667, ans=0.2 2024-09-24 09:06:51,908 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.073e+02 1.262e+02 1.341e+02 1.444e+02 3.624e+02, threshold=2.682e+02, percent-clipped=1.0 2024-09-24 09:06:57,805 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=10.06 vs. limit=15.0 2024-09-24 09:07:18,143 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=467324.6666666667, ans=0.1 2024-09-24 09:07:27,270 INFO [train.py:1198] (2/4) Epoch 26, batch 2750, loss[loss=0.219, ctc_loss=0.1443, cr_loss=0.3734, over 17263.00 frames. ], tot_loss[loss=0.2056, ctc_loss=0.1352, cr_loss=0.3519, over 3341690.51 frames. ], batch size: 44, lr: 4.58e-03, grad_scale: 16.0 2024-09-24 09:08:00,246 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=467464.6666666667, ans=0.0 2024-09-24 09:08:00,769 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=7.80 vs. limit=15.0 2024-09-24 09:08:43,793 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=467558.0, ans=0.125 2024-09-24 09:08:52,542 INFO [train.py:1198] (2/4) Epoch 26, batch 2800, loss[loss=0.2071, ctc_loss=0.137, cr_loss=0.3503, over 17212.00 frames. ], tot_loss[loss=0.2053, ctc_loss=0.135, cr_loss=0.3514, over 3342151.28 frames. ], batch size: 47, lr: 4.58e-03, grad_scale: 32.0 2024-09-24 09:09:28,380 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=467698.0, ans=0.2 2024-09-24 09:09:37,154 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=3.49 vs. limit=15.0 2024-09-24 09:09:37,848 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.041e+02 1.261e+02 1.357e+02 1.478e+02 1.924e+02, threshold=2.714e+02, percent-clipped=0.0 2024-09-24 09:09:44,903 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=6.22 vs. limit=15.0 2024-09-24 09:10:12,860 INFO [train.py:1198] (2/4) Epoch 26, batch 2850, loss[loss=0.175, ctc_loss=0.1125, cr_loss=0.3122, over 17248.00 frames. ], tot_loss[loss=0.2049, ctc_loss=0.1348, cr_loss=0.3507, over 3332993.97 frames. ], batch size: 42, lr: 4.58e-03, grad_scale: 32.0 2024-09-24 09:10:16,351 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=467838.0, ans=0.1 2024-09-24 09:10:16,682 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.10 vs. limit=6.0 2024-09-24 09:10:30,792 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=467884.6666666667, ans=0.2 2024-09-24 09:10:32,423 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=467884.6666666667, ans=0.125 2024-09-24 09:10:38,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=467884.6666666667, ans=0.125 2024-09-24 09:10:49,943 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.19 vs. limit=15.0 2024-09-24 09:11:01,879 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=467978.0, ans=0.125 2024-09-24 09:11:03,594 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=467978.0, ans=0.0 2024-09-24 09:11:14,746 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=467978.0, ans=0.125 2024-09-24 09:11:35,003 INFO [train.py:1198] (2/4) Epoch 26, batch 2900, loss[loss=0.2104, ctc_loss=0.1391, cr_loss=0.3562, over 16716.00 frames. ], tot_loss[loss=0.2063, ctc_loss=0.1359, cr_loss=0.3521, over 3335435.08 frames. ], batch size: 61, lr: 4.58e-03, grad_scale: 32.0 2024-09-24 09:11:48,930 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=468071.3333333333, ans=0.125 2024-09-24 09:12:08,865 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.43 vs. limit=10.0 2024-09-24 09:12:22,358 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.321e+02 1.382e+02 1.512e+02 2.485e+02, threshold=2.764e+02, percent-clipped=0.0 2024-09-24 09:12:48,661 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=11.92 vs. limit=15.0 2024-09-24 09:13:00,441 INFO [train.py:1198] (2/4) Epoch 26, batch 2950, loss[loss=0.2222, ctc_loss=0.146, cr_loss=0.3811, over 16141.00 frames. ], tot_loss[loss=0.2056, ctc_loss=0.1352, cr_loss=0.3518, over 3353508.89 frames. ], batch size: 74, lr: 4.57e-03, grad_scale: 32.0 2024-09-24 09:13:04,390 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.50 vs. limit=15.0 2024-09-24 09:13:13,754 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=468304.6666666667, ans=0.2 2024-09-24 09:13:21,764 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=468351.3333333333, ans=0.1 2024-09-24 09:13:33,011 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=468398.0, ans=0.125 2024-09-24 09:13:53,310 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=468444.6666666667, ans=0.1 2024-09-24 09:14:15,267 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=468491.3333333333, ans=0.125 2024-09-24 09:14:15,295 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=468491.3333333333, ans=0.125 2024-09-24 09:14:23,227 INFO [train.py:1198] (2/4) Epoch 26, batch 3000, loss[loss=0.1923, ctc_loss=0.1244, cr_loss=0.3395, over 17012.00 frames. ], tot_loss[loss=0.2055, ctc_loss=0.1351, cr_loss=0.3521, over 3344913.75 frames. ], batch size: 44, lr: 4.57e-03, grad_scale: 32.0 2024-09-24 09:14:23,227 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 09:14:38,566 INFO [train.py:1230] (2/4) Epoch 26, validation: loss=0.03742, ctc_loss=0.03742, cr_loss=8.706e-15, over 944034.00 frames. 2024-09-24 09:14:38,567 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 09:14:45,060 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=468538.0, ans=0.125 2024-09-24 09:15:11,457 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=468631.3333333333, ans=0.1 2024-09-24 09:15:22,387 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.270e+02 1.354e+02 1.456e+02 4.080e+02, threshold=2.708e+02, percent-clipped=1.0 2024-09-24 09:15:49,169 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=468724.6666666667, ans=0.1 2024-09-24 09:15:56,790 INFO [train.py:1198] (2/4) Epoch 26, batch 3050, loss[loss=0.2162, ctc_loss=0.1407, cr_loss=0.3777, over 17297.00 frames. ], tot_loss[loss=0.2046, ctc_loss=0.1344, cr_loss=0.3508, over 3350062.60 frames. ], batch size: 49, lr: 4.57e-03, grad_scale: 32.0 2024-09-24 09:16:19,200 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=468818.0, ans=0.2 2024-09-24 09:16:22,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=468818.0, ans=0.125 2024-09-24 09:16:31,528 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=468864.6666666667, ans=0.125 2024-09-24 09:16:38,024 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=468864.6666666667, ans=0.125 2024-09-24 09:16:40,353 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.78 vs. limit=6.0 2024-09-24 09:17:14,927 INFO [train.py:1198] (2/4) Epoch 26, batch 3100, loss[loss=0.206, ctc_loss=0.1333, cr_loss=0.3635, over 17292.00 frames. ], tot_loss[loss=0.2049, ctc_loss=0.1347, cr_loss=0.3513, over 3357667.82 frames. ], batch size: 49, lr: 4.57e-03, grad_scale: 32.0 2024-09-24 09:17:18,979 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.03 vs. limit=15.0 2024-09-24 09:17:19,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=469004.6666666667, ans=0.0 2024-09-24 09:17:19,901 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=469004.6666666667, ans=0.0 2024-09-24 09:17:53,849 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=469098.0, ans=0.1 2024-09-24 09:18:01,341 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.099e+02 1.241e+02 1.330e+02 1.444e+02 2.208e+02, threshold=2.660e+02, percent-clipped=0.0 2024-09-24 09:18:03,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=469144.6666666667, ans=0.125 2024-09-24 09:18:03,217 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=469144.6666666667, ans=0.0 2024-09-24 09:18:03,306 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=469144.6666666667, ans=0.1 2024-09-24 09:18:14,360 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 09:18:15,744 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=469144.6666666667, ans=0.125 2024-09-24 09:18:35,771 INFO [train.py:1198] (2/4) Epoch 26, batch 3150, loss[loss=0.2431, ctc_loss=0.1664, cr_loss=0.3835, over 15071.00 frames. ], tot_loss[loss=0.2054, ctc_loss=0.1351, cr_loss=0.3517, over 3358905.89 frames. ], batch size: 89, lr: 4.57e-03, grad_scale: 32.0 2024-09-24 09:18:50,619 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=469284.6666666667, ans=0.0 2024-09-24 09:19:27,126 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=469378.0, ans=0.1 2024-09-24 09:19:42,540 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=469424.6666666667, ans=0.025 2024-09-24 09:19:50,471 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=469424.6666666667, ans=0.0 2024-09-24 09:19:50,506 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=469424.6666666667, ans=0.125 2024-09-24 09:19:56,681 INFO [train.py:1198] (2/4) Epoch 26, batch 3200, loss[loss=0.187, ctc_loss=0.1247, cr_loss=0.3117, over 17204.00 frames. ], tot_loss[loss=0.2054, ctc_loss=0.1349, cr_loss=0.3527, over 3366751.37 frames. ], batch size: 50, lr: 4.57e-03, grad_scale: 32.0 2024-09-24 09:20:02,797 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=8.62 vs. limit=15.0 2024-09-24 09:20:04,337 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=3.69 vs. limit=15.0 2024-09-24 09:20:11,563 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=469518.0, ans=0.0 2024-09-24 09:20:16,344 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=469518.0, ans=0.125 2024-09-24 09:20:28,897 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=469564.6666666667, ans=0.125 2024-09-24 09:20:44,065 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.082e+02 1.250e+02 1.395e+02 1.507e+02 2.152e+02, threshold=2.790e+02, percent-clipped=0.0 2024-09-24 09:21:11,324 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.76 vs. limit=6.0 2024-09-24 09:21:15,229 INFO [train.py:1198] (2/4) Epoch 26, batch 3250, loss[loss=0.2038, ctc_loss=0.1336, cr_loss=0.351, over 17030.00 frames. ], tot_loss[loss=0.2056, ctc_loss=0.1351, cr_loss=0.3526, over 3356733.55 frames. ], batch size: 44, lr: 4.57e-03, grad_scale: 16.0 2024-09-24 09:21:42,006 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=469751.3333333333, ans=0.1 2024-09-24 09:21:56,192 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=469798.0, ans=0.125 2024-09-24 09:22:01,048 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=469844.6666666667, ans=0.125 2024-09-24 09:22:15,993 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=6.36 vs. limit=15.0 2024-09-24 09:22:35,465 INFO [train.py:1198] (2/4) Epoch 26, batch 3300, loss[loss=0.2249, ctc_loss=0.1488, cr_loss=0.3801, over 17228.00 frames. ], tot_loss[loss=0.2049, ctc_loss=0.1345, cr_loss=0.3521, over 3367196.99 frames. ], batch size: 55, lr: 4.57e-03, grad_scale: 16.0 2024-09-24 09:22:46,842 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=469938.0, ans=0.0 2024-09-24 09:23:03,760 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=469984.6666666667, ans=0.125 2024-09-24 09:23:06,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=470031.3333333333, ans=0.0 2024-09-24 09:23:16,450 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 09:23:24,578 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.271e+02 1.364e+02 1.523e+02 2.164e+02, threshold=2.728e+02, percent-clipped=0.0 2024-09-24 09:23:26,477 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=470078.0, ans=0.2 2024-09-24 09:23:34,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=470078.0, ans=0.025 2024-09-24 09:23:42,022 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=470124.6666666667, ans=0.125 2024-09-24 09:23:45,017 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=470124.6666666667, ans=0.0 2024-09-24 09:23:55,639 INFO [train.py:1198] (2/4) Epoch 26, batch 3350, loss[loss=0.22, ctc_loss=0.1444, cr_loss=0.3782, over 17028.00 frames. ], tot_loss[loss=0.2054, ctc_loss=0.1348, cr_loss=0.3528, over 3369175.61 frames. ], batch size: 56, lr: 4.57e-03, grad_scale: 16.0 2024-09-24 09:24:58,699 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=470358.0, ans=0.125 2024-09-24 09:25:06,627 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=470358.0, ans=0.125 2024-09-24 09:25:14,414 INFO [train.py:1198] (2/4) Epoch 26, batch 3400, loss[loss=0.2211, ctc_loss=0.1452, cr_loss=0.3793, over 16993.00 frames. ], tot_loss[loss=0.2062, ctc_loss=0.1354, cr_loss=0.3539, over 3366105.02 frames. ], batch size: 53, lr: 4.56e-03, grad_scale: 16.0 2024-09-24 09:25:22,620 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=470404.6666666667, ans=0.125 2024-09-24 09:25:44,132 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=13.72 vs. limit=15.0 2024-09-24 09:25:54,014 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=470498.0, ans=0.0 2024-09-24 09:26:00,244 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=470544.6666666667, ans=0.1 2024-09-24 09:26:01,626 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.254e+02 1.315e+02 1.422e+02 2.277e+02, threshold=2.630e+02, percent-clipped=0.0 2024-09-24 09:26:16,184 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.59 vs. limit=15.0 2024-09-24 09:26:32,734 INFO [train.py:1198] (2/4) Epoch 26, batch 3450, loss[loss=0.2111, ctc_loss=0.1409, cr_loss=0.3512, over 17146.00 frames. ], tot_loss[loss=0.2061, ctc_loss=0.1354, cr_loss=0.3535, over 3367329.35 frames. ], batch size: 48, lr: 4.56e-03, grad_scale: 16.0 2024-09-24 09:26:45,619 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=470638.0, ans=0.125 2024-09-24 09:27:10,739 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=470731.3333333333, ans=0.125 2024-09-24 09:27:33,037 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.87 vs. limit=10.0 2024-09-24 09:27:34,140 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=470824.6666666667, ans=0.0 2024-09-24 09:27:53,197 INFO [train.py:1198] (2/4) Epoch 26, batch 3500, loss[loss=0.2351, ctc_loss=0.1576, cr_loss=0.3875, over 14958.00 frames. ], tot_loss[loss=0.205, ctc_loss=0.1347, cr_loss=0.3516, over 3361969.58 frames. ], batch size: 89, lr: 4.56e-03, grad_scale: 16.0 2024-09-24 09:28:05,971 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=470871.3333333333, ans=0.07 2024-09-24 09:28:10,745 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=470918.0, ans=0.125 2024-09-24 09:28:39,878 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.160e+02 1.277e+02 1.357e+02 1.514e+02 2.797e+02, threshold=2.715e+02, percent-clipped=1.0 2024-09-24 09:28:51,190 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 09:29:03,958 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=471058.0, ans=0.1 2024-09-24 09:29:08,520 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=471058.0, ans=0.2 2024-09-24 09:29:11,280 INFO [train.py:1198] (2/4) Epoch 26, batch 3550, loss[loss=0.2566, ctc_loss=0.176, cr_loss=0.4031, over 11957.00 frames. ], tot_loss[loss=0.2055, ctc_loss=0.135, cr_loss=0.3524, over 3356321.65 frames. ], batch size: 124, lr: 4.56e-03, grad_scale: 16.0 2024-09-24 09:29:12,051 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=13.64 vs. limit=22.5 2024-09-24 09:29:43,310 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=471198.0, ans=0.1 2024-09-24 09:29:54,187 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=471198.0, ans=0.0 2024-09-24 09:29:55,610 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=471198.0, ans=0.2 2024-09-24 09:29:57,212 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=471198.0, ans=0.04949747468305833 2024-09-24 09:30:18,287 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.88 vs. limit=6.0 2024-09-24 09:30:21,024 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=471291.3333333333, ans=0.125 2024-09-24 09:30:32,069 INFO [train.py:1198] (2/4) Epoch 26, batch 3600, loss[loss=0.198, ctc_loss=0.1335, cr_loss=0.3225, over 17035.00 frames. ], tot_loss[loss=0.2047, ctc_loss=0.1343, cr_loss=0.3518, over 3363355.17 frames. ], batch size: 56, lr: 4.56e-03, grad_scale: 32.0 2024-09-24 09:30:47,731 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=471384.6666666667, ans=0.125 2024-09-24 09:30:58,689 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 09:31:03,369 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=471431.3333333333, ans=0.1 2024-09-24 09:31:06,611 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.25 vs. limit=15.0 2024-09-24 09:31:12,551 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=471431.3333333333, ans=0.125 2024-09-24 09:31:14,267 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=471431.3333333333, ans=0.125 2024-09-24 09:31:17,171 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=471478.0, ans=0.025 2024-09-24 09:31:17,719 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.40 vs. limit=12.0 2024-09-24 09:31:20,017 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.297e+02 1.443e+02 1.634e+02 2.383e+02, threshold=2.886e+02, percent-clipped=0.0 2024-09-24 09:31:21,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=471478.0, ans=0.2 2024-09-24 09:31:49,546 INFO [train.py:1198] (2/4) Epoch 26, batch 3650, loss[loss=0.2008, ctc_loss=0.1299, cr_loss=0.3544, over 17149.00 frames. ], tot_loss[loss=0.2038, ctc_loss=0.1337, cr_loss=0.3506, over 3363588.09 frames. ], batch size: 45, lr: 4.56e-03, grad_scale: 16.0 2024-09-24 09:31:55,899 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=471571.3333333333, ans=0.07 2024-09-24 09:32:00,859 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=471571.3333333333, ans=0.0 2024-09-24 09:32:29,259 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 09:33:05,278 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.59 vs. limit=15.0 2024-09-24 09:33:12,494 INFO [train.py:1198] (2/4) Epoch 26, batch 3700, loss[loss=0.1668, ctc_loss=0.1042, cr_loss=0.3131, over 17188.00 frames. ], tot_loss[loss=0.2039, ctc_loss=0.1337, cr_loss=0.3512, over 3367960.77 frames. ], batch size: 41, lr: 4.56e-03, grad_scale: 16.0 2024-09-24 09:33:18,932 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=471804.6666666667, ans=0.125 2024-09-24 09:33:20,665 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=471804.6666666667, ans=0.0 2024-09-24 09:34:01,194 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.253e+02 1.341e+02 1.434e+02 1.966e+02, threshold=2.682e+02, percent-clipped=0.0 2024-09-24 09:34:18,769 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=471991.3333333333, ans=0.125 2024-09-24 09:34:20,296 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=471991.3333333333, ans=0.0 2024-09-24 09:34:30,806 INFO [train.py:1198] (2/4) Epoch 26, batch 3750, loss[loss=0.1969, ctc_loss=0.1316, cr_loss=0.3267, over 17326.00 frames. ], tot_loss[loss=0.2044, ctc_loss=0.1342, cr_loss=0.3514, over 3359504.78 frames. ], batch size: 52, lr: 4.56e-03, grad_scale: 16.0 2024-09-24 09:34:55,456 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=472084.6666666667, ans=0.125 2024-09-24 09:35:12,215 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=472131.3333333333, ans=0.0 2024-09-24 09:35:36,485 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=472224.6666666667, ans=0.125 2024-09-24 09:35:47,118 INFO [train.py:1198] (2/4) Epoch 26, batch 3800, loss[loss=0.2076, ctc_loss=0.1369, cr_loss=0.3534, over 17044.00 frames. ], tot_loss[loss=0.2051, ctc_loss=0.1348, cr_loss=0.3515, over 3350415.50 frames. ], batch size: 52, lr: 4.56e-03, grad_scale: 16.0 2024-09-24 09:36:00,352 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.whiten, num_groups=1, num_channels=192, metric=3.85 vs. limit=12.0 2024-09-24 09:36:11,763 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 09:36:18,050 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.82 vs. limit=15.0 2024-09-24 09:36:34,126 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.081e+02 1.296e+02 1.371e+02 1.525e+02 1.900e+02, threshold=2.741e+02, percent-clipped=0.0 2024-09-24 09:36:46,892 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=472458.0, ans=0.0 2024-09-24 09:37:03,758 INFO [train.py:1198] (2/4) Epoch 26, batch 3850, loss[loss=0.1902, ctc_loss=0.125, cr_loss=0.326, over 17017.00 frames. ], tot_loss[loss=0.2083, ctc_loss=0.1373, cr_loss=0.3549, over 3306480.11 frames. ], batch size: 44, lr: 4.55e-03, grad_scale: 16.0 2024-09-24 09:37:11,665 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=472504.6666666667, ans=0.0 2024-09-24 09:37:14,716 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=472504.6666666667, ans=0.0 2024-09-24 09:37:24,130 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.73 vs. limit=22.5 2024-09-24 09:38:01,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=472644.6666666667, ans=0.125 2024-09-24 09:38:03,388 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=472691.3333333333, ans=0.2 2024-09-24 09:39:03,326 INFO [train.py:1198] (2/4) Epoch 27, batch 0, loss[loss=0.2029, ctc_loss=0.1327, cr_loss=0.351, over 17009.00 frames. ], tot_loss[loss=0.2029, ctc_loss=0.1327, cr_loss=0.351, over 17009.00 frames. ], batch size: 56, lr: 4.47e-03, grad_scale: 32.0 2024-09-24 09:39:03,327 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 09:39:21,541 INFO [train.py:1230] (2/4) Epoch 27, validation: loss=0.03741, ctc_loss=0.03741, cr_loss=8.388e-15, over 944034.00 frames. 2024-09-24 09:39:21,542 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 09:39:24,879 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=472719.3333333333, ans=0.04949747468305833 2024-09-24 09:39:28,608 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.88 vs. limit=15.0 2024-09-24 09:40:04,287 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=472812.6666666667, ans=0.025 2024-09-24 09:40:05,699 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=472812.6666666667, ans=0.0 2024-09-24 09:40:21,657 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.122e+02 1.362e+02 1.530e+02 1.663e+02 2.257e+02, threshold=3.060e+02, percent-clipped=0.0 2024-09-24 09:40:39,731 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=472906.0, ans=0.2 2024-09-24 09:40:44,916 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=16.76 vs. limit=22.5 2024-09-24 09:40:45,608 INFO [train.py:1198] (2/4) Epoch 27, batch 50, loss[loss=0.2042, ctc_loss=0.1351, cr_loss=0.3451, over 17297.00 frames. ], tot_loss[loss=0.2066, ctc_loss=0.1363, cr_loss=0.3519, over 749734.08 frames. ], batch size: 46, lr: 4.47e-03, grad_scale: 32.0 2024-09-24 09:40:46,216 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=11.47 vs. limit=15.0 2024-09-24 09:40:52,148 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=472952.6666666667, ans=0.125 2024-09-24 09:41:12,229 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=472999.3333333333, ans=0.125 2024-09-24 09:41:53,050 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.33 vs. limit=15.0 2024-09-24 09:41:54,134 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=473139.3333333333, ans=0.125 2024-09-24 09:41:58,852 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=473139.3333333333, ans=0.1 2024-09-24 09:42:00,323 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=473139.3333333333, ans=0.1 2024-09-24 09:42:04,847 INFO [train.py:1198] (2/4) Epoch 27, batch 100, loss[loss=0.2135, ctc_loss=0.1423, cr_loss=0.3561, over 17072.00 frames. ], tot_loss[loss=0.205, ctc_loss=0.1349, cr_loss=0.3502, over 1326366.61 frames. ], batch size: 46, lr: 4.46e-03, grad_scale: 32.0 2024-09-24 09:42:11,134 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=13.14 vs. limit=22.5 2024-09-24 09:42:16,616 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=473186.0, ans=0.0 2024-09-24 09:42:17,295 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.31 vs. limit=15.0 2024-09-24 09:42:24,850 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=473232.6666666667, ans=0.125 2024-09-24 09:42:45,047 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.min_positive, batch_count=473279.3333333333, ans=0.025 2024-09-24 09:42:56,150 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=473326.0, ans=0.125 2024-09-24 09:43:03,945 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.054e+02 1.215e+02 1.307e+02 1.417e+02 1.891e+02, threshold=2.615e+02, percent-clipped=0.0 2024-09-24 09:43:13,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=473372.6666666667, ans=0.0 2024-09-24 09:43:28,090 INFO [train.py:1198] (2/4) Epoch 27, batch 150, loss[loss=0.2334, ctc_loss=0.1542, cr_loss=0.3957, over 17019.00 frames. ], tot_loss[loss=0.2029, ctc_loss=0.1331, cr_loss=0.3488, over 1788487.78 frames. ], batch size: 53, lr: 4.46e-03, grad_scale: 32.0 2024-09-24 09:43:37,875 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=473419.3333333333, ans=0.025 2024-09-24 09:43:48,318 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=4.76 vs. limit=15.0 2024-09-24 09:44:10,828 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=3.82 vs. limit=10.0 2024-09-24 09:44:11,978 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=473512.6666666667, ans=0.1 2024-09-24 09:44:15,164 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=473559.3333333333, ans=0.1 2024-09-24 09:44:29,079 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=473559.3333333333, ans=0.1 2024-09-24 09:44:53,617 INFO [train.py:1198] (2/4) Epoch 27, batch 200, loss[loss=0.232, ctc_loss=0.157, cr_loss=0.3748, over 16450.00 frames. ], tot_loss[loss=0.2045, ctc_loss=0.1344, cr_loss=0.3505, over 2127351.53 frames. ], batch size: 66, lr: 4.46e-03, grad_scale: 32.0 2024-09-24 09:45:46,450 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=473792.6666666667, ans=0.0 2024-09-24 09:45:51,083 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=473792.6666666667, ans=0.125 2024-09-24 09:45:51,111 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=473792.6666666667, ans=0.07 2024-09-24 09:45:52,247 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.169e+02 1.251e+02 1.322e+02 1.422e+02 2.046e+02, threshold=2.645e+02, percent-clipped=0.0 2024-09-24 09:46:02,151 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=473839.3333333333, ans=0.025 2024-09-24 09:46:13,522 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten.whitening_limit, batch_count=473839.3333333333, ans=15.0 2024-09-24 09:46:16,197 INFO [train.py:1198] (2/4) Epoch 27, batch 250, loss[loss=0.1611, ctc_loss=0.1028, cr_loss=0.2914, over 16335.00 frames. ], tot_loss[loss=0.2035, ctc_loss=0.1336, cr_loss=0.3494, over 2401065.75 frames. ], batch size: 36, lr: 4.46e-03, grad_scale: 32.0 2024-09-24 09:46:43,527 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=473932.6666666667, ans=0.0 2024-09-24 09:46:43,868 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.67 vs. limit=15.0 2024-09-24 09:46:49,775 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=473979.3333333333, ans=0.125 2024-09-24 09:47:07,426 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.min_positive, batch_count=474026.0, ans=0.025 2024-09-24 09:47:20,113 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=474072.6666666667, ans=0.0 2024-09-24 09:47:21,753 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=474072.6666666667, ans=0.1 2024-09-24 09:47:38,636 INFO [train.py:1198] (2/4) Epoch 27, batch 300, loss[loss=0.1898, ctc_loss=0.1225, cr_loss=0.3365, over 17293.00 frames. ], tot_loss[loss=0.2035, ctc_loss=0.1336, cr_loss=0.3496, over 2623004.24 frames. ], batch size: 51, lr: 4.46e-03, grad_scale: 32.0 2024-09-24 09:47:40,621 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=474119.3333333333, ans=0.125 2024-09-24 09:47:40,752 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 09:47:52,110 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=474119.3333333333, ans=0.0 2024-09-24 09:47:53,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=474166.0, ans=0.0 2024-09-24 09:48:00,340 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=474166.0, ans=0.125 2024-09-24 09:48:05,426 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=474166.0, ans=0.125 2024-09-24 09:48:27,758 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=474259.3333333333, ans=0.125 2024-09-24 09:48:34,664 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.55 vs. limit=15.0 2024-09-24 09:48:35,476 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.098e+02 1.322e+02 1.414e+02 1.594e+02 2.687e+02, threshold=2.828e+02, percent-clipped=1.0 2024-09-24 09:48:50,822 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=8.54 vs. limit=15.0 2024-09-24 09:48:51,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=474306.0, ans=0.0 2024-09-24 09:48:59,274 INFO [train.py:1198] (2/4) Epoch 27, batch 350, loss[loss=0.2356, ctc_loss=0.157, cr_loss=0.3926, over 17002.00 frames. ], tot_loss[loss=0.204, ctc_loss=0.134, cr_loss=0.3499, over 2774335.02 frames. ], batch size: 53, lr: 4.46e-03, grad_scale: 32.0 2024-09-24 09:49:21,376 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=474399.3333333333, ans=0.2 2024-09-24 09:49:33,191 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=13.60 vs. limit=15.0 2024-09-24 09:49:35,532 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=474446.0, ans=0.0 2024-09-24 09:49:53,090 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=10.78 vs. limit=15.0 2024-09-24 09:49:54,077 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=474492.6666666667, ans=0.125 2024-09-24 09:50:01,089 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.74 vs. limit=10.0 2024-09-24 09:50:23,400 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.89 vs. limit=12.0 2024-09-24 09:50:27,284 INFO [train.py:1198] (2/4) Epoch 27, batch 400, loss[loss=0.1896, ctc_loss=0.1219, cr_loss=0.3387, over 17221.00 frames. ], tot_loss[loss=0.2034, ctc_loss=0.1334, cr_loss=0.3496, over 2915208.37 frames. ], batch size: 47, lr: 4.46e-03, grad_scale: 32.0 2024-09-24 09:50:30,944 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=474586.0, ans=0.0 2024-09-24 09:50:34,416 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten.whitening_limit, batch_count=474586.0, ans=15.0 2024-09-24 09:50:40,753 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=6.65 vs. limit=15.0 2024-09-24 09:51:20,366 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=474726.0, ans=0.125 2024-09-24 09:51:23,172 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.110e+02 1.248e+02 1.347e+02 1.469e+02 2.188e+02, threshold=2.694e+02, percent-clipped=0.0 2024-09-24 09:51:28,461 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=474726.0, ans=0.025 2024-09-24 09:51:31,673 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=474772.6666666667, ans=0.125 2024-09-24 09:51:38,189 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=474772.6666666667, ans=0.2 2024-09-24 09:51:47,497 INFO [train.py:1198] (2/4) Epoch 27, batch 450, loss[loss=0.2176, ctc_loss=0.1404, cr_loss=0.3861, over 17301.00 frames. ], tot_loss[loss=0.2033, ctc_loss=0.1333, cr_loss=0.3497, over 3019274.81 frames. ], batch size: 49, lr: 4.46e-03, grad_scale: 32.0 2024-09-24 09:52:07,121 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=3.88 vs. limit=10.0 2024-09-24 09:52:16,200 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=474866.0, ans=0.0 2024-09-24 09:52:25,992 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=474912.6666666667, ans=0.1 2024-09-24 09:52:51,011 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.min_abs, batch_count=474959.3333333333, ans=0.5 2024-09-24 09:52:59,444 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.16 vs. limit=22.5 2024-09-24 09:53:09,777 INFO [train.py:1198] (2/4) Epoch 27, batch 500, loss[loss=0.1519, ctc_loss=0.09942, cr_loss=0.2623, over 17035.00 frames. ], tot_loss[loss=0.2027, ctc_loss=0.133, cr_loss=0.3486, over 3084446.40 frames. ], batch size: 39, lr: 4.46e-03, grad_scale: 32.0 2024-09-24 09:53:34,572 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=475099.3333333333, ans=0.125 2024-09-24 09:54:06,519 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.150e+02 1.284e+02 1.368e+02 1.499e+02 2.424e+02, threshold=2.736e+02, percent-clipped=0.0 2024-09-24 09:54:23,825 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=475239.3333333333, ans=0.0 2024-09-24 09:54:33,116 INFO [train.py:1198] (2/4) Epoch 27, batch 550, loss[loss=0.1986, ctc_loss=0.1294, cr_loss=0.3458, over 17007.00 frames. ], tot_loss[loss=0.2038, ctc_loss=0.1337, cr_loss=0.3505, over 3152146.21 frames. ], batch size: 51, lr: 4.45e-03, grad_scale: 32.0 2024-09-24 09:54:48,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=475286.0, ans=0.125 2024-09-24 09:54:54,948 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=475332.6666666667, ans=0.125 2024-09-24 09:54:54,970 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=475332.6666666667, ans=0.125 2024-09-24 09:55:14,559 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.38 vs. limit=15.0 2024-09-24 09:55:45,824 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.71 vs. limit=10.0 2024-09-24 09:55:50,229 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=475472.6666666667, ans=0.5 2024-09-24 09:55:57,941 INFO [train.py:1198] (2/4) Epoch 27, batch 600, loss[loss=0.2216, ctc_loss=0.1451, cr_loss=0.3822, over 16756.00 frames. ], tot_loss[loss=0.2043, ctc_loss=0.1339, cr_loss=0.3519, over 3208309.10 frames. ], batch size: 61, lr: 4.45e-03, grad_scale: 32.0 2024-09-24 09:56:17,268 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=475566.0, ans=0.125 2024-09-24 09:56:20,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=475566.0, ans=0.0 2024-09-24 09:56:53,666 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.260e+02 1.326e+02 1.393e+02 1.864e+02, threshold=2.652e+02, percent-clipped=0.0 2024-09-24 09:57:07,809 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.95 vs. limit=6.0 2024-09-24 09:57:14,917 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=475706.0, ans=0.125 2024-09-24 09:57:17,880 INFO [train.py:1198] (2/4) Epoch 27, batch 650, loss[loss=0.2001, ctc_loss=0.1319, cr_loss=0.3411, over 17019.00 frames. ], tot_loss[loss=0.2059, ctc_loss=0.1351, cr_loss=0.3538, over 3248166.82 frames. ], batch size: 51, lr: 4.45e-03, grad_scale: 32.0 2024-09-24 09:57:44,331 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=475799.3333333333, ans=0.07 2024-09-24 09:57:46,458 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.84 vs. limit=15.0 2024-09-24 09:57:51,440 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn2.whiten, num_groups=1, num_channels=192, metric=8.02 vs. limit=22.5 2024-09-24 09:58:12,952 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=475892.6666666667, ans=0.1 2024-09-24 09:58:16,062 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=475892.6666666667, ans=0.125 2024-09-24 09:58:23,176 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=5.67 vs. limit=15.0 2024-09-24 09:58:25,640 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=475939.3333333333, ans=0.025 2024-09-24 09:58:39,736 INFO [train.py:1198] (2/4) Epoch 27, batch 700, loss[loss=0.2172, ctc_loss=0.1389, cr_loss=0.3914, over 17293.00 frames. ], tot_loss[loss=0.2059, ctc_loss=0.1352, cr_loss=0.3536, over 3272379.41 frames. ], batch size: 46, lr: 4.45e-03, grad_scale: 32.0 2024-09-24 09:58:49,899 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=475986.0, ans=0.125 2024-09-24 09:58:51,612 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=475986.0, ans=0.125 2024-09-24 09:59:09,417 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=11.76 vs. limit=15.0 2024-09-24 09:59:15,632 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=476079.3333333333, ans=0.125 2024-09-24 09:59:25,743 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=476079.3333333333, ans=0.025 2024-09-24 09:59:40,876 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.083e+02 1.256e+02 1.371e+02 1.501e+02 2.687e+02, threshold=2.742e+02, percent-clipped=1.0 2024-09-24 10:00:04,597 INFO [train.py:1198] (2/4) Epoch 27, batch 750, loss[loss=0.2118, ctc_loss=0.1404, cr_loss=0.3568, over 17001.00 frames. ], tot_loss[loss=0.2049, ctc_loss=0.1345, cr_loss=0.3518, over 3286263.64 frames. ], batch size: 56, lr: 4.45e-03, grad_scale: 32.0 2024-09-24 10:00:04,875 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 10:00:09,752 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=476219.3333333333, ans=0.125 2024-09-24 10:00:19,140 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=9.20 vs. limit=15.0 2024-09-24 10:00:28,557 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=476266.0, ans=0.125 2024-09-24 10:00:34,909 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=476266.0, ans=0.0 2024-09-24 10:00:41,904 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.19 vs. limit=15.0 2024-09-24 10:00:46,057 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=476312.6666666667, ans=0.1 2024-09-24 10:01:00,734 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=476359.3333333333, ans=0.0 2024-09-24 10:01:08,340 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=476359.3333333333, ans=0.125 2024-09-24 10:01:17,138 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=8.04 vs. limit=10.0 2024-09-24 10:01:18,073 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.61 vs. limit=22.5 2024-09-24 10:01:25,840 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=476452.6666666667, ans=0.05 2024-09-24 10:01:27,179 INFO [train.py:1198] (2/4) Epoch 27, batch 800, loss[loss=0.1565, ctc_loss=0.102, cr_loss=0.2726, over 17042.00 frames. ], tot_loss[loss=0.2044, ctc_loss=0.1341, cr_loss=0.3515, over 3299522.14 frames. ], batch size: 39, lr: 4.45e-03, grad_scale: 32.0 2024-09-24 10:02:06,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=476546.0, ans=0.09899494936611666 2024-09-24 10:02:21,067 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=476592.6666666667, ans=0.125 2024-09-24 10:02:22,820 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=476592.6666666667, ans=0.0 2024-09-24 10:02:23,974 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.060e+02 1.235e+02 1.338e+02 1.405e+02 1.662e+02, threshold=2.676e+02, percent-clipped=0.0 2024-09-24 10:02:29,420 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=476592.6666666667, ans=0.1 2024-09-24 10:02:43,167 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=476639.3333333333, ans=0.125 2024-09-24 10:02:48,018 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=476639.3333333333, ans=0.0 2024-09-24 10:02:50,859 INFO [train.py:1198] (2/4) Epoch 27, batch 850, loss[loss=0.1478, ctc_loss=0.09522, cr_loss=0.2627, over 17270.00 frames. ], tot_loss[loss=0.2034, ctc_loss=0.1333, cr_loss=0.3505, over 3320809.92 frames. ], batch size: 42, lr: 4.45e-03, grad_scale: 32.0 2024-09-24 10:02:57,503 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=476686.0, ans=0.125 2024-09-24 10:03:02,571 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.25 vs. limit=15.0 2024-09-24 10:04:02,764 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=476872.6666666667, ans=0.0 2024-09-24 10:04:10,718 INFO [train.py:1198] (2/4) Epoch 27, batch 900, loss[loss=0.2334, ctc_loss=0.1531, cr_loss=0.4015, over 17043.00 frames. ], tot_loss[loss=0.2036, ctc_loss=0.1335, cr_loss=0.3506, over 3339075.99 frames. ], batch size: 52, lr: 4.45e-03, grad_scale: 32.0 2024-09-24 10:04:16,761 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=476919.3333333333, ans=0.09899494936611666 2024-09-24 10:04:38,748 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=476966.0, ans=0.0 2024-09-24 10:04:40,284 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=476966.0, ans=0.1 2024-09-24 10:04:43,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=476966.0, ans=0.125 2024-09-24 10:05:14,403 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.081e+02 1.238e+02 1.345e+02 1.503e+02 3.181e+02, threshold=2.691e+02, percent-clipped=1.0 2024-09-24 10:05:16,382 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=477059.3333333333, ans=0.0 2024-09-24 10:05:38,931 INFO [train.py:1198] (2/4) Epoch 27, batch 950, loss[loss=0.2058, ctc_loss=0.1331, cr_loss=0.3635, over 17291.00 frames. ], tot_loss[loss=0.2035, ctc_loss=0.1333, cr_loss=0.3508, over 3350257.15 frames. ], batch size: 46, lr: 4.45e-03, grad_scale: 32.0 2024-09-24 10:06:16,348 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.77 vs. limit=15.0 2024-09-24 10:06:17,733 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=477246.0, ans=0.125 2024-09-24 10:06:18,090 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=4.42 vs. limit=12.0 2024-09-24 10:06:58,798 INFO [train.py:1198] (2/4) Epoch 27, batch 1000, loss[loss=0.1978, ctc_loss=0.1267, cr_loss=0.3553, over 17056.00 frames. ], tot_loss[loss=0.2044, ctc_loss=0.1341, cr_loss=0.3514, over 3339912.88 frames. ], batch size: 46, lr: 4.44e-03, grad_scale: 8.0 2024-09-24 10:07:02,289 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=477386.0, ans=0.09899494936611666 2024-09-24 10:07:59,668 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=477526.0, ans=0.025 2024-09-24 10:08:00,491 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=4.75 vs. limit=15.0 2024-09-24 10:08:00,887 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.228e+02 1.308e+02 1.402e+02 1.814e+02, threshold=2.617e+02, percent-clipped=0.0 2024-09-24 10:08:01,155 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=477526.0, ans=0.1 2024-09-24 10:08:01,583 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=9.51 vs. limit=15.0 2024-09-24 10:08:07,993 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=477572.6666666667, ans=0.125 2024-09-24 10:08:12,766 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=477572.6666666667, ans=0.07 2024-09-24 10:08:22,251 INFO [train.py:1198] (2/4) Epoch 27, batch 1050, loss[loss=0.1894, ctc_loss=0.1228, cr_loss=0.3333, over 17217.00 frames. ], tot_loss[loss=0.2045, ctc_loss=0.1342, cr_loss=0.3517, over 3349286.18 frames. ], batch size: 47, lr: 4.44e-03, grad_scale: 8.0 2024-09-24 10:08:27,277 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=477619.3333333333, ans=0.125 2024-09-24 10:08:28,948 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=477619.3333333333, ans=0.125 2024-09-24 10:08:41,881 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=477666.0, ans=0.2 2024-09-24 10:08:46,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=477666.0, ans=0.0 2024-09-24 10:09:03,591 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=16.11 vs. limit=22.5 2024-09-24 10:09:38,143 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=477806.0, ans=0.125 2024-09-24 10:09:43,626 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.70 vs. limit=15.0 2024-09-24 10:09:43,955 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=9.33 vs. limit=15.0 2024-09-24 10:09:44,673 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 10:09:47,484 INFO [train.py:1198] (2/4) Epoch 27, batch 1100, loss[loss=0.2112, ctc_loss=0.1435, cr_loss=0.3384, over 16760.00 frames. ], tot_loss[loss=0.2033, ctc_loss=0.1334, cr_loss=0.3498, over 3351952.07 frames. ], batch size: 61, lr: 4.44e-03, grad_scale: 8.0 2024-09-24 10:10:20,154 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=3.79 vs. limit=10.0 2024-09-24 10:10:21,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=477946.0, ans=0.1 2024-09-24 10:10:24,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=477946.0, ans=0.2 2024-09-24 10:10:36,127 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.83 vs. limit=15.0 2024-09-24 10:10:49,361 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.034e+02 1.279e+02 1.377e+02 1.520e+02 1.966e+02, threshold=2.755e+02, percent-clipped=0.0 2024-09-24 10:10:51,378 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=477992.6666666667, ans=0.0 2024-09-24 10:10:51,701 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.59 vs. limit=22.5 2024-09-24 10:10:52,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=478039.3333333333, ans=0.2 2024-09-24 10:11:01,059 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=478039.3333333333, ans=0.125 2024-09-24 10:11:10,470 INFO [train.py:1198] (2/4) Epoch 27, batch 1150, loss[loss=0.1852, ctc_loss=0.12, cr_loss=0.3257, over 17056.00 frames. ], tot_loss[loss=0.2033, ctc_loss=0.1333, cr_loss=0.3499, over 3357677.03 frames. ], batch size: 39, lr: 4.44e-03, grad_scale: 8.0 2024-09-24 10:11:17,241 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=478086.0, ans=0.1 2024-09-24 10:11:39,790 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=478132.6666666667, ans=0.2 2024-09-24 10:11:44,806 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.26 vs. limit=6.0 2024-09-24 10:12:24,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=478272.6666666667, ans=0.2 2024-09-24 10:12:32,237 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=478319.3333333333, ans=0.125 2024-09-24 10:12:33,336 INFO [train.py:1198] (2/4) Epoch 27, batch 1200, loss[loss=0.1812, ctc_loss=0.1155, cr_loss=0.3286, over 16237.00 frames. ], tot_loss[loss=0.2029, ctc_loss=0.1331, cr_loss=0.3493, over 3363692.45 frames. ], batch size: 36, lr: 4.44e-03, grad_scale: 16.0 2024-09-24 10:12:35,839 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=12.09 vs. limit=15.0 2024-09-24 10:12:43,287 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=478319.3333333333, ans=0.09899494936611666 2024-09-24 10:12:46,664 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 10:12:56,489 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=4.28 vs. limit=15.0 2024-09-24 10:13:14,456 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.whiten, num_groups=1, num_channels=192, metric=4.00 vs. limit=12.0 2024-09-24 10:13:18,801 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=11.71 vs. limit=22.5 2024-09-24 10:13:32,437 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.237e+02 1.313e+02 1.395e+02 2.791e+02, threshold=2.627e+02, percent-clipped=1.0 2024-09-24 10:13:53,064 INFO [train.py:1198] (2/4) Epoch 27, batch 1250, loss[loss=0.1831, ctc_loss=0.1186, cr_loss=0.3225, over 17264.00 frames. ], tot_loss[loss=0.203, ctc_loss=0.1331, cr_loss=0.3494, over 3358068.81 frames. ], batch size: 44, lr: 4.44e-03, grad_scale: 16.0 2024-09-24 10:14:21,125 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.86 vs. limit=6.0 2024-09-24 10:14:59,975 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=478692.6666666667, ans=0.1 2024-09-24 10:15:19,936 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=478786.0, ans=0.125 2024-09-24 10:15:21,232 INFO [train.py:1198] (2/4) Epoch 27, batch 1300, loss[loss=0.1953, ctc_loss=0.1273, cr_loss=0.34, over 17010.00 frames. ], tot_loss[loss=0.2038, ctc_loss=0.1337, cr_loss=0.3508, over 3361562.31 frames. ], batch size: 44, lr: 4.44e-03, grad_scale: 16.0 2024-09-24 10:15:24,871 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.03 vs. limit=15.0 2024-09-24 10:15:55,783 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=17.98 vs. limit=22.5 2024-09-24 10:16:09,950 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=478926.0, ans=0.09899494936611666 2024-09-24 10:16:15,272 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.06 vs. limit=6.0 2024-09-24 10:16:22,438 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.264e+02 1.365e+02 1.488e+02 1.950e+02, threshold=2.729e+02, percent-clipped=0.0 2024-09-24 10:16:35,674 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=478972.6666666667, ans=0.125 2024-09-24 10:16:37,345 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=478972.6666666667, ans=0.1 2024-09-24 10:16:39,005 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=478972.6666666667, ans=0.2 2024-09-24 10:16:41,760 INFO [train.py:1198] (2/4) Epoch 27, batch 1350, loss[loss=0.2502, ctc_loss=0.1674, cr_loss=0.414, over 17011.00 frames. ], tot_loss[loss=0.2035, ctc_loss=0.1334, cr_loss=0.3506, over 3371215.16 frames. ], batch size: 56, lr: 4.44e-03, grad_scale: 8.0 2024-09-24 10:16:44,105 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.44 vs. limit=15.0 2024-09-24 10:17:25,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=479112.6666666667, ans=0.2 2024-09-24 10:17:40,665 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=479159.3333333333, ans=0.1 2024-09-24 10:17:41,615 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=8.88 vs. limit=10.0 2024-09-24 10:17:48,664 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=479206.0, ans=0.125 2024-09-24 10:17:50,284 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=479206.0, ans=0.0 2024-09-24 10:18:01,449 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=479206.0, ans=0.125 2024-09-24 10:18:04,479 INFO [train.py:1198] (2/4) Epoch 27, batch 1400, loss[loss=0.2275, ctc_loss=0.1517, cr_loss=0.3792, over 16923.00 frames. ], tot_loss[loss=0.2042, ctc_loss=0.1339, cr_loss=0.3515, over 3377882.07 frames. ], batch size: 58, lr: 4.44e-03, grad_scale: 8.0 2024-09-24 10:18:22,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=479299.3333333333, ans=0.2 2024-09-24 10:18:30,666 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=479299.3333333333, ans=0.0 2024-09-24 10:18:43,552 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=479346.0, ans=0.07 2024-09-24 10:18:59,913 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.35 vs. limit=10.0 2024-09-24 10:19:01,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=479392.6666666667, ans=0.125 2024-09-24 10:19:04,182 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=479392.6666666667, ans=0.0 2024-09-24 10:19:08,018 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.080e+02 1.257e+02 1.359e+02 1.482e+02 2.377e+02, threshold=2.717e+02, percent-clipped=0.0 2024-09-24 10:19:13,150 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=479439.3333333333, ans=0.125 2024-09-24 10:19:14,579 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=479439.3333333333, ans=0.0 2024-09-24 10:19:24,732 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=18.95 vs. limit=22.5 2024-09-24 10:19:27,140 INFO [train.py:1198] (2/4) Epoch 27, batch 1450, loss[loss=0.1903, ctc_loss=0.1228, cr_loss=0.3371, over 17272.00 frames. ], tot_loss[loss=0.2035, ctc_loss=0.1334, cr_loss=0.3504, over 3379764.29 frames. ], batch size: 44, lr: 4.43e-03, grad_scale: 8.0 2024-09-24 10:19:28,986 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=479486.0, ans=0.2 2024-09-24 10:19:32,287 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=479486.0, ans=0.0 2024-09-24 10:19:36,273 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=479486.0, ans=0.0 2024-09-24 10:19:52,261 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-24 10:19:55,293 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=479532.6666666667, ans=0.0 2024-09-24 10:20:00,190 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=479579.3333333333, ans=0.2 2024-09-24 10:20:36,705 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=479672.6666666667, ans=0.0 2024-09-24 10:20:36,710 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=479672.6666666667, ans=0.125 2024-09-24 10:20:52,645 INFO [train.py:1198] (2/4) Epoch 27, batch 1500, loss[loss=0.2046, ctc_loss=0.1317, cr_loss=0.3647, over 17160.00 frames. ], tot_loss[loss=0.2039, ctc_loss=0.1336, cr_loss=0.3512, over 3371362.28 frames. ], batch size: 48, lr: 4.43e-03, grad_scale: 8.0 2024-09-24 10:21:54,047 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.112e+02 1.289e+02 1.371e+02 1.496e+02 2.046e+02, threshold=2.742e+02, percent-clipped=0.0 2024-09-24 10:22:03,611 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=479906.0, ans=0.125 2024-09-24 10:22:13,081 INFO [train.py:1198] (2/4) Epoch 27, batch 1550, loss[loss=0.1785, ctc_loss=0.1132, cr_loss=0.3267, over 17198.00 frames. ], tot_loss[loss=0.2035, ctc_loss=0.1333, cr_loss=0.3508, over 3378665.30 frames. ], batch size: 47, lr: 4.43e-03, grad_scale: 8.0 2024-09-24 10:23:12,759 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.52 vs. limit=15.0 2024-09-24 10:23:15,663 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=17.72 vs. limit=22.5 2024-09-24 10:23:23,431 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=480139.3333333333, ans=0.1 2024-09-24 10:23:35,000 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=480186.0, ans=0.2 2024-09-24 10:23:36,295 INFO [train.py:1198] (2/4) Epoch 27, batch 1600, loss[loss=0.1889, ctc_loss=0.121, cr_loss=0.3393, over 17068.00 frames. ], tot_loss[loss=0.2038, ctc_loss=0.1337, cr_loss=0.3507, over 3368982.63 frames. ], batch size: 39, lr: 4.43e-03, grad_scale: 16.0 2024-09-24 10:24:11,698 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.whiten, num_groups=1, num_channels=512, metric=4.93 vs. limit=12.0 2024-09-24 10:24:25,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=480326.0, ans=0.2 2024-09-24 10:24:40,461 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=480326.0, ans=0.035 2024-09-24 10:24:41,899 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.022e+02 1.244e+02 1.330e+02 1.456e+02 2.026e+02, threshold=2.660e+02, percent-clipped=0.0 2024-09-24 10:24:53,167 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=480372.6666666667, ans=0.0 2024-09-24 10:25:03,406 INFO [train.py:1198] (2/4) Epoch 27, batch 1650, loss[loss=0.2204, ctc_loss=0.1468, cr_loss=0.3681, over 16883.00 frames. ], tot_loss[loss=0.204, ctc_loss=0.1339, cr_loss=0.3508, over 3354066.44 frames. ], batch size: 58, lr: 4.43e-03, grad_scale: 16.0 2024-09-24 10:25:32,295 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=480466.0, ans=0.125 2024-09-24 10:25:34,020 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=480512.6666666667, ans=0.125 2024-09-24 10:25:41,934 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=480512.6666666667, ans=0.2 2024-09-24 10:25:56,613 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=480559.3333333333, ans=0.0 2024-09-24 10:25:58,897 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=13.97 vs. limit=22.5 2024-09-24 10:26:00,169 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.26 vs. limit=15.0 2024-09-24 10:26:16,507 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=10.52 vs. limit=22.5 2024-09-24 10:26:23,697 INFO [train.py:1198] (2/4) Epoch 27, batch 1700, loss[loss=0.2026, ctc_loss=0.1324, cr_loss=0.3508, over 17093.00 frames. ], tot_loss[loss=0.204, ctc_loss=0.1337, cr_loss=0.3514, over 3350394.64 frames. ], batch size: 49, lr: 4.43e-03, grad_scale: 16.0 2024-09-24 10:26:29,438 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=12.11 vs. limit=15.0 2024-09-24 10:26:36,968 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=480652.6666666667, ans=0.0 2024-09-24 10:26:46,399 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=480699.3333333333, ans=0.1 2024-09-24 10:26:51,321 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=480699.3333333333, ans=0.0 2024-09-24 10:26:57,703 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=480746.0, ans=0.125 2024-09-24 10:27:04,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=480746.0, ans=0.125 2024-09-24 10:27:04,588 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=512, metric=4.81 vs. limit=15.0 2024-09-24 10:27:26,754 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.079e+02 1.231e+02 1.318e+02 1.430e+02 1.905e+02, threshold=2.636e+02, percent-clipped=0.0 2024-09-24 10:27:27,606 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.53 vs. limit=6.0 2024-09-24 10:27:30,295 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=480839.3333333333, ans=0.1 2024-09-24 10:27:45,874 INFO [train.py:1198] (2/4) Epoch 27, batch 1750, loss[loss=0.1773, ctc_loss=0.1162, cr_loss=0.3053, over 16299.00 frames. ], tot_loss[loss=0.2029, ctc_loss=0.133, cr_loss=0.3494, over 3347277.13 frames. ], batch size: 36, lr: 4.43e-03, grad_scale: 16.0 2024-09-24 10:27:46,222 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=480886.0, ans=0.95 2024-09-24 10:27:49,351 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=480886.0, ans=0.125 2024-09-24 10:28:13,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=480932.6666666667, ans=0.125 2024-09-24 10:28:47,085 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=481026.0, ans=0.1 2024-09-24 10:28:47,156 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=481026.0, ans=0.0 2024-09-24 10:28:50,415 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=481072.6666666667, ans=0.1 2024-09-24 10:29:07,562 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=481119.3333333333, ans=0.125 2024-09-24 10:29:08,830 INFO [train.py:1198] (2/4) Epoch 27, batch 1800, loss[loss=0.2204, ctc_loss=0.1486, cr_loss=0.3592, over 17345.00 frames. ], tot_loss[loss=0.2035, ctc_loss=0.1335, cr_loss=0.3504, over 3347435.13 frames. ], batch size: 48, lr: 4.43e-03, grad_scale: 16.0 2024-09-24 10:29:09,043 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=481119.3333333333, ans=0.0 2024-09-24 10:30:14,559 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.138e+02 1.286e+02 1.379e+02 1.526e+02 2.061e+02, threshold=2.757e+02, percent-clipped=0.0 2024-09-24 10:30:21,253 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=481306.0, ans=0.125 2024-09-24 10:30:32,946 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=17.19 vs. limit=22.5 2024-09-24 10:30:33,713 INFO [train.py:1198] (2/4) Epoch 27, batch 1850, loss[loss=0.1758, ctc_loss=0.1125, cr_loss=0.3168, over 17087.00 frames. ], tot_loss[loss=0.2037, ctc_loss=0.1336, cr_loss=0.3507, over 3352489.51 frames. ], batch size: 43, lr: 4.43e-03, grad_scale: 16.0 2024-09-24 10:30:37,729 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.05 vs. limit=15.0 2024-09-24 10:31:28,808 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=14.42 vs. limit=22.5 2024-09-24 10:31:31,638 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=481492.6666666667, ans=0.125 2024-09-24 10:31:41,107 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.min_positive, batch_count=481539.3333333333, ans=0.025 2024-09-24 10:31:45,848 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=481539.3333333333, ans=0.125 2024-09-24 10:31:53,594 INFO [train.py:1198] (2/4) Epoch 27, batch 1900, loss[loss=0.2036, ctc_loss=0.1359, cr_loss=0.3384, over 16816.00 frames. ], tot_loss[loss=0.2037, ctc_loss=0.1336, cr_loss=0.3503, over 3347139.72 frames. ], batch size: 61, lr: 4.43e-03, grad_scale: 16.0 2024-09-24 10:31:57,163 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=481586.0, ans=0.125 2024-09-24 10:32:56,253 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=481726.0, ans=0.0 2024-09-24 10:32:57,446 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.257e+02 1.350e+02 1.464e+02 2.291e+02, threshold=2.700e+02, percent-clipped=0.0 2024-09-24 10:33:01,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=481772.6666666667, ans=0.1 2024-09-24 10:33:10,677 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.30 vs. limit=6.0 2024-09-24 10:33:16,470 INFO [train.py:1198] (2/4) Epoch 27, batch 1950, loss[loss=0.1805, ctc_loss=0.116, cr_loss=0.3227, over 17144.00 frames. ], tot_loss[loss=0.204, ctc_loss=0.1338, cr_loss=0.3507, over 3345542.18 frames. ], batch size: 45, lr: 4.42e-03, grad_scale: 16.0 2024-09-24 10:33:33,509 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.07 vs. limit=15.0 2024-09-24 10:33:41,164 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=481866.0, ans=0.125 2024-09-24 10:33:57,420 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=481912.6666666667, ans=0.125 2024-09-24 10:34:14,238 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=481959.3333333333, ans=0.1 2024-09-24 10:34:20,730 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=481959.3333333333, ans=0.125 2024-09-24 10:34:42,089 INFO [train.py:1198] (2/4) Epoch 27, batch 2000, loss[loss=0.1911, ctc_loss=0.1245, cr_loss=0.3333, over 17300.00 frames. ], tot_loss[loss=0.2039, ctc_loss=0.1339, cr_loss=0.3501, over 3331678.01 frames. ], batch size: 49, lr: 4.42e-03, grad_scale: 32.0 2024-09-24 10:34:49,402 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.42 vs. limit=15.0 2024-09-24 10:34:58,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=482099.3333333333, ans=0.125 2024-09-24 10:35:16,673 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=482146.0, ans=0.1 2024-09-24 10:35:19,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=482146.0, ans=0.0 2024-09-24 10:35:46,434 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.048e+02 1.275e+02 1.345e+02 1.450e+02 1.969e+02, threshold=2.691e+02, percent-clipped=0.0 2024-09-24 10:35:56,519 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=482239.3333333333, ans=0.0 2024-09-24 10:36:01,040 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=482239.3333333333, ans=0.1 2024-09-24 10:36:04,144 INFO [train.py:1198] (2/4) Epoch 27, batch 2050, loss[loss=0.1895, ctc_loss=0.1207, cr_loss=0.3441, over 17135.00 frames. ], tot_loss[loss=0.2041, ctc_loss=0.1339, cr_loss=0.3511, over 3340140.35 frames. ], batch size: 48, lr: 4.42e-03, grad_scale: 16.0 2024-09-24 10:36:26,815 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.min_positive, batch_count=482332.6666666667, ans=0.025 2024-09-24 10:36:30,049 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=482332.6666666667, ans=0.2 2024-09-24 10:36:30,078 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=482332.6666666667, ans=0.125 2024-09-24 10:36:36,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=482379.3333333333, ans=0.07 2024-09-24 10:36:44,811 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=14.55 vs. limit=22.5 2024-09-24 10:36:57,191 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 10:37:10,136 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=482472.6666666667, ans=0.0 2024-09-24 10:37:27,021 INFO [train.py:1198] (2/4) Epoch 27, batch 2100, loss[loss=0.192, ctc_loss=0.1248, cr_loss=0.3358, over 16970.00 frames. ], tot_loss[loss=0.2051, ctc_loss=0.1348, cr_loss=0.3514, over 3326097.88 frames. ], batch size: 42, lr: 4.42e-03, grad_scale: 16.0 2024-09-24 10:37:44,530 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=5.85 vs. limit=15.0 2024-09-24 10:38:09,212 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=482612.6666666667, ans=0.125 2024-09-24 10:38:26,692 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=482659.3333333333, ans=0.0 2024-09-24 10:38:29,420 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.227e+02 1.308e+02 1.433e+02 1.787e+02, threshold=2.617e+02, percent-clipped=0.0 2024-09-24 10:38:29,731 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=482706.0, ans=0.125 2024-09-24 10:38:34,654 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=482706.0, ans=0.125 2024-09-24 10:38:47,142 INFO [train.py:1198] (2/4) Epoch 27, batch 2150, loss[loss=0.2082, ctc_loss=0.1384, cr_loss=0.349, over 16480.00 frames. ], tot_loss[loss=0.2052, ctc_loss=0.1348, cr_loss=0.3519, over 3342583.96 frames. ], batch size: 66, lr: 4.42e-03, grad_scale: 16.0 2024-09-24 10:38:55,673 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=482752.6666666667, ans=0.0 2024-09-24 10:39:00,357 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=482752.6666666667, ans=0.125 2024-09-24 10:39:14,362 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=4.77 vs. limit=12.0 2024-09-24 10:39:44,901 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 10:39:51,755 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.42 vs. limit=22.5 2024-09-24 10:40:08,092 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 10:40:14,291 INFO [train.py:1198] (2/4) Epoch 27, batch 2200, loss[loss=0.1708, ctc_loss=0.1103, cr_loss=0.3028, over 16931.00 frames. ], tot_loss[loss=0.2042, ctc_loss=0.134, cr_loss=0.3507, over 3354200.86 frames. ], batch size: 42, lr: 4.42e-03, grad_scale: 16.0 2024-09-24 10:40:14,945 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.78 vs. limit=15.0 2024-09-24 10:40:16,709 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.35 vs. limit=15.0 2024-09-24 10:40:30,410 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=483032.6666666667, ans=0.125 2024-09-24 10:40:41,984 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.96 vs. limit=10.0 2024-09-24 10:40:51,212 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=483079.3333333333, ans=0.2 2024-09-24 10:41:04,194 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.67 vs. limit=15.0 2024-09-24 10:41:10,127 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=483126.0, ans=0.125 2024-09-24 10:41:16,366 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.255e+02 1.303e+02 1.378e+02 1.644e+02, threshold=2.606e+02, percent-clipped=0.0 2024-09-24 10:41:26,281 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=483172.6666666667, ans=0.0 2024-09-24 10:41:32,840 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=483219.3333333333, ans=0.0 2024-09-24 10:41:34,084 INFO [train.py:1198] (2/4) Epoch 27, batch 2250, loss[loss=0.1783, ctc_loss=0.1149, cr_loss=0.3168, over 17066.00 frames. ], tot_loss[loss=0.2031, ctc_loss=0.1331, cr_loss=0.3498, over 3357886.09 frames. ], batch size: 46, lr: 4.42e-03, grad_scale: 16.0 2024-09-24 10:41:45,862 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=483219.3333333333, ans=0.125 2024-09-24 10:41:57,362 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.min_positive, batch_count=483266.0, ans=0.025 2024-09-24 10:42:01,186 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=6.47 vs. limit=15.0 2024-09-24 10:42:11,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=483312.6666666667, ans=0.0 2024-09-24 10:42:56,647 INFO [train.py:1198] (2/4) Epoch 27, batch 2300, loss[loss=0.1973, ctc_loss=0.1316, cr_loss=0.3285, over 17215.00 frames. ], tot_loss[loss=0.2033, ctc_loss=0.1333, cr_loss=0.3502, over 3366113.83 frames. ], batch size: 47, lr: 4.42e-03, grad_scale: 16.0 2024-09-24 10:43:01,895 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=483452.6666666667, ans=0.0 2024-09-24 10:43:23,961 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=483499.3333333333, ans=0.0 2024-09-24 10:43:25,986 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1.whitening_limit, batch_count=483499.3333333333, ans=10.0 2024-09-24 10:43:42,967 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=483592.6666666667, ans=0.0 2024-09-24 10:43:46,277 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=483592.6666666667, ans=0.1 2024-09-24 10:43:58,586 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.254e+02 1.343e+02 1.462e+02 2.563e+02, threshold=2.686e+02, percent-clipped=0.0 2024-09-24 10:44:08,061 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=483639.3333333333, ans=0.125 2024-09-24 10:44:18,919 INFO [train.py:1198] (2/4) Epoch 27, batch 2350, loss[loss=0.1894, ctc_loss=0.1279, cr_loss=0.3075, over 17032.00 frames. ], tot_loss[loss=0.204, ctc_loss=0.1338, cr_loss=0.3512, over 3363740.70 frames. ], batch size: 39, lr: 4.42e-03, grad_scale: 16.0 2024-09-24 10:44:22,427 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=483686.0, ans=0.2 2024-09-24 10:44:25,760 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=483686.0, ans=0.1 2024-09-24 10:44:36,155 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=483732.6666666667, ans=0.5 2024-09-24 10:44:41,326 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=7.15 vs. limit=15.0 2024-09-24 10:45:27,113 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.66 vs. limit=15.0 2024-09-24 10:45:42,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=483919.3333333333, ans=0.2 2024-09-24 10:45:43,652 INFO [train.py:1198] (2/4) Epoch 27, batch 2400, loss[loss=0.2083, ctc_loss=0.1334, cr_loss=0.3745, over 17217.00 frames. ], tot_loss[loss=0.2034, ctc_loss=0.1333, cr_loss=0.3504, over 3363541.08 frames. ], batch size: 50, lr: 4.41e-03, grad_scale: 32.0 2024-09-24 10:46:09,096 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=483966.0, ans=0.125 2024-09-24 10:46:15,644 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=484012.6666666667, ans=0.05 2024-09-24 10:46:18,750 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=484012.6666666667, ans=0.025 2024-09-24 10:46:22,747 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=4.36 vs. limit=15.0 2024-09-24 10:46:45,619 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.113e+02 1.266e+02 1.332e+02 1.424e+02 3.115e+02, threshold=2.664e+02, percent-clipped=1.0 2024-09-24 10:47:03,206 INFO [train.py:1198] (2/4) Epoch 27, batch 2450, loss[loss=0.2047, ctc_loss=0.1315, cr_loss=0.3658, over 17172.00 frames. ], tot_loss[loss=0.2037, ctc_loss=0.1335, cr_loss=0.3509, over 3361118.97 frames. ], batch size: 45, lr: 4.41e-03, grad_scale: 32.0 2024-09-24 10:47:11,466 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=484152.6666666667, ans=0.2 2024-09-24 10:47:33,757 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=10.00 vs. limit=15.0 2024-09-24 10:47:34,833 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=484199.3333333333, ans=0.125 2024-09-24 10:47:39,627 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=484246.0, ans=0.025 2024-09-24 10:47:44,554 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=484246.0, ans=0.1 2024-09-24 10:48:02,222 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=484292.6666666667, ans=0.125 2024-09-24 10:48:06,724 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=484292.6666666667, ans=0.035 2024-09-24 10:48:06,809 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=484292.6666666667, ans=0.0 2024-09-24 10:48:17,336 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=5.90 vs. limit=15.0 2024-09-24 10:48:25,694 INFO [train.py:1198] (2/4) Epoch 27, batch 2500, loss[loss=0.1886, ctc_loss=0.1209, cr_loss=0.3386, over 17165.00 frames. ], tot_loss[loss=0.204, ctc_loss=0.1338, cr_loss=0.3513, over 3353787.55 frames. ], batch size: 45, lr: 4.41e-03, grad_scale: 32.0 2024-09-24 10:48:34,275 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=484386.0, ans=0.125 2024-09-24 10:48:56,422 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=484479.3333333333, ans=0.0 2024-09-24 10:49:13,133 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=484479.3333333333, ans=0.125 2024-09-24 10:49:16,554 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=484526.0, ans=0.125 2024-09-24 10:49:30,485 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.109e+02 1.304e+02 1.383e+02 1.470e+02 1.966e+02, threshold=2.767e+02, percent-clipped=0.0 2024-09-24 10:49:46,403 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 10:49:50,882 INFO [train.py:1198] (2/4) Epoch 27, batch 2550, loss[loss=0.2088, ctc_loss=0.1365, cr_loss=0.3613, over 17290.00 frames. ], tot_loss[loss=0.2044, ctc_loss=0.1341, cr_loss=0.3514, over 3340703.67 frames. ], batch size: 46, lr: 4.41e-03, grad_scale: 32.0 2024-09-24 10:49:52,890 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=484619.3333333333, ans=0.125 2024-09-24 10:50:05,032 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=484619.3333333333, ans=0.1 2024-09-24 10:50:40,083 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=484759.3333333333, ans=0.125 2024-09-24 10:50:53,019 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=484759.3333333333, ans=0.0 2024-09-24 10:51:09,199 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=9.99 vs. limit=12.0 2024-09-24 10:51:13,299 INFO [train.py:1198] (2/4) Epoch 27, batch 2600, loss[loss=0.2108, ctc_loss=0.1394, cr_loss=0.3573, over 17303.00 frames. ], tot_loss[loss=0.2044, ctc_loss=0.1341, cr_loss=0.3515, over 3346113.28 frames. ], batch size: 49, lr: 4.41e-03, grad_scale: 32.0 2024-09-24 10:51:29,821 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=484899.3333333333, ans=0.0 2024-09-24 10:51:39,310 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=484899.3333333333, ans=0.1 2024-09-24 10:52:00,149 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=484992.6666666667, ans=0.125 2024-09-24 10:52:00,177 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=484992.6666666667, ans=0.125 2024-09-24 10:52:03,305 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.min_abs, batch_count=484992.6666666667, ans=0.5 2024-09-24 10:52:15,856 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.081e+02 1.258e+02 1.335e+02 1.464e+02 4.634e+02, threshold=2.669e+02, percent-clipped=1.0 2024-09-24 10:52:19,805 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.39 vs. limit=15.0 2024-09-24 10:52:34,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=485086.0, ans=0.125 2024-09-24 10:52:36,179 INFO [train.py:1198] (2/4) Epoch 27, batch 2650, loss[loss=0.1783, ctc_loss=0.1128, cr_loss=0.3274, over 17196.00 frames. ], tot_loss[loss=0.2048, ctc_loss=0.1343, cr_loss=0.3523, over 3352120.35 frames. ], batch size: 41, lr: 4.41e-03, grad_scale: 32.0 2024-09-24 10:53:08,939 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.67 vs. limit=10.0 2024-09-24 10:53:49,895 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=485272.6666666667, ans=0.125 2024-09-24 10:53:54,546 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=485319.3333333333, ans=0.025 2024-09-24 10:53:55,756 INFO [train.py:1198] (2/4) Epoch 27, batch 2700, loss[loss=0.2084, ctc_loss=0.1382, cr_loss=0.3513, over 17009.00 frames. ], tot_loss[loss=0.2038, ctc_loss=0.1336, cr_loss=0.351, over 3354861.08 frames. ], batch size: 52, lr: 4.41e-03, grad_scale: 32.0 2024-09-24 10:53:59,382 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=485319.3333333333, ans=0.125 2024-09-24 10:54:18,601 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=485366.0, ans=0.125 2024-09-24 10:54:24,064 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=3.86 vs. limit=15.0 2024-09-24 10:54:53,461 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=485459.3333333333, ans=0.0 2024-09-24 10:55:01,295 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.71 vs. limit=15.0 2024-09-24 10:55:03,223 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=8.04 vs. limit=15.0 2024-09-24 10:55:08,209 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.085e+02 1.258e+02 1.322e+02 1.410e+02 2.487e+02, threshold=2.644e+02, percent-clipped=0.0 2024-09-24 10:55:14,890 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=485506.0, ans=0.125 2024-09-24 10:55:25,584 INFO [train.py:1198] (2/4) Epoch 27, batch 2750, loss[loss=0.2016, ctc_loss=0.1335, cr_loss=0.3405, over 17304.00 frames. ], tot_loss[loss=0.2036, ctc_loss=0.1336, cr_loss=0.3503, over 3357677.92 frames. ], batch size: 46, lr: 4.41e-03, grad_scale: 32.0 2024-09-24 10:55:26,099 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=10.38 vs. limit=12.0 2024-09-24 10:55:32,258 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=485552.6666666667, ans=0.125 2024-09-24 10:55:48,097 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=485599.3333333333, ans=0.2 2024-09-24 10:56:00,894 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=485646.0, ans=0.1 2024-09-24 10:56:19,893 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=485692.6666666667, ans=0.1 2024-09-24 10:56:39,116 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=485739.3333333333, ans=0.1 2024-09-24 10:56:45,243 INFO [train.py:1198] (2/4) Epoch 27, batch 2800, loss[loss=0.2752, ctc_loss=0.195, cr_loss=0.4007, over 11339.00 frames. ], tot_loss[loss=0.2041, ctc_loss=0.134, cr_loss=0.3507, over 3354273.52 frames. ], batch size: 123, lr: 4.41e-03, grad_scale: 32.0 2024-09-24 10:57:10,555 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.16 vs. limit=15.0 2024-09-24 10:57:50,314 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.262e+02 1.376e+02 1.500e+02 2.364e+02, threshold=2.751e+02, percent-clipped=0.0 2024-09-24 10:58:07,806 INFO [train.py:1198] (2/4) Epoch 27, batch 2850, loss[loss=0.1616, ctc_loss=0.1039, cr_loss=0.2885, over 17147.00 frames. ], tot_loss[loss=0.2041, ctc_loss=0.134, cr_loss=0.3503, over 3338932.03 frames. ], batch size: 41, lr: 4.41e-03, grad_scale: 32.0 2024-09-24 10:58:38,699 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.84 vs. limit=6.0 2024-09-24 10:59:30,062 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=486206.0, ans=0.125 2024-09-24 10:59:32,995 INFO [train.py:1198] (2/4) Epoch 27, batch 2900, loss[loss=0.224, ctc_loss=0.1487, cr_loss=0.3766, over 17005.00 frames. ], tot_loss[loss=0.205, ctc_loss=0.1347, cr_loss=0.3517, over 3339341.41 frames. ], batch size: 53, lr: 4.40e-03, grad_scale: 32.0 2024-09-24 10:59:54,956 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=486299.3333333333, ans=0.035 2024-09-24 11:00:23,624 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=486392.6666666667, ans=0.0 2024-09-24 11:00:37,817 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.059e+02 1.250e+02 1.339e+02 1.437e+02 2.331e+02, threshold=2.678e+02, percent-clipped=0.0 2024-09-24 11:00:44,549 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=486439.3333333333, ans=0.0 2024-09-24 11:00:55,785 INFO [train.py:1198] (2/4) Epoch 27, batch 2950, loss[loss=0.2413, ctc_loss=0.1689, cr_loss=0.3617, over 11777.00 frames. ], tot_loss[loss=0.204, ctc_loss=0.1338, cr_loss=0.3509, over 3347810.40 frames. ], batch size: 123, lr: 4.40e-03, grad_scale: 32.0 2024-09-24 11:01:17,471 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.out_whiten.whitening_limit, batch_count=486532.6666666667, ans=15.0 2024-09-24 11:01:23,497 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=11.54 vs. limit=15.0 2024-09-24 11:01:29,517 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=486579.3333333333, ans=0.0 2024-09-24 11:01:48,546 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=486626.0, ans=0.125 2024-09-24 11:02:05,593 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=486672.6666666667, ans=0.125 2024-09-24 11:02:14,940 INFO [train.py:1198] (2/4) Epoch 27, batch 3000, loss[loss=0.2357, ctc_loss=0.1541, cr_loss=0.4082, over 17016.00 frames. ], tot_loss[loss=0.2039, ctc_loss=0.1337, cr_loss=0.3509, over 3356110.85 frames. ], batch size: 56, lr: 4.40e-03, grad_scale: 32.0 2024-09-24 11:02:14,941 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 11:02:26,062 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.2.encoder.layers.1.self_attn_weights, attn_weights_entropy = tensor([5.4180, 4.8740, 4.7597, 5.1178], device='cuda:2') 2024-09-24 11:02:30,459 INFO [train.py:1230] (2/4) Epoch 27, validation: loss=0.03681, ctc_loss=0.03681, cr_loss=8.353e-15, over 944034.00 frames. 2024-09-24 11:02:30,460 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 11:02:40,127 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=486719.3333333333, ans=0.0 2024-09-24 11:02:46,660 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer_na.min_abs, batch_count=486766.0, ans=0.02 2024-09-24 11:03:17,956 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=486859.3333333333, ans=0.1 2024-09-24 11:03:32,005 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.087e+02 1.263e+02 1.339e+02 1.435e+02 2.051e+02, threshold=2.678e+02, percent-clipped=0.0 2024-09-24 11:03:45,710 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=4.97 vs. limit=10.0 2024-09-24 11:03:46,283 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=486906.0, ans=0.125 2024-09-24 11:03:46,392 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=486906.0, ans=0.125 2024-09-24 11:03:49,182 INFO [train.py:1198] (2/4) Epoch 27, batch 3050, loss[loss=0.2175, ctc_loss=0.1435, cr_loss=0.3698, over 17010.00 frames. ], tot_loss[loss=0.2045, ctc_loss=0.1341, cr_loss=0.3517, over 3357203.26 frames. ], batch size: 52, lr: 4.40e-03, grad_scale: 32.0 2024-09-24 11:04:00,345 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=486952.6666666667, ans=0.125 2024-09-24 11:04:00,422 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=486952.6666666667, ans=0.2 2024-09-24 11:04:04,935 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.max_abs, batch_count=486999.3333333333, ans=10.0 2024-09-24 11:04:15,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer_ff3.min_abs, batch_count=486999.3333333333, ans=0.2 2024-09-24 11:04:17,703 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=486999.3333333333, ans=0.125 2024-09-24 11:04:20,939 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=487046.0, ans=0.125 2024-09-24 11:04:55,421 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=487139.3333333333, ans=0.2 2024-09-24 11:05:00,261 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=487139.3333333333, ans=0.125 2024-09-24 11:05:03,118 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=487139.3333333333, ans=0.125 2024-09-24 11:05:07,554 INFO [train.py:1198] (2/4) Epoch 27, batch 3100, loss[loss=0.2031, ctc_loss=0.1332, cr_loss=0.3495, over 16996.00 frames. ], tot_loss[loss=0.2038, ctc_loss=0.1336, cr_loss=0.3511, over 3358285.87 frames. ], batch size: 53, lr: 4.40e-03, grad_scale: 32.0 2024-09-24 11:05:22,349 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=487232.6666666667, ans=0.125 2024-09-24 11:05:48,508 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=487279.3333333333, ans=0.125 2024-09-24 11:06:05,623 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=487326.0, ans=0.035 2024-09-24 11:06:11,764 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.251e+02 1.335e+02 1.455e+02 2.261e+02, threshold=2.670e+02, percent-clipped=0.0 2024-09-24 11:06:12,083 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=487372.6666666667, ans=0.125 2024-09-24 11:06:13,690 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=487372.6666666667, ans=0.025 2024-09-24 11:06:16,668 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=487372.6666666667, ans=0.0 2024-09-24 11:06:22,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=487372.6666666667, ans=0.125 2024-09-24 11:06:28,978 INFO [train.py:1198] (2/4) Epoch 27, batch 3150, loss[loss=0.2269, ctc_loss=0.1501, cr_loss=0.3841, over 16690.00 frames. ], tot_loss[loss=0.2038, ctc_loss=0.1337, cr_loss=0.3506, over 3353872.59 frames. ], batch size: 61, lr: 4.40e-03, grad_scale: 32.0 2024-09-24 11:06:56,714 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=487466.0, ans=0.0 2024-09-24 11:07:10,502 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=487512.6666666667, ans=0.125 2024-09-24 11:07:23,765 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=487559.3333333333, ans=0.125 2024-09-24 11:07:42,339 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=487606.0, ans=0.1 2024-09-24 11:07:48,439 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=487606.0, ans=0.0 2024-09-24 11:07:51,271 INFO [train.py:1198] (2/4) Epoch 27, batch 3200, loss[loss=0.1635, ctc_loss=0.1013, cr_loss=0.3106, over 17200.00 frames. ], tot_loss[loss=0.2033, ctc_loss=0.1332, cr_loss=0.35, over 3358096.05 frames. ], batch size: 41, lr: 4.40e-03, grad_scale: 32.0 2024-09-24 11:08:27,846 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=487746.0, ans=0.0 2024-09-24 11:08:52,484 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.142e+02 1.262e+02 1.380e+02 1.502e+02 1.892e+02, threshold=2.760e+02, percent-clipped=0.0 2024-09-24 11:09:09,782 INFO [train.py:1198] (2/4) Epoch 27, batch 3250, loss[loss=0.1642, ctc_loss=0.1031, cr_loss=0.3055, over 17102.00 frames. ], tot_loss[loss=0.2037, ctc_loss=0.1336, cr_loss=0.3507, over 3356552.19 frames. ], batch size: 40, lr: 4.40e-03, grad_scale: 32.0 2024-09-24 11:09:33,676 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.min_positive, batch_count=487932.6666666667, ans=0.025 2024-09-24 11:09:33,722 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=487932.6666666667, ans=0.0 2024-09-24 11:09:38,419 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=487932.6666666667, ans=0.125 2024-09-24 11:09:47,620 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=487979.3333333333, ans=0.2 2024-09-24 11:09:52,318 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=487979.3333333333, ans=0.2 2024-09-24 11:09:53,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=487979.3333333333, ans=0.125 2024-09-24 11:10:01,582 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=488026.0, ans=0.125 2024-09-24 11:10:10,826 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=488072.6666666667, ans=0.125 2024-09-24 11:10:20,369 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=488072.6666666667, ans=0.1 2024-09-24 11:10:27,841 INFO [train.py:1198] (2/4) Epoch 27, batch 3300, loss[loss=0.2245, ctc_loss=0.15, cr_loss=0.3723, over 16893.00 frames. ], tot_loss[loss=0.2032, ctc_loss=0.1331, cr_loss=0.3502, over 3354793.42 frames. ], batch size: 58, lr: 4.40e-03, grad_scale: 32.0 2024-09-24 11:10:48,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=488166.0, ans=0.2 2024-09-24 11:11:04,317 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=488212.6666666667, ans=0.125 2024-09-24 11:11:07,338 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=488212.6666666667, ans=0.0 2024-09-24 11:11:15,356 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=488259.3333333333, ans=0.125 2024-09-24 11:11:28,980 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.251e+02 1.338e+02 1.445e+02 2.480e+02, threshold=2.676e+02, percent-clipped=0.0 2024-09-24 11:11:40,437 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=488306.0, ans=0.0 2024-09-24 11:11:46,441 INFO [train.py:1198] (2/4) Epoch 27, batch 3350, loss[loss=0.2145, ctc_loss=0.1381, cr_loss=0.3822, over 17022.00 frames. ], tot_loss[loss=0.2029, ctc_loss=0.133, cr_loss=0.3494, over 3358351.77 frames. ], batch size: 52, lr: 4.39e-03, grad_scale: 32.0 2024-09-24 11:12:21,225 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=488446.0, ans=0.125 2024-09-24 11:12:29,537 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=488446.0, ans=0.125 2024-09-24 11:12:31,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=488446.0, ans=0.0 2024-09-24 11:13:07,096 INFO [train.py:1198] (2/4) Epoch 27, batch 3400, loss[loss=0.2142, ctc_loss=0.1418, cr_loss=0.362, over 17018.00 frames. ], tot_loss[loss=0.2029, ctc_loss=0.1331, cr_loss=0.3492, over 3355785.11 frames. ], batch size: 44, lr: 4.39e-03, grad_scale: 32.0 2024-09-24 11:13:10,556 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=488586.0, ans=0.125 2024-09-24 11:13:10,574 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=488586.0, ans=0.5 2024-09-24 11:13:11,277 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.91 vs. limit=10.0 2024-09-24 11:13:23,033 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=488632.6666666667, ans=0.0 2024-09-24 11:13:39,256 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.83 vs. limit=10.0 2024-09-24 11:13:48,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=488679.3333333333, ans=0.125 2024-09-24 11:13:49,793 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=488679.3333333333, ans=0.125 2024-09-24 11:14:08,004 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.079e+02 1.290e+02 1.371e+02 1.502e+02 1.850e+02, threshold=2.743e+02, percent-clipped=0.0 2024-09-24 11:14:08,289 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=488772.6666666667, ans=0.125 2024-09-24 11:14:25,084 INFO [train.py:1198] (2/4) Epoch 27, batch 3450, loss[loss=0.1697, ctc_loss=0.1092, cr_loss=0.3024, over 17169.00 frames. ], tot_loss[loss=0.2023, ctc_loss=0.1327, cr_loss=0.3479, over 3353195.99 frames. ], batch size: 41, lr: 4.39e-03, grad_scale: 32.0 2024-09-24 11:14:25,335 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=488819.3333333333, ans=0.0 2024-09-24 11:15:24,740 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=488959.3333333333, ans=0.125 2024-09-24 11:15:44,983 INFO [train.py:1198] (2/4) Epoch 27, batch 3500, loss[loss=0.181, ctc_loss=0.1157, cr_loss=0.3267, over 17084.00 frames. ], tot_loss[loss=0.2033, ctc_loss=0.1335, cr_loss=0.349, over 3341251.48 frames. ], batch size: 40, lr: 4.39e-03, grad_scale: 32.0 2024-09-24 11:16:10,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=489099.3333333333, ans=0.0 2024-09-24 11:16:13,973 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=489099.3333333333, ans=0.125 2024-09-24 11:16:31,311 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=489192.6666666667, ans=0.125 2024-09-24 11:16:32,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=489192.6666666667, ans=0.125 2024-09-24 11:16:46,472 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.188e+02 1.271e+02 1.363e+02 1.500e+02 3.531e+02, threshold=2.727e+02, percent-clipped=1.0 2024-09-24 11:16:48,270 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=489239.3333333333, ans=0.2 2024-09-24 11:16:55,034 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=489239.3333333333, ans=0.125 2024-09-24 11:17:05,734 INFO [train.py:1198] (2/4) Epoch 27, batch 3550, loss[loss=0.2159, ctc_loss=0.1446, cr_loss=0.3569, over 16980.00 frames. ], tot_loss[loss=0.203, ctc_loss=0.1332, cr_loss=0.349, over 3352095.48 frames. ], batch size: 56, lr: 4.39e-03, grad_scale: 32.0 2024-09-24 11:17:07,477 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=489286.0, ans=0.0 2024-09-24 11:17:10,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=489286.0, ans=0.2 2024-09-24 11:17:23,939 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=12.11 vs. limit=15.0 2024-09-24 11:17:37,810 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=13.30 vs. limit=15.0 2024-09-24 11:17:48,418 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=489379.3333333333, ans=0.1 2024-09-24 11:17:54,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=489426.0, ans=0.125 2024-09-24 11:18:12,038 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=489472.6666666667, ans=0.125 2024-09-24 11:18:18,726 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=3.82 vs. limit=15.0 2024-09-24 11:18:22,841 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=489472.6666666667, ans=0.0 2024-09-24 11:18:25,766 INFO [train.py:1198] (2/4) Epoch 27, batch 3600, loss[loss=0.1791, ctc_loss=0.1165, cr_loss=0.3128, over 16693.00 frames. ], tot_loss[loss=0.2029, ctc_loss=0.1331, cr_loss=0.3491, over 3355895.71 frames. ], batch size: 37, lr: 4.39e-03, grad_scale: 32.0 2024-09-24 11:18:30,639 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=489519.3333333333, ans=0.125 2024-09-24 11:18:51,013 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=489566.0, ans=0.0 2024-09-24 11:19:26,809 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.098e+02 1.270e+02 1.354e+02 1.435e+02 1.974e+02, threshold=2.707e+02, percent-clipped=0.0 2024-09-24 11:19:44,335 INFO [train.py:1198] (2/4) Epoch 27, batch 3650, loss[loss=0.1994, ctc_loss=0.1321, cr_loss=0.3362, over 17071.00 frames. ], tot_loss[loss=0.2034, ctc_loss=0.1336, cr_loss=0.3494, over 3336901.14 frames. ], batch size: 46, lr: 4.39e-03, grad_scale: 32.0 2024-09-24 11:19:57,242 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=489752.6666666667, ans=0.125 2024-09-24 11:19:58,888 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=489799.3333333333, ans=0.0 2024-09-24 11:20:03,528 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=489799.3333333333, ans=0.125 2024-09-24 11:20:08,192 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=489799.3333333333, ans=0.125 2024-09-24 11:20:09,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=489799.3333333333, ans=0.1 2024-09-24 11:20:25,879 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=489846.0, ans=0.125 2024-09-24 11:20:59,415 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=8.37 vs. limit=15.0 2024-09-24 11:21:00,657 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=489939.3333333333, ans=0.125 2024-09-24 11:21:03,530 INFO [train.py:1198] (2/4) Epoch 27, batch 3700, loss[loss=0.1612, ctc_loss=0.102, cr_loss=0.296, over 16664.00 frames. ], tot_loss[loss=0.2029, ctc_loss=0.1331, cr_loss=0.349, over 3347349.23 frames. ], batch size: 37, lr: 4.39e-03, grad_scale: 32.0 2024-09-24 11:22:05,030 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.056e+02 1.302e+02 1.450e+02 1.561e+02 3.629e+02, threshold=2.900e+02, percent-clipped=2.0 2024-09-24 11:22:06,801 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=490172.6666666667, ans=0.0 2024-09-24 11:22:06,960 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=490172.6666666667, ans=0.07 2024-09-24 11:22:09,799 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=490172.6666666667, ans=0.1 2024-09-24 11:22:21,896 INFO [train.py:1198] (2/4) Epoch 27, batch 3750, loss[loss=0.1606, ctc_loss=0.1038, cr_loss=0.2843, over 17169.00 frames. ], tot_loss[loss=0.202, ctc_loss=0.1325, cr_loss=0.3476, over 3346622.50 frames. ], batch size: 41, lr: 4.39e-03, grad_scale: 32.0 2024-09-24 11:22:23,535 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=490219.3333333333, ans=0.125 2024-09-24 11:22:44,619 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=15.92 vs. limit=22.5 2024-09-24 11:23:21,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=490359.3333333333, ans=0.125 2024-09-24 11:23:39,567 INFO [train.py:1198] (2/4) Epoch 27, batch 3800, loss[loss=0.1935, ctc_loss=0.1263, cr_loss=0.3362, over 17044.00 frames. ], tot_loss[loss=0.2022, ctc_loss=0.1327, cr_loss=0.3476, over 3324235.34 frames. ], batch size: 39, lr: 4.39e-03, grad_scale: 32.0 2024-09-24 11:24:02,229 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=490499.3333333333, ans=0.125 2024-09-24 11:24:41,922 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.297e+02 1.423e+02 1.568e+02 2.554e+02, threshold=2.846e+02, percent-clipped=0.0 2024-09-24 11:24:43,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=490639.3333333333, ans=0.125 2024-09-24 11:24:59,148 INFO [train.py:1198] (2/4) Epoch 27, batch 3850, loss[loss=0.2745, ctc_loss=0.1951, cr_loss=0.3971, over 11085.00 frames. ], tot_loss[loss=0.2046, ctc_loss=0.1346, cr_loss=0.3503, over 3279944.24 frames. ], batch size: 123, lr: 4.38e-03, grad_scale: 32.0 2024-09-24 11:25:01,005 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=490686.0, ans=0.125 2024-09-24 11:25:25,404 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=490732.6666666667, ans=0.1 2024-09-24 11:25:28,541 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=490779.3333333333, ans=0.025 2024-09-24 11:25:56,843 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=9.45 vs. limit=12.0 2024-09-24 11:26:02,559 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=490872.6666666667, ans=0.2 2024-09-24 11:27:00,313 INFO [train.py:1198] (2/4) Epoch 28, batch 0, loss[loss=0.1956, ctc_loss=0.1277, cr_loss=0.3393, over 17231.00 frames. ], tot_loss[loss=0.1956, ctc_loss=0.1277, cr_loss=0.3393, over 17231.00 frames. ], batch size: 44, lr: 4.30e-03, grad_scale: 32.0 2024-09-24 11:27:00,314 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 11:27:15,942 INFO [train.py:1230] (2/4) Epoch 28, validation: loss=0.03666, ctc_loss=0.03666, cr_loss=9.126e-15, over 944034.00 frames. 2024-09-24 11:27:15,943 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 11:27:59,334 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=490994.0, ans=0.2 2024-09-24 11:28:10,962 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.89 vs. limit=15.0 2024-09-24 11:28:27,520 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.102e+02 1.316e+02 1.494e+02 1.657e+02 3.455e+02, threshold=2.987e+02, percent-clipped=1.0 2024-09-24 11:28:38,879 INFO [train.py:1198] (2/4) Epoch 28, batch 50, loss[loss=0.1938, ctc_loss=0.1253, cr_loss=0.3428, over 17306.00 frames. ], tot_loss[loss=0.2072, ctc_loss=0.1359, cr_loss=0.3568, over 755606.84 frames. ], batch size: 51, lr: 4.30e-03, grad_scale: 32.0 2024-09-24 11:28:44,197 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=10.57 vs. limit=22.5 2024-09-24 11:29:10,301 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=11.40 vs. limit=22.5 2024-09-24 11:29:11,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=491227.3333333333, ans=0.125 2024-09-24 11:29:17,412 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=491227.3333333333, ans=0.125 2024-09-24 11:29:29,956 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=491274.0, ans=0.0 2024-09-24 11:29:31,559 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=491274.0, ans=0.035 2024-09-24 11:29:57,349 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 11:29:58,609 INFO [train.py:1198] (2/4) Epoch 28, batch 100, loss[loss=0.2202, ctc_loss=0.1454, cr_loss=0.3743, over 17317.00 frames. ], tot_loss[loss=0.2061, ctc_loss=0.1351, cr_loss=0.3552, over 1339320.47 frames. ], batch size: 46, lr: 4.30e-03, grad_scale: 32.0 2024-09-24 11:30:17,792 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=491414.0, ans=0.1 2024-09-24 11:30:19,415 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=491414.0, ans=0.125 2024-09-24 11:30:28,886 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=491414.0, ans=0.125 2024-09-24 11:30:40,089 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=491460.6666666667, ans=0.1 2024-09-24 11:30:48,187 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=491507.3333333333, ans=0.125 2024-09-24 11:31:00,529 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=491507.3333333333, ans=0.125 2024-09-24 11:31:11,239 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.069e+02 1.221e+02 1.303e+02 1.406e+02 2.036e+02, threshold=2.607e+02, percent-clipped=0.0 2024-09-24 11:31:20,724 INFO [train.py:1198] (2/4) Epoch 28, batch 150, loss[loss=0.2183, ctc_loss=0.1459, cr_loss=0.3622, over 17369.00 frames. ], tot_loss[loss=0.2052, ctc_loss=0.1344, cr_loss=0.3539, over 1782240.60 frames. ], batch size: 48, lr: 4.30e-03, grad_scale: 32.0 2024-09-24 11:31:25,713 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=491600.6666666667, ans=0.125 2024-09-24 11:31:46,414 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=491647.3333333333, ans=0.125 2024-09-24 11:31:58,832 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=491694.0, ans=0.2 2024-09-24 11:32:12,503 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 11:32:21,913 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.12 vs. limit=10.0 2024-09-24 11:32:48,350 INFO [train.py:1198] (2/4) Epoch 28, batch 200, loss[loss=0.2165, ctc_loss=0.1451, cr_loss=0.3573, over 17108.00 frames. ], tot_loss[loss=0.2046, ctc_loss=0.1341, cr_loss=0.3523, over 2130966.32 frames. ], batch size: 49, lr: 4.30e-03, grad_scale: 16.0 2024-09-24 11:32:50,333 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=491834.0, ans=0.125 2024-09-24 11:33:15,649 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=491880.6666666667, ans=0.125 2024-09-24 11:33:28,408 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=491927.3333333333, ans=0.125 2024-09-24 11:34:00,088 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.107e+02 1.241e+02 1.331e+02 1.443e+02 1.741e+02, threshold=2.662e+02, percent-clipped=0.0 2024-09-24 11:34:08,075 INFO [train.py:1198] (2/4) Epoch 28, batch 250, loss[loss=0.1805, ctc_loss=0.1156, cr_loss=0.3247, over 17057.00 frames. ], tot_loss[loss=0.2044, ctc_loss=0.1339, cr_loss=0.3528, over 2409586.13 frames. ], batch size: 39, lr: 4.30e-03, grad_scale: 16.0 2024-09-24 11:34:35,724 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=8.21 vs. limit=15.0 2024-09-24 11:34:38,584 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=492160.6666666667, ans=0.1 2024-09-24 11:35:03,933 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=492207.3333333333, ans=0.2 2024-09-24 11:35:30,770 INFO [train.py:1198] (2/4) Epoch 28, batch 300, loss[loss=0.2205, ctc_loss=0.1441, cr_loss=0.3822, over 17230.00 frames. ], tot_loss[loss=0.2027, ctc_loss=0.1328, cr_loss=0.3497, over 2622791.87 frames. ], batch size: 55, lr: 4.30e-03, grad_scale: 16.0 2024-09-24 11:35:37,246 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=492300.6666666667, ans=0.125 2024-09-24 11:35:45,292 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=492347.3333333333, ans=0.0 2024-09-24 11:35:50,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=492347.3333333333, ans=0.0 2024-09-24 11:36:19,367 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=9.90 vs. limit=12.0 2024-09-24 11:36:20,645 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=492440.6666666667, ans=0.125 2024-09-24 11:36:28,779 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=492440.6666666667, ans=0.125 2024-09-24 11:36:36,725 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=492487.3333333333, ans=0.2 2024-09-24 11:36:39,754 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=492487.3333333333, ans=0.125 2024-09-24 11:36:42,613 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.243e+02 1.313e+02 1.459e+02 2.703e+02, threshold=2.626e+02, percent-clipped=1.0 2024-09-24 11:36:43,503 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=512, metric=2.74 vs. limit=15.0 2024-09-24 11:36:50,619 INFO [train.py:1198] (2/4) Epoch 28, batch 350, loss[loss=0.1777, ctc_loss=0.1136, cr_loss=0.3205, over 17174.00 frames. ], tot_loss[loss=0.2041, ctc_loss=0.1339, cr_loss=0.351, over 2782551.63 frames. ], batch size: 41, lr: 4.30e-03, grad_scale: 16.0 2024-09-24 11:36:50,744 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=492534.0, ans=0.125 2024-09-24 11:37:15,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=492580.6666666667, ans=0.0 2024-09-24 11:37:40,662 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=492627.3333333333, ans=0.0 2024-09-24 11:37:46,869 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=492674.0, ans=0.125 2024-09-24 11:38:05,850 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=492720.6666666667, ans=0.125 2024-09-24 11:38:17,399 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=492767.3333333333, ans=0.0 2024-09-24 11:38:18,683 INFO [train.py:1198] (2/4) Epoch 28, batch 400, loss[loss=0.2209, ctc_loss=0.145, cr_loss=0.3794, over 17037.00 frames. ], tot_loss[loss=0.2033, ctc_loss=0.1333, cr_loss=0.3499, over 2915357.81 frames. ], batch size: 52, lr: 4.29e-03, grad_scale: 32.0 2024-09-24 11:38:19,310 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.36 vs. limit=6.0 2024-09-24 11:38:30,567 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=492767.3333333333, ans=0.0 2024-09-24 11:38:40,049 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 11:38:43,091 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=492814.0, ans=0.125 2024-09-24 11:39:16,569 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=492907.3333333333, ans=0.1 2024-09-24 11:39:20,014 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.28 vs. limit=15.0 2024-09-24 11:39:30,616 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.218e+02 1.284e+02 1.430e+02 3.216e+02, threshold=2.568e+02, percent-clipped=1.0 2024-09-24 11:39:38,502 INFO [train.py:1198] (2/4) Epoch 28, batch 450, loss[loss=0.2138, ctc_loss=0.1434, cr_loss=0.3518, over 16843.00 frames. ], tot_loss[loss=0.2033, ctc_loss=0.1334, cr_loss=0.3494, over 3011784.70 frames. ], batch size: 58, lr: 4.29e-03, grad_scale: 32.0 2024-09-24 11:39:53,113 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 11:39:53,237 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 11:40:23,401 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=14.62 vs. limit=22.5 2024-09-24 11:40:42,114 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 11:40:48,535 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=493187.3333333333, ans=0.1 2024-09-24 11:40:58,095 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=493187.3333333333, ans=0.0 2024-09-24 11:41:00,829 INFO [train.py:1198] (2/4) Epoch 28, batch 500, loss[loss=0.2421, ctc_loss=0.1577, cr_loss=0.4219, over 17195.00 frames. ], tot_loss[loss=0.2039, ctc_loss=0.1337, cr_loss=0.3507, over 3094769.14 frames. ], batch size: 55, lr: 4.29e-03, grad_scale: 16.0 2024-09-24 11:41:17,046 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=493280.6666666667, ans=0.07 2024-09-24 11:41:23,306 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=493280.6666666667, ans=0.125 2024-09-24 11:41:39,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=493327.3333333333, ans=0.0 2024-09-24 11:41:50,418 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=493374.0, ans=0.125 2024-09-24 11:42:00,846 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=493374.0, ans=0.125 2024-09-24 11:42:21,660 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.063e+02 1.292e+02 1.373e+02 1.533e+02 1.928e+02, threshold=2.747e+02, percent-clipped=0.0 2024-09-24 11:42:26,651 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=493467.3333333333, ans=0.0 2024-09-24 11:42:27,935 INFO [train.py:1198] (2/4) Epoch 28, batch 550, loss[loss=0.233, ctc_loss=0.1548, cr_loss=0.3912, over 16137.00 frames. ], tot_loss[loss=0.2046, ctc_loss=0.1343, cr_loss=0.3514, over 3135462.71 frames. ], batch size: 74, lr: 4.29e-03, grad_scale: 16.0 2024-09-24 11:42:54,348 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.70 vs. limit=15.0 2024-09-24 11:42:59,288 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.26 vs. limit=12.0 2024-09-24 11:43:05,500 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=13.50 vs. limit=15.0 2024-09-24 11:43:47,659 INFO [train.py:1198] (2/4) Epoch 28, batch 600, loss[loss=0.2328, ctc_loss=0.1503, cr_loss=0.4121, over 16553.00 frames. ], tot_loss[loss=0.2056, ctc_loss=0.1351, cr_loss=0.3525, over 3174734.84 frames. ], batch size: 66, lr: 4.29e-03, grad_scale: 16.0 2024-09-24 11:44:28,411 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=493794.0, ans=0.0 2024-09-24 11:44:30,355 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=10.86 vs. limit=22.5 2024-09-24 11:45:04,823 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.115e+02 1.264e+02 1.373e+02 1.490e+02 2.458e+02, threshold=2.746e+02, percent-clipped=0.0 2024-09-24 11:45:11,168 INFO [train.py:1198] (2/4) Epoch 28, batch 650, loss[loss=0.1867, ctc_loss=0.1194, cr_loss=0.3367, over 17273.00 frames. ], tot_loss[loss=0.2045, ctc_loss=0.1342, cr_loss=0.351, over 3212829.03 frames. ], batch size: 46, lr: 4.29e-03, grad_scale: 16.0 2024-09-24 11:45:11,529 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=493934.0, ans=0.125 2024-09-24 11:45:29,396 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.99 vs. limit=15.0 2024-09-24 11:45:30,769 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=493980.6666666667, ans=0.125 2024-09-24 11:46:07,491 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=494074.0, ans=0.0 2024-09-24 11:46:07,573 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=494074.0, ans=0.0 2024-09-24 11:46:15,501 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=494120.6666666667, ans=0.025 2024-09-24 11:46:23,666 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.min_positive, batch_count=494120.6666666667, ans=0.05 2024-09-24 11:46:31,266 INFO [train.py:1198] (2/4) Epoch 28, batch 700, loss[loss=0.2167, ctc_loss=0.1467, cr_loss=0.3501, over 16064.00 frames. ], tot_loss[loss=0.2046, ctc_loss=0.1343, cr_loss=0.3512, over 3241939.17 frames. ], batch size: 74, lr: 4.29e-03, grad_scale: 16.0 2024-09-24 11:46:36,337 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.max_abs, batch_count=494167.3333333333, ans=10.0 2024-09-24 11:47:02,970 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=19.32 vs. limit=22.5 2024-09-24 11:47:18,288 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=512, metric=3.07 vs. limit=15.0 2024-09-24 11:47:41,580 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=494354.0, ans=0.1 2024-09-24 11:47:52,208 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.257e+02 1.354e+02 1.480e+02 2.179e+02, threshold=2.707e+02, percent-clipped=0.0 2024-09-24 11:47:58,822 INFO [train.py:1198] (2/4) Epoch 28, batch 750, loss[loss=0.1855, ctc_loss=0.1183, cr_loss=0.336, over 16954.00 frames. ], tot_loss[loss=0.2029, ctc_loss=0.133, cr_loss=0.3498, over 3275022.69 frames. ], batch size: 42, lr: 4.29e-03, grad_scale: 16.0 2024-09-24 11:48:04,127 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=494400.6666666667, ans=0.1 2024-09-24 11:48:04,357 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.40 vs. limit=10.0 2024-09-24 11:48:09,600 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=4.26 vs. limit=15.0 2024-09-24 11:48:24,717 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=494447.3333333333, ans=0.125 2024-09-24 11:48:35,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=494494.0, ans=0.0 2024-09-24 11:49:06,157 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=494587.3333333333, ans=0.1 2024-09-24 11:49:18,698 INFO [train.py:1198] (2/4) Epoch 28, batch 800, loss[loss=0.1852, ctc_loss=0.1207, cr_loss=0.3226, over 17189.00 frames. ], tot_loss[loss=0.2005, ctc_loss=0.1313, cr_loss=0.3462, over 3298423.61 frames. ], batch size: 41, lr: 4.29e-03, grad_scale: 32.0 2024-09-24 11:49:49,115 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=494727.3333333333, ans=0.0 2024-09-24 11:50:02,505 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=494727.3333333333, ans=0.05 2024-09-24 11:50:13,538 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=494774.0, ans=0.0 2024-09-24 11:50:29,322 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=494820.6666666667, ans=0.125 2024-09-24 11:50:33,710 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.239e+02 1.301e+02 1.424e+02 2.595e+02, threshold=2.602e+02, percent-clipped=0.0 2024-09-24 11:50:34,192 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 11:50:40,254 INFO [train.py:1198] (2/4) Epoch 28, batch 850, loss[loss=0.1866, ctc_loss=0.1206, cr_loss=0.33, over 17217.00 frames. ], tot_loss[loss=0.2009, ctc_loss=0.1315, cr_loss=0.3468, over 3318192.45 frames. ], batch size: 47, lr: 4.29e-03, grad_scale: 32.0 2024-09-24 11:50:40,651 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=494867.3333333333, ans=0.125 2024-09-24 11:50:42,240 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=494867.3333333333, ans=0.1 2024-09-24 11:51:01,254 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=494914.0, ans=0.125 2024-09-24 11:51:17,169 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=494960.6666666667, ans=0.125 2024-09-24 11:51:39,210 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=495007.3333333333, ans=0.125 2024-09-24 11:51:52,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=495054.0, ans=0.125 2024-09-24 11:51:52,967 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=495054.0, ans=0.125 2024-09-24 11:52:04,921 INFO [train.py:1198] (2/4) Epoch 28, batch 900, loss[loss=0.1844, ctc_loss=0.1189, cr_loss=0.3274, over 17082.00 frames. ], tot_loss[loss=0.2021, ctc_loss=0.1324, cr_loss=0.3485, over 3328498.06 frames. ], batch size: 46, lr: 4.28e-03, grad_scale: 16.0 2024-09-24 11:52:10,034 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=495100.6666666667, ans=0.125 2024-09-24 11:52:17,866 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=495100.6666666667, ans=0.0 2024-09-24 11:52:35,813 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 11:53:17,347 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=495287.3333333333, ans=0.025 2024-09-24 11:53:20,248 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.272e+02 1.349e+02 1.497e+02 2.522e+02, threshold=2.699e+02, percent-clipped=0.0 2024-09-24 11:53:25,042 INFO [train.py:1198] (2/4) Epoch 28, batch 950, loss[loss=0.1886, ctc_loss=0.1228, cr_loss=0.3289, over 17229.00 frames. ], tot_loss[loss=0.2025, ctc_loss=0.1326, cr_loss=0.3492, over 3346352.65 frames. ], batch size: 47, lr: 4.28e-03, grad_scale: 16.0 2024-09-24 11:53:43,111 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=495380.6666666667, ans=0.1 2024-09-24 11:54:10,803 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.whiten.whitening_limit, batch_count=495427.3333333333, ans=12.0 2024-09-24 11:54:47,629 INFO [train.py:1198] (2/4) Epoch 28, batch 1000, loss[loss=0.2302, ctc_loss=0.1492, cr_loss=0.4048, over 17014.00 frames. ], tot_loss[loss=0.2017, ctc_loss=0.132, cr_loss=0.3484, over 3347355.01 frames. ], batch size: 44, lr: 4.28e-03, grad_scale: 16.0 2024-09-24 11:54:57,645 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 11:55:00,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=495567.3333333333, ans=0.125 2024-09-24 11:55:34,418 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.max_positive, batch_count=495707.3333333333, ans=0.95 2024-09-24 11:55:40,947 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=495707.3333333333, ans=0.125 2024-09-24 11:56:02,811 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.044e+02 1.229e+02 1.298e+02 1.405e+02 1.761e+02, threshold=2.595e+02, percent-clipped=0.0 2024-09-24 11:56:07,659 INFO [train.py:1198] (2/4) Epoch 28, batch 1050, loss[loss=0.1934, ctc_loss=0.1262, cr_loss=0.3357, over 17183.00 frames. ], tot_loss[loss=0.2022, ctc_loss=0.1324, cr_loss=0.3492, over 3352092.23 frames. ], batch size: 41, lr: 4.28e-03, grad_scale: 16.0 2024-09-24 11:56:15,069 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=14.26 vs. limit=22.5 2024-09-24 11:56:38,850 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.whiten.whitening_limit, batch_count=495894.0, ans=12.0 2024-09-24 11:56:47,327 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=495894.0, ans=0.125 2024-09-24 11:57:35,041 INFO [train.py:1198] (2/4) Epoch 28, batch 1100, loss[loss=0.1885, ctc_loss=0.1219, cr_loss=0.3327, over 17072.00 frames. ], tot_loss[loss=0.2018, ctc_loss=0.132, cr_loss=0.3486, over 3352837.68 frames. ], batch size: 46, lr: 4.28e-03, grad_scale: 16.0 2024-09-24 11:57:48,092 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=496034.0, ans=0.125 2024-09-24 11:58:08,464 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=496127.3333333333, ans=0.125 2024-09-24 11:58:11,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=496127.3333333333, ans=0.125 2024-09-24 11:58:14,311 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.76 vs. limit=6.0 2024-09-24 11:58:21,206 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=496174.0, ans=0.2 2024-09-24 11:58:26,258 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=496174.0, ans=0.0 2024-09-24 11:58:41,465 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=4.95 vs. limit=15.0 2024-09-24 11:58:47,384 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=496220.6666666667, ans=0.1 2024-09-24 11:58:50,153 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.267e+02 1.365e+02 1.472e+02 3.756e+02, threshold=2.729e+02, percent-clipped=1.0 2024-09-24 11:58:54,888 INFO [train.py:1198] (2/4) Epoch 28, batch 1150, loss[loss=0.2123, ctc_loss=0.1401, cr_loss=0.3612, over 17262.00 frames. ], tot_loss[loss=0.2022, ctc_loss=0.1323, cr_loss=0.3492, over 3351627.00 frames. ], batch size: 44, lr: 4.28e-03, grad_scale: 16.0 2024-09-24 11:59:03,369 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=496267.3333333333, ans=0.1 2024-09-24 11:59:07,386 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=5.27 vs. limit=10.0 2024-09-24 11:59:13,186 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=496314.0, ans=0.2 2024-09-24 11:59:35,272 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=496360.6666666667, ans=0.125 2024-09-24 11:59:44,172 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=496407.3333333333, ans=0.1 2024-09-24 11:59:47,274 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=496407.3333333333, ans=0.125 2024-09-24 11:59:58,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=496407.3333333333, ans=0.125 2024-09-24 12:00:17,102 INFO [train.py:1198] (2/4) Epoch 28, batch 1200, loss[loss=0.1736, ctc_loss=0.1124, cr_loss=0.306, over 17094.00 frames. ], tot_loss[loss=0.2014, ctc_loss=0.1318, cr_loss=0.3481, over 3355537.13 frames. ], batch size: 43, lr: 4.28e-03, grad_scale: 32.0 2024-09-24 12:00:20,664 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=496500.6666666667, ans=0.0 2024-09-24 12:00:22,232 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=496500.6666666667, ans=0.125 2024-09-24 12:00:50,107 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.70 vs. limit=6.0 2024-09-24 12:00:54,109 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=496594.0, ans=0.125 2024-09-24 12:00:55,698 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=496594.0, ans=0.125 2024-09-24 12:00:57,231 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.min_positive, batch_count=496594.0, ans=0.05 2024-09-24 12:01:07,407 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=5.45 vs. limit=12.0 2024-09-24 12:01:13,076 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=496640.6666666667, ans=0.125 2024-09-24 12:01:26,150 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=496687.3333333333, ans=0.125 2024-09-24 12:01:32,308 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.095e+02 1.289e+02 1.355e+02 1.454e+02 3.122e+02, threshold=2.710e+02, percent-clipped=1.0 2024-09-24 12:01:37,158 INFO [train.py:1198] (2/4) Epoch 28, batch 1250, loss[loss=0.1805, ctc_loss=0.1175, cr_loss=0.3153, over 17164.00 frames. ], tot_loss[loss=0.2011, ctc_loss=0.1316, cr_loss=0.3476, over 3356651.92 frames. ], batch size: 45, lr: 4.28e-03, grad_scale: 32.0 2024-09-24 12:01:37,501 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=496734.0, ans=0.07 2024-09-24 12:02:13,196 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=496780.6666666667, ans=0.125 2024-09-24 12:02:28,051 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=16.51 vs. limit=22.5 2024-09-24 12:03:03,730 INFO [train.py:1198] (2/4) Epoch 28, batch 1300, loss[loss=0.1693, ctc_loss=0.1085, cr_loss=0.3041, over 16297.00 frames. ], tot_loss[loss=0.2019, ctc_loss=0.1322, cr_loss=0.3489, over 3349248.94 frames. ], batch size: 36, lr: 4.28e-03, grad_scale: 32.0 2024-09-24 12:03:21,433 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=497014.0, ans=0.0 2024-09-24 12:03:29,191 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=497014.0, ans=0.125 2024-09-24 12:04:11,615 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten.whitening_limit, batch_count=497154.0, ans=22.5 2024-09-24 12:04:18,678 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.141e+02 1.268e+02 1.342e+02 1.439e+02 2.491e+02, threshold=2.683e+02, percent-clipped=0.0 2024-09-24 12:04:20,631 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=497154.0, ans=0.1 2024-09-24 12:04:23,559 INFO [train.py:1198] (2/4) Epoch 28, batch 1350, loss[loss=0.2178, ctc_loss=0.1435, cr_loss=0.3714, over 16537.00 frames. ], tot_loss[loss=0.2017, ctc_loss=0.132, cr_loss=0.3486, over 3346416.38 frames. ], batch size: 66, lr: 4.28e-03, grad_scale: 32.0 2024-09-24 12:04:27,221 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=497200.6666666667, ans=0.125 2024-09-24 12:04:36,342 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=497200.6666666667, ans=0.5 2024-09-24 12:04:42,795 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=497247.3333333333, ans=0.025 2024-09-24 12:05:12,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=497340.6666666667, ans=0.125 2024-09-24 12:05:17,514 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=497340.6666666667, ans=0.125 2024-09-24 12:05:17,525 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=497340.6666666667, ans=0.07 2024-09-24 12:05:36,730 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=497387.3333333333, ans=0.125 2024-09-24 12:05:39,932 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=497387.3333333333, ans=0.125 2024-09-24 12:05:45,903 INFO [train.py:1198] (2/4) Epoch 28, batch 1400, loss[loss=0.1992, ctc_loss=0.1314, cr_loss=0.339, over 17057.00 frames. ], tot_loss[loss=0.2031, ctc_loss=0.133, cr_loss=0.3504, over 3341938.43 frames. ], batch size: 46, lr: 4.27e-03, grad_scale: 16.0 2024-09-24 12:05:46,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=497434.0, ans=0.2 2024-09-24 12:06:07,154 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=497480.6666666667, ans=0.125 2024-09-24 12:06:18,271 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=497527.3333333333, ans=0.0 2024-09-24 12:06:19,130 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=6.49 vs. limit=15.0 2024-09-24 12:06:21,462 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=497527.3333333333, ans=0.125 2024-09-24 12:06:24,755 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=497527.3333333333, ans=0.0 2024-09-24 12:06:31,255 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=497527.3333333333, ans=0.125 2024-09-24 12:06:43,109 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=497574.0, ans=0.0 2024-09-24 12:06:44,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=497574.0, ans=0.2 2024-09-24 12:07:05,080 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=497620.6666666667, ans=0.125 2024-09-24 12:07:08,054 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.131e+02 1.287e+02 1.409e+02 1.543e+02 2.036e+02, threshold=2.818e+02, percent-clipped=0.0 2024-09-24 12:07:11,207 INFO [train.py:1198] (2/4) Epoch 28, batch 1450, loss[loss=0.22, ctc_loss=0.1429, cr_loss=0.3856, over 17012.00 frames. ], tot_loss[loss=0.2037, ctc_loss=0.1332, cr_loss=0.3523, over 3350171.03 frames. ], batch size: 51, lr: 4.27e-03, grad_scale: 16.0 2024-09-24 12:07:12,294 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=12.06 vs. limit=22.5 2024-09-24 12:07:38,567 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=497714.0, ans=0.2 2024-09-24 12:07:43,430 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=497760.6666666667, ans=0.125 2024-09-24 12:07:49,814 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=497760.6666666667, ans=0.1 2024-09-24 12:08:04,076 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=497807.3333333333, ans=0.2 2024-09-24 12:08:10,381 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=497807.3333333333, ans=0.125 2024-09-24 12:08:15,165 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=497854.0, ans=0.125 2024-09-24 12:08:30,997 INFO [train.py:1198] (2/4) Epoch 28, batch 1500, loss[loss=0.1478, ctc_loss=0.0947, cr_loss=0.2656, over 17256.00 frames. ], tot_loss[loss=0.2024, ctc_loss=0.1323, cr_loss=0.3502, over 3354717.55 frames. ], batch size: 42, lr: 4.27e-03, grad_scale: 16.0 2024-09-24 12:08:49,566 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=5.09 vs. limit=10.0 2024-09-24 12:08:58,817 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=497947.3333333333, ans=0.0 2024-09-24 12:09:44,588 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=498087.3333333333, ans=0.125 2024-09-24 12:09:50,698 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.155e+02 1.264e+02 1.369e+02 1.478e+02 2.054e+02, threshold=2.738e+02, percent-clipped=0.0 2024-09-24 12:09:53,873 INFO [train.py:1198] (2/4) Epoch 28, batch 1550, loss[loss=0.2199, ctc_loss=0.1457, cr_loss=0.371, over 16870.00 frames. ], tot_loss[loss=0.2023, ctc_loss=0.1323, cr_loss=0.3501, over 3354719.48 frames. ], batch size: 58, lr: 4.27e-03, grad_scale: 16.0 2024-09-24 12:10:02,248 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=498134.0, ans=0.125 2024-09-24 12:10:24,800 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=498227.3333333333, ans=0.5 2024-09-24 12:10:31,790 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=7.57 vs. limit=12.0 2024-09-24 12:10:43,904 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=498274.0, ans=0.04949747468305833 2024-09-24 12:10:45,532 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=498274.0, ans=0.125 2024-09-24 12:10:48,779 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=498274.0, ans=0.125 2024-09-24 12:11:00,036 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=498320.6666666667, ans=0.125 2024-09-24 12:11:13,886 INFO [train.py:1198] (2/4) Epoch 28, batch 1600, loss[loss=0.2192, ctc_loss=0.1428, cr_loss=0.3819, over 17356.00 frames. ], tot_loss[loss=0.2024, ctc_loss=0.1323, cr_loss=0.3502, over 3348565.93 frames. ], batch size: 48, lr: 4.27e-03, grad_scale: 32.0 2024-09-24 12:11:55,596 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=498460.6666666667, ans=0.125 2024-09-24 12:12:11,261 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=498507.3333333333, ans=0.1 2024-09-24 12:12:16,099 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=498507.3333333333, ans=0.0 2024-09-24 12:12:36,985 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=498554.0, ans=0.1 2024-09-24 12:12:38,198 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.239e+02 1.329e+02 1.427e+02 2.041e+02, threshold=2.658e+02, percent-clipped=0.0 2024-09-24 12:12:41,510 INFO [train.py:1198] (2/4) Epoch 28, batch 1650, loss[loss=0.2364, ctc_loss=0.1566, cr_loss=0.399, over 17023.00 frames. ], tot_loss[loss=0.2014, ctc_loss=0.1316, cr_loss=0.3492, over 3356445.67 frames. ], batch size: 53, lr: 4.27e-03, grad_scale: 32.0 2024-09-24 12:12:45,035 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=498600.6666666667, ans=0.125 2024-09-24 12:13:03,877 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=498647.3333333333, ans=0.0 2024-09-24 12:13:05,439 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=498647.3333333333, ans=0.125 2024-09-24 12:13:15,182 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.min_positive, batch_count=498694.0, ans=0.05 2024-09-24 12:13:29,416 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=498740.6666666667, ans=0.1 2024-09-24 12:13:32,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=498740.6666666667, ans=0.125 2024-09-24 12:13:39,094 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=498740.6666666667, ans=0.1 2024-09-24 12:13:51,548 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=498787.3333333333, ans=0.0 2024-09-24 12:13:56,169 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=498787.3333333333, ans=0.025 2024-09-24 12:14:00,723 INFO [train.py:1198] (2/4) Epoch 28, batch 1700, loss[loss=0.2158, ctc_loss=0.1423, cr_loss=0.3675, over 17009.00 frames. ], tot_loss[loss=0.2015, ctc_loss=0.1316, cr_loss=0.3492, over 3361589.78 frames. ], batch size: 53, lr: 4.27e-03, grad_scale: 32.0 2024-09-24 12:14:08,096 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten.whitening_limit, batch_count=498834.0, ans=22.5 2024-09-24 12:14:40,860 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.42 vs. limit=15.0 2024-09-24 12:14:44,794 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=498927.3333333333, ans=0.125 2024-09-24 12:14:57,646 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=498974.0, ans=0.1 2024-09-24 12:14:59,231 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=498974.0, ans=0.0 2024-09-24 12:15:00,767 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=498974.0, ans=0.0 2024-09-24 12:15:19,451 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.071e+02 1.240e+02 1.314e+02 1.418e+02 1.778e+02, threshold=2.628e+02, percent-clipped=0.0 2024-09-24 12:15:22,615 INFO [train.py:1198] (2/4) Epoch 28, batch 1750, loss[loss=0.2019, ctc_loss=0.1325, cr_loss=0.347, over 17200.00 frames. ], tot_loss[loss=0.2019, ctc_loss=0.1321, cr_loss=0.3494, over 3357572.05 frames. ], batch size: 47, lr: 4.27e-03, grad_scale: 32.0 2024-09-24 12:15:35,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=499067.3333333333, ans=0.125 2024-09-24 12:15:43,529 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=499114.0, ans=0.0 2024-09-24 12:16:47,611 INFO [train.py:1198] (2/4) Epoch 28, batch 1800, loss[loss=0.1628, ctc_loss=0.1023, cr_loss=0.3021, over 17016.00 frames. ], tot_loss[loss=0.2028, ctc_loss=0.1327, cr_loss=0.3505, over 3355400.03 frames. ], batch size: 39, lr: 4.27e-03, grad_scale: 32.0 2024-09-24 12:16:54,259 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=499300.6666666667, ans=0.125 2024-09-24 12:17:05,404 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=499347.3333333333, ans=0.0 2024-09-24 12:17:11,880 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=499347.3333333333, ans=0.125 2024-09-24 12:17:59,333 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=499487.3333333333, ans=0.125 2024-09-24 12:17:59,456 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 12:18:03,603 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.272e+02 1.347e+02 1.424e+02 2.124e+02, threshold=2.693e+02, percent-clipped=0.0 2024-09-24 12:18:06,858 INFO [train.py:1198] (2/4) Epoch 28, batch 1850, loss[loss=0.1739, ctc_loss=0.1123, cr_loss=0.308, over 17304.00 frames. ], tot_loss[loss=0.2031, ctc_loss=0.1329, cr_loss=0.351, over 3361636.34 frames. ], batch size: 42, lr: 4.27e-03, grad_scale: 32.0 2024-09-24 12:18:10,379 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=499534.0, ans=0.07 2024-09-24 12:18:16,754 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=499534.0, ans=0.1 2024-09-24 12:18:29,403 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=499580.6666666667, ans=0.0 2024-09-24 12:19:12,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=499720.6666666667, ans=0.125 2024-09-24 12:19:17,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=499720.6666666667, ans=0.1 2024-09-24 12:19:29,608 INFO [train.py:1198] (2/4) Epoch 28, batch 1900, loss[loss=0.2103, ctc_loss=0.1385, cr_loss=0.3587, over 17060.00 frames. ], tot_loss[loss=0.2028, ctc_loss=0.1327, cr_loss=0.3504, over 3366120.14 frames. ], batch size: 52, lr: 4.26e-03, grad_scale: 32.0 2024-09-24 12:19:29,983 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=499767.3333333333, ans=0.1 2024-09-24 12:19:34,862 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=499767.3333333333, ans=0.0 2024-09-24 12:19:57,401 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=499814.0, ans=0.1 2024-09-24 12:19:59,002 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=499814.0, ans=0.125 2024-09-24 12:20:32,302 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=499954.0, ans=0.1 2024-09-24 12:20:32,369 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=499954.0, ans=0.0 2024-09-24 12:20:46,566 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.102e+02 1.256e+02 1.349e+02 1.450e+02 2.234e+02, threshold=2.698e+02, percent-clipped=0.0 2024-09-24 12:20:49,741 INFO [train.py:1198] (2/4) Epoch 28, batch 1950, loss[loss=0.1747, ctc_loss=0.111, cr_loss=0.3188, over 17102.00 frames. ], tot_loss[loss=0.2021, ctc_loss=0.1323, cr_loss=0.3493, over 3362302.87 frames. ], batch size: 40, lr: 4.26e-03, grad_scale: 32.0 2024-09-24 12:21:42,077 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=500140.6666666667, ans=0.0 2024-09-24 12:22:06,973 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=500187.3333333333, ans=0.2 2024-09-24 12:22:17,929 INFO [train.py:1198] (2/4) Epoch 28, batch 2000, loss[loss=0.205, ctc_loss=0.134, cr_loss=0.355, over 17043.00 frames. ], tot_loss[loss=0.2025, ctc_loss=0.1326, cr_loss=0.3497, over 3350179.31 frames. ], batch size: 56, lr: 4.26e-03, grad_scale: 32.0 2024-09-24 12:22:31,354 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=500234.0, ans=0.125 2024-09-24 12:22:35,137 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.87 vs. limit=6.0 2024-09-24 12:22:41,063 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=500280.6666666667, ans=0.05 2024-09-24 12:22:58,787 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=500327.3333333333, ans=0.025 2024-09-24 12:23:08,484 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.08 vs. limit=6.0 2024-09-24 12:23:21,135 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=500420.6666666667, ans=0.2 2024-09-24 12:23:29,009 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=500420.6666666667, ans=0.125 2024-09-24 12:23:35,003 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.274e+02 1.350e+02 1.459e+02 2.226e+02, threshold=2.700e+02, percent-clipped=0.0 2024-09-24 12:23:38,212 INFO [train.py:1198] (2/4) Epoch 28, batch 2050, loss[loss=0.2074, ctc_loss=0.1342, cr_loss=0.3657, over 17175.00 frames. ], tot_loss[loss=0.2014, ctc_loss=0.1319, cr_loss=0.3476, over 3348245.85 frames. ], batch size: 45, lr: 4.26e-03, grad_scale: 32.0 2024-09-24 12:24:00,737 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=500514.0, ans=0.125 2024-09-24 12:24:05,347 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=500514.0, ans=0.2 2024-09-24 12:24:33,649 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=500607.3333333333, ans=0.125 2024-09-24 12:24:35,332 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=500607.3333333333, ans=0.0 2024-09-24 12:24:41,684 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=500607.3333333333, ans=0.125 2024-09-24 12:24:46,513 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 12:24:56,090 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=500654.0, ans=0.125 2024-09-24 12:24:56,149 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=500654.0, ans=0.0 2024-09-24 12:25:00,539 INFO [train.py:1198] (2/4) Epoch 28, batch 2100, loss[loss=0.2076, ctc_loss=0.1344, cr_loss=0.3658, over 17264.00 frames. ], tot_loss[loss=0.2021, ctc_loss=0.1324, cr_loss=0.3487, over 3349069.44 frames. ], batch size: 44, lr: 4.26e-03, grad_scale: 32.0 2024-09-24 12:25:26,367 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=500747.3333333333, ans=0.125 2024-09-24 12:25:48,881 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.54 vs. limit=12.0 2024-09-24 12:25:54,620 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=500840.6666666667, ans=0.0 2024-09-24 12:26:04,079 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=500887.3333333333, ans=0.1 2024-09-24 12:26:16,783 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.122e+02 1.293e+02 1.395e+02 1.538e+02 2.200e+02, threshold=2.790e+02, percent-clipped=0.0 2024-09-24 12:26:25,163 INFO [train.py:1198] (2/4) Epoch 28, batch 2150, loss[loss=0.1845, ctc_loss=0.1206, cr_loss=0.3192, over 17021.00 frames. ], tot_loss[loss=0.2029, ctc_loss=0.1329, cr_loss=0.3501, over 3348114.00 frames. ], batch size: 44, lr: 4.26e-03, grad_scale: 32.0 2024-09-24 12:26:43,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=500980.6666666667, ans=0.125 2024-09-24 12:26:45,445 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=500980.6666666667, ans=0.125 2024-09-24 12:26:57,011 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.90 vs. limit=15.0 2024-09-24 12:27:01,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.max_positive, batch_count=501027.3333333333, ans=0.95 2024-09-24 12:27:17,260 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=501074.0, ans=0.125 2024-09-24 12:27:28,760 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=501074.0, ans=0.2 2024-09-24 12:27:47,640 INFO [train.py:1198] (2/4) Epoch 28, batch 2200, loss[loss=0.2267, ctc_loss=0.1501, cr_loss=0.3831, over 17196.00 frames. ], tot_loss[loss=0.2025, ctc_loss=0.1326, cr_loss=0.3495, over 3348132.91 frames. ], batch size: 55, lr: 4.26e-03, grad_scale: 16.0 2024-09-24 12:28:20,068 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=501260.6666666667, ans=0.125 2024-09-24 12:28:31,276 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=501260.6666666667, ans=0.125 2024-09-24 12:28:47,036 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=501307.3333333333, ans=0.2 2024-09-24 12:28:51,141 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=19.63 vs. limit=22.5 2024-09-24 12:28:57,326 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.17 vs. limit=15.0 2024-09-24 12:29:06,263 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.156e+02 1.288e+02 1.410e+02 1.537e+02 2.040e+02, threshold=2.819e+02, percent-clipped=0.0 2024-09-24 12:29:10,548 INFO [train.py:1198] (2/4) Epoch 28, batch 2250, loss[loss=0.2335, ctc_loss=0.157, cr_loss=0.3824, over 17036.00 frames. ], tot_loss[loss=0.2039, ctc_loss=0.1336, cr_loss=0.3515, over 3350043.81 frames. ], batch size: 51, lr: 4.26e-03, grad_scale: 16.0 2024-09-24 12:29:23,765 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=501400.6666666667, ans=0.2 2024-09-24 12:29:33,585 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=501447.3333333333, ans=0.1 2024-09-24 12:29:36,672 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=501447.3333333333, ans=0.1 2024-09-24 12:29:41,502 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=501494.0, ans=0.0 2024-09-24 12:29:54,301 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 12:29:54,384 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=501494.0, ans=0.1 2024-09-24 12:30:00,877 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=501540.6666666667, ans=0.07 2024-09-24 12:30:02,390 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-24 12:30:03,775 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=501540.6666666667, ans=0.0 2024-09-24 12:30:26,639 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=10.61 vs. limit=15.0 2024-09-24 12:30:30,751 INFO [train.py:1198] (2/4) Epoch 28, batch 2300, loss[loss=0.2275, ctc_loss=0.1514, cr_loss=0.3807, over 17046.00 frames. ], tot_loss[loss=0.2035, ctc_loss=0.1333, cr_loss=0.3511, over 3354684.35 frames. ], batch size: 52, lr: 4.26e-03, grad_scale: 16.0 2024-09-24 12:31:04,781 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=501727.3333333333, ans=0.125 2024-09-24 12:31:27,507 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=501774.0, ans=0.125 2024-09-24 12:31:50,795 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=501820.6666666667, ans=0.125 2024-09-24 12:31:52,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=501820.6666666667, ans=0.0 2024-09-24 12:31:56,841 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.285e+02 1.365e+02 1.487e+02 3.397e+02, threshold=2.730e+02, percent-clipped=1.0 2024-09-24 12:31:58,415 INFO [train.py:1198] (2/4) Epoch 28, batch 2350, loss[loss=0.2098, ctc_loss=0.1414, cr_loss=0.3421, over 16765.00 frames. ], tot_loss[loss=0.2036, ctc_loss=0.1333, cr_loss=0.3514, over 3360579.12 frames. ], batch size: 61, lr: 4.26e-03, grad_scale: 16.0 2024-09-24 12:32:09,684 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=501867.3333333333, ans=0.05 2024-09-24 12:32:11,343 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=501867.3333333333, ans=0.2 2024-09-24 12:32:13,088 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=501914.0, ans=0.125 2024-09-24 12:32:19,516 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 12:32:27,229 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=501914.0, ans=0.1 2024-09-24 12:32:35,230 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=501960.6666666667, ans=0.125 2024-09-24 12:33:07,351 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=11.79 vs. limit=15.0 2024-09-24 12:33:17,990 INFO [train.py:1198] (2/4) Epoch 28, batch 2400, loss[loss=0.2079, ctc_loss=0.1374, cr_loss=0.3524, over 17012.00 frames. ], tot_loss[loss=0.2024, ctc_loss=0.1325, cr_loss=0.3498, over 3359658.28 frames. ], batch size: 51, lr: 4.25e-03, grad_scale: 32.0 2024-09-24 12:33:31,477 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=502100.6666666667, ans=0.2 2024-09-24 12:33:47,205 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=502147.3333333333, ans=0.125 2024-09-24 12:34:09,910 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.12 vs. limit=15.0 2024-09-24 12:34:39,437 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.025e+02 1.258e+02 1.381e+02 1.475e+02 2.172e+02, threshold=2.761e+02, percent-clipped=0.0 2024-09-24 12:34:41,086 INFO [train.py:1198] (2/4) Epoch 28, batch 2450, loss[loss=0.1981, ctc_loss=0.1295, cr_loss=0.3432, over 17348.00 frames. ], tot_loss[loss=0.2022, ctc_loss=0.1323, cr_loss=0.3497, over 3358527.39 frames. ], batch size: 48, lr: 4.25e-03, grad_scale: 32.0 2024-09-24 12:34:52,785 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=13.17 vs. limit=22.5 2024-09-24 12:35:05,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=502380.6666666667, ans=0.125 2024-09-24 12:35:08,319 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=502380.6666666667, ans=0.125 2024-09-24 12:35:45,489 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=4.60 vs. limit=12.0 2024-09-24 12:35:48,274 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=502520.6666666667, ans=0.0 2024-09-24 12:36:00,605 INFO [train.py:1198] (2/4) Epoch 28, batch 2500, loss[loss=0.2273, ctc_loss=0.1504, cr_loss=0.3844, over 17258.00 frames. ], tot_loss[loss=0.2024, ctc_loss=0.1325, cr_loss=0.3494, over 3359763.08 frames. ], batch size: 44, lr: 4.25e-03, grad_scale: 16.0 2024-09-24 12:36:00,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=502567.3333333333, ans=0.025 2024-09-24 12:36:22,171 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 12:36:27,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=502614.0, ans=0.09899494936611666 2024-09-24 12:37:24,887 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=8.13 vs. limit=8.0 2024-09-24 12:37:28,344 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.066e+02 1.269e+02 1.387e+02 1.478e+02 2.068e+02, threshold=2.774e+02, percent-clipped=0.0 2024-09-24 12:37:28,369 INFO [train.py:1198] (2/4) Epoch 28, batch 2550, loss[loss=0.2034, ctc_loss=0.1315, cr_loss=0.3596, over 17289.00 frames. ], tot_loss[loss=0.2017, ctc_loss=0.132, cr_loss=0.3486, over 3360752.95 frames. ], batch size: 46, lr: 4.25e-03, grad_scale: 16.0 2024-09-24 12:37:32,788 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=4.71 vs. limit=15.0 2024-09-24 12:37:38,452 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=502800.6666666667, ans=0.125 2024-09-24 12:37:47,870 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=502847.3333333333, ans=0.2 2024-09-24 12:37:52,569 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=502847.3333333333, ans=0.0 2024-09-24 12:38:00,544 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=502894.0, ans=0.125 2024-09-24 12:38:24,280 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=502940.6666666667, ans=0.025 2024-09-24 12:38:29,165 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=502940.6666666667, ans=0.0 2024-09-24 12:38:48,319 INFO [train.py:1198] (2/4) Epoch 28, batch 2600, loss[loss=0.1795, ctc_loss=0.1171, cr_loss=0.312, over 17168.00 frames. ], tot_loss[loss=0.2012, ctc_loss=0.1316, cr_loss=0.3477, over 3362090.93 frames. ], batch size: 45, lr: 4.25e-03, grad_scale: 16.0 2024-09-24 12:39:00,887 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=503034.0, ans=0.125 2024-09-24 12:39:21,589 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=503127.3333333333, ans=0.2 2024-09-24 12:39:34,603 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=503127.3333333333, ans=0.2 2024-09-24 12:39:53,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=503220.6666666667, ans=0.0 2024-09-24 12:40:01,699 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=503220.6666666667, ans=0.125 2024-09-24 12:40:03,283 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=503220.6666666667, ans=0.0 2024-09-24 12:40:04,947 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=503220.6666666667, ans=0.1 2024-09-24 12:40:10,912 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.254e+02 1.337e+02 1.417e+02 2.017e+02, threshold=2.675e+02, percent-clipped=0.0 2024-09-24 12:40:10,936 INFO [train.py:1198] (2/4) Epoch 28, batch 2650, loss[loss=0.186, ctc_loss=0.1179, cr_loss=0.3406, over 17291.00 frames. ], tot_loss[loss=0.2016, ctc_loss=0.132, cr_loss=0.3481, over 3360887.98 frames. ], batch size: 46, lr: 4.25e-03, grad_scale: 16.0 2024-09-24 12:40:22,423 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=503267.3333333333, ans=0.0 2024-09-24 12:40:35,300 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=503314.0, ans=0.125 2024-09-24 12:40:38,486 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=503314.0, ans=0.125 2024-09-24 12:40:46,932 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.62 vs. limit=15.0 2024-09-24 12:41:12,741 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=503407.3333333333, ans=0.125 2024-09-24 12:41:21,681 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=503454.0, ans=0.0 2024-09-24 12:41:38,425 INFO [train.py:1198] (2/4) Epoch 28, batch 2700, loss[loss=0.2117, ctc_loss=0.1449, cr_loss=0.3341, over 12263.00 frames. ], tot_loss[loss=0.2022, ctc_loss=0.1324, cr_loss=0.3489, over 3354018.56 frames. ], batch size: 123, lr: 4.25e-03, grad_scale: 16.0 2024-09-24 12:41:43,596 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=503500.6666666667, ans=0.125 2024-09-24 12:41:48,453 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 12:42:12,401 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=503594.0, ans=0.125 2024-09-24 12:42:47,681 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=503687.3333333333, ans=0.04949747468305833 2024-09-24 12:42:58,695 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.145e+02 1.278e+02 1.354e+02 1.486e+02 3.287e+02, threshold=2.708e+02, percent-clipped=2.0 2024-09-24 12:42:58,720 INFO [train.py:1198] (2/4) Epoch 28, batch 2750, loss[loss=0.2223, ctc_loss=0.1477, cr_loss=0.3729, over 17104.00 frames. ], tot_loss[loss=0.2023, ctc_loss=0.1325, cr_loss=0.3491, over 3351815.70 frames. ], batch size: 49, lr: 4.25e-03, grad_scale: 16.0 2024-09-24 12:43:00,636 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=503734.0, ans=0.125 2024-09-24 12:43:02,178 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=503734.0, ans=0.2 2024-09-24 12:43:05,186 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=503734.0, ans=0.125 2024-09-24 12:43:27,246 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=503780.6666666667, ans=0.125 2024-09-24 12:43:38,468 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=503827.3333333333, ans=0.1 2024-09-24 12:43:46,487 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=503874.0, ans=0.125 2024-09-24 12:43:49,872 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 12:44:00,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=503874.0, ans=0.125 2024-09-24 12:44:20,597 INFO [train.py:1198] (2/4) Epoch 28, batch 2800, loss[loss=0.1995, ctc_loss=0.1285, cr_loss=0.3553, over 16990.00 frames. ], tot_loss[loss=0.2018, ctc_loss=0.132, cr_loss=0.3487, over 3349794.43 frames. ], batch size: 51, lr: 4.25e-03, grad_scale: 32.0 2024-09-24 12:44:28,951 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=503967.3333333333, ans=0.125 2024-09-24 12:44:50,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=504014.0, ans=0.125 2024-09-24 12:45:15,051 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=5.29 vs. limit=10.0 2024-09-24 12:45:32,583 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.36 vs. limit=10.0 2024-09-24 12:45:33,868 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=504154.0, ans=0.125 2024-09-24 12:45:43,156 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.102e+02 1.267e+02 1.390e+02 1.555e+02 2.170e+02, threshold=2.781e+02, percent-clipped=0.0 2024-09-24 12:45:43,181 INFO [train.py:1198] (2/4) Epoch 28, batch 2850, loss[loss=0.1917, ctc_loss=0.1248, cr_loss=0.3345, over 17252.00 frames. ], tot_loss[loss=0.2019, ctc_loss=0.1321, cr_loss=0.3488, over 3349136.99 frames. ], batch size: 44, lr: 4.25e-03, grad_scale: 32.0 2024-09-24 12:45:49,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=504200.6666666667, ans=0.125 2024-09-24 12:45:53,024 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=504200.6666666667, ans=0.0 2024-09-24 12:46:42,480 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=504340.6666666667, ans=0.1 2024-09-24 12:47:01,397 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=504387.3333333333, ans=0.125 2024-09-24 12:47:10,934 INFO [train.py:1198] (2/4) Epoch 28, batch 2900, loss[loss=0.1721, ctc_loss=0.1087, cr_loss=0.3172, over 16329.00 frames. ], tot_loss[loss=0.2019, ctc_loss=0.1322, cr_loss=0.3486, over 3351689.72 frames. ], batch size: 36, lr: 4.24e-03, grad_scale: 32.0 2024-09-24 12:47:14,515 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.min_positive, batch_count=504434.0, ans=0.05 2024-09-24 12:47:16,009 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=504434.0, ans=0.025 2024-09-24 12:47:22,251 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=504434.0, ans=0.125 2024-09-24 12:47:37,725 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=15.10 vs. limit=22.5 2024-09-24 12:47:43,343 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=504527.3333333333, ans=0.125 2024-09-24 12:47:48,135 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=504527.3333333333, ans=0.0 2024-09-24 12:48:02,678 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=504574.0, ans=0.2 2024-09-24 12:48:18,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=504620.6666666667, ans=0.1 2024-09-24 12:48:23,906 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.59 vs. limit=15.0 2024-09-24 12:48:31,271 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.289e+02 1.394e+02 1.570e+02 3.133e+02, threshold=2.787e+02, percent-clipped=1.0 2024-09-24 12:48:31,295 INFO [train.py:1198] (2/4) Epoch 28, batch 2950, loss[loss=0.1876, ctc_loss=0.1192, cr_loss=0.3419, over 17029.00 frames. ], tot_loss[loss=0.2025, ctc_loss=0.1326, cr_loss=0.3494, over 3353676.70 frames. ], batch size: 39, lr: 4.24e-03, grad_scale: 32.0 2024-09-24 12:48:37,959 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=504667.3333333333, ans=0.125 2024-09-24 12:48:54,800 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=504714.0, ans=0.1 2024-09-24 12:49:20,053 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=504807.3333333333, ans=0.035 2024-09-24 12:49:53,070 INFO [train.py:1198] (2/4) Epoch 28, batch 3000, loss[loss=0.249, ctc_loss=0.1716, cr_loss=0.3873, over 11409.00 frames. ], tot_loss[loss=0.2032, ctc_loss=0.1332, cr_loss=0.3499, over 3345430.74 frames. ], batch size: 123, lr: 4.24e-03, grad_scale: 32.0 2024-09-24 12:49:53,071 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 12:50:08,094 INFO [train.py:1230] (2/4) Epoch 28, validation: loss=0.03718, ctc_loss=0.03718, cr_loss=8.452e-15, over 944034.00 frames. 2024-09-24 12:50:08,094 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 12:50:10,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=504900.6666666667, ans=0.0 2024-09-24 12:50:57,447 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=505040.6666666667, ans=0.0 2024-09-24 12:50:58,931 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=505040.6666666667, ans=0.125 2024-09-24 12:51:03,627 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=505040.6666666667, ans=0.125 2024-09-24 12:51:06,685 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=505040.6666666667, ans=0.1 2024-09-24 12:51:20,617 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=505087.3333333333, ans=0.07 2024-09-24 12:51:22,178 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=505087.3333333333, ans=0.125 2024-09-24 12:51:22,835 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.56 vs. limit=12.0 2024-09-24 12:51:25,426 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=505134.0, ans=0.2 2024-09-24 12:51:26,594 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.079e+02 1.294e+02 1.362e+02 1.471e+02 2.139e+02, threshold=2.723e+02, percent-clipped=0.0 2024-09-24 12:51:26,618 INFO [train.py:1198] (2/4) Epoch 28, batch 3050, loss[loss=0.2035, ctc_loss=0.1334, cr_loss=0.3503, over 17356.00 frames. ], tot_loss[loss=0.2022, ctc_loss=0.1324, cr_loss=0.3493, over 3360113.50 frames. ], batch size: 48, lr: 4.24e-03, grad_scale: 32.0 2024-09-24 12:51:28,531 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=505134.0, ans=0.2 2024-09-24 12:51:32,270 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.62 vs. limit=15.0 2024-09-24 12:51:48,787 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=505180.6666666667, ans=0.125 2024-09-24 12:51:51,932 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=505180.6666666667, ans=0.2 2024-09-24 12:51:56,393 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=505227.3333333333, ans=0.2 2024-09-24 12:52:22,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=505274.0, ans=0.125 2024-09-24 12:52:28,250 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=11.89 vs. limit=22.5 2024-09-24 12:52:48,943 INFO [train.py:1198] (2/4) Epoch 28, batch 3100, loss[loss=0.1747, ctc_loss=0.112, cr_loss=0.3134, over 17135.00 frames. ], tot_loss[loss=0.2027, ctc_loss=0.1326, cr_loss=0.3502, over 3360452.77 frames. ], batch size: 40, lr: 4.24e-03, grad_scale: 16.0 2024-09-24 12:52:55,928 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.31 vs. limit=12.0 2024-09-24 12:53:00,094 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=505367.3333333333, ans=0.125 2024-09-24 12:53:12,459 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=505414.0, ans=0.0 2024-09-24 12:53:55,664 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=505554.0, ans=0.125 2024-09-24 12:54:09,375 INFO [train.py:1198] (2/4) Epoch 28, batch 3150, loss[loss=0.218, ctc_loss=0.1422, cr_loss=0.3793, over 17217.00 frames. ], tot_loss[loss=0.2025, ctc_loss=0.1325, cr_loss=0.3502, over 3361852.22 frames. ], batch size: 47, lr: 4.24e-03, grad_scale: 16.0 2024-09-24 12:54:10,884 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.260e+02 1.340e+02 1.441e+02 3.228e+02, threshold=2.680e+02, percent-clipped=2.0 2024-09-24 12:54:51,738 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=505694.0, ans=0.09899494936611666 2024-09-24 12:54:58,478 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=9.88 vs. limit=15.0 2024-09-24 12:55:04,145 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=505740.6666666667, ans=0.125 2024-09-24 12:55:27,146 INFO [train.py:1198] (2/4) Epoch 28, batch 3200, loss[loss=0.2032, ctc_loss=0.1309, cr_loss=0.3617, over 17238.00 frames. ], tot_loss[loss=0.202, ctc_loss=0.132, cr_loss=0.3501, over 3361081.77 frames. ], batch size: 47, lr: 4.24e-03, grad_scale: 32.0 2024-09-24 12:55:28,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=505834.0, ans=0.0 2024-09-24 12:55:30,368 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=505834.0, ans=0.0 2024-09-24 12:55:43,053 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=505880.6666666667, ans=0.0 2024-09-24 12:56:11,854 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.58 vs. limit=10.0 2024-09-24 12:56:12,943 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=505974.0, ans=0.125 2024-09-24 12:56:25,686 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=10.48 vs. limit=22.5 2024-09-24 12:56:40,758 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=506020.6666666667, ans=0.125 2024-09-24 12:56:45,209 INFO [train.py:1198] (2/4) Epoch 28, batch 3250, loss[loss=0.1874, ctc_loss=0.1198, cr_loss=0.338, over 17202.00 frames. ], tot_loss[loss=0.2016, ctc_loss=0.1317, cr_loss=0.3495, over 3361394.47 frames. ], batch size: 41, lr: 4.24e-03, grad_scale: 32.0 2024-09-24 12:56:46,875 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.036e+02 1.261e+02 1.336e+02 1.430e+02 2.422e+02, threshold=2.672e+02, percent-clipped=0.0 2024-09-24 12:56:49,195 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.29 vs. limit=15.0 2024-09-24 12:57:01,514 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.70 vs. limit=10.0 2024-09-24 12:57:23,142 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=506160.6666666667, ans=0.125 2024-09-24 12:57:26,836 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=10.46 vs. limit=22.5 2024-09-24 12:57:34,331 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=506207.3333333333, ans=0.07 2024-09-24 12:57:45,761 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=8.15 vs. limit=15.0 2024-09-24 12:58:04,252 INFO [train.py:1198] (2/4) Epoch 28, batch 3300, loss[loss=0.2123, ctc_loss=0.1376, cr_loss=0.3733, over 17366.00 frames. ], tot_loss[loss=0.2021, ctc_loss=0.1322, cr_loss=0.3496, over 3356276.00 frames. ], batch size: 48, lr: 4.24e-03, grad_scale: 32.0 2024-09-24 12:58:08,051 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=5.18 vs. limit=12.0 2024-09-24 12:58:44,168 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=506394.0, ans=0.125 2024-09-24 12:58:48,846 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=506394.0, ans=0.125 2024-09-24 12:58:50,870 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=8.31 vs. limit=15.0 2024-09-24 12:59:04,541 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 12:59:13,738 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=506487.3333333333, ans=0.125 2024-09-24 12:59:21,507 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=506487.3333333333, ans=0.125 2024-09-24 12:59:24,424 INFO [train.py:1198] (2/4) Epoch 28, batch 3350, loss[loss=0.1717, ctc_loss=0.1108, cr_loss=0.3043, over 17125.00 frames. ], tot_loss[loss=0.2006, ctc_loss=0.1311, cr_loss=0.3474, over 3367481.53 frames. ], batch size: 40, lr: 4.24e-03, grad_scale: 32.0 2024-09-24 12:59:25,944 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.130e+02 1.246e+02 1.321e+02 1.387e+02 1.674e+02, threshold=2.642e+02, percent-clipped=0.0 2024-09-24 12:59:29,383 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=506534.0, ans=0.125 2024-09-24 12:59:29,397 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=506534.0, ans=0.125 2024-09-24 12:59:37,211 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=506534.0, ans=0.125 2024-09-24 12:59:38,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=506580.6666666667, ans=0.0 2024-09-24 12:59:56,357 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2.whitening_limit, batch_count=506627.3333333333, ans=15.0 2024-09-24 12:59:59,044 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=506627.3333333333, ans=0.125 2024-09-24 13:00:02,085 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=506627.3333333333, ans=0.015 2024-09-24 13:00:33,366 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.max_positive, batch_count=506720.6666666667, ans=0.95 2024-09-24 13:00:42,379 INFO [train.py:1198] (2/4) Epoch 28, batch 3400, loss[loss=0.2093, ctc_loss=0.1398, cr_loss=0.3475, over 17254.00 frames. ], tot_loss[loss=0.2007, ctc_loss=0.1313, cr_loss=0.3474, over 3357785.25 frames. ], batch size: 44, lr: 4.24e-03, grad_scale: 32.0 2024-09-24 13:00:47,338 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=506767.3333333333, ans=0.125 2024-09-24 13:01:03,609 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten.whitening_limit, batch_count=506814.0, ans=15.0 2024-09-24 13:01:12,572 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=506860.6666666667, ans=0.0 2024-09-24 13:01:18,892 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=506860.6666666667, ans=0.0 2024-09-24 13:01:19,311 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=6.10 vs. limit=12.0 2024-09-24 13:01:47,208 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.07 vs. limit=15.0 2024-09-24 13:01:53,128 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=506954.0, ans=0.125 2024-09-24 13:02:00,540 INFO [train.py:1198] (2/4) Epoch 28, batch 3450, loss[loss=0.1923, ctc_loss=0.1263, cr_loss=0.3303, over 17012.00 frames. ], tot_loss[loss=0.2015, ctc_loss=0.1317, cr_loss=0.3485, over 3356993.38 frames. ], batch size: 51, lr: 4.23e-03, grad_scale: 32.0 2024-09-24 13:02:02,018 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.326e+02 1.438e+02 1.586e+02 2.934e+02, threshold=2.877e+02, percent-clipped=1.0 2024-09-24 13:02:28,845 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=507047.3333333333, ans=0.0 2024-09-24 13:02:33,616 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=507094.0, ans=0.125 2024-09-24 13:02:43,335 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=507094.0, ans=0.125 2024-09-24 13:02:46,354 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=507140.6666666667, ans=0.0 2024-09-24 13:02:54,598 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=507140.6666666667, ans=0.125 2024-09-24 13:03:25,082 INFO [train.py:1198] (2/4) Epoch 28, batch 3500, loss[loss=0.2021, ctc_loss=0.1326, cr_loss=0.3476, over 17150.00 frames. ], tot_loss[loss=0.202, ctc_loss=0.1321, cr_loss=0.3496, over 3364975.73 frames. ], batch size: 48, lr: 4.23e-03, grad_scale: 32.0 2024-09-24 13:03:40,910 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=507280.6666666667, ans=0.1 2024-09-24 13:03:45,572 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=507280.6666666667, ans=0.1 2024-09-24 13:03:51,796 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 13:03:53,526 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=507280.6666666667, ans=0.125 2024-09-24 13:03:56,505 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=507327.3333333333, ans=0.125 2024-09-24 13:04:11,433 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=9.25 vs. limit=10.0 2024-09-24 13:04:16,710 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=507374.0, ans=0.2 2024-09-24 13:04:35,800 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=14.13 vs. limit=22.5 2024-09-24 13:04:41,701 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=507467.3333333333, ans=0.125 2024-09-24 13:04:42,941 INFO [train.py:1198] (2/4) Epoch 28, batch 3550, loss[loss=0.2125, ctc_loss=0.1385, cr_loss=0.3701, over 16732.00 frames. ], tot_loss[loss=0.203, ctc_loss=0.1329, cr_loss=0.3502, over 3357264.51 frames. ], batch size: 61, lr: 4.23e-03, grad_scale: 32.0 2024-09-24 13:04:44,467 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.084e+02 1.292e+02 1.411e+02 1.555e+02 1.879e+02, threshold=2.822e+02, percent-clipped=0.0 2024-09-24 13:05:10,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=507514.0, ans=0.1 2024-09-24 13:05:38,049 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=507607.3333333333, ans=0.125 2024-09-24 13:06:00,780 INFO [train.py:1198] (2/4) Epoch 28, batch 3600, loss[loss=0.1484, ctc_loss=0.09331, cr_loss=0.2757, over 17109.00 frames. ], tot_loss[loss=0.203, ctc_loss=0.1329, cr_loss=0.3504, over 3360500.80 frames. ], batch size: 40, lr: 4.23e-03, grad_scale: 32.0 2024-09-24 13:06:04,124 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=507700.6666666667, ans=0.0 2024-09-24 13:06:18,880 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=9.08 vs. limit=12.0 2024-09-24 13:06:20,150 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=507747.3333333333, ans=0.125 2024-09-24 13:06:31,154 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=507794.0, ans=0.05 2024-09-24 13:06:42,009 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=507794.0, ans=0.025 2024-09-24 13:06:45,255 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=507794.0, ans=0.125 2024-09-24 13:06:54,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=507840.6666666667, ans=0.125 2024-09-24 13:07:03,222 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=8.00 vs. limit=15.0 2024-09-24 13:07:11,659 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=507887.3333333333, ans=0.125 2024-09-24 13:07:19,235 INFO [train.py:1198] (2/4) Epoch 28, batch 3650, loss[loss=0.2134, ctc_loss=0.1404, cr_loss=0.3647, over 17135.00 frames. ], tot_loss[loss=0.2027, ctc_loss=0.1327, cr_loss=0.3502, over 3373027.23 frames. ], batch size: 48, lr: 4.23e-03, grad_scale: 32.0 2024-09-24 13:07:20,720 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.115e+02 1.240e+02 1.329e+02 1.473e+02 2.274e+02, threshold=2.658e+02, percent-clipped=0.0 2024-09-24 13:07:21,621 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=9.97 vs. limit=15.0 2024-09-24 13:07:30,458 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=507934.0, ans=0.2 2024-09-24 13:07:35,721 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=5.60 vs. limit=12.0 2024-09-24 13:08:01,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=508027.3333333333, ans=0.0 2024-09-24 13:08:40,141 INFO [train.py:1198] (2/4) Epoch 28, batch 3700, loss[loss=0.215, ctc_loss=0.1431, cr_loss=0.3597, over 16719.00 frames. ], tot_loss[loss=0.202, ctc_loss=0.1322, cr_loss=0.349, over 3367504.66 frames. ], batch size: 61, lr: 4.23e-03, grad_scale: 32.0 2024-09-24 13:08:43,561 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=508167.3333333333, ans=0.2 2024-09-24 13:08:55,235 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=21.45 vs. limit=22.5 2024-09-24 13:08:56,127 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=508214.0, ans=0.0 2024-09-24 13:08:59,461 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=508214.0, ans=0.2 2024-09-24 13:09:12,079 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=508260.6666666667, ans=0.05 2024-09-24 13:09:34,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=508307.3333333333, ans=0.0 2024-09-24 13:09:43,497 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.min_positive, batch_count=508354.0, ans=0.025 2024-09-24 13:09:46,822 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=2.687e-03 2024-09-24 13:09:58,944 INFO [train.py:1198] (2/4) Epoch 28, batch 3750, loss[loss=0.2482, ctc_loss=0.1688, cr_loss=0.3971, over 15003.00 frames. ], tot_loss[loss=0.2024, ctc_loss=0.1326, cr_loss=0.3491, over 3362265.73 frames. ], batch size: 89, lr: 4.23e-03, grad_scale: 32.0 2024-09-24 13:10:00,434 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.083e+02 1.248e+02 1.336e+02 1.422e+02 1.841e+02, threshold=2.673e+02, percent-clipped=0.0 2024-09-24 13:10:48,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys.whitening_limit, batch_count=508540.6666666667, ans=6.0 2024-09-24 13:10:49,708 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=508540.6666666667, ans=0.2 2024-09-24 13:11:05,471 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=508587.3333333333, ans=0.125 2024-09-24 13:11:15,133 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 13:11:17,814 INFO [train.py:1198] (2/4) Epoch 28, batch 3800, loss[loss=0.2372, ctc_loss=0.1592, cr_loss=0.3898, over 14882.00 frames. ], tot_loss[loss=0.2051, ctc_loss=0.1346, cr_loss=0.3528, over 3330961.53 frames. ], batch size: 89, lr: 4.23e-03, grad_scale: 16.0 2024-09-24 13:11:21,393 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=508634.0, ans=0.025 2024-09-24 13:11:27,500 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=508634.0, ans=0.125 2024-09-24 13:11:54,508 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=508727.3333333333, ans=0.125 2024-09-24 13:12:38,536 INFO [train.py:1198] (2/4) Epoch 28, batch 3850, loss[loss=0.2192, ctc_loss=0.1457, cr_loss=0.3675, over 17294.00 frames. ], tot_loss[loss=0.2073, ctc_loss=0.1363, cr_loss=0.3547, over 3283679.31 frames. ], batch size: 51, lr: 4.23e-03, grad_scale: 16.0 2024-09-24 13:12:42,159 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.266e+02 1.354e+02 1.491e+02 2.044e+02, threshold=2.707e+02, percent-clipped=0.0 2024-09-24 13:12:53,566 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=508914.0, ans=0.0 2024-09-24 13:12:59,929 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=508914.0, ans=0.0 2024-09-24 13:13:04,969 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.65 vs. limit=6.0 2024-09-24 13:13:26,118 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer_ff3.min_abs, batch_count=509007.3333333333, ans=0.2 2024-09-24 13:13:38,326 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=509007.3333333333, ans=0.2 2024-09-24 13:14:43,486 INFO [train.py:1198] (2/4) Epoch 29, batch 0, loss[loss=0.2195, ctc_loss=0.1438, cr_loss=0.3788, over 16721.00 frames. ], tot_loss[loss=0.2195, ctc_loss=0.1438, cr_loss=0.3788, over 16721.00 frames. ], batch size: 61, lr: 4.15e-03, grad_scale: 32.0 2024-09-24 13:14:43,486 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 13:14:56,596 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.4.encoder.layers.0.self_attn_weights, attn_weights_entropy = tensor([2.3491, 3.4925, 2.9447, 2.5847], device='cuda:2') 2024-09-24 13:14:58,953 INFO [train.py:1230] (2/4) Epoch 29, validation: loss=0.03615, ctc_loss=0.03615, cr_loss=9.405e-15, over 944034.00 frames. 2024-09-24 13:14:58,954 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 13:14:59,332 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=509082.0, ans=0.1 2024-09-24 13:16:21,069 INFO [train.py:1198] (2/4) Epoch 29, batch 50, loss[loss=0.2009, ctc_loss=0.1337, cr_loss=0.3358, over 17001.00 frames. ], tot_loss[loss=0.2006, ctc_loss=0.131, cr_loss=0.3479, over 766186.89 frames. ], batch size: 51, lr: 4.15e-03, grad_scale: 32.0 2024-09-24 13:16:21,963 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.31 vs. limit=10.0 2024-09-24 13:16:22,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.min_positive, batch_count=509315.3333333333, ans=0.025 2024-09-24 13:16:24,583 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=509315.3333333333, ans=0.125 2024-09-24 13:16:30,903 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.316e+02 1.466e+02 1.618e+02 2.901e+02, threshold=2.933e+02, percent-clipped=1.0 2024-09-24 13:16:38,251 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.64 vs. limit=22.5 2024-09-24 13:16:39,176 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=509362.0, ans=0.125 2024-09-24 13:16:50,756 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 13:16:55,432 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=509408.6666666667, ans=0.025 2024-09-24 13:16:58,609 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=509408.6666666667, ans=0.125 2024-09-24 13:17:09,822 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=509455.3333333333, ans=0.125 2024-09-24 13:17:44,484 INFO [train.py:1198] (2/4) Epoch 29, batch 100, loss[loss=0.1731, ctc_loss=0.1094, cr_loss=0.3188, over 17267.00 frames. ], tot_loss[loss=0.197, ctc_loss=0.1285, cr_loss=0.3426, over 1346116.97 frames. ], batch size: 42, lr: 4.15e-03, grad_scale: 32.0 2024-09-24 13:18:00,707 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=509595.3333333333, ans=0.125 2024-09-24 13:18:10,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=509595.3333333333, ans=0.125 2024-09-24 13:18:12,864 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=16.10 vs. limit=15.0 2024-09-24 13:18:28,629 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=509642.0, ans=0.2 2024-09-24 13:18:44,697 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 13:18:46,400 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=8.28 vs. limit=15.0 2024-09-24 13:19:09,195 INFO [train.py:1198] (2/4) Epoch 29, batch 150, loss[loss=0.1915, ctc_loss=0.1248, cr_loss=0.3336, over 17299.00 frames. ], tot_loss[loss=0.1987, ctc_loss=0.1297, cr_loss=0.3449, over 1794371.79 frames. ], batch size: 49, lr: 4.15e-03, grad_scale: 32.0 2024-09-24 13:19:14,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=509782.0, ans=0.125 2024-09-24 13:19:18,660 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.009e+02 1.258e+02 1.349e+02 1.479e+02 2.103e+02, threshold=2.697e+02, percent-clipped=0.0 2024-09-24 13:19:20,616 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=509782.0, ans=0.125 2024-09-24 13:19:45,056 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.42 vs. limit=15.0 2024-09-24 13:19:56,157 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.whiten, num_groups=1, num_channels=192, metric=4.70 vs. limit=12.0 2024-09-24 13:20:12,913 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 13:20:17,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=509968.6666666667, ans=0.0 2024-09-24 13:20:23,935 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=509968.6666666667, ans=0.0 2024-09-24 13:20:31,475 INFO [train.py:1198] (2/4) Epoch 29, batch 200, loss[loss=0.208, ctc_loss=0.1373, cr_loss=0.3533, over 17208.00 frames. ], tot_loss[loss=0.2012, ctc_loss=0.1316, cr_loss=0.3482, over 2124402.84 frames. ], batch size: 55, lr: 4.15e-03, grad_scale: 32.0 2024-09-24 13:20:49,598 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=510062.0, ans=0.0 2024-09-24 13:21:07,127 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.65 vs. limit=15.0 2024-09-24 13:21:09,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=510108.6666666667, ans=0.2 2024-09-24 13:21:14,631 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=510108.6666666667, ans=0.125 2024-09-24 13:21:51,164 INFO [train.py:1198] (2/4) Epoch 29, batch 250, loss[loss=0.2155, ctc_loss=0.1419, cr_loss=0.3679, over 16933.00 frames. ], tot_loss[loss=0.2005, ctc_loss=0.131, cr_loss=0.3477, over 2400574.63 frames. ], batch size: 58, lr: 4.15e-03, grad_scale: 32.0 2024-09-24 13:22:00,734 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.112e+02 1.248e+02 1.376e+02 1.477e+02 2.189e+02, threshold=2.751e+02, percent-clipped=0.0 2024-09-24 13:22:05,698 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=510295.3333333333, ans=0.025 2024-09-24 13:22:14,180 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=4.83 vs. limit=15.0 2024-09-24 13:22:18,629 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=7.39 vs. limit=12.0 2024-09-24 13:22:24,871 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=510342.0, ans=0.1 2024-09-24 13:22:56,281 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=510435.3333333333, ans=0.0 2024-09-24 13:23:07,493 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=510435.3333333333, ans=0.125 2024-09-24 13:23:16,525 INFO [train.py:1198] (2/4) Epoch 29, batch 300, loss[loss=0.207, ctc_loss=0.1378, cr_loss=0.3461, over 16907.00 frames. ], tot_loss[loss=0.2022, ctc_loss=0.1322, cr_loss=0.3502, over 2615011.65 frames. ], batch size: 58, lr: 4.14e-03, grad_scale: 16.0 2024-09-24 13:23:17,295 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=10.20 vs. limit=22.5 2024-09-24 13:23:20,022 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=510482.0, ans=0.125 2024-09-24 13:23:24,887 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=510482.0, ans=0.09899494936611666 2024-09-24 13:23:49,062 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=510575.3333333333, ans=0.0 2024-09-24 13:24:07,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=510622.0, ans=0.125 2024-09-24 13:24:09,375 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=510622.0, ans=0.125 2024-09-24 13:24:17,363 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=510622.0, ans=0.1 2024-09-24 13:24:18,851 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=510622.0, ans=0.0 2024-09-24 13:24:20,518 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=510622.0, ans=0.0 2024-09-24 13:24:39,066 INFO [train.py:1198] (2/4) Epoch 29, batch 350, loss[loss=0.2084, ctc_loss=0.136, cr_loss=0.3616, over 17014.00 frames. ], tot_loss[loss=0.2023, ctc_loss=0.1322, cr_loss=0.3508, over 2783147.03 frames. ], batch size: 51, lr: 4.14e-03, grad_scale: 16.0 2024-09-24 13:24:50,234 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.082e+02 1.267e+02 1.359e+02 1.490e+02 2.133e+02, threshold=2.717e+02, percent-clipped=0.0 2024-09-24 13:24:50,502 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=510715.3333333333, ans=0.035 2024-09-24 13:25:15,826 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=510808.6666666667, ans=0.1 2024-09-24 13:25:44,724 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=3.97 vs. limit=15.0 2024-09-24 13:26:01,404 INFO [train.py:1198] (2/4) Epoch 29, batch 400, loss[loss=0.1776, ctc_loss=0.113, cr_loss=0.323, over 17162.00 frames. ], tot_loss[loss=0.2017, ctc_loss=0.1316, cr_loss=0.3505, over 2918538.42 frames. ], batch size: 41, lr: 4.14e-03, grad_scale: 32.0 2024-09-24 13:26:07,810 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=510948.6666666667, ans=0.125 2024-09-24 13:26:18,167 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.81 vs. limit=12.0 2024-09-24 13:27:09,485 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.93 vs. limit=15.0 2024-09-24 13:27:21,139 INFO [train.py:1198] (2/4) Epoch 29, batch 450, loss[loss=0.2008, ctc_loss=0.1311, cr_loss=0.3484, over 17304.00 frames. ], tot_loss[loss=0.201, ctc_loss=0.1311, cr_loss=0.3498, over 3021699.42 frames. ], batch size: 46, lr: 4.14e-03, grad_scale: 32.0 2024-09-24 13:27:35,217 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.066e+02 1.269e+02 1.351e+02 1.464e+02 1.902e+02, threshold=2.702e+02, percent-clipped=0.0 2024-09-24 13:27:49,553 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=511228.6666666667, ans=0.125 2024-09-24 13:27:49,831 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=511228.6666666667, ans=0.125 2024-09-24 13:27:52,932 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=511228.6666666667, ans=0.1 2024-09-24 13:27:56,085 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=511275.3333333333, ans=0.0 2024-09-24 13:28:15,195 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=511322.0, ans=0.125 2024-09-24 13:28:46,768 INFO [train.py:1198] (2/4) Epoch 29, batch 500, loss[loss=0.2204, ctc_loss=0.144, cr_loss=0.3821, over 17229.00 frames. ], tot_loss[loss=0.2014, ctc_loss=0.1314, cr_loss=0.3503, over 3102737.70 frames. ], batch size: 55, lr: 4.14e-03, grad_scale: 32.0 2024-09-24 13:28:51,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=511415.3333333333, ans=0.0 2024-09-24 13:29:12,218 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=511462.0, ans=0.2 2024-09-24 13:29:32,860 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 13:29:56,762 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=511602.0, ans=0.2 2024-09-24 13:29:58,498 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=511602.0, ans=0.09899494936611666 2024-09-24 13:30:09,199 INFO [train.py:1198] (2/4) Epoch 29, batch 550, loss[loss=0.1981, ctc_loss=0.1295, cr_loss=0.3432, over 17308.00 frames. ], tot_loss[loss=0.2012, ctc_loss=0.1314, cr_loss=0.3493, over 3155104.08 frames. ], batch size: 51, lr: 4.14e-03, grad_scale: 32.0 2024-09-24 13:30:20,287 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.249e+02 1.312e+02 1.438e+02 1.848e+02, threshold=2.623e+02, percent-clipped=0.0 2024-09-24 13:30:56,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=511742.0, ans=0.0 2024-09-24 13:31:04,840 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.90 vs. limit=6.0 2024-09-24 13:31:24,774 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=511835.3333333333, ans=0.09899494936611666 2024-09-24 13:31:30,726 INFO [train.py:1198] (2/4) Epoch 29, batch 600, loss[loss=0.2616, ctc_loss=0.1905, cr_loss=0.3555, over 11413.00 frames. ], tot_loss[loss=0.2011, ctc_loss=0.1313, cr_loss=0.349, over 3204164.29 frames. ], batch size: 123, lr: 4.14e-03, grad_scale: 32.0 2024-09-24 13:31:43,806 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=511882.0, ans=0.0 2024-09-24 13:31:45,625 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=13.15 vs. limit=15.0 2024-09-24 13:32:04,306 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=511975.3333333333, ans=0.0 2024-09-24 13:32:32,836 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=512022.0, ans=0.125 2024-09-24 13:32:47,202 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=512068.6666666667, ans=0.125 2024-09-24 13:32:53,244 INFO [train.py:1198] (2/4) Epoch 29, batch 650, loss[loss=0.1989, ctc_loss=0.1276, cr_loss=0.3565, over 17080.00 frames. ], tot_loss[loss=0.2007, ctc_loss=0.1311, cr_loss=0.3481, over 3239910.27 frames. ], batch size: 43, lr: 4.14e-03, grad_scale: 32.0 2024-09-24 13:33:04,449 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.092e+02 1.266e+02 1.354e+02 1.448e+02 2.374e+02, threshold=2.709e+02, percent-clipped=0.0 2024-09-24 13:33:04,681 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=512115.3333333333, ans=0.0 2024-09-24 13:33:08,507 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=5.92 vs. limit=15.0 2024-09-24 13:33:15,689 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=512162.0, ans=0.125 2024-09-24 13:33:17,891 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=10.77 vs. limit=15.0 2024-09-24 13:33:35,160 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 13:33:41,696 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=512208.6666666667, ans=0.2 2024-09-24 13:33:43,682 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=11.54 vs. limit=15.0 2024-09-24 13:33:52,697 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=512255.3333333333, ans=0.125 2024-09-24 13:34:19,042 INFO [train.py:1198] (2/4) Epoch 29, batch 700, loss[loss=0.1919, ctc_loss=0.123, cr_loss=0.3444, over 17303.00 frames. ], tot_loss[loss=0.2008, ctc_loss=0.1312, cr_loss=0.3477, over 3254718.61 frames. ], batch size: 46, lr: 4.14e-03, grad_scale: 32.0 2024-09-24 13:34:24,091 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=512348.6666666667, ans=0.025 2024-09-24 13:34:29,435 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.92 vs. limit=15.0 2024-09-24 13:34:32,700 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=4.75 vs. limit=15.0 2024-09-24 13:35:01,285 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.55 vs. limit=15.0 2024-09-24 13:35:02,441 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=512442.0, ans=0.2 2024-09-24 13:35:41,600 INFO [train.py:1198] (2/4) Epoch 29, batch 750, loss[loss=0.2629, ctc_loss=0.187, cr_loss=0.3792, over 11737.00 frames. ], tot_loss[loss=0.201, ctc_loss=0.1314, cr_loss=0.3479, over 3274355.19 frames. ], batch size: 123, lr: 4.14e-03, grad_scale: 32.0 2024-09-24 13:35:52,785 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.055e+02 1.251e+02 1.333e+02 1.428e+02 1.733e+02, threshold=2.666e+02, percent-clipped=0.0 2024-09-24 13:36:45,880 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.03 vs. limit=22.5 2024-09-24 13:36:53,599 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=512768.6666666667, ans=0.125 2024-09-24 13:36:58,948 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.74 vs. limit=15.0 2024-09-24 13:37:01,403 INFO [train.py:1198] (2/4) Epoch 29, batch 800, loss[loss=0.2193, ctc_loss=0.1458, cr_loss=0.3674, over 17108.00 frames. ], tot_loss[loss=0.2012, ctc_loss=0.1315, cr_loss=0.3482, over 3281981.11 frames. ], batch size: 49, lr: 4.14e-03, grad_scale: 32.0 2024-09-24 13:37:03,278 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=512815.3333333333, ans=0.1 2024-09-24 13:37:11,432 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=512815.3333333333, ans=0.125 2024-09-24 13:37:22,752 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=512862.0, ans=0.0 2024-09-24 13:37:44,867 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=512908.6666666667, ans=0.2 2024-09-24 13:37:51,206 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=512955.3333333333, ans=0.125 2024-09-24 13:38:14,792 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=513002.0, ans=0.0 2024-09-24 13:38:25,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=513048.6666666667, ans=0.125 2024-09-24 13:38:27,133 INFO [train.py:1198] (2/4) Epoch 29, batch 850, loss[loss=0.1746, ctc_loss=0.1151, cr_loss=0.2972, over 17200.00 frames. ], tot_loss[loss=0.201, ctc_loss=0.1315, cr_loss=0.3479, over 3295299.42 frames. ], batch size: 47, lr: 4.13e-03, grad_scale: 32.0 2024-09-24 13:38:29,051 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=513048.6666666667, ans=0.0 2024-09-24 13:38:34,242 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.55 vs. limit=22.5 2024-09-24 13:38:38,336 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.286e+02 1.364e+02 1.497e+02 3.898e+02, threshold=2.729e+02, percent-clipped=1.0 2024-09-24 13:38:45,676 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=2.78 vs. limit=6.0 2024-09-24 13:38:54,550 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=513095.3333333333, ans=0.1 2024-09-24 13:39:38,526 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=513235.3333333333, ans=0.0 2024-09-24 13:39:44,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=513235.3333333333, ans=0.1 2024-09-24 13:39:47,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=513282.0, ans=0.0 2024-09-24 13:39:49,250 INFO [train.py:1198] (2/4) Epoch 29, batch 900, loss[loss=0.1723, ctc_loss=0.1098, cr_loss=0.3129, over 17029.00 frames. ], tot_loss[loss=0.2001, ctc_loss=0.1307, cr_loss=0.347, over 3309687.20 frames. ], batch size: 44, lr: 4.13e-03, grad_scale: 32.0 2024-09-24 13:39:57,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=513282.0, ans=0.0 2024-09-24 13:39:58,574 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=9.49 vs. limit=10.0 2024-09-24 13:40:04,258 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=9.53 vs. limit=15.0 2024-09-24 13:40:29,168 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=513375.3333333333, ans=0.125 2024-09-24 13:40:30,783 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=513375.3333333333, ans=0.025 2024-09-24 13:40:32,646 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 13:40:37,485 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=513375.3333333333, ans=0.2 2024-09-24 13:40:51,912 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 13:41:06,494 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=513468.6666666667, ans=0.025 2024-09-24 13:41:12,250 INFO [train.py:1198] (2/4) Epoch 29, batch 950, loss[loss=0.2086, ctc_loss=0.1373, cr_loss=0.3565, over 17088.00 frames. ], tot_loss[loss=0.2011, ctc_loss=0.1314, cr_loss=0.3486, over 3322319.20 frames. ], batch size: 46, lr: 4.13e-03, grad_scale: 32.0 2024-09-24 13:41:23,505 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.274e+02 1.388e+02 1.487e+02 2.628e+02, threshold=2.776e+02, percent-clipped=0.0 2024-09-24 13:41:24,228 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.62 vs. limit=22.5 2024-09-24 13:41:27,329 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.14 vs. limit=10.0 2024-09-24 13:41:36,814 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=513562.0, ans=0.0 2024-09-24 13:41:56,644 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.39 vs. limit=12.0 2024-09-24 13:42:10,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=513655.3333333333, ans=0.125 2024-09-24 13:42:33,820 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=513748.6666666667, ans=0.125 2024-09-24 13:42:35,064 INFO [train.py:1198] (2/4) Epoch 29, batch 1000, loss[loss=0.1645, ctc_loss=0.1053, cr_loss=0.2956, over 17088.00 frames. ], tot_loss[loss=0.2001, ctc_loss=0.1306, cr_loss=0.3475, over 3338551.10 frames. ], batch size: 40, lr: 4.13e-03, grad_scale: 16.0 2024-09-24 13:43:00,608 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=513795.3333333333, ans=0.07 2024-09-24 13:43:45,297 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=20.52 vs. limit=22.5 2024-09-24 13:43:59,529 INFO [train.py:1198] (2/4) Epoch 29, batch 1050, loss[loss=0.2013, ctc_loss=0.1343, cr_loss=0.335, over 17228.00 frames. ], tot_loss[loss=0.2011, ctc_loss=0.1314, cr_loss=0.3484, over 3339864.03 frames. ], batch size: 50, lr: 4.13e-03, grad_scale: 16.0 2024-09-24 13:44:12,020 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.285e+02 1.349e+02 1.436e+02 3.120e+02, threshold=2.698e+02, percent-clipped=1.0 2024-09-24 13:44:57,264 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=514122.0, ans=0.0 2024-09-24 13:45:02,063 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=514168.6666666667, ans=0.0 2024-09-24 13:45:21,711 INFO [train.py:1198] (2/4) Epoch 29, batch 1100, loss[loss=0.2183, ctc_loss=0.1427, cr_loss=0.3781, over 16984.00 frames. ], tot_loss[loss=0.2026, ctc_loss=0.1324, cr_loss=0.351, over 3342508.04 frames. ], batch size: 53, lr: 4.13e-03, grad_scale: 16.0 2024-09-24 13:45:31,758 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.04 vs. limit=15.0 2024-09-24 13:45:35,587 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=5.05 vs. limit=10.0 2024-09-24 13:45:36,289 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=514262.0, ans=0.025 2024-09-24 13:45:38,585 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=4.86 vs. limit=15.0 2024-09-24 13:45:40,318 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=14.35 vs. limit=22.5 2024-09-24 13:45:41,536 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=514262.0, ans=0.125 2024-09-24 13:46:10,491 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=514355.3333333333, ans=0.125 2024-09-24 13:46:10,518 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=514355.3333333333, ans=0.0 2024-09-24 13:46:13,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=514355.3333333333, ans=0.0 2024-09-24 13:46:42,098 INFO [train.py:1198] (2/4) Epoch 29, batch 1150, loss[loss=0.2066, ctc_loss=0.1299, cr_loss=0.3836, over 17068.00 frames. ], tot_loss[loss=0.2011, ctc_loss=0.1313, cr_loss=0.3491, over 3356208.48 frames. ], batch size: 46, lr: 4.13e-03, grad_scale: 16.0 2024-09-24 13:46:42,358 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=514448.6666666667, ans=0.125 2024-09-24 13:46:50,990 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=13.63 vs. limit=15.0 2024-09-24 13:46:54,882 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.257e+02 1.322e+02 1.401e+02 2.069e+02, threshold=2.644e+02, percent-clipped=0.0 2024-09-24 13:47:38,645 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=514588.6666666667, ans=0.0 2024-09-24 13:47:40,491 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=514588.6666666667, ans=0.0 2024-09-24 13:47:46,745 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=514635.3333333333, ans=0.125 2024-09-24 13:47:47,109 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.28 vs. limit=6.0 2024-09-24 13:48:04,033 INFO [train.py:1198] (2/4) Epoch 29, batch 1200, loss[loss=0.1878, ctc_loss=0.1218, cr_loss=0.3299, over 16940.00 frames. ], tot_loss[loss=0.1997, ctc_loss=0.1303, cr_loss=0.3473, over 3363170.57 frames. ], batch size: 42, lr: 4.13e-03, grad_scale: 16.0 2024-09-24 13:48:19,484 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=514682.0, ans=0.125 2024-09-24 13:48:22,739 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=514728.6666666667, ans=0.2 2024-09-24 13:48:38,606 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=514775.3333333333, ans=0.0 2024-09-24 13:49:00,891 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=3.71 vs. limit=15.0 2024-09-24 13:49:04,987 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=514822.0, ans=0.125 2024-09-24 13:49:25,574 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=514868.6666666667, ans=0.04949747468305833 2024-09-24 13:49:28,541 INFO [train.py:1198] (2/4) Epoch 29, batch 1250, loss[loss=0.1907, ctc_loss=0.1226, cr_loss=0.3405, over 17020.00 frames. ], tot_loss[loss=0.2004, ctc_loss=0.1307, cr_loss=0.3485, over 3358270.52 frames. ], batch size: 44, lr: 4.13e-03, grad_scale: 16.0 2024-09-24 13:49:42,754 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.062e+02 1.263e+02 1.323e+02 1.423e+02 2.408e+02, threshold=2.646e+02, percent-clipped=0.0 2024-09-24 13:49:44,684 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=514962.0, ans=0.0 2024-09-24 13:49:54,729 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.46 vs. limit=15.0 2024-09-24 13:49:56,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=514962.0, ans=0.0 2024-09-24 13:50:41,599 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=515102.0, ans=0.1 2024-09-24 13:50:43,065 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=515102.0, ans=0.2 2024-09-24 13:50:50,682 INFO [train.py:1198] (2/4) Epoch 29, batch 1300, loss[loss=0.2165, ctc_loss=0.1436, cr_loss=0.3644, over 16889.00 frames. ], tot_loss[loss=0.2, ctc_loss=0.1305, cr_loss=0.3475, over 3366265.67 frames. ], batch size: 58, lr: 4.13e-03, grad_scale: 16.0 2024-09-24 13:50:52,990 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=5.30 vs. limit=12.0 2024-09-24 13:50:57,477 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=515148.6666666667, ans=0.125 2024-09-24 13:50:59,250 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.93 vs. limit=15.0 2024-09-24 13:51:07,083 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=515195.3333333333, ans=0.2 2024-09-24 13:51:07,426 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.17 vs. limit=15.0 2024-09-24 13:51:09,113 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=515195.3333333333, ans=0.05 2024-09-24 13:51:33,341 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=9.21 vs. limit=15.0 2024-09-24 13:52:10,986 INFO [train.py:1198] (2/4) Epoch 29, batch 1350, loss[loss=0.2148, ctc_loss=0.1419, cr_loss=0.3646, over 17050.00 frames. ], tot_loss[loss=0.2007, ctc_loss=0.131, cr_loss=0.3485, over 3358699.47 frames. ], batch size: 52, lr: 4.13e-03, grad_scale: 16.0 2024-09-24 13:52:25,431 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.241e+02 1.323e+02 1.445e+02 2.039e+02, threshold=2.645e+02, percent-clipped=0.0 2024-09-24 13:52:51,889 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=515475.3333333333, ans=0.1 2024-09-24 13:53:00,170 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=515522.0, ans=0.0 2024-09-24 13:53:22,081 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=12.92 vs. limit=22.5 2024-09-24 13:53:35,817 INFO [train.py:1198] (2/4) Epoch 29, batch 1400, loss[loss=0.1683, ctc_loss=0.1081, cr_loss=0.3015, over 17162.00 frames. ], tot_loss[loss=0.1999, ctc_loss=0.1304, cr_loss=0.3477, over 3365845.82 frames. ], batch size: 45, lr: 4.12e-03, grad_scale: 16.0 2024-09-24 13:53:39,165 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=515615.3333333333, ans=0.125 2024-09-24 13:53:53,474 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=515662.0, ans=0.125 2024-09-24 13:54:15,094 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=515708.6666666667, ans=0.0 2024-09-24 13:54:22,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=515708.6666666667, ans=0.1 2024-09-24 13:54:26,210 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=515755.3333333333, ans=0.2 2024-09-24 13:54:40,412 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=515802.0, ans=0.125 2024-09-24 13:54:57,872 INFO [train.py:1198] (2/4) Epoch 29, batch 1450, loss[loss=0.2253, ctc_loss=0.1479, cr_loss=0.387, over 17232.00 frames. ], tot_loss[loss=0.2003, ctc_loss=0.1306, cr_loss=0.3483, over 3366830.22 frames. ], batch size: 55, lr: 4.12e-03, grad_scale: 8.0 2024-09-24 13:54:58,189 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=515848.6666666667, ans=0.1 2024-09-24 13:55:16,336 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.107e+02 1.288e+02 1.365e+02 1.470e+02 2.158e+02, threshold=2.730e+02, percent-clipped=0.0 2024-09-24 13:55:18,048 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=515895.3333333333, ans=0.015 2024-09-24 13:55:26,807 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.92 vs. limit=15.0 2024-09-24 13:55:27,866 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=515895.3333333333, ans=0.0 2024-09-24 13:55:31,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=515942.0, ans=0.125 2024-09-24 13:55:32,827 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=515942.0, ans=0.125 2024-09-24 13:55:39,018 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 13:56:19,905 INFO [train.py:1198] (2/4) Epoch 29, batch 1500, loss[loss=0.1559, ctc_loss=0.09486, cr_loss=0.3052, over 16240.00 frames. ], tot_loss[loss=0.201, ctc_loss=0.1313, cr_loss=0.3486, over 3360163.42 frames. ], batch size: 36, lr: 4.12e-03, grad_scale: 8.0 2024-09-24 13:56:23,622 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=12.03 vs. limit=12.0 2024-09-24 13:56:33,990 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module1.whiten, num_groups=1, num_channels=192, metric=7.81 vs. limit=15.0 2024-09-24 13:56:41,408 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=516128.6666666667, ans=0.0 2024-09-24 13:56:45,084 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=14.59 vs. limit=22.5 2024-09-24 13:57:01,155 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=9.77 vs. limit=15.0 2024-09-24 13:57:42,715 INFO [train.py:1198] (2/4) Epoch 29, batch 1550, loss[loss=0.1935, ctc_loss=0.1263, cr_loss=0.336, over 17099.00 frames. ], tot_loss[loss=0.2009, ctc_loss=0.1313, cr_loss=0.3481, over 3356384.59 frames. ], batch size: 43, lr: 4.12e-03, grad_scale: 8.0 2024-09-24 13:57:48,038 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=516315.3333333333, ans=0.1 2024-09-24 13:57:54,338 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=516315.3333333333, ans=0.09899494936611666 2024-09-24 13:57:58,708 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.272e+02 1.340e+02 1.430e+02 4.940e+02, threshold=2.681e+02, percent-clipped=1.0 2024-09-24 13:58:22,110 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=516408.6666666667, ans=0.125 2024-09-24 13:58:26,897 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=516408.6666666667, ans=0.125 2024-09-24 13:59:07,816 INFO [train.py:1198] (2/4) Epoch 29, batch 1600, loss[loss=0.1897, ctc_loss=0.1215, cr_loss=0.341, over 17227.00 frames. ], tot_loss[loss=0.2008, ctc_loss=0.1311, cr_loss=0.3485, over 3356153.21 frames. ], batch size: 50, lr: 4.12e-03, grad_scale: 16.0 2024-09-24 13:59:32,145 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=516595.3333333333, ans=0.125 2024-09-24 13:59:44,709 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=516642.0, ans=0.2 2024-09-24 13:59:52,572 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=516642.0, ans=0.0 2024-09-24 14:00:16,278 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=6.62 vs. limit=15.0 2024-09-24 14:00:30,025 INFO [train.py:1198] (2/4) Epoch 29, batch 1650, loss[loss=0.2247, ctc_loss=0.1492, cr_loss=0.3772, over 17217.00 frames. ], tot_loss[loss=0.2005, ctc_loss=0.131, cr_loss=0.3476, over 3355013.20 frames. ], batch size: 55, lr: 4.12e-03, grad_scale: 16.0 2024-09-24 14:00:30,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=516782.0, ans=0.125 2024-09-24 14:00:41,396 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=516782.0, ans=0.035 2024-09-24 14:00:45,987 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.269e+02 1.357e+02 1.497e+02 2.178e+02, threshold=2.713e+02, percent-clipped=0.0 2024-09-24 14:00:54,900 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.59 vs. limit=10.0 2024-09-24 14:00:57,376 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=516828.6666666667, ans=0.0 2024-09-24 14:01:38,860 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=516968.6666666667, ans=0.0 2024-09-24 14:01:42,023 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=516968.6666666667, ans=0.125 2024-09-24 14:01:42,116 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=516968.6666666667, ans=0.0 2024-09-24 14:01:49,852 INFO [train.py:1198] (2/4) Epoch 29, batch 1700, loss[loss=0.16, ctc_loss=0.104, cr_loss=0.2802, over 15841.00 frames. ], tot_loss[loss=0.2018, ctc_loss=0.132, cr_loss=0.3492, over 3344753.19 frames. ], batch size: 35, lr: 4.12e-03, grad_scale: 16.0 2024-09-24 14:01:57,328 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.70 vs. limit=15.0 2024-09-24 14:02:02,930 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=517015.3333333333, ans=0.125 2024-09-24 14:02:04,505 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=517062.0, ans=0.0 2024-09-24 14:02:06,850 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=12.09 vs. limit=22.5 2024-09-24 14:02:09,478 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=517062.0, ans=0.0 2024-09-24 14:02:17,279 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=517062.0, ans=0.1 2024-09-24 14:02:20,638 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=8.87 vs. limit=15.0 2024-09-24 14:02:25,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=517108.6666666667, ans=0.125 2024-09-24 14:02:38,729 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=517155.3333333333, ans=0.125 2024-09-24 14:02:40,250 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer_na.min_abs, batch_count=517155.3333333333, ans=0.02 2024-09-24 14:02:48,910 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=13.73 vs. limit=22.5 2024-09-24 14:03:08,118 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=517202.0, ans=0.0 2024-09-24 14:03:13,479 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1.whitening_limit, batch_count=517248.6666666667, ans=10.0 2024-09-24 14:03:14,198 INFO [train.py:1198] (2/4) Epoch 29, batch 1750, loss[loss=0.2052, ctc_loss=0.1315, cr_loss=0.3688, over 17263.00 frames. ], tot_loss[loss=0.2013, ctc_loss=0.1316, cr_loss=0.3482, over 3333117.72 frames. ], batch size: 44, lr: 4.12e-03, grad_scale: 16.0 2024-09-24 14:03:28,873 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=517295.3333333333, ans=0.07 2024-09-24 14:03:30,182 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.095e+02 1.263e+02 1.353e+02 1.492e+02 2.426e+02, threshold=2.707e+02, percent-clipped=0.0 2024-09-24 14:03:30,567 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=517295.3333333333, ans=0.0 2024-09-24 14:04:00,189 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=517342.0, ans=0.125 2024-09-24 14:04:03,330 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=517388.6666666667, ans=0.0 2024-09-24 14:04:03,412 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=517388.6666666667, ans=0.125 2024-09-24 14:04:08,109 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=517388.6666666667, ans=0.0 2024-09-24 14:04:09,858 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=517388.6666666667, ans=0.125 2024-09-24 14:04:15,264 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=14.22 vs. limit=15.0 2024-09-24 14:04:25,563 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=517435.3333333333, ans=0.0 2024-09-24 14:04:36,175 INFO [train.py:1198] (2/4) Epoch 29, batch 1800, loss[loss=0.1848, ctc_loss=0.1191, cr_loss=0.3285, over 17168.00 frames. ], tot_loss[loss=0.2012, ctc_loss=0.1317, cr_loss=0.3478, over 3340709.34 frames. ], batch size: 41, lr: 4.12e-03, grad_scale: 16.0 2024-09-24 14:05:17,670 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=517575.3333333333, ans=0.1 2024-09-24 14:05:36,823 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=517622.0, ans=0.0 2024-09-24 14:05:38,364 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=517622.0, ans=0.2 2024-09-24 14:05:42,497 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=11.06 vs. limit=22.5 2024-09-24 14:05:58,951 INFO [train.py:1198] (2/4) Epoch 29, batch 1850, loss[loss=0.2265, ctc_loss=0.1457, cr_loss=0.4042, over 16957.00 frames. ], tot_loss[loss=0.2013, ctc_loss=0.1317, cr_loss=0.3482, over 3344628.92 frames. ], batch size: 58, lr: 4.12e-03, grad_scale: 8.0 2024-09-24 14:06:16,451 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.270e+02 1.382e+02 1.482e+02 2.420e+02, threshold=2.764e+02, percent-clipped=0.0 2024-09-24 14:06:18,731 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=10.29 vs. limit=15.0 2024-09-24 14:06:26,806 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.59 vs. limit=15.0 2024-09-24 14:06:29,659 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=517808.6666666667, ans=0.07 2024-09-24 14:06:52,130 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=517855.3333333333, ans=0.125 2024-09-24 14:06:56,880 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=517855.3333333333, ans=0.0 2024-09-24 14:07:06,728 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=13.74 vs. limit=22.5 2024-09-24 14:07:15,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=517902.0, ans=0.0 2024-09-24 14:07:21,463 INFO [train.py:1198] (2/4) Epoch 29, batch 1900, loss[loss=0.1918, ctc_loss=0.1265, cr_loss=0.3264, over 17073.00 frames. ], tot_loss[loss=0.2034, ctc_loss=0.1332, cr_loss=0.351, over 3347680.06 frames. ], batch size: 46, lr: 4.12e-03, grad_scale: 8.0 2024-09-24 14:07:47,794 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=517995.3333333333, ans=0.125 2024-09-24 14:08:02,079 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.62 vs. limit=15.0 2024-09-24 14:08:16,962 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=518088.6666666667, ans=0.1 2024-09-24 14:08:18,592 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=518088.6666666667, ans=0.0 2024-09-24 14:08:43,830 INFO [train.py:1198] (2/4) Epoch 29, batch 1950, loss[loss=0.2349, ctc_loss=0.1593, cr_loss=0.3777, over 12204.00 frames. ], tot_loss[loss=0.2027, ctc_loss=0.1328, cr_loss=0.3498, over 3348118.24 frames. ], batch size: 123, lr: 4.11e-03, grad_scale: 8.0 2024-09-24 14:09:03,910 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.107e+02 1.275e+02 1.365e+02 1.439e+02 2.080e+02, threshold=2.731e+02, percent-clipped=0.0 2024-09-24 14:09:34,629 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=518322.0, ans=0.0 2024-09-24 14:09:41,230 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=518322.0, ans=0.0 2024-09-24 14:09:46,036 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=518322.0, ans=0.2 2024-09-24 14:09:53,508 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=518368.6666666667, ans=0.125 2024-09-24 14:10:08,969 INFO [train.py:1198] (2/4) Epoch 29, batch 2000, loss[loss=0.1899, ctc_loss=0.1252, cr_loss=0.3234, over 17163.00 frames. ], tot_loss[loss=0.2029, ctc_loss=0.1328, cr_loss=0.3502, over 3358724.85 frames. ], batch size: 45, lr: 4.11e-03, grad_scale: 16.0 2024-09-24 14:10:36,707 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=518462.0, ans=0.1 2024-09-24 14:10:57,468 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=13.52 vs. limit=15.0 2024-09-24 14:11:18,032 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=518602.0, ans=0.125 2024-09-24 14:11:28,975 INFO [train.py:1198] (2/4) Epoch 29, batch 2050, loss[loss=0.1774, ctc_loss=0.1155, cr_loss=0.3093, over 17065.00 frames. ], tot_loss[loss=0.2021, ctc_loss=0.1323, cr_loss=0.3495, over 3355098.20 frames. ], batch size: 46, lr: 4.11e-03, grad_scale: 16.0 2024-09-24 14:11:40,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=518648.6666666667, ans=0.0 2024-09-24 14:11:46,520 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.263e+02 1.339e+02 1.463e+02 3.835e+02, threshold=2.678e+02, percent-clipped=1.0 2024-09-24 14:11:58,820 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=5.97 vs. limit=15.0 2024-09-24 14:12:26,446 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.35 vs. limit=15.0 2024-09-24 14:12:32,943 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.20 vs. limit=22.5 2024-09-24 14:12:43,780 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 14:12:54,001 INFO [train.py:1198] (2/4) Epoch 29, batch 2100, loss[loss=0.2593, ctc_loss=0.1823, cr_loss=0.3855, over 11804.00 frames. ], tot_loss[loss=0.2014, ctc_loss=0.1317, cr_loss=0.3483, over 3354268.93 frames. ], batch size: 123, lr: 4.11e-03, grad_scale: 16.0 2024-09-24 14:13:00,636 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=518882.0, ans=0.125 2024-09-24 14:13:27,610 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=518975.3333333333, ans=0.0 2024-09-24 14:13:37,023 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=518975.3333333333, ans=0.125 2024-09-24 14:13:54,239 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=11.27 vs. limit=15.0 2024-09-24 14:13:54,698 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=3.68 vs. limit=15.0 2024-09-24 14:14:15,827 INFO [train.py:1198] (2/4) Epoch 29, batch 2150, loss[loss=0.1996, ctc_loss=0.1279, cr_loss=0.3584, over 17297.00 frames. ], tot_loss[loss=0.2016, ctc_loss=0.1319, cr_loss=0.3486, over 3362037.14 frames. ], batch size: 49, lr: 4.11e-03, grad_scale: 16.0 2024-09-24 14:14:33,679 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.060e+02 1.281e+02 1.376e+02 1.508e+02 1.841e+02, threshold=2.752e+02, percent-clipped=0.0 2024-09-24 14:14:56,219 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=9.18 vs. limit=15.0 2024-09-24 14:15:02,053 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=519208.6666666667, ans=0.125 2024-09-24 14:15:10,346 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=519255.3333333333, ans=0.125 2024-09-24 14:15:22,449 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.54 vs. limit=15.0 2024-09-24 14:15:27,910 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=519302.0, ans=0.125 2024-09-24 14:15:38,897 INFO [train.py:1198] (2/4) Epoch 29, batch 2200, loss[loss=0.2026, ctc_loss=0.1334, cr_loss=0.3458, over 17144.00 frames. ], tot_loss[loss=0.2021, ctc_loss=0.1322, cr_loss=0.3494, over 3363214.26 frames. ], batch size: 48, lr: 4.11e-03, grad_scale: 16.0 2024-09-24 14:15:50,622 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=519348.6666666667, ans=0.125 2024-09-24 14:16:20,210 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.63 vs. limit=6.0 2024-09-24 14:16:59,444 INFO [train.py:1198] (2/4) Epoch 29, batch 2250, loss[loss=0.2165, ctc_loss=0.1362, cr_loss=0.4017, over 17099.00 frames. ], tot_loss[loss=0.202, ctc_loss=0.132, cr_loss=0.3502, over 3364670.09 frames. ], batch size: 49, lr: 4.11e-03, grad_scale: 16.0 2024-09-24 14:17:19,626 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.297e+02 1.378e+02 1.443e+02 1.753e+02, threshold=2.756e+02, percent-clipped=0.0 2024-09-24 14:18:23,697 INFO [train.py:1198] (2/4) Epoch 29, batch 2300, loss[loss=0.2334, ctc_loss=0.1602, cr_loss=0.3658, over 11789.00 frames. ], tot_loss[loss=0.2026, ctc_loss=0.1325, cr_loss=0.3506, over 3350709.82 frames. ], batch size: 123, lr: 4.11e-03, grad_scale: 16.0 2024-09-24 14:18:43,395 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=519862.0, ans=0.125 2024-09-24 14:18:48,811 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=7.44 vs. limit=15.0 2024-09-24 14:19:11,823 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=9.74 vs. limit=15.0 2024-09-24 14:19:18,050 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=519955.3333333333, ans=0.125 2024-09-24 14:19:24,262 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=519955.3333333333, ans=0.1 2024-09-24 14:19:27,496 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=519955.3333333333, ans=0.125 2024-09-24 14:19:40,073 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=520002.0, ans=0.1 2024-09-24 14:19:46,204 INFO [train.py:1198] (2/4) Epoch 29, batch 2350, loss[loss=0.1784, ctc_loss=0.1163, cr_loss=0.3105, over 17019.00 frames. ], tot_loss[loss=0.201, ctc_loss=0.1314, cr_loss=0.3483, over 3341399.75 frames. ], batch size: 39, lr: 4.11e-03, grad_scale: 16.0 2024-09-24 14:19:53,633 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=520048.6666666667, ans=0.125 2024-09-24 14:20:06,253 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.131e+02 1.272e+02 1.348e+02 1.492e+02 2.219e+02, threshold=2.696e+02, percent-clipped=0.0 2024-09-24 14:20:12,881 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=520095.3333333333, ans=0.125 2024-09-24 14:20:26,191 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.05 vs. limit=6.0 2024-09-24 14:20:32,070 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=520142.0, ans=0.125 2024-09-24 14:21:04,494 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=520235.3333333333, ans=0.125 2024-09-24 14:21:07,584 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=520282.0, ans=0.0 2024-09-24 14:21:08,865 INFO [train.py:1198] (2/4) Epoch 29, batch 2400, loss[loss=0.2257, ctc_loss=0.1489, cr_loss=0.3841, over 16910.00 frames. ], tot_loss[loss=0.2017, ctc_loss=0.132, cr_loss=0.3488, over 3337524.62 frames. ], batch size: 58, lr: 4.11e-03, grad_scale: 32.0 2024-09-24 14:22:28,621 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=520468.6666666667, ans=0.2 2024-09-24 14:22:31,467 INFO [train.py:1198] (2/4) Epoch 29, batch 2450, loss[loss=0.1903, ctc_loss=0.1249, cr_loss=0.327, over 17266.00 frames. ], tot_loss[loss=0.2029, ctc_loss=0.1328, cr_loss=0.3502, over 3330703.26 frames. ], batch size: 44, lr: 4.10e-03, grad_scale: 16.0 2024-09-24 14:22:53,218 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.297e+02 1.387e+02 1.500e+02 2.305e+02, threshold=2.773e+02, percent-clipped=0.0 2024-09-24 14:22:59,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=520562.0, ans=0.1 2024-09-24 14:23:06,352 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=520608.6666666667, ans=0.2 2024-09-24 14:23:31,854 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=520655.3333333333, ans=0.125 2024-09-24 14:23:35,136 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=520655.3333333333, ans=0.1 2024-09-24 14:23:56,403 INFO [train.py:1198] (2/4) Epoch 29, batch 2500, loss[loss=0.2196, ctc_loss=0.1487, cr_loss=0.3546, over 15976.00 frames. ], tot_loss[loss=0.2035, ctc_loss=0.1332, cr_loss=0.3513, over 3342904.15 frames. ], batch size: 74, lr: 4.10e-03, grad_scale: 16.0 2024-09-24 14:23:57,291 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.39 vs. limit=15.0 2024-09-24 14:24:25,659 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=520795.3333333333, ans=0.05 2024-09-24 14:24:29,496 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=13.99 vs. limit=15.0 2024-09-24 14:24:31,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=520842.0, ans=10.0 2024-09-24 14:24:35,169 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=520842.0, ans=0.025 2024-09-24 14:24:43,741 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.07 vs. limit=10.0 2024-09-24 14:25:00,610 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=10.56 vs. limit=15.0 2024-09-24 14:25:08,020 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=520935.3333333333, ans=0.025 2024-09-24 14:25:11,255 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=520935.3333333333, ans=0.0 2024-09-24 14:25:18,997 INFO [train.py:1198] (2/4) Epoch 29, batch 2550, loss[loss=0.2031, ctc_loss=0.1316, cr_loss=0.3574, over 17086.00 frames. ], tot_loss[loss=0.2026, ctc_loss=0.1325, cr_loss=0.3503, over 3348724.23 frames. ], batch size: 49, lr: 4.10e-03, grad_scale: 16.0 2024-09-24 14:25:24,419 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten.whitening_limit, batch_count=520982.0, ans=22.5 2024-09-24 14:25:33,761 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=521028.6666666667, ans=0.1 2024-09-24 14:25:38,171 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.116e+02 1.280e+02 1.353e+02 1.438e+02 1.912e+02, threshold=2.707e+02, percent-clipped=0.0 2024-09-24 14:25:38,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=521028.6666666667, ans=0.125 2024-09-24 14:25:51,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=521075.3333333333, ans=0.1 2024-09-24 14:26:21,192 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=521168.6666666667, ans=0.125 2024-09-24 14:26:37,036 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=521215.3333333333, ans=0.0 2024-09-24 14:26:37,067 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=521215.3333333333, ans=0.125 2024-09-24 14:26:38,263 INFO [train.py:1198] (2/4) Epoch 29, batch 2600, loss[loss=0.2389, ctc_loss=0.1591, cr_loss=0.3989, over 16214.00 frames. ], tot_loss[loss=0.2025, ctc_loss=0.1325, cr_loss=0.35, over 3352894.94 frames. ], batch size: 74, lr: 4.10e-03, grad_scale: 16.0 2024-09-24 14:27:28,357 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.28 vs. limit=15.0 2024-09-24 14:27:29,106 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=521355.3333333333, ans=0.0 2024-09-24 14:27:36,887 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=521355.3333333333, ans=0.0 2024-09-24 14:27:42,953 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=521355.3333333333, ans=0.0 2024-09-24 14:27:49,471 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer_na.min_abs, batch_count=521402.0, ans=0.02 2024-09-24 14:28:03,407 INFO [train.py:1198] (2/4) Epoch 29, batch 2650, loss[loss=0.2068, ctc_loss=0.1346, cr_loss=0.3613, over 17285.00 frames. ], tot_loss[loss=0.2033, ctc_loss=0.133, cr_loss=0.3517, over 3348999.85 frames. ], batch size: 46, lr: 4.10e-03, grad_scale: 16.0 2024-09-24 14:28:15,308 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.57 vs. limit=6.0 2024-09-24 14:28:22,577 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.102e+02 1.267e+02 1.389e+02 1.480e+02 2.068e+02, threshold=2.779e+02, percent-clipped=0.0 2024-09-24 14:28:28,401 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn2.whiten, num_groups=1, num_channels=192, metric=7.89 vs. limit=22.5 2024-09-24 14:28:32,636 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=521495.3333333333, ans=0.0 2024-09-24 14:29:14,986 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=521635.3333333333, ans=0.125 2024-09-24 14:29:25,644 INFO [train.py:1198] (2/4) Epoch 29, batch 2700, loss[loss=0.1896, ctc_loss=0.123, cr_loss=0.333, over 17023.00 frames. ], tot_loss[loss=0.2028, ctc_loss=0.1325, cr_loss=0.3512, over 3346063.57 frames. ], batch size: 44, lr: 4.10e-03, grad_scale: 16.0 2024-09-24 14:30:01,233 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=521775.3333333333, ans=0.125 2024-09-24 14:30:10,893 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=521775.3333333333, ans=0.125 2024-09-24 14:30:15,795 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=521822.0, ans=0.0 2024-09-24 14:30:20,455 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=521822.0, ans=0.125 2024-09-24 14:30:24,991 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=521822.0, ans=0.0 2024-09-24 14:30:33,281 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=521868.6666666667, ans=0.125 2024-09-24 14:30:49,227 INFO [train.py:1198] (2/4) Epoch 29, batch 2750, loss[loss=0.1867, ctc_loss=0.1212, cr_loss=0.3274, over 17227.00 frames. ], tot_loss[loss=0.2024, ctc_loss=0.1323, cr_loss=0.3507, over 3357803.41 frames. ], batch size: 50, lr: 4.10e-03, grad_scale: 16.0 2024-09-24 14:30:59,649 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=5.47 vs. limit=10.0 2024-09-24 14:31:08,254 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.228e+02 1.306e+02 1.411e+02 3.014e+02, threshold=2.612e+02, percent-clipped=1.0 2024-09-24 14:31:10,317 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=1.019e-02 2024-09-24 14:31:29,226 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=522008.6666666667, ans=0.2 2024-09-24 14:31:43,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=522055.3333333333, ans=0.0 2024-09-24 14:31:45,293 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=522055.3333333333, ans=0.04949747468305833 2024-09-24 14:31:57,388 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=522102.0, ans=0.1 2024-09-24 14:32:11,736 INFO [train.py:1198] (2/4) Epoch 29, batch 2800, loss[loss=0.1878, ctc_loss=0.122, cr_loss=0.3288, over 17074.00 frames. ], tot_loss[loss=0.2005, ctc_loss=0.1309, cr_loss=0.3483, over 3363038.87 frames. ], batch size: 46, lr: 4.10e-03, grad_scale: 32.0 2024-09-24 14:32:15,172 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=522148.6666666667, ans=0.0 2024-09-24 14:32:18,938 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.80 vs. limit=15.0 2024-09-24 14:32:37,039 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=522195.3333333333, ans=0.2 2024-09-24 14:32:38,641 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=522195.3333333333, ans=0.1 2024-09-24 14:33:11,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=522288.6666666667, ans=0.0 2024-09-24 14:33:34,159 INFO [train.py:1198] (2/4) Epoch 29, batch 2850, loss[loss=0.2064, ctc_loss=0.1338, cr_loss=0.3629, over 16909.00 frames. ], tot_loss[loss=0.2006, ctc_loss=0.1309, cr_loss=0.3485, over 3362456.34 frames. ], batch size: 58, lr: 4.10e-03, grad_scale: 32.0 2024-09-24 14:33:46,981 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=522382.0, ans=0.125 2024-09-24 14:33:57,876 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.283e+02 1.399e+02 1.566e+02 1.810e+02, threshold=2.799e+02, percent-clipped=0.0 2024-09-24 14:33:59,743 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=522428.6666666667, ans=0.125 2024-09-24 14:34:20,825 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=522475.3333333333, ans=0.2 2024-09-24 14:35:00,062 INFO [train.py:1198] (2/4) Epoch 29, batch 2900, loss[loss=0.1746, ctc_loss=0.1141, cr_loss=0.3026, over 17019.00 frames. ], tot_loss[loss=0.2005, ctc_loss=0.1309, cr_loss=0.3481, over 3366816.11 frames. ], batch size: 44, lr: 4.10e-03, grad_scale: 16.0 2024-09-24 14:35:45,152 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=20.32 vs. limit=22.5 2024-09-24 14:36:03,904 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=522755.3333333333, ans=0.0 2024-09-24 14:36:13,403 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=522802.0, ans=0.1 2024-09-24 14:36:22,782 INFO [train.py:1198] (2/4) Epoch 29, batch 2950, loss[loss=0.1906, ctc_loss=0.1217, cr_loss=0.3443, over 17029.00 frames. ], tot_loss[loss=0.1998, ctc_loss=0.1304, cr_loss=0.3474, over 3368361.30 frames. ], batch size: 44, lr: 4.10e-03, grad_scale: 16.0 2024-09-24 14:36:31,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=522848.6666666667, ans=0.125 2024-09-24 14:36:34,376 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=522848.6666666667, ans=0.125 2024-09-24 14:36:43,418 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.270e+02 1.367e+02 1.488e+02 1.985e+02, threshold=2.734e+02, percent-clipped=0.0 2024-09-24 14:36:50,029 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=522895.3333333333, ans=0.07 2024-09-24 14:37:18,403 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=522988.6666666667, ans=0.0 2024-09-24 14:37:44,729 INFO [train.py:1198] (2/4) Epoch 29, batch 3000, loss[loss=0.1681, ctc_loss=0.1079, cr_loss=0.3009, over 17179.00 frames. ], tot_loss[loss=0.2003, ctc_loss=0.1307, cr_loss=0.348, over 3364053.66 frames. ], batch size: 41, lr: 4.09e-03, grad_scale: 16.0 2024-09-24 14:37:44,729 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 14:38:00,320 INFO [train.py:1230] (2/4) Epoch 29, validation: loss=0.03658, ctc_loss=0.03658, cr_loss=8.731e-15, over 944034.00 frames. 2024-09-24 14:38:00,321 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 14:38:02,103 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=523082.0, ans=0.125 2024-09-24 14:38:16,422 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=523128.6666666667, ans=0.125 2024-09-24 14:38:24,456 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=523128.6666666667, ans=0.125 2024-09-24 14:38:35,308 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=523175.3333333333, ans=0.0 2024-09-24 14:38:46,347 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=523222.0, ans=0.125 2024-09-24 14:38:49,639 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=523222.0, ans=0.0 2024-09-24 14:39:19,059 INFO [train.py:1198] (2/4) Epoch 29, batch 3050, loss[loss=0.1921, ctc_loss=0.1277, cr_loss=0.3222, over 17274.00 frames. ], tot_loss[loss=0.2008, ctc_loss=0.1311, cr_loss=0.3486, over 3370669.00 frames. ], batch size: 44, lr: 4.09e-03, grad_scale: 16.0 2024-09-24 14:39:21,515 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.26 vs. limit=12.0 2024-09-24 14:39:26,661 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=523315.3333333333, ans=0.125 2024-09-24 14:39:31,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=523315.3333333333, ans=0.04949747468305833 2024-09-24 14:39:42,288 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.064e+02 1.250e+02 1.367e+02 1.506e+02 1.984e+02, threshold=2.734e+02, percent-clipped=0.0 2024-09-24 14:39:48,850 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=523362.0, ans=0.125 2024-09-24 14:39:50,382 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=523362.0, ans=0.0 2024-09-24 14:39:58,344 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=523408.6666666667, ans=0.2 2024-09-24 14:40:01,445 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=523408.6666666667, ans=0.2 2024-09-24 14:40:04,481 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=523408.6666666667, ans=0.1 2024-09-24 14:40:10,852 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=523455.3333333333, ans=0.5 2024-09-24 14:40:12,584 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=523455.3333333333, ans=0.0 2024-09-24 14:40:25,072 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=523502.0, ans=0.125 2024-09-24 14:40:40,602 INFO [train.py:1198] (2/4) Epoch 29, batch 3100, loss[loss=0.2323, ctc_loss=0.1535, cr_loss=0.3937, over 16537.00 frames. ], tot_loss[loss=0.2005, ctc_loss=0.1309, cr_loss=0.348, over 3368579.71 frames. ], batch size: 66, lr: 4.09e-03, grad_scale: 16.0 2024-09-24 14:40:40,853 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=523548.6666666667, ans=0.125 2024-09-24 14:40:54,908 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=523595.3333333333, ans=0.1 2024-09-24 14:41:02,217 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.53 vs. limit=6.0 2024-09-24 14:41:06,255 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=523595.3333333333, ans=0.1 2024-09-24 14:41:18,966 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=523642.0, ans=0.125 2024-09-24 14:41:28,076 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=523688.6666666667, ans=0.025 2024-09-24 14:42:00,681 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.79 vs. limit=15.0 2024-09-24 14:42:01,583 INFO [train.py:1198] (2/4) Epoch 29, batch 3150, loss[loss=0.1741, ctc_loss=0.1109, cr_loss=0.3164, over 17097.00 frames. ], tot_loss[loss=0.2001, ctc_loss=0.1307, cr_loss=0.3474, over 3371202.23 frames. ], batch size: 43, lr: 4.09e-03, grad_scale: 16.0 2024-09-24 14:42:21,628 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.246e+02 1.357e+02 1.527e+02 1.959e+02, threshold=2.714e+02, percent-clipped=0.0 2024-09-24 14:42:43,458 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=523875.3333333333, ans=0.1 2024-09-24 14:42:43,469 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.min_positive, batch_count=523875.3333333333, ans=0.025 2024-09-24 14:42:48,254 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=523922.0, ans=0.0 2024-09-24 14:43:18,236 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 14:43:19,437 INFO [train.py:1198] (2/4) Epoch 29, batch 3200, loss[loss=0.204, ctc_loss=0.1317, cr_loss=0.3615, over 17086.00 frames. ], tot_loss[loss=0.1998, ctc_loss=0.1304, cr_loss=0.3467, over 3369107.65 frames. ], batch size: 49, lr: 4.09e-03, grad_scale: 32.0 2024-09-24 14:43:21,339 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=524015.3333333333, ans=0.0 2024-09-24 14:43:45,979 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=524062.0, ans=0.125 2024-09-24 14:44:17,284 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=524155.3333333333, ans=0.2 2024-09-24 14:44:28,286 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=524202.0, ans=0.125 2024-09-24 14:44:31,249 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=524202.0, ans=0.125 2024-09-24 14:44:34,615 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=11.81 vs. limit=15.0 2024-09-24 14:44:37,176 INFO [train.py:1198] (2/4) Epoch 29, batch 3250, loss[loss=0.1649, ctc_loss=0.1056, cr_loss=0.2964, over 16976.00 frames. ], tot_loss[loss=0.2003, ctc_loss=0.1308, cr_loss=0.3474, over 3365296.89 frames. ], batch size: 42, lr: 4.09e-03, grad_scale: 32.0 2024-09-24 14:44:57,321 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.092e+02 1.269e+02 1.350e+02 1.434e+02 1.655e+02, threshold=2.700e+02, percent-clipped=0.0 2024-09-24 14:45:00,913 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=524295.3333333334, ans=0.125 2024-09-24 14:45:39,878 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=524435.3333333334, ans=0.0 2024-09-24 14:45:55,497 INFO [train.py:1198] (2/4) Epoch 29, batch 3300, loss[loss=0.2223, ctc_loss=0.1453, cr_loss=0.3852, over 17141.00 frames. ], tot_loss[loss=0.2006, ctc_loss=0.1311, cr_loss=0.3477, over 3359271.11 frames. ], batch size: 48, lr: 4.09e-03, grad_scale: 32.0 2024-09-24 14:45:55,754 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=524482.0, ans=0.1 2024-09-24 14:46:01,957 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=524482.0, ans=0.125 2024-09-24 14:46:06,738 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=524482.0, ans=0.04949747468305833 2024-09-24 14:46:15,108 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=524528.6666666666, ans=0.1 2024-09-24 14:46:18,262 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=524528.6666666666, ans=0.125 2024-09-24 14:46:32,395 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=524575.3333333334, ans=0.0 2024-09-24 14:46:33,880 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=524575.3333333334, ans=0.1 2024-09-24 14:46:34,607 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.57 vs. limit=22.5 2024-09-24 14:46:37,284 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn1.whiten.whitening_limit, batch_count=524575.3333333334, ans=22.5 2024-09-24 14:46:46,366 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=524622.0, ans=0.1 2024-09-24 14:46:57,041 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.max_abs, batch_count=524622.0, ans=10.0 2024-09-24 14:46:59,076 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten.whitening_limit, batch_count=524668.6666666666, ans=22.5 2024-09-24 14:47:15,322 INFO [train.py:1198] (2/4) Epoch 29, batch 3350, loss[loss=0.171, ctc_loss=0.1104, cr_loss=0.3032, over 17091.00 frames. ], tot_loss[loss=0.2009, ctc_loss=0.1313, cr_loss=0.3481, over 3353609.36 frames. ], batch size: 43, lr: 4.09e-03, grad_scale: 32.0 2024-09-24 14:47:34,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=524762.0, ans=0.2 2024-09-24 14:47:35,614 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.319e+02 1.369e+02 1.450e+02 1.942e+02, threshold=2.738e+02, percent-clipped=0.0 2024-09-24 14:48:01,336 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=524808.6666666666, ans=0.125 2024-09-24 14:48:09,760 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.31 vs. limit=12.0 2024-09-24 14:48:14,014 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=524855.3333333334, ans=0.125 2024-09-24 14:48:17,114 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=524855.3333333334, ans=0.125 2024-09-24 14:48:20,265 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=524902.0, ans=0.125 2024-09-24 14:48:35,632 INFO [train.py:1198] (2/4) Epoch 29, batch 3400, loss[loss=0.1747, ctc_loss=0.1113, cr_loss=0.317, over 16725.00 frames. ], tot_loss[loss=0.2018, ctc_loss=0.1319, cr_loss=0.3492, over 3349082.05 frames. ], batch size: 37, lr: 4.09e-03, grad_scale: 32.0 2024-09-24 14:48:50,069 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=524995.3333333334, ans=0.0 2024-09-24 14:49:00,875 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=524995.3333333334, ans=0.2 2024-09-24 14:49:07,796 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.69 vs. limit=15.0 2024-09-24 14:49:39,576 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=525135.3333333334, ans=0.0 2024-09-24 14:49:55,465 INFO [train.py:1198] (2/4) Epoch 29, batch 3450, loss[loss=0.1728, ctc_loss=0.1123, cr_loss=0.3025, over 17091.00 frames. ], tot_loss[loss=0.2024, ctc_loss=0.1324, cr_loss=0.35, over 3356599.93 frames. ], batch size: 40, lr: 4.09e-03, grad_scale: 32.0 2024-09-24 14:49:57,289 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=525182.0, ans=0.1 2024-09-24 14:50:15,930 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.258e+02 1.341e+02 1.460e+02 2.344e+02, threshold=2.681e+02, percent-clipped=0.0 2024-09-24 14:50:16,167 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=525228.6666666666, ans=0.125 2024-09-24 14:50:17,793 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=525228.6666666666, ans=0.1 2024-09-24 14:50:28,657 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=525275.3333333334, ans=0.0 2024-09-24 14:50:39,702 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=525275.3333333334, ans=0.125 2024-09-24 14:50:48,111 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=512, metric=2.95 vs. limit=15.0 2024-09-24 14:51:09,792 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=525368.6666666666, ans=0.0 2024-09-24 14:51:14,074 INFO [train.py:1198] (2/4) Epoch 29, batch 3500, loss[loss=0.2599, ctc_loss=0.1747, cr_loss=0.4261, over 14974.00 frames. ], tot_loss[loss=0.2018, ctc_loss=0.132, cr_loss=0.349, over 3339221.38 frames. ], batch size: 89, lr: 4.09e-03, grad_scale: 32.0 2024-09-24 14:52:18,617 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=525602.0, ans=0.1 2024-09-24 14:52:31,425 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=525602.0, ans=0.05 2024-09-24 14:52:34,350 INFO [train.py:1198] (2/4) Epoch 29, batch 3550, loss[loss=0.2036, ctc_loss=0.132, cr_loss=0.358, over 17158.00 frames. ], tot_loss[loss=0.2012, ctc_loss=0.1314, cr_loss=0.3487, over 3342270.94 frames. ], batch size: 45, lr: 4.08e-03, grad_scale: 32.0 2024-09-24 14:52:54,489 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.066e+02 1.276e+02 1.353e+02 1.468e+02 1.866e+02, threshold=2.706e+02, percent-clipped=0.0 2024-09-24 14:53:32,107 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=525788.6666666666, ans=0.125 2024-09-24 14:53:41,615 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=525835.3333333334, ans=0.1 2024-09-24 14:53:46,511 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.02 vs. limit=6.0 2024-09-24 14:53:52,065 INFO [train.py:1198] (2/4) Epoch 29, batch 3600, loss[loss=0.1835, ctc_loss=0.1161, cr_loss=0.3366, over 17254.00 frames. ], tot_loss[loss=0.2017, ctc_loss=0.1318, cr_loss=0.3495, over 3338827.97 frames. ], batch size: 44, lr: 4.08e-03, grad_scale: 32.0 2024-09-24 14:54:00,461 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.71 vs. limit=15.0 2024-09-24 14:54:34,369 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=525975.3333333334, ans=0.125 2024-09-24 14:54:42,292 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=526022.0, ans=0.0 2024-09-24 14:55:07,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=526068.6666666666, ans=0.5 2024-09-24 14:55:10,106 INFO [train.py:1198] (2/4) Epoch 29, batch 3650, loss[loss=0.164, ctc_loss=0.103, cr_loss=0.3053, over 16946.00 frames. ], tot_loss[loss=0.2017, ctc_loss=0.1318, cr_loss=0.3496, over 3346458.62 frames. ], batch size: 42, lr: 4.08e-03, grad_scale: 32.0 2024-09-24 14:55:31,931 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.135e+02 1.251e+02 1.330e+02 1.491e+02 2.044e+02, threshold=2.661e+02, percent-clipped=0.0 2024-09-24 14:55:37,242 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 14:55:37,276 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 14:55:44,930 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=526208.6666666666, ans=0.1 2024-09-24 14:55:54,926 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=526208.6666666666, ans=0.125 2024-09-24 14:56:17,443 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=526302.0, ans=0.125 2024-09-24 14:56:31,268 INFO [train.py:1198] (2/4) Epoch 29, batch 3700, loss[loss=0.1971, ctc_loss=0.1288, cr_loss=0.3417, over 17142.00 frames. ], tot_loss[loss=0.2021, ctc_loss=0.1321, cr_loss=0.35, over 3342652.10 frames. ], batch size: 48, lr: 4.08e-03, grad_scale: 16.0 2024-09-24 14:56:46,945 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=526395.3333333334, ans=0.125 2024-09-24 14:57:51,171 INFO [train.py:1198] (2/4) Epoch 29, batch 3750, loss[loss=0.2428, ctc_loss=0.165, cr_loss=0.3887, over 14878.00 frames. ], tot_loss[loss=0.2033, ctc_loss=0.1331, cr_loss=0.3514, over 3343246.60 frames. ], batch size: 90, lr: 4.08e-03, grad_scale: 16.0 2024-09-24 14:57:52,935 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=526582.0, ans=0.0 2024-09-24 14:57:54,896 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.72 vs. limit=10.0 2024-09-24 14:58:13,097 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.112e+02 1.272e+02 1.350e+02 1.451e+02 2.081e+02, threshold=2.700e+02, percent-clipped=0.0 2024-09-24 14:58:18,027 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=526628.6666666666, ans=0.0 2024-09-24 14:58:18,183 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=526628.6666666666, ans=0.0 2024-09-24 14:58:18,787 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=6.16 vs. limit=15.0 2024-09-24 14:58:29,047 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=526675.3333333334, ans=0.025 2024-09-24 14:58:46,122 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=526722.0, ans=0.0 2024-09-24 14:59:05,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=526768.6666666666, ans=0.0 2024-09-24 14:59:10,334 INFO [train.py:1198] (2/4) Epoch 29, batch 3800, loss[loss=0.2265, ctc_loss=0.1587, cr_loss=0.3388, over 14770.00 frames. ], tot_loss[loss=0.2035, ctc_loss=0.1333, cr_loss=0.351, over 3319021.16 frames. ], batch size: 89, lr: 4.08e-03, grad_scale: 16.0 2024-09-24 14:59:16,968 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=526815.3333333334, ans=0.125 2024-09-24 14:59:43,677 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=526908.6666666666, ans=0.125 2024-09-24 14:59:45,595 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=6.77 vs. limit=15.0 2024-09-24 15:00:03,648 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=526955.3333333334, ans=0.125 2024-09-24 15:00:27,867 INFO [train.py:1198] (2/4) Epoch 29, batch 3850, loss[loss=0.2395, ctc_loss=0.1607, cr_loss=0.3938, over 15114.00 frames. ], tot_loss[loss=0.2071, ctc_loss=0.1361, cr_loss=0.3551, over 3263008.02 frames. ], batch size: 89, lr: 4.08e-03, grad_scale: 16.0 2024-09-24 15:00:49,020 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.095e+02 1.344e+02 1.460e+02 1.620e+02 2.302e+02, threshold=2.919e+02, percent-clipped=0.0 2024-09-24 15:00:55,025 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=10.73 vs. limit=15.0 2024-09-24 15:00:57,591 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=527142.0, ans=0.0 2024-09-24 15:01:03,822 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=527142.0, ans=0.0 2024-09-24 15:01:07,176 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=11.19 vs. limit=12.0 2024-09-24 15:01:14,763 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=14.52 vs. limit=22.5 2024-09-24 15:02:21,332 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=527263.3333333334, ans=0.0 2024-09-24 15:02:28,429 INFO [train.py:1198] (2/4) Epoch 30, batch 0, loss[loss=0.2142, ctc_loss=0.1403, cr_loss=0.3695, over 17302.00 frames. ], tot_loss[loss=0.2142, ctc_loss=0.1403, cr_loss=0.3695, over 17302.00 frames. ], batch size: 51, lr: 4.01e-03, grad_scale: 32.0 2024-09-24 15:02:28,429 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 15:02:37,739 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.2.encoder.layers.0.self_attn_weights, attn_weights_entropy = tensor([4.3766, 4.0471, 4.0771, 3.9023], device='cuda:2') 2024-09-24 15:02:43,738 INFO [train.py:1230] (2/4) Epoch 30, validation: loss=0.0352, ctc_loss=0.0352, cr_loss=9.262e-15, over 944034.00 frames. 2024-09-24 15:02:43,739 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 15:03:06,391 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=527310.0, ans=0.125 2024-09-24 15:03:28,247 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.49 vs. limit=15.0 2024-09-24 15:03:32,504 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=527356.6666666666, ans=0.2 2024-09-24 15:04:04,172 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=527450.0, ans=0.125 2024-09-24 15:04:07,010 INFO [train.py:1198] (2/4) Epoch 30, batch 50, loss[loss=0.2258, ctc_loss=0.1497, cr_loss=0.3806, over 17096.00 frames. ], tot_loss[loss=0.2024, ctc_loss=0.1321, cr_loss=0.3514, over 756878.12 frames. ], batch size: 49, lr: 4.01e-03, grad_scale: 32.0 2024-09-24 15:04:32,899 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=527543.3333333334, ans=0.2 2024-09-24 15:04:35,823 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.064e+02 1.261e+02 1.395e+02 1.583e+02 2.602e+02, threshold=2.789e+02, percent-clipped=0.0 2024-09-24 15:04:37,733 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=527590.0, ans=0.125 2024-09-24 15:04:39,497 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-24 15:04:52,434 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.86 vs. limit=10.0 2024-09-24 15:05:19,407 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=527683.3333333334, ans=0.125 2024-09-24 15:05:30,395 INFO [train.py:1198] (2/4) Epoch 30, batch 100, loss[loss=0.1971, ctc_loss=0.1303, cr_loss=0.334, over 17232.00 frames. ], tot_loss[loss=0.2011, ctc_loss=0.1312, cr_loss=0.3496, over 1339949.40 frames. ], batch size: 55, lr: 4.01e-03, grad_scale: 32.0 2024-09-24 15:06:18,291 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=527870.0, ans=0.125 2024-09-24 15:06:20,441 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=11.69 vs. limit=22.5 2024-09-24 15:06:28,854 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.51 vs. limit=5.0 2024-09-24 15:06:55,899 INFO [train.py:1198] (2/4) Epoch 30, batch 150, loss[loss=0.2172, ctc_loss=0.1468, cr_loss=0.3518, over 11264.00 frames. ], tot_loss[loss=0.2005, ctc_loss=0.1308, cr_loss=0.3486, over 1791373.36 frames. ], batch size: 123, lr: 4.01e-03, grad_scale: 32.0 2024-09-24 15:06:57,847 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=527963.3333333334, ans=0.0 2024-09-24 15:07:09,423 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=512, metric=6.37 vs. limit=15.0 2024-09-24 15:07:13,759 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=528010.0, ans=0.0 2024-09-24 15:07:16,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=528010.0, ans=10.0 2024-09-24 15:07:22,149 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.06 vs. limit=6.0 2024-09-24 15:07:24,702 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.139e+02 1.297e+02 1.402e+02 1.523e+02 2.631e+02, threshold=2.805e+02, percent-clipped=0.0 2024-09-24 15:07:34,636 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=528056.6666666666, ans=0.125 2024-09-24 15:07:37,239 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=7.51 vs. limit=8.0 2024-09-24 15:08:19,311 INFO [train.py:1198] (2/4) Epoch 30, batch 200, loss[loss=0.1647, ctc_loss=0.104, cr_loss=0.3032, over 17078.00 frames. ], tot_loss[loss=0.1998, ctc_loss=0.1303, cr_loss=0.3476, over 2142908.62 frames. ], batch size: 39, lr: 4.01e-03, grad_scale: 32.0 2024-09-24 15:08:22,958 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=528196.6666666666, ans=0.95 2024-09-24 15:08:26,090 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=528196.6666666666, ans=0.035 2024-09-24 15:08:27,730 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=528196.6666666666, ans=0.125 2024-09-24 15:08:37,673 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.06 vs. limit=6.0 2024-09-24 15:08:58,611 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=528290.0, ans=0.2 2024-09-24 15:09:14,406 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=528336.6666666666, ans=0.125 2024-09-24 15:09:19,183 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.max_positive, batch_count=528336.6666666666, ans=0.95 2024-09-24 15:09:25,616 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.min_abs, batch_count=528383.3333333334, ans=0.5 2024-09-24 15:09:42,473 INFO [train.py:1198] (2/4) Epoch 30, batch 250, loss[loss=0.2224, ctc_loss=0.1441, cr_loss=0.3913, over 16782.00 frames. ], tot_loss[loss=0.2005, ctc_loss=0.1308, cr_loss=0.3483, over 2417352.54 frames. ], batch size: 61, lr: 4.00e-03, grad_scale: 32.0 2024-09-24 15:09:49,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=528430.0, ans=0.0 2024-09-24 15:09:52,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=528430.0, ans=0.1 2024-09-24 15:10:02,867 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=13.57 vs. limit=22.5 2024-09-24 15:10:11,297 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.080e+02 1.266e+02 1.350e+02 1.451e+02 1.821e+02, threshold=2.699e+02, percent-clipped=0.0 2024-09-24 15:10:26,283 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=528523.3333333334, ans=0.125 2024-09-24 15:10:30,922 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=528570.0, ans=0.0 2024-09-24 15:10:35,818 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=528570.0, ans=0.125 2024-09-24 15:10:59,930 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=528616.6666666666, ans=0.125 2024-09-24 15:11:02,796 INFO [train.py:1198] (2/4) Epoch 30, batch 300, loss[loss=0.2321, ctc_loss=0.1534, cr_loss=0.3931, over 17346.00 frames. ], tot_loss[loss=0.2004, ctc_loss=0.1308, cr_loss=0.348, over 2619615.48 frames. ], batch size: 48, lr: 4.00e-03, grad_scale: 32.0 2024-09-24 15:11:08,543 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=9.82 vs. limit=15.0 2024-09-24 15:11:19,283 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.min_positive, batch_count=528710.0, ans=0.025 2024-09-24 15:11:30,547 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=528710.0, ans=0.125 2024-09-24 15:11:36,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=528756.6666666666, ans=0.1 2024-09-24 15:12:14,342 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=528850.0, ans=0.125 2024-09-24 15:12:14,911 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=9.97 vs. limit=15.0 2024-09-24 15:12:28,505 INFO [train.py:1198] (2/4) Epoch 30, batch 350, loss[loss=0.2103, ctc_loss=0.1364, cr_loss=0.3692, over 17026.00 frames. ], tot_loss[loss=0.1994, ctc_loss=0.1301, cr_loss=0.3466, over 2780909.15 frames. ], batch size: 53, lr: 4.00e-03, grad_scale: 32.0 2024-09-24 15:12:43,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=528943.3333333334, ans=0.125 2024-09-24 15:12:46,300 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=528943.3333333334, ans=0.125 2024-09-24 15:13:00,083 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.261e+02 1.365e+02 1.554e+02 1.989e+02, threshold=2.731e+02, percent-clipped=0.0 2024-09-24 15:13:16,419 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=528990.0, ans=0.0 2024-09-24 15:13:22,815 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=529036.6666666666, ans=0.125 2024-09-24 15:13:24,317 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=529036.6666666666, ans=0.0 2024-09-24 15:13:27,436 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=529036.6666666666, ans=0.2 2024-09-24 15:13:37,165 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=529083.3333333334, ans=0.1 2024-09-24 15:13:51,535 INFO [train.py:1198] (2/4) Epoch 30, batch 400, loss[loss=0.2214, ctc_loss=0.1459, cr_loss=0.3775, over 16693.00 frames. ], tot_loss[loss=0.2002, ctc_loss=0.1308, cr_loss=0.3472, over 2899904.89 frames. ], batch size: 61, lr: 4.00e-03, grad_scale: 32.0 2024-09-24 15:13:58,998 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.63 vs. limit=15.0 2024-09-24 15:14:11,344 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=11.68 vs. limit=15.0 2024-09-24 15:14:17,711 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=529176.6666666666, ans=0.1 2024-09-24 15:14:33,697 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=529223.3333333334, ans=0.2 2024-09-24 15:14:38,431 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=529270.0, ans=0.125 2024-09-24 15:15:14,475 INFO [train.py:1198] (2/4) Epoch 30, batch 450, loss[loss=0.1831, ctc_loss=0.1192, cr_loss=0.3196, over 17311.00 frames. ], tot_loss[loss=0.2003, ctc_loss=0.1308, cr_loss=0.3474, over 3006676.08 frames. ], batch size: 46, lr: 4.00e-03, grad_scale: 16.0 2024-09-24 15:15:17,078 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.91 vs. limit=10.0 2024-09-24 15:15:29,298 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=529410.0, ans=0.125 2024-09-24 15:15:44,835 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.260e+02 1.335e+02 1.450e+02 2.256e+02, threshold=2.670e+02, percent-clipped=0.0 2024-09-24 15:16:06,438 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=14.36 vs. limit=22.5 2024-09-24 15:16:07,504 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=529503.3333333334, ans=0.2 2024-09-24 15:16:19,100 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=4.42 vs. limit=12.0 2024-09-24 15:16:20,166 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=529550.0, ans=0.125 2024-09-24 15:16:34,281 INFO [train.py:1198] (2/4) Epoch 30, batch 500, loss[loss=0.2081, ctc_loss=0.1349, cr_loss=0.3661, over 17365.00 frames. ], tot_loss[loss=0.2007, ctc_loss=0.1311, cr_loss=0.3482, over 3093953.91 frames. ], batch size: 48, lr: 4.00e-03, grad_scale: 16.0 2024-09-24 15:17:16,731 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=529690.0, ans=0.0 2024-09-24 15:17:31,521 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.97 vs. limit=15.0 2024-09-24 15:18:02,945 INFO [train.py:1198] (2/4) Epoch 30, batch 550, loss[loss=0.189, ctc_loss=0.1216, cr_loss=0.337, over 17299.00 frames. ], tot_loss[loss=0.2014, ctc_loss=0.1316, cr_loss=0.3491, over 3147520.23 frames. ], batch size: 46, lr: 4.00e-03, grad_scale: 16.0 2024-09-24 15:18:04,710 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=529830.0, ans=0.05 2024-09-24 15:18:14,934 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.whiten, num_groups=1, num_channels=512, metric=4.82 vs. limit=12.0 2024-09-24 15:18:17,831 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 15:18:19,337 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=529876.6666666666, ans=0.125 2024-09-24 15:18:24,363 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=529876.6666666666, ans=0.025 2024-09-24 15:18:24,385 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=529876.6666666666, ans=0.0 2024-09-24 15:18:29,089 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=529876.6666666666, ans=0.2 2024-09-24 15:18:33,366 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.121e+02 1.259e+02 1.342e+02 1.453e+02 2.055e+02, threshold=2.683e+02, percent-clipped=0.0 2024-09-24 15:18:40,002 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=529923.3333333334, ans=0.125 2024-09-24 15:18:44,799 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=529923.3333333334, ans=0.2 2024-09-24 15:18:54,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=529970.0, ans=0.125 2024-09-24 15:19:08,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=530016.6666666666, ans=0.0 2024-09-24 15:19:20,507 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=530016.6666666666, ans=0.1 2024-09-24 15:19:23,319 INFO [train.py:1198] (2/4) Epoch 30, batch 600, loss[loss=0.1487, ctc_loss=0.09259, cr_loss=0.2805, over 16216.00 frames. ], tot_loss[loss=0.2016, ctc_loss=0.1318, cr_loss=0.3494, over 3186551.91 frames. ], batch size: 36, lr: 4.00e-03, grad_scale: 16.0 2024-09-24 15:19:36,075 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=530063.3333333334, ans=0.125 2024-09-24 15:19:43,290 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.70 vs. limit=15.0 2024-09-24 15:19:51,720 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=530110.0, ans=0.2 2024-09-24 15:19:57,050 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=8.47 vs. limit=15.0 2024-09-24 15:19:58,194 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=530156.6666666666, ans=0.0 2024-09-24 15:19:58,851 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=7.96 vs. limit=15.0 2024-09-24 15:20:04,735 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=530156.6666666666, ans=0.1 2024-09-24 15:20:09,383 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=530156.6666666666, ans=0.0 2024-09-24 15:20:32,152 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=11.79 vs. limit=15.0 2024-09-24 15:20:45,716 INFO [train.py:1198] (2/4) Epoch 30, batch 650, loss[loss=0.1715, ctc_loss=0.1118, cr_loss=0.2987, over 17052.00 frames. ], tot_loss[loss=0.2008, ctc_loss=0.1311, cr_loss=0.3486, over 3227245.10 frames. ], batch size: 39, lr: 4.00e-03, grad_scale: 16.0 2024-09-24 15:21:04,041 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=6.97 vs. limit=15.0 2024-09-24 15:21:14,467 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=530343.3333333334, ans=0.0 2024-09-24 15:21:15,828 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 1.275e+02 1.371e+02 1.445e+02 2.518e+02, threshold=2.741e+02, percent-clipped=0.0 2024-09-24 15:21:44,276 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=530436.6666666666, ans=0.2 2024-09-24 15:22:10,436 INFO [train.py:1198] (2/4) Epoch 30, batch 700, loss[loss=0.2093, ctc_loss=0.1361, cr_loss=0.3657, over 17194.00 frames. ], tot_loss[loss=0.2004, ctc_loss=0.1308, cr_loss=0.3479, over 3257299.33 frames. ], batch size: 55, lr: 4.00e-03, grad_scale: 16.0 2024-09-24 15:22:29,848 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=530576.6666666666, ans=0.2 2024-09-24 15:22:34,807 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=530576.6666666666, ans=0.125 2024-09-24 15:22:35,167 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.99 vs. limit=10.0 2024-09-24 15:23:08,557 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn2.whiten, num_groups=1, num_channels=192, metric=8.28 vs. limit=22.5 2024-09-24 15:23:09,464 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=530670.0, ans=0.0 2024-09-24 15:23:10,152 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.72 vs. limit=15.0 2024-09-24 15:23:12,557 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=530670.0, ans=0.0 2024-09-24 15:23:19,106 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=530716.6666666666, ans=0.1 2024-09-24 15:23:22,679 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=14.63 vs. limit=15.0 2024-09-24 15:23:33,170 INFO [train.py:1198] (2/4) Epoch 30, batch 750, loss[loss=0.1874, ctc_loss=0.121, cr_loss=0.3319, over 17251.00 frames. ], tot_loss[loss=0.2004, ctc_loss=0.1308, cr_loss=0.3482, over 3283225.57 frames. ], batch size: 44, lr: 4.00e-03, grad_scale: 16.0 2024-09-24 15:23:36,728 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=530763.3333333334, ans=0.125 2024-09-24 15:23:44,722 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=530763.3333333334, ans=0.035 2024-09-24 15:23:59,572 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=530810.0, ans=0.125 2024-09-24 15:24:04,035 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.243e+02 1.346e+02 1.462e+02 2.306e+02, threshold=2.692e+02, percent-clipped=0.0 2024-09-24 15:24:15,505 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=530856.6666666666, ans=0.125 2024-09-24 15:24:17,321 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=530856.6666666666, ans=0.0 2024-09-24 15:24:23,705 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=530903.3333333334, ans=0.125 2024-09-24 15:24:33,879 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.02 vs. limit=22.5 2024-09-24 15:24:45,943 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=530950.0, ans=0.125 2024-09-24 15:24:50,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=530950.0, ans=0.025 2024-09-24 15:24:52,164 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=12.09 vs. limit=15.0 2024-09-24 15:24:56,184 INFO [train.py:1198] (2/4) Epoch 30, batch 800, loss[loss=0.2328, ctc_loss=0.155, cr_loss=0.3893, over 15016.00 frames. ], tot_loss[loss=0.2001, ctc_loss=0.1307, cr_loss=0.3469, over 3301952.39 frames. ], batch size: 89, lr: 3.99e-03, grad_scale: 16.0 2024-09-24 15:25:04,426 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=530996.6666666666, ans=0.2 2024-09-24 15:25:06,270 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.35 vs. limit=10.0 2024-09-24 15:25:08,859 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=530996.6666666666, ans=0.125 2024-09-24 15:25:50,811 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=531136.6666666666, ans=0.125 2024-09-24 15:26:05,508 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=9.78 vs. limit=15.0 2024-09-24 15:26:14,797 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=531230.0, ans=0.1 2024-09-24 15:26:16,195 INFO [train.py:1198] (2/4) Epoch 30, batch 850, loss[loss=0.1661, ctc_loss=0.1061, cr_loss=0.3003, over 17165.00 frames. ], tot_loss[loss=0.1992, ctc_loss=0.13, cr_loss=0.3463, over 3316634.46 frames. ], batch size: 45, lr: 3.99e-03, grad_scale: 16.0 2024-09-24 15:26:18,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=531230.0, ans=0.0 2024-09-24 15:26:26,137 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=531230.0, ans=0.025 2024-09-24 15:26:53,452 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.016e+02 1.252e+02 1.339e+02 1.409e+02 2.350e+02, threshold=2.679e+02, percent-clipped=0.0 2024-09-24 15:27:41,873 INFO [train.py:1198] (2/4) Epoch 30, batch 900, loss[loss=0.2266, ctc_loss=0.149, cr_loss=0.388, over 17004.00 frames. ], tot_loss[loss=0.2001, ctc_loss=0.1306, cr_loss=0.3477, over 3333055.38 frames. ], batch size: 53, lr: 3.99e-03, grad_scale: 16.0 2024-09-24 15:27:53,132 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=531463.3333333334, ans=0.0 2024-09-24 15:28:05,314 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=531510.0, ans=0.0 2024-09-24 15:28:13,337 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.max_abs, batch_count=531510.0, ans=10.0 2024-09-24 15:29:04,215 INFO [train.py:1198] (2/4) Epoch 30, batch 950, loss[loss=0.1706, ctc_loss=0.11, cr_loss=0.3029, over 17243.00 frames. ], tot_loss[loss=0.2007, ctc_loss=0.131, cr_loss=0.3483, over 3339360.53 frames. ], batch size: 42, lr: 3.99e-03, grad_scale: 16.0 2024-09-24 15:29:34,557 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=531790.0, ans=0.0 2024-09-24 15:29:35,837 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.143e+02 1.249e+02 1.333e+02 1.422e+02 1.714e+02, threshold=2.667e+02, percent-clipped=0.0 2024-09-24 15:29:36,187 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=531790.0, ans=0.0 2024-09-24 15:30:02,933 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=531836.6666666666, ans=0.125 2024-09-24 15:30:18,772 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=531883.3333333334, ans=0.0 2024-09-24 15:30:26,508 INFO [train.py:1198] (2/4) Epoch 30, batch 1000, loss[loss=0.214, ctc_loss=0.1405, cr_loss=0.3677, over 17291.00 frames. ], tot_loss[loss=0.2007, ctc_loss=0.131, cr_loss=0.3482, over 3343755.22 frames. ], batch size: 46, lr: 3.99e-03, grad_scale: 16.0 2024-09-24 15:30:43,171 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=531976.6666666666, ans=0.125 2024-09-24 15:31:13,736 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=532070.0, ans=0.125 2024-09-24 15:31:14,071 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.90 vs. limit=15.0 2024-09-24 15:31:18,499 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=532070.0, ans=0.2 2024-09-24 15:31:51,971 INFO [train.py:1198] (2/4) Epoch 30, batch 1050, loss[loss=0.2294, ctc_loss=0.1528, cr_loss=0.3831, over 17003.00 frames. ], tot_loss[loss=0.2017, ctc_loss=0.1318, cr_loss=0.3494, over 3344601.44 frames. ], batch size: 53, lr: 3.99e-03, grad_scale: 16.0 2024-09-24 15:32:21,057 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=532210.0, ans=0.0 2024-09-24 15:32:23,922 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.138e+02 1.266e+02 1.377e+02 1.519e+02 2.640e+02, threshold=2.754e+02, percent-clipped=0.0 2024-09-24 15:32:29,179 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=532256.6666666666, ans=0.0 2024-09-24 15:32:54,348 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=532350.0, ans=0.125 2024-09-24 15:33:03,201 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.min_positive, batch_count=532350.0, ans=0.05 2024-09-24 15:33:14,101 INFO [train.py:1198] (2/4) Epoch 30, batch 1100, loss[loss=0.1589, ctc_loss=0.09996, cr_loss=0.2946, over 17005.00 frames. ], tot_loss[loss=0.2008, ctc_loss=0.1311, cr_loss=0.3486, over 3354570.25 frames. ], batch size: 44, lr: 3.99e-03, grad_scale: 16.0 2024-09-24 15:33:16,068 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=532396.6666666666, ans=0.2 2024-09-24 15:33:22,449 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=532396.6666666666, ans=0.125 2024-09-24 15:33:23,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=532396.6666666666, ans=0.125 2024-09-24 15:33:25,500 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=532396.6666666666, ans=0.07 2024-09-24 15:33:47,530 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=532490.0, ans=0.1 2024-09-24 15:34:03,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=532536.6666666666, ans=0.125 2024-09-24 15:34:03,586 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=532536.6666666666, ans=0.1 2024-09-24 15:34:24,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=532583.3333333334, ans=0.125 2024-09-24 15:34:33,773 INFO [train.py:1198] (2/4) Epoch 30, batch 1150, loss[loss=0.1965, ctc_loss=0.1284, cr_loss=0.3405, over 17023.00 frames. ], tot_loss[loss=0.1995, ctc_loss=0.1301, cr_loss=0.3467, over 3355283.47 frames. ], batch size: 53, lr: 3.99e-03, grad_scale: 16.0 2024-09-24 15:35:09,705 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.070e+02 1.299e+02 1.402e+02 1.532e+02 2.058e+02, threshold=2.804e+02, percent-clipped=0.0 2024-09-24 15:35:21,314 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=532723.3333333334, ans=0.09899494936611666 2024-09-24 15:35:37,046 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=532770.0, ans=0.0 2024-09-24 15:35:40,241 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=532816.6666666666, ans=0.0 2024-09-24 15:35:42,022 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=532816.6666666666, ans=0.125 2024-09-24 15:35:50,814 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module1.whiten, num_groups=1, num_channels=192, metric=7.44 vs. limit=15.0 2024-09-24 15:35:56,125 INFO [train.py:1198] (2/4) Epoch 30, batch 1200, loss[loss=0.2246, ctc_loss=0.1454, cr_loss=0.3958, over 17226.00 frames. ], tot_loss[loss=0.2003, ctc_loss=0.1308, cr_loss=0.3477, over 3358030.64 frames. ], batch size: 50, lr: 3.99e-03, grad_scale: 16.0 2024-09-24 15:36:36,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=532956.6666666666, ans=0.125 2024-09-24 15:36:49,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=533003.3333333334, ans=0.125 2024-09-24 15:36:57,951 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=533003.3333333334, ans=0.0 2024-09-24 15:37:09,046 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=533050.0, ans=0.125 2024-09-24 15:37:13,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=533050.0, ans=0.0 2024-09-24 15:37:21,569 INFO [train.py:1198] (2/4) Epoch 30, batch 1250, loss[loss=0.1935, ctc_loss=0.126, cr_loss=0.3372, over 17090.00 frames. ], tot_loss[loss=0.2005, ctc_loss=0.1309, cr_loss=0.3481, over 3362426.25 frames. ], batch size: 43, lr: 3.99e-03, grad_scale: 16.0 2024-09-24 15:37:21,912 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=533096.6666666666, ans=0.0 2024-09-24 15:37:26,590 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=533096.6666666666, ans=0.125 2024-09-24 15:37:57,234 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.068e+02 1.280e+02 1.378e+02 1.490e+02 1.846e+02, threshold=2.756e+02, percent-clipped=0.0 2024-09-24 15:38:07,953 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module1.whiten, num_groups=1, num_channels=192, metric=8.12 vs. limit=15.0 2024-09-24 15:38:17,029 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=4.62 vs. limit=15.0 2024-09-24 15:38:32,514 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=533283.3333333334, ans=0.2 2024-09-24 15:38:35,902 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=18.58 vs. limit=22.5 2024-09-24 15:38:43,374 INFO [train.py:1198] (2/4) Epoch 30, batch 1300, loss[loss=0.244, ctc_loss=0.1618, cr_loss=0.411, over 16993.00 frames. ], tot_loss[loss=0.2009, ctc_loss=0.1311, cr_loss=0.3489, over 3352907.05 frames. ], batch size: 53, lr: 3.99e-03, grad_scale: 16.0 2024-09-24 15:38:51,717 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=533330.0, ans=0.125 2024-09-24 15:38:56,340 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=533330.0, ans=0.0 2024-09-24 15:39:31,609 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten.whitening_limit, batch_count=533470.0, ans=15.0 2024-09-24 15:39:54,587 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=533516.6666666666, ans=0.125 2024-09-24 15:40:05,652 INFO [train.py:1198] (2/4) Epoch 30, batch 1350, loss[loss=0.2135, ctc_loss=0.1403, cr_loss=0.3657, over 17156.00 frames. ], tot_loss[loss=0.2002, ctc_loss=0.1306, cr_loss=0.3479, over 3361324.72 frames. ], batch size: 48, lr: 3.99e-03, grad_scale: 16.0 2024-09-24 15:40:10,625 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=533563.3333333334, ans=0.125 2024-09-24 15:40:32,794 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=533610.0, ans=0.125 2024-09-24 15:40:38,858 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.284e+02 1.374e+02 1.506e+02 2.096e+02, threshold=2.749e+02, percent-clipped=0.0 2024-09-24 15:40:55,453 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=533703.3333333334, ans=0.125 2024-09-24 15:41:01,967 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=533703.3333333334, ans=0.125 2024-09-24 15:41:26,034 INFO [train.py:1198] (2/4) Epoch 30, batch 1400, loss[loss=0.2156, ctc_loss=0.1397, cr_loss=0.3796, over 16725.00 frames. ], tot_loss[loss=0.2017, ctc_loss=0.1317, cr_loss=0.3502, over 3353942.59 frames. ], batch size: 61, lr: 3.98e-03, grad_scale: 16.0 2024-09-24 15:41:29,494 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=533796.6666666666, ans=0.125 2024-09-24 15:42:00,728 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=533843.3333333334, ans=0.0 2024-09-24 15:42:12,166 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=533890.0, ans=0.025 2024-09-24 15:42:31,526 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=533936.6666666666, ans=0.1 2024-09-24 15:42:33,711 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.whiten, num_groups=1, num_channels=512, metric=4.99 vs. limit=12.0 2024-09-24 15:42:36,494 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=533983.3333333334, ans=0.0 2024-09-24 15:42:54,635 INFO [train.py:1198] (2/4) Epoch 30, batch 1450, loss[loss=0.2038, ctc_loss=0.1321, cr_loss=0.3585, over 16787.00 frames. ], tot_loss[loss=0.2012, ctc_loss=0.1313, cr_loss=0.3497, over 3363173.44 frames. ], batch size: 61, lr: 3.98e-03, grad_scale: 16.0 2024-09-24 15:43:01,461 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=534030.0, ans=0.025 2024-09-24 15:43:03,368 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=5.84 vs. limit=15.0 2024-09-24 15:43:10,738 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=534076.6666666666, ans=0.0 2024-09-24 15:43:15,557 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=534076.6666666666, ans=0.1 2024-09-24 15:43:25,328 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=534123.3333333334, ans=0.125 2024-09-24 15:43:28,224 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.340e+02 1.457e+02 1.567e+02 2.615e+02, threshold=2.914e+02, percent-clipped=0.0 2024-09-24 15:44:02,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=534216.6666666666, ans=0.1 2024-09-24 15:44:05,823 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=6.96 vs. limit=15.0 2024-09-24 15:44:14,826 INFO [train.py:1198] (2/4) Epoch 30, batch 1500, loss[loss=0.1794, ctc_loss=0.1155, cr_loss=0.3197, over 17005.00 frames. ], tot_loss[loss=0.1999, ctc_loss=0.1304, cr_loss=0.3476, over 3367602.43 frames. ], batch size: 44, lr: 3.98e-03, grad_scale: 16.0 2024-09-24 15:44:34,084 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=534310.0, ans=0.1 2024-09-24 15:44:34,192 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=534310.0, ans=0.1 2024-09-24 15:44:40,745 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=534310.0, ans=0.0 2024-09-24 15:45:19,988 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=7.56 vs. limit=15.0 2024-09-24 15:45:26,302 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=534450.0, ans=0.125 2024-09-24 15:45:27,790 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=534450.0, ans=0.125 2024-09-24 15:45:27,876 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=534450.0, ans=0.1 2024-09-24 15:45:32,669 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=534450.0, ans=0.125 2024-09-24 15:45:37,368 INFO [train.py:1198] (2/4) Epoch 30, batch 1550, loss[loss=0.1972, ctc_loss=0.1291, cr_loss=0.3408, over 17301.00 frames. ], tot_loss[loss=0.1992, ctc_loss=0.1299, cr_loss=0.3464, over 3360063.37 frames. ], batch size: 46, lr: 3.98e-03, grad_scale: 16.0 2024-09-24 15:45:48,885 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=534496.6666666666, ans=0.125 2024-09-24 15:46:04,266 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module1.whiten, num_groups=1, num_channels=192, metric=7.73 vs. limit=15.0 2024-09-24 15:46:10,123 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=10.96 vs. limit=22.5 2024-09-24 15:46:11,088 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.255e+02 1.331e+02 1.418e+02 1.761e+02, threshold=2.662e+02, percent-clipped=0.0 2024-09-24 15:46:36,477 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.90 vs. limit=15.0 2024-09-24 15:46:46,829 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=534683.3333333334, ans=0.0 2024-09-24 15:46:48,330 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=534683.3333333334, ans=0.07 2024-09-24 15:46:53,344 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-24 15:47:00,125 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=10.15 vs. limit=15.0 2024-09-24 15:47:01,134 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=534730.0, ans=0.125 2024-09-24 15:47:02,574 INFO [train.py:1198] (2/4) Epoch 30, batch 1600, loss[loss=0.1829, ctc_loss=0.1173, cr_loss=0.3279, over 16245.00 frames. ], tot_loss[loss=0.1985, ctc_loss=0.1293, cr_loss=0.346, over 3368349.60 frames. ], batch size: 36, lr: 3.98e-03, grad_scale: 32.0 2024-09-24 15:47:04,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=534730.0, ans=0.0 2024-09-24 15:47:53,135 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=11.77 vs. limit=22.5 2024-09-24 15:48:25,593 INFO [train.py:1198] (2/4) Epoch 30, batch 1650, loss[loss=0.1789, ctc_loss=0.1133, cr_loss=0.3279, over 17293.00 frames. ], tot_loss[loss=0.1985, ctc_loss=0.1293, cr_loss=0.3458, over 3376192.01 frames. ], batch size: 42, lr: 3.98e-03, grad_scale: 32.0 2024-09-24 15:48:32,147 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=534963.3333333334, ans=0.025 2024-09-24 15:48:40,039 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=535010.0, ans=0.0 2024-09-24 15:48:41,713 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=535010.0, ans=0.125 2024-09-24 15:48:46,708 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=535010.0, ans=0.125 2024-09-24 15:48:59,209 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.272e+02 1.342e+02 1.418e+02 1.836e+02, threshold=2.683e+02, percent-clipped=0.0 2024-09-24 15:49:04,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=535056.6666666666, ans=0.1 2024-09-24 15:49:18,865 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=9.71 vs. limit=15.0 2024-09-24 15:49:21,572 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=535103.3333333334, ans=0.125 2024-09-24 15:49:45,317 INFO [train.py:1198] (2/4) Epoch 30, batch 1700, loss[loss=0.2058, ctc_loss=0.1355, cr_loss=0.3514, over 17306.00 frames. ], tot_loss[loss=0.198, ctc_loss=0.1291, cr_loss=0.3448, over 3380416.87 frames. ], batch size: 51, lr: 3.98e-03, grad_scale: 32.0 2024-09-24 15:50:26,840 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.min_abs, batch_count=535290.0, ans=0.5 2024-09-24 15:50:52,507 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=535383.3333333334, ans=0.125 2024-09-24 15:51:08,583 INFO [train.py:1198] (2/4) Epoch 30, batch 1750, loss[loss=0.2093, ctc_loss=0.1375, cr_loss=0.3587, over 16475.00 frames. ], tot_loss[loss=0.199, ctc_loss=0.1298, cr_loss=0.3458, over 3379607.88 frames. ], batch size: 66, lr: 3.98e-03, grad_scale: 32.0 2024-09-24 15:51:32,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=535476.6666666666, ans=0.125 2024-09-24 15:51:47,335 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.161e+02 1.310e+02 1.375e+02 1.459e+02 2.196e+02, threshold=2.749e+02, percent-clipped=0.0 2024-09-24 15:51:52,367 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 15:52:27,289 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=535616.6666666666, ans=0.0 2024-09-24 15:52:33,547 INFO [train.py:1198] (2/4) Epoch 30, batch 1800, loss[loss=0.2245, ctc_loss=0.1466, cr_loss=0.3898, over 17033.00 frames. ], tot_loss[loss=0.1997, ctc_loss=0.1303, cr_loss=0.3468, over 3372548.63 frames. ], batch size: 51, lr: 3.98e-03, grad_scale: 32.0 2024-09-24 15:52:54,031 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=535710.0, ans=0.1 2024-09-24 15:53:12,170 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=535756.6666666666, ans=0.125 2024-09-24 15:53:20,596 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.18 vs. limit=22.5 2024-09-24 15:53:56,383 INFO [train.py:1198] (2/4) Epoch 30, batch 1850, loss[loss=0.2168, ctc_loss=0.1451, cr_loss=0.3583, over 17023.00 frames. ], tot_loss[loss=0.2, ctc_loss=0.1305, cr_loss=0.3477, over 3376950.58 frames. ], batch size: 56, lr: 3.98e-03, grad_scale: 32.0 2024-09-24 15:54:09,554 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=535896.6666666666, ans=0.0 2024-09-24 15:54:22,464 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=535943.3333333334, ans=0.1 2024-09-24 15:54:28,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=535990.0, ans=0.1 2024-09-24 15:54:30,248 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.250e+02 1.313e+02 1.396e+02 1.985e+02, threshold=2.627e+02, percent-clipped=0.0 2024-09-24 15:54:35,170 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=535990.0, ans=0.125 2024-09-24 15:55:16,418 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=536083.3333333334, ans=0.125 2024-09-24 15:55:19,283 INFO [train.py:1198] (2/4) Epoch 30, batch 1900, loss[loss=0.1824, ctc_loss=0.1187, cr_loss=0.3182, over 17039.00 frames. ], tot_loss[loss=0.2008, ctc_loss=0.131, cr_loss=0.3493, over 3380588.67 frames. ], batch size: 44, lr: 3.98e-03, grad_scale: 32.0 2024-09-24 15:55:22,706 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=536130.0, ans=0.125 2024-09-24 15:55:32,662 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=536130.0, ans=0.0 2024-09-24 15:56:01,820 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=4.77 vs. limit=12.0 2024-09-24 15:56:14,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=536270.0, ans=0.125 2024-09-24 15:56:34,037 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=536316.6666666666, ans=0.125 2024-09-24 15:56:44,474 INFO [train.py:1198] (2/4) Epoch 30, batch 1950, loss[loss=0.1881, ctc_loss=0.1227, cr_loss=0.327, over 16671.00 frames. ], tot_loss[loss=0.2003, ctc_loss=0.1306, cr_loss=0.3486, over 3376624.10 frames. ], batch size: 37, lr: 3.97e-03, grad_scale: 32.0 2024-09-24 15:56:47,934 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=536363.3333333334, ans=0.125 2024-09-24 15:56:53,409 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.57 vs. limit=15.0 2024-09-24 15:56:54,322 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=536363.3333333334, ans=0.125 2024-09-24 15:57:00,809 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=536410.0, ans=0.125 2024-09-24 15:57:05,386 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=536410.0, ans=0.125 2024-09-24 15:57:19,471 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.270e+02 1.356e+02 1.450e+02 2.095e+02, threshold=2.713e+02, percent-clipped=0.0 2024-09-24 15:57:44,652 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=536503.3333333334, ans=0.1 2024-09-24 15:57:48,512 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.92 vs. limit=15.0 2024-09-24 15:57:52,741 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=536550.0, ans=0.1 2024-09-24 15:58:00,641 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=1.533e-01 2024-09-24 15:58:06,788 INFO [train.py:1198] (2/4) Epoch 30, batch 2000, loss[loss=0.1575, ctc_loss=0.09741, cr_loss=0.3006, over 17055.00 frames. ], tot_loss[loss=0.1986, ctc_loss=0.1294, cr_loss=0.3462, over 3371127.27 frames. ], batch size: 39, lr: 3.97e-03, grad_scale: 32.0 2024-09-24 15:58:13,420 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=536596.6666666666, ans=0.025 2024-09-24 15:58:19,745 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=536596.6666666666, ans=0.125 2024-09-24 15:58:53,566 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=536736.6666666666, ans=0.1 2024-09-24 15:59:00,117 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=536736.6666666666, ans=0.05 2024-09-24 15:59:14,453 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=536783.3333333334, ans=0.125 2024-09-24 15:59:15,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=536783.3333333334, ans=0.125 2024-09-24 15:59:24,140 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=536783.3333333334, ans=0.0 2024-09-24 15:59:26,920 INFO [train.py:1198] (2/4) Epoch 30, batch 2050, loss[loss=0.1857, ctc_loss=0.1171, cr_loss=0.3426, over 17113.00 frames. ], tot_loss[loss=0.1987, ctc_loss=0.1295, cr_loss=0.3461, over 3370051.73 frames. ], batch size: 43, lr: 3.97e-03, grad_scale: 32.0 2024-09-24 15:59:32,047 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=536830.0, ans=0.125 2024-09-24 15:59:57,033 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=536876.6666666666, ans=0.05 2024-09-24 16:00:04,534 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.272e+02 1.366e+02 1.463e+02 2.391e+02, threshold=2.732e+02, percent-clipped=0.0 2024-09-24 16:00:26,461 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.whiten, num_groups=1, num_channels=192, metric=4.86 vs. limit=12.0 2024-09-24 16:00:32,265 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=537016.6666666666, ans=0.125 2024-09-24 16:00:33,768 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=537016.6666666666, ans=0.0 2024-09-24 16:00:49,745 INFO [train.py:1198] (2/4) Epoch 30, batch 2100, loss[loss=0.1932, ctc_loss=0.126, cr_loss=0.3363, over 17144.00 frames. ], tot_loss[loss=0.1982, ctc_loss=0.1291, cr_loss=0.3456, over 3380222.79 frames. ], batch size: 48, lr: 3.97e-03, grad_scale: 32.0 2024-09-24 16:01:37,210 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer_ff2.min_abs, batch_count=537156.6666666666, ans=0.1 2024-09-24 16:01:51,855 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=13.44 vs. limit=22.5 2024-09-24 16:02:11,945 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=537250.0, ans=0.125 2024-09-24 16:02:14,603 INFO [train.py:1198] (2/4) Epoch 30, batch 2150, loss[loss=0.215, ctc_loss=0.1389, cr_loss=0.3807, over 15840.00 frames. ], tot_loss[loss=0.1978, ctc_loss=0.1287, cr_loss=0.3452, over 3372577.61 frames. ], batch size: 74, lr: 3.97e-03, grad_scale: 32.0 2024-09-24 16:02:18,227 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 16:02:38,262 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=537343.3333333334, ans=0.125 2024-09-24 16:02:44,508 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=537343.3333333334, ans=0.125 2024-09-24 16:02:47,888 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=537390.0, ans=0.2 2024-09-24 16:02:52,195 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.110e+02 1.245e+02 1.327e+02 1.449e+02 2.310e+02, threshold=2.654e+02, percent-clipped=0.0 2024-09-24 16:03:12,344 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=6.47 vs. limit=15.0 2024-09-24 16:03:17,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=537436.6666666666, ans=0.1 2024-09-24 16:03:35,925 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=13.34 vs. limit=22.5 2024-09-24 16:03:36,774 INFO [train.py:1198] (2/4) Epoch 30, batch 2200, loss[loss=0.1758, ctc_loss=0.1154, cr_loss=0.3017, over 17188.00 frames. ], tot_loss[loss=0.1984, ctc_loss=0.1293, cr_loss=0.3454, over 3367993.08 frames. ], batch size: 41, lr: 3.97e-03, grad_scale: 32.0 2024-09-24 16:03:37,185 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=537530.0, ans=0.0 2024-09-24 16:03:45,118 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=537530.0, ans=0.125 2024-09-24 16:03:48,299 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=537530.0, ans=0.0 2024-09-24 16:03:56,168 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=537576.6666666666, ans=0.2 2024-09-24 16:04:46,023 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=537716.6666666666, ans=0.125 2024-09-24 16:05:00,238 INFO [train.py:1198] (2/4) Epoch 30, batch 2250, loss[loss=0.1834, ctc_loss=0.1197, cr_loss=0.3183, over 17306.00 frames. ], tot_loss[loss=0.1984, ctc_loss=0.1293, cr_loss=0.3453, over 3375397.35 frames. ], batch size: 46, lr: 3.97e-03, grad_scale: 32.0 2024-09-24 16:05:02,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=537763.3333333334, ans=0.2 2024-09-24 16:05:11,786 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=537763.3333333334, ans=0.2 2024-09-24 16:05:18,458 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=537810.0, ans=0.2 2024-09-24 16:05:35,468 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.139e+02 1.295e+02 1.394e+02 1.566e+02 2.386e+02, threshold=2.787e+02, percent-clipped=0.0 2024-09-24 16:05:40,579 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=537856.6666666666, ans=0.1 2024-09-24 16:05:51,952 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=537903.3333333334, ans=0.2 2024-09-24 16:06:03,238 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=537950.0, ans=0.125 2024-09-24 16:06:11,674 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.77 vs. limit=10.0 2024-09-24 16:06:14,243 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=537950.0, ans=0.125 2024-09-24 16:06:20,401 INFO [train.py:1198] (2/4) Epoch 30, batch 2300, loss[loss=0.2203, ctc_loss=0.1446, cr_loss=0.3782, over 16895.00 frames. ], tot_loss[loss=0.1985, ctc_loss=0.1294, cr_loss=0.3455, over 3361579.25 frames. ], batch size: 58, lr: 3.97e-03, grad_scale: 32.0 2024-09-24 16:06:44,800 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=538043.3333333334, ans=0.125 2024-09-24 16:06:52,064 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn1.whiten, num_groups=1, num_channels=192, metric=9.28 vs. limit=22.5 2024-09-24 16:06:54,579 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=538043.3333333334, ans=0.1 2024-09-24 16:07:47,753 INFO [train.py:1198] (2/4) Epoch 30, batch 2350, loss[loss=0.188, ctc_loss=0.123, cr_loss=0.325, over 17297.00 frames. ], tot_loss[loss=0.1999, ctc_loss=0.1304, cr_loss=0.3473, over 3355789.21 frames. ], batch size: 49, lr: 3.97e-03, grad_scale: 32.0 2024-09-24 16:08:15,662 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten.whitening_limit, batch_count=538276.6666666666, ans=15.0 2024-09-24 16:08:23,164 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.120e+02 1.275e+02 1.344e+02 1.471e+02 2.396e+02, threshold=2.688e+02, percent-clipped=0.0 2024-09-24 16:08:49,348 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=538370.0, ans=0.125 2024-09-24 16:09:08,042 INFO [train.py:1198] (2/4) Epoch 30, batch 2400, loss[loss=0.1939, ctc_loss=0.1277, cr_loss=0.3309, over 16068.00 frames. ], tot_loss[loss=0.2002, ctc_loss=0.1306, cr_loss=0.3479, over 3363283.58 frames. ], batch size: 74, lr: 3.97e-03, grad_scale: 32.0 2024-09-24 16:09:22,849 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=538510.0, ans=0.1 2024-09-24 16:09:49,273 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=538556.6666666666, ans=0.0 2024-09-24 16:09:55,602 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=538556.6666666666, ans=0.125 2024-09-24 16:10:21,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=538650.0, ans=0.125 2024-09-24 16:10:29,060 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=538696.6666666666, ans=0.125 2024-09-24 16:10:29,681 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.91 vs. limit=6.0 2024-09-24 16:10:30,237 INFO [train.py:1198] (2/4) Epoch 30, batch 2450, loss[loss=0.1994, ctc_loss=0.1296, cr_loss=0.3489, over 17290.00 frames. ], tot_loss[loss=0.1992, ctc_loss=0.1299, cr_loss=0.3466, over 3361524.65 frames. ], batch size: 51, lr: 3.97e-03, grad_scale: 32.0 2024-09-24 16:10:32,143 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=538696.6666666666, ans=0.125 2024-09-24 16:10:41,755 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=538696.6666666666, ans=0.0 2024-09-24 16:11:05,480 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.006e+02 1.261e+02 1.331e+02 1.438e+02 1.772e+02, threshold=2.662e+02, percent-clipped=0.0 2024-09-24 16:11:55,427 INFO [train.py:1198] (2/4) Epoch 30, batch 2500, loss[loss=0.1668, ctc_loss=0.1051, cr_loss=0.3084, over 16977.00 frames. ], tot_loss[loss=0.1993, ctc_loss=0.1299, cr_loss=0.3469, over 3355880.75 frames. ], batch size: 42, lr: 3.97e-03, grad_scale: 32.0 2024-09-24 16:12:33,066 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=539023.3333333334, ans=0.2 2024-09-24 16:13:07,083 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=539116.6666666666, ans=0.09899494936611666 2024-09-24 16:13:17,973 INFO [train.py:1198] (2/4) Epoch 30, batch 2550, loss[loss=0.189, ctc_loss=0.1226, cr_loss=0.3321, over 17328.00 frames. ], tot_loss[loss=0.1997, ctc_loss=0.1302, cr_loss=0.3475, over 3361585.34 frames. ], batch size: 52, lr: 3.96e-03, grad_scale: 32.0 2024-09-24 16:13:29,651 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=539163.3333333334, ans=0.2 2024-09-24 16:13:53,195 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.247e+02 1.346e+02 1.482e+02 2.162e+02, threshold=2.691e+02, percent-clipped=0.0 2024-09-24 16:14:17,079 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=539303.3333333334, ans=0.125 2024-09-24 16:14:24,080 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=14.77 vs. limit=22.5 2024-09-24 16:14:25,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=539350.0, ans=0.0 2024-09-24 16:14:40,187 INFO [train.py:1198] (2/4) Epoch 30, batch 2600, loss[loss=0.2158, ctc_loss=0.1405, cr_loss=0.3767, over 17137.00 frames. ], tot_loss[loss=0.2003, ctc_loss=0.1306, cr_loss=0.3485, over 3356964.96 frames. ], batch size: 48, lr: 3.96e-03, grad_scale: 16.0 2024-09-24 16:14:40,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=539396.6666666666, ans=0.2 2024-09-24 16:14:40,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=539396.6666666666, ans=0.125 2024-09-24 16:14:40,827 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=7.69 vs. limit=15.0 2024-09-24 16:15:47,762 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=539583.3333333334, ans=0.125 2024-09-24 16:16:00,286 INFO [train.py:1198] (2/4) Epoch 30, batch 2650, loss[loss=0.1879, ctc_loss=0.1194, cr_loss=0.3426, over 17192.00 frames. ], tot_loss[loss=0.2005, ctc_loss=0.1308, cr_loss=0.3488, over 3355766.15 frames. ], batch size: 41, lr: 3.96e-03, grad_scale: 16.0 2024-09-24 16:16:00,622 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=539630.0, ans=0.2 2024-09-24 16:16:29,169 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=539676.6666666666, ans=0.125 2024-09-24 16:16:30,958 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=539676.6666666666, ans=0.125 2024-09-24 16:16:42,933 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.102e+02 1.284e+02 1.354e+02 1.501e+02 2.171e+02, threshold=2.708e+02, percent-clipped=0.0 2024-09-24 16:16:57,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=539770.0, ans=0.0 2024-09-24 16:17:15,186 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=539816.6666666666, ans=0.125 2024-09-24 16:17:16,702 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=539816.6666666666, ans=0.125 2024-09-24 16:17:28,543 INFO [train.py:1198] (2/4) Epoch 30, batch 2700, loss[loss=0.1986, ctc_loss=0.1251, cr_loss=0.3674, over 17311.00 frames. ], tot_loss[loss=0.1999, ctc_loss=0.1303, cr_loss=0.3478, over 3363116.56 frames. ], batch size: 46, lr: 3.96e-03, grad_scale: 16.0 2024-09-24 16:17:30,692 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=11.99 vs. limit=15.0 2024-09-24 16:17:41,345 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=539863.3333333334, ans=0.0 2024-09-24 16:17:43,058 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=539910.0, ans=0.0 2024-09-24 16:18:15,396 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=16.17 vs. limit=22.5 2024-09-24 16:18:21,946 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.39 vs. limit=15.0 2024-09-24 16:18:45,032 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 16:18:46,714 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=540096.6666666666, ans=0.125 2024-09-24 16:18:48,019 INFO [train.py:1198] (2/4) Epoch 30, batch 2750, loss[loss=0.2386, ctc_loss=0.1603, cr_loss=0.3915, over 16901.00 frames. ], tot_loss[loss=0.1998, ctc_loss=0.1303, cr_loss=0.3473, over 3363599.75 frames. ], batch size: 58, lr: 3.96e-03, grad_scale: 16.0 2024-09-24 16:18:48,719 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.35 vs. limit=15.0 2024-09-24 16:19:04,282 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=540143.3333333334, ans=0.125 2024-09-24 16:19:26,135 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.162e+02 1.254e+02 1.332e+02 1.454e+02 2.287e+02, threshold=2.664e+02, percent-clipped=0.0 2024-09-24 16:19:26,493 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=540190.0, ans=0.025 2024-09-24 16:19:32,852 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=540190.0, ans=0.1 2024-09-24 16:19:42,044 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=540236.6666666666, ans=0.0 2024-09-24 16:19:43,788 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=540236.6666666666, ans=0.125 2024-09-24 16:20:10,880 INFO [train.py:1198] (2/4) Epoch 30, batch 2800, loss[loss=0.2003, ctc_loss=0.1284, cr_loss=0.3593, over 17094.00 frames. ], tot_loss[loss=0.1996, ctc_loss=0.1301, cr_loss=0.3473, over 3369989.86 frames. ], batch size: 43, lr: 3.96e-03, grad_scale: 16.0 2024-09-24 16:20:17,710 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=540330.0, ans=0.125 2024-09-24 16:20:35,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=540376.6666666666, ans=0.125 2024-09-24 16:20:44,410 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.95 vs. limit=15.0 2024-09-24 16:20:51,642 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=540423.3333333334, ans=0.125 2024-09-24 16:21:01,363 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=540470.0, ans=0.125 2024-09-24 16:21:06,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=540470.0, ans=0.0 2024-09-24 16:21:25,012 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=540516.6666666666, ans=0.1 2024-09-24 16:21:26,739 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=540516.6666666666, ans=0.0 2024-09-24 16:21:30,882 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=540516.6666666666, ans=0.1 2024-09-24 16:21:36,739 INFO [train.py:1198] (2/4) Epoch 30, batch 2850, loss[loss=0.1935, ctc_loss=0.1283, cr_loss=0.326, over 17158.00 frames. ], tot_loss[loss=0.1994, ctc_loss=0.13, cr_loss=0.3471, over 3375226.85 frames. ], batch size: 45, lr: 3.96e-03, grad_scale: 16.0 2024-09-24 16:22:15,506 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.298e+02 1.387e+02 1.478e+02 2.436e+02, threshold=2.774e+02, percent-clipped=0.0 2024-09-24 16:22:17,566 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=540656.6666666666, ans=0.025 2024-09-24 16:22:26,396 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=540703.3333333334, ans=0.2 2024-09-24 16:22:28,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=540703.3333333334, ans=0.025 2024-09-24 16:22:32,948 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=540703.3333333334, ans=0.0 2024-09-24 16:22:33,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=540703.3333333334, ans=0.2 2024-09-24 16:22:36,214 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=540703.3333333334, ans=0.2 2024-09-24 16:22:56,124 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.67 vs. limit=15.0 2024-09-24 16:22:58,780 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=540796.6666666666, ans=0.125 2024-09-24 16:23:00,143 INFO [train.py:1198] (2/4) Epoch 30, batch 2900, loss[loss=0.2139, ctc_loss=0.1412, cr_loss=0.3635, over 17069.00 frames. ], tot_loss[loss=0.2, ctc_loss=0.1304, cr_loss=0.3482, over 3374378.73 frames. ], batch size: 52, lr: 3.96e-03, grad_scale: 16.0 2024-09-24 16:23:11,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=540796.6666666666, ans=0.0 2024-09-24 16:23:36,157 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=13.00 vs. limit=22.5 2024-09-24 16:23:40,734 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.min_positive, batch_count=540890.0, ans=0.05 2024-09-24 16:23:40,803 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.max_abs, batch_count=540890.0, ans=10.0 2024-09-24 16:24:20,582 INFO [train.py:1198] (2/4) Epoch 30, batch 2950, loss[loss=0.1552, ctc_loss=0.09751, cr_loss=0.2883, over 17253.00 frames. ], tot_loss[loss=0.1998, ctc_loss=0.1303, cr_loss=0.3475, over 3369054.24 frames. ], batch size: 42, lr: 3.96e-03, grad_scale: 16.0 2024-09-24 16:24:52,684 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.87 vs. limit=15.0 2024-09-24 16:24:55,586 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=541123.3333333334, ans=0.2 2024-09-24 16:25:01,511 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.034e+02 1.276e+02 1.368e+02 1.488e+02 1.810e+02, threshold=2.735e+02, percent-clipped=0.0 2024-09-24 16:25:01,952 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=541123.3333333334, ans=0.125 2024-09-24 16:25:25,416 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=541216.6666666666, ans=0.0 2024-09-24 16:25:42,514 INFO [train.py:1198] (2/4) Epoch 30, batch 3000, loss[loss=0.2009, ctc_loss=0.1322, cr_loss=0.3438, over 17211.00 frames. ], tot_loss[loss=0.2004, ctc_loss=0.1307, cr_loss=0.3488, over 3360672.29 frames. ], batch size: 47, lr: 3.96e-03, grad_scale: 16.0 2024-09-24 16:25:42,514 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 16:25:57,421 INFO [train.py:1230] (2/4) Epoch 30, validation: loss=0.03649, ctc_loss=0.03649, cr_loss=8.522e-15, over 944034.00 frames. 2024-09-24 16:25:57,422 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 16:26:07,203 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=541263.3333333334, ans=0.0 2024-09-24 16:26:08,938 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=541263.3333333334, ans=0.0 2024-09-24 16:26:14,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=541310.0, ans=0.1 2024-09-24 16:26:27,788 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=541310.0, ans=0.1 2024-09-24 16:26:34,226 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 16:26:36,276 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.29 vs. limit=15.0 2024-09-24 16:26:40,356 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=541356.6666666666, ans=0.125 2024-09-24 16:26:56,015 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=541403.3333333334, ans=0.125 2024-09-24 16:26:57,646 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=541403.3333333334, ans=0.0 2024-09-24 16:26:57,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=541403.3333333334, ans=0.125 2024-09-24 16:26:59,418 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=541403.3333333334, ans=0.025 2024-09-24 16:27:04,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=541403.3333333334, ans=0.125 2024-09-24 16:27:17,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=541450.0, ans=0.125 2024-09-24 16:27:23,552 INFO [train.py:1198] (2/4) Epoch 30, batch 3050, loss[loss=0.2294, ctc_loss=0.1518, cr_loss=0.3881, over 17224.00 frames. ], tot_loss[loss=0.2016, ctc_loss=0.1315, cr_loss=0.3508, over 3350551.31 frames. ], batch size: 55, lr: 3.96e-03, grad_scale: 16.0 2024-09-24 16:27:31,549 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=541496.6666666666, ans=0.0 2024-09-24 16:27:56,371 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=541590.0, ans=0.1 2024-09-24 16:28:00,733 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.275e+02 1.345e+02 1.423e+02 1.790e+02, threshold=2.690e+02, percent-clipped=0.0 2024-09-24 16:28:07,183 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=541590.0, ans=0.125 2024-09-24 16:28:15,067 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=541636.6666666666, ans=0.1 2024-09-24 16:28:30,556 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=541683.3333333334, ans=0.125 2024-09-24 16:28:41,094 INFO [train.py:1198] (2/4) Epoch 30, batch 3100, loss[loss=0.193, ctc_loss=0.1231, cr_loss=0.3494, over 17270.00 frames. ], tot_loss[loss=0.2008, ctc_loss=0.1309, cr_loss=0.3495, over 3358024.45 frames. ], batch size: 42, lr: 3.96e-03, grad_scale: 16.0 2024-09-24 16:28:44,555 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=541730.0, ans=0.125 2024-09-24 16:28:58,879 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=541776.6666666666, ans=0.1 2024-09-24 16:29:21,335 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=541823.3333333334, ans=0.0 2024-09-24 16:29:30,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=541870.0, ans=0.125 2024-09-24 16:29:59,059 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=541916.6666666666, ans=0.2 2024-09-24 16:29:59,098 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=541916.6666666666, ans=0.125 2024-09-24 16:30:01,866 INFO [train.py:1198] (2/4) Epoch 30, batch 3150, loss[loss=0.189, ctc_loss=0.1234, cr_loss=0.3281, over 17010.00 frames. ], tot_loss[loss=0.1997, ctc_loss=0.1302, cr_loss=0.3478, over 3364807.03 frames. ], batch size: 44, lr: 3.95e-03, grad_scale: 16.0 2024-09-24 16:30:15,912 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=542010.0, ans=0.125 2024-09-24 16:30:15,917 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=542010.0, ans=0.125 2024-09-24 16:30:19,111 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=542010.0, ans=0.125 2024-09-24 16:30:28,454 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=542010.0, ans=0.0 2024-09-24 16:30:29,140 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=4.75 vs. limit=10.0 2024-09-24 16:30:39,238 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.276e+02 1.352e+02 1.455e+02 1.845e+02, threshold=2.704e+02, percent-clipped=0.0 2024-09-24 16:30:46,336 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.87 vs. limit=15.0 2024-09-24 16:30:50,600 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=542103.3333333334, ans=0.025 2024-09-24 16:31:20,115 INFO [train.py:1198] (2/4) Epoch 30, batch 3200, loss[loss=0.1814, ctc_loss=0.1161, cr_loss=0.3264, over 17261.00 frames. ], tot_loss[loss=0.1996, ctc_loss=0.1301, cr_loss=0.3476, over 3369200.22 frames. ], batch size: 42, lr: 3.95e-03, grad_scale: 32.0 2024-09-24 16:31:28,342 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=542196.6666666666, ans=0.1 2024-09-24 16:31:28,637 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=7.46 vs. limit=15.0 2024-09-24 16:31:54,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=542290.0, ans=0.0 2024-09-24 16:32:01,096 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=542290.0, ans=0.0 2024-09-24 16:32:13,585 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=542336.6666666666, ans=0.5 2024-09-24 16:32:15,025 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=542336.6666666666, ans=0.025 2024-09-24 16:32:34,221 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=542383.3333333334, ans=0.125 2024-09-24 16:32:38,559 INFO [train.py:1198] (2/4) Epoch 30, batch 3250, loss[loss=0.173, ctc_loss=0.1098, cr_loss=0.3158, over 16238.00 frames. ], tot_loss[loss=0.1992, ctc_loss=0.1298, cr_loss=0.3468, over 3372583.78 frames. ], batch size: 36, lr: 3.95e-03, grad_scale: 32.0 2024-09-24 16:33:16,568 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.131e+02 1.257e+02 1.346e+02 1.457e+02 3.617e+02, threshold=2.692e+02, percent-clipped=1.0 2024-09-24 16:33:27,709 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=542570.0, ans=0.1 2024-09-24 16:33:59,504 INFO [train.py:1198] (2/4) Epoch 30, batch 3300, loss[loss=0.2316, ctc_loss=0.1555, cr_loss=0.3805, over 16525.00 frames. ], tot_loss[loss=0.1981, ctc_loss=0.1291, cr_loss=0.3453, over 3370814.40 frames. ], batch size: 66, lr: 3.95e-03, grad_scale: 32.0 2024-09-24 16:34:31,075 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=542756.6666666666, ans=0.0 2024-09-24 16:34:36,395 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=4.73 vs. limit=15.0 2024-09-24 16:35:17,843 INFO [train.py:1198] (2/4) Epoch 30, batch 3350, loss[loss=0.1936, ctc_loss=0.1239, cr_loss=0.3485, over 17348.00 frames. ], tot_loss[loss=0.1983, ctc_loss=0.1292, cr_loss=0.3454, over 3366469.08 frames. ], batch size: 48, lr: 3.95e-03, grad_scale: 32.0 2024-09-24 16:35:18,117 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=542896.6666666666, ans=0.0 2024-09-24 16:35:18,844 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=4.04 vs. limit=12.0 2024-09-24 16:35:55,302 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.300e+02 1.382e+02 1.466e+02 2.312e+02, threshold=2.764e+02, percent-clipped=0.0 2024-09-24 16:35:57,107 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=542990.0, ans=0.1 2024-09-24 16:35:57,210 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=542990.0, ans=0.0 2024-09-24 16:36:28,748 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=543083.3333333334, ans=0.125 2024-09-24 16:36:36,354 INFO [train.py:1198] (2/4) Epoch 30, batch 3400, loss[loss=0.1744, ctc_loss=0.1129, cr_loss=0.3077, over 16958.00 frames. ], tot_loss[loss=0.1993, ctc_loss=0.13, cr_loss=0.3468, over 3350697.32 frames. ], batch size: 42, lr: 3.95e-03, grad_scale: 32.0 2024-09-24 16:36:39,763 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=543130.0, ans=0.0 2024-09-24 16:36:44,403 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=543130.0, ans=0.125 2024-09-24 16:36:59,152 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=9.11 vs. limit=15.0 2024-09-24 16:37:27,455 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=543270.0, ans=0.025 2024-09-24 16:37:27,482 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=543270.0, ans=0.125 2024-09-24 16:37:53,726 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=543316.6666666666, ans=0.0 2024-09-24 16:37:56,612 INFO [train.py:1198] (2/4) Epoch 30, batch 3450, loss[loss=0.1873, ctc_loss=0.1198, cr_loss=0.3377, over 17024.00 frames. ], tot_loss[loss=0.1996, ctc_loss=0.1302, cr_loss=0.3469, over 3349457.14 frames. ], batch size: 44, lr: 3.95e-03, grad_scale: 32.0 2024-09-24 16:38:35,837 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.132e+02 1.271e+02 1.366e+02 1.468e+02 2.089e+02, threshold=2.732e+02, percent-clipped=0.0 2024-09-24 16:38:42,545 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=543456.6666666666, ans=0.125 2024-09-24 16:38:50,385 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=543503.3333333334, ans=0.0 2024-09-24 16:38:51,883 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=543503.3333333334, ans=0.1 2024-09-24 16:39:05,323 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=12.76 vs. limit=22.5 2024-09-24 16:39:16,035 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=10.35 vs. limit=15.0 2024-09-24 16:39:16,716 INFO [train.py:1198] (2/4) Epoch 30, batch 3500, loss[loss=0.2038, ctc_loss=0.1344, cr_loss=0.347, over 17088.00 frames. ], tot_loss[loss=0.1997, ctc_loss=0.1302, cr_loss=0.3476, over 3357561.48 frames. ], batch size: 43, lr: 3.95e-03, grad_scale: 32.0 2024-09-24 16:39:51,139 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.50 vs. limit=15.0 2024-09-24 16:40:10,569 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=543736.6666666666, ans=0.0 2024-09-24 16:40:23,010 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=543783.3333333334, ans=0.125 2024-09-24 16:40:34,119 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=543783.3333333334, ans=0.125 2024-09-24 16:40:36,961 INFO [train.py:1198] (2/4) Epoch 30, batch 3550, loss[loss=0.1736, ctc_loss=0.1114, cr_loss=0.3107, over 17086.00 frames. ], tot_loss[loss=0.1993, ctc_loss=0.1299, cr_loss=0.3467, over 3369778.97 frames. ], batch size: 43, lr: 3.95e-03, grad_scale: 32.0 2024-09-24 16:40:41,929 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=543830.0, ans=0.125 2024-09-24 16:40:42,775 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.69 vs. limit=22.5 2024-09-24 16:40:48,709 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=5.99 vs. limit=10.0 2024-09-24 16:40:58,173 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=8.93 vs. limit=15.0 2024-09-24 16:41:04,373 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=15.56 vs. limit=22.5 2024-09-24 16:41:14,200 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.113e+02 1.268e+02 1.349e+02 1.434e+02 2.153e+02, threshold=2.697e+02, percent-clipped=0.0 2024-09-24 16:41:26,105 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=7.51 vs. limit=15.0 2024-09-24 16:41:32,096 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 16:41:43,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=544016.6666666666, ans=0.125 2024-09-24 16:41:44,908 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=14.04 vs. limit=15.0 2024-09-24 16:41:55,341 INFO [train.py:1198] (2/4) Epoch 30, batch 3600, loss[loss=0.2156, ctc_loss=0.1469, cr_loss=0.3436, over 15962.00 frames. ], tot_loss[loss=0.1999, ctc_loss=0.1304, cr_loss=0.3477, over 3372709.00 frames. ], batch size: 74, lr: 3.95e-03, grad_scale: 32.0 2024-09-24 16:42:28,524 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.max_positive, batch_count=544156.6666666666, ans=0.95 2024-09-24 16:42:33,170 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=544156.6666666666, ans=0.125 2024-09-24 16:42:53,349 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.max_abs, batch_count=544203.3333333334, ans=10.0 2024-09-24 16:43:13,172 INFO [train.py:1198] (2/4) Epoch 30, batch 3650, loss[loss=0.1889, ctc_loss=0.121, cr_loss=0.3397, over 17154.00 frames. ], tot_loss[loss=0.1998, ctc_loss=0.1303, cr_loss=0.3475, over 3379019.78 frames. ], batch size: 45, lr: 3.95e-03, grad_scale: 16.0 2024-09-24 16:43:13,531 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.max_abs, batch_count=544296.6666666666, ans=10.0 2024-09-24 16:43:16,631 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=544296.6666666666, ans=0.0 2024-09-24 16:43:16,698 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=544296.6666666666, ans=0.1 2024-09-24 16:43:38,768 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 16:43:54,825 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.034e+02 1.277e+02 1.338e+02 1.462e+02 2.145e+02, threshold=2.677e+02, percent-clipped=0.0 2024-09-24 16:44:17,293 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=544483.3333333334, ans=0.1 2024-09-24 16:44:26,948 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=544483.3333333334, ans=0.2 2024-09-24 16:44:32,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=544483.3333333334, ans=0.1 2024-09-24 16:44:35,114 INFO [train.py:1198] (2/4) Epoch 30, batch 3700, loss[loss=0.1805, ctc_loss=0.1127, cr_loss=0.3389, over 17285.00 frames. ], tot_loss[loss=0.199, ctc_loss=0.1297, cr_loss=0.3465, over 3371708.80 frames. ], batch size: 46, lr: 3.95e-03, grad_scale: 8.0 2024-09-24 16:44:35,816 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.84 vs. limit=6.0 2024-09-24 16:44:45,005 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=544530.0, ans=0.1 2024-09-24 16:45:04,468 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=3.64 vs. limit=15.0 2024-09-24 16:45:27,352 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=544670.0, ans=0.025 2024-09-24 16:45:29,100 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.60 vs. limit=15.0 2024-09-24 16:45:35,414 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.33 vs. limit=15.0 2024-09-24 16:45:53,702 INFO [train.py:1198] (2/4) Epoch 30, batch 3750, loss[loss=0.1921, ctc_loss=0.1251, cr_loss=0.3347, over 17175.00 frames. ], tot_loss[loss=0.1989, ctc_loss=0.1296, cr_loss=0.3462, over 3366492.30 frames. ], batch size: 45, lr: 3.94e-03, grad_scale: 8.0 2024-09-24 16:46:13,031 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 16:46:35,276 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.265e+02 1.357e+02 1.446e+02 2.507e+02, threshold=2.713e+02, percent-clipped=0.0 2024-09-24 16:46:35,603 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=544856.6666666666, ans=0.0 2024-09-24 16:46:43,207 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=544903.3333333334, ans=0.0 2024-09-24 16:46:46,307 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=544903.3333333334, ans=0.1 2024-09-24 16:46:51,706 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=9.61 vs. limit=15.0 2024-09-24 16:47:12,371 INFO [train.py:1198] (2/4) Epoch 30, batch 3800, loss[loss=0.16, ctc_loss=0.09956, cr_loss=0.3023, over 17031.00 frames. ], tot_loss[loss=0.2, ctc_loss=0.1306, cr_loss=0.3472, over 3336994.18 frames. ], batch size: 39, lr: 3.94e-03, grad_scale: 8.0 2024-09-24 16:47:29,167 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=545043.3333333334, ans=0.125 2024-09-24 16:47:36,195 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.66 vs. limit=15.0 2024-09-24 16:48:14,602 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=545183.3333333334, ans=0.025 2024-09-24 16:48:14,644 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=545183.3333333334, ans=0.07 2024-09-24 16:48:31,293 INFO [train.py:1198] (2/4) Epoch 30, batch 3850, loss[loss=0.2481, ctc_loss=0.1651, cr_loss=0.4151, over 15182.00 frames. ], tot_loss[loss=0.2049, ctc_loss=0.1344, cr_loss=0.3523, over 3264450.74 frames. ], batch size: 90, lr: 3.94e-03, grad_scale: 8.0 2024-09-24 16:48:55,514 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=13.81 vs. limit=15.0 2024-09-24 16:49:01,103 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=545323.3333333334, ans=0.125 2024-09-24 16:49:03,930 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=545323.3333333334, ans=0.1 2024-09-24 16:49:11,169 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.380e+02 1.551e+02 1.671e+02 2.623e+02, threshold=3.102e+02, percent-clipped=0.0 2024-09-24 16:49:16,028 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=545370.0, ans=0.0 2024-09-24 16:50:31,769 INFO [train.py:1198] (2/4) Epoch 31, batch 0, loss[loss=0.215, ctc_loss=0.1389, cr_loss=0.3802, over 17050.00 frames. ], tot_loss[loss=0.215, ctc_loss=0.1389, cr_loss=0.3802, over 17050.00 frames. ], batch size: 52, lr: 3.88e-03, grad_scale: 16.0 2024-09-24 16:50:31,769 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 16:50:43,679 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.1.encoder.layers.1.self_attn_weights, attn_weights_entropy = tensor([5.5025, 5.3068, 4.6545, 5.2537], device='cuda:2') 2024-09-24 16:50:45,201 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.3.encoder.layers.2.self_attn_weights, attn_weights_entropy = tensor([2.8168, 3.3066, 3.1250, 3.5162, 3.0582, 3.0191, 3.5702, 3.7849], device='cuda:2') 2024-09-24 16:50:47,103 INFO [train.py:1230] (2/4) Epoch 31, validation: loss=0.03594, ctc_loss=0.03594, cr_loss=9.065e-15, over 944034.00 frames. 2024-09-24 16:50:47,104 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 16:51:34,147 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=545584.6666666666, ans=0.035 2024-09-24 16:51:40,719 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=19.72 vs. limit=22.5 2024-09-24 16:51:45,122 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=545584.6666666666, ans=0.07 2024-09-24 16:51:58,420 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.93 vs. limit=15.0 2024-09-24 16:51:59,550 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=545631.3333333334, ans=0.2 2024-09-24 16:52:08,061 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=6.92 vs. limit=15.0 2024-09-24 16:52:12,148 INFO [train.py:1198] (2/4) Epoch 31, batch 50, loss[loss=0.1803, ctc_loss=0.1198, cr_loss=0.3025, over 17109.00 frames. ], tot_loss[loss=0.1995, ctc_loss=0.1302, cr_loss=0.3462, over 758053.16 frames. ], batch size: 49, lr: 3.88e-03, grad_scale: 16.0 2024-09-24 16:52:24,207 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.47 vs. limit=15.0 2024-09-24 16:52:29,913 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=10.89 vs. limit=15.0 2024-09-24 16:52:31,256 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=545724.6666666666, ans=0.2 2024-09-24 16:52:37,024 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=545724.6666666666, ans=0.5 2024-09-24 16:52:37,088 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=545724.6666666666, ans=0.125 2024-09-24 16:52:43,353 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=545724.6666666666, ans=0.0 2024-09-24 16:52:54,957 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.whiten.whitening_limit, batch_count=545771.3333333334, ans=15.0 2024-09-24 16:53:01,962 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.275e+02 1.374e+02 1.534e+02 1.966e+02, threshold=2.749e+02, percent-clipped=0.0 2024-09-24 16:53:12,353 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=10.13 vs. limit=22.5 2024-09-24 16:53:23,166 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=545864.6666666666, ans=0.125 2024-09-24 16:53:26,241 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=545864.6666666666, ans=0.0 2024-09-24 16:53:33,384 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.77 vs. limit=10.0 2024-09-24 16:53:34,045 INFO [train.py:1198] (2/4) Epoch 31, batch 100, loss[loss=0.2098, ctc_loss=0.1392, cr_loss=0.3532, over 17027.00 frames. ], tot_loss[loss=0.2027, ctc_loss=0.1326, cr_loss=0.3504, over 1324542.81 frames. ], batch size: 53, lr: 3.87e-03, grad_scale: 16.0 2024-09-24 16:54:25,485 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=546051.3333333334, ans=0.05 2024-09-24 16:54:31,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=546051.3333333334, ans=0.0 2024-09-24 16:54:47,773 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.96 vs. limit=15.0 2024-09-24 16:54:53,805 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=546098.0, ans=0.125 2024-09-24 16:54:56,691 INFO [train.py:1198] (2/4) Epoch 31, batch 150, loss[loss=0.1868, ctc_loss=0.1193, cr_loss=0.3376, over 16733.00 frames. ], tot_loss[loss=0.203, ctc_loss=0.1326, cr_loss=0.3516, over 1780100.92 frames. ], batch size: 37, lr: 3.87e-03, grad_scale: 16.0 2024-09-24 16:55:01,902 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=546144.6666666666, ans=0.125 2024-09-24 16:55:13,329 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=546191.3333333334, ans=0.1 2024-09-24 16:55:18,160 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=546191.3333333334, ans=0.125 2024-09-24 16:55:44,680 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.122e+02 1.256e+02 1.336e+02 1.452e+02 1.976e+02, threshold=2.673e+02, percent-clipped=0.0 2024-09-24 16:55:53,024 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=546284.6666666666, ans=0.0 2024-09-24 16:56:16,759 INFO [train.py:1198] (2/4) Epoch 31, batch 200, loss[loss=0.2424, ctc_loss=0.161, cr_loss=0.407, over 17029.00 frames. ], tot_loss[loss=0.2025, ctc_loss=0.132, cr_loss=0.3523, over 2131699.07 frames. ], batch size: 52, lr: 3.87e-03, grad_scale: 16.0 2024-09-24 16:56:44,227 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=546424.6666666666, ans=0.0 2024-09-24 16:57:44,191 INFO [train.py:1198] (2/4) Epoch 31, batch 250, loss[loss=0.1558, ctc_loss=0.09869, cr_loss=0.2853, over 17273.00 frames. ], tot_loss[loss=0.2029, ctc_loss=0.1324, cr_loss=0.3523, over 2391063.54 frames. ], batch size: 42, lr: 3.87e-03, grad_scale: 16.0 2024-09-24 16:58:14,678 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=546704.6666666666, ans=0.09899494936611666 2024-09-24 16:58:31,699 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.093e+02 1.278e+02 1.364e+02 1.509e+02 2.036e+02, threshold=2.729e+02, percent-clipped=0.0 2024-09-24 16:58:41,666 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=546751.3333333334, ans=0.125 2024-09-24 16:58:49,495 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer_ff3.min_abs, batch_count=546798.0, ans=0.2 2024-09-24 16:59:03,342 INFO [train.py:1198] (2/4) Epoch 31, batch 300, loss[loss=0.2254, ctc_loss=0.1474, cr_loss=0.3899, over 17146.00 frames. ], tot_loss[loss=0.2014, ctc_loss=0.1313, cr_loss=0.3507, over 2612415.23 frames. ], batch size: 48, lr: 3.87e-03, grad_scale: 16.0 2024-09-24 16:59:37,050 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=546938.0, ans=0.125 2024-09-24 17:00:24,748 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=547078.0, ans=0.5 2024-09-24 17:00:26,012 INFO [train.py:1198] (2/4) Epoch 31, batch 350, loss[loss=0.1908, ctc_loss=0.1218, cr_loss=0.345, over 17145.00 frames. ], tot_loss[loss=0.2014, ctc_loss=0.1314, cr_loss=0.3496, over 2762297.78 frames. ], batch size: 48, lr: 3.87e-03, grad_scale: 16.0 2024-09-24 17:00:31,200 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-24 17:00:39,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=547078.0, ans=0.2 2024-09-24 17:00:42,394 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=547124.6666666666, ans=0.125 2024-09-24 17:00:45,728 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=547124.6666666666, ans=0.2 2024-09-24 17:01:01,636 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=547171.3333333334, ans=0.125 2024-09-24 17:01:03,293 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=547171.3333333334, ans=0.125 2024-09-24 17:01:14,093 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.126e+02 1.224e+02 1.318e+02 1.423e+02 1.795e+02, threshold=2.635e+02, percent-clipped=0.0 2024-09-24 17:01:28,417 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=547218.0, ans=0.125 2024-09-24 17:01:51,998 INFO [train.py:1198] (2/4) Epoch 31, batch 400, loss[loss=0.215, ctc_loss=0.1374, cr_loss=0.388, over 17104.00 frames. ], tot_loss[loss=0.2011, ctc_loss=0.1312, cr_loss=0.3493, over 2892455.97 frames. ], batch size: 49, lr: 3.87e-03, grad_scale: 32.0 2024-09-24 17:01:53,788 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=547311.3333333334, ans=0.125 2024-09-24 17:01:53,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=547311.3333333334, ans=0.05 2024-09-24 17:01:55,506 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=547311.3333333334, ans=10.0 2024-09-24 17:02:03,364 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=547311.3333333334, ans=0.125 2024-09-24 17:02:09,614 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=547358.0, ans=0.125 2024-09-24 17:02:58,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=547498.0, ans=0.0 2024-09-24 17:03:05,287 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=547498.0, ans=0.125 2024-09-24 17:03:14,667 INFO [train.py:1198] (2/4) Epoch 31, batch 450, loss[loss=0.2005, ctc_loss=0.1309, cr_loss=0.348, over 17000.00 frames. ], tot_loss[loss=0.2012, ctc_loss=0.1313, cr_loss=0.3496, over 3005129.71 frames. ], batch size: 39, lr: 3.87e-03, grad_scale: 32.0 2024-09-24 17:03:26,364 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=547544.6666666666, ans=0.125 2024-09-24 17:03:33,650 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.72 vs. limit=22.5 2024-09-24 17:03:34,647 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=547591.3333333334, ans=0.0 2024-09-24 17:03:58,931 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=9.44 vs. limit=15.0 2024-09-24 17:04:02,909 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.095e+02 1.314e+02 1.420e+02 1.522e+02 1.945e+02, threshold=2.840e+02, percent-clipped=0.0 2024-09-24 17:04:04,909 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.min_positive, batch_count=547684.6666666666, ans=0.05 2024-09-24 17:04:06,861 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.82 vs. limit=6.0 2024-09-24 17:04:15,786 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=547684.6666666666, ans=0.0 2024-09-24 17:04:17,429 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=547731.3333333334, ans=0.0 2024-09-24 17:04:22,133 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=547731.3333333334, ans=0.125 2024-09-24 17:04:23,683 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=547731.3333333334, ans=0.125 2024-09-24 17:04:28,061 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=547731.3333333334, ans=0.0 2024-09-24 17:04:37,650 INFO [train.py:1198] (2/4) Epoch 31, batch 500, loss[loss=0.188, ctc_loss=0.1225, cr_loss=0.3276, over 16969.00 frames. ], tot_loss[loss=0.2002, ctc_loss=0.1305, cr_loss=0.3486, over 3090120.06 frames. ], batch size: 42, lr: 3.87e-03, grad_scale: 32.0 2024-09-24 17:05:22,720 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.19 vs. limit=6.0 2024-09-24 17:05:47,918 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=547964.6666666666, ans=0.125 2024-09-24 17:05:58,745 INFO [train.py:1198] (2/4) Epoch 31, batch 550, loss[loss=0.1893, ctc_loss=0.1252, cr_loss=0.3205, over 17184.00 frames. ], tot_loss[loss=0.201, ctc_loss=0.1311, cr_loss=0.3494, over 3148305.56 frames. ], batch size: 45, lr: 3.87e-03, grad_scale: 32.0 2024-09-24 17:06:14,945 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=548058.0, ans=0.025 2024-09-24 17:06:43,659 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.94 vs. limit=15.0 2024-09-24 17:06:44,714 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=548104.6666666666, ans=0.125 2024-09-24 17:06:47,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=548104.6666666666, ans=0.125 2024-09-24 17:06:48,507 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.48 vs. limit=15.0 2024-09-24 17:06:52,449 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.269e+02 1.346e+02 1.471e+02 2.079e+02, threshold=2.692e+02, percent-clipped=0.0 2024-09-24 17:07:02,168 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=548151.3333333334, ans=0.0 2024-09-24 17:07:04,148 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=4.33 vs. limit=15.0 2024-09-24 17:07:08,413 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=548198.0, ans=0.2 2024-09-24 17:07:08,493 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=548198.0, ans=0.125 2024-09-24 17:07:15,156 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys.whitening_limit, batch_count=548198.0, ans=6.0 2024-09-24 17:07:16,650 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=13.39 vs. limit=15.0 2024-09-24 17:07:23,919 INFO [train.py:1198] (2/4) Epoch 31, batch 600, loss[loss=0.2123, ctc_loss=0.1422, cr_loss=0.3506, over 16898.00 frames. ], tot_loss[loss=0.2006, ctc_loss=0.1308, cr_loss=0.349, over 3202423.45 frames. ], batch size: 58, lr: 3.87e-03, grad_scale: 32.0 2024-09-24 17:07:58,480 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=548338.0, ans=0.0 2024-09-24 17:08:01,599 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=548338.0, ans=0.0 2024-09-24 17:08:09,860 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=548338.0, ans=0.125 2024-09-24 17:08:30,752 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=548431.3333333334, ans=0.1 2024-09-24 17:08:38,739 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=548431.3333333334, ans=0.125 2024-09-24 17:08:38,987 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.22 vs. limit=22.5 2024-09-24 17:08:46,119 INFO [train.py:1198] (2/4) Epoch 31, batch 650, loss[loss=0.2059, ctc_loss=0.1349, cr_loss=0.3547, over 15961.00 frames. ], tot_loss[loss=0.1995, ctc_loss=0.13, cr_loss=0.3475, over 3243309.52 frames. ], batch size: 74, lr: 3.87e-03, grad_scale: 32.0 2024-09-24 17:08:51,327 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=548478.0, ans=0.05 2024-09-24 17:09:19,949 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=548571.3333333334, ans=0.015 2024-09-24 17:09:30,724 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=548571.3333333334, ans=0.0 2024-09-24 17:09:35,555 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=548618.0, ans=0.125 2024-09-24 17:09:36,765 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.059e+02 1.261e+02 1.352e+02 1.465e+02 2.749e+02, threshold=2.705e+02, percent-clipped=1.0 2024-09-24 17:09:56,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=548664.6666666666, ans=0.125 2024-09-24 17:10:08,668 INFO [train.py:1198] (2/4) Epoch 31, batch 700, loss[loss=0.1897, ctc_loss=0.1233, cr_loss=0.332, over 17301.00 frames. ], tot_loss[loss=0.199, ctc_loss=0.1297, cr_loss=0.347, over 3258550.38 frames. ], batch size: 51, lr: 3.86e-03, grad_scale: 32.0 2024-09-24 17:10:23,331 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=548758.0, ans=0.125 2024-09-24 17:10:51,252 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=7.18 vs. limit=12.0 2024-09-24 17:11:31,808 INFO [train.py:1198] (2/4) Epoch 31, batch 750, loss[loss=0.1657, ctc_loss=0.1073, cr_loss=0.292, over 16326.00 frames. ], tot_loss[loss=0.1992, ctc_loss=0.1298, cr_loss=0.3472, over 3276285.05 frames. ], batch size: 36, lr: 3.86e-03, grad_scale: 32.0 2024-09-24 17:11:52,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=548991.3333333334, ans=0.015 2024-09-24 17:12:08,457 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=549038.0, ans=0.0 2024-09-24 17:12:08,574 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=549038.0, ans=0.0 2024-09-24 17:12:25,357 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.062e+02 1.270e+02 1.339e+02 1.427e+02 2.075e+02, threshold=2.677e+02, percent-clipped=0.0 2024-09-24 17:12:35,382 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=549084.6666666666, ans=0.0 2024-09-24 17:12:57,347 INFO [train.py:1198] (2/4) Epoch 31, batch 800, loss[loss=0.1848, ctc_loss=0.1198, cr_loss=0.3254, over 17027.00 frames. ], tot_loss[loss=0.1991, ctc_loss=0.1297, cr_loss=0.347, over 3290948.62 frames. ], batch size: 44, lr: 3.86e-03, grad_scale: 32.0 2024-09-24 17:13:27,985 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=549271.3333333334, ans=0.1 2024-09-24 17:14:09,564 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=549364.6666666666, ans=0.04949747468305833 2024-09-24 17:14:17,350 INFO [train.py:1198] (2/4) Epoch 31, batch 850, loss[loss=0.1834, ctc_loss=0.1186, cr_loss=0.3242, over 17129.00 frames. ], tot_loss[loss=0.199, ctc_loss=0.1296, cr_loss=0.3473, over 3305464.40 frames. ], batch size: 48, lr: 3.86e-03, grad_scale: 32.0 2024-09-24 17:14:25,805 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=549411.3333333334, ans=0.125 2024-09-24 17:14:52,260 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=549504.6666666666, ans=0.1 2024-09-24 17:15:09,431 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.115e+02 1.257e+02 1.348e+02 1.447e+02 2.367e+02, threshold=2.696e+02, percent-clipped=0.0 2024-09-24 17:15:11,218 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=549551.3333333334, ans=0.2 2024-09-24 17:15:27,600 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.52 vs. limit=6.0 2024-09-24 17:15:33,508 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=549598.0, ans=0.2 2024-09-24 17:15:38,237 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=549644.6666666666, ans=0.0 2024-09-24 17:15:39,449 INFO [train.py:1198] (2/4) Epoch 31, batch 900, loss[loss=0.229, ctc_loss=0.1488, cr_loss=0.4011, over 16994.00 frames. ], tot_loss[loss=0.1995, ctc_loss=0.13, cr_loss=0.3476, over 3318634.25 frames. ], batch size: 56, lr: 3.86e-03, grad_scale: 16.0 2024-09-24 17:16:21,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=549738.0, ans=0.125 2024-09-24 17:16:47,822 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=549831.3333333334, ans=0.125 2024-09-24 17:17:05,204 INFO [train.py:1198] (2/4) Epoch 31, batch 950, loss[loss=0.2125, ctc_loss=0.1405, cr_loss=0.3601, over 16101.00 frames. ], tot_loss[loss=0.1992, ctc_loss=0.1298, cr_loss=0.3472, over 3326283.32 frames. ], batch size: 74, lr: 3.86e-03, grad_scale: 16.0 2024-09-24 17:17:36,777 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=549924.6666666666, ans=0.125 2024-09-24 17:17:46,174 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=549971.3333333334, ans=0.125 2024-09-24 17:17:46,603 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.25 vs. limit=15.0 2024-09-24 17:17:57,027 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.099e+02 1.281e+02 1.359e+02 1.441e+02 1.895e+02, threshold=2.717e+02, percent-clipped=0.0 2024-09-24 17:18:03,713 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=550018.0, ans=0.1 2024-09-24 17:18:04,435 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.62 vs. limit=10.0 2024-09-24 17:18:27,163 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=8.05 vs. limit=15.0 2024-09-24 17:18:27,476 INFO [train.py:1198] (2/4) Epoch 31, batch 1000, loss[loss=0.1948, ctc_loss=0.1275, cr_loss=0.3366, over 16737.00 frames. ], tot_loss[loss=0.1984, ctc_loss=0.1292, cr_loss=0.346, over 3338028.13 frames. ], batch size: 61, lr: 3.86e-03, grad_scale: 16.0 2024-09-24 17:19:04,948 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.98 vs. limit=22.5 2024-09-24 17:19:50,155 INFO [train.py:1198] (2/4) Epoch 31, batch 1050, loss[loss=0.2217, ctc_loss=0.1475, cr_loss=0.3711, over 16589.00 frames. ], tot_loss[loss=0.1984, ctc_loss=0.1292, cr_loss=0.346, over 3343864.63 frames. ], batch size: 66, lr: 3.86e-03, grad_scale: 16.0 2024-09-24 17:19:55,201 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=550344.6666666666, ans=0.125 2024-09-24 17:19:56,746 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=550344.6666666666, ans=0.2 2024-09-24 17:20:03,709 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=10.14 vs. limit=15.0 2024-09-24 17:20:24,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=550438.0, ans=0.1 2024-09-24 17:20:39,885 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.258e+02 1.325e+02 1.421e+02 1.846e+02, threshold=2.651e+02, percent-clipped=0.0 2024-09-24 17:21:10,255 INFO [train.py:1198] (2/4) Epoch 31, batch 1100, loss[loss=0.2192, ctc_loss=0.1458, cr_loss=0.367, over 17094.00 frames. ], tot_loss[loss=0.1984, ctc_loss=0.1292, cr_loss=0.3461, over 3347244.25 frames. ], batch size: 49, lr: 3.86e-03, grad_scale: 16.0 2024-09-24 17:21:12,144 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.max_positive, batch_count=550578.0, ans=0.95 2024-09-24 17:21:13,794 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=550578.0, ans=0.125 2024-09-24 17:21:20,124 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=550578.0, ans=0.0 2024-09-24 17:21:58,463 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 17:22:19,150 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=550764.6666666666, ans=0.0 2024-09-24 17:22:23,215 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=550764.6666666666, ans=0.0 2024-09-24 17:22:28,032 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=550764.6666666666, ans=0.0 2024-09-24 17:22:39,182 INFO [train.py:1198] (2/4) Epoch 31, batch 1150, loss[loss=0.2067, ctc_loss=0.1339, cr_loss=0.3636, over 17157.00 frames. ], tot_loss[loss=0.1987, ctc_loss=0.1295, cr_loss=0.3461, over 3343230.54 frames. ], batch size: 45, lr: 3.86e-03, grad_scale: 16.0 2024-09-24 17:22:52,258 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=550811.3333333334, ans=0.2 2024-09-24 17:22:55,319 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=550858.0, ans=0.1 2024-09-24 17:22:56,935 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=550858.0, ans=0.0 2024-09-24 17:23:03,287 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=550858.0, ans=0.0 2024-09-24 17:23:28,709 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.121e+02 1.270e+02 1.354e+02 1.493e+02 2.055e+02, threshold=2.708e+02, percent-clipped=0.0 2024-09-24 17:23:51,620 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=550998.0, ans=0.125 2024-09-24 17:23:59,053 INFO [train.py:1198] (2/4) Epoch 31, batch 1200, loss[loss=0.2036, ctc_loss=0.1315, cr_loss=0.3602, over 17061.00 frames. ], tot_loss[loss=0.2002, ctc_loss=0.1305, cr_loss=0.3483, over 3344984.11 frames. ], batch size: 46, lr: 3.86e-03, grad_scale: 32.0 2024-09-24 17:24:10,530 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=551044.6666666666, ans=0.125 2024-09-24 17:24:28,750 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=13.70 vs. limit=22.5 2024-09-24 17:24:32,625 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=551138.0, ans=0.0 2024-09-24 17:24:50,264 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=551184.6666666666, ans=0.0 2024-09-24 17:25:12,818 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=551231.3333333334, ans=0.125 2024-09-24 17:25:20,598 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=551278.0, ans=0.125 2024-09-24 17:25:21,965 INFO [train.py:1198] (2/4) Epoch 31, batch 1250, loss[loss=0.1637, ctc_loss=0.1035, cr_loss=0.3011, over 16956.00 frames. ], tot_loss[loss=0.1992, ctc_loss=0.1299, cr_loss=0.3466, over 3350189.69 frames. ], batch size: 42, lr: 3.86e-03, grad_scale: 16.0 2024-09-24 17:25:56,172 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=13.37 vs. limit=15.0 2024-09-24 17:25:57,682 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.whiten.whitening_limit, batch_count=551371.3333333334, ans=12.0 2024-09-24 17:26:13,265 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.139e+02 1.296e+02 1.398e+02 1.491e+02 2.236e+02, threshold=2.796e+02, percent-clipped=0.0 2024-09-24 17:26:42,468 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=551464.6666666666, ans=0.0 2024-09-24 17:26:45,572 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=551511.3333333334, ans=0.015 2024-09-24 17:26:46,959 INFO [train.py:1198] (2/4) Epoch 31, batch 1300, loss[loss=0.2521, ctc_loss=0.1703, cr_loss=0.409, over 14804.00 frames. ], tot_loss[loss=0.199, ctc_loss=0.1298, cr_loss=0.3462, over 3351135.46 frames. ], batch size: 88, lr: 3.86e-03, grad_scale: 16.0 2024-09-24 17:27:47,987 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=551651.3333333334, ans=0.125 2024-09-24 17:27:55,965 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.94 vs. limit=10.0 2024-09-24 17:28:09,624 INFO [train.py:1198] (2/4) Epoch 31, batch 1350, loss[loss=0.1981, ctc_loss=0.1291, cr_loss=0.3451, over 17137.00 frames. ], tot_loss[loss=0.2004, ctc_loss=0.1309, cr_loss=0.3476, over 3331454.68 frames. ], batch size: 48, lr: 3.85e-03, grad_scale: 16.0 2024-09-24 17:28:11,481 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=551744.6666666666, ans=0.125 2024-09-24 17:28:32,287 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=551791.3333333334, ans=0.0 2024-09-24 17:28:41,807 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=551838.0, ans=0.125 2024-09-24 17:29:00,715 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.265e+02 1.360e+02 1.447e+02 1.964e+02, threshold=2.719e+02, percent-clipped=0.0 2024-09-24 17:29:08,591 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=7.01 vs. limit=8.0 2024-09-24 17:29:32,591 INFO [train.py:1198] (2/4) Epoch 31, batch 1400, loss[loss=0.228, ctc_loss=0.1494, cr_loss=0.3929, over 17009.00 frames. ], tot_loss[loss=0.2, ctc_loss=0.1306, cr_loss=0.3473, over 3343423.64 frames. ], batch size: 53, lr: 3.85e-03, grad_scale: 16.0 2024-09-24 17:29:50,766 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=552024.6666666666, ans=0.125 2024-09-24 17:29:58,600 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=552024.6666666666, ans=0.125 2024-09-24 17:30:01,882 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=552024.6666666666, ans=0.125 2024-09-24 17:30:09,650 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=552071.3333333334, ans=0.1 2024-09-24 17:30:38,586 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=552164.6666666666, ans=0.0 2024-09-24 17:30:50,049 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=552164.6666666666, ans=0.04949747468305833 2024-09-24 17:30:52,825 INFO [train.py:1198] (2/4) Epoch 31, batch 1450, loss[loss=0.1828, ctc_loss=0.1193, cr_loss=0.318, over 17119.00 frames. ], tot_loss[loss=0.1988, ctc_loss=0.1296, cr_loss=0.3458, over 3358861.16 frames. ], batch size: 40, lr: 3.85e-03, grad_scale: 16.0 2024-09-24 17:30:56,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=552211.3333333334, ans=0.125 2024-09-24 17:31:04,550 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=552211.3333333334, ans=0.125 2024-09-24 17:31:45,919 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=552351.3333333334, ans=0.0 2024-09-24 17:31:48,822 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.089e+02 1.229e+02 1.325e+02 1.469e+02 3.189e+02, threshold=2.649e+02, percent-clipped=1.0 2024-09-24 17:32:19,916 INFO [train.py:1198] (2/4) Epoch 31, batch 1500, loss[loss=0.2056, ctc_loss=0.1331, cr_loss=0.3624, over 16695.00 frames. ], tot_loss[loss=0.1976, ctc_loss=0.1288, cr_loss=0.3443, over 3362869.16 frames. ], batch size: 61, lr: 3.85e-03, grad_scale: 16.0 2024-09-24 17:32:32,876 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=552444.6666666666, ans=0.0 2024-09-24 17:32:58,307 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=552538.0, ans=0.0 2024-09-24 17:33:28,543 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=552631.3333333334, ans=0.125 2024-09-24 17:33:33,552 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=552631.3333333334, ans=0.125 2024-09-24 17:33:39,657 INFO [train.py:1198] (2/4) Epoch 31, batch 1550, loss[loss=0.2085, ctc_loss=0.1385, cr_loss=0.3498, over 16995.00 frames. ], tot_loss[loss=0.1981, ctc_loss=0.1291, cr_loss=0.3454, over 3360789.56 frames. ], batch size: 56, lr: 3.85e-03, grad_scale: 16.0 2024-09-24 17:33:41,509 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=552678.0, ans=0.2 2024-09-24 17:33:53,484 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=10.93 vs. limit=15.0 2024-09-24 17:34:05,591 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=552724.6666666666, ans=0.0 2024-09-24 17:34:33,295 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.274e+02 1.344e+02 1.453e+02 2.165e+02, threshold=2.688e+02, percent-clipped=0.0 2024-09-24 17:34:35,162 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=552818.0, ans=0.125 2024-09-24 17:34:45,020 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=552864.6666666666, ans=0.1 2024-09-24 17:34:57,677 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=552864.6666666666, ans=0.035 2024-09-24 17:35:02,238 INFO [train.py:1198] (2/4) Epoch 31, batch 1600, loss[loss=0.1908, ctc_loss=0.1254, cr_loss=0.3272, over 17152.00 frames. ], tot_loss[loss=0.199, ctc_loss=0.1297, cr_loss=0.3463, over 3351062.30 frames. ], batch size: 48, lr: 3.85e-03, grad_scale: 32.0 2024-09-24 17:35:10,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=552911.3333333334, ans=0.0 2024-09-24 17:35:20,211 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=552958.0, ans=0.125 2024-09-24 17:36:15,910 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=553098.0, ans=0.0 2024-09-24 17:36:25,120 INFO [train.py:1198] (2/4) Epoch 31, batch 1650, loss[loss=0.1727, ctc_loss=0.1094, cr_loss=0.3166, over 17059.00 frames. ], tot_loss[loss=0.199, ctc_loss=0.1298, cr_loss=0.3461, over 3353561.00 frames. ], batch size: 46, lr: 3.85e-03, grad_scale: 32.0 2024-09-24 17:37:00,351 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.14 vs. limit=6.0 2024-09-24 17:37:02,012 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.04 vs. limit=15.0 2024-09-24 17:37:20,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.min_abs, batch_count=553284.6666666666, ans=0.5 2024-09-24 17:37:21,451 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.162e+02 1.255e+02 1.333e+02 1.430e+02 2.111e+02, threshold=2.665e+02, percent-clipped=0.0 2024-09-24 17:37:39,505 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=553331.3333333334, ans=0.125 2024-09-24 17:37:39,514 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.min_positive, batch_count=553331.3333333334, ans=0.025 2024-09-24 17:37:50,371 INFO [train.py:1198] (2/4) Epoch 31, batch 1700, loss[loss=0.2139, ctc_loss=0.1425, cr_loss=0.3573, over 17012.00 frames. ], tot_loss[loss=0.1996, ctc_loss=0.1302, cr_loss=0.3474, over 3360444.89 frames. ], batch size: 51, lr: 3.85e-03, grad_scale: 32.0 2024-09-24 17:37:56,430 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=8.02 vs. limit=15.0 2024-09-24 17:37:58,902 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=553378.0, ans=0.0 2024-09-24 17:38:08,627 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=553424.6666666666, ans=0.125 2024-09-24 17:38:10,312 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=553424.6666666666, ans=0.0 2024-09-24 17:38:13,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=553424.6666666666, ans=0.0 2024-09-24 17:38:25,712 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.63 vs. limit=6.0 2024-09-24 17:38:39,328 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=553518.0, ans=0.125 2024-09-24 17:38:48,910 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=553518.0, ans=0.125 2024-09-24 17:38:53,893 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=553564.6666666666, ans=0.125 2024-09-24 17:38:54,045 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=553564.6666666666, ans=0.0 2024-09-24 17:39:03,642 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=553564.6666666666, ans=0.0 2024-09-24 17:39:13,913 INFO [train.py:1198] (2/4) Epoch 31, batch 1750, loss[loss=0.1745, ctc_loss=0.1125, cr_loss=0.31, over 16973.00 frames. ], tot_loss[loss=0.2008, ctc_loss=0.1311, cr_loss=0.3485, over 3331423.96 frames. ], batch size: 42, lr: 3.85e-03, grad_scale: 32.0 2024-09-24 17:39:30,045 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=553658.0, ans=0.0 2024-09-24 17:39:50,744 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=553704.6666666666, ans=0.0 2024-09-24 17:40:04,937 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.152e+02 1.305e+02 1.399e+02 1.556e+02 3.053e+02, threshold=2.798e+02, percent-clipped=1.0 2024-09-24 17:40:11,706 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=553751.3333333334, ans=0.0 2024-09-24 17:40:22,804 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=553798.0, ans=0.125 2024-09-24 17:40:27,646 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=553798.0, ans=0.5 2024-09-24 17:40:33,711 INFO [train.py:1198] (2/4) Epoch 31, batch 1800, loss[loss=0.1924, ctc_loss=0.1242, cr_loss=0.341, over 17356.00 frames. ], tot_loss[loss=0.2013, ctc_loss=0.1315, cr_loss=0.3491, over 3324002.94 frames. ], batch size: 48, lr: 3.85e-03, grad_scale: 16.0 2024-09-24 17:40:43,525 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=553844.6666666666, ans=0.0 2024-09-24 17:40:50,076 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.min_positive, batch_count=553891.3333333334, ans=0.05 2024-09-24 17:40:50,210 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 17:40:51,555 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=553891.3333333334, ans=0.1 2024-09-24 17:40:56,133 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=553891.3333333334, ans=0.125 2024-09-24 17:41:38,192 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.29 vs. limit=10.0 2024-09-24 17:42:01,013 INFO [train.py:1198] (2/4) Epoch 31, batch 1850, loss[loss=0.2114, ctc_loss=0.1366, cr_loss=0.3741, over 17043.00 frames. ], tot_loss[loss=0.1992, ctc_loss=0.1298, cr_loss=0.3468, over 3334660.94 frames. ], batch size: 46, lr: 3.85e-03, grad_scale: 16.0 2024-09-24 17:42:07,685 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=554078.0, ans=0.035 2024-09-24 17:42:28,390 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=554124.6666666666, ans=0.05 2024-09-24 17:42:47,527 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=554218.0, ans=0.125 2024-09-24 17:42:53,413 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.087e+02 1.284e+02 1.348e+02 1.466e+02 2.241e+02, threshold=2.697e+02, percent-clipped=0.0 2024-09-24 17:42:56,887 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=554218.0, ans=0.2 2024-09-24 17:43:13,070 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=554264.6666666666, ans=0.125 2024-09-24 17:43:13,095 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=554264.6666666666, ans=0.0 2024-09-24 17:43:20,557 INFO [train.py:1198] (2/4) Epoch 31, batch 1900, loss[loss=0.1976, ctc_loss=0.1314, cr_loss=0.3309, over 17047.00 frames. ], tot_loss[loss=0.2002, ctc_loss=0.1306, cr_loss=0.3483, over 3336100.08 frames. ], batch size: 51, lr: 3.85e-03, grad_scale: 16.0 2024-09-24 17:43:21,833 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=10.20 vs. limit=15.0 2024-09-24 17:44:05,116 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=554404.6666666666, ans=0.025 2024-09-24 17:44:24,298 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=554451.3333333334, ans=0.1 2024-09-24 17:44:42,947 INFO [train.py:1198] (2/4) Epoch 31, batch 1950, loss[loss=0.1674, ctc_loss=0.1062, cr_loss=0.3063, over 17119.00 frames. ], tot_loss[loss=0.1997, ctc_loss=0.1303, cr_loss=0.3474, over 3334276.65 frames. ], batch size: 40, lr: 3.84e-03, grad_scale: 16.0 2024-09-24 17:44:43,614 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.07 vs. limit=6.0 2024-09-24 17:44:49,809 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=554544.6666666666, ans=0.125 2024-09-24 17:44:56,760 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.39 vs. limit=10.0 2024-09-24 17:45:20,284 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=554638.0, ans=0.2 2024-09-24 17:45:20,330 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=554638.0, ans=0.0 2024-09-24 17:45:35,729 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.109e+02 1.290e+02 1.379e+02 1.473e+02 2.110e+02, threshold=2.758e+02, percent-clipped=0.0 2024-09-24 17:45:45,666 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=554731.3333333334, ans=0.125 2024-09-24 17:45:56,324 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=554731.3333333334, ans=0.1 2024-09-24 17:46:05,598 INFO [train.py:1198] (2/4) Epoch 31, batch 2000, loss[loss=0.1959, ctc_loss=0.1264, cr_loss=0.3474, over 17027.00 frames. ], tot_loss[loss=0.2014, ctc_loss=0.1315, cr_loss=0.3496, over 3339434.69 frames. ], batch size: 52, lr: 3.84e-03, grad_scale: 32.0 2024-09-24 17:46:13,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=554778.0, ans=0.0 2024-09-24 17:46:18,902 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=554778.0, ans=0.125 2024-09-24 17:46:31,094 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=554824.6666666666, ans=0.125 2024-09-24 17:46:31,104 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=554824.6666666666, ans=0.035 2024-09-24 17:46:32,685 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=554824.6666666666, ans=0.0 2024-09-24 17:46:39,716 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.91 vs. limit=6.0 2024-09-24 17:47:05,554 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=554918.0, ans=0.1 2024-09-24 17:47:05,688 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=554918.0, ans=0.0 2024-09-24 17:47:15,311 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=554964.6666666666, ans=0.125 2024-09-24 17:47:31,370 INFO [train.py:1198] (2/4) Epoch 31, batch 2050, loss[loss=0.2124, ctc_loss=0.1381, cr_loss=0.3718, over 16898.00 frames. ], tot_loss[loss=0.2013, ctc_loss=0.1313, cr_loss=0.3499, over 3347479.73 frames. ], batch size: 58, lr: 3.84e-03, grad_scale: 32.0 2024-09-24 17:47:57,338 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=555058.0, ans=0.0 2024-09-24 17:48:06,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=555104.6666666666, ans=0.035 2024-09-24 17:48:24,148 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.069e+02 1.242e+02 1.319e+02 1.408e+02 1.614e+02, threshold=2.639e+02, percent-clipped=0.0 2024-09-24 17:48:45,285 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=555198.0, ans=0.125 2024-09-24 17:48:50,166 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=555244.6666666666, ans=0.04949747468305833 2024-09-24 17:48:51,333 INFO [train.py:1198] (2/4) Epoch 31, batch 2100, loss[loss=0.2367, ctc_loss=0.1566, cr_loss=0.4004, over 15134.00 frames. ], tot_loss[loss=0.2007, ctc_loss=0.1308, cr_loss=0.3494, over 3349465.64 frames. ], batch size: 89, lr: 3.84e-03, grad_scale: 32.0 2024-09-24 17:48:57,352 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=555244.6666666666, ans=0.0 2024-09-24 17:48:59,057 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=555244.6666666666, ans=0.0 2024-09-24 17:49:03,805 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=555244.6666666666, ans=0.0 2024-09-24 17:50:10,336 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=555431.3333333334, ans=0.125 2024-09-24 17:50:14,773 INFO [train.py:1198] (2/4) Epoch 31, batch 2150, loss[loss=0.2171, ctc_loss=0.141, cr_loss=0.3805, over 17366.00 frames. ], tot_loss[loss=0.1997, ctc_loss=0.1302, cr_loss=0.3477, over 3350868.37 frames. ], batch size: 48, lr: 3.84e-03, grad_scale: 32.0 2024-09-24 17:50:42,610 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=10.35 vs. limit=22.5 2024-09-24 17:50:55,158 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.45 vs. limit=15.0 2024-09-24 17:51:10,125 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.047e+02 1.261e+02 1.344e+02 1.456e+02 2.277e+02, threshold=2.688e+02, percent-clipped=0.0 2024-09-24 17:51:10,459 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=555618.0, ans=0.0 2024-09-24 17:51:33,717 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=555664.6666666666, ans=0.0 2024-09-24 17:51:39,714 INFO [train.py:1198] (2/4) Epoch 31, batch 2200, loss[loss=0.249, ctc_loss=0.1727, cr_loss=0.3816, over 12078.00 frames. ], tot_loss[loss=0.1994, ctc_loss=0.1299, cr_loss=0.3478, over 3355915.62 frames. ], batch size: 123, lr: 3.84e-03, grad_scale: 32.0 2024-09-24 17:52:09,908 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=7.75 vs. limit=15.0 2024-09-24 17:52:27,343 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.13 vs. limit=6.0 2024-09-24 17:53:02,094 INFO [train.py:1198] (2/4) Epoch 31, batch 2250, loss[loss=0.1609, ctc_loss=0.1037, cr_loss=0.2859, over 17018.00 frames. ], tot_loss[loss=0.1994, ctc_loss=0.1299, cr_loss=0.3477, over 3355782.95 frames. ], batch size: 44, lr: 3.84e-03, grad_scale: 32.0 2024-09-24 17:53:37,397 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.max_abs, batch_count=556038.0, ans=10.0 2024-09-24 17:53:57,559 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.055e+02 1.242e+02 1.333e+02 1.428e+02 2.130e+02, threshold=2.666e+02, percent-clipped=0.0 2024-09-24 17:54:24,672 INFO [train.py:1198] (2/4) Epoch 31, batch 2300, loss[loss=0.2207, ctc_loss=0.147, cr_loss=0.3685, over 16677.00 frames. ], tot_loss[loss=0.1992, ctc_loss=0.1298, cr_loss=0.347, over 3346606.66 frames. ], batch size: 61, lr: 3.84e-03, grad_scale: 32.0 2024-09-24 17:55:32,299 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.min_abs, batch_count=556364.6666666666, ans=0.5 2024-09-24 17:55:47,357 INFO [train.py:1198] (2/4) Epoch 31, batch 2350, loss[loss=0.1822, ctc_loss=0.1145, cr_loss=0.3388, over 17026.00 frames. ], tot_loss[loss=0.1984, ctc_loss=0.1292, cr_loss=0.3459, over 3348455.75 frames. ], batch size: 39, lr: 3.84e-03, grad_scale: 32.0 2024-09-24 17:55:57,377 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=556411.3333333334, ans=0.125 2024-09-24 17:56:03,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=556458.0, ans=0.1 2024-09-24 17:56:43,026 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.265e+02 1.349e+02 1.461e+02 1.759e+02, threshold=2.698e+02, percent-clipped=0.0 2024-09-24 17:57:12,805 INFO [train.py:1198] (2/4) Epoch 31, batch 2400, loss[loss=0.239, ctc_loss=0.1585, cr_loss=0.4023, over 17026.00 frames. ], tot_loss[loss=0.1981, ctc_loss=0.129, cr_loss=0.3457, over 3352625.92 frames. ], batch size: 52, lr: 3.84e-03, grad_scale: 32.0 2024-09-24 17:57:49,155 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.34 vs. limit=15.0 2024-09-24 17:57:58,091 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=556738.0, ans=0.2 2024-09-24 17:58:07,650 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=556784.6666666666, ans=0.2 2024-09-24 17:58:23,632 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=556831.3333333334, ans=0.125 2024-09-24 17:58:32,915 INFO [train.py:1198] (2/4) Epoch 31, batch 2450, loss[loss=0.1999, ctc_loss=0.1311, cr_loss=0.344, over 17357.00 frames. ], tot_loss[loss=0.1986, ctc_loss=0.1293, cr_loss=0.3464, over 3354289.63 frames. ], batch size: 48, lr: 3.84e-03, grad_scale: 32.0 2024-09-24 17:59:14,567 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=556971.3333333334, ans=0.125 2024-09-24 17:59:25,478 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=557018.0, ans=0.0 2024-09-24 17:59:28,402 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.122e+02 1.235e+02 1.328e+02 1.443e+02 2.383e+02, threshold=2.655e+02, percent-clipped=0.0 2024-09-24 17:59:39,627 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=557064.6666666666, ans=0.125 2024-09-24 17:59:50,008 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn2.whiten, num_groups=1, num_channels=192, metric=8.09 vs. limit=22.5 2024-09-24 17:59:52,318 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=557064.6666666666, ans=0.0 2024-09-24 17:59:55,317 INFO [train.py:1198] (2/4) Epoch 31, batch 2500, loss[loss=0.1918, ctc_loss=0.124, cr_loss=0.3389, over 17014.00 frames. ], tot_loss[loss=0.1988, ctc_loss=0.1295, cr_loss=0.3467, over 3361471.34 frames. ], batch size: 44, lr: 3.84e-03, grad_scale: 32.0 2024-09-24 17:59:57,800 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.88 vs. limit=12.0 2024-09-24 18:00:01,998 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=557111.3333333334, ans=0.0 2024-09-24 18:00:20,203 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.34 vs. limit=10.0 2024-09-24 18:00:37,833 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=557204.6666666666, ans=0.0 2024-09-24 18:01:04,329 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=557298.0, ans=0.0 2024-09-24 18:01:05,903 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=557298.0, ans=0.125 2024-09-24 18:01:15,756 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=557298.0, ans=0.1 2024-09-24 18:01:18,463 INFO [train.py:1198] (2/4) Epoch 31, batch 2550, loss[loss=0.1908, ctc_loss=0.1233, cr_loss=0.3372, over 17266.00 frames. ], tot_loss[loss=0.198, ctc_loss=0.1288, cr_loss=0.346, over 3366978.25 frames. ], batch size: 44, lr: 3.83e-03, grad_scale: 32.0 2024-09-24 18:01:34,457 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=557344.6666666666, ans=0.125 2024-09-24 18:01:40,688 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=557391.3333333334, ans=0.125 2024-09-24 18:01:56,487 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.09 vs. limit=6.0 2024-09-24 18:02:16,588 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.233e+02 1.314e+02 1.407e+02 1.652e+02, threshold=2.628e+02, percent-clipped=0.0 2024-09-24 18:02:20,109 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=557484.6666666666, ans=0.0 2024-09-24 18:02:23,379 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=557484.6666666666, ans=0.1 2024-09-24 18:02:31,874 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=7.85 vs. limit=15.0 2024-09-24 18:02:37,537 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=557531.3333333334, ans=0.04949747468305833 2024-09-24 18:02:43,701 INFO [train.py:1198] (2/4) Epoch 31, batch 2600, loss[loss=0.1942, ctc_loss=0.1272, cr_loss=0.3351, over 17004.00 frames. ], tot_loss[loss=0.1985, ctc_loss=0.1292, cr_loss=0.3467, over 3364993.15 frames. ], batch size: 44, lr: 3.83e-03, grad_scale: 32.0 2024-09-24 18:02:44,032 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=557578.0, ans=0.2 2024-09-24 18:03:22,866 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=10.67 vs. limit=15.0 2024-09-24 18:03:24,970 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=11.34 vs. limit=22.5 2024-09-24 18:03:32,983 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=13.90 vs. limit=22.5 2024-09-24 18:03:54,466 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=557764.6666666666, ans=0.0 2024-09-24 18:04:02,333 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=557764.6666666666, ans=0.0 2024-09-24 18:04:05,511 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=557811.3333333334, ans=0.125 2024-09-24 18:04:06,716 INFO [train.py:1198] (2/4) Epoch 31, batch 2650, loss[loss=0.2158, ctc_loss=0.1436, cr_loss=0.3613, over 17020.00 frames. ], tot_loss[loss=0.2008, ctc_loss=0.131, cr_loss=0.349, over 3358372.78 frames. ], batch size: 53, lr: 3.83e-03, grad_scale: 32.0 2024-09-24 18:04:08,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=557811.3333333334, ans=0.125 2024-09-24 18:04:08,682 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=557811.3333333334, ans=0.125 2024-09-24 18:04:30,803 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=557858.0, ans=0.1 2024-09-24 18:04:32,744 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=11.87 vs. limit=15.0 2024-09-24 18:04:38,892 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=557904.6666666666, ans=0.125 2024-09-24 18:04:45,198 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=557904.6666666666, ans=0.0 2024-09-24 18:04:54,853 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=557951.3333333334, ans=0.125 2024-09-24 18:04:59,304 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.276e+02 1.391e+02 1.491e+02 3.639e+02, threshold=2.781e+02, percent-clipped=1.0 2024-09-24 18:05:14,088 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=557998.0, ans=0.125 2024-09-24 18:05:26,482 INFO [train.py:1198] (2/4) Epoch 31, batch 2700, loss[loss=0.179, ctc_loss=0.1124, cr_loss=0.3329, over 17289.00 frames. ], tot_loss[loss=0.2003, ctc_loss=0.1306, cr_loss=0.3485, over 3362537.33 frames. ], batch size: 51, lr: 3.83e-03, grad_scale: 32.0 2024-09-24 18:05:45,330 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=558091.3333333334, ans=0.125 2024-09-24 18:05:46,937 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=558091.3333333334, ans=0.2 2024-09-24 18:05:48,669 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=558091.3333333334, ans=0.0 2024-09-24 18:06:10,998 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=558138.0, ans=0.125 2024-09-24 18:06:47,219 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=10.49 vs. limit=15.0 2024-09-24 18:06:54,369 INFO [train.py:1198] (2/4) Epoch 31, batch 2750, loss[loss=0.18, ctc_loss=0.1153, cr_loss=0.3233, over 16949.00 frames. ], tot_loss[loss=0.1996, ctc_loss=0.1301, cr_loss=0.3474, over 3357845.56 frames. ], batch size: 42, lr: 3.83e-03, grad_scale: 16.0 2024-09-24 18:07:18,605 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=558324.6666666666, ans=0.125 2024-09-24 18:07:28,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=558371.3333333334, ans=0.0 2024-09-24 18:07:34,426 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=558371.3333333334, ans=0.125 2024-09-24 18:07:34,521 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=558371.3333333334, ans=0.125 2024-09-24 18:07:48,590 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.267e+02 1.363e+02 1.488e+02 1.840e+02, threshold=2.726e+02, percent-clipped=0.0 2024-09-24 18:08:03,770 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=9.54 vs. limit=15.0 2024-09-24 18:08:08,273 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=558464.6666666666, ans=0.1 2024-09-24 18:08:14,311 INFO [train.py:1198] (2/4) Epoch 31, batch 2800, loss[loss=0.2431, ctc_loss=0.1615, cr_loss=0.4079, over 15091.00 frames. ], tot_loss[loss=0.2001, ctc_loss=0.1305, cr_loss=0.3479, over 3348510.97 frames. ], batch size: 89, lr: 3.83e-03, grad_scale: 32.0 2024-09-24 18:08:17,893 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=558511.3333333334, ans=0.125 2024-09-24 18:08:25,767 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=558511.3333333334, ans=0.125 2024-09-24 18:08:28,860 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=558558.0, ans=0.2 2024-09-24 18:08:38,637 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=558558.0, ans=0.2 2024-09-24 18:08:50,527 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.min_positive, batch_count=558604.6666666666, ans=0.025 2024-09-24 18:08:53,832 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=558604.6666666666, ans=0.125 2024-09-24 18:09:10,034 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=8.24 vs. limit=15.0 2024-09-24 18:09:17,636 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=558651.3333333334, ans=0.0 2024-09-24 18:09:22,641 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=11.08 vs. limit=12.0 2024-09-24 18:09:25,553 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=558698.0, ans=0.1 2024-09-24 18:09:36,682 INFO [train.py:1198] (2/4) Epoch 31, batch 2850, loss[loss=0.2173, ctc_loss=0.1422, cr_loss=0.3755, over 16067.00 frames. ], tot_loss[loss=0.2004, ctc_loss=0.1307, cr_loss=0.3481, over 3345164.30 frames. ], batch size: 74, lr: 3.83e-03, grad_scale: 32.0 2024-09-24 18:10:12,221 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=558838.0, ans=0.025 2024-09-24 18:10:35,374 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.168e+02 1.299e+02 1.361e+02 1.459e+02 2.038e+02, threshold=2.722e+02, percent-clipped=0.0 2024-09-24 18:10:59,635 INFO [train.py:1198] (2/4) Epoch 31, batch 2900, loss[loss=0.2403, ctc_loss=0.1562, cr_loss=0.4207, over 15007.00 frames. ], tot_loss[loss=0.201, ctc_loss=0.1311, cr_loss=0.3497, over 3347598.75 frames. ], batch size: 89, lr: 3.83e-03, grad_scale: 16.0 2024-09-24 18:11:11,157 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=558978.0, ans=0.025 2024-09-24 18:11:58,556 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=11.92 vs. limit=15.0 2024-09-24 18:12:11,250 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=559164.6666666666, ans=0.125 2024-09-24 18:12:17,819 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=559164.6666666666, ans=0.09899494936611666 2024-09-24 18:12:25,454 INFO [train.py:1198] (2/4) Epoch 31, batch 2950, loss[loss=0.204, ctc_loss=0.1318, cr_loss=0.361, over 15972.00 frames. ], tot_loss[loss=0.1996, ctc_loss=0.1301, cr_loss=0.3475, over 3335800.54 frames. ], batch size: 74, lr: 3.83e-03, grad_scale: 16.0 2024-09-24 18:13:02,135 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=3.88 vs. limit=15.0 2024-09-24 18:13:07,851 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=559304.6666666666, ans=0.125 2024-09-24 18:13:18,943 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=559351.3333333334, ans=0.05 2024-09-24 18:13:21,742 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.279e+02 1.362e+02 1.485e+02 2.605e+02, threshold=2.724e+02, percent-clipped=0.0 2024-09-24 18:13:26,562 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=559351.3333333334, ans=0.125 2024-09-24 18:13:33,632 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=15.48 vs. limit=22.5 2024-09-24 18:13:45,365 INFO [train.py:1198] (2/4) Epoch 31, batch 3000, loss[loss=0.24, ctc_loss=0.1689, cr_loss=0.3556, over 11605.00 frames. ], tot_loss[loss=0.1994, ctc_loss=0.1299, cr_loss=0.3478, over 3346508.23 frames. ], batch size: 123, lr: 3.83e-03, grad_scale: 16.0 2024-09-24 18:13:45,365 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 18:14:00,844 INFO [train.py:1230] (2/4) Epoch 31, validation: loss=0.03667, ctc_loss=0.03667, cr_loss=9.013e-15, over 944034.00 frames. 2024-09-24 18:14:00,845 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 18:14:15,238 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=559491.3333333334, ans=0.2 2024-09-24 18:14:17,373 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.72 vs. limit=10.0 2024-09-24 18:14:26,186 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=559491.3333333334, ans=0.125 2024-09-24 18:14:40,413 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.24 vs. limit=10.0 2024-09-24 18:14:57,592 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=559584.6666666666, ans=0.125 2024-09-24 18:14:57,985 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=7.59 vs. limit=15.0 2024-09-24 18:15:19,066 INFO [train.py:1198] (2/4) Epoch 31, batch 3050, loss[loss=0.234, ctc_loss=0.1591, cr_loss=0.3744, over 11884.00 frames. ], tot_loss[loss=0.1986, ctc_loss=0.1293, cr_loss=0.3466, over 3353282.14 frames. ], batch size: 123, lr: 3.83e-03, grad_scale: 16.0 2024-09-24 18:16:13,651 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.155e+02 1.275e+02 1.357e+02 1.505e+02 2.128e+02, threshold=2.715e+02, percent-clipped=0.0 2024-09-24 18:16:37,063 INFO [train.py:1198] (2/4) Epoch 31, batch 3100, loss[loss=0.1631, ctc_loss=0.1015, cr_loss=0.3082, over 17055.00 frames. ], tot_loss[loss=0.1978, ctc_loss=0.1287, cr_loss=0.3454, over 3357360.65 frames. ], batch size: 39, lr: 3.83e-03, grad_scale: 16.0 2024-09-24 18:16:46,509 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=559911.3333333334, ans=0.125 2024-09-24 18:16:51,960 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.80 vs. limit=15.0 2024-09-24 18:16:52,739 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=559958.0, ans=0.125 2024-09-24 18:16:54,385 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=559958.0, ans=0.1 2024-09-24 18:17:05,584 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=559958.0, ans=0.125 2024-09-24 18:17:34,425 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=560051.3333333334, ans=0.125 2024-09-24 18:17:41,182 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten.whitening_limit, batch_count=560098.0, ans=15.0 2024-09-24 18:17:57,415 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=10.91 vs. limit=22.5 2024-09-24 18:18:00,009 INFO [train.py:1198] (2/4) Epoch 31, batch 3150, loss[loss=0.1692, ctc_loss=0.1123, cr_loss=0.2844, over 17179.00 frames. ], tot_loss[loss=0.1974, ctc_loss=0.1285, cr_loss=0.3448, over 3362312.14 frames. ], batch size: 41, lr: 3.83e-03, grad_scale: 16.0 2024-09-24 18:18:00,359 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=560144.6666666666, ans=0.025 2024-09-24 18:18:11,553 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=560144.6666666666, ans=0.125 2024-09-24 18:18:13,594 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.55 vs. limit=6.0 2024-09-24 18:18:42,961 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.91 vs. limit=10.0 2024-09-24 18:18:44,419 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=560238.0, ans=0.125 2024-09-24 18:18:52,344 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=560284.6666666666, ans=0.2 2024-09-24 18:18:55,088 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.106e+02 1.279e+02 1.355e+02 1.423e+02 2.365e+02, threshold=2.709e+02, percent-clipped=0.0 2024-09-24 18:19:09,434 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=560331.3333333334, ans=0.1 2024-09-24 18:19:18,857 INFO [train.py:1198] (2/4) Epoch 31, batch 3200, loss[loss=0.1666, ctc_loss=0.1055, cr_loss=0.3055, over 16305.00 frames. ], tot_loss[loss=0.1979, ctc_loss=0.1288, cr_loss=0.3454, over 3351781.59 frames. ], batch size: 36, lr: 3.82e-03, grad_scale: 32.0 2024-09-24 18:19:30,101 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=560378.0, ans=0.025 2024-09-24 18:19:56,197 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.whiten.whitening_limit, batch_count=560471.3333333334, ans=15.0 2024-09-24 18:19:57,079 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=560471.3333333334, ans=0.125 2024-09-24 18:19:58,669 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=560471.3333333334, ans=0.125 2024-09-24 18:20:03,336 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=560471.3333333334, ans=0.0 2024-09-24 18:20:13,686 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.99 vs. limit=15.0 2024-09-24 18:20:29,082 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=560564.6666666666, ans=0.1 2024-09-24 18:20:29,087 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=560564.6666666666, ans=0.0 2024-09-24 18:20:35,319 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=560564.6666666666, ans=0.07 2024-09-24 18:20:41,360 INFO [train.py:1198] (2/4) Epoch 31, batch 3250, loss[loss=0.2288, ctc_loss=0.1528, cr_loss=0.3797, over 16417.00 frames. ], tot_loss[loss=0.1977, ctc_loss=0.1286, cr_loss=0.3451, over 3362922.27 frames. ], batch size: 66, lr: 3.82e-03, grad_scale: 16.0 2024-09-24 18:20:56,246 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=9.05 vs. limit=15.0 2024-09-24 18:21:22,285 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=560704.6666666666, ans=0.025 2024-09-24 18:21:37,469 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.092e+02 1.237e+02 1.355e+02 1.495e+02 1.938e+02, threshold=2.711e+02, percent-clipped=0.0 2024-09-24 18:21:44,282 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=13.03 vs. limit=15.0 2024-09-24 18:21:59,316 INFO [train.py:1198] (2/4) Epoch 31, batch 3300, loss[loss=0.2046, ctc_loss=0.1324, cr_loss=0.3609, over 17008.00 frames. ], tot_loss[loss=0.1969, ctc_loss=0.1281, cr_loss=0.3443, over 3364886.36 frames. ], batch size: 51, lr: 3.82e-03, grad_scale: 16.0 2024-09-24 18:22:01,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=560844.6666666666, ans=0.125 2024-09-24 18:22:18,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=560891.3333333334, ans=0.1 2024-09-24 18:22:32,624 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=560938.0, ans=0.0 2024-09-24 18:22:40,521 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=560938.0, ans=0.0 2024-09-24 18:22:45,360 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=560984.6666666666, ans=0.2 2024-09-24 18:22:47,083 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=560984.6666666666, ans=0.125 2024-09-24 18:23:10,694 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=12.99 vs. limit=15.0 2024-09-24 18:23:13,581 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=561031.3333333334, ans=0.0 2024-09-24 18:23:13,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=561031.3333333334, ans=0.125 2024-09-24 18:23:18,163 INFO [train.py:1198] (2/4) Epoch 31, batch 3350, loss[loss=0.1753, ctc_loss=0.1126, cr_loss=0.3133, over 17290.00 frames. ], tot_loss[loss=0.197, ctc_loss=0.1281, cr_loss=0.3447, over 3367898.34 frames. ], batch size: 42, lr: 3.82e-03, grad_scale: 16.0 2024-09-24 18:23:24,749 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=561078.0, ans=0.1 2024-09-24 18:23:44,988 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=561124.6666666666, ans=0.09899494936611666 2024-09-24 18:23:53,090 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=561171.3333333334, ans=0.125 2024-09-24 18:24:04,386 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.79 vs. limit=15.0 2024-09-24 18:24:05,097 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=561218.0, ans=0.125 2024-09-24 18:24:14,290 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.283e+02 1.367e+02 1.477e+02 4.748e+02, threshold=2.734e+02, percent-clipped=1.0 2024-09-24 18:24:36,277 INFO [train.py:1198] (2/4) Epoch 31, batch 3400, loss[loss=0.1846, ctc_loss=0.1206, cr_loss=0.32, over 17155.00 frames. ], tot_loss[loss=0.1972, ctc_loss=0.1283, cr_loss=0.3444, over 3363369.92 frames. ], batch size: 45, lr: 3.82e-03, grad_scale: 16.0 2024-09-24 18:24:57,440 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=561358.0, ans=0.1 2024-09-24 18:24:59,055 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=561358.0, ans=0.2 2024-09-24 18:25:05,240 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=561358.0, ans=0.1 2024-09-24 18:25:24,016 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=561451.3333333334, ans=0.1 2024-09-24 18:25:47,254 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=561498.0, ans=0.0 2024-09-24 18:25:48,633 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=561498.0, ans=0.2 2024-09-24 18:25:56,288 INFO [train.py:1198] (2/4) Epoch 31, batch 3450, loss[loss=0.184, ctc_loss=0.1169, cr_loss=0.3357, over 17213.00 frames. ], tot_loss[loss=0.1981, ctc_loss=0.129, cr_loss=0.3455, over 3359270.01 frames. ], batch size: 47, lr: 3.82e-03, grad_scale: 16.0 2024-09-24 18:26:13,792 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=561591.3333333334, ans=0.125 2024-09-24 18:26:24,797 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=561591.3333333334, ans=0.125 2024-09-24 18:26:27,860 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=561638.0, ans=0.125 2024-09-24 18:26:51,157 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=561684.6666666666, ans=0.1 2024-09-24 18:26:52,251 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.286e+02 1.363e+02 1.449e+02 2.372e+02, threshold=2.726e+02, percent-clipped=0.0 2024-09-24 18:27:13,004 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.max_abs, batch_count=561778.0, ans=10.0 2024-09-24 18:27:14,199 INFO [train.py:1198] (2/4) Epoch 31, batch 3500, loss[loss=0.1676, ctc_loss=0.108, cr_loss=0.2981, over 16963.00 frames. ], tot_loss[loss=0.1983, ctc_loss=0.1291, cr_loss=0.3461, over 3368139.33 frames. ], batch size: 42, lr: 3.82e-03, grad_scale: 16.0 2024-09-24 18:27:24,164 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=561778.0, ans=0.0 2024-09-24 18:27:31,866 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=561824.6666666666, ans=0.1 2024-09-24 18:27:33,460 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=561824.6666666666, ans=0.0 2024-09-24 18:28:34,275 INFO [train.py:1198] (2/4) Epoch 31, batch 3550, loss[loss=0.2579, ctc_loss=0.1825, cr_loss=0.3773, over 11384.00 frames. ], tot_loss[loss=0.2004, ctc_loss=0.1307, cr_loss=0.3484, over 3337285.73 frames. ], batch size: 125, lr: 3.82e-03, grad_scale: 16.0 2024-09-24 18:28:53,350 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-24 18:29:03,395 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=5.46 vs. limit=15.0 2024-09-24 18:29:18,715 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.85 vs. limit=15.0 2024-09-24 18:29:33,956 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.64 vs. limit=6.0 2024-09-24 18:29:34,934 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.099e+02 1.302e+02 1.378e+02 1.478e+02 2.528e+02, threshold=2.755e+02, percent-clipped=0.0 2024-09-24 18:29:43,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=562198.0, ans=0.125 2024-09-24 18:29:53,975 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=562198.0, ans=0.125 2024-09-24 18:29:56,784 INFO [train.py:1198] (2/4) Epoch 31, batch 3600, loss[loss=0.2462, ctc_loss=0.1728, cr_loss=0.3672, over 12437.00 frames. ], tot_loss[loss=0.1997, ctc_loss=0.1303, cr_loss=0.3471, over 3330950.58 frames. ], batch size: 123, lr: 3.82e-03, grad_scale: 32.0 2024-09-24 18:30:10,924 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=562291.3333333334, ans=0.125 2024-09-24 18:30:45,065 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=562384.6666666666, ans=0.125 2024-09-24 18:31:02,330 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=562431.3333333334, ans=0.125 2024-09-24 18:31:14,555 INFO [train.py:1198] (2/4) Epoch 31, batch 3650, loss[loss=0.1624, ctc_loss=0.1057, cr_loss=0.2837, over 17087.00 frames. ], tot_loss[loss=0.1993, ctc_loss=0.13, cr_loss=0.3463, over 3333018.22 frames. ], batch size: 43, lr: 3.82e-03, grad_scale: 32.0 2024-09-24 18:31:19,968 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=5.56 vs. limit=12.0 2024-09-24 18:31:35,170 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=562524.6666666666, ans=0.125 2024-09-24 18:31:43,760 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.97 vs. limit=15.0 2024-09-24 18:31:56,250 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=562571.3333333334, ans=0.1 2024-09-24 18:32:08,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=562618.0, ans=0.0 2024-09-24 18:32:11,871 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.260e+02 1.374e+02 1.513e+02 2.191e+02, threshold=2.748e+02, percent-clipped=0.0 2024-09-24 18:32:13,838 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=562618.0, ans=0.125 2024-09-24 18:32:26,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=562664.6666666666, ans=0.125 2024-09-24 18:32:33,922 INFO [train.py:1198] (2/4) Epoch 31, batch 3700, loss[loss=0.2287, ctc_loss=0.1531, cr_loss=0.378, over 16028.00 frames. ], tot_loss[loss=0.1991, ctc_loss=0.1299, cr_loss=0.346, over 3333577.90 frames. ], batch size: 74, lr: 3.82e-03, grad_scale: 32.0 2024-09-24 18:32:35,704 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=562711.3333333334, ans=0.125 2024-09-24 18:33:12,691 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten.whitening_limit, batch_count=562804.6666666666, ans=22.5 2024-09-24 18:33:18,894 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=8.00 vs. limit=15.0 2024-09-24 18:33:33,768 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=562851.3333333334, ans=0.025 2024-09-24 18:33:45,101 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.67 vs. limit=22.5 2024-09-24 18:33:52,139 INFO [train.py:1198] (2/4) Epoch 31, batch 3750, loss[loss=0.1959, ctc_loss=0.128, cr_loss=0.3395, over 17309.00 frames. ], tot_loss[loss=0.1997, ctc_loss=0.1304, cr_loss=0.3465, over 3321619.30 frames. ], batch size: 51, lr: 3.82e-03, grad_scale: 32.0 2024-09-24 18:34:21,393 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=10.97 vs. limit=15.0 2024-09-24 18:34:48,370 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.261e+02 1.326e+02 1.431e+02 2.182e+02, threshold=2.652e+02, percent-clipped=0.0 2024-09-24 18:35:02,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=563131.3333333334, ans=0.0 2024-09-24 18:35:09,879 INFO [train.py:1198] (2/4) Epoch 31, batch 3800, loss[loss=0.2263, ctc_loss=0.1477, cr_loss=0.393, over 17022.00 frames. ], tot_loss[loss=0.1993, ctc_loss=0.13, cr_loss=0.3462, over 3327010.43 frames. ], batch size: 56, lr: 3.82e-03, grad_scale: 16.0 2024-09-24 18:35:36,707 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.39 vs. limit=12.0 2024-09-24 18:35:39,429 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=563271.3333333334, ans=0.2 2024-09-24 18:35:40,914 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=563271.3333333334, ans=0.2 2024-09-24 18:36:27,172 INFO [train.py:1198] (2/4) Epoch 31, batch 3850, loss[loss=0.1897, ctc_loss=0.1245, cr_loss=0.3262, over 17313.00 frames. ], tot_loss[loss=0.2002, ctc_loss=0.1308, cr_loss=0.3468, over 3292293.45 frames. ], batch size: 51, lr: 3.81e-03, grad_scale: 16.0 2024-09-24 18:36:44,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=563458.0, ans=0.125 2024-09-24 18:36:46,414 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=563458.0, ans=0.1 2024-09-24 18:37:09,726 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=563504.6666666666, ans=0.07 2024-09-24 18:37:21,690 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=563551.3333333334, ans=0.0 2024-09-24 18:37:24,225 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.110e+02 1.313e+02 1.427e+02 1.596e+02 2.274e+02, threshold=2.854e+02, percent-clipped=0.0 2024-09-24 18:37:24,978 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module1.whiten, num_groups=1, num_channels=512, metric=11.04 vs. limit=15.0 2024-09-24 18:37:32,858 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=6.49 vs. limit=15.0 2024-09-24 18:38:22,073 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=12.80 vs. limit=22.5 2024-09-24 18:38:29,743 INFO [train.py:1198] (2/4) Epoch 32, batch 0, loss[loss=0.2286, ctc_loss=0.1516, cr_loss=0.3846, over 16512.00 frames. ], tot_loss[loss=0.2286, ctc_loss=0.1516, cr_loss=0.3846, over 16512.00 frames. ], batch size: 66, lr: 3.75e-03, grad_scale: 32.0 2024-09-24 18:38:29,744 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 18:38:45,165 INFO [train.py:1230] (2/4) Epoch 32, validation: loss=0.03599, ctc_loss=0.03599, cr_loss=9.022e-15, over 944034.00 frames. 2024-09-24 18:38:45,166 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 18:38:58,122 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=563626.0, ans=0.125 2024-09-24 18:39:17,293 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=563719.3333333334, ans=0.125 2024-09-24 18:39:20,431 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=563719.3333333334, ans=0.0 2024-09-24 18:39:32,314 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=563766.0, ans=0.125 2024-09-24 18:39:49,403 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=563766.0, ans=0.1 2024-09-24 18:40:10,385 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=563859.3333333334, ans=0.125 2024-09-24 18:40:14,197 INFO [train.py:1198] (2/4) Epoch 32, batch 50, loss[loss=0.1735, ctc_loss=0.1111, cr_loss=0.312, over 17272.00 frames. ], tot_loss[loss=0.1984, ctc_loss=0.1288, cr_loss=0.3478, over 766794.87 frames. ], batch size: 42, lr: 3.75e-03, grad_scale: 32.0 2024-09-24 18:40:22,329 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=563859.3333333334, ans=0.0 2024-09-24 18:40:38,392 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer_ff2.min_abs, batch_count=563906.0, ans=0.1 2024-09-24 18:40:39,938 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=563906.0, ans=0.125 2024-09-24 18:40:44,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=563952.6666666666, ans=0.125 2024-09-24 18:41:19,567 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.079e+02 1.253e+02 1.338e+02 1.477e+02 2.326e+02, threshold=2.677e+02, percent-clipped=0.0 2024-09-24 18:41:33,988 INFO [train.py:1198] (2/4) Epoch 32, batch 100, loss[loss=0.233, ctc_loss=0.1555, cr_loss=0.3874, over 15013.00 frames. ], tot_loss[loss=0.2006, ctc_loss=0.1303, cr_loss=0.3517, over 1349692.16 frames. ], batch size: 89, lr: 3.75e-03, grad_scale: 32.0 2024-09-24 18:41:34,273 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=564092.6666666666, ans=0.0 2024-09-24 18:41:58,418 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=564139.3333333334, ans=0.09899494936611666 2024-09-24 18:42:14,627 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=564186.0, ans=0.2 2024-09-24 18:42:16,206 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=564186.0, ans=0.1 2024-09-24 18:42:38,721 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 18:42:56,205 INFO [train.py:1198] (2/4) Epoch 32, batch 150, loss[loss=0.181, ctc_loss=0.1137, cr_loss=0.3365, over 17099.00 frames. ], tot_loss[loss=0.1999, ctc_loss=0.1297, cr_loss=0.3512, over 1803548.29 frames. ], batch size: 40, lr: 3.75e-03, grad_scale: 32.0 2024-09-24 18:43:44,559 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=564466.0, ans=0.125 2024-09-24 18:43:49,314 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=564466.0, ans=0.0 2024-09-24 18:44:01,705 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.094e+02 1.312e+02 1.410e+02 1.549e+02 1.890e+02, threshold=2.819e+02, percent-clipped=0.0 2024-09-24 18:44:11,569 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=564512.6666666666, ans=0.125 2024-09-24 18:44:11,726 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=564512.6666666666, ans=0.05 2024-09-24 18:44:13,155 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=564512.6666666666, ans=0.125 2024-09-24 18:44:16,127 INFO [train.py:1198] (2/4) Epoch 32, batch 200, loss[loss=0.2293, ctc_loss=0.1534, cr_loss=0.3797, over 16981.00 frames. ], tot_loss[loss=0.1992, ctc_loss=0.1293, cr_loss=0.3493, over 2147108.41 frames. ], batch size: 53, lr: 3.75e-03, grad_scale: 32.0 2024-09-24 18:44:16,431 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=564559.3333333334, ans=0.0 2024-09-24 18:44:16,528 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=564559.3333333334, ans=0.125 2024-09-24 18:44:31,103 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=564606.0, ans=0.0 2024-09-24 18:44:45,316 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 18:45:33,759 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=564746.0, ans=0.1 2024-09-24 18:45:40,013 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=564746.0, ans=0.125 2024-09-24 18:45:46,132 INFO [train.py:1198] (2/4) Epoch 32, batch 250, loss[loss=0.198, ctc_loss=0.1273, cr_loss=0.3539, over 16980.00 frames. ], tot_loss[loss=0.2008, ctc_loss=0.1307, cr_loss=0.3506, over 2407365.53 frames. ], batch size: 56, lr: 3.75e-03, grad_scale: 32.0 2024-09-24 18:45:46,508 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=564792.6666666666, ans=0.1 2024-09-24 18:46:30,031 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module1.whiten, num_groups=1, num_channels=192, metric=9.10 vs. limit=15.0 2024-09-24 18:46:42,078 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=564932.6666666666, ans=0.125 2024-09-24 18:46:52,819 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.110e+02 1.269e+02 1.327e+02 1.431e+02 2.007e+02, threshold=2.654e+02, percent-clipped=0.0 2024-09-24 18:47:03,285 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.86 vs. limit=22.5 2024-09-24 18:47:05,424 INFO [train.py:1198] (2/4) Epoch 32, batch 300, loss[loss=0.2196, ctc_loss=0.148, cr_loss=0.3581, over 15388.00 frames. ], tot_loss[loss=0.1998, ctc_loss=0.13, cr_loss=0.3491, over 2620791.20 frames. ], batch size: 89, lr: 3.75e-03, grad_scale: 16.0 2024-09-24 18:47:26,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=565072.6666666666, ans=0.2 2024-09-24 18:47:37,812 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.89 vs. limit=6.0 2024-09-24 18:47:55,724 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer_na.min_abs, batch_count=565166.0, ans=0.02 2024-09-24 18:47:58,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=565166.0, ans=0.0 2024-09-24 18:48:02,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=565166.0, ans=0.1 2024-09-24 18:48:08,604 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=565166.0, ans=0.125 2024-09-24 18:48:13,296 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=565212.6666666666, ans=0.125 2024-09-24 18:48:26,329 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=565212.6666666666, ans=0.0 2024-09-24 18:48:29,163 INFO [train.py:1198] (2/4) Epoch 32, batch 350, loss[loss=0.2102, ctc_loss=0.136, cr_loss=0.3712, over 17298.00 frames. ], tot_loss[loss=0.2003, ctc_loss=0.1305, cr_loss=0.3494, over 2781681.83 frames. ], batch size: 51, lr: 3.75e-03, grad_scale: 16.0 2024-09-24 18:48:42,554 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=565259.3333333334, ans=0.025 2024-09-24 18:48:53,680 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer_ff3.min_abs, batch_count=565306.0, ans=0.2 2024-09-24 18:48:58,341 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=565306.0, ans=0.125 2024-09-24 18:49:20,913 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=565399.3333333334, ans=0.125 2024-09-24 18:49:36,802 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.109e+02 1.263e+02 1.319e+02 1.416e+02 2.178e+02, threshold=2.638e+02, percent-clipped=0.0 2024-09-24 18:49:51,959 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=4.43 vs. limit=15.0 2024-09-24 18:49:52,384 INFO [train.py:1198] (2/4) Epoch 32, batch 400, loss[loss=0.2032, ctc_loss=0.1328, cr_loss=0.3523, over 17312.00 frames. ], tot_loss[loss=0.1998, ctc_loss=0.1301, cr_loss=0.3482, over 2901734.57 frames. ], batch size: 46, lr: 3.75e-03, grad_scale: 32.0 2024-09-24 18:50:05,335 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=565492.6666666666, ans=0.04949747468305833 2024-09-24 18:50:24,146 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=565539.3333333334, ans=0.125 2024-09-24 18:50:45,467 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=565632.6666666666, ans=0.1 2024-09-24 18:51:18,645 INFO [train.py:1198] (2/4) Epoch 32, batch 450, loss[loss=0.2081, ctc_loss=0.1341, cr_loss=0.3699, over 17027.00 frames. ], tot_loss[loss=0.2, ctc_loss=0.1302, cr_loss=0.3487, over 2991241.01 frames. ], batch size: 56, lr: 3.75e-03, grad_scale: 32.0 2024-09-24 18:51:21,371 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.66 vs. limit=6.0 2024-09-24 18:51:30,438 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=565726.0, ans=0.2 2024-09-24 18:51:43,539 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=3.84 vs. limit=10.0 2024-09-24 18:51:50,906 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=565819.3333333334, ans=0.0 2024-09-24 18:52:03,706 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.min_positive, batch_count=565819.3333333334, ans=0.05 2024-09-24 18:52:09,974 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=565866.0, ans=0.125 2024-09-24 18:52:25,962 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.138e+02 1.290e+02 1.389e+02 1.522e+02 2.571e+02, threshold=2.778e+02, percent-clipped=0.0 2024-09-24 18:52:27,985 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=565912.6666666666, ans=0.125 2024-09-24 18:52:28,035 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=565912.6666666666, ans=0.125 2024-09-24 18:52:29,555 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=565912.6666666666, ans=0.0 2024-09-24 18:52:31,162 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=565912.6666666666, ans=0.2 2024-09-24 18:52:36,649 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=5.69 vs. limit=15.0 2024-09-24 18:52:38,788 INFO [train.py:1198] (2/4) Epoch 32, batch 500, loss[loss=0.2205, ctc_loss=0.1435, cr_loss=0.3851, over 17142.00 frames. ], tot_loss[loss=0.2001, ctc_loss=0.1305, cr_loss=0.3483, over 3064602.27 frames. ], batch size: 48, lr: 3.74e-03, grad_scale: 32.0 2024-09-24 18:52:58,255 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=566006.0, ans=0.125 2024-09-24 18:53:05,089 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.02 vs. limit=15.0 2024-09-24 18:53:32,545 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.76 vs. limit=6.0 2024-09-24 18:53:37,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=566099.3333333334, ans=0.125 2024-09-24 18:53:44,817 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=5.28 vs. limit=15.0 2024-09-24 18:54:01,197 INFO [train.py:1198] (2/4) Epoch 32, batch 550, loss[loss=0.1929, ctc_loss=0.1246, cr_loss=0.3413, over 17031.00 frames. ], tot_loss[loss=0.1996, ctc_loss=0.13, cr_loss=0.3482, over 3133731.57 frames. ], batch size: 56, lr: 3.74e-03, grad_scale: 32.0 2024-09-24 18:54:16,063 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=566239.3333333334, ans=0.125 2024-09-24 18:54:28,728 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=566239.3333333334, ans=0.125 2024-09-24 18:54:39,422 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=566286.0, ans=0.0 2024-09-24 18:54:41,046 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=566286.0, ans=0.1 2024-09-24 18:54:44,276 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=566286.0, ans=0.125 2024-09-24 18:55:16,236 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.095e+02 1.263e+02 1.353e+02 1.425e+02 2.013e+02, threshold=2.706e+02, percent-clipped=0.0 2024-09-24 18:55:19,845 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=566379.3333333334, ans=0.0 2024-09-24 18:55:23,372 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=6.57 vs. limit=15.0 2024-09-24 18:55:29,022 INFO [train.py:1198] (2/4) Epoch 32, batch 600, loss[loss=0.2022, ctc_loss=0.1332, cr_loss=0.3449, over 16777.00 frames. ], tot_loss[loss=0.1987, ctc_loss=0.1294, cr_loss=0.3462, over 3177249.98 frames. ], batch size: 61, lr: 3.74e-03, grad_scale: 32.0 2024-09-24 18:55:45,601 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=566472.6666666666, ans=0.1 2024-09-24 18:55:53,307 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=566472.6666666666, ans=0.125 2024-09-24 18:55:53,731 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=11.70 vs. limit=15.0 2024-09-24 18:56:03,016 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=566519.3333333334, ans=0.125 2024-09-24 18:56:43,287 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=566612.6666666666, ans=0.125 2024-09-24 18:56:49,244 INFO [train.py:1198] (2/4) Epoch 32, batch 650, loss[loss=0.1903, ctc_loss=0.1235, cr_loss=0.3339, over 17144.00 frames. ], tot_loss[loss=0.1986, ctc_loss=0.1294, cr_loss=0.3463, over 3219734.03 frames. ], batch size: 45, lr: 3.74e-03, grad_scale: 32.0 2024-09-24 18:56:49,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=566659.3333333334, ans=0.1 2024-09-24 18:56:55,996 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 18:57:06,425 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=5.04 vs. limit=10.0 2024-09-24 18:57:08,893 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=566706.0, ans=0.05 2024-09-24 18:57:28,081 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=566752.6666666666, ans=0.125 2024-09-24 18:57:28,195 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=566752.6666666666, ans=0.1 2024-09-24 18:57:55,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=566846.0, ans=0.125 2024-09-24 18:58:00,289 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=566846.0, ans=0.125 2024-09-24 18:58:01,565 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.265e+02 1.321e+02 1.437e+02 2.053e+02, threshold=2.643e+02, percent-clipped=0.0 2024-09-24 18:58:03,581 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=566846.0, ans=0.0 2024-09-24 18:58:05,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=566846.0, ans=0.125 2024-09-24 18:58:13,018 INFO [train.py:1198] (2/4) Epoch 32, batch 700, loss[loss=0.1908, ctc_loss=0.1253, cr_loss=0.3276, over 17119.00 frames. ], tot_loss[loss=0.1978, ctc_loss=0.1288, cr_loss=0.345, over 3247883.64 frames. ], batch size: 40, lr: 3.74e-03, grad_scale: 16.0 2024-09-24 18:58:14,946 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=566892.6666666666, ans=0.125 2024-09-24 18:58:26,592 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.58 vs. limit=15.0 2024-09-24 18:58:55,110 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=566986.0, ans=0.1 2024-09-24 18:58:56,914 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=14.32 vs. limit=22.5 2024-09-24 18:58:58,757 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=6.69 vs. limit=15.0 2024-09-24 18:59:19,765 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=5.79 vs. limit=12.0 2024-09-24 18:59:32,124 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=7.22 vs. limit=12.0 2024-09-24 18:59:33,158 INFO [train.py:1198] (2/4) Epoch 32, batch 750, loss[loss=0.2067, ctc_loss=0.1321, cr_loss=0.3728, over 17352.00 frames. ], tot_loss[loss=0.1982, ctc_loss=0.1291, cr_loss=0.3457, over 3270720.98 frames. ], batch size: 48, lr: 3.74e-03, grad_scale: 16.0 2024-09-24 18:59:54,886 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=10.59 vs. limit=15.0 2024-09-24 19:00:01,010 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=567172.6666666666, ans=0.125 2024-09-24 19:00:14,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=567219.3333333334, ans=0.2 2024-09-24 19:00:21,654 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.08 vs. limit=12.0 2024-09-24 19:00:40,157 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=567266.0, ans=0.0 2024-09-24 19:00:41,576 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=567266.0, ans=0.1 2024-09-24 19:00:45,572 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten.whitening_limit, batch_count=567312.6666666666, ans=22.5 2024-09-24 19:00:49,277 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.130e+02 1.286e+02 1.352e+02 1.497e+02 2.163e+02, threshold=2.705e+02, percent-clipped=0.0 2024-09-24 19:01:00,447 INFO [train.py:1198] (2/4) Epoch 32, batch 800, loss[loss=0.2071, ctc_loss=0.1345, cr_loss=0.3634, over 17244.00 frames. ], tot_loss[loss=0.1984, ctc_loss=0.1291, cr_loss=0.3464, over 3294104.63 frames. ], batch size: 55, lr: 3.74e-03, grad_scale: 32.0 2024-09-24 19:01:14,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=567406.0, ans=0.1 2024-09-24 19:01:29,053 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=567406.0, ans=0.125 2024-09-24 19:01:33,806 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=567452.6666666666, ans=0.125 2024-09-24 19:01:56,665 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer_na.min_abs, batch_count=567499.3333333334, ans=0.02 2024-09-24 19:01:59,859 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=567499.3333333334, ans=0.1 2024-09-24 19:02:12,634 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=567546.0, ans=0.125 2024-09-24 19:02:14,108 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=567546.0, ans=0.0 2024-09-24 19:02:20,178 INFO [train.py:1198] (2/4) Epoch 32, batch 850, loss[loss=0.1766, ctc_loss=0.1131, cr_loss=0.3175, over 17091.00 frames. ], tot_loss[loss=0.197, ctc_loss=0.128, cr_loss=0.345, over 3313490.67 frames. ], batch size: 40, lr: 3.74e-03, grad_scale: 32.0 2024-09-24 19:02:28,872 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.34 vs. limit=15.0 2024-09-24 19:02:57,071 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.03 vs. limit=15.0 2024-09-24 19:03:02,028 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2.whitening_limit, batch_count=567686.0, ans=15.0 2024-09-24 19:03:11,414 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 19:03:31,999 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.300e+02 1.382e+02 1.535e+02 3.730e+02, threshold=2.764e+02, percent-clipped=1.0 2024-09-24 19:03:43,402 INFO [train.py:1198] (2/4) Epoch 32, batch 900, loss[loss=0.2176, ctc_loss=0.1439, cr_loss=0.3686, over 16847.00 frames. ], tot_loss[loss=0.1982, ctc_loss=0.1289, cr_loss=0.3468, over 3329952.91 frames. ], batch size: 58, lr: 3.74e-03, grad_scale: 32.0 2024-09-24 19:03:45,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=567826.0, ans=0.0 2024-09-24 19:03:53,420 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=567826.0, ans=0.125 2024-09-24 19:03:58,237 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=567872.6666666666, ans=0.025 2024-09-24 19:05:02,171 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=13.01 vs. limit=15.0 2024-09-24 19:05:09,169 INFO [train.py:1198] (2/4) Epoch 32, batch 950, loss[loss=0.2178, ctc_loss=0.144, cr_loss=0.3691, over 17001.00 frames. ], tot_loss[loss=0.2002, ctc_loss=0.1304, cr_loss=0.3488, over 3328865.15 frames. ], batch size: 53, lr: 3.74e-03, grad_scale: 32.0 2024-09-24 19:05:17,188 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.33 vs. limit=15.0 2024-09-24 19:05:23,258 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=568059.3333333334, ans=0.025 2024-09-24 19:05:37,955 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=568106.0, ans=0.2 2024-09-24 19:05:40,920 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=568106.0, ans=0.125 2024-09-24 19:05:41,092 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=568106.0, ans=0.125 2024-09-24 19:05:44,584 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.33 vs. limit=10.0 2024-09-24 19:06:06,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=568199.3333333334, ans=0.125 2024-09-24 19:06:15,853 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=568246.0, ans=0.125 2024-09-24 19:06:20,503 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.287e+02 1.382e+02 1.480e+02 1.903e+02, threshold=2.765e+02, percent-clipped=0.0 2024-09-24 19:06:31,838 INFO [train.py:1198] (2/4) Epoch 32, batch 1000, loss[loss=0.2082, ctc_loss=0.1364, cr_loss=0.3588, over 17354.00 frames. ], tot_loss[loss=0.1991, ctc_loss=0.1296, cr_loss=0.3475, over 3330367.93 frames. ], batch size: 48, lr: 3.74e-03, grad_scale: 32.0 2024-09-24 19:06:35,894 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=5.21 vs. limit=12.0 2024-09-24 19:07:15,685 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=568386.0, ans=0.0 2024-09-24 19:07:36,703 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=6.21 vs. limit=15.0 2024-09-24 19:07:54,840 INFO [train.py:1198] (2/4) Epoch 32, batch 1050, loss[loss=0.1894, ctc_loss=0.1219, cr_loss=0.3377, over 17009.00 frames. ], tot_loss[loss=0.1993, ctc_loss=0.1297, cr_loss=0.3477, over 3340070.36 frames. ], batch size: 44, lr: 3.74e-03, grad_scale: 32.0 2024-09-24 19:08:01,338 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.max_abs, batch_count=568526.0, ans=10.0 2024-09-24 19:08:10,969 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=568572.6666666666, ans=0.125 2024-09-24 19:08:18,789 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=568572.6666666666, ans=0.125 2024-09-24 19:08:35,284 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.24 vs. limit=15.0 2024-09-24 19:09:00,664 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=568712.6666666666, ans=0.0 2024-09-24 19:09:03,546 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.269e+02 1.394e+02 1.503e+02 2.012e+02, threshold=2.787e+02, percent-clipped=0.0 2024-09-24 19:09:15,034 INFO [train.py:1198] (2/4) Epoch 32, batch 1100, loss[loss=0.1634, ctc_loss=0.1033, cr_loss=0.3006, over 17210.00 frames. ], tot_loss[loss=0.1993, ctc_loss=0.1298, cr_loss=0.3475, over 3342769.05 frames. ], batch size: 50, lr: 3.74e-03, grad_scale: 32.0 2024-09-24 19:09:16,763 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=568759.3333333334, ans=0.125 2024-09-24 19:09:53,931 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=5.08 vs. limit=10.0 2024-09-24 19:09:59,998 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=568852.6666666666, ans=0.125 2024-09-24 19:10:05,339 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.02 vs. limit=15.0 2024-09-24 19:10:23,001 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=568899.3333333334, ans=0.125 2024-09-24 19:10:27,814 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=568946.0, ans=0.2 2024-09-24 19:10:32,510 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=568946.0, ans=0.0 2024-09-24 19:10:35,820 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=568946.0, ans=0.1 2024-09-24 19:10:41,723 INFO [train.py:1198] (2/4) Epoch 32, batch 1150, loss[loss=0.1929, ctc_loss=0.1264, cr_loss=0.3327, over 17295.00 frames. ], tot_loss[loss=0.1989, ctc_loss=0.1295, cr_loss=0.347, over 3333946.22 frames. ], batch size: 46, lr: 3.73e-03, grad_scale: 32.0 2024-09-24 19:10:45,329 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=568992.6666666666, ans=0.125 2024-09-24 19:10:53,269 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=568992.6666666666, ans=0.125 2024-09-24 19:10:54,894 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=568992.6666666666, ans=0.1 2024-09-24 19:11:21,922 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=569086.0, ans=0.0 2024-09-24 19:11:28,364 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=569132.6666666666, ans=0.1 2024-09-24 19:11:36,419 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=569132.6666666666, ans=0.125 2024-09-24 19:11:45,237 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=4.26 vs. limit=15.0 2024-09-24 19:11:50,710 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.032e+02 1.271e+02 1.362e+02 1.461e+02 3.149e+02, threshold=2.725e+02, percent-clipped=1.0 2024-09-24 19:11:51,617 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.58 vs. limit=12.0 2024-09-24 19:12:00,654 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=569226.0, ans=0.1 2024-09-24 19:12:01,961 INFO [train.py:1198] (2/4) Epoch 32, batch 1200, loss[loss=0.204, ctc_loss=0.1317, cr_loss=0.3614, over 17344.00 frames. ], tot_loss[loss=0.1986, ctc_loss=0.1293, cr_loss=0.347, over 3338932.27 frames. ], batch size: 48, lr: 3.73e-03, grad_scale: 32.0 2024-09-24 19:12:02,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=569226.0, ans=0.125 2024-09-24 19:12:44,244 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=6.43 vs. limit=15.0 2024-09-24 19:12:55,971 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=569366.0, ans=0.025 2024-09-24 19:13:05,453 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=569366.0, ans=0.125 2024-09-24 19:13:24,566 INFO [train.py:1198] (2/4) Epoch 32, batch 1250, loss[loss=0.207, ctc_loss=0.1358, cr_loss=0.3563, over 17250.00 frames. ], tot_loss[loss=0.1989, ctc_loss=0.1295, cr_loss=0.3469, over 3346226.58 frames. ], batch size: 44, lr: 3.73e-03, grad_scale: 32.0 2024-09-24 19:13:32,762 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=569459.3333333334, ans=0.1 2024-09-24 19:13:44,134 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.min_positive, batch_count=569506.0, ans=0.025 2024-09-24 19:13:55,379 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=569552.6666666666, ans=0.0 2024-09-24 19:13:57,025 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=569552.6666666666, ans=0.125 2024-09-24 19:13:59,132 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=6.04 vs. limit=12.0 2024-09-24 19:14:34,070 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.275e+02 1.363e+02 1.497e+02 1.949e+02, threshold=2.726e+02, percent-clipped=0.0 2024-09-24 19:14:42,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=569646.0, ans=0.2 2024-09-24 19:14:50,538 INFO [train.py:1198] (2/4) Epoch 32, batch 1300, loss[loss=0.2235, ctc_loss=0.1459, cr_loss=0.3882, over 17093.00 frames. ], tot_loss[loss=0.1986, ctc_loss=0.1292, cr_loss=0.3467, over 3351368.39 frames. ], batch size: 49, lr: 3.73e-03, grad_scale: 32.0 2024-09-24 19:14:57,263 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=569692.6666666666, ans=0.025 2024-09-24 19:14:58,836 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=569692.6666666666, ans=0.1 2024-09-24 19:15:02,261 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=569692.6666666666, ans=0.2 2024-09-24 19:15:10,165 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=569739.3333333334, ans=0.1 2024-09-24 19:15:39,665 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=569832.6666666666, ans=0.125 2024-09-24 19:15:49,437 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=569832.6666666666, ans=0.1 2024-09-24 19:16:05,601 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=569879.3333333334, ans=0.0 2024-09-24 19:16:13,218 INFO [train.py:1198] (2/4) Epoch 32, batch 1350, loss[loss=0.1872, ctc_loss=0.1204, cr_loss=0.3343, over 17295.00 frames. ], tot_loss[loss=0.1984, ctc_loss=0.129, cr_loss=0.3467, over 3355674.11 frames. ], batch size: 51, lr: 3.73e-03, grad_scale: 16.0 2024-09-24 19:16:31,287 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=569972.6666666666, ans=0.125 2024-09-24 19:16:32,877 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=569972.6666666666, ans=0.2 2024-09-24 19:16:37,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=569972.6666666666, ans=0.125 2024-09-24 19:16:40,804 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-24 19:17:21,695 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=6.46 vs. limit=15.0 2024-09-24 19:17:23,671 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.251e+02 1.359e+02 1.461e+02 2.063e+02, threshold=2.718e+02, percent-clipped=0.0 2024-09-24 19:17:24,847 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.69 vs. limit=6.0 2024-09-24 19:17:33,282 INFO [train.py:1198] (2/4) Epoch 32, batch 1400, loss[loss=0.212, ctc_loss=0.1381, cr_loss=0.3697, over 17362.00 frames. ], tot_loss[loss=0.1991, ctc_loss=0.1295, cr_loss=0.3476, over 3358060.70 frames. ], batch size: 48, lr: 3.73e-03, grad_scale: 16.0 2024-09-24 19:17:35,700 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.58 vs. limit=15.0 2024-09-24 19:17:44,760 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=570159.3333333334, ans=0.2 2024-09-24 19:18:07,042 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.89 vs. limit=15.0 2024-09-24 19:18:16,700 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=570252.6666666666, ans=0.125 2024-09-24 19:18:31,348 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=570299.3333333334, ans=0.5 2024-09-24 19:18:56,762 INFO [train.py:1198] (2/4) Epoch 32, batch 1450, loss[loss=0.1798, ctc_loss=0.1178, cr_loss=0.31, over 17121.00 frames. ], tot_loss[loss=0.1981, ctc_loss=0.1289, cr_loss=0.3463, over 3348898.70 frames. ], batch size: 40, lr: 3.73e-03, grad_scale: 16.0 2024-09-24 19:19:03,415 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=570392.6666666666, ans=0.125 2024-09-24 19:19:16,562 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=570439.3333333334, ans=0.125 2024-09-24 19:19:16,623 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=570439.3333333334, ans=0.07 2024-09-24 19:19:28,174 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.51 vs. limit=15.0 2024-09-24 19:19:56,949 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=570532.6666666666, ans=0.125 2024-09-24 19:20:12,693 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.051e+02 1.251e+02 1.341e+02 1.423e+02 1.681e+02, threshold=2.682e+02, percent-clipped=0.0 2024-09-24 19:20:23,576 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=570626.0, ans=0.2 2024-09-24 19:20:24,787 INFO [train.py:1198] (2/4) Epoch 32, batch 1500, loss[loss=0.2247, ctc_loss=0.1497, cr_loss=0.3753, over 14752.00 frames. ], tot_loss[loss=0.1979, ctc_loss=0.1287, cr_loss=0.346, over 3353551.96 frames. ], batch size: 89, lr: 3.73e-03, grad_scale: 16.0 2024-09-24 19:20:47,503 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer_ff3.min_abs, batch_count=570672.6666666666, ans=0.2 2024-09-24 19:21:44,249 INFO [train.py:1198] (2/4) Epoch 32, batch 1550, loss[loss=0.1685, ctc_loss=0.1069, cr_loss=0.3081, over 17270.00 frames. ], tot_loss[loss=0.1983, ctc_loss=0.129, cr_loss=0.3467, over 3363085.02 frames. ], batch size: 42, lr: 3.73e-03, grad_scale: 16.0 2024-09-24 19:22:08,946 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=570906.0, ans=0.125 2024-09-24 19:22:13,840 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 19:22:31,853 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=12.19 vs. limit=15.0 2024-09-24 19:22:33,285 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=570999.3333333334, ans=0.125 2024-09-24 19:22:57,880 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.269e+02 1.362e+02 1.460e+02 2.798e+02, threshold=2.724e+02, percent-clipped=1.0 2024-09-24 19:23:07,636 INFO [train.py:1198] (2/4) Epoch 32, batch 1600, loss[loss=0.2229, ctc_loss=0.145, cr_loss=0.3892, over 17007.00 frames. ], tot_loss[loss=0.1978, ctc_loss=0.1286, cr_loss=0.3457, over 3366023.23 frames. ], batch size: 53, lr: 3.73e-03, grad_scale: 32.0 2024-09-24 19:23:53,000 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=571186.0, ans=0.0 2024-09-24 19:24:12,190 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=571279.3333333334, ans=0.125 2024-09-24 19:24:18,540 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=571279.3333333334, ans=0.125 2024-09-24 19:24:27,902 INFO [train.py:1198] (2/4) Epoch 32, batch 1650, loss[loss=0.2289, ctc_loss=0.1506, cr_loss=0.3916, over 17029.00 frames. ], tot_loss[loss=0.1975, ctc_loss=0.1284, cr_loss=0.3451, over 3364853.53 frames. ], batch size: 52, lr: 3.73e-03, grad_scale: 32.0 2024-09-24 19:25:15,115 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=571419.3333333334, ans=0.125 2024-09-24 19:25:15,264 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=571419.3333333334, ans=0.0 2024-09-24 19:25:34,293 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.23 vs. limit=15.0 2024-09-24 19:25:46,145 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.081e+02 1.279e+02 1.364e+02 1.457e+02 2.649e+02, threshold=2.728e+02, percent-clipped=0.0 2024-09-24 19:25:55,639 INFO [train.py:1198] (2/4) Epoch 32, batch 1700, loss[loss=0.216, ctc_loss=0.1418, cr_loss=0.3707, over 16959.00 frames. ], tot_loss[loss=0.1977, ctc_loss=0.1285, cr_loss=0.3459, over 3366780.15 frames. ], batch size: 58, lr: 3.73e-03, grad_scale: 32.0 2024-09-24 19:26:15,641 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=14.49 vs. limit=22.5 2024-09-24 19:26:34,248 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=571652.6666666666, ans=0.1 2024-09-24 19:26:43,953 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=571699.3333333334, ans=0.0 2024-09-24 19:26:59,957 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=571746.0, ans=0.125 2024-09-24 19:27:09,523 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 19:27:11,038 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=571746.0, ans=0.2 2024-09-24 19:27:15,594 INFO [train.py:1198] (2/4) Epoch 32, batch 1750, loss[loss=0.2111, ctc_loss=0.1379, cr_loss=0.366, over 17345.00 frames. ], tot_loss[loss=0.1988, ctc_loss=0.1293, cr_loss=0.3474, over 3363067.42 frames. ], batch size: 48, lr: 3.73e-03, grad_scale: 32.0 2024-09-24 19:27:51,966 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=571886.0, ans=0.1 2024-09-24 19:27:52,574 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.45 vs. limit=10.0 2024-09-24 19:28:11,088 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=571932.6666666666, ans=0.1 2024-09-24 19:28:26,328 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=8.98 vs. limit=10.0 2024-09-24 19:28:28,286 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.082e+02 1.257e+02 1.347e+02 1.444e+02 1.830e+02, threshold=2.693e+02, percent-clipped=0.0 2024-09-24 19:28:37,948 INFO [train.py:1198] (2/4) Epoch 32, batch 1800, loss[loss=0.238, ctc_loss=0.157, cr_loss=0.4051, over 16972.00 frames. ], tot_loss[loss=0.1993, ctc_loss=0.1297, cr_loss=0.3481, over 3364753.18 frames. ], batch size: 53, lr: 3.72e-03, grad_scale: 32.0 2024-09-24 19:28:51,434 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=6.68 vs. limit=15.0 2024-09-24 19:29:13,977 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.31 vs. limit=15.0 2024-09-24 19:29:20,293 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=572119.3333333334, ans=0.0 2024-09-24 19:29:25,031 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=572166.0, ans=0.2 2024-09-24 19:29:28,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=572166.0, ans=0.125 2024-09-24 19:29:50,661 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=572212.6666666666, ans=0.125 2024-09-24 19:29:51,204 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.97 vs. limit=15.0 2024-09-24 19:29:52,167 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=572212.6666666666, ans=0.125 2024-09-24 19:30:02,992 INFO [train.py:1198] (2/4) Epoch 32, batch 1850, loss[loss=0.24, ctc_loss=0.1677, cr_loss=0.3614, over 11630.00 frames. ], tot_loss[loss=0.1986, ctc_loss=0.1291, cr_loss=0.3473, over 3365188.39 frames. ], batch size: 123, lr: 3.72e-03, grad_scale: 32.0 2024-09-24 19:30:09,659 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=572259.3333333334, ans=0.025 2024-09-24 19:30:29,170 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.64 vs. limit=12.0 2024-09-24 19:30:48,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.min_abs, batch_count=572352.6666666666, ans=0.5 2024-09-24 19:31:09,762 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=572446.0, ans=0.125 2024-09-24 19:31:11,247 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=572446.0, ans=0.125 2024-09-24 19:31:13,029 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=572446.0, ans=0.0 2024-09-24 19:31:17,413 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.271e+02 1.345e+02 1.449e+02 2.207e+02, threshold=2.691e+02, percent-clipped=0.0 2024-09-24 19:31:25,383 INFO [train.py:1198] (2/4) Epoch 32, batch 1900, loss[loss=0.166, ctc_loss=0.1053, cr_loss=0.3035, over 17010.00 frames. ], tot_loss[loss=0.1991, ctc_loss=0.1294, cr_loss=0.3485, over 3371632.84 frames. ], batch size: 39, lr: 3.72e-03, grad_scale: 16.0 2024-09-24 19:31:27,626 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.26 vs. limit=22.5 2024-09-24 19:31:57,705 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.min_positive, batch_count=572586.0, ans=0.025 2024-09-24 19:32:20,163 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=572632.6666666666, ans=0.07 2024-09-24 19:32:29,825 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=572679.3333333334, ans=0.125 2024-09-24 19:32:47,861 INFO [train.py:1198] (2/4) Epoch 32, batch 1950, loss[loss=0.1576, ctc_loss=0.09892, cr_loss=0.2933, over 16397.00 frames. ], tot_loss[loss=0.198, ctc_loss=0.1287, cr_loss=0.3461, over 3362563.59 frames. ], batch size: 36, lr: 3.72e-03, grad_scale: 16.0 2024-09-24 19:33:07,323 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=572772.6666666666, ans=0.125 2024-09-24 19:33:32,847 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=572819.3333333334, ans=0.025 2024-09-24 19:33:36,051 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=572866.0, ans=0.0 2024-09-24 19:33:44,200 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=572866.0, ans=0.125 2024-09-24 19:33:50,856 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=572912.6666666666, ans=0.0 2024-09-24 19:33:57,424 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=572912.6666666666, ans=0.0 2024-09-24 19:34:00,277 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.147e+02 1.299e+02 1.379e+02 1.509e+02 1.987e+02, threshold=2.758e+02, percent-clipped=0.0 2024-09-24 19:34:08,189 INFO [train.py:1198] (2/4) Epoch 32, batch 2000, loss[loss=0.2479, ctc_loss=0.1704, cr_loss=0.3875, over 11704.00 frames. ], tot_loss[loss=0.1985, ctc_loss=0.1292, cr_loss=0.3465, over 3356753.95 frames. ], batch size: 123, lr: 3.72e-03, grad_scale: 32.0 2024-09-24 19:34:10,192 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 19:34:31,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=573006.0, ans=0.0 2024-09-24 19:34:57,691 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=573099.3333333334, ans=0.0 2024-09-24 19:35:03,017 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=7.14 vs. limit=15.0 2024-09-24 19:35:20,778 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=573146.0, ans=0.2 2024-09-24 19:35:33,440 INFO [train.py:1198] (2/4) Epoch 32, batch 2050, loss[loss=0.1882, ctc_loss=0.1207, cr_loss=0.3373, over 16888.00 frames. ], tot_loss[loss=0.1979, ctc_loss=0.1287, cr_loss=0.346, over 3361751.43 frames. ], batch size: 58, lr: 3.72e-03, grad_scale: 32.0 2024-09-24 19:35:49,870 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=573239.3333333334, ans=0.1 2024-09-24 19:36:42,818 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=573379.3333333334, ans=0.2 2024-09-24 19:36:42,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=573379.3333333334, ans=0.1 2024-09-24 19:36:45,681 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.089e+02 1.270e+02 1.344e+02 1.458e+02 2.417e+02, threshold=2.689e+02, percent-clipped=0.0 2024-09-24 19:36:52,496 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=573426.0, ans=0.125 2024-09-24 19:36:53,624 INFO [train.py:1198] (2/4) Epoch 32, batch 2100, loss[loss=0.1779, ctc_loss=0.1141, cr_loss=0.3188, over 17153.00 frames. ], tot_loss[loss=0.1985, ctc_loss=0.1292, cr_loss=0.3466, over 3353023.67 frames. ], batch size: 45, lr: 3.72e-03, grad_scale: 32.0 2024-09-24 19:36:55,538 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=573426.0, ans=10.0 2024-09-24 19:37:06,641 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=573426.0, ans=0.125 2024-09-24 19:37:26,187 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=512, metric=3.49 vs. limit=15.0 2024-09-24 19:37:28,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=573519.3333333334, ans=0.0 2024-09-24 19:37:32,799 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=3.79 vs. limit=15.0 2024-09-24 19:38:04,457 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=12.46 vs. limit=22.5 2024-09-24 19:38:05,117 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=573612.6666666666, ans=0.125 2024-09-24 19:38:15,960 INFO [train.py:1198] (2/4) Epoch 32, batch 2150, loss[loss=0.2207, ctc_loss=0.1453, cr_loss=0.3768, over 17164.00 frames. ], tot_loss[loss=0.1998, ctc_loss=0.1301, cr_loss=0.3485, over 3348064.79 frames. ], batch size: 45, lr: 3.72e-03, grad_scale: 32.0 2024-09-24 19:38:43,504 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=12.80 vs. limit=15.0 2024-09-24 19:38:44,888 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=573706.0, ans=0.125 2024-09-24 19:39:11,385 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=573799.3333333334, ans=0.125 2024-09-24 19:39:33,825 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.018e+02 1.275e+02 1.342e+02 1.446e+02 2.934e+02, threshold=2.683e+02, percent-clipped=1.0 2024-09-24 19:39:40,380 INFO [train.py:1198] (2/4) Epoch 32, batch 2200, loss[loss=0.2023, ctc_loss=0.1301, cr_loss=0.361, over 17307.00 frames. ], tot_loss[loss=0.1979, ctc_loss=0.1287, cr_loss=0.346, over 3349757.25 frames. ], batch size: 46, lr: 3.72e-03, grad_scale: 16.0 2024-09-24 19:39:42,143 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=573892.6666666666, ans=0.0 2024-09-24 19:40:16,620 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=573986.0, ans=0.1 2024-09-24 19:40:21,824 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=573986.0, ans=0.125 2024-09-24 19:40:51,169 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=574079.3333333334, ans=0.125 2024-09-24 19:40:57,550 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=574079.3333333334, ans=0.0 2024-09-24 19:40:59,168 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=574079.3333333334, ans=0.0 2024-09-24 19:41:03,500 INFO [train.py:1198] (2/4) Epoch 32, batch 2250, loss[loss=0.1634, ctc_loss=0.1029, cr_loss=0.3024, over 17088.00 frames. ], tot_loss[loss=0.1977, ctc_loss=0.1286, cr_loss=0.3454, over 3348300.08 frames. ], batch size: 40, lr: 3.72e-03, grad_scale: 16.0 2024-09-24 19:41:16,689 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=574126.0, ans=0.125 2024-09-24 19:41:26,216 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=574172.6666666666, ans=0.125 2024-09-24 19:41:39,541 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.38 vs. limit=6.0 2024-09-24 19:41:41,939 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=574219.3333333334, ans=0.0 2024-09-24 19:41:43,513 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=574219.3333333334, ans=0.125 2024-09-24 19:41:51,629 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=574266.0, ans=0.05 2024-09-24 19:42:07,673 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.08 vs. limit=15.0 2024-09-24 19:42:10,499 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=574312.6666666666, ans=0.125 2024-09-24 19:42:12,130 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.max_positive, batch_count=574312.6666666666, ans=0.95 2024-09-24 19:42:15,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=574312.6666666666, ans=0.0 2024-09-24 19:42:16,564 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.263e+02 1.326e+02 1.428e+02 1.960e+02, threshold=2.652e+02, percent-clipped=0.0 2024-09-24 19:42:22,902 INFO [train.py:1198] (2/4) Epoch 32, batch 2300, loss[loss=0.2153, ctc_loss=0.1408, cr_loss=0.3726, over 16490.00 frames. ], tot_loss[loss=0.198, ctc_loss=0.1288, cr_loss=0.346, over 3357329.81 frames. ], batch size: 66, lr: 3.72e-03, grad_scale: 16.0 2024-09-24 19:42:31,511 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=574359.3333333334, ans=0.025 2024-09-24 19:43:17,985 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.88 vs. limit=6.0 2024-09-24 19:43:44,365 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=574592.6666666666, ans=0.125 2024-09-24 19:43:45,787 INFO [train.py:1198] (2/4) Epoch 32, batch 2350, loss[loss=0.2243, ctc_loss=0.1446, cr_loss=0.3984, over 17046.00 frames. ], tot_loss[loss=0.1983, ctc_loss=0.129, cr_loss=0.3465, over 3355890.53 frames. ], batch size: 52, lr: 3.72e-03, grad_scale: 16.0 2024-09-24 19:43:47,725 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=574592.6666666666, ans=0.1 2024-09-24 19:44:10,305 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=574639.3333333334, ans=0.0 2024-09-24 19:44:36,332 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=574686.0, ans=0.125 2024-09-24 19:44:37,810 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=574732.6666666666, ans=0.0 2024-09-24 19:44:43,268 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=7.94 vs. limit=15.0 2024-09-24 19:44:47,430 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=574732.6666666666, ans=0.125 2024-09-24 19:45:00,124 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=574779.3333333334, ans=0.125 2024-09-24 19:45:04,053 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=5.90 vs. limit=15.0 2024-09-24 19:45:04,495 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.074e+02 1.259e+02 1.339e+02 1.484e+02 2.121e+02, threshold=2.678e+02, percent-clipped=0.0 2024-09-24 19:45:04,747 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=574779.3333333334, ans=0.0 2024-09-24 19:45:12,239 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=574826.0, ans=0.125 2024-09-24 19:45:13,560 INFO [train.py:1198] (2/4) Epoch 32, batch 2400, loss[loss=0.1921, ctc_loss=0.1267, cr_loss=0.3271, over 17202.00 frames. ], tot_loss[loss=0.1976, ctc_loss=0.1285, cr_loss=0.3456, over 3359132.50 frames. ], batch size: 47, lr: 3.72e-03, grad_scale: 32.0 2024-09-24 19:45:15,508 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=574826.0, ans=0.125 2024-09-24 19:45:50,740 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=574919.3333333334, ans=0.0 2024-09-24 19:46:33,599 INFO [train.py:1198] (2/4) Epoch 32, batch 2450, loss[loss=0.2285, ctc_loss=0.1513, cr_loss=0.3858, over 15940.00 frames. ], tot_loss[loss=0.1978, ctc_loss=0.1285, cr_loss=0.3465, over 3350351.99 frames. ], batch size: 74, lr: 3.72e-03, grad_scale: 32.0 2024-09-24 19:46:54,527 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=575106.0, ans=0.2 2024-09-24 19:47:01,132 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=575106.0, ans=0.2 2024-09-24 19:47:05,920 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=575152.6666666666, ans=0.025 2024-09-24 19:47:17,069 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=575152.6666666666, ans=0.1 2024-09-24 19:47:23,501 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=575199.3333333334, ans=0.125 2024-09-24 19:47:38,889 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-24 19:47:49,729 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.250e+02 1.318e+02 1.420e+02 2.123e+02, threshold=2.637e+02, percent-clipped=0.0 2024-09-24 19:47:53,335 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=575246.0, ans=0.1 2024-09-24 19:47:56,339 INFO [train.py:1198] (2/4) Epoch 32, batch 2500, loss[loss=0.1668, ctc_loss=0.1084, cr_loss=0.2918, over 17075.00 frames. ], tot_loss[loss=0.198, ctc_loss=0.1287, cr_loss=0.3464, over 3350889.40 frames. ], batch size: 46, lr: 3.71e-03, grad_scale: 32.0 2024-09-24 19:47:58,259 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=575292.6666666666, ans=0.09899494936611666 2024-09-24 19:48:09,512 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=575292.6666666666, ans=0.2 2024-09-24 19:48:09,546 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=575292.6666666666, ans=0.1 2024-09-24 19:48:30,337 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=575386.0, ans=0.2 2024-09-24 19:48:32,105 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=575386.0, ans=0.0 2024-09-24 19:49:18,663 INFO [train.py:1198] (2/4) Epoch 32, batch 2550, loss[loss=0.2104, ctc_loss=0.1377, cr_loss=0.3636, over 17134.00 frames. ], tot_loss[loss=0.1984, ctc_loss=0.129, cr_loss=0.3466, over 3336212.13 frames. ], batch size: 48, lr: 3.71e-03, grad_scale: 32.0 2024-09-24 19:49:39,211 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=575572.6666666666, ans=0.125 2024-09-24 19:50:20,442 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=575666.0, ans=0.0 2024-09-24 19:50:28,218 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=575712.6666666666, ans=0.125 2024-09-24 19:50:39,009 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.044e+02 1.280e+02 1.336e+02 1.446e+02 2.286e+02, threshold=2.672e+02, percent-clipped=0.0 2024-09-24 19:50:43,838 INFO [train.py:1198] (2/4) Epoch 32, batch 2600, loss[loss=0.1967, ctc_loss=0.1275, cr_loss=0.3458, over 17352.00 frames. ], tot_loss[loss=0.2, ctc_loss=0.1302, cr_loss=0.3487, over 3336723.43 frames. ], batch size: 48, lr: 3.71e-03, grad_scale: 16.0 2024-09-24 19:50:50,593 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=575759.3333333334, ans=0.1 2024-09-24 19:51:13,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=575806.0, ans=0.1 2024-09-24 19:51:51,506 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=7.36 vs. limit=12.0 2024-09-24 19:51:57,725 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=575946.0, ans=0.125 2024-09-24 19:52:03,679 INFO [train.py:1198] (2/4) Epoch 32, batch 2650, loss[loss=0.186, ctc_loss=0.1201, cr_loss=0.3295, over 17279.00 frames. ], tot_loss[loss=0.2002, ctc_loss=0.1304, cr_loss=0.3489, over 3350414.46 frames. ], batch size: 44, lr: 3.71e-03, grad_scale: 16.0 2024-09-24 19:52:07,321 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=575992.6666666666, ans=0.125 2024-09-24 19:52:13,683 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=575992.6666666666, ans=0.125 2024-09-24 19:52:13,767 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=575992.6666666666, ans=0.125 2024-09-24 19:52:20,427 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=8.43 vs. limit=15.0 2024-09-24 19:52:21,726 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=576039.3333333334, ans=0.125 2024-09-24 19:52:23,428 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=576039.3333333334, ans=0.1 2024-09-24 19:52:53,789 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=576132.6666666666, ans=0.125 2024-09-24 19:53:22,461 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.112e+02 1.282e+02 1.356e+02 1.452e+02 2.905e+02, threshold=2.711e+02, percent-clipped=3.0 2024-09-24 19:53:27,413 INFO [train.py:1198] (2/4) Epoch 32, batch 2700, loss[loss=0.1595, ctc_loss=0.1013, cr_loss=0.2911, over 17114.00 frames. ], tot_loss[loss=0.1982, ctc_loss=0.129, cr_loss=0.3461, over 3361026.02 frames. ], batch size: 40, lr: 3.71e-03, grad_scale: 16.0 2024-09-24 19:53:27,757 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=576226.0, ans=0.125 2024-09-24 19:53:32,538 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=576226.0, ans=0.125 2024-09-24 19:53:40,407 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=576226.0, ans=0.0 2024-09-24 19:53:47,020 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=576272.6666666666, ans=0.0 2024-09-24 19:53:48,536 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=576272.6666666666, ans=0.0 2024-09-24 19:53:59,799 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=576319.3333333334, ans=0.125 2024-09-24 19:54:07,967 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=576319.3333333334, ans=0.125 2024-09-24 19:54:20,332 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=576366.0, ans=0.125 2024-09-24 19:54:50,197 INFO [train.py:1198] (2/4) Epoch 32, batch 2750, loss[loss=0.1945, ctc_loss=0.1251, cr_loss=0.3468, over 17140.00 frames. ], tot_loss[loss=0.1974, ctc_loss=0.1283, cr_loss=0.3456, over 3366440.87 frames. ], batch size: 48, lr: 3.71e-03, grad_scale: 16.0 2024-09-24 19:54:50,953 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=13.60 vs. limit=15.0 2024-09-24 19:54:58,945 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.67 vs. limit=22.5 2024-09-24 19:55:41,124 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=576599.3333333334, ans=0.0 2024-09-24 19:55:50,781 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=576599.3333333334, ans=0.125 2024-09-24 19:55:50,904 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=576599.3333333334, ans=0.125 2024-09-24 19:55:52,910 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.84 vs. limit=15.0 2024-09-24 19:56:08,200 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.172e+02 1.294e+02 1.403e+02 1.500e+02 1.947e+02, threshold=2.805e+02, percent-clipped=0.0 2024-09-24 19:56:13,017 INFO [train.py:1198] (2/4) Epoch 32, batch 2800, loss[loss=0.227, ctc_loss=0.1523, cr_loss=0.3735, over 16594.00 frames. ], tot_loss[loss=0.1972, ctc_loss=0.1281, cr_loss=0.3452, over 3368803.87 frames. ], batch size: 66, lr: 3.71e-03, grad_scale: 32.0 2024-09-24 19:57:17,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=576879.3333333334, ans=0.0 2024-09-24 19:57:19,518 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=576879.3333333334, ans=0.1 2024-09-24 19:57:27,366 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=576879.3333333334, ans=0.1 2024-09-24 19:57:36,057 INFO [train.py:1198] (2/4) Epoch 32, batch 2850, loss[loss=0.2008, ctc_loss=0.1288, cr_loss=0.3598, over 17238.00 frames. ], tot_loss[loss=0.1971, ctc_loss=0.128, cr_loss=0.3454, over 3369215.39 frames. ], batch size: 44, lr: 3.71e-03, grad_scale: 32.0 2024-09-24 19:57:38,026 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=576926.0, ans=0.2 2024-09-24 19:57:42,775 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=576926.0, ans=0.125 2024-09-24 19:58:05,641 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=7.24 vs. limit=15.0 2024-09-24 19:58:06,773 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=577019.3333333334, ans=0.1 2024-09-24 19:58:14,701 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=577019.3333333334, ans=0.125 2024-09-24 19:58:45,043 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=577112.6666666666, ans=0.09899494936611666 2024-09-24 19:58:45,108 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=577112.6666666666, ans=0.125 2024-09-24 19:58:51,150 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.067e+02 1.250e+02 1.340e+02 1.459e+02 3.976e+02, threshold=2.680e+02, percent-clipped=1.0 2024-09-24 19:58:55,897 INFO [train.py:1198] (2/4) Epoch 32, batch 2900, loss[loss=0.1798, ctc_loss=0.114, cr_loss=0.3289, over 17206.00 frames. ], tot_loss[loss=0.1973, ctc_loss=0.1282, cr_loss=0.3456, over 3367452.75 frames. ], batch size: 47, lr: 3.71e-03, grad_scale: 32.0 2024-09-24 19:59:36,730 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=577252.6666666666, ans=0.125 2024-09-24 19:59:43,457 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.05 vs. limit=6.0 2024-09-24 19:59:54,432 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=577299.3333333334, ans=0.125 2024-09-24 20:00:08,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=577346.0, ans=0.035 2024-09-24 20:00:15,674 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=577346.0, ans=0.0 2024-09-24 20:00:20,979 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.out_whiten.whitening_limit, batch_count=577346.0, ans=15.0 2024-09-24 20:00:23,209 INFO [train.py:1198] (2/4) Epoch 32, batch 2950, loss[loss=0.2225, ctc_loss=0.1467, cr_loss=0.379, over 16874.00 frames. ], tot_loss[loss=0.1969, ctc_loss=0.1279, cr_loss=0.3448, over 3369550.23 frames. ], batch size: 58, lr: 3.71e-03, grad_scale: 32.0 2024-09-24 20:00:32,949 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=577392.6666666666, ans=0.125 2024-09-24 20:00:53,713 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=577486.0, ans=0.035 2024-09-24 20:01:39,276 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.060e+02 1.244e+02 1.340e+02 1.434e+02 1.849e+02, threshold=2.679e+02, percent-clipped=0.0 2024-09-24 20:01:39,564 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 20:01:42,520 INFO [train.py:1198] (2/4) Epoch 32, batch 3000, loss[loss=0.2376, ctc_loss=0.1583, cr_loss=0.3966, over 15191.00 frames. ], tot_loss[loss=0.1975, ctc_loss=0.1285, cr_loss=0.3453, over 3355128.43 frames. ], batch size: 89, lr: 3.71e-03, grad_scale: 16.0 2024-09-24 20:01:42,521 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 20:01:57,944 INFO [train.py:1230] (2/4) Epoch 32, validation: loss=0.03608, ctc_loss=0.03608, cr_loss=9.027e-15, over 944034.00 frames. 2024-09-24 20:01:57,945 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 20:02:15,593 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=577672.6666666666, ans=0.0 2024-09-24 20:02:34,690 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=577719.3333333334, ans=0.1 2024-09-24 20:02:39,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=577719.3333333334, ans=0.1 2024-09-24 20:02:49,089 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=577766.0, ans=0.0 2024-09-24 20:03:00,014 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=577812.6666666666, ans=0.0 2024-09-24 20:03:15,216 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 20:03:19,565 INFO [train.py:1198] (2/4) Epoch 32, batch 3050, loss[loss=0.2829, ctc_loss=0.1972, cr_loss=0.4282, over 11849.00 frames. ], tot_loss[loss=0.1969, ctc_loss=0.128, cr_loss=0.3445, over 3351570.99 frames. ], batch size: 123, lr: 3.71e-03, grad_scale: 16.0 2024-09-24 20:03:19,789 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=577859.3333333334, ans=0.125 2024-09-24 20:03:19,898 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 20:03:49,430 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=577952.6666666666, ans=0.125 2024-09-24 20:04:34,696 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.093e+02 1.228e+02 1.314e+02 1.465e+02 2.452e+02, threshold=2.627e+02, percent-clipped=0.0 2024-09-24 20:04:37,879 INFO [train.py:1198] (2/4) Epoch 32, batch 3100, loss[loss=0.1927, ctc_loss=0.123, cr_loss=0.3482, over 17177.00 frames. ], tot_loss[loss=0.1962, ctc_loss=0.1275, cr_loss=0.3436, over 3352960.25 frames. ], batch size: 45, lr: 3.71e-03, grad_scale: 16.0 2024-09-24 20:04:45,064 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.68 vs. limit=6.0 2024-09-24 20:04:47,569 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=578092.6666666666, ans=0.125 2024-09-24 20:04:49,440 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=11.48 vs. limit=12.0 2024-09-24 20:05:11,360 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=578186.0, ans=0.125 2024-09-24 20:05:16,812 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=4.10 vs. limit=15.0 2024-09-24 20:05:17,703 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=578186.0, ans=0.0 2024-09-24 20:05:30,217 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=578232.6666666666, ans=0.0 2024-09-24 20:05:34,831 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=578232.6666666666, ans=0.0 2024-09-24 20:05:37,047 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.whiten, num_groups=1, num_channels=192, metric=3.96 vs. limit=12.0 2024-09-24 20:05:56,195 INFO [train.py:1198] (2/4) Epoch 32, batch 3150, loss[loss=0.1772, ctc_loss=0.113, cr_loss=0.3212, over 17027.00 frames. ], tot_loss[loss=0.1959, ctc_loss=0.1273, cr_loss=0.3431, over 3353251.88 frames. ], batch size: 44, lr: 3.70e-03, grad_scale: 16.0 2024-09-24 20:05:58,190 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=578326.0, ans=0.0 2024-09-24 20:06:33,920 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=578419.3333333334, ans=0.125 2024-09-24 20:06:40,117 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=578419.3333333334, ans=0.125 2024-09-24 20:06:45,357 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 20:06:48,734 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.06 vs. limit=6.0 2024-09-24 20:06:50,022 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=578466.0, ans=0.0 2024-09-24 20:07:03,250 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=578512.6666666666, ans=0.125 2024-09-24 20:07:04,984 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=578512.6666666666, ans=0.1 2024-09-24 20:07:15,658 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.274e+02 1.414e+02 1.514e+02 2.013e+02, threshold=2.828e+02, percent-clipped=0.0 2024-09-24 20:07:18,909 INFO [train.py:1198] (2/4) Epoch 32, batch 3200, loss[loss=0.2069, ctc_loss=0.1374, cr_loss=0.3475, over 16996.00 frames. ], tot_loss[loss=0.1961, ctc_loss=0.1275, cr_loss=0.343, over 3355347.30 frames. ], batch size: 53, lr: 3.70e-03, grad_scale: 32.0 2024-09-24 20:07:30,116 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=578559.3333333334, ans=0.1 2024-09-24 20:07:43,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=578606.0, ans=0.125 2024-09-24 20:07:57,317 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=578652.6666666666, ans=0.2 2024-09-24 20:08:05,111 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=578652.6666666666, ans=0.125 2024-09-24 20:08:21,251 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=4.31 vs. limit=12.0 2024-09-24 20:08:39,238 INFO [train.py:1198] (2/4) Epoch 32, batch 3250, loss[loss=0.1895, ctc_loss=0.1231, cr_loss=0.332, over 17235.00 frames. ], tot_loss[loss=0.1959, ctc_loss=0.1274, cr_loss=0.3429, over 3361474.09 frames. ], batch size: 44, lr: 3.70e-03, grad_scale: 32.0 2024-09-24 20:09:00,406 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=578839.3333333334, ans=0.125 2024-09-24 20:09:52,220 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=578979.3333333334, ans=0.2 2024-09-24 20:09:54,517 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=5.37 vs. limit=15.0 2024-09-24 20:09:56,536 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.279e+02 1.356e+02 1.447e+02 1.834e+02, threshold=2.713e+02, percent-clipped=0.0 2024-09-24 20:09:59,775 INFO [train.py:1198] (2/4) Epoch 32, batch 3300, loss[loss=0.1914, ctc_loss=0.1239, cr_loss=0.3372, over 17296.00 frames. ], tot_loss[loss=0.1965, ctc_loss=0.1278, cr_loss=0.3435, over 3347195.08 frames. ], batch size: 51, lr: 3.70e-03, grad_scale: 32.0 2024-09-24 20:10:20,456 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 20:10:54,585 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=579166.0, ans=0.1 2024-09-24 20:11:02,395 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=579212.6666666666, ans=0.0 2024-09-24 20:11:17,458 INFO [train.py:1198] (2/4) Epoch 32, batch 3350, loss[loss=0.1538, ctc_loss=0.09607, cr_loss=0.2889, over 17037.00 frames. ], tot_loss[loss=0.198, ctc_loss=0.1289, cr_loss=0.3456, over 3352195.74 frames. ], batch size: 39, lr: 3.70e-03, grad_scale: 32.0 2024-09-24 20:11:33,200 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=579306.0, ans=0.125 2024-09-24 20:12:01,358 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=579352.6666666666, ans=0.125 2024-09-24 20:12:32,489 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.280e+02 1.354e+02 1.467e+02 2.390e+02, threshold=2.709e+02, percent-clipped=0.0 2024-09-24 20:12:35,625 INFO [train.py:1198] (2/4) Epoch 32, batch 3400, loss[loss=0.2162, ctc_loss=0.1415, cr_loss=0.3736, over 17236.00 frames. ], tot_loss[loss=0.1991, ctc_loss=0.1296, cr_loss=0.3474, over 3354626.58 frames. ], batch size: 55, lr: 3.70e-03, grad_scale: 32.0 2024-09-24 20:12:51,383 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=579539.3333333334, ans=0.1 2024-09-24 20:12:56,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=579539.3333333334, ans=0.0 2024-09-24 20:13:24,470 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=579632.6666666666, ans=0.04949747468305833 2024-09-24 20:13:32,920 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.06 vs. limit=15.0 2024-09-24 20:13:40,128 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=579679.3333333334, ans=0.125 2024-09-24 20:13:55,923 INFO [train.py:1198] (2/4) Epoch 32, batch 3450, loss[loss=0.1694, ctc_loss=0.1052, cr_loss=0.3212, over 17117.00 frames. ], tot_loss[loss=0.1988, ctc_loss=0.1293, cr_loss=0.3476, over 3352790.11 frames. ], batch size: 40, lr: 3.70e-03, grad_scale: 32.0 2024-09-24 20:14:23,907 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=579772.6666666666, ans=0.125 2024-09-24 20:15:07,688 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=579912.6666666666, ans=0.125 2024-09-24 20:15:10,635 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.167e+02 1.299e+02 1.417e+02 1.544e+02 2.266e+02, threshold=2.834e+02, percent-clipped=0.0 2024-09-24 20:15:13,759 INFO [train.py:1198] (2/4) Epoch 32, batch 3500, loss[loss=0.2349, ctc_loss=0.1538, cr_loss=0.4055, over 17001.00 frames. ], tot_loss[loss=0.1993, ctc_loss=0.1296, cr_loss=0.3483, over 3343681.62 frames. ], batch size: 56, lr: 3.70e-03, grad_scale: 32.0 2024-09-24 20:15:17,293 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=579959.3333333334, ans=0.125 2024-09-24 20:15:20,846 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.89 vs. limit=15.0 2024-09-24 20:15:22,339 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=3.80 vs. limit=10.0 2024-09-24 20:15:42,215 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=580006.0, ans=0.125 2024-09-24 20:16:15,869 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.45 vs. limit=15.0 2024-09-24 20:16:18,355 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=580146.0, ans=0.125 2024-09-24 20:16:27,863 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=580146.0, ans=0.0 2024-09-24 20:16:32,167 INFO [train.py:1198] (2/4) Epoch 32, batch 3550, loss[loss=0.1863, ctc_loss=0.1197, cr_loss=0.3331, over 17110.00 frames. ], tot_loss[loss=0.197, ctc_loss=0.1279, cr_loss=0.3452, over 3358453.72 frames. ], batch size: 40, lr: 3.70e-03, grad_scale: 32.0 2024-09-24 20:16:35,647 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=580192.6666666666, ans=0.1 2024-09-24 20:16:40,406 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=580192.6666666666, ans=0.125 2024-09-24 20:16:54,807 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=580239.3333333334, ans=0.125 2024-09-24 20:17:06,057 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=580286.0, ans=0.125 2024-09-24 20:17:06,177 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=580286.0, ans=0.125 2024-09-24 20:17:08,198 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.59 vs. limit=6.0 2024-09-24 20:17:29,951 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=580332.6666666666, ans=0.0 2024-09-24 20:17:33,211 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=580332.6666666666, ans=0.07 2024-09-24 20:17:51,751 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.152e+02 1.264e+02 1.345e+02 1.464e+02 2.321e+02, threshold=2.690e+02, percent-clipped=0.0 2024-09-24 20:17:53,768 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=580426.0, ans=0.2 2024-09-24 20:17:54,936 INFO [train.py:1198] (2/4) Epoch 32, batch 3600, loss[loss=0.1786, ctc_loss=0.1123, cr_loss=0.3314, over 17029.00 frames. ], tot_loss[loss=0.1974, ctc_loss=0.1283, cr_loss=0.3454, over 3362024.27 frames. ], batch size: 53, lr: 3.70e-03, grad_scale: 32.0 2024-09-24 20:17:56,761 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=580426.0, ans=0.1 2024-09-24 20:18:44,542 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=580566.0, ans=0.125 2024-09-24 20:18:50,740 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=580566.0, ans=0.025 2024-09-24 20:19:06,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=580612.6666666666, ans=0.125 2024-09-24 20:19:15,281 INFO [train.py:1198] (2/4) Epoch 32, batch 3650, loss[loss=0.2408, ctc_loss=0.1665, cr_loss=0.3714, over 11208.00 frames. ], tot_loss[loss=0.1973, ctc_loss=0.1283, cr_loss=0.3447, over 3352666.06 frames. ], batch size: 124, lr: 3.70e-03, grad_scale: 32.0 2024-09-24 20:19:56,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=580752.6666666666, ans=0.125 2024-09-24 20:20:02,154 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=580799.3333333334, ans=0.125 2024-09-24 20:20:02,264 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=580799.3333333334, ans=0.025 2024-09-24 20:20:07,082 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=580799.3333333334, ans=0.125 2024-09-24 20:20:08,381 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=580799.3333333334, ans=0.2 2024-09-24 20:20:17,706 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=580846.0, ans=0.0 2024-09-24 20:20:26,904 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=580846.0, ans=0.125 2024-09-24 20:20:30,685 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.112e+02 1.254e+02 1.367e+02 1.508e+02 1.700e+02, threshold=2.733e+02, percent-clipped=0.0 2024-09-24 20:20:33,724 INFO [train.py:1198] (2/4) Epoch 32, batch 3700, loss[loss=0.203, ctc_loss=0.1297, cr_loss=0.3663, over 17243.00 frames. ], tot_loss[loss=0.1974, ctc_loss=0.1284, cr_loss=0.345, over 3350034.44 frames. ], batch size: 44, lr: 3.70e-03, grad_scale: 32.0 2024-09-24 20:20:46,716 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=580892.6666666666, ans=0.125 2024-09-24 20:20:52,943 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer_ff3.min_abs, batch_count=580939.3333333334, ans=0.2 2024-09-24 20:20:59,503 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=580939.3333333334, ans=0.2 2024-09-24 20:21:16,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=580986.0, ans=0.1 2024-09-24 20:21:51,884 INFO [train.py:1198] (2/4) Epoch 32, batch 3750, loss[loss=0.2014, ctc_loss=0.1302, cr_loss=0.3559, over 17238.00 frames. ], tot_loss[loss=0.1979, ctc_loss=0.1287, cr_loss=0.3461, over 3344112.64 frames. ], batch size: 44, lr: 3.70e-03, grad_scale: 32.0 2024-09-24 20:22:37,487 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=581266.0, ans=0.0 2024-09-24 20:22:53,352 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 20:23:04,520 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=10.39 vs. limit=15.0 2024-09-24 20:23:06,879 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.283e+02 1.375e+02 1.482e+02 1.854e+02, threshold=2.749e+02, percent-clipped=0.0 2024-09-24 20:23:10,821 INFO [train.py:1198] (2/4) Epoch 32, batch 3800, loss[loss=0.1502, ctc_loss=0.09463, cr_loss=0.2777, over 17029.00 frames. ], tot_loss[loss=0.1963, ctc_loss=0.1276, cr_loss=0.3432, over 3319919.65 frames. ], batch size: 39, lr: 3.69e-03, grad_scale: 32.0 2024-09-24 20:23:31,045 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=581406.0, ans=0.2 2024-09-24 20:23:34,156 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=581406.0, ans=0.125 2024-09-24 20:23:38,771 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=581406.0, ans=0.125 2024-09-24 20:24:00,827 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.90 vs. limit=10.0 2024-09-24 20:24:01,939 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=581499.3333333334, ans=0.1 2024-09-24 20:24:06,541 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=581499.3333333334, ans=0.125 2024-09-24 20:24:09,673 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=581499.3333333334, ans=0.2 2024-09-24 20:24:11,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=581546.0, ans=0.0 2024-09-24 20:24:28,217 INFO [train.py:1198] (2/4) Epoch 32, batch 3850, loss[loss=0.2291, ctc_loss=0.1565, cr_loss=0.3632, over 11975.00 frames. ], tot_loss[loss=0.1999, ctc_loss=0.1306, cr_loss=0.3464, over 3264900.17 frames. ], batch size: 125, lr: 3.69e-03, grad_scale: 32.0 2024-09-24 20:24:47,112 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=581639.3333333334, ans=0.125 2024-09-24 20:24:50,073 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=581639.3333333334, ans=0.125 2024-09-24 20:24:51,764 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=581639.3333333334, ans=0.1 2024-09-24 20:24:55,542 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=9.82 vs. limit=22.5 2024-09-24 20:25:22,136 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=581732.6666666666, ans=0.125 2024-09-24 20:25:30,031 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=581779.3333333334, ans=0.125 2024-09-24 20:25:34,425 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=581779.3333333334, ans=0.125 2024-09-24 20:26:29,985 INFO [train.py:1198] (2/4) Epoch 33, batch 0, loss[loss=0.2191, ctc_loss=0.1424, cr_loss=0.3834, over 17045.00 frames. ], tot_loss[loss=0.2191, ctc_loss=0.1424, cr_loss=0.3834, over 17045.00 frames. ], batch size: 56, lr: 3.64e-03, grad_scale: 32.0 2024-09-24 20:26:29,986 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 20:26:46,725 INFO [train.py:1230] (2/4) Epoch 33, validation: loss=0.03608, ctc_loss=0.03608, cr_loss=9.001e-15, over 944034.00 frames. 2024-09-24 20:26:46,726 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 20:26:52,642 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.455e+02 1.559e+02 1.655e+02 2.375e+02, threshold=3.119e+02, percent-clipped=0.0 2024-09-24 20:27:42,764 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=581947.3333333334, ans=0.125 2024-09-24 20:27:44,346 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=581947.3333333334, ans=0.0 2024-09-24 20:27:55,428 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=581994.0, ans=0.125 2024-09-24 20:27:56,983 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=581994.0, ans=0.0 2024-09-24 20:28:04,861 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=581994.0, ans=0.2 2024-09-24 20:28:09,435 INFO [train.py:1198] (2/4) Epoch 33, batch 50, loss[loss=0.1593, ctc_loss=0.1012, cr_loss=0.2908, over 17088.00 frames. ], tot_loss[loss=0.1976, ctc_loss=0.1288, cr_loss=0.3444, over 744401.45 frames. ], batch size: 43, lr: 3.64e-03, grad_scale: 32.0 2024-09-24 20:28:19,420 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=582040.6666666666, ans=0.125 2024-09-24 20:28:24,106 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=582087.3333333334, ans=0.125 2024-09-24 20:28:28,747 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=582087.3333333334, ans=0.125 2024-09-24 20:29:06,656 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=582180.6666666666, ans=0.1 2024-09-24 20:29:14,418 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=582227.3333333334, ans=0.125 2024-09-24 20:29:17,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer_ff3.min_abs, batch_count=582227.3333333334, ans=0.2 2024-09-24 20:29:31,539 INFO [train.py:1198] (2/4) Epoch 33, batch 100, loss[loss=0.2301, ctc_loss=0.1606, cr_loss=0.3472, over 12053.00 frames. ], tot_loss[loss=0.2004, ctc_loss=0.1308, cr_loss=0.3483, over 1312811.15 frames. ], batch size: 123, lr: 3.63e-03, grad_scale: 32.0 2024-09-24 20:29:34,671 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.274e+02 1.348e+02 1.462e+02 2.671e+02, threshold=2.696e+02, percent-clipped=0.0 2024-09-24 20:29:47,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=582320.6666666666, ans=0.0 2024-09-24 20:30:15,493 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=582367.3333333334, ans=0.0 2024-09-24 20:30:20,213 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=582414.0, ans=0.025 2024-09-24 20:30:54,479 INFO [train.py:1198] (2/4) Epoch 33, batch 150, loss[loss=0.2, ctc_loss=0.1319, cr_loss=0.3406, over 16151.00 frames. ], tot_loss[loss=0.201, ctc_loss=0.1311, cr_loss=0.3495, over 1764198.38 frames. ], batch size: 74, lr: 3.63e-03, grad_scale: 32.0 2024-09-24 20:31:03,012 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=582507.3333333334, ans=0.125 2024-09-24 20:31:10,760 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=582554.0, ans=0.125 2024-09-24 20:31:25,376 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1.whitening_limit, batch_count=582600.6666666666, ans=10.0 2024-09-24 20:31:59,919 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=582647.3333333334, ans=0.1 2024-09-24 20:32:06,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=582694.0, ans=0.125 2024-09-24 20:32:20,259 INFO [train.py:1198] (2/4) Epoch 33, batch 200, loss[loss=0.2238, ctc_loss=0.1473, cr_loss=0.3826, over 17013.00 frames. ], tot_loss[loss=0.2006, ctc_loss=0.1307, cr_loss=0.3494, over 2121038.88 frames. ], batch size: 56, lr: 3.63e-03, grad_scale: 32.0 2024-09-24 20:32:20,538 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=582740.6666666666, ans=0.05 2024-09-24 20:32:20,600 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=582740.6666666666, ans=0.125 2024-09-24 20:32:23,423 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.261e+02 1.350e+02 1.462e+02 5.443e+02, threshold=2.700e+02, percent-clipped=2.0 2024-09-24 20:32:44,670 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=582787.3333333334, ans=0.0 2024-09-24 20:32:50,920 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=582834.0, ans=0.125 2024-09-24 20:32:52,495 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=582834.0, ans=0.0 2024-09-24 20:32:54,301 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-24 20:32:55,702 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=582834.0, ans=0.125 2024-09-24 20:33:13,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=582880.6666666666, ans=0.0 2024-09-24 20:33:42,545 INFO [train.py:1198] (2/4) Epoch 33, batch 250, loss[loss=0.1811, ctc_loss=0.1175, cr_loss=0.3181, over 16958.00 frames. ], tot_loss[loss=0.1984, ctc_loss=0.129, cr_loss=0.3469, over 2396486.43 frames. ], batch size: 42, lr: 3.63e-03, grad_scale: 32.0 2024-09-24 20:33:55,268 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=582974.0, ans=0.125 2024-09-24 20:34:01,678 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=583020.6666666666, ans=0.125 2024-09-24 20:34:06,699 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.33 vs. limit=10.0 2024-09-24 20:34:11,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.min_positive, batch_count=583020.6666666666, ans=0.025 2024-09-24 20:34:55,509 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=583160.6666666666, ans=0.1 2024-09-24 20:35:01,848 INFO [train.py:1198] (2/4) Epoch 33, batch 300, loss[loss=0.1841, ctc_loss=0.1202, cr_loss=0.3197, over 17149.00 frames. ], tot_loss[loss=0.1975, ctc_loss=0.1284, cr_loss=0.3456, over 2612049.05 frames. ], batch size: 48, lr: 3.63e-03, grad_scale: 32.0 2024-09-24 20:35:05,041 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.095e+02 1.274e+02 1.352e+02 1.473e+02 1.783e+02, threshold=2.703e+02, percent-clipped=0.0 2024-09-24 20:35:14,958 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=583207.3333333334, ans=0.0 2024-09-24 20:35:38,716 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=583300.6666666666, ans=0.0 2024-09-24 20:35:41,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=583300.6666666666, ans=0.0 2024-09-24 20:35:56,526 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=583347.3333333334, ans=0.2 2024-09-24 20:35:56,648 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=583347.3333333334, ans=0.0 2024-09-24 20:36:25,026 INFO [train.py:1198] (2/4) Epoch 33, batch 350, loss[loss=0.1809, ctc_loss=0.1151, cr_loss=0.3291, over 17065.00 frames. ], tot_loss[loss=0.1966, ctc_loss=0.1277, cr_loss=0.3445, over 2774047.94 frames. ], batch size: 39, lr: 3.63e-03, grad_scale: 32.0 2024-09-24 20:36:57,050 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=10.78 vs. limit=15.0 2024-09-24 20:36:59,429 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=583487.3333333334, ans=0.125 2024-09-24 20:37:00,154 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=6.09 vs. limit=15.0 2024-09-24 20:37:13,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=583534.0, ans=0.1 2024-09-24 20:37:42,689 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=583627.3333333334, ans=0.2 2024-09-24 20:37:47,736 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=10.12 vs. limit=22.5 2024-09-24 20:37:49,120 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=583674.0, ans=0.2 2024-09-24 20:37:50,367 INFO [train.py:1198] (2/4) Epoch 33, batch 400, loss[loss=0.2079, ctc_loss=0.1327, cr_loss=0.3761, over 17282.00 frames. ], tot_loss[loss=0.1959, ctc_loss=0.1271, cr_loss=0.3439, over 2903356.80 frames. ], batch size: 46, lr: 3.63e-03, grad_scale: 32.0 2024-09-24 20:37:50,608 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=583674.0, ans=0.1 2024-09-24 20:37:53,545 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.274e+02 1.352e+02 1.470e+02 2.470e+02, threshold=2.705e+02, percent-clipped=0.0 2024-09-24 20:37:56,978 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=583674.0, ans=10.0 2024-09-24 20:38:01,840 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=583674.0, ans=0.125 2024-09-24 20:38:08,041 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=583720.6666666666, ans=0.125 2024-09-24 20:38:11,246 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=583720.6666666666, ans=0.0 2024-09-24 20:38:14,500 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 20:38:50,612 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=583814.0, ans=0.0 2024-09-24 20:38:54,421 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=5.73 vs. limit=15.0 2024-09-24 20:38:58,500 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=583860.6666666666, ans=0.0 2024-09-24 20:39:05,485 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=5.59 vs. limit=15.0 2024-09-24 20:39:12,683 INFO [train.py:1198] (2/4) Epoch 33, batch 450, loss[loss=0.1919, ctc_loss=0.126, cr_loss=0.3292, over 16677.00 frames. ], tot_loss[loss=0.1964, ctc_loss=0.1276, cr_loss=0.3443, over 3005488.43 frames. ], batch size: 61, lr: 3.63e-03, grad_scale: 32.0 2024-09-24 20:39:17,848 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=583907.3333333334, ans=0.125 2024-09-24 20:39:57,483 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=584000.6666666666, ans=0.2 2024-09-24 20:40:31,385 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=5.12 vs. limit=15.0 2024-09-24 20:40:35,213 INFO [train.py:1198] (2/4) Epoch 33, batch 500, loss[loss=0.1618, ctc_loss=0.1028, cr_loss=0.2948, over 17177.00 frames. ], tot_loss[loss=0.1975, ctc_loss=0.1283, cr_loss=0.3459, over 3086968.10 frames. ], batch size: 41, lr: 3.63e-03, grad_scale: 32.0 2024-09-24 20:40:37,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=584140.6666666666, ans=0.2 2024-09-24 20:40:38,401 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.093e+02 1.260e+02 1.369e+02 1.441e+02 1.904e+02, threshold=2.739e+02, percent-clipped=0.0 2024-09-24 20:40:49,800 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=584187.3333333334, ans=0.0 2024-09-24 20:41:02,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=584187.3333333334, ans=0.125 2024-09-24 20:41:05,516 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=584234.0, ans=0.1 2024-09-24 20:41:27,957 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=584280.6666666666, ans=0.1 2024-09-24 20:41:48,047 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=584327.3333333334, ans=0.125 2024-09-24 20:41:49,696 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=584327.3333333334, ans=0.125 2024-09-24 20:41:59,485 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=584374.0, ans=0.2 2024-09-24 20:42:00,692 INFO [train.py:1198] (2/4) Epoch 33, batch 550, loss[loss=0.2048, ctc_loss=0.1324, cr_loss=0.3619, over 17019.00 frames. ], tot_loss[loss=0.1974, ctc_loss=0.1282, cr_loss=0.3461, over 3149463.71 frames. ], batch size: 56, lr: 3.63e-03, grad_scale: 32.0 2024-09-24 20:42:03,301 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=4.91 vs. limit=10.0 2024-09-24 20:42:15,283 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=584420.6666666666, ans=0.125 2024-09-24 20:42:33,144 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=584467.3333333334, ans=0.125 2024-09-24 20:42:33,278 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=584467.3333333334, ans=0.0 2024-09-24 20:42:41,249 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=584467.3333333334, ans=0.125 2024-09-24 20:42:45,946 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=584467.3333333334, ans=0.125 2024-09-24 20:42:47,807 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=584514.0, ans=0.2 2024-09-24 20:42:48,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten.whitening_limit, batch_count=584514.0, ans=22.5 2024-09-24 20:43:03,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=584560.6666666666, ans=0.125 2024-09-24 20:43:11,670 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=584560.6666666666, ans=0.125 2024-09-24 20:43:14,886 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=584560.6666666666, ans=0.025 2024-09-24 20:43:20,875 INFO [train.py:1198] (2/4) Epoch 33, batch 600, loss[loss=0.199, ctc_loss=0.1322, cr_loss=0.334, over 17094.00 frames. ], tot_loss[loss=0.1969, ctc_loss=0.1277, cr_loss=0.3457, over 3193986.31 frames. ], batch size: 49, lr: 3.63e-03, grad_scale: 32.0 2024-09-24 20:43:26,765 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.082e+02 1.269e+02 1.358e+02 1.468e+02 2.109e+02, threshold=2.715e+02, percent-clipped=0.0 2024-09-24 20:43:33,579 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=584607.3333333334, ans=0.2 2024-09-24 20:44:02,725 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=584700.6666666666, ans=0.025 2024-09-24 20:44:12,430 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=584747.3333333334, ans=0.125 2024-09-24 20:44:25,918 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=12.40 vs. limit=15.0 2024-09-24 20:44:26,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=584794.0, ans=0.0 2024-09-24 20:44:31,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=584794.0, ans=0.0 2024-09-24 20:44:40,315 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.35 vs. limit=15.0 2024-09-24 20:44:43,686 INFO [train.py:1198] (2/4) Epoch 33, batch 650, loss[loss=0.2272, ctc_loss=0.1511, cr_loss=0.3804, over 15990.00 frames. ], tot_loss[loss=0.1974, ctc_loss=0.1282, cr_loss=0.3464, over 3231428.12 frames. ], batch size: 74, lr: 3.63e-03, grad_scale: 32.0 2024-09-24 20:44:51,968 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=584840.6666666666, ans=0.0 2024-09-24 20:45:47,358 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=584980.6666666666, ans=0.125 2024-09-24 20:45:52,205 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=585027.3333333334, ans=0.125 2024-09-24 20:45:58,545 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=585027.3333333334, ans=0.025 2024-09-24 20:45:58,872 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2.whitening_limit, batch_count=585027.3333333334, ans=15.0 2024-09-24 20:46:06,204 INFO [train.py:1198] (2/4) Epoch 33, batch 700, loss[loss=0.2074, ctc_loss=0.1372, cr_loss=0.3509, over 17170.00 frames. ], tot_loss[loss=0.1979, ctc_loss=0.1285, cr_loss=0.3469, over 3255583.28 frames. ], batch size: 45, lr: 3.63e-03, grad_scale: 16.0 2024-09-24 20:46:10,908 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.093e+02 1.247e+02 1.323e+02 1.439e+02 2.225e+02, threshold=2.645e+02, percent-clipped=0.0 2024-09-24 20:46:47,103 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=585167.3333333334, ans=0.125 2024-09-24 20:46:53,664 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=13.19 vs. limit=22.5 2024-09-24 20:47:01,461 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=585214.0, ans=0.125 2024-09-24 20:47:03,876 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.18 vs. limit=22.5 2024-09-24 20:47:04,673 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=585214.0, ans=0.125 2024-09-24 20:47:08,142 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=585214.0, ans=0.0 2024-09-24 20:47:24,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=585260.6666666666, ans=0.0 2024-09-24 20:47:31,641 INFO [train.py:1198] (2/4) Epoch 33, batch 750, loss[loss=0.2007, ctc_loss=0.1313, cr_loss=0.3474, over 17288.00 frames. ], tot_loss[loss=0.1976, ctc_loss=0.1284, cr_loss=0.3464, over 3276151.44 frames. ], batch size: 49, lr: 3.63e-03, grad_scale: 16.0 2024-09-24 20:47:35,312 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=585307.3333333334, ans=0.125 2024-09-24 20:47:44,696 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=585307.3333333334, ans=0.125 2024-09-24 20:47:49,359 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=585354.0, ans=0.125 2024-09-24 20:48:21,108 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=585447.3333333334, ans=0.0 2024-09-24 20:48:52,591 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=585540.6666666666, ans=0.125 2024-09-24 20:48:53,865 INFO [train.py:1198] (2/4) Epoch 33, batch 800, loss[loss=0.2115, ctc_loss=0.1374, cr_loss=0.3704, over 17070.00 frames. ], tot_loss[loss=0.1971, ctc_loss=0.128, cr_loss=0.3453, over 3285839.77 frames. ], batch size: 46, lr: 3.62e-03, grad_scale: 32.0 2024-09-24 20:48:58,602 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.132e+02 1.263e+02 1.335e+02 1.404e+02 2.398e+02, threshold=2.669e+02, percent-clipped=0.0 2024-09-24 20:49:21,290 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=585587.3333333334, ans=0.0 2024-09-24 20:49:43,699 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=585680.6666666666, ans=0.025 2024-09-24 20:50:01,458 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=585727.3333333334, ans=0.1 2024-09-24 20:50:14,050 INFO [train.py:1198] (2/4) Epoch 33, batch 850, loss[loss=0.2221, ctc_loss=0.1445, cr_loss=0.3879, over 17206.00 frames. ], tot_loss[loss=0.1974, ctc_loss=0.1282, cr_loss=0.3461, over 3304676.00 frames. ], batch size: 47, lr: 3.62e-03, grad_scale: 32.0 2024-09-24 20:50:33,512 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=585820.6666666666, ans=0.2 2024-09-24 20:51:39,162 INFO [train.py:1198] (2/4) Epoch 33, batch 900, loss[loss=0.2507, ctc_loss=0.1669, cr_loss=0.419, over 14902.00 frames. ], tot_loss[loss=0.1975, ctc_loss=0.1283, cr_loss=0.346, over 3320602.09 frames. ], batch size: 89, lr: 3.62e-03, grad_scale: 32.0 2024-09-24 20:51:43,303 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.02 vs. limit=6.0 2024-09-24 20:51:44,052 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.262e+02 1.327e+02 1.436e+02 1.975e+02, threshold=2.653e+02, percent-clipped=0.0 2024-09-24 20:51:44,405 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=586007.3333333334, ans=0.0 2024-09-24 20:52:00,060 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=586054.0, ans=0.125 2024-09-24 20:52:15,901 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=586100.6666666666, ans=0.125 2024-09-24 20:52:24,691 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=13.11 vs. limit=22.5 2024-09-24 20:52:38,823 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=586147.3333333334, ans=0.125 2024-09-24 20:52:45,393 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=586194.0, ans=0.125 2024-09-24 20:52:46,810 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=586194.0, ans=0.125 2024-09-24 20:52:55,863 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=13.04 vs. limit=22.5 2024-09-24 20:52:59,224 INFO [train.py:1198] (2/4) Epoch 33, batch 950, loss[loss=0.1889, ctc_loss=0.1214, cr_loss=0.3376, over 17148.00 frames. ], tot_loss[loss=0.198, ctc_loss=0.1286, cr_loss=0.3468, over 3328053.63 frames. ], batch size: 45, lr: 3.62e-03, grad_scale: 32.0 2024-09-24 20:53:05,729 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=586240.6666666666, ans=0.125 2024-09-24 20:53:34,606 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=15.57 vs. limit=22.5 2024-09-24 20:53:45,104 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=586334.0, ans=0.125 2024-09-24 20:54:08,936 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=586427.3333333334, ans=0.0 2024-09-24 20:54:17,063 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=8.96 vs. limit=15.0 2024-09-24 20:54:17,421 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.04 vs. limit=15.0 2024-09-24 20:54:21,207 INFO [train.py:1198] (2/4) Epoch 33, batch 1000, loss[loss=0.2057, ctc_loss=0.1351, cr_loss=0.3531, over 17234.00 frames. ], tot_loss[loss=0.1983, ctc_loss=0.1288, cr_loss=0.3477, over 3339599.67 frames. ], batch size: 50, lr: 3.62e-03, grad_scale: 32.0 2024-09-24 20:54:26,138 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.300e+02 1.389e+02 1.469e+02 2.721e+02, threshold=2.778e+02, percent-clipped=0.0 2024-09-24 20:54:31,231 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=586474.0, ans=0.125 2024-09-24 20:55:02,318 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.84 vs. limit=22.5 2024-09-24 20:55:08,357 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=586614.0, ans=0.025 2024-09-24 20:55:14,782 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=586614.0, ans=0.0 2024-09-24 20:55:39,981 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=586660.6666666666, ans=0.125 2024-09-24 20:55:44,467 INFO [train.py:1198] (2/4) Epoch 33, batch 1050, loss[loss=0.1963, ctc_loss=0.1247, cr_loss=0.3578, over 17310.00 frames. ], tot_loss[loss=0.1979, ctc_loss=0.1286, cr_loss=0.3468, over 3341400.88 frames. ], batch size: 51, lr: 3.62e-03, grad_scale: 32.0 2024-09-24 20:55:44,873 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=586707.3333333334, ans=0.1 2024-09-24 20:56:08,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=586754.0, ans=0.1 2024-09-24 20:56:10,079 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=586754.0, ans=0.2 2024-09-24 20:56:32,516 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=586800.6666666666, ans=0.0 2024-09-24 20:56:35,764 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=586847.3333333334, ans=0.0 2024-09-24 20:56:37,224 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=586847.3333333334, ans=0.1 2024-09-24 20:56:42,737 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=4.01 vs. limit=15.0 2024-09-24 20:56:50,555 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.39 vs. limit=12.0 2024-09-24 20:56:55,075 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=586894.0, ans=0.125 2024-09-24 20:57:09,128 INFO [train.py:1198] (2/4) Epoch 33, batch 1100, loss[loss=0.1982, ctc_loss=0.1263, cr_loss=0.3598, over 17288.00 frames. ], tot_loss[loss=0.1985, ctc_loss=0.129, cr_loss=0.3478, over 3346169.22 frames. ], batch size: 51, lr: 3.62e-03, grad_scale: 32.0 2024-09-24 20:57:13,958 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.249e+02 1.328e+02 1.446e+02 1.774e+02, threshold=2.656e+02, percent-clipped=0.0 2024-09-24 20:57:20,728 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=586940.6666666666, ans=0.0 2024-09-24 20:57:33,639 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=586987.3333333334, ans=0.1 2024-09-24 20:57:41,963 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.whiten, num_groups=1, num_channels=512, metric=4.66 vs. limit=12.0 2024-09-24 20:57:43,183 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=587034.0, ans=0.0 2024-09-24 20:58:11,121 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=512, metric=4.06 vs. limit=15.0 2024-09-24 20:58:32,050 INFO [train.py:1198] (2/4) Epoch 33, batch 1150, loss[loss=0.2055, ctc_loss=0.1339, cr_loss=0.3581, over 16983.00 frames. ], tot_loss[loss=0.1995, ctc_loss=0.1297, cr_loss=0.3489, over 3344695.73 frames. ], batch size: 53, lr: 3.62e-03, grad_scale: 32.0 2024-09-24 20:58:32,271 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=587174.0, ans=0.125 2024-09-24 20:58:34,019 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=587174.0, ans=0.0 2024-09-24 20:58:53,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=587220.6666666666, ans=0.125 2024-09-24 20:58:56,593 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=11.13 vs. limit=15.0 2024-09-24 20:58:57,874 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=587220.6666666666, ans=0.2 2024-09-24 20:58:57,973 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=587220.6666666666, ans=0.125 2024-09-24 20:59:51,714 INFO [train.py:1198] (2/4) Epoch 33, batch 1200, loss[loss=0.2014, ctc_loss=0.1283, cr_loss=0.3655, over 17031.00 frames. ], tot_loss[loss=0.1986, ctc_loss=0.129, cr_loss=0.3482, over 3350292.44 frames. ], batch size: 39, lr: 3.62e-03, grad_scale: 32.0 2024-09-24 20:59:56,479 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.281e+02 1.363e+02 1.447e+02 2.223e+02, threshold=2.727e+02, percent-clipped=0.0 2024-09-24 21:00:01,440 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=587407.3333333334, ans=0.125 2024-09-24 21:00:06,183 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=587454.0, ans=0.125 2024-09-24 21:00:43,914 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=587547.3333333334, ans=0.2 2024-09-24 21:01:13,623 INFO [train.py:1198] (2/4) Epoch 33, batch 1250, loss[loss=0.2369, ctc_loss=0.1595, cr_loss=0.387, over 12270.00 frames. ], tot_loss[loss=0.1987, ctc_loss=0.1292, cr_loss=0.3476, over 3328768.65 frames. ], batch size: 124, lr: 3.62e-03, grad_scale: 32.0 2024-09-24 21:01:50,911 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=587734.0, ans=0.0 2024-09-24 21:01:54,349 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=11.07 vs. limit=15.0 2024-09-24 21:02:11,499 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=587780.6666666666, ans=0.07 2024-09-24 21:02:16,333 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=587780.6666666666, ans=0.025 2024-09-24 21:02:35,633 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=587827.3333333334, ans=0.125 2024-09-24 21:02:38,674 INFO [train.py:1198] (2/4) Epoch 33, batch 1300, loss[loss=0.1937, ctc_loss=0.1263, cr_loss=0.337, over 17286.00 frames. ], tot_loss[loss=0.1982, ctc_loss=0.1288, cr_loss=0.3468, over 3327416.48 frames. ], batch size: 49, lr: 3.62e-03, grad_scale: 32.0 2024-09-24 21:02:43,488 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.064e+02 1.271e+02 1.359e+02 1.460e+02 2.870e+02, threshold=2.719e+02, percent-clipped=1.0 2024-09-24 21:02:50,076 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=587874.0, ans=0.125 2024-09-24 21:03:09,918 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=587920.6666666666, ans=0.0 2024-09-24 21:03:21,136 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=587967.3333333334, ans=0.2 2024-09-24 21:03:32,825 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 21:03:40,906 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=11.80 vs. limit=22.5 2024-09-24 21:03:45,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=588060.6666666666, ans=0.125 2024-09-24 21:04:00,912 INFO [train.py:1198] (2/4) Epoch 33, batch 1350, loss[loss=0.1882, ctc_loss=0.1224, cr_loss=0.3291, over 17229.00 frames. ], tot_loss[loss=0.1961, ctc_loss=0.1274, cr_loss=0.3436, over 3341449.13 frames. ], batch size: 50, lr: 3.62e-03, grad_scale: 32.0 2024-09-24 21:04:08,013 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=13.76 vs. limit=15.0 2024-09-24 21:04:11,300 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=14.17 vs. limit=15.0 2024-09-24 21:04:17,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=588154.0, ans=0.1 2024-09-24 21:04:18,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=588154.0, ans=0.1 2024-09-24 21:04:24,194 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=5.29 vs. limit=15.0 2024-09-24 21:04:38,797 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=9.68 vs. limit=22.5 2024-09-24 21:04:39,957 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=588200.6666666666, ans=0.0 2024-09-24 21:05:01,488 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=3.85 vs. limit=15.0 2024-09-24 21:05:21,444 INFO [train.py:1198] (2/4) Epoch 33, batch 1400, loss[loss=0.2098, ctc_loss=0.1384, cr_loss=0.357, over 16865.00 frames. ], tot_loss[loss=0.1959, ctc_loss=0.1271, cr_loss=0.3436, over 3343321.71 frames. ], batch size: 58, lr: 3.62e-03, grad_scale: 32.0 2024-09-24 21:05:26,392 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.268e+02 1.325e+02 1.452e+02 1.895e+02, threshold=2.651e+02, percent-clipped=0.0 2024-09-24 21:05:38,810 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=588387.3333333334, ans=0.125 2024-09-24 21:05:45,242 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=588387.3333333334, ans=0.2 2024-09-24 21:05:45,606 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=14.44 vs. limit=15.0 2024-09-24 21:05:51,595 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=588387.3333333334, ans=0.125 2024-09-24 21:05:53,345 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.88 vs. limit=10.0 2024-09-24 21:05:56,179 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=588434.0, ans=0.125 2024-09-24 21:06:02,598 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=588434.0, ans=0.2 2024-09-24 21:06:44,051 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=6.18 vs. limit=15.0 2024-09-24 21:06:44,962 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=588527.3333333334, ans=0.0 2024-09-24 21:06:49,513 INFO [train.py:1198] (2/4) Epoch 33, batch 1450, loss[loss=0.1801, ctc_loss=0.1129, cr_loss=0.336, over 15872.00 frames. ], tot_loss[loss=0.1957, ctc_loss=0.127, cr_loss=0.3433, over 3344392.15 frames. ], batch size: 35, lr: 3.62e-03, grad_scale: 32.0 2024-09-24 21:06:54,580 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=588574.0, ans=0.0 2024-09-24 21:06:57,791 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=588574.0, ans=0.125 2024-09-24 21:07:02,491 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=588574.0, ans=0.1 2024-09-24 21:07:17,557 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.59 vs. limit=12.0 2024-09-24 21:07:21,954 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=588667.3333333334, ans=0.125 2024-09-24 21:07:44,226 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=588714.0, ans=0.025 2024-09-24 21:07:55,528 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=588760.6666666666, ans=0.0 2024-09-24 21:07:58,750 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=588760.6666666666, ans=0.1 2024-09-24 21:07:59,402 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.78 vs. limit=6.0 2024-09-24 21:08:02,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten.whitening_limit, batch_count=588760.6666666666, ans=22.5 2024-09-24 21:08:04,062 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=9.29 vs. limit=15.0 2024-09-24 21:08:08,547 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=588807.3333333334, ans=0.025 2024-09-24 21:08:12,481 INFO [train.py:1198] (2/4) Epoch 33, batch 1500, loss[loss=0.2242, ctc_loss=0.1447, cr_loss=0.3975, over 17021.00 frames. ], tot_loss[loss=0.1966, ctc_loss=0.1278, cr_loss=0.3444, over 3343074.38 frames. ], batch size: 53, lr: 3.61e-03, grad_scale: 32.0 2024-09-24 21:08:12,689 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=588807.3333333334, ans=0.0 2024-09-24 21:08:16,414 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=13.84 vs. limit=22.5 2024-09-24 21:08:16,538 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.13 vs. limit=15.0 2024-09-24 21:08:17,327 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.121e+02 1.288e+02 1.351e+02 1.447e+02 2.720e+02, threshold=2.702e+02, percent-clipped=1.0 2024-09-24 21:08:33,798 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=588854.0, ans=0.1 2024-09-24 21:09:12,901 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.82 vs. limit=10.0 2024-09-24 21:09:21,755 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=588994.0, ans=0.2 2024-09-24 21:09:32,673 INFO [train.py:1198] (2/4) Epoch 33, batch 1550, loss[loss=0.1619, ctc_loss=0.1015, cr_loss=0.302, over 16946.00 frames. ], tot_loss[loss=0.1966, ctc_loss=0.1278, cr_loss=0.3441, over 3350905.42 frames. ], batch size: 42, lr: 3.61e-03, grad_scale: 32.0 2024-09-24 21:09:36,198 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=589040.6666666666, ans=0.09899494936611666 2024-09-24 21:09:44,286 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=589040.6666666666, ans=0.1 2024-09-24 21:09:52,347 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=589087.3333333334, ans=0.125 2024-09-24 21:10:00,324 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-24 21:10:00,343 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=589087.3333333334, ans=0.0 2024-09-24 21:10:16,157 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=589134.0, ans=0.0 2024-09-24 21:10:28,455 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.04 vs. limit=15.0 2024-09-24 21:10:42,935 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=589227.3333333334, ans=0.125 2024-09-24 21:10:48,174 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=4.70 vs. limit=12.0 2024-09-24 21:10:55,334 INFO [train.py:1198] (2/4) Epoch 33, batch 1600, loss[loss=0.1817, ctc_loss=0.1156, cr_loss=0.3303, over 17304.00 frames. ], tot_loss[loss=0.1954, ctc_loss=0.1269, cr_loss=0.3423, over 3355772.44 frames. ], batch size: 51, lr: 3.61e-03, grad_scale: 32.0 2024-09-24 21:11:00,334 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.046e+02 1.243e+02 1.338e+02 1.455e+02 2.020e+02, threshold=2.676e+02, percent-clipped=0.0 2024-09-24 21:11:43,903 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=589367.3333333334, ans=0.125 2024-09-24 21:12:06,220 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=589460.6666666666, ans=0.0 2024-09-24 21:12:15,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=589460.6666666666, ans=0.1 2024-09-24 21:12:20,303 INFO [train.py:1198] (2/4) Epoch 33, batch 1650, loss[loss=0.1837, ctc_loss=0.1173, cr_loss=0.3321, over 17166.00 frames. ], tot_loss[loss=0.1971, ctc_loss=0.1282, cr_loss=0.3445, over 3349498.26 frames. ], batch size: 45, lr: 3.61e-03, grad_scale: 16.0 2024-09-24 21:12:34,848 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=589554.0, ans=0.125 2024-09-24 21:12:38,614 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=5.09 vs. limit=12.0 2024-09-24 21:13:43,006 INFO [train.py:1198] (2/4) Epoch 33, batch 1700, loss[loss=0.1699, ctc_loss=0.1083, cr_loss=0.3081, over 17107.00 frames. ], tot_loss[loss=0.1958, ctc_loss=0.1273, cr_loss=0.3427, over 3353439.13 frames. ], batch size: 40, lr: 3.61e-03, grad_scale: 16.0 2024-09-24 21:13:43,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=589740.6666666666, ans=0.125 2024-09-24 21:13:49,389 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.055e+02 1.272e+02 1.348e+02 1.489e+02 2.581e+02, threshold=2.697e+02, percent-clipped=0.0 2024-09-24 21:13:57,891 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=589787.3333333334, ans=0.0 2024-09-24 21:13:59,425 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=589787.3333333334, ans=0.025 2024-09-24 21:14:00,109 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.04 vs. limit=22.5 2024-09-24 21:14:20,226 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=589834.0, ans=0.0 2024-09-24 21:14:36,829 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=589880.6666666666, ans=0.1 2024-09-24 21:14:38,448 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=589880.6666666666, ans=0.1 2024-09-24 21:14:54,628 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=5.16 vs. limit=10.0 2024-09-24 21:15:01,467 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.75 vs. limit=15.0 2024-09-24 21:15:03,862 INFO [train.py:1198] (2/4) Epoch 33, batch 1750, loss[loss=0.1972, ctc_loss=0.1305, cr_loss=0.3334, over 17216.00 frames. ], tot_loss[loss=0.1966, ctc_loss=0.1278, cr_loss=0.3438, over 3356897.24 frames. ], batch size: 47, lr: 3.61e-03, grad_scale: 16.0 2024-09-24 21:15:18,529 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=590020.6666666666, ans=0.0 2024-09-24 21:15:21,601 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=590020.6666666666, ans=0.1 2024-09-24 21:16:13,518 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=590160.6666666666, ans=0.125 2024-09-24 21:16:22,401 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.min_positive, batch_count=590160.6666666666, ans=0.05 2024-09-24 21:16:31,141 INFO [train.py:1198] (2/4) Epoch 33, batch 1800, loss[loss=0.2396, ctc_loss=0.1669, cr_loss=0.3636, over 11363.00 frames. ], tot_loss[loss=0.1982, ctc_loss=0.129, cr_loss=0.3461, over 3351147.34 frames. ], batch size: 123, lr: 3.61e-03, grad_scale: 16.0 2024-09-24 21:16:37,337 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.113e+02 1.268e+02 1.341e+02 1.421e+02 2.295e+02, threshold=2.681e+02, percent-clipped=0.0 2024-09-24 21:16:40,149 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.27 vs. limit=15.0 2024-09-24 21:16:47,349 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=590254.0, ans=0.0 2024-09-24 21:17:14,718 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=4.33 vs. limit=12.0 2024-09-24 21:17:18,213 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.95 vs. limit=15.0 2024-09-24 21:17:50,740 INFO [train.py:1198] (2/4) Epoch 33, batch 1850, loss[loss=0.2456, ctc_loss=0.1667, cr_loss=0.3947, over 15232.00 frames. ], tot_loss[loss=0.1996, ctc_loss=0.13, cr_loss=0.3478, over 3347823.27 frames. ], batch size: 89, lr: 3.61e-03, grad_scale: 16.0 2024-09-24 21:17:58,766 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=590440.6666666666, ans=0.1 2024-09-24 21:18:17,403 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=590487.3333333334, ans=0.025 2024-09-24 21:18:33,708 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer_ff2.min_abs, batch_count=590534.0, ans=0.1 2024-09-24 21:19:13,088 INFO [train.py:1198] (2/4) Epoch 33, batch 1900, loss[loss=0.201, ctc_loss=0.1301, cr_loss=0.3544, over 17007.00 frames. ], tot_loss[loss=0.1984, ctc_loss=0.1292, cr_loss=0.3464, over 3357354.28 frames. ], batch size: 51, lr: 3.61e-03, grad_scale: 16.0 2024-09-24 21:19:13,355 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=590674.0, ans=0.2 2024-09-24 21:19:19,404 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.258e+02 1.359e+02 1.471e+02 2.658e+02, threshold=2.718e+02, percent-clipped=0.0 2024-09-24 21:19:53,287 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=590767.3333333334, ans=0.125 2024-09-24 21:20:09,953 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=9.62 vs. limit=15.0 2024-09-24 21:20:17,684 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=590860.6666666666, ans=0.125 2024-09-24 21:20:31,988 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=590907.3333333334, ans=0.125 2024-09-24 21:20:33,375 INFO [train.py:1198] (2/4) Epoch 33, batch 1950, loss[loss=0.1888, ctc_loss=0.125, cr_loss=0.3189, over 17207.00 frames. ], tot_loss[loss=0.1976, ctc_loss=0.1285, cr_loss=0.3456, over 3364867.31 frames. ], batch size: 47, lr: 3.61e-03, grad_scale: 8.0 2024-09-24 21:20:41,740 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=590907.3333333334, ans=0.125 2024-09-24 21:20:50,722 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=590954.0, ans=0.0 2024-09-24 21:20:58,662 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=590954.0, ans=0.125 2024-09-24 21:21:05,794 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.whiten, num_groups=1, num_channels=192, metric=4.06 vs. limit=12.0 2024-09-24 21:21:24,218 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=591000.6666666666, ans=0.125 2024-09-24 21:21:53,453 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.47 vs. limit=15.0 2024-09-24 21:21:56,038 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=591094.0, ans=0.125 2024-09-24 21:21:59,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=591140.6666666666, ans=0.0 2024-09-24 21:22:00,619 INFO [train.py:1198] (2/4) Epoch 33, batch 2000, loss[loss=0.2002, ctc_loss=0.1324, cr_loss=0.339, over 17359.00 frames. ], tot_loss[loss=0.1971, ctc_loss=0.128, cr_loss=0.3454, over 3365106.12 frames. ], batch size: 48, lr: 3.61e-03, grad_scale: 16.0 2024-09-24 21:22:08,625 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.162e+02 1.300e+02 1.359e+02 1.463e+02 1.672e+02, threshold=2.718e+02, percent-clipped=0.0 2024-09-24 21:22:12,013 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=591140.6666666666, ans=0.125 2024-09-24 21:22:55,075 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=591280.6666666666, ans=0.0 2024-09-24 21:23:07,139 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=591327.3333333334, ans=0.0 2024-09-24 21:23:17,202 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=8.51 vs. limit=15.0 2024-09-24 21:23:22,761 INFO [train.py:1198] (2/4) Epoch 33, batch 2050, loss[loss=0.1966, ctc_loss=0.1252, cr_loss=0.3571, over 16987.00 frames. ], tot_loss[loss=0.1975, ctc_loss=0.1283, cr_loss=0.3461, over 3360527.97 frames. ], batch size: 51, lr: 3.61e-03, grad_scale: 8.0 2024-09-24 21:23:22,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=591374.0, ans=0.0 2024-09-24 21:23:30,894 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=591374.0, ans=0.025 2024-09-24 21:23:42,212 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=591420.6666666666, ans=0.125 2024-09-24 21:23:56,780 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=591467.3333333334, ans=0.0 2024-09-24 21:24:39,937 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=591560.6666666666, ans=0.125 2024-09-24 21:24:42,769 INFO [train.py:1198] (2/4) Epoch 33, batch 2100, loss[loss=0.2135, ctc_loss=0.1397, cr_loss=0.3687, over 17216.00 frames. ], tot_loss[loss=0.1976, ctc_loss=0.1283, cr_loss=0.3464, over 3356012.37 frames. ], batch size: 50, lr: 3.61e-03, grad_scale: 8.0 2024-09-24 21:24:49,552 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=591607.3333333334, ans=0.07 2024-09-24 21:24:52,446 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.121e+02 1.286e+02 1.370e+02 1.480e+02 2.165e+02, threshold=2.740e+02, percent-clipped=0.0 2024-09-24 21:24:54,941 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=9.01 vs. limit=15.0 2024-09-24 21:25:06,125 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=5.28 vs. limit=12.0 2024-09-24 21:25:07,903 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=4.02 vs. limit=15.0 2024-09-24 21:25:13,705 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=591700.6666666666, ans=0.125 2024-09-24 21:25:13,929 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 21:25:56,607 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=591794.0, ans=0.125 2024-09-24 21:26:03,087 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 21:26:05,919 INFO [train.py:1198] (2/4) Epoch 33, batch 2150, loss[loss=0.1983, ctc_loss=0.1286, cr_loss=0.3482, over 17347.00 frames. ], tot_loss[loss=0.1975, ctc_loss=0.1283, cr_loss=0.3461, over 3357206.22 frames. ], batch size: 52, lr: 3.61e-03, grad_scale: 8.0 2024-09-24 21:26:22,921 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 21:26:43,555 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=591934.0, ans=0.125 2024-09-24 21:26:58,171 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=591980.6666666666, ans=0.0 2024-09-24 21:27:02,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=591980.6666666666, ans=0.0 2024-09-24 21:27:16,006 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=592027.3333333334, ans=0.125 2024-09-24 21:27:26,320 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.29 vs. limit=15.0 2024-09-24 21:27:27,732 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=592027.3333333334, ans=0.125 2024-09-24 21:27:32,258 INFO [train.py:1198] (2/4) Epoch 33, batch 2200, loss[loss=0.1854, ctc_loss=0.1202, cr_loss=0.326, over 17301.00 frames. ], tot_loss[loss=0.1968, ctc_loss=0.1278, cr_loss=0.3454, over 3361682.38 frames. ], batch size: 49, lr: 3.60e-03, grad_scale: 8.0 2024-09-24 21:27:41,951 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.112e+02 1.250e+02 1.335e+02 1.414e+02 1.894e+02, threshold=2.669e+02, percent-clipped=0.0 2024-09-24 21:27:50,439 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=592120.6666666666, ans=0.125 2024-09-24 21:28:07,338 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=592167.3333333334, ans=0.09899494936611666 2024-09-24 21:28:16,071 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.28 vs. limit=22.5 2024-09-24 21:28:40,568 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=592260.6666666666, ans=0.2 2024-09-24 21:28:43,690 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=592260.6666666666, ans=0.125 2024-09-24 21:28:50,054 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=592260.6666666666, ans=0.0 2024-09-24 21:28:54,690 INFO [train.py:1198] (2/4) Epoch 33, batch 2250, loss[loss=0.1946, ctc_loss=0.127, cr_loss=0.3381, over 17090.00 frames. ], tot_loss[loss=0.1974, ctc_loss=0.1282, cr_loss=0.3457, over 3356748.63 frames. ], batch size: 49, lr: 3.60e-03, grad_scale: 8.0 2024-09-24 21:28:58,263 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=592307.3333333334, ans=0.0 2024-09-24 21:29:43,368 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=592447.3333333334, ans=0.125 2024-09-24 21:29:58,286 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.00 vs. limit=22.5 2024-09-24 21:30:14,784 INFO [train.py:1198] (2/4) Epoch 33, batch 2300, loss[loss=0.2014, ctc_loss=0.1341, cr_loss=0.3368, over 15123.00 frames. ], tot_loss[loss=0.1975, ctc_loss=0.1283, cr_loss=0.3456, over 3346383.79 frames. ], batch size: 89, lr: 3.60e-03, grad_scale: 8.0 2024-09-24 21:30:15,110 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=592540.6666666666, ans=0.125 2024-09-24 21:30:22,067 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=6.20 vs. limit=15.0 2024-09-24 21:30:24,326 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.141e+02 1.313e+02 1.398e+02 1.523e+02 4.380e+02, threshold=2.797e+02, percent-clipped=1.0 2024-09-24 21:30:54,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=592634.0, ans=0.125 2024-09-24 21:31:07,155 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=592680.6666666666, ans=0.125 2024-09-24 21:31:24,570 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=9.88 vs. limit=15.0 2024-09-24 21:31:29,389 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=14.36 vs. limit=15.0 2024-09-24 21:31:32,746 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=9.29 vs. limit=15.0 2024-09-24 21:31:43,090 INFO [train.py:1198] (2/4) Epoch 33, batch 2350, loss[loss=0.1739, ctc_loss=0.1092, cr_loss=0.3235, over 17036.00 frames. ], tot_loss[loss=0.1972, ctc_loss=0.1282, cr_loss=0.3451, over 3354649.06 frames. ], batch size: 39, lr: 3.60e-03, grad_scale: 8.0 2024-09-24 21:31:45,167 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 21:31:53,148 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=592774.0, ans=0.125 2024-09-24 21:32:20,043 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=592867.3333333334, ans=0.125 2024-09-24 21:32:42,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=592914.0, ans=0.025 2024-09-24 21:32:42,222 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=592914.0, ans=0.2 2024-09-24 21:32:43,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=592914.0, ans=0.125 2024-09-24 21:32:55,735 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.min_positive, batch_count=592960.6666666666, ans=0.025 2024-09-24 21:33:05,026 INFO [train.py:1198] (2/4) Epoch 33, batch 2400, loss[loss=0.1884, ctc_loss=0.1205, cr_loss=0.3397, over 17177.00 frames. ], tot_loss[loss=0.1975, ctc_loss=0.1283, cr_loss=0.346, over 3358143.20 frames. ], batch size: 45, lr: 3.60e-03, grad_scale: 16.0 2024-09-24 21:33:14,695 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.083e+02 1.266e+02 1.359e+02 1.474e+02 2.316e+02, threshold=2.718e+02, percent-clipped=0.0 2024-09-24 21:33:37,400 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=593100.6666666666, ans=0.0 2024-09-24 21:33:53,854 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.73 vs. limit=10.0 2024-09-24 21:34:12,853 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=8.83 vs. limit=15.0 2024-09-24 21:34:24,926 INFO [train.py:1198] (2/4) Epoch 33, batch 2450, loss[loss=0.1841, ctc_loss=0.1171, cr_loss=0.3351, over 17215.00 frames. ], tot_loss[loss=0.1974, ctc_loss=0.1282, cr_loss=0.3457, over 3366174.04 frames. ], batch size: 47, lr: 3.60e-03, grad_scale: 16.0 2024-09-24 21:35:25,793 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=593380.6666666666, ans=0.125 2024-09-24 21:35:48,034 INFO [train.py:1198] (2/4) Epoch 33, batch 2500, loss[loss=0.1863, ctc_loss=0.1202, cr_loss=0.3304, over 17326.00 frames. ], tot_loss[loss=0.1982, ctc_loss=0.1287, cr_loss=0.3473, over 3364756.64 frames. ], batch size: 46, lr: 3.60e-03, grad_scale: 16.0 2024-09-24 21:35:51,667 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=593474.0, ans=0.1 2024-09-24 21:35:53,170 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=593474.0, ans=0.0 2024-09-24 21:36:00,299 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.106e+02 1.280e+02 1.356e+02 1.432e+02 1.776e+02, threshold=2.712e+02, percent-clipped=0.0 2024-09-24 21:36:19,222 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=593520.6666666666, ans=0.1 2024-09-24 21:36:19,229 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=593520.6666666666, ans=0.125 2024-09-24 21:36:31,986 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=593567.3333333334, ans=0.125 2024-09-24 21:36:47,286 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.84 vs. limit=15.0 2024-09-24 21:37:13,486 INFO [train.py:1198] (2/4) Epoch 33, batch 2550, loss[loss=0.2292, ctc_loss=0.1529, cr_loss=0.3815, over 17348.00 frames. ], tot_loss[loss=0.1988, ctc_loss=0.1292, cr_loss=0.3477, over 3352227.58 frames. ], batch size: 48, lr: 3.60e-03, grad_scale: 16.0 2024-09-24 21:37:20,578 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=9.99 vs. limit=22.5 2024-09-24 21:37:42,379 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=593754.0, ans=0.125 2024-09-24 21:37:51,247 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=593800.6666666666, ans=0.1 2024-09-24 21:37:57,422 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=593800.6666666666, ans=0.125 2024-09-24 21:38:00,518 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=593800.6666666666, ans=0.2 2024-09-24 21:38:05,407 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=593847.3333333334, ans=0.125 2024-09-24 21:38:11,827 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=593847.3333333334, ans=0.0 2024-09-24 21:38:35,207 INFO [train.py:1198] (2/4) Epoch 33, batch 2600, loss[loss=0.1986, ctc_loss=0.1318, cr_loss=0.334, over 16989.00 frames. ], tot_loss[loss=0.1982, ctc_loss=0.1288, cr_loss=0.3472, over 3355903.97 frames. ], batch size: 53, lr: 3.60e-03, grad_scale: 16.0 2024-09-24 21:38:44,724 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.263e+02 1.356e+02 1.473e+02 2.021e+02, threshold=2.712e+02, percent-clipped=0.0 2024-09-24 21:38:45,149 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=593940.6666666666, ans=0.2 2024-09-24 21:39:03,943 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=593987.3333333334, ans=0.125 2024-09-24 21:39:37,966 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.14 vs. limit=15.0 2024-09-24 21:39:54,548 INFO [train.py:1198] (2/4) Epoch 33, batch 2650, loss[loss=0.2291, ctc_loss=0.1538, cr_loss=0.3766, over 11371.00 frames. ], tot_loss[loss=0.1972, ctc_loss=0.1281, cr_loss=0.3456, over 3354555.58 frames. ], batch size: 124, lr: 3.60e-03, grad_scale: 16.0 2024-09-24 21:39:56,372 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=594174.0, ans=0.125 2024-09-24 21:40:14,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=594220.6666666666, ans=0.2 2024-09-24 21:40:27,084 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 21:41:14,967 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.41 vs. limit=15.0 2024-09-24 21:41:19,333 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=594360.6666666666, ans=0.0 2024-09-24 21:41:20,737 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=594407.3333333334, ans=0.0 2024-09-24 21:41:22,173 INFO [train.py:1198] (2/4) Epoch 33, batch 2700, loss[loss=0.2117, ctc_loss=0.1378, cr_loss=0.3692, over 17060.00 frames. ], tot_loss[loss=0.1963, ctc_loss=0.1274, cr_loss=0.3445, over 3356214.63 frames. ], batch size: 46, lr: 3.60e-03, grad_scale: 16.0 2024-09-24 21:41:31,684 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.245e+02 1.329e+02 1.418e+02 2.018e+02, threshold=2.657e+02, percent-clipped=0.0 2024-09-24 21:41:40,016 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=594454.0, ans=0.1 2024-09-24 21:41:44,912 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=8.66 vs. limit=15.0 2024-09-24 21:42:44,868 INFO [train.py:1198] (2/4) Epoch 33, batch 2750, loss[loss=0.2271, ctc_loss=0.1509, cr_loss=0.3811, over 14878.00 frames. ], tot_loss[loss=0.1969, ctc_loss=0.1279, cr_loss=0.3451, over 3356596.69 frames. ], batch size: 89, lr: 3.60e-03, grad_scale: 16.0 2024-09-24 21:43:12,909 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=13.30 vs. limit=15.0 2024-09-24 21:43:14,515 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=10.35 vs. limit=22.5 2024-09-24 21:43:39,850 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=594780.6666666666, ans=0.05 2024-09-24 21:43:55,506 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=594827.3333333334, ans=0.1 2024-09-24 21:44:04,857 INFO [train.py:1198] (2/4) Epoch 33, batch 2800, loss[loss=0.2171, ctc_loss=0.1435, cr_loss=0.3681, over 15760.00 frames. ], tot_loss[loss=0.1976, ctc_loss=0.1283, cr_loss=0.3461, over 3360504.66 frames. ], batch size: 74, lr: 3.60e-03, grad_scale: 32.0 2024-09-24 21:44:14,277 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.107e+02 1.285e+02 1.356e+02 1.483e+02 1.883e+02, threshold=2.713e+02, percent-clipped=0.0 2024-09-24 21:44:38,512 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=594967.3333333334, ans=0.125 2024-09-24 21:44:38,528 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=594967.3333333334, ans=0.025 2024-09-24 21:44:43,770 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.00 vs. limit=22.5 2024-09-24 21:44:56,782 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.52 vs. limit=12.0 2024-09-24 21:44:59,610 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer_na.min_abs, batch_count=595014.0, ans=0.02 2024-09-24 21:45:13,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=595060.6666666666, ans=0.0 2024-09-24 21:45:22,472 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.04 vs. limit=15.0 2024-09-24 21:45:24,681 INFO [train.py:1198] (2/4) Epoch 33, batch 2850, loss[loss=0.1907, ctc_loss=0.1242, cr_loss=0.3328, over 17360.00 frames. ], tot_loss[loss=0.1976, ctc_loss=0.1284, cr_loss=0.346, over 3363078.37 frames. ], batch size: 48, lr: 3.60e-03, grad_scale: 32.0 2024-09-24 21:45:36,966 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=595107.3333333334, ans=0.125 2024-09-24 21:45:49,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=595154.0, ans=0.0 2024-09-24 21:45:59,701 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=595154.0, ans=0.1 2024-09-24 21:46:07,604 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=595200.6666666666, ans=0.125 2024-09-24 21:46:26,657 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=595247.3333333334, ans=0.0 2024-09-24 21:46:52,041 INFO [train.py:1198] (2/4) Epoch 33, batch 2900, loss[loss=0.1886, ctc_loss=0.123, cr_loss=0.3283, over 17016.00 frames. ], tot_loss[loss=0.1975, ctc_loss=0.1283, cr_loss=0.3459, over 3367999.44 frames. ], batch size: 51, lr: 3.59e-03, grad_scale: 32.0 2024-09-24 21:46:59,350 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten.whitening_limit, batch_count=595340.6666666666, ans=15.0 2024-09-24 21:47:01,744 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.073e+02 1.278e+02 1.378e+02 1.507e+02 2.269e+02, threshold=2.756e+02, percent-clipped=0.0 2024-09-24 21:47:17,850 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=595387.3333333334, ans=0.125 2024-09-24 21:47:19,967 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.20 vs. limit=22.5 2024-09-24 21:48:06,099 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=595527.3333333334, ans=0.125 2024-09-24 21:48:15,330 INFO [train.py:1198] (2/4) Epoch 33, batch 2950, loss[loss=0.2083, ctc_loss=0.1374, cr_loss=0.3545, over 17141.00 frames. ], tot_loss[loss=0.1974, ctc_loss=0.1282, cr_loss=0.3461, over 3370298.08 frames. ], batch size: 48, lr: 3.59e-03, grad_scale: 32.0 2024-09-24 21:48:31,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=595620.6666666666, ans=0.125 2024-09-24 21:48:49,370 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=595667.3333333334, ans=0.125 2024-09-24 21:49:22,818 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.09 vs. limit=15.0 2024-09-24 21:49:28,602 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=595760.6666666666, ans=0.025 2024-09-24 21:49:34,546 INFO [train.py:1198] (2/4) Epoch 33, batch 3000, loss[loss=0.2039, ctc_loss=0.1344, cr_loss=0.3477, over 16518.00 frames. ], tot_loss[loss=0.1992, ctc_loss=0.1295, cr_loss=0.3484, over 3363248.71 frames. ], batch size: 66, lr: 3.59e-03, grad_scale: 32.0 2024-09-24 21:49:34,547 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 21:49:50,274 INFO [train.py:1230] (2/4) Epoch 33, validation: loss=0.03597, ctc_loss=0.03597, cr_loss=9.382e-15, over 944034.00 frames. 2024-09-24 21:49:50,274 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 21:49:56,791 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=595807.3333333334, ans=0.0 2024-09-24 21:49:59,653 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.292e+02 1.353e+02 1.495e+02 2.152e+02, threshold=2.707e+02, percent-clipped=0.0 2024-09-24 21:50:00,055 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=595807.3333333334, ans=0.0 2024-09-24 21:50:06,227 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=595854.0, ans=0.1 2024-09-24 21:50:33,481 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=11.33 vs. limit=15.0 2024-09-24 21:50:34,812 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-24 21:50:34,836 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=595900.6666666666, ans=0.125 2024-09-24 21:50:34,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=595900.6666666666, ans=0.2 2024-09-24 21:50:57,886 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=595994.0, ans=0.0 2024-09-24 21:50:59,418 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=595994.0, ans=0.025 2024-09-24 21:50:59,437 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=595994.0, ans=0.2 2024-09-24 21:51:08,635 INFO [train.py:1198] (2/4) Epoch 33, batch 3050, loss[loss=0.2034, ctc_loss=0.1302, cr_loss=0.366, over 17263.00 frames. ], tot_loss[loss=0.1995, ctc_loss=0.1297, cr_loss=0.3487, over 3356525.10 frames. ], batch size: 44, lr: 3.59e-03, grad_scale: 32.0 2024-09-24 21:51:13,773 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=596040.6666666666, ans=0.125 2024-09-24 21:51:16,708 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=596040.6666666666, ans=0.0 2024-09-24 21:52:01,725 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=16.51 vs. limit=22.5 2024-09-24 21:52:12,394 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.min_positive, batch_count=596227.3333333334, ans=0.05 2024-09-24 21:52:18,527 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=596227.3333333334, ans=0.0 2024-09-24 21:52:34,217 INFO [train.py:1198] (2/4) Epoch 33, batch 3100, loss[loss=0.1571, ctc_loss=0.09797, cr_loss=0.2956, over 17080.00 frames. ], tot_loss[loss=0.1981, ctc_loss=0.1287, cr_loss=0.347, over 3362860.92 frames. ], batch size: 43, lr: 3.59e-03, grad_scale: 32.0 2024-09-24 21:52:40,880 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=596274.0, ans=0.0 2024-09-24 21:52:43,659 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.278e+02 1.346e+02 1.474e+02 1.974e+02, threshold=2.692e+02, percent-clipped=0.0 2024-09-24 21:52:48,726 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=596320.6666666666, ans=0.2 2024-09-24 21:53:10,944 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=596367.3333333334, ans=0.125 2024-09-24 21:53:20,518 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=596414.0, ans=0.0 2024-09-24 21:53:28,649 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=596414.0, ans=0.1 2024-09-24 21:53:53,433 INFO [train.py:1198] (2/4) Epoch 33, batch 3150, loss[loss=0.2016, ctc_loss=0.1298, cr_loss=0.3592, over 17215.00 frames. ], tot_loss[loss=0.1979, ctc_loss=0.1286, cr_loss=0.3465, over 3357814.61 frames. ], batch size: 50, lr: 3.59e-03, grad_scale: 32.0 2024-09-24 21:53:56,949 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=596507.3333333334, ans=0.125 2024-09-24 21:55:02,987 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.51 vs. limit=10.0 2024-09-24 21:55:11,608 INFO [train.py:1198] (2/4) Epoch 33, batch 3200, loss[loss=0.1821, ctc_loss=0.116, cr_loss=0.3306, over 17108.00 frames. ], tot_loss[loss=0.1988, ctc_loss=0.1292, cr_loss=0.3481, over 3359485.35 frames. ], batch size: 40, lr: 3.59e-03, grad_scale: 32.0 2024-09-24 21:55:22,508 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.266e+02 1.354e+02 1.465e+02 1.945e+02, threshold=2.709e+02, percent-clipped=0.0 2024-09-24 21:55:38,308 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=596787.3333333334, ans=0.125 2024-09-24 21:55:49,302 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=596834.0, ans=0.0 2024-09-24 21:55:57,178 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=596880.6666666666, ans=0.125 2024-09-24 21:55:58,609 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=596880.6666666666, ans=0.0 2024-09-24 21:55:58,763 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=596880.6666666666, ans=0.2 2024-09-24 21:56:31,845 INFO [train.py:1198] (2/4) Epoch 33, batch 3250, loss[loss=0.1609, ctc_loss=0.1053, cr_loss=0.2779, over 16692.00 frames. ], tot_loss[loss=0.1974, ctc_loss=0.1282, cr_loss=0.3462, over 3364333.61 frames. ], batch size: 37, lr: 3.59e-03, grad_scale: 32.0 2024-09-24 21:56:34,574 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.whiten, num_groups=1, num_channels=192, metric=4.11 vs. limit=12.0 2024-09-24 21:56:37,281 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.20 vs. limit=22.5 2024-09-24 21:57:03,564 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=597067.3333333334, ans=0.1 2024-09-24 21:57:05,220 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 21:57:06,846 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=597067.3333333334, ans=0.125 2024-09-24 21:57:08,395 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=597067.3333333334, ans=0.07 2024-09-24 21:57:12,987 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=597067.3333333334, ans=0.125 2024-09-24 21:57:19,381 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=597114.0, ans=0.0 2024-09-24 21:57:24,009 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.min_abs, batch_count=597114.0, ans=0.5 2024-09-24 21:57:31,926 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=597114.0, ans=0.0 2024-09-24 21:57:50,337 INFO [train.py:1198] (2/4) Epoch 33, batch 3300, loss[loss=0.148, ctc_loss=0.0938, cr_loss=0.2711, over 17100.00 frames. ], tot_loss[loss=0.1958, ctc_loss=0.1271, cr_loss=0.3436, over 3372705.50 frames. ], batch size: 40, lr: 3.59e-03, grad_scale: 32.0 2024-09-24 21:58:00,051 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=597207.3333333334, ans=0.1 2024-09-24 21:58:01,413 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.145e+02 1.317e+02 1.415e+02 1.516e+02 3.233e+02, threshold=2.830e+02, percent-clipped=1.0 2024-09-24 21:58:23,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=597300.6666666666, ans=0.2 2024-09-24 21:59:11,412 INFO [train.py:1198] (2/4) Epoch 33, batch 3350, loss[loss=0.2293, ctc_loss=0.1494, cr_loss=0.3991, over 17070.00 frames. ], tot_loss[loss=0.195, ctc_loss=0.1265, cr_loss=0.3428, over 3377230.11 frames. ], batch size: 56, lr: 3.59e-03, grad_scale: 32.0 2024-09-24 21:59:13,367 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=597440.6666666666, ans=0.0 2024-09-24 21:59:26,021 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=597487.3333333334, ans=0.0 2024-09-24 22:00:09,751 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=597580.6666666666, ans=0.1 2024-09-24 22:00:29,915 INFO [train.py:1198] (2/4) Epoch 33, batch 3400, loss[loss=0.2245, ctc_loss=0.147, cr_loss=0.3873, over 16497.00 frames. ], tot_loss[loss=0.1944, ctc_loss=0.126, cr_loss=0.3421, over 3377083.24 frames. ], batch size: 66, lr: 3.59e-03, grad_scale: 32.0 2024-09-24 22:00:40,768 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.046e+02 1.249e+02 1.331e+02 1.438e+02 2.060e+02, threshold=2.663e+02, percent-clipped=0.0 2024-09-24 22:01:26,482 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=597814.0, ans=0.0 2024-09-24 22:01:35,977 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.max_abs, batch_count=597860.6666666666, ans=10.0 2024-09-24 22:01:48,346 INFO [train.py:1198] (2/4) Epoch 33, batch 3450, loss[loss=0.1527, ctc_loss=0.09683, cr_loss=0.2794, over 16629.00 frames. ], tot_loss[loss=0.194, ctc_loss=0.1257, cr_loss=0.3415, over 3378897.91 frames. ], batch size: 37, lr: 3.59e-03, grad_scale: 32.0 2024-09-24 22:02:10,635 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=597954.0, ans=0.125 2024-09-24 22:02:19,852 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=598000.6666666666, ans=0.1 2024-09-24 22:02:27,597 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=598000.6666666666, ans=0.0 2024-09-24 22:02:39,109 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=598047.3333333334, ans=0.125 2024-09-24 22:02:42,314 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=598047.3333333334, ans=0.0 2024-09-24 22:03:01,648 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=598094.0, ans=0.125 2024-09-24 22:03:03,080 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer_ff2.min_abs, batch_count=598094.0, ans=0.1 2024-09-24 22:03:04,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.max_positive, batch_count=598094.0, ans=0.95 2024-09-24 22:03:12,223 INFO [train.py:1198] (2/4) Epoch 33, batch 3500, loss[loss=0.2023, ctc_loss=0.131, cr_loss=0.3568, over 16970.00 frames. ], tot_loss[loss=0.1947, ctc_loss=0.1262, cr_loss=0.3423, over 3373315.72 frames. ], batch size: 58, lr: 3.59e-03, grad_scale: 32.0 2024-09-24 22:03:23,014 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.259e+02 1.352e+02 1.466e+02 4.097e+02, threshold=2.703e+02, percent-clipped=1.0 2024-09-24 22:03:29,762 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=598187.3333333334, ans=0.125 2024-09-24 22:03:35,935 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=598187.3333333334, ans=0.1 2024-09-24 22:03:54,835 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=2.373e-02 2024-09-24 22:04:30,420 INFO [train.py:1198] (2/4) Epoch 33, batch 3550, loss[loss=0.1806, ctc_loss=0.118, cr_loss=0.3129, over 17199.00 frames. ], tot_loss[loss=0.1948, ctc_loss=0.1263, cr_loss=0.3426, over 3373567.38 frames. ], batch size: 41, lr: 3.59e-03, grad_scale: 32.0 2024-09-24 22:04:30,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=598374.0, ans=0.0 2024-09-24 22:04:49,377 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=598420.6666666666, ans=0.0 2024-09-24 22:04:52,459 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=598420.6666666666, ans=0.125 2024-09-24 22:05:12,197 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.70 vs. limit=6.0 2024-09-24 22:05:16,247 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=598514.0, ans=0.125 2024-09-24 22:05:24,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=598514.0, ans=0.0 2024-09-24 22:05:24,454 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=4.29 vs. limit=12.0 2024-09-24 22:05:27,100 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=598514.0, ans=0.0 2024-09-24 22:05:48,756 INFO [train.py:1198] (2/4) Epoch 33, batch 3600, loss[loss=0.2235, ctc_loss=0.146, cr_loss=0.3878, over 17025.00 frames. ], tot_loss[loss=0.1957, ctc_loss=0.1269, cr_loss=0.3438, over 3380692.41 frames. ], batch size: 52, lr: 3.58e-03, grad_scale: 32.0 2024-09-24 22:05:59,708 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.085e+02 1.266e+02 1.339e+02 1.451e+02 1.959e+02, threshold=2.678e+02, percent-clipped=0.0 2024-09-24 22:06:06,405 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=598654.0, ans=0.2 2024-09-24 22:06:36,403 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer_ff3.min_abs, batch_count=598747.3333333334, ans=0.2 2024-09-24 22:07:08,975 INFO [train.py:1198] (2/4) Epoch 33, batch 3650, loss[loss=0.2374, ctc_loss=0.1571, cr_loss=0.4017, over 15223.00 frames. ], tot_loss[loss=0.1955, ctc_loss=0.1268, cr_loss=0.3436, over 3369706.32 frames. ], batch size: 89, lr: 3.58e-03, grad_scale: 32.0 2024-09-24 22:08:01,850 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=598980.6666666666, ans=0.025 2024-09-24 22:08:11,197 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.min_positive, batch_count=599027.3333333334, ans=0.05 2024-09-24 22:08:25,142 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=599027.3333333334, ans=0.125 2024-09-24 22:08:26,810 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=599074.0, ans=0.125 2024-09-24 22:08:27,969 INFO [train.py:1198] (2/4) Epoch 33, batch 3700, loss[loss=0.2099, ctc_loss=0.1407, cr_loss=0.3461, over 14779.00 frames. ], tot_loss[loss=0.1966, ctc_loss=0.1276, cr_loss=0.3449, over 3367979.87 frames. ], batch size: 89, lr: 3.58e-03, grad_scale: 32.0 2024-09-24 22:08:39,001 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.089e+02 1.282e+02 1.323e+02 1.479e+02 2.496e+02, threshold=2.646e+02, percent-clipped=0.0 2024-09-24 22:09:00,576 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=599167.3333333334, ans=0.0 2024-09-24 22:09:16,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=599214.0, ans=0.125 2024-09-24 22:09:18,192 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=599214.0, ans=0.1 2024-09-24 22:09:23,521 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.94 vs. limit=12.0 2024-09-24 22:09:36,876 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=599260.6666666666, ans=0.035 2024-09-24 22:09:45,897 INFO [train.py:1198] (2/4) Epoch 33, batch 3750, loss[loss=0.2149, ctc_loss=0.1397, cr_loss=0.3757, over 16917.00 frames. ], tot_loss[loss=0.1968, ctc_loss=0.1277, cr_loss=0.3452, over 3354017.23 frames. ], batch size: 58, lr: 3.58e-03, grad_scale: 32.0 2024-09-24 22:09:55,956 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=11.46 vs. limit=22.5 2024-09-24 22:10:03,869 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=15.92 vs. limit=22.5 2024-09-24 22:10:46,810 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=5.29 vs. limit=15.0 2024-09-24 22:11:02,778 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=599540.6666666666, ans=0.125 2024-09-24 22:11:03,989 INFO [train.py:1198] (2/4) Epoch 33, batch 3800, loss[loss=0.2067, ctc_loss=0.1336, cr_loss=0.3657, over 17023.00 frames. ], tot_loss[loss=0.1988, ctc_loss=0.1292, cr_loss=0.3479, over 3349598.81 frames. ], batch size: 51, lr: 3.58e-03, grad_scale: 32.0 2024-09-24 22:11:14,753 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.301e+02 1.426e+02 1.535e+02 2.880e+02, threshold=2.852e+02, percent-clipped=1.0 2024-09-24 22:11:16,709 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=599540.6666666666, ans=0.0 2024-09-24 22:11:31,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=599587.3333333334, ans=0.125 2024-09-24 22:11:53,408 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.81 vs. limit=6.0 2024-09-24 22:12:10,765 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=599727.3333333334, ans=0.0 2024-09-24 22:12:22,914 INFO [train.py:1198] (2/4) Epoch 33, batch 3850, loss[loss=0.2405, ctc_loss=0.1645, cr_loss=0.3799, over 11852.00 frames. ], tot_loss[loss=0.2013, ctc_loss=0.1312, cr_loss=0.3506, over 3312254.98 frames. ], batch size: 124, lr: 3.58e-03, grad_scale: 16.0 2024-09-24 22:13:09,521 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=599914.0, ans=0.0 2024-09-24 22:13:11,054 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=599914.0, ans=0.05 2024-09-24 22:13:12,476 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=599914.0, ans=0.125 2024-09-24 22:13:25,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=599960.6666666666, ans=0.1 2024-09-24 22:14:24,100 INFO [train.py:1198] (2/4) Epoch 34, batch 0, loss[loss=0.1977, ctc_loss=0.1247, cr_loss=0.3654, over 17167.00 frames. ], tot_loss[loss=0.1977, ctc_loss=0.1247, cr_loss=0.3654, over 17167.00 frames. ], batch size: 45, lr: 3.53e-03, grad_scale: 32.0 2024-09-24 22:14:24,100 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 22:14:39,382 INFO [train.py:1230] (2/4) Epoch 34, validation: loss=0.03567, ctc_loss=0.03567, cr_loss=1.032e-14, over 944034.00 frames. 2024-09-24 22:14:39,382 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 22:14:58,491 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.377e+02 1.527e+02 1.748e+02 2.707e+02, threshold=3.055e+02, percent-clipped=0.0 2024-09-24 22:15:03,748 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=600035.3333333334, ans=0.1 2024-09-24 22:15:15,527 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=9.96 vs. limit=15.0 2024-09-24 22:15:29,494 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=600128.6666666666, ans=0.2 2024-09-24 22:15:59,188 INFO [train.py:1198] (2/4) Epoch 34, batch 50, loss[loss=0.1973, ctc_loss=0.1245, cr_loss=0.3643, over 17002.00 frames. ], tot_loss[loss=0.1922, ctc_loss=0.1242, cr_loss=0.34, over 766791.13 frames. ], batch size: 44, lr: 3.53e-03, grad_scale: 32.0 2024-09-24 22:16:18,451 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=600268.6666666666, ans=0.125 2024-09-24 22:16:20,126 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=600268.6666666666, ans=0.1 2024-09-24 22:16:20,245 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=600268.6666666666, ans=0.125 2024-09-24 22:16:30,863 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=600268.6666666666, ans=0.0 2024-09-24 22:16:36,895 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=600315.3333333334, ans=0.1 2024-09-24 22:16:37,599 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.47 vs. limit=12.0 2024-09-24 22:17:03,676 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=600362.0, ans=0.125 2024-09-24 22:17:07,334 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.79 vs. limit=6.0 2024-09-24 22:17:13,340 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=600408.6666666666, ans=0.125 2024-09-24 22:17:29,796 INFO [train.py:1198] (2/4) Epoch 34, batch 100, loss[loss=0.1844, ctc_loss=0.1186, cr_loss=0.3292, over 17160.00 frames. ], tot_loss[loss=0.1921, ctc_loss=0.1243, cr_loss=0.3389, over 1330369.72 frames. ], batch size: 45, lr: 3.53e-03, grad_scale: 32.0 2024-09-24 22:17:49,214 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.286e+02 1.376e+02 1.498e+02 2.035e+02, threshold=2.751e+02, percent-clipped=0.0 2024-09-24 22:17:59,243 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=600502.0, ans=0.0 2024-09-24 22:18:00,691 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=600548.6666666666, ans=0.125 2024-09-24 22:18:49,899 INFO [train.py:1198] (2/4) Epoch 34, batch 150, loss[loss=0.208, ctc_loss=0.1341, cr_loss=0.3694, over 16768.00 frames. ], tot_loss[loss=0.1912, ctc_loss=0.1236, cr_loss=0.3383, over 1785820.99 frames. ], batch size: 61, lr: 3.52e-03, grad_scale: 32.0 2024-09-24 22:18:58,107 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=600688.6666666666, ans=0.0 2024-09-24 22:19:04,562 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=600735.3333333334, ans=0.125 2024-09-24 22:19:07,606 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=600735.3333333334, ans=0.0 2024-09-24 22:19:24,854 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=600782.0, ans=0.0 2024-09-24 22:19:31,138 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=600782.0, ans=0.125 2024-09-24 22:19:58,272 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=600875.3333333334, ans=0.0 2024-09-24 22:19:58,989 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.74 vs. limit=12.0 2024-09-24 22:20:09,192 INFO [train.py:1198] (2/4) Epoch 34, batch 200, loss[loss=0.2123, ctc_loss=0.139, cr_loss=0.3668, over 17207.00 frames. ], tot_loss[loss=0.1924, ctc_loss=0.1244, cr_loss=0.3399, over 2139458.41 frames. ], batch size: 50, lr: 3.52e-03, grad_scale: 32.0 2024-09-24 22:20:27,508 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.23 vs. limit=15.0 2024-09-24 22:20:30,036 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.131e+02 1.259e+02 1.333e+02 1.425e+02 2.058e+02, threshold=2.665e+02, percent-clipped=0.0 2024-09-24 22:20:44,718 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=601015.3333333334, ans=0.0 2024-09-24 22:20:51,162 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=601015.3333333334, ans=0.125 2024-09-24 22:21:04,293 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=601062.0, ans=0.0 2024-09-24 22:21:12,662 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=512, metric=6.01 vs. limit=15.0 2024-09-24 22:21:32,286 INFO [train.py:1198] (2/4) Epoch 34, batch 250, loss[loss=0.1819, ctc_loss=0.116, cr_loss=0.3297, over 17286.00 frames. ], tot_loss[loss=0.1928, ctc_loss=0.1246, cr_loss=0.3409, over 2401719.73 frames. ], batch size: 46, lr: 3.52e-03, grad_scale: 16.0 2024-09-24 22:22:04,162 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=601202.0, ans=0.1 2024-09-24 22:22:21,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=601248.6666666666, ans=0.0 2024-09-24 22:23:00,622 INFO [train.py:1198] (2/4) Epoch 34, batch 300, loss[loss=0.2155, ctc_loss=0.1428, cr_loss=0.3632, over 11507.00 frames. ], tot_loss[loss=0.195, ctc_loss=0.1263, cr_loss=0.3434, over 2604340.63 frames. ], batch size: 123, lr: 3.52e-03, grad_scale: 16.0 2024-09-24 22:23:01,480 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=10.52 vs. limit=15.0 2024-09-24 22:23:05,697 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=601388.6666666666, ans=0.1 2024-09-24 22:23:21,391 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.136e+02 1.266e+02 1.387e+02 1.525e+02 2.483e+02, threshold=2.773e+02, percent-clipped=0.0 2024-09-24 22:23:37,725 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=601482.0, ans=0.125 2024-09-24 22:23:45,915 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.54 vs. limit=15.0 2024-09-24 22:23:55,294 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=601528.6666666666, ans=0.125 2024-09-24 22:23:58,433 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=601528.6666666666, ans=0.125 2024-09-24 22:24:07,792 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=601575.3333333334, ans=0.125 2024-09-24 22:24:19,110 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=601622.0, ans=0.025 2024-09-24 22:24:20,383 INFO [train.py:1198] (2/4) Epoch 34, batch 350, loss[loss=0.2228, ctc_loss=0.1449, cr_loss=0.3892, over 17295.00 frames. ], tot_loss[loss=0.1959, ctc_loss=0.1271, cr_loss=0.3439, over 2765926.30 frames. ], batch size: 51, lr: 3.52e-03, grad_scale: 16.0 2024-09-24 22:24:36,540 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=601668.6666666666, ans=0.125 2024-09-24 22:24:41,588 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=601668.6666666666, ans=0.125 2024-09-24 22:24:54,219 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=601715.3333333334, ans=0.0 2024-09-24 22:24:55,854 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=601715.3333333334, ans=0.0 2024-09-24 22:24:58,292 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=12.13 vs. limit=15.0 2024-09-24 22:25:18,092 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=601762.0, ans=0.125 2024-09-24 22:25:19,602 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=601762.0, ans=0.07 2024-09-24 22:25:35,986 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.36 vs. limit=15.0 2024-09-24 22:25:39,904 INFO [train.py:1198] (2/4) Epoch 34, batch 400, loss[loss=0.2106, ctc_loss=0.137, cr_loss=0.3683, over 17361.00 frames. ], tot_loss[loss=0.1964, ctc_loss=0.1274, cr_loss=0.3453, over 2898445.82 frames. ], batch size: 48, lr: 3.52e-03, grad_scale: 16.0 2024-09-24 22:25:44,277 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.83 vs. limit=6.0 2024-09-24 22:25:46,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=601855.3333333334, ans=0.125 2024-09-24 22:25:54,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=601902.0, ans=0.125 2024-09-24 22:26:02,214 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.142e+02 1.275e+02 1.410e+02 1.501e+02 2.362e+02, threshold=2.821e+02, percent-clipped=0.0 2024-09-24 22:26:32,819 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer_ff2.min_abs, batch_count=601995.3333333334, ans=0.1 2024-09-24 22:26:44,057 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=601995.3333333334, ans=0.2 2024-09-24 22:27:05,426 INFO [train.py:1198] (2/4) Epoch 34, batch 450, loss[loss=0.2206, ctc_loss=0.1419, cr_loss=0.3934, over 17022.00 frames. ], tot_loss[loss=0.1973, ctc_loss=0.1281, cr_loss=0.3463, over 3002566.53 frames. ], batch size: 52, lr: 3.52e-03, grad_scale: 16.0 2024-09-24 22:27:46,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=602182.0, ans=0.1 2024-09-24 22:27:49,445 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 22:27:52,880 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=10.98 vs. limit=15.0 2024-09-24 22:28:00,291 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=602228.6666666666, ans=0.2 2024-09-24 22:28:09,779 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=602228.6666666666, ans=0.0 2024-09-24 22:28:30,683 INFO [train.py:1198] (2/4) Epoch 34, batch 500, loss[loss=0.1535, ctc_loss=0.0956, cr_loss=0.2895, over 17264.00 frames. ], tot_loss[loss=0.1971, ctc_loss=0.128, cr_loss=0.3458, over 3078993.13 frames. ], batch size: 42, lr: 3.52e-03, grad_scale: 16.0 2024-09-24 22:28:52,903 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.146e+02 1.302e+02 1.372e+02 1.474e+02 2.066e+02, threshold=2.745e+02, percent-clipped=0.0 2024-09-24 22:29:12,925 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.00 vs. limit=15.0 2024-09-24 22:29:31,221 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=602462.0, ans=0.05 2024-09-24 22:29:39,558 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=5.16 vs. limit=10.0 2024-09-24 22:29:49,895 INFO [train.py:1198] (2/4) Epoch 34, batch 550, loss[loss=0.1904, ctc_loss=0.1258, cr_loss=0.323, over 17129.00 frames. ], tot_loss[loss=0.196, ctc_loss=0.1272, cr_loss=0.3442, over 3149825.13 frames. ], batch size: 48, lr: 3.52e-03, grad_scale: 16.0 2024-09-24 22:30:09,638 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.max_positive, batch_count=602602.0, ans=0.95 2024-09-24 22:30:52,910 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=602742.0, ans=0.125 2024-09-24 22:31:02,171 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=602742.0, ans=0.1 2024-09-24 22:31:06,760 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=602742.0, ans=0.2 2024-09-24 22:31:07,485 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=11.05 vs. limit=15.0 2024-09-24 22:31:09,783 INFO [train.py:1198] (2/4) Epoch 34, batch 600, loss[loss=0.1843, ctc_loss=0.1191, cr_loss=0.326, over 17188.00 frames. ], tot_loss[loss=0.1958, ctc_loss=0.127, cr_loss=0.3439, over 3202131.72 frames. ], batch size: 47, lr: 3.52e-03, grad_scale: 16.0 2024-09-24 22:31:13,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=602788.6666666666, ans=0.025 2024-09-24 22:31:25,126 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=602788.6666666666, ans=0.125 2024-09-24 22:31:34,416 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.259e+02 1.339e+02 1.429e+02 3.195e+02, threshold=2.679e+02, percent-clipped=1.0 2024-09-24 22:31:36,308 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 22:32:17,961 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=602928.6666666666, ans=0.125 2024-09-24 22:32:32,162 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=602975.3333333334, ans=0.5 2024-09-24 22:32:39,990 INFO [train.py:1198] (2/4) Epoch 34, batch 650, loss[loss=0.1924, ctc_loss=0.1245, cr_loss=0.3398, over 17308.00 frames. ], tot_loss[loss=0.1969, ctc_loss=0.1278, cr_loss=0.3455, over 3226764.83 frames. ], batch size: 46, lr: 3.52e-03, grad_scale: 16.0 2024-09-24 22:32:46,808 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=603022.0, ans=0.125 2024-09-24 22:32:48,271 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=603022.0, ans=0.2 2024-09-24 22:32:49,832 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=603022.0, ans=0.0 2024-09-24 22:33:17,454 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=603115.3333333334, ans=0.125 2024-09-24 22:33:17,462 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=603115.3333333334, ans=0.05 2024-09-24 22:33:19,683 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.39 vs. limit=15.0 2024-09-24 22:34:00,563 INFO [train.py:1198] (2/4) Epoch 34, batch 700, loss[loss=0.2064, ctc_loss=0.1356, cr_loss=0.3539, over 17156.00 frames. ], tot_loss[loss=0.1975, ctc_loss=0.1281, cr_loss=0.3467, over 3269318.72 frames. ], batch size: 45, lr: 3.52e-03, grad_scale: 16.0 2024-09-24 22:34:15,354 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=603302.0, ans=0.2 2024-09-24 22:34:23,177 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.142e+02 1.279e+02 1.394e+02 1.540e+02 2.036e+02, threshold=2.788e+02, percent-clipped=0.0 2024-09-24 22:34:33,622 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=6.16 vs. limit=15.0 2024-09-24 22:34:36,928 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.80 vs. limit=10.0 2024-09-24 22:35:21,219 INFO [train.py:1198] (2/4) Epoch 34, batch 750, loss[loss=0.1609, ctc_loss=0.1036, cr_loss=0.2865, over 17047.00 frames. ], tot_loss[loss=0.197, ctc_loss=0.1278, cr_loss=0.3461, over 3291331.23 frames. ], batch size: 39, lr: 3.52e-03, grad_scale: 16.0 2024-09-24 22:35:24,691 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=603488.6666666666, ans=0.0 2024-09-24 22:35:45,467 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=603535.3333333334, ans=0.025 2024-09-24 22:35:59,556 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=603582.0, ans=0.025 2024-09-24 22:36:06,357 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=13.80 vs. limit=15.0 2024-09-24 22:36:15,547 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=603628.6666666666, ans=0.125 2024-09-24 22:36:17,135 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=603628.6666666666, ans=0.125 2024-09-24 22:36:24,561 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=603628.6666666666, ans=0.125 2024-09-24 22:36:27,822 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=603675.3333333334, ans=0.125 2024-09-24 22:36:34,229 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=603675.3333333334, ans=0.125 2024-09-24 22:36:40,624 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=603675.3333333334, ans=0.125 2024-09-24 22:36:43,570 INFO [train.py:1198] (2/4) Epoch 34, batch 800, loss[loss=0.2301, ctc_loss=0.1525, cr_loss=0.3878, over 16973.00 frames. ], tot_loss[loss=0.1964, ctc_loss=0.1274, cr_loss=0.3455, over 3309584.86 frames. ], batch size: 53, lr: 3.52e-03, grad_scale: 32.0 2024-09-24 22:37:08,640 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.110e+02 1.269e+02 1.369e+02 1.482e+02 1.915e+02, threshold=2.738e+02, percent-clipped=0.0 2024-09-24 22:37:22,192 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=603815.3333333334, ans=0.05 2024-09-24 22:37:48,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=603862.0, ans=0.2 2024-09-24 22:38:09,954 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=5.61 vs. limit=15.0 2024-09-24 22:38:12,078 INFO [train.py:1198] (2/4) Epoch 34, batch 850, loss[loss=0.2259, ctc_loss=0.1453, cr_loss=0.403, over 17019.00 frames. ], tot_loss[loss=0.1962, ctc_loss=0.1272, cr_loss=0.3453, over 3330440.07 frames. ], batch size: 52, lr: 3.52e-03, grad_scale: 32.0 2024-09-24 22:38:18,743 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=603955.3333333334, ans=0.125 2024-09-24 22:38:26,773 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=604002.0, ans=0.0 2024-09-24 22:38:50,536 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=604048.6666666666, ans=0.0 2024-09-24 22:38:56,864 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=604048.6666666666, ans=0.0 2024-09-24 22:39:17,691 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=604142.0, ans=0.2 2024-09-24 22:39:31,802 INFO [train.py:1198] (2/4) Epoch 34, batch 900, loss[loss=0.207, ctc_loss=0.1358, cr_loss=0.3561, over 17006.00 frames. ], tot_loss[loss=0.1958, ctc_loss=0.1269, cr_loss=0.3445, over 3340146.11 frames. ], batch size: 51, lr: 3.51e-03, grad_scale: 32.0 2024-09-24 22:39:33,009 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=16.93 vs. limit=22.5 2024-09-24 22:39:54,076 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.145e+02 1.293e+02 1.407e+02 1.509e+02 3.747e+02, threshold=2.814e+02, percent-clipped=1.0 2024-09-24 22:40:01,471 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.44 vs. limit=22.5 2024-09-24 22:40:10,735 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=604282.0, ans=0.125 2024-09-24 22:40:14,066 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.min_positive, batch_count=604282.0, ans=0.05 2024-09-24 22:40:27,016 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=604328.6666666666, ans=0.0 2024-09-24 22:40:31,836 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=604328.6666666666, ans=0.0 2024-09-24 22:40:33,940 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.95 vs. limit=10.0 2024-09-24 22:40:52,440 INFO [train.py:1198] (2/4) Epoch 34, batch 950, loss[loss=0.2044, ctc_loss=0.1329, cr_loss=0.3573, over 16749.00 frames. ], tot_loss[loss=0.1953, ctc_loss=0.1265, cr_loss=0.344, over 3338658.22 frames. ], batch size: 61, lr: 3.51e-03, grad_scale: 32.0 2024-09-24 22:41:08,884 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=604468.6666666666, ans=0.0 2024-09-24 22:42:23,021 INFO [train.py:1198] (2/4) Epoch 34, batch 1000, loss[loss=0.2025, ctc_loss=0.1305, cr_loss=0.3595, over 17162.00 frames. ], tot_loss[loss=0.195, ctc_loss=0.1263, cr_loss=0.3436, over 3355146.20 frames. ], batch size: 45, lr: 3.51e-03, grad_scale: 32.0 2024-09-24 22:42:45,412 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.115e+02 1.270e+02 1.349e+02 1.472e+02 1.904e+02, threshold=2.698e+02, percent-clipped=0.0 2024-09-24 22:42:47,361 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=604702.0, ans=0.125 2024-09-24 22:42:50,477 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=604702.0, ans=0.125 2024-09-24 22:43:40,623 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=604842.0, ans=0.125 2024-09-24 22:43:43,609 INFO [train.py:1198] (2/4) Epoch 34, batch 1050, loss[loss=0.1952, ctc_loss=0.1301, cr_loss=0.3254, over 17017.00 frames. ], tot_loss[loss=0.1947, ctc_loss=0.1262, cr_loss=0.3428, over 3352529.13 frames. ], batch size: 44, lr: 3.51e-03, grad_scale: 32.0 2024-09-24 22:43:45,492 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=604888.6666666666, ans=0.125 2024-09-24 22:43:46,894 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=604888.6666666666, ans=0.0 2024-09-24 22:45:03,651 INFO [train.py:1198] (2/4) Epoch 34, batch 1100, loss[loss=0.2161, ctc_loss=0.1438, cr_loss=0.3615, over 17242.00 frames. ], tot_loss[loss=0.1944, ctc_loss=0.1259, cr_loss=0.3424, over 3363806.27 frames. ], batch size: 50, lr: 3.51e-03, grad_scale: 32.0 2024-09-24 22:45:08,968 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=605122.0, ans=0.07 2024-09-24 22:45:26,104 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.266e+02 1.354e+02 1.438e+02 1.919e+02, threshold=2.709e+02, percent-clipped=0.0 2024-09-24 22:46:21,379 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=605308.6666666666, ans=0.125 2024-09-24 22:46:26,898 INFO [train.py:1198] (2/4) Epoch 34, batch 1150, loss[loss=0.2331, ctc_loss=0.1603, cr_loss=0.3642, over 11822.00 frames. ], tot_loss[loss=0.1945, ctc_loss=0.126, cr_loss=0.3425, over 3361770.09 frames. ], batch size: 124, lr: 3.51e-03, grad_scale: 32.0 2024-09-24 22:46:32,490 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=4.47 vs. limit=15.0 2024-09-24 22:46:50,762 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.93 vs. limit=15.0 2024-09-24 22:46:58,363 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=605402.0, ans=0.2 2024-09-24 22:47:16,789 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=605448.6666666666, ans=0.1 2024-09-24 22:47:16,902 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=605448.6666666666, ans=0.125 2024-09-24 22:47:50,069 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.62 vs. limit=10.0 2024-09-24 22:47:54,033 INFO [train.py:1198] (2/4) Epoch 34, batch 1200, loss[loss=0.2334, ctc_loss=0.1546, cr_loss=0.3942, over 15070.00 frames. ], tot_loss[loss=0.1948, ctc_loss=0.1262, cr_loss=0.3428, over 3358182.66 frames. ], batch size: 89, lr: 3.51e-03, grad_scale: 32.0 2024-09-24 22:48:16,699 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.120e+02 1.285e+02 1.360e+02 1.428e+02 2.074e+02, threshold=2.720e+02, percent-clipped=0.0 2024-09-24 22:48:21,903 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=605635.3333333334, ans=0.1 2024-09-24 22:48:28,616 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=17.64 vs. limit=22.5 2024-09-24 22:49:14,482 INFO [train.py:1198] (2/4) Epoch 34, batch 1250, loss[loss=0.203, ctc_loss=0.1331, cr_loss=0.3496, over 16986.00 frames. ], tot_loss[loss=0.195, ctc_loss=0.1265, cr_loss=0.3428, over 3349977.41 frames. ], batch size: 56, lr: 3.51e-03, grad_scale: 32.0 2024-09-24 22:49:24,490 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=605822.0, ans=0.2 2024-09-24 22:49:29,120 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=605868.6666666666, ans=0.2 2024-09-24 22:49:29,136 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=605868.6666666666, ans=0.125 2024-09-24 22:49:37,165 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=605868.6666666666, ans=0.0 2024-09-24 22:49:40,763 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=14.67 vs. limit=15.0 2024-09-24 22:49:42,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=605868.6666666666, ans=0.09899494936611666 2024-09-24 22:49:53,120 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=605915.3333333334, ans=0.1 2024-09-24 22:50:18,912 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=606008.6666666666, ans=0.0 2024-09-24 22:50:28,590 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=606008.6666666666, ans=0.2 2024-09-24 22:50:34,624 INFO [train.py:1198] (2/4) Epoch 34, batch 1300, loss[loss=0.2092, ctc_loss=0.1397, cr_loss=0.3475, over 17092.00 frames. ], tot_loss[loss=0.1947, ctc_loss=0.1263, cr_loss=0.3421, over 3355487.67 frames. ], batch size: 49, lr: 3.51e-03, grad_scale: 32.0 2024-09-24 22:50:39,699 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=606055.3333333334, ans=0.0 2024-09-24 22:50:45,146 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=16.01 vs. limit=22.5 2024-09-24 22:50:47,745 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=606055.3333333334, ans=0.0 2024-09-24 22:50:57,098 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.157e+02 1.261e+02 1.315e+02 1.424e+02 2.011e+02, threshold=2.630e+02, percent-clipped=0.0 2024-09-24 22:51:03,789 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=606102.0, ans=0.2 2024-09-24 22:51:16,837 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=606148.6666666666, ans=0.0 2024-09-24 22:51:20,223 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=9.17 vs. limit=15.0 2024-09-24 22:51:59,708 INFO [train.py:1198] (2/4) Epoch 34, batch 1350, loss[loss=0.1848, ctc_loss=0.1187, cr_loss=0.3307, over 17181.00 frames. ], tot_loss[loss=0.1952, ctc_loss=0.1266, cr_loss=0.343, over 3359259.47 frames. ], batch size: 41, lr: 3.51e-03, grad_scale: 32.0 2024-09-24 22:52:06,410 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=606288.6666666666, ans=0.125 2024-09-24 22:52:22,145 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten.whitening_limit, batch_count=606335.3333333334, ans=15.0 2024-09-24 22:52:43,335 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.min_positive, batch_count=606382.0, ans=0.025 2024-09-24 22:53:01,603 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.97 vs. limit=15.0 2024-09-24 22:53:02,666 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=606428.6666666666, ans=0.0 2024-09-24 22:53:24,833 INFO [train.py:1198] (2/4) Epoch 34, batch 1400, loss[loss=0.1786, ctc_loss=0.1139, cr_loss=0.3235, over 17080.00 frames. ], tot_loss[loss=0.1951, ctc_loss=0.1265, cr_loss=0.343, over 3356346.42 frames. ], batch size: 43, lr: 3.51e-03, grad_scale: 32.0 2024-09-24 22:53:47,247 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.283e+02 1.396e+02 1.529e+02 1.918e+02, threshold=2.793e+02, percent-clipped=0.0 2024-09-24 22:53:49,267 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=606568.6666666666, ans=0.125 2024-09-24 22:54:03,532 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=606615.3333333334, ans=0.2 2024-09-24 22:54:05,094 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=606615.3333333334, ans=0.025 2024-09-24 22:54:08,115 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=606615.3333333334, ans=0.0 2024-09-24 22:54:09,891 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=606615.3333333334, ans=0.0 2024-09-24 22:54:16,640 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.73 vs. limit=22.5 2024-09-24 22:54:19,619 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=606662.0, ans=0.0 2024-09-24 22:54:32,189 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=606708.6666666666, ans=0.125 2024-09-24 22:54:44,663 INFO [train.py:1198] (2/4) Epoch 34, batch 1450, loss[loss=0.1752, ctc_loss=0.1139, cr_loss=0.3069, over 16694.00 frames. ], tot_loss[loss=0.1944, ctc_loss=0.126, cr_loss=0.3419, over 3363288.40 frames. ], batch size: 61, lr: 3.51e-03, grad_scale: 32.0 2024-09-24 22:55:08,858 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=606802.0, ans=0.125 2024-09-24 22:55:10,474 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=606802.0, ans=0.0 2024-09-24 22:55:20,100 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=606848.6666666666, ans=0.0 2024-09-24 22:55:50,529 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=606942.0, ans=0.0 2024-09-24 22:55:58,647 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=606942.0, ans=0.1 2024-09-24 22:56:04,599 INFO [train.py:1198] (2/4) Epoch 34, batch 1500, loss[loss=0.2219, ctc_loss=0.147, cr_loss=0.3744, over 15288.00 frames. ], tot_loss[loss=0.1949, ctc_loss=0.1265, cr_loss=0.3422, over 3355514.17 frames. ], batch size: 89, lr: 3.51e-03, grad_scale: 32.0 2024-09-24 22:56:26,520 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.min_abs, batch_count=607035.3333333334, ans=0.5 2024-09-24 22:56:28,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=607035.3333333334, ans=0.125 2024-09-24 22:56:29,447 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.138e+02 1.258e+02 1.350e+02 1.442e+02 1.821e+02, threshold=2.700e+02, percent-clipped=0.0 2024-09-24 22:56:48,566 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=607082.0, ans=0.1 2024-09-24 22:56:56,562 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=607128.6666666666, ans=0.2 2024-09-24 22:57:34,685 INFO [train.py:1198] (2/4) Epoch 34, batch 1550, loss[loss=0.1787, ctc_loss=0.1169, cr_loss=0.3092, over 17161.00 frames. ], tot_loss[loss=0.1952, ctc_loss=0.1267, cr_loss=0.3427, over 3362618.61 frames. ], batch size: 48, lr: 3.51e-03, grad_scale: 32.0 2024-09-24 22:57:34,908 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=607222.0, ans=0.1 2024-09-24 22:57:38,100 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=607222.0, ans=0.2 2024-09-24 22:57:39,614 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=607222.0, ans=0.1 2024-09-24 22:57:50,940 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=607268.6666666666, ans=0.0 2024-09-24 22:58:01,969 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=607268.6666666666, ans=0.0 2024-09-24 22:58:27,682 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=607362.0, ans=0.125 2024-09-24 22:58:54,426 INFO [train.py:1198] (2/4) Epoch 34, batch 1600, loss[loss=0.2221, ctc_loss=0.143, cr_loss=0.3955, over 17305.00 frames. ], tot_loss[loss=0.1954, ctc_loss=0.1268, cr_loss=0.3431, over 3367425.99 frames. ], batch size: 49, lr: 3.51e-03, grad_scale: 32.0 2024-09-24 22:59:06,261 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.63 vs. limit=10.0 2024-09-24 22:59:16,751 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.116e+02 1.263e+02 1.329e+02 1.418e+02 2.097e+02, threshold=2.657e+02, percent-clipped=0.0 2024-09-24 22:59:36,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=607548.6666666666, ans=0.2 2024-09-24 22:59:36,819 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=512, metric=3.29 vs. limit=15.0 2024-09-24 23:00:14,909 INFO [train.py:1198] (2/4) Epoch 34, batch 1650, loss[loss=0.2141, ctc_loss=0.1385, cr_loss=0.3781, over 17206.00 frames. ], tot_loss[loss=0.196, ctc_loss=0.1272, cr_loss=0.344, over 3360698.64 frames. ], batch size: 55, lr: 3.50e-03, grad_scale: 16.0 2024-09-24 23:00:34,471 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=607735.3333333334, ans=0.1 2024-09-24 23:00:58,066 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=607782.0, ans=0.0 2024-09-24 23:01:10,552 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=607828.6666666666, ans=0.025 2024-09-24 23:01:15,406 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=607828.6666666666, ans=0.125 2024-09-24 23:01:25,950 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=607875.3333333334, ans=0.125 2024-09-24 23:01:29,016 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=607875.3333333334, ans=0.0 2024-09-24 23:01:33,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=607875.3333333334, ans=0.125 2024-09-24 23:01:36,802 INFO [train.py:1198] (2/4) Epoch 34, batch 1700, loss[loss=0.1766, ctc_loss=0.1097, cr_loss=0.3347, over 17185.00 frames. ], tot_loss[loss=0.1968, ctc_loss=0.1278, cr_loss=0.3452, over 3352491.99 frames. ], batch size: 41, lr: 3.50e-03, grad_scale: 16.0 2024-09-24 23:02:05,582 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.051e+02 1.278e+02 1.343e+02 1.428e+02 2.095e+02, threshold=2.685e+02, percent-clipped=0.0 2024-09-24 23:02:05,947 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=607968.6666666666, ans=0.125 2024-09-24 23:02:22,952 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=608015.3333333334, ans=0.1 2024-09-24 23:02:38,874 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=608062.0, ans=0.025 2024-09-24 23:02:45,424 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=608062.0, ans=0.0 2024-09-24 23:03:04,327 INFO [train.py:1198] (2/4) Epoch 34, batch 1750, loss[loss=0.1937, ctc_loss=0.1264, cr_loss=0.3365, over 17077.00 frames. ], tot_loss[loss=0.1979, ctc_loss=0.1286, cr_loss=0.3463, over 3339621.98 frames. ], batch size: 46, lr: 3.50e-03, grad_scale: 16.0 2024-09-24 23:03:14,196 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.min_abs, batch_count=608155.3333333334, ans=0.5 2024-09-24 23:03:40,077 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=608248.6666666666, ans=0.0 2024-09-24 23:03:52,793 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=608295.3333333334, ans=0.125 2024-09-24 23:04:24,156 INFO [train.py:1198] (2/4) Epoch 34, batch 1800, loss[loss=0.2106, ctc_loss=0.1387, cr_loss=0.3592, over 17028.00 frames. ], tot_loss[loss=0.1966, ctc_loss=0.1277, cr_loss=0.3448, over 3339194.87 frames. ], batch size: 52, lr: 3.50e-03, grad_scale: 16.0 2024-09-24 23:04:35,927 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.97 vs. limit=15.0 2024-09-24 23:04:45,283 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=608435.3333333334, ans=0.1 2024-09-24 23:04:48,241 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.131e+02 1.244e+02 1.332e+02 1.441e+02 2.602e+02, threshold=2.663e+02, percent-clipped=0.0 2024-09-24 23:04:50,088 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=608435.3333333334, ans=0.1 2024-09-24 23:05:06,138 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=608482.0, ans=0.0 2024-09-24 23:05:44,354 INFO [train.py:1198] (2/4) Epoch 34, batch 1850, loss[loss=0.2118, ctc_loss=0.1399, cr_loss=0.3594, over 17305.00 frames. ], tot_loss[loss=0.1967, ctc_loss=0.1276, cr_loss=0.3454, over 3353528.34 frames. ], batch size: 51, lr: 3.50e-03, grad_scale: 16.0 2024-09-24 23:05:53,393 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=8.94 vs. limit=10.0 2024-09-24 23:06:02,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=608668.6666666666, ans=0.0 2024-09-24 23:06:25,609 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=608715.3333333334, ans=0.125 2024-09-24 23:06:46,077 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=608762.0, ans=0.1 2024-09-24 23:07:15,202 INFO [train.py:1198] (2/4) Epoch 34, batch 1900, loss[loss=0.1691, ctc_loss=0.1067, cr_loss=0.3118, over 17302.00 frames. ], tot_loss[loss=0.1965, ctc_loss=0.1275, cr_loss=0.3448, over 3358224.02 frames. ], batch size: 46, lr: 3.50e-03, grad_scale: 16.0 2024-09-24 23:07:21,123 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.21 vs. limit=15.0 2024-09-24 23:07:23,731 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 23:07:39,345 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.279e+02 1.348e+02 1.451e+02 1.794e+02, threshold=2.695e+02, percent-clipped=0.0 2024-09-24 23:07:52,279 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=608948.6666666666, ans=0.5 2024-09-24 23:07:57,245 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=608948.6666666666, ans=0.2 2024-09-24 23:08:07,294 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=512, metric=3.00 vs. limit=15.0 2024-09-24 23:08:35,254 INFO [train.py:1198] (2/4) Epoch 34, batch 1950, loss[loss=0.229, ctc_loss=0.1521, cr_loss=0.3843, over 15948.00 frames. ], tot_loss[loss=0.1972, ctc_loss=0.128, cr_loss=0.3458, over 3355494.38 frames. ], batch size: 74, lr: 3.50e-03, grad_scale: 16.0 2024-09-24 23:08:38,693 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=609088.6666666666, ans=0.125 2024-09-24 23:08:39,295 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.96 vs. limit=15.0 2024-09-24 23:09:20,742 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=609182.0, ans=0.125 2024-09-24 23:09:32,362 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=609228.6666666666, ans=0.2 2024-09-24 23:09:56,092 INFO [train.py:1198] (2/4) Epoch 34, batch 2000, loss[loss=0.2121, ctc_loss=0.1364, cr_loss=0.379, over 17062.00 frames. ], tot_loss[loss=0.1979, ctc_loss=0.1287, cr_loss=0.3459, over 3346135.61 frames. ], batch size: 46, lr: 3.50e-03, grad_scale: 32.0 2024-09-24 23:09:59,581 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=609322.0, ans=0.125 2024-09-24 23:10:07,617 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=609322.0, ans=0.125 2024-09-24 23:10:16,987 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=609368.6666666666, ans=0.125 2024-09-24 23:10:19,956 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.131e+02 1.293e+02 1.367e+02 1.523e+02 2.152e+02, threshold=2.734e+02, percent-clipped=0.0 2024-09-24 23:11:18,783 INFO [train.py:1198] (2/4) Epoch 34, batch 2050, loss[loss=0.2005, ctc_loss=0.1285, cr_loss=0.3601, over 16992.00 frames. ], tot_loss[loss=0.1969, ctc_loss=0.128, cr_loss=0.3447, over 3353433.69 frames. ], batch size: 53, lr: 3.50e-03, grad_scale: 32.0 2024-09-24 23:12:21,977 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=609695.3333333334, ans=0.2 2024-09-24 23:12:44,519 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=609788.6666666666, ans=0.1 2024-09-24 23:12:45,148 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=21.32 vs. limit=22.5 2024-09-24 23:12:45,743 INFO [train.py:1198] (2/4) Epoch 34, batch 2100, loss[loss=0.1576, ctc_loss=0.09974, cr_loss=0.2892, over 17063.00 frames. ], tot_loss[loss=0.1978, ctc_loss=0.1287, cr_loss=0.3455, over 3333795.71 frames. ], batch size: 39, lr: 3.50e-03, grad_scale: 32.0 2024-09-24 23:13:10,053 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.285e+02 1.366e+02 1.478e+02 2.142e+02, threshold=2.733e+02, percent-clipped=0.0 2024-09-24 23:13:19,937 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=609882.0, ans=0.2 2024-09-24 23:13:23,924 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=15.68 vs. limit=22.5 2024-09-24 23:13:28,187 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=609882.0, ans=0.1 2024-09-24 23:13:29,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=609882.0, ans=0.0 2024-09-24 23:13:48,978 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=609975.3333333334, ans=0.125 2024-09-24 23:13:53,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=609975.3333333334, ans=0.125 2024-09-24 23:13:55,427 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=609975.3333333334, ans=0.1 2024-09-24 23:14:00,156 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=609975.3333333334, ans=0.125 2024-09-24 23:14:06,284 INFO [train.py:1198] (2/4) Epoch 34, batch 2150, loss[loss=0.1808, ctc_loss=0.1183, cr_loss=0.3129, over 17252.00 frames. ], tot_loss[loss=0.1967, ctc_loss=0.1278, cr_loss=0.3444, over 3340254.48 frames. ], batch size: 44, lr: 3.50e-03, grad_scale: 32.0 2024-09-24 23:14:09,607 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=610022.0, ans=0.125 2024-09-24 23:14:57,518 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=610162.0, ans=0.125 2024-09-24 23:15:08,749 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=610208.6666666666, ans=0.0 2024-09-24 23:15:26,282 INFO [train.py:1198] (2/4) Epoch 34, batch 2200, loss[loss=0.2, ctc_loss=0.1321, cr_loss=0.3398, over 17008.00 frames. ], tot_loss[loss=0.1962, ctc_loss=0.1275, cr_loss=0.3432, over 3337789.56 frames. ], batch size: 51, lr: 3.50e-03, grad_scale: 32.0 2024-09-24 23:15:41,029 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=610302.0, ans=0.1 2024-09-24 23:15:50,349 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.250e+02 1.358e+02 1.486e+02 2.433e+02, threshold=2.717e+02, percent-clipped=0.0 2024-09-24 23:15:52,065 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=610302.0, ans=0.125 2024-09-24 23:16:02,500 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.94 vs. limit=6.0 2024-09-24 23:16:06,764 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.min_abs, batch_count=610348.6666666666, ans=0.5 2024-09-24 23:16:24,317 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn2.whiten.whitening_limit, batch_count=610395.3333333334, ans=22.5 2024-09-24 23:16:33,461 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=610442.0, ans=0.125 2024-09-24 23:16:51,473 INFO [train.py:1198] (2/4) Epoch 34, batch 2250, loss[loss=0.1967, ctc_loss=0.1282, cr_loss=0.3425, over 17185.00 frames. ], tot_loss[loss=0.1965, ctc_loss=0.1278, cr_loss=0.3437, over 3344714.38 frames. ], batch size: 45, lr: 3.50e-03, grad_scale: 32.0 2024-09-24 23:17:19,894 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=610535.3333333334, ans=0.2 2024-09-24 23:17:25,130 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.56 vs. limit=15.0 2024-09-24 23:17:30,179 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=9.69 vs. limit=15.0 2024-09-24 23:17:52,320 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=11.56 vs. limit=22.5 2024-09-24 23:17:56,710 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=610675.3333333334, ans=0.125 2024-09-24 23:17:56,863 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=610675.3333333334, ans=0.07 2024-09-24 23:18:14,097 INFO [train.py:1198] (2/4) Epoch 34, batch 2300, loss[loss=0.1936, ctc_loss=0.1261, cr_loss=0.3375, over 17023.00 frames. ], tot_loss[loss=0.1968, ctc_loss=0.1279, cr_loss=0.3446, over 3354199.76 frames. ], batch size: 51, lr: 3.50e-03, grad_scale: 32.0 2024-09-24 23:18:38,051 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.297e+02 1.390e+02 1.515e+02 2.091e+02, threshold=2.780e+02, percent-clipped=0.0 2024-09-24 23:18:40,084 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=610768.6666666666, ans=10.0 2024-09-24 23:18:59,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=610815.3333333334, ans=0.125 2024-09-24 23:19:07,566 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.46 vs. limit=6.0 2024-09-24 23:19:11,948 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=610862.0, ans=0.0 2024-09-24 23:19:34,055 INFO [train.py:1198] (2/4) Epoch 34, batch 2350, loss[loss=0.2065, ctc_loss=0.1381, cr_loss=0.3423, over 17215.00 frames. ], tot_loss[loss=0.1966, ctc_loss=0.1277, cr_loss=0.3447, over 3364340.86 frames. ], batch size: 50, lr: 3.49e-03, grad_scale: 32.0 2024-09-24 23:19:36,344 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=9.12 vs. limit=15.0 2024-09-24 23:19:48,921 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=10.32 vs. limit=15.0 2024-09-24 23:19:53,115 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=611002.0, ans=0.0 2024-09-24 23:20:05,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=611048.6666666666, ans=0.1 2024-09-24 23:20:17,171 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.43 vs. limit=15.0 2024-09-24 23:20:17,235 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=9.14 vs. limit=15.0 2024-09-24 23:20:42,374 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=611142.0, ans=0.125 2024-09-24 23:20:53,277 INFO [train.py:1198] (2/4) Epoch 34, batch 2400, loss[loss=0.2282, ctc_loss=0.1499, cr_loss=0.3917, over 16990.00 frames. ], tot_loss[loss=0.196, ctc_loss=0.1273, cr_loss=0.3439, over 3364521.73 frames. ], batch size: 53, lr: 3.49e-03, grad_scale: 32.0 2024-09-24 23:21:01,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=611188.6666666666, ans=0.0 2024-09-24 23:21:04,873 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=611188.6666666666, ans=0.125 2024-09-24 23:21:19,778 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.075e+02 1.247e+02 1.313e+02 1.426e+02 1.860e+02, threshold=2.625e+02, percent-clipped=0.0 2024-09-24 23:21:28,320 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=611282.0, ans=0.125 2024-09-24 23:21:32,187 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.66 vs. limit=22.5 2024-09-24 23:21:40,383 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=611282.0, ans=0.125 2024-09-24 23:21:50,138 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=611328.6666666666, ans=0.04949747468305833 2024-09-24 23:21:50,546 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=10.23 vs. limit=12.0 2024-09-24 23:22:11,338 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=611375.3333333334, ans=0.0 2024-09-24 23:22:15,388 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=8.61 vs. limit=10.0 2024-09-24 23:22:23,849 INFO [train.py:1198] (2/4) Epoch 34, batch 2450, loss[loss=0.21, ctc_loss=0.135, cr_loss=0.3747, over 17087.00 frames. ], tot_loss[loss=0.1965, ctc_loss=0.1275, cr_loss=0.3449, over 3370989.86 frames. ], batch size: 49, lr: 3.49e-03, grad_scale: 32.0 2024-09-24 23:22:52,914 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=611468.6666666666, ans=0.125 2024-09-24 23:23:42,786 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=611655.3333333334, ans=0.1 2024-09-24 23:23:43,934 INFO [train.py:1198] (2/4) Epoch 34, batch 2500, loss[loss=0.1952, ctc_loss=0.1278, cr_loss=0.3369, over 17074.00 frames. ], tot_loss[loss=0.197, ctc_loss=0.1278, cr_loss=0.3461, over 3367855.31 frames. ], batch size: 43, lr: 3.49e-03, grad_scale: 32.0 2024-09-24 23:23:49,095 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=611655.3333333334, ans=0.0 2024-09-24 23:23:57,182 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=611655.3333333334, ans=0.025 2024-09-24 23:23:59,156 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=13.23 vs. limit=22.5 2024-09-24 23:24:08,084 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.121e+02 1.282e+02 1.373e+02 1.484e+02 1.977e+02, threshold=2.747e+02, percent-clipped=0.0 2024-09-24 23:24:16,424 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=611748.6666666666, ans=0.125 2024-09-24 23:25:04,064 INFO [train.py:1198] (2/4) Epoch 34, batch 2550, loss[loss=0.1651, ctc_loss=0.1066, cr_loss=0.2927, over 17011.00 frames. ], tot_loss[loss=0.1967, ctc_loss=0.1276, cr_loss=0.3454, over 3371690.56 frames. ], batch size: 51, lr: 3.49e-03, grad_scale: 32.0 2024-09-24 23:25:08,085 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.49 vs. limit=10.0 2024-09-24 23:25:31,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=611935.3333333334, ans=0.0 2024-09-24 23:25:52,259 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=612028.6666666666, ans=0.125 2024-09-24 23:26:01,701 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=612028.6666666666, ans=0.125 2024-09-24 23:26:01,725 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=612028.6666666666, ans=0.0 2024-09-24 23:26:29,169 INFO [train.py:1198] (2/4) Epoch 34, batch 2600, loss[loss=0.214, ctc_loss=0.1422, cr_loss=0.359, over 16195.00 frames. ], tot_loss[loss=0.1971, ctc_loss=0.128, cr_loss=0.3458, over 3356994.83 frames. ], batch size: 74, lr: 3.49e-03, grad_scale: 32.0 2024-09-24 23:26:34,314 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=612122.0, ans=0.1 2024-09-24 23:26:39,119 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=612122.0, ans=0.0 2024-09-24 23:26:54,380 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=612168.6666666666, ans=0.1 2024-09-24 23:26:55,518 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.160e+02 1.298e+02 1.376e+02 1.484e+02 2.452e+02, threshold=2.751e+02, percent-clipped=0.0 2024-09-24 23:27:07,061 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=612215.3333333334, ans=0.1 2024-09-24 23:27:31,767 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=10.39 vs. limit=15.0 2024-09-24 23:27:51,787 INFO [train.py:1198] (2/4) Epoch 34, batch 2650, loss[loss=0.1893, ctc_loss=0.12, cr_loss=0.3466, over 16887.00 frames. ], tot_loss[loss=0.198, ctc_loss=0.1286, cr_loss=0.3472, over 3357997.02 frames. ], batch size: 58, lr: 3.49e-03, grad_scale: 32.0 2024-09-24 23:28:03,216 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=612355.3333333334, ans=0.125 2024-09-24 23:28:09,634 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=612402.0, ans=0.125 2024-09-24 23:28:13,099 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1.whitening_limit, batch_count=612402.0, ans=10.0 2024-09-24 23:28:51,541 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=612495.3333333334, ans=0.0 2024-09-24 23:29:02,868 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=612542.0, ans=0.1 2024-09-24 23:29:07,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=612542.0, ans=0.2 2024-09-24 23:29:10,868 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.min_positive, batch_count=612588.6666666666, ans=0.025 2024-09-24 23:29:12,127 INFO [train.py:1198] (2/4) Epoch 34, batch 2700, loss[loss=0.2006, ctc_loss=0.1298, cr_loss=0.3538, over 17275.00 frames. ], tot_loss[loss=0.1966, ctc_loss=0.1276, cr_loss=0.3451, over 3360297.29 frames. ], batch size: 44, lr: 3.49e-03, grad_scale: 32.0 2024-09-24 23:29:13,984 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=612588.6666666666, ans=0.0 2024-09-24 23:29:36,054 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.146e+02 1.298e+02 1.372e+02 1.486e+02 2.496e+02, threshold=2.744e+02, percent-clipped=0.0 2024-09-24 23:29:52,526 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=612682.0, ans=0.125 2024-09-24 23:29:55,573 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=612682.0, ans=0.125 2024-09-24 23:29:56,069 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=9.08 vs. limit=15.0 2024-09-24 23:29:57,180 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=612682.0, ans=0.125 2024-09-24 23:30:11,607 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.66 vs. limit=15.0 2024-09-24 23:30:17,999 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 23:30:31,910 INFO [train.py:1198] (2/4) Epoch 34, batch 2750, loss[loss=0.2621, ctc_loss=0.1811, cr_loss=0.4046, over 11537.00 frames. ], tot_loss[loss=0.1967, ctc_loss=0.1277, cr_loss=0.3449, over 3358013.78 frames. ], batch size: 123, lr: 3.49e-03, grad_scale: 16.0 2024-09-24 23:30:33,949 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=612822.0, ans=0.0 2024-09-24 23:30:37,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=612822.0, ans=0.0 2024-09-24 23:31:54,788 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.46 vs. limit=10.0 2024-09-24 23:32:02,048 INFO [train.py:1198] (2/4) Epoch 34, batch 2800, loss[loss=0.1567, ctc_loss=0.09837, cr_loss=0.2919, over 17132.00 frames. ], tot_loss[loss=0.1959, ctc_loss=0.1272, cr_loss=0.3433, over 3355006.49 frames. ], batch size: 40, lr: 3.49e-03, grad_scale: 32.0 2024-09-24 23:32:04,026 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=613055.3333333334, ans=0.125 2024-09-24 23:32:08,708 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=613055.3333333334, ans=0.1 2024-09-24 23:32:10,470 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=613055.3333333334, ans=0.125 2024-09-24 23:32:27,769 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.048e+02 1.265e+02 1.343e+02 1.457e+02 2.911e+02, threshold=2.687e+02, percent-clipped=1.0 2024-09-24 23:32:42,476 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=613148.6666666666, ans=0.0 2024-09-24 23:32:45,723 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=613148.6666666666, ans=0.0 2024-09-24 23:33:00,311 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=613195.3333333334, ans=0.125 2024-09-24 23:33:13,049 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=613242.0, ans=0.125 2024-09-24 23:33:22,293 INFO [train.py:1198] (2/4) Epoch 34, batch 2850, loss[loss=0.1935, ctc_loss=0.1265, cr_loss=0.3346, over 17363.00 frames. ], tot_loss[loss=0.1951, ctc_loss=0.1267, cr_loss=0.342, over 3345418.31 frames. ], batch size: 48, lr: 3.49e-03, grad_scale: 32.0 2024-09-24 23:33:22,637 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=613288.6666666666, ans=0.0 2024-09-24 23:33:25,723 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=613288.6666666666, ans=0.125 2024-09-24 23:33:37,130 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=613335.3333333334, ans=0.1 2024-09-24 23:34:01,555 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.40 vs. limit=6.0 2024-09-24 23:34:24,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=613475.3333333334, ans=0.125 2024-09-24 23:34:26,583 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=613475.3333333334, ans=0.125 2024-09-24 23:34:39,171 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=613475.3333333334, ans=0.0 2024-09-24 23:34:42,221 INFO [train.py:1198] (2/4) Epoch 34, batch 2900, loss[loss=0.2013, ctc_loss=0.1334, cr_loss=0.3396, over 16078.00 frames. ], tot_loss[loss=0.1959, ctc_loss=0.1273, cr_loss=0.3428, over 3337840.80 frames. ], batch size: 74, lr: 3.49e-03, grad_scale: 32.0 2024-09-24 23:34:50,436 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=613522.0, ans=0.0 2024-09-24 23:34:57,039 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 23:35:01,689 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=613568.6666666666, ans=0.2 2024-09-24 23:35:03,207 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=613568.6666666666, ans=0.1 2024-09-24 23:35:07,635 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.262e+02 1.340e+02 1.438e+02 2.197e+02, threshold=2.679e+02, percent-clipped=0.0 2024-09-24 23:35:07,953 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=613568.6666666666, ans=0.2 2024-09-24 23:35:14,699 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.99 vs. limit=15.0 2024-09-24 23:35:21,082 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=7.36 vs. limit=15.0 2024-09-24 23:36:04,802 INFO [train.py:1198] (2/4) Epoch 34, batch 2950, loss[loss=0.2081, ctc_loss=0.1342, cr_loss=0.3695, over 17232.00 frames. ], tot_loss[loss=0.1962, ctc_loss=0.1275, cr_loss=0.3434, over 3330305.68 frames. ], batch size: 55, lr: 3.49e-03, grad_scale: 32.0 2024-09-24 23:36:13,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=613755.3333333334, ans=0.125 2024-09-24 23:36:14,678 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=613755.3333333334, ans=0.0 2024-09-24 23:37:10,462 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=613895.3333333334, ans=0.2 2024-09-24 23:37:19,752 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=613942.0, ans=0.2 2024-09-24 23:37:19,773 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=613942.0, ans=0.0 2024-09-24 23:37:32,108 INFO [train.py:1198] (2/4) Epoch 34, batch 3000, loss[loss=0.1909, ctc_loss=0.1235, cr_loss=0.3372, over 17252.00 frames. ], tot_loss[loss=0.1946, ctc_loss=0.1264, cr_loss=0.341, over 3334570.54 frames. ], batch size: 44, lr: 3.49e-03, grad_scale: 32.0 2024-09-24 23:37:32,109 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-24 23:37:47,933 INFO [train.py:1230] (2/4) Epoch 34, validation: loss=0.03583, ctc_loss=0.03583, cr_loss=9.471e-15, over 944034.00 frames. 2024-09-24 23:37:47,934 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-24 23:37:51,378 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=613988.6666666666, ans=10.0 2024-09-24 23:38:12,760 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.143e+02 1.290e+02 1.384e+02 1.469e+02 2.229e+02, threshold=2.767e+02, percent-clipped=0.0 2024-09-24 23:38:39,657 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=614128.6666666666, ans=0.1 2024-09-24 23:38:52,812 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=5.06 vs. limit=15.0 2024-09-24 23:38:58,452 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=614175.3333333334, ans=0.1 2024-09-24 23:39:01,440 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=614175.3333333334, ans=0.2 2024-09-24 23:39:05,968 INFO [train.py:1198] (2/4) Epoch 34, batch 3050, loss[loss=0.1963, ctc_loss=0.1302, cr_loss=0.3306, over 17299.00 frames. ], tot_loss[loss=0.1942, ctc_loss=0.1259, cr_loss=0.341, over 3339964.95 frames. ], batch size: 51, lr: 3.49e-03, grad_scale: 32.0 2024-09-24 23:39:34,231 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=614268.6666666666, ans=0.0 2024-09-24 23:39:34,304 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=614268.6666666666, ans=0.05 2024-09-24 23:39:46,711 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=614315.3333333334, ans=0.07 2024-09-24 23:40:02,314 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=614362.0, ans=0.125 2024-09-24 23:40:18,031 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=614408.6666666666, ans=0.1 2024-09-24 23:40:23,998 INFO [train.py:1198] (2/4) Epoch 34, batch 3100, loss[loss=0.2035, ctc_loss=0.1313, cr_loss=0.3608, over 17211.00 frames. ], tot_loss[loss=0.1943, ctc_loss=0.1261, cr_loss=0.3414, over 3348807.62 frames. ], batch size: 50, lr: 3.49e-03, grad_scale: 16.0 2024-09-24 23:40:24,268 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=614455.3333333334, ans=0.0 2024-09-24 23:40:30,343 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=614455.3333333334, ans=0.035 2024-09-24 23:40:41,366 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=614502.0, ans=0.015 2024-09-24 23:40:49,214 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=614502.0, ans=0.125 2024-09-24 23:40:49,262 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=614502.0, ans=0.0 2024-09-24 23:40:50,475 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.130e+02 1.258e+02 1.350e+02 1.404e+02 1.862e+02, threshold=2.700e+02, percent-clipped=0.0 2024-09-24 23:41:10,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=614595.3333333334, ans=0.125 2024-09-24 23:41:16,647 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=12.19 vs. limit=22.5 2024-09-24 23:41:33,003 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.46 vs. limit=15.0 2024-09-24 23:41:41,891 INFO [train.py:1198] (2/4) Epoch 34, batch 3150, loss[loss=0.195, ctc_loss=0.1247, cr_loss=0.3518, over 17158.00 frames. ], tot_loss[loss=0.1946, ctc_loss=0.1262, cr_loss=0.3421, over 3347178.49 frames. ], batch size: 48, lr: 3.48e-03, grad_scale: 16.0 2024-09-24 23:41:53,050 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=614688.6666666666, ans=0.125 2024-09-24 23:41:56,885 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.17 vs. limit=15.0 2024-09-24 23:42:17,014 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=614782.0, ans=0.1 2024-09-24 23:42:23,511 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 23:42:24,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=614782.0, ans=0.125 2024-09-24 23:42:26,687 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=614782.0, ans=0.125 2024-09-24 23:42:35,693 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=614828.6666666666, ans=0.125 2024-09-24 23:42:46,938 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=614875.3333333334, ans=0.125 2024-09-24 23:42:59,369 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=614922.0, ans=0.125 2024-09-24 23:43:00,718 INFO [train.py:1198] (2/4) Epoch 34, batch 3200, loss[loss=0.1848, ctc_loss=0.1194, cr_loss=0.3272, over 17289.00 frames. ], tot_loss[loss=0.1947, ctc_loss=0.1261, cr_loss=0.3429, over 3351695.85 frames. ], batch size: 46, lr: 3.48e-03, grad_scale: 32.0 2024-09-24 23:43:01,785 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.37 vs. limit=12.0 2024-09-24 23:43:14,977 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=614968.6666666666, ans=0.125 2024-09-24 23:43:27,299 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.284e+02 1.377e+02 1.475e+02 3.177e+02, threshold=2.753e+02, percent-clipped=2.0 2024-09-24 23:43:49,422 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=615062.0, ans=0.1 2024-09-24 23:43:55,701 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=615062.0, ans=0.125 2024-09-24 23:44:19,049 INFO [train.py:1198] (2/4) Epoch 34, batch 3250, loss[loss=0.1764, ctc_loss=0.1159, cr_loss=0.3026, over 17086.00 frames. ], tot_loss[loss=0.1957, ctc_loss=0.1269, cr_loss=0.3438, over 3356195.37 frames. ], batch size: 43, lr: 3.48e-03, grad_scale: 32.0 2024-09-24 23:44:52,612 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=615248.6666666666, ans=0.125 2024-09-24 23:44:54,638 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=13.42 vs. limit=15.0 2024-09-24 23:45:33,989 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=9.76 vs. limit=15.0 2024-09-24 23:45:39,574 INFO [train.py:1198] (2/4) Epoch 34, batch 3300, loss[loss=0.2077, ctc_loss=0.1333, cr_loss=0.3722, over 17139.00 frames. ], tot_loss[loss=0.1961, ctc_loss=0.1272, cr_loss=0.3441, over 3349538.07 frames. ], batch size: 48, lr: 3.48e-03, grad_scale: 32.0 2024-09-24 23:45:55,352 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=615435.3333333334, ans=0.025 2024-09-24 23:45:59,015 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=615435.3333333334, ans=0.025 2024-09-24 23:46:10,485 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.168e+02 1.276e+02 1.373e+02 1.543e+02 3.468e+02, threshold=2.745e+02, percent-clipped=1.0 2024-09-24 23:46:23,843 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=615482.0, ans=0.125 2024-09-24 23:46:51,053 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.57 vs. limit=15.0 2024-09-24 23:47:04,474 INFO [train.py:1198] (2/4) Epoch 34, batch 3350, loss[loss=0.1824, ctc_loss=0.1167, cr_loss=0.3286, over 17265.00 frames. ], tot_loss[loss=0.1965, ctc_loss=0.1277, cr_loss=0.344, over 3336415.83 frames. ], batch size: 44, lr: 3.48e-03, grad_scale: 32.0 2024-09-24 23:47:35,963 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=615715.3333333334, ans=0.0 2024-09-24 23:47:45,256 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=615715.3333333334, ans=0.125 2024-09-24 23:47:53,033 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=615762.0, ans=0.0 2024-09-24 23:48:12,317 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.08 vs. limit=22.5 2024-09-24 23:48:22,374 INFO [train.py:1198] (2/4) Epoch 34, batch 3400, loss[loss=0.1662, ctc_loss=0.1056, cr_loss=0.3032, over 16961.00 frames. ], tot_loss[loss=0.1953, ctc_loss=0.127, cr_loss=0.3416, over 3343339.18 frames. ], batch size: 42, lr: 3.48e-03, grad_scale: 32.0 2024-09-24 23:48:27,413 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=615855.3333333334, ans=0.2 2024-09-24 23:48:41,742 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.71 vs. limit=15.0 2024-09-24 23:48:48,722 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.112e+02 1.295e+02 1.404e+02 1.516e+02 2.292e+02, threshold=2.807e+02, percent-clipped=0.0 2024-09-24 23:49:17,640 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=615995.3333333334, ans=0.0 2024-09-24 23:49:23,640 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=615995.3333333334, ans=0.0 2024-09-24 23:49:42,037 INFO [train.py:1198] (2/4) Epoch 34, batch 3450, loss[loss=0.2036, ctc_loss=0.1286, cr_loss=0.3753, over 17297.00 frames. ], tot_loss[loss=0.1967, ctc_loss=0.128, cr_loss=0.3436, over 3333527.30 frames. ], batch size: 46, lr: 3.48e-03, grad_scale: 32.0 2024-09-24 23:50:05,685 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-24 23:50:25,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=616182.0, ans=0.2 2024-09-24 23:50:25,955 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=616182.0, ans=0.1 2024-09-24 23:50:37,166 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=616228.6666666666, ans=0.2 2024-09-24 23:50:54,502 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 23:50:54,768 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.56 vs. limit=10.0 2024-09-24 23:51:00,312 INFO [train.py:1198] (2/4) Epoch 34, batch 3500, loss[loss=0.1966, ctc_loss=0.1269, cr_loss=0.3486, over 17016.00 frames. ], tot_loss[loss=0.196, ctc_loss=0.1276, cr_loss=0.3418, over 3329867.56 frames. ], batch size: 44, lr: 3.48e-03, grad_scale: 32.0 2024-09-24 23:51:00,676 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=616322.0, ans=0.125 2024-09-24 23:51:05,708 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.06 vs. limit=15.0 2024-09-24 23:51:08,370 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=616322.0, ans=0.1 2024-09-24 23:51:10,547 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.52 vs. limit=10.0 2024-09-24 23:51:26,844 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.270e+02 1.396e+02 1.524e+02 2.184e+02, threshold=2.793e+02, percent-clipped=0.0 2024-09-24 23:51:28,893 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=616368.6666666666, ans=0.2 2024-09-24 23:51:31,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=616415.3333333334, ans=0.125 2024-09-24 23:51:33,528 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=616415.3333333334, ans=0.2 2024-09-24 23:51:42,804 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=616415.3333333334, ans=0.125 2024-09-24 23:51:55,922 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=3.45 vs. limit=15.0 2024-09-24 23:52:03,734 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.84 vs. limit=22.5 2024-09-24 23:52:10,844 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=616508.6666666666, ans=0.125 2024-09-24 23:52:17,301 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=616555.3333333334, ans=0.125 2024-09-24 23:52:18,496 INFO [train.py:1198] (2/4) Epoch 34, batch 3550, loss[loss=0.2426, ctc_loss=0.1721, cr_loss=0.3525, over 11646.00 frames. ], tot_loss[loss=0.1964, ctc_loss=0.1278, cr_loss=0.3428, over 3337362.72 frames. ], batch size: 123, lr: 3.48e-03, grad_scale: 32.0 2024-09-24 23:52:26,526 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=616555.3333333334, ans=0.125 2024-09-24 23:52:28,552 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.06 vs. limit=15.0 2024-09-24 23:52:41,863 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=616602.0, ans=0.125 2024-09-24 23:53:09,209 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.27 vs. limit=15.0 2024-09-24 23:53:36,430 INFO [train.py:1198] (2/4) Epoch 34, batch 3600, loss[loss=0.1944, ctc_loss=0.1274, cr_loss=0.3348, over 17208.00 frames. ], tot_loss[loss=0.1961, ctc_loss=0.1276, cr_loss=0.3426, over 3339391.91 frames. ], batch size: 55, lr: 3.48e-03, grad_scale: 32.0 2024-09-24 23:53:55,540 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=616835.3333333334, ans=0.2 2024-09-24 23:54:03,052 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.287e+02 1.340e+02 1.449e+02 1.947e+02, threshold=2.680e+02, percent-clipped=0.0 2024-09-24 23:54:19,037 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=616882.0, ans=0.1 2024-09-24 23:54:41,836 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=616975.3333333334, ans=0.125 2024-09-24 23:54:56,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=617022.0, ans=0.0 2024-09-24 23:54:57,437 INFO [train.py:1198] (2/4) Epoch 34, batch 3650, loss[loss=0.195, ctc_loss=0.1278, cr_loss=0.3356, over 17173.00 frames. ], tot_loss[loss=0.1971, ctc_loss=0.1282, cr_loss=0.3444, over 3341174.61 frames. ], batch size: 45, lr: 3.48e-03, grad_scale: 32.0 2024-09-24 23:56:09,367 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=617208.6666666666, ans=0.125 2024-09-24 23:56:10,801 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=617208.6666666666, ans=0.1 2024-09-24 23:56:21,544 INFO [train.py:1198] (2/4) Epoch 34, batch 3700, loss[loss=0.181, ctc_loss=0.1153, cr_loss=0.3285, over 17065.00 frames. ], tot_loss[loss=0.1976, ctc_loss=0.1287, cr_loss=0.3448, over 3327358.26 frames. ], batch size: 46, lr: 3.48e-03, grad_scale: 32.0 2024-09-24 23:56:32,667 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=617255.3333333334, ans=0.1 2024-09-24 23:56:46,872 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=617302.0, ans=0.125 2024-09-24 23:56:48,014 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.065e+02 1.259e+02 1.354e+02 1.435e+02 3.016e+02, threshold=2.708e+02, percent-clipped=2.0 2024-09-24 23:56:51,429 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 23:57:39,592 INFO [train.py:1198] (2/4) Epoch 34, batch 3750, loss[loss=0.1926, ctc_loss=0.1266, cr_loss=0.3299, over 17166.00 frames. ], tot_loss[loss=0.1971, ctc_loss=0.1283, cr_loss=0.3442, over 3328458.50 frames. ], batch size: 45, lr: 3.48e-03, grad_scale: 32.0 2024-09-24 23:57:44,509 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=617488.6666666666, ans=0.1 2024-09-24 23:57:57,464 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.27 vs. limit=10.0 2024-09-24 23:58:08,051 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=15.32 vs. limit=22.5 2024-09-24 23:58:13,768 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=617582.0, ans=0.09899494936611666 2024-09-24 23:58:27,859 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=617628.6666666666, ans=0.125 2024-09-24 23:58:30,906 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=617628.6666666666, ans=0.0 2024-09-24 23:58:56,794 INFO [train.py:1198] (2/4) Epoch 34, batch 3800, loss[loss=0.2049, ctc_loss=0.138, cr_loss=0.3347, over 16879.00 frames. ], tot_loss[loss=0.197, ctc_loss=0.1283, cr_loss=0.3439, over 3322072.25 frames. ], batch size: 58, lr: 3.48e-03, grad_scale: 32.0 2024-09-24 23:59:09,702 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-24 23:59:23,297 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.092e+02 1.281e+02 1.379e+02 1.537e+02 2.661e+02, threshold=2.757e+02, percent-clipped=0.0 2024-09-24 23:59:26,820 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-24 23:59:42,618 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.max_positive, batch_count=617862.0, ans=0.95 2024-09-24 23:59:48,016 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=11.80 vs. limit=22.5 2024-09-24 23:59:49,926 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=9.71 vs. limit=15.0 2024-09-25 00:00:08,272 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=617908.6666666666, ans=0.125 2024-09-25 00:00:14,447 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=617955.3333333334, ans=0.2 2024-09-25 00:00:15,668 INFO [train.py:1198] (2/4) Epoch 34, batch 3850, loss[loss=0.2332, ctc_loss=0.1567, cr_loss=0.3829, over 12585.00 frames. ], tot_loss[loss=0.199, ctc_loss=0.1297, cr_loss=0.3463, over 3284546.29 frames. ], batch size: 123, lr: 3.48e-03, grad_scale: 32.0 2024-09-25 00:00:23,707 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=617955.3333333334, ans=0.04949747468305833 2024-09-25 00:00:30,344 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=7.40 vs. limit=15.0 2024-09-25 00:00:37,770 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=618002.0, ans=0.0 2024-09-25 00:00:39,133 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=618002.0, ans=0.125 2024-09-25 00:00:39,136 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=618002.0, ans=0.125 2024-09-25 00:00:55,868 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=618048.6666666666, ans=0.1 2024-09-25 00:01:03,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=618095.3333333334, ans=0.1 2024-09-25 00:01:03,486 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=618095.3333333334, ans=0.0 2024-09-25 00:01:07,868 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=618095.3333333334, ans=0.025 2024-09-25 00:01:07,922 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=618095.3333333334, ans=0.0 2024-09-25 00:01:12,360 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=618095.3333333334, ans=0.0 2024-09-25 00:02:16,955 INFO [train.py:1198] (2/4) Epoch 35, batch 0, loss[loss=0.2105, ctc_loss=0.1349, cr_loss=0.3779, over 17062.00 frames. ], tot_loss[loss=0.2105, ctc_loss=0.1349, cr_loss=0.3779, over 17062.00 frames. ], batch size: 46, lr: 3.42e-03, grad_scale: 32.0 2024-09-25 00:02:16,956 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 00:02:32,188 INFO [train.py:1230] (2/4) Epoch 35, validation: loss=0.03449, ctc_loss=0.03449, cr_loss=9.757e-15, over 944034.00 frames. 2024-09-25 00:02:32,189 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 00:02:47,236 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=11.18 vs. limit=22.5 2024-09-25 00:03:08,932 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=618263.3333333334, ans=0.09899494936611666 2024-09-25 00:03:10,088 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.150e+02 1.401e+02 1.522e+02 1.667e+02 2.435e+02, threshold=3.044e+02, percent-clipped=0.0 2024-09-25 00:03:17,067 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 00:03:56,376 INFO [train.py:1198] (2/4) Epoch 35, batch 50, loss[loss=0.1817, ctc_loss=0.1129, cr_loss=0.3439, over 17097.00 frames. ], tot_loss[loss=0.2016, ctc_loss=0.1308, cr_loss=0.3537, over 764802.91 frames. ], batch size: 40, lr: 3.42e-03, grad_scale: 32.0 2024-09-25 00:04:08,058 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=618403.3333333334, ans=0.0 2024-09-25 00:04:09,727 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=618403.3333333334, ans=0.2 2024-09-25 00:04:47,824 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=618543.3333333334, ans=0.125 2024-09-25 00:05:16,501 INFO [train.py:1198] (2/4) Epoch 35, batch 100, loss[loss=0.2057, ctc_loss=0.1325, cr_loss=0.3663, over 16909.00 frames. ], tot_loss[loss=0.1964, ctc_loss=0.1271, cr_loss=0.3464, over 1344240.02 frames. ], batch size: 58, lr: 3.42e-03, grad_scale: 32.0 2024-09-25 00:05:18,428 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 00:05:27,874 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=618636.6666666666, ans=0.125 2024-09-25 00:05:29,926 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.43 vs. limit=12.0 2024-09-25 00:05:34,220 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=618683.3333333334, ans=0.0 2024-09-25 00:05:37,626 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 00:05:49,884 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.242e+02 1.304e+02 1.413e+02 1.730e+02, threshold=2.607e+02, percent-clipped=0.0 2024-09-25 00:06:10,025 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.66 vs. limit=15.0 2024-09-25 00:06:17,717 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=618776.6666666666, ans=0.0 2024-09-25 00:06:31,381 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=618823.3333333334, ans=0.025 2024-09-25 00:06:38,903 INFO [train.py:1198] (2/4) Epoch 35, batch 150, loss[loss=0.1865, ctc_loss=0.1203, cr_loss=0.3306, over 17220.00 frames. ], tot_loss[loss=0.1964, ctc_loss=0.1272, cr_loss=0.3458, over 1780235.34 frames. ], batch size: 50, lr: 3.42e-03, grad_scale: 32.0 2024-09-25 00:06:42,561 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=618870.0, ans=0.0 2024-09-25 00:06:43,167 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.64 vs. limit=12.0 2024-09-25 00:07:08,069 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=618916.6666666666, ans=0.1 2024-09-25 00:07:22,810 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=618963.3333333334, ans=0.2 2024-09-25 00:08:05,713 INFO [train.py:1198] (2/4) Epoch 35, batch 200, loss[loss=0.1842, ctc_loss=0.1214, cr_loss=0.314, over 17100.00 frames. ], tot_loss[loss=0.1971, ctc_loss=0.1277, cr_loss=0.3474, over 2132091.70 frames. ], batch size: 49, lr: 3.42e-03, grad_scale: 32.0 2024-09-25 00:08:12,797 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=619103.3333333334, ans=0.09899494936611666 2024-09-25 00:08:14,302 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=619103.3333333334, ans=0.125 2024-09-25 00:08:27,732 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.89 vs. limit=10.0 2024-09-25 00:08:34,914 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=619150.0, ans=0.1 2024-09-25 00:08:41,112 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.092e+02 1.252e+02 1.342e+02 1.492e+02 1.753e+02, threshold=2.683e+02, percent-clipped=0.0 2024-09-25 00:08:57,480 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=619243.3333333334, ans=0.125 2024-09-25 00:09:02,270 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=619243.3333333334, ans=0.1 2024-09-25 00:09:05,457 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=619243.3333333334, ans=0.125 2024-09-25 00:09:11,747 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=619290.0, ans=0.125 2024-09-25 00:09:27,867 INFO [train.py:1198] (2/4) Epoch 35, batch 250, loss[loss=0.1945, ctc_loss=0.1251, cr_loss=0.347, over 16806.00 frames. ], tot_loss[loss=0.1958, ctc_loss=0.1267, cr_loss=0.3454, over 2408376.03 frames. ], batch size: 61, lr: 3.42e-03, grad_scale: 32.0 2024-09-25 00:09:39,913 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=13.20 vs. limit=15.0 2024-09-25 00:10:04,725 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=619430.0, ans=0.0 2024-09-25 00:10:20,479 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.min_positive, batch_count=619476.6666666666, ans=0.025 2024-09-25 00:10:39,764 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=619523.3333333334, ans=0.125 2024-09-25 00:10:47,175 INFO [train.py:1198] (2/4) Epoch 35, batch 300, loss[loss=0.2056, ctc_loss=0.1356, cr_loss=0.3496, over 17125.00 frames. ], tot_loss[loss=0.1946, ctc_loss=0.1258, cr_loss=0.344, over 2627286.85 frames. ], batch size: 48, lr: 3.42e-03, grad_scale: 32.0 2024-09-25 00:11:13,309 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=619616.6666666666, ans=0.1 2024-09-25 00:11:20,985 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.276e+02 1.333e+02 1.416e+02 3.334e+02, threshold=2.666e+02, percent-clipped=1.0 2024-09-25 00:11:43,414 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=619710.0, ans=0.125 2024-09-25 00:12:02,596 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=619756.6666666666, ans=0.0 2024-09-25 00:12:10,250 INFO [train.py:1198] (2/4) Epoch 35, batch 350, loss[loss=0.2303, ctc_loss=0.1503, cr_loss=0.4001, over 16412.00 frames. ], tot_loss[loss=0.195, ctc_loss=0.1262, cr_loss=0.344, over 2784813.94 frames. ], batch size: 66, lr: 3.42e-03, grad_scale: 32.0 2024-09-25 00:12:15,797 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=13.30 vs. limit=22.5 2024-09-25 00:13:32,912 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=619990.0, ans=0.5 2024-09-25 00:13:33,750 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.54 vs. limit=12.0 2024-09-25 00:13:39,191 INFO [train.py:1198] (2/4) Epoch 35, batch 400, loss[loss=0.1861, ctc_loss=0.121, cr_loss=0.3256, over 17183.00 frames. ], tot_loss[loss=0.1952, ctc_loss=0.1263, cr_loss=0.3444, over 2912521.48 frames. ], batch size: 41, lr: 3.42e-03, grad_scale: 32.0 2024-09-25 00:13:47,741 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.63 vs. limit=15.0 2024-09-25 00:13:58,544 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=620083.3333333334, ans=0.95 2024-09-25 00:14:09,433 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=620130.0, ans=0.125 2024-09-25 00:14:12,440 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.132e+02 1.297e+02 1.357e+02 1.460e+02 2.001e+02, threshold=2.715e+02, percent-clipped=0.0 2024-09-25 00:14:20,799 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=620130.0, ans=0.1 2024-09-25 00:14:26,220 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=7.99 vs. limit=15.0 2024-09-25 00:14:39,006 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.02 vs. limit=6.0 2024-09-25 00:14:52,026 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=5.56 vs. limit=15.0 2024-09-25 00:14:58,006 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=620270.0, ans=0.125 2024-09-25 00:14:59,149 INFO [train.py:1198] (2/4) Epoch 35, batch 450, loss[loss=0.2631, ctc_loss=0.1815, cr_loss=0.4078, over 12114.00 frames. ], tot_loss[loss=0.194, ctc_loss=0.1256, cr_loss=0.3418, over 3009562.81 frames. ], batch size: 123, lr: 3.42e-03, grad_scale: 16.0 2024-09-25 00:15:17,016 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=620316.6666666666, ans=0.0 2024-09-25 00:15:20,250 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=620316.6666666666, ans=0.1 2024-09-25 00:15:20,714 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.67 vs. limit=15.0 2024-09-25 00:15:30,041 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=620363.3333333334, ans=0.125 2024-09-25 00:16:01,956 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=620456.6666666666, ans=0.125 2024-09-25 00:16:06,722 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=620456.6666666666, ans=0.125 2024-09-25 00:16:08,421 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=620456.6666666666, ans=0.0 2024-09-25 00:16:14,754 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=620456.6666666666, ans=0.0 2024-09-25 00:16:19,362 INFO [train.py:1198] (2/4) Epoch 35, batch 500, loss[loss=0.2129, ctc_loss=0.1399, cr_loss=0.365, over 17028.00 frames. ], tot_loss[loss=0.1956, ctc_loss=0.1267, cr_loss=0.3445, over 3087537.24 frames. ], batch size: 56, lr: 3.42e-03, grad_scale: 16.0 2024-09-25 00:16:39,810 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=620550.0, ans=0.2 2024-09-25 00:16:57,098 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.242e+02 1.333e+02 1.438e+02 2.516e+02, threshold=2.666e+02, percent-clipped=0.0 2024-09-25 00:17:11,903 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=620643.3333333334, ans=0.125 2024-09-25 00:17:20,026 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=620643.3333333334, ans=0.09899494936611666 2024-09-25 00:17:20,769 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=8.37 vs. limit=15.0 2024-09-25 00:17:22,089 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.80 vs. limit=15.0 2024-09-25 00:17:32,319 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=620690.0, ans=0.125 2024-09-25 00:17:39,009 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=620690.0, ans=0.0 2024-09-25 00:17:44,911 INFO [train.py:1198] (2/4) Epoch 35, batch 550, loss[loss=0.1843, ctc_loss=0.1189, cr_loss=0.3269, over 17058.00 frames. ], tot_loss[loss=0.1959, ctc_loss=0.1269, cr_loss=0.3449, over 3144703.61 frames. ], batch size: 46, lr: 3.42e-03, grad_scale: 16.0 2024-09-25 00:17:48,396 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=620736.6666666666, ans=0.1 2024-09-25 00:18:11,618 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=620783.3333333334, ans=0.0 2024-09-25 00:18:25,820 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=620830.0, ans=0.2 2024-09-25 00:18:37,065 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=620876.6666666666, ans=0.125 2024-09-25 00:18:38,759 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=620876.6666666666, ans=0.0 2024-09-25 00:19:06,483 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=15.08 vs. limit=22.5 2024-09-25 00:19:10,361 INFO [train.py:1198] (2/4) Epoch 35, batch 600, loss[loss=0.1882, ctc_loss=0.1205, cr_loss=0.3387, over 17179.00 frames. ], tot_loss[loss=0.1952, ctc_loss=0.1265, cr_loss=0.3438, over 3185624.62 frames. ], batch size: 41, lr: 3.42e-03, grad_scale: 16.0 2024-09-25 00:19:20,890 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.17 vs. limit=15.0 2024-09-25 00:19:42,622 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=621063.3333333334, ans=0.0 2024-09-25 00:19:45,442 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.145e+02 1.279e+02 1.403e+02 1.511e+02 1.952e+02, threshold=2.806e+02, percent-clipped=0.0 2024-09-25 00:20:24,025 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=621156.6666666666, ans=0.125 2024-09-25 00:20:30,168 INFO [train.py:1198] (2/4) Epoch 35, batch 650, loss[loss=0.1961, ctc_loss=0.1267, cr_loss=0.3471, over 17147.00 frames. ], tot_loss[loss=0.1955, ctc_loss=0.1267, cr_loss=0.3441, over 3220398.75 frames. ], batch size: 48, lr: 3.42e-03, grad_scale: 16.0 2024-09-25 00:20:30,438 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=621203.3333333334, ans=0.025 2024-09-25 00:20:30,563 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=621203.3333333334, ans=0.2 2024-09-25 00:20:37,505 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=13.83 vs. limit=22.5 2024-09-25 00:21:12,246 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=621296.6666666666, ans=0.0 2024-09-25 00:21:32,120 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=13.05 vs. limit=22.5 2024-09-25 00:21:52,479 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=18.89 vs. limit=22.5 2024-09-25 00:21:53,028 INFO [train.py:1198] (2/4) Epoch 35, batch 700, loss[loss=0.229, ctc_loss=0.1517, cr_loss=0.3864, over 15146.00 frames. ], tot_loss[loss=0.1954, ctc_loss=0.1266, cr_loss=0.344, over 3244900.95 frames. ], batch size: 89, lr: 3.41e-03, grad_scale: 16.0 2024-09-25 00:21:58,250 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=621436.6666666666, ans=0.2 2024-09-25 00:22:27,884 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=621530.0, ans=0.1 2024-09-25 00:22:30,421 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=12.49 vs. limit=22.5 2024-09-25 00:22:30,895 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.276e+02 1.349e+02 1.443e+02 1.700e+02, threshold=2.699e+02, percent-clipped=0.0 2024-09-25 00:22:55,556 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.75 vs. limit=10.0 2024-09-25 00:23:18,473 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=621623.3333333334, ans=0.025 2024-09-25 00:23:21,254 INFO [train.py:1198] (2/4) Epoch 35, batch 750, loss[loss=0.1687, ctc_loss=0.1073, cr_loss=0.307, over 17071.00 frames. ], tot_loss[loss=0.1946, ctc_loss=0.126, cr_loss=0.343, over 3279580.20 frames. ], batch size: 46, lr: 3.41e-03, grad_scale: 16.0 2024-09-25 00:23:53,731 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 00:24:41,687 INFO [train.py:1198] (2/4) Epoch 35, batch 800, loss[loss=0.1924, ctc_loss=0.125, cr_loss=0.3371, over 17022.00 frames. ], tot_loss[loss=0.195, ctc_loss=0.1263, cr_loss=0.3437, over 3299181.78 frames. ], batch size: 44, lr: 3.41e-03, grad_scale: 32.0 2024-09-25 00:24:41,791 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=621903.3333333334, ans=0.125 2024-09-25 00:24:43,646 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=621903.3333333334, ans=0.09899494936611666 2024-09-25 00:24:44,098 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.86 vs. limit=6.0 2024-09-25 00:24:50,703 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.20 vs. limit=22.5 2024-09-25 00:25:16,940 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.280e+02 1.366e+02 1.481e+02 2.443e+02, threshold=2.732e+02, percent-clipped=0.0 2024-09-25 00:25:18,859 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=621996.6666666666, ans=0.125 2024-09-25 00:25:31,778 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=622043.3333333334, ans=0.125 2024-09-25 00:25:34,961 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=622043.3333333334, ans=0.1 2024-09-25 00:25:46,939 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.36 vs. limit=15.0 2024-09-25 00:26:01,105 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.88 vs. limit=6.0 2024-09-25 00:26:01,918 INFO [train.py:1198] (2/4) Epoch 35, batch 850, loss[loss=0.1586, ctc_loss=0.09798, cr_loss=0.3032, over 16952.00 frames. ], tot_loss[loss=0.1946, ctc_loss=0.126, cr_loss=0.3431, over 3308107.24 frames. ], batch size: 42, lr: 3.41e-03, grad_scale: 16.0 2024-09-25 00:26:02,461 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=10.36 vs. limit=15.0 2024-09-25 00:26:33,995 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=622230.0, ans=0.125 2024-09-25 00:26:44,921 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=622230.0, ans=0.125 2024-09-25 00:27:15,441 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=13.40 vs. limit=22.5 2024-09-25 00:27:17,384 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.74 vs. limit=5.0 2024-09-25 00:27:25,339 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=622370.0, ans=0.125 2024-09-25 00:27:26,639 INFO [train.py:1198] (2/4) Epoch 35, batch 900, loss[loss=0.1805, ctc_loss=0.1186, cr_loss=0.3091, over 15888.00 frames. ], tot_loss[loss=0.1954, ctc_loss=0.1266, cr_loss=0.344, over 3313431.84 frames. ], batch size: 74, lr: 3.41e-03, grad_scale: 16.0 2024-09-25 00:28:06,014 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=622463.3333333334, ans=0.04949747468305833 2024-09-25 00:28:08,746 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.186e+02 1.305e+02 1.377e+02 1.455e+02 1.762e+02, threshold=2.753e+02, percent-clipped=0.0 2024-09-25 00:28:15,501 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=622463.3333333334, ans=0.125 2024-09-25 00:28:21,837 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=622510.0, ans=0.125 2024-09-25 00:28:52,251 INFO [train.py:1198] (2/4) Epoch 35, batch 950, loss[loss=0.1929, ctc_loss=0.1239, cr_loss=0.3453, over 17094.00 frames. ], tot_loss[loss=0.1953, ctc_loss=0.1265, cr_loss=0.3439, over 3323658.86 frames. ], batch size: 49, lr: 3.41e-03, grad_scale: 16.0 2024-09-25 00:28:57,323 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=622603.3333333334, ans=0.125 2024-09-25 00:29:20,080 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=622650.0, ans=0.125 2024-09-25 00:29:24,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=622696.6666666666, ans=0.125 2024-09-25 00:29:35,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.min_positive, batch_count=622696.6666666666, ans=0.05 2024-09-25 00:29:42,511 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=622743.3333333334, ans=0.125 2024-09-25 00:29:46,275 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.22 vs. limit=15.0 2024-09-25 00:29:55,377 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=622790.0, ans=0.1 2024-09-25 00:30:03,612 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=9.83 vs. limit=12.0 2024-09-25 00:30:12,544 INFO [train.py:1198] (2/4) Epoch 35, batch 1000, loss[loss=0.1738, ctc_loss=0.1104, cr_loss=0.3172, over 16977.00 frames. ], tot_loss[loss=0.1951, ctc_loss=0.1265, cr_loss=0.3434, over 3324405.42 frames. ], batch size: 42, lr: 3.41e-03, grad_scale: 16.0 2024-09-25 00:30:38,293 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=622883.3333333334, ans=0.125 2024-09-25 00:30:49,042 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.266e+02 1.340e+02 1.444e+02 2.744e+02, threshold=2.680e+02, percent-clipped=0.0 2024-09-25 00:31:10,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=622976.6666666666, ans=0.125 2024-09-25 00:31:35,000 INFO [train.py:1198] (2/4) Epoch 35, batch 1050, loss[loss=0.1817, ctc_loss=0.1162, cr_loss=0.3277, over 17010.00 frames. ], tot_loss[loss=0.1956, ctc_loss=0.1268, cr_loss=0.3441, over 3331751.67 frames. ], batch size: 44, lr: 3.41e-03, grad_scale: 16.0 2024-09-25 00:31:49,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=623116.6666666666, ans=0.2 2024-09-25 00:31:51,087 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=623116.6666666666, ans=0.125 2024-09-25 00:32:46,688 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=623256.6666666666, ans=0.05 2024-09-25 00:32:50,817 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=623256.6666666666, ans=0.0 2024-09-25 00:32:55,622 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=623256.6666666666, ans=0.2 2024-09-25 00:33:02,768 INFO [train.py:1198] (2/4) Epoch 35, batch 1100, loss[loss=0.1947, ctc_loss=0.1259, cr_loss=0.3441, over 17066.00 frames. ], tot_loss[loss=0.1955, ctc_loss=0.1266, cr_loss=0.3443, over 3336150.55 frames. ], batch size: 46, lr: 3.41e-03, grad_scale: 16.0 2024-09-25 00:33:20,863 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.96 vs. limit=6.0 2024-09-25 00:33:37,860 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=623396.6666666666, ans=0.1 2024-09-25 00:33:39,206 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.286e+02 1.347e+02 1.466e+02 2.468e+02, threshold=2.694e+02, percent-clipped=0.0 2024-09-25 00:33:55,591 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=623443.3333333334, ans=0.0 2024-09-25 00:34:17,028 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.43 vs. limit=15.0 2024-09-25 00:34:22,932 INFO [train.py:1198] (2/4) Epoch 35, batch 1150, loss[loss=0.1798, ctc_loss=0.116, cr_loss=0.3191, over 17235.00 frames. ], tot_loss[loss=0.195, ctc_loss=0.1263, cr_loss=0.3431, over 3328793.12 frames. ], batch size: 50, lr: 3.41e-03, grad_scale: 16.0 2024-09-25 00:34:32,650 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=623536.6666666666, ans=0.0 2024-09-25 00:34:46,907 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=623583.3333333334, ans=0.1 2024-09-25 00:35:42,222 INFO [train.py:1198] (2/4) Epoch 35, batch 1200, loss[loss=0.1964, ctc_loss=0.1254, cr_loss=0.355, over 16512.00 frames. ], tot_loss[loss=0.1946, ctc_loss=0.126, cr_loss=0.3428, over 3338456.80 frames. ], batch size: 66, lr: 3.41e-03, grad_scale: 32.0 2024-09-25 00:35:58,500 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 00:35:58,530 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=623816.6666666666, ans=0.125 2024-09-25 00:36:18,955 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.278e+02 1.375e+02 1.481e+02 3.565e+02, threshold=2.751e+02, percent-clipped=1.0 2024-09-25 00:36:27,267 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=623863.3333333334, ans=0.0 2024-09-25 00:36:36,364 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=623910.0, ans=0.125 2024-09-25 00:36:39,635 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 00:37:04,232 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=8.44 vs. limit=15.0 2024-09-25 00:37:04,858 INFO [train.py:1198] (2/4) Epoch 35, batch 1250, loss[loss=0.1987, ctc_loss=0.1307, cr_loss=0.3397, over 17229.00 frames. ], tot_loss[loss=0.1955, ctc_loss=0.1267, cr_loss=0.3442, over 3348368.12 frames. ], batch size: 55, lr: 3.41e-03, grad_scale: 32.0 2024-09-25 00:37:10,087 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=624003.3333333334, ans=0.0 2024-09-25 00:37:18,665 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=624003.3333333334, ans=0.125 2024-09-25 00:37:34,506 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=624050.0, ans=0.0 2024-09-25 00:37:42,679 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=624096.6666666666, ans=0.125 2024-09-25 00:37:44,236 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=624096.6666666666, ans=0.1 2024-09-25 00:37:54,770 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=624096.6666666666, ans=0.125 2024-09-25 00:37:56,339 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=624143.3333333334, ans=0.125 2024-09-25 00:37:56,464 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=624143.3333333334, ans=0.0 2024-09-25 00:38:08,413 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=624143.3333333334, ans=0.015 2024-09-25 00:38:13,432 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=624143.3333333334, ans=0.04949747468305833 2024-09-25 00:38:17,424 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.83 vs. limit=6.0 2024-09-25 00:38:31,007 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 00:38:32,164 INFO [train.py:1198] (2/4) Epoch 35, batch 1300, loss[loss=0.1951, ctc_loss=0.1288, cr_loss=0.3319, over 16993.00 frames. ], tot_loss[loss=0.1941, ctc_loss=0.1257, cr_loss=0.342, over 3357862.28 frames. ], batch size: 53, lr: 3.41e-03, grad_scale: 32.0 2024-09-25 00:38:42,043 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=624236.6666666666, ans=0.0 2024-09-25 00:38:48,215 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.max_positive, batch_count=624283.3333333334, ans=0.95 2024-09-25 00:38:57,984 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=624283.3333333334, ans=0.125 2024-09-25 00:39:08,850 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.267e+02 1.378e+02 1.452e+02 1.774e+02, threshold=2.755e+02, percent-clipped=0.0 2024-09-25 00:39:09,118 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=624330.0, ans=0.0 2024-09-25 00:39:12,490 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=624330.0, ans=0.0 2024-09-25 00:39:52,134 INFO [train.py:1198] (2/4) Epoch 35, batch 1350, loss[loss=0.1956, ctc_loss=0.1259, cr_loss=0.3487, over 17296.00 frames. ], tot_loss[loss=0.1948, ctc_loss=0.1262, cr_loss=0.3431, over 3362930.80 frames. ], batch size: 49, lr: 3.41e-03, grad_scale: 32.0 2024-09-25 00:39:58,561 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=624470.0, ans=0.1 2024-09-25 00:39:59,659 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=12.46 vs. limit=22.5 2024-09-25 00:40:10,278 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=624516.6666666666, ans=0.125 2024-09-25 00:40:23,200 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=624563.3333333334, ans=0.025 2024-09-25 00:40:39,771 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=5.70 vs. limit=15.0 2024-09-25 00:40:44,066 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.48 vs. limit=15.0 2024-09-25 00:40:45,436 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=624610.0, ans=0.0 2024-09-25 00:41:07,882 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=4.49 vs. limit=12.0 2024-09-25 00:41:11,914 INFO [train.py:1198] (2/4) Epoch 35, batch 1400, loss[loss=0.1964, ctc_loss=0.1253, cr_loss=0.3556, over 17167.00 frames. ], tot_loss[loss=0.1946, ctc_loss=0.126, cr_loss=0.3431, over 3371537.08 frames. ], batch size: 45, lr: 3.41e-03, grad_scale: 32.0 2024-09-25 00:41:12,221 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=624703.3333333334, ans=0.125 2024-09-25 00:41:20,401 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=624703.3333333334, ans=0.125 2024-09-25 00:41:23,662 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=624703.3333333334, ans=0.125 2024-09-25 00:41:51,231 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.153e+02 1.292e+02 1.373e+02 1.479e+02 2.692e+02, threshold=2.745e+02, percent-clipped=0.0 2024-09-25 00:41:51,673 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=624796.6666666666, ans=0.0 2024-09-25 00:41:54,888 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=624796.6666666666, ans=0.0 2024-09-25 00:41:59,490 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=624796.6666666666, ans=0.125 2024-09-25 00:42:02,724 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=624843.3333333334, ans=0.125 2024-09-25 00:42:04,767 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=5.32 vs. limit=12.0 2024-09-25 00:42:24,787 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=624890.0, ans=0.0 2024-09-25 00:42:33,105 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=5.19 vs. limit=15.0 2024-09-25 00:42:37,117 INFO [train.py:1198] (2/4) Epoch 35, batch 1450, loss[loss=0.1679, ctc_loss=0.1071, cr_loss=0.3038, over 17210.00 frames. ], tot_loss[loss=0.1949, ctc_loss=0.1262, cr_loss=0.3435, over 3379097.17 frames. ], batch size: 41, lr: 3.41e-03, grad_scale: 32.0 2024-09-25 00:43:14,356 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=625030.0, ans=0.07 2024-09-25 00:43:22,423 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=625030.0, ans=0.0 2024-09-25 00:43:38,537 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=625076.6666666666, ans=0.125 2024-09-25 00:44:00,007 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.49 vs. limit=6.0 2024-09-25 00:44:02,341 INFO [train.py:1198] (2/4) Epoch 35, batch 1500, loss[loss=0.1835, ctc_loss=0.1153, cr_loss=0.3409, over 17303.00 frames. ], tot_loss[loss=0.195, ctc_loss=0.1262, cr_loss=0.3438, over 3380586.06 frames. ], batch size: 46, lr: 3.40e-03, grad_scale: 32.0 2024-09-25 00:44:05,783 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=625170.0, ans=0.0 2024-09-25 00:44:39,192 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.069e+02 1.249e+02 1.364e+02 1.434e+02 2.230e+02, threshold=2.728e+02, percent-clipped=0.0 2024-09-25 00:44:42,597 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=625263.3333333334, ans=0.0 2024-09-25 00:44:53,949 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=625310.0, ans=0.0 2024-09-25 00:45:00,844 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=625310.0, ans=0.1 2024-09-25 00:45:08,974 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=625356.6666666666, ans=0.125 2024-09-25 00:45:23,010 INFO [train.py:1198] (2/4) Epoch 35, batch 1550, loss[loss=0.1679, ctc_loss=0.1056, cr_loss=0.3117, over 17279.00 frames. ], tot_loss[loss=0.1958, ctc_loss=0.1269, cr_loss=0.3446, over 3365533.06 frames. ], batch size: 42, lr: 3.40e-03, grad_scale: 32.0 2024-09-25 00:45:24,970 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=625403.3333333334, ans=0.125 2024-09-25 00:45:42,542 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=625450.0, ans=0.0 2024-09-25 00:46:22,776 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=625543.3333333334, ans=0.125 2024-09-25 00:46:29,048 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=625590.0, ans=0.125 2024-09-25 00:46:45,798 INFO [train.py:1198] (2/4) Epoch 35, batch 1600, loss[loss=0.1851, ctc_loss=0.1184, cr_loss=0.3333, over 17175.00 frames. ], tot_loss[loss=0.1963, ctc_loss=0.1272, cr_loss=0.3453, over 3366955.08 frames. ], batch size: 45, lr: 3.40e-03, grad_scale: 32.0 2024-09-25 00:47:00,426 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=625683.3333333334, ans=0.125 2024-09-25 00:47:25,309 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.121e+02 1.273e+02 1.348e+02 1.440e+02 1.761e+02, threshold=2.695e+02, percent-clipped=0.0 2024-09-25 00:47:27,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=625730.0, ans=0.0 2024-09-25 00:47:47,245 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=10.54 vs. limit=15.0 2024-09-25 00:48:09,822 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=625823.3333333334, ans=0.125 2024-09-25 00:48:13,207 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=625870.0, ans=0.125 2024-09-25 00:48:14,374 INFO [train.py:1198] (2/4) Epoch 35, batch 1650, loss[loss=0.1643, ctc_loss=0.1062, cr_loss=0.2908, over 17085.00 frames. ], tot_loss[loss=0.1953, ctc_loss=0.1265, cr_loss=0.3441, over 3373197.15 frames. ], batch size: 46, lr: 3.40e-03, grad_scale: 32.0 2024-09-25 00:48:25,741 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=625870.0, ans=0.125 2024-09-25 00:48:25,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=625870.0, ans=0.0 2024-09-25 00:48:32,276 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=625916.6666666666, ans=0.1 2024-09-25 00:48:37,100 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=625916.6666666666, ans=0.1 2024-09-25 00:49:12,349 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=626010.0, ans=0.125 2024-09-25 00:49:15,366 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=626010.0, ans=0.125 2024-09-25 00:49:15,967 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=13.97 vs. limit=15.0 2024-09-25 00:49:34,599 INFO [train.py:1198] (2/4) Epoch 35, batch 1700, loss[loss=0.2181, ctc_loss=0.1445, cr_loss=0.3679, over 17074.00 frames. ], tot_loss[loss=0.1944, ctc_loss=0.1259, cr_loss=0.3428, over 3356135.69 frames. ], batch size: 46, lr: 3.40e-03, grad_scale: 32.0 2024-09-25 00:49:44,333 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=626103.3333333334, ans=0.125 2024-09-25 00:49:49,004 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=626150.0, ans=0.125 2024-09-25 00:49:52,464 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 00:49:55,407 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=626150.0, ans=0.1 2024-09-25 00:50:00,430 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 00:50:04,911 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=626196.6666666666, ans=0.025 2024-09-25 00:50:10,954 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.095e+02 1.273e+02 1.359e+02 1.469e+02 2.264e+02, threshold=2.717e+02, percent-clipped=0.0 2024-09-25 00:50:16,592 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=18.17 vs. limit=22.5 2024-09-25 00:50:27,107 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=626243.3333333334, ans=0.0 2024-09-25 00:50:32,989 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=5.19 vs. limit=15.0 2024-09-25 00:50:37,224 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=626290.0, ans=0.125 2024-09-25 00:50:51,910 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=4.78 vs. limit=12.0 2024-09-25 00:50:52,279 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=4.00 vs. limit=15.0 2024-09-25 00:50:54,471 INFO [train.py:1198] (2/4) Epoch 35, batch 1750, loss[loss=0.2148, ctc_loss=0.1436, cr_loss=0.356, over 16568.00 frames. ], tot_loss[loss=0.1951, ctc_loss=0.1264, cr_loss=0.3435, over 3365815.96 frames. ], batch size: 66, lr: 3.40e-03, grad_scale: 32.0 2024-09-25 00:51:15,542 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=626383.3333333334, ans=0.0 2024-09-25 00:51:56,936 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=18.24 vs. limit=22.5 2024-09-25 00:52:19,236 INFO [train.py:1198] (2/4) Epoch 35, batch 1800, loss[loss=0.1824, ctc_loss=0.12, cr_loss=0.312, over 17045.00 frames. ], tot_loss[loss=0.1954, ctc_loss=0.1266, cr_loss=0.344, over 3366673.53 frames. ], batch size: 52, lr: 3.40e-03, grad_scale: 32.0 2024-09-25 00:52:41,788 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=626616.6666666666, ans=0.0 2024-09-25 00:52:49,806 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=626663.3333333334, ans=0.0 2024-09-25 00:53:00,986 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.081e+02 1.248e+02 1.342e+02 1.447e+02 1.901e+02, threshold=2.684e+02, percent-clipped=0.0 2024-09-25 00:53:10,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=626710.0, ans=0.125 2024-09-25 00:53:28,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=626756.6666666666, ans=0.125 2024-09-25 00:53:44,284 INFO [train.py:1198] (2/4) Epoch 35, batch 1850, loss[loss=0.1596, ctc_loss=0.1022, cr_loss=0.2873, over 17024.00 frames. ], tot_loss[loss=0.1951, ctc_loss=0.1263, cr_loss=0.3439, over 3376915.45 frames. ], batch size: 44, lr: 3.40e-03, grad_scale: 32.0 2024-09-25 00:53:46,269 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=626803.3333333334, ans=0.125 2024-09-25 00:53:55,715 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=626803.3333333334, ans=0.2 2024-09-25 00:54:02,113 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=626850.0, ans=0.1 2024-09-25 00:54:05,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=626850.0, ans=0.125 2024-09-25 00:54:13,301 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=626850.0, ans=0.0 2024-09-25 00:54:30,874 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=626943.3333333334, ans=0.09899494936611666 2024-09-25 00:54:43,880 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 00:55:04,162 INFO [train.py:1198] (2/4) Epoch 35, batch 1900, loss[loss=0.2011, ctc_loss=0.1322, cr_loss=0.3444, over 17042.00 frames. ], tot_loss[loss=0.195, ctc_loss=0.1262, cr_loss=0.3441, over 3378337.75 frames. ], batch size: 56, lr: 3.40e-03, grad_scale: 32.0 2024-09-25 00:55:21,989 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=627083.3333333334, ans=0.1 2024-09-25 00:55:26,774 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=627083.3333333334, ans=0.025 2024-09-25 00:55:28,413 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=627083.3333333334, ans=0.125 2024-09-25 00:55:35,108 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 00:55:41,024 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.172e+02 1.333e+02 1.409e+02 1.508e+02 2.528e+02, threshold=2.819e+02, percent-clipped=0.0 2024-09-25 00:55:42,849 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=627130.0, ans=0.125 2024-09-25 00:56:19,993 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=627223.3333333334, ans=0.125 2024-09-25 00:56:24,333 INFO [train.py:1198] (2/4) Epoch 35, batch 1950, loss[loss=0.1753, ctc_loss=0.1119, cr_loss=0.3169, over 16340.00 frames. ], tot_loss[loss=0.1947, ctc_loss=0.126, cr_loss=0.3433, over 3378104.12 frames. ], batch size: 36, lr: 3.40e-03, grad_scale: 32.0 2024-09-25 00:56:32,536 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.54 vs. limit=10.0 2024-09-25 00:56:49,631 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=627316.6666666666, ans=0.125 2024-09-25 00:57:08,684 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=627363.3333333334, ans=0.2 2024-09-25 00:57:24,040 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=627410.0, ans=0.125 2024-09-25 00:57:30,594 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=627410.0, ans=0.0 2024-09-25 00:57:49,087 INFO [train.py:1198] (2/4) Epoch 35, batch 2000, loss[loss=0.231, ctc_loss=0.1519, cr_loss=0.3956, over 17033.00 frames. ], tot_loss[loss=0.1963, ctc_loss=0.1272, cr_loss=0.3456, over 3373971.93 frames. ], batch size: 52, lr: 3.40e-03, grad_scale: 32.0 2024-09-25 00:57:51,910 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=9.63 vs. limit=15.0 2024-09-25 00:58:30,772 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.112e+02 1.295e+02 1.390e+02 1.531e+02 2.683e+02, threshold=2.780e+02, percent-clipped=0.0 2024-09-25 00:58:53,733 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.17 vs. limit=15.0 2024-09-25 00:59:13,923 INFO [train.py:1198] (2/4) Epoch 35, batch 2050, loss[loss=0.1634, ctc_loss=0.1016, cr_loss=0.3093, over 17180.00 frames. ], tot_loss[loss=0.1958, ctc_loss=0.1269, cr_loss=0.3448, over 3377822.09 frames. ], batch size: 41, lr: 3.40e-03, grad_scale: 32.0 2024-09-25 01:00:10,493 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.12 vs. limit=6.0 2024-09-25 01:00:33,290 INFO [train.py:1198] (2/4) Epoch 35, batch 2100, loss[loss=0.1661, ctc_loss=0.1033, cr_loss=0.3141, over 17114.00 frames. ], tot_loss[loss=0.1953, ctc_loss=0.1265, cr_loss=0.3437, over 3373827.93 frames. ], batch size: 40, lr: 3.40e-03, grad_scale: 32.0 2024-09-25 01:01:10,508 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.271e+02 1.342e+02 1.455e+02 1.760e+02, threshold=2.684e+02, percent-clipped=0.0 2024-09-25 01:01:10,925 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=628063.3333333334, ans=0.1 2024-09-25 01:01:46,303 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.34 vs. limit=15.0 2024-09-25 01:01:51,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=628156.6666666666, ans=0.2 2024-09-25 01:01:56,292 INFO [train.py:1198] (2/4) Epoch 35, batch 2150, loss[loss=0.2121, ctc_loss=0.1383, cr_loss=0.369, over 16479.00 frames. ], tot_loss[loss=0.1951, ctc_loss=0.1263, cr_loss=0.3437, over 3367637.52 frames. ], batch size: 66, lr: 3.40e-03, grad_scale: 32.0 2024-09-25 01:02:06,247 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=628203.3333333334, ans=0.125 2024-09-25 01:02:07,789 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=628203.3333333334, ans=0.0 2024-09-25 01:02:18,375 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=628250.0, ans=0.0 2024-09-25 01:02:19,929 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=628250.0, ans=0.05 2024-09-25 01:02:55,464 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.86 vs. limit=6.0 2024-09-25 01:03:21,447 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=628390.0, ans=0.125 2024-09-25 01:03:24,302 INFO [train.py:1198] (2/4) Epoch 35, batch 2200, loss[loss=0.1957, ctc_loss=0.1282, cr_loss=0.3374, over 17020.00 frames. ], tot_loss[loss=0.1948, ctc_loss=0.1262, cr_loss=0.3429, over 3365524.14 frames. ], batch size: 51, lr: 3.40e-03, grad_scale: 32.0 2024-09-25 01:03:26,798 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.99 vs. limit=6.0 2024-09-25 01:04:00,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=628530.0, ans=0.125 2024-09-25 01:04:01,427 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.102e+02 1.291e+02 1.359e+02 1.454e+02 2.045e+02, threshold=2.717e+02, percent-clipped=0.0 2024-09-25 01:04:03,223 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=628530.0, ans=0.125 2024-09-25 01:04:04,112 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.46 vs. limit=15.0 2024-09-25 01:04:27,497 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=628623.3333333334, ans=0.125 2024-09-25 01:04:45,059 INFO [train.py:1198] (2/4) Epoch 35, batch 2250, loss[loss=0.2082, ctc_loss=0.1351, cr_loss=0.366, over 17059.00 frames. ], tot_loss[loss=0.1964, ctc_loss=0.1274, cr_loss=0.3453, over 3362931.36 frames. ], batch size: 56, lr: 3.40e-03, grad_scale: 32.0 2024-09-25 01:04:53,334 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=628670.0, ans=0.0 2024-09-25 01:05:30,260 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=628763.3333333334, ans=0.09899494936611666 2024-09-25 01:05:48,277 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=9.85 vs. limit=15.0 2024-09-25 01:06:05,296 INFO [train.py:1198] (2/4) Epoch 35, batch 2300, loss[loss=0.2083, ctc_loss=0.1343, cr_loss=0.3702, over 17224.00 frames. ], tot_loss[loss=0.1954, ctc_loss=0.1266, cr_loss=0.3441, over 3370448.11 frames. ], batch size: 50, lr: 3.39e-03, grad_scale: 32.0 2024-09-25 01:06:07,143 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=628903.3333333334, ans=0.125 2024-09-25 01:06:10,456 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=628903.3333333334, ans=0.07 2024-09-25 01:06:23,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=628950.0, ans=0.125 2024-09-25 01:06:38,370 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.min_positive, batch_count=628996.6666666666, ans=0.025 2024-09-25 01:06:39,979 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=628996.6666666666, ans=0.2 2024-09-25 01:06:44,417 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.311e+02 1.390e+02 1.542e+02 2.527e+02, threshold=2.780e+02, percent-clipped=0.0 2024-09-25 01:07:18,146 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=13.12 vs. limit=22.5 2024-09-25 01:07:30,082 INFO [train.py:1198] (2/4) Epoch 35, batch 2350, loss[loss=0.2007, ctc_loss=0.128, cr_loss=0.3636, over 16961.00 frames. ], tot_loss[loss=0.1942, ctc_loss=0.1258, cr_loss=0.3421, over 3375973.45 frames. ], batch size: 53, lr: 3.39e-03, grad_scale: 32.0 2024-09-25 01:07:43,139 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=629136.6666666666, ans=0.1 2024-09-25 01:08:26,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=629276.6666666666, ans=0.07 2024-09-25 01:08:30,007 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=629276.6666666666, ans=0.125 2024-09-25 01:08:44,484 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=629323.3333333334, ans=0.1 2024-09-25 01:08:50,932 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=629323.3333333334, ans=0.1 2024-09-25 01:08:55,602 INFO [train.py:1198] (2/4) Epoch 35, batch 2400, loss[loss=0.2162, ctc_loss=0.139, cr_loss=0.3863, over 17229.00 frames. ], tot_loss[loss=0.1942, ctc_loss=0.1258, cr_loss=0.3419, over 3379360.39 frames. ], batch size: 55, lr: 3.39e-03, grad_scale: 32.0 2024-09-25 01:09:01,285 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.79 vs. limit=15.0 2024-09-25 01:09:05,545 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=629370.0, ans=0.0 2024-09-25 01:09:17,278 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.45 vs. limit=10.0 2024-09-25 01:09:32,456 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.057e+02 1.256e+02 1.334e+02 1.415e+02 2.339e+02, threshold=2.669e+02, percent-clipped=0.0 2024-09-25 01:09:39,013 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=629463.3333333334, ans=0.2 2024-09-25 01:10:15,371 INFO [train.py:1198] (2/4) Epoch 35, batch 2450, loss[loss=0.1777, ctc_loss=0.112, cr_loss=0.3284, over 17101.00 frames. ], tot_loss[loss=0.1946, ctc_loss=0.1261, cr_loss=0.3427, over 3371625.93 frames. ], batch size: 43, lr: 3.39e-03, grad_scale: 32.0 2024-09-25 01:10:17,803 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=14.78 vs. limit=22.5 2024-09-25 01:10:18,794 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=629603.3333333334, ans=0.0 2024-09-25 01:10:50,963 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=629696.6666666666, ans=0.125 2024-09-25 01:10:57,829 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=2.98 vs. limit=6.0 2024-09-25 01:11:11,548 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=629743.3333333334, ans=0.0 2024-09-25 01:11:11,567 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=629743.3333333334, ans=0.1 2024-09-25 01:11:25,239 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=629790.0, ans=0.0 2024-09-25 01:11:37,756 INFO [train.py:1198] (2/4) Epoch 35, batch 2500, loss[loss=0.1828, ctc_loss=0.1154, cr_loss=0.3367, over 16964.00 frames. ], tot_loss[loss=0.1943, ctc_loss=0.1259, cr_loss=0.3419, over 3366433.34 frames. ], batch size: 42, lr: 3.39e-03, grad_scale: 32.0 2024-09-25 01:11:54,108 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 01:12:10,833 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=629930.0, ans=0.125 2024-09-25 01:12:16,801 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.245e+02 1.351e+02 1.450e+02 1.965e+02, threshold=2.702e+02, percent-clipped=0.0 2024-09-25 01:12:18,998 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.90 vs. limit=15.0 2024-09-25 01:12:28,492 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.min_positive, batch_count=629976.6666666666, ans=0.025 2024-09-25 01:13:05,480 INFO [train.py:1198] (2/4) Epoch 35, batch 2550, loss[loss=0.2143, ctc_loss=0.1355, cr_loss=0.3937, over 17172.00 frames. ], tot_loss[loss=0.1954, ctc_loss=0.1267, cr_loss=0.3435, over 3355845.59 frames. ], batch size: 45, lr: 3.39e-03, grad_scale: 32.0 2024-09-25 01:13:05,859 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=630070.0, ans=0.0 2024-09-25 01:13:10,275 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=630070.0, ans=0.035 2024-09-25 01:13:13,460 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=630070.0, ans=0.0 2024-09-25 01:13:21,694 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=630116.6666666666, ans=0.0 2024-09-25 01:13:39,126 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=630163.3333333334, ans=0.125 2024-09-25 01:13:40,552 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=630163.3333333334, ans=0.2 2024-09-25 01:14:02,924 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=630210.0, ans=0.1 2024-09-25 01:14:12,745 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=630256.6666666666, ans=0.125 2024-09-25 01:14:25,424 INFO [train.py:1198] (2/4) Epoch 35, batch 2600, loss[loss=0.1863, ctc_loss=0.1202, cr_loss=0.3307, over 17311.00 frames. ], tot_loss[loss=0.1962, ctc_loss=0.1273, cr_loss=0.3446, over 3358721.82 frames. ], batch size: 51, lr: 3.39e-03, grad_scale: 32.0 2024-09-25 01:14:39,114 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=9.01 vs. limit=15.0 2024-09-25 01:15:02,358 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.098e+02 1.261e+02 1.318e+02 1.420e+02 1.911e+02, threshold=2.636e+02, percent-clipped=0.0 2024-09-25 01:15:05,961 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=630396.6666666666, ans=0.125 2024-09-25 01:15:19,236 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=3.93 vs. limit=15.0 2024-09-25 01:15:22,110 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=13.24 vs. limit=15.0 2024-09-25 01:15:45,580 INFO [train.py:1198] (2/4) Epoch 35, batch 2650, loss[loss=0.2215, ctc_loss=0.1457, cr_loss=0.3788, over 16942.00 frames. ], tot_loss[loss=0.1962, ctc_loss=0.1273, cr_loss=0.3447, over 3369281.18 frames. ], batch size: 58, lr: 3.39e-03, grad_scale: 32.0 2024-09-25 01:15:50,682 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=630536.6666666666, ans=0.125 2024-09-25 01:16:00,148 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=630583.3333333334, ans=0.05 2024-09-25 01:16:05,507 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=16.39 vs. limit=22.5 2024-09-25 01:17:03,783 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=630723.3333333334, ans=0.125 2024-09-25 01:17:05,346 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=630723.3333333334, ans=0.125 2024-09-25 01:17:08,143 INFO [train.py:1198] (2/4) Epoch 35, batch 2700, loss[loss=0.2327, ctc_loss=0.1528, cr_loss=0.3993, over 16982.00 frames. ], tot_loss[loss=0.1954, ctc_loss=0.1265, cr_loss=0.3441, over 3374499.61 frames. ], batch size: 53, lr: 3.39e-03, grad_scale: 32.0 2024-09-25 01:17:52,694 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.085e+02 1.276e+02 1.350e+02 1.459e+02 1.855e+02, threshold=2.700e+02, percent-clipped=0.0 2024-09-25 01:17:57,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=630863.3333333334, ans=0.1 2024-09-25 01:18:04,481 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=10.36 vs. limit=22.5 2024-09-25 01:18:20,217 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=630956.6666666666, ans=0.125 2024-09-25 01:18:20,386 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.min_abs, batch_count=630956.6666666666, ans=0.5 2024-09-25 01:18:23,875 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=12.94 vs. limit=22.5 2024-09-25 01:18:34,743 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=631003.3333333334, ans=0.125 2024-09-25 01:18:36,127 INFO [train.py:1198] (2/4) Epoch 35, batch 2750, loss[loss=0.1996, ctc_loss=0.1285, cr_loss=0.3553, over 17364.00 frames. ], tot_loss[loss=0.1946, ctc_loss=0.1259, cr_loss=0.3434, over 3381250.38 frames. ], batch size: 48, lr: 3.39e-03, grad_scale: 32.0 2024-09-25 01:18:39,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=631003.3333333334, ans=0.125 2024-09-25 01:18:46,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=631003.3333333334, ans=0.125 2024-09-25 01:19:14,454 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=12.33 vs. limit=22.5 2024-09-25 01:19:36,968 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module1.whiten, num_groups=1, num_channels=192, metric=7.67 vs. limit=15.0 2024-09-25 01:19:39,283 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=631190.0, ans=0.0 2024-09-25 01:19:50,870 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=9.51 vs. limit=15.0 2024-09-25 01:19:53,552 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=631190.0, ans=0.125 2024-09-25 01:19:56,334 INFO [train.py:1198] (2/4) Epoch 35, batch 2800, loss[loss=0.1709, ctc_loss=0.1091, cr_loss=0.3091, over 17168.00 frames. ], tot_loss[loss=0.1943, ctc_loss=0.1257, cr_loss=0.343, over 3381407.79 frames. ], batch size: 45, lr: 3.39e-03, grad_scale: 32.0 2024-09-25 01:20:23,917 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=631283.3333333334, ans=0.125 2024-09-25 01:20:31,873 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=631330.0, ans=0.125 2024-09-25 01:20:33,122 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.248e+02 1.324e+02 1.411e+02 2.015e+02, threshold=2.647e+02, percent-clipped=0.0 2024-09-25 01:20:47,677 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=631376.6666666666, ans=0.125 2024-09-25 01:20:57,351 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=631376.6666666666, ans=0.125 2024-09-25 01:21:06,971 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=631423.3333333334, ans=0.2 2024-09-25 01:21:16,333 INFO [train.py:1198] (2/4) Epoch 35, batch 2850, loss[loss=0.1664, ctc_loss=0.1075, cr_loss=0.2943, over 17156.00 frames. ], tot_loss[loss=0.195, ctc_loss=0.1263, cr_loss=0.3437, over 3372852.62 frames. ], batch size: 45, lr: 3.39e-03, grad_scale: 64.0 2024-09-25 01:21:27,131 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=631470.0, ans=0.0 2024-09-25 01:21:35,355 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.88 vs. limit=15.0 2024-09-25 01:21:54,138 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=631563.3333333334, ans=0.125 2024-09-25 01:22:17,664 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=631610.0, ans=0.0 2024-09-25 01:22:18,171 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.35 vs. limit=10.0 2024-09-25 01:22:26,308 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.81 vs. limit=12.0 2024-09-25 01:22:46,461 INFO [train.py:1198] (2/4) Epoch 35, batch 2900, loss[loss=0.1735, ctc_loss=0.1095, cr_loss=0.32, over 17056.00 frames. ], tot_loss[loss=0.196, ctc_loss=0.1271, cr_loss=0.3447, over 3359391.35 frames. ], batch size: 39, lr: 3.39e-03, grad_scale: 32.0 2024-09-25 01:23:07,937 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.29 vs. limit=6.0 2024-09-25 01:23:24,647 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.138e+02 1.283e+02 1.348e+02 1.424e+02 1.926e+02, threshold=2.696e+02, percent-clipped=0.0 2024-09-25 01:24:04,572 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.47 vs. limit=15.0 2024-09-25 01:24:06,673 INFO [train.py:1198] (2/4) Epoch 35, batch 2950, loss[loss=0.2191, ctc_loss=0.1488, cr_loss=0.3515, over 11799.00 frames. ], tot_loss[loss=0.1952, ctc_loss=0.1267, cr_loss=0.3428, over 3351109.10 frames. ], batch size: 123, lr: 3.39e-03, grad_scale: 32.0 2024-09-25 01:24:22,063 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=10.63 vs. limit=15.0 2024-09-25 01:24:40,827 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=632030.0, ans=0.2 2024-09-25 01:24:42,261 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=632030.0, ans=0.0 2024-09-25 01:25:09,789 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=632123.3333333334, ans=0.125 2024-09-25 01:25:26,966 INFO [train.py:1198] (2/4) Epoch 35, batch 3000, loss[loss=0.1627, ctc_loss=0.1017, cr_loss=0.305, over 17276.00 frames. ], tot_loss[loss=0.1944, ctc_loss=0.1261, cr_loss=0.3419, over 3354305.53 frames. ], batch size: 42, lr: 3.39e-03, grad_scale: 32.0 2024-09-25 01:25:26,967 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 01:25:42,182 INFO [train.py:1230] (2/4) Epoch 35, validation: loss=0.03538, ctc_loss=0.03538, cr_loss=9.094e-15, over 944034.00 frames. 2024-09-25 01:25:42,183 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 01:26:19,534 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.295e+02 1.357e+02 1.448e+02 2.181e+02, threshold=2.715e+02, percent-clipped=0.0 2024-09-25 01:26:43,217 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.min_positive, batch_count=632356.6666666666, ans=0.05 2024-09-25 01:26:49,178 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=632356.6666666666, ans=0.125 2024-09-25 01:27:02,789 INFO [train.py:1198] (2/4) Epoch 35, batch 3050, loss[loss=0.1615, ctc_loss=0.105, cr_loss=0.2826, over 16722.00 frames. ], tot_loss[loss=0.1955, ctc_loss=0.1268, cr_loss=0.3437, over 3342364.98 frames. ], batch size: 37, lr: 3.39e-03, grad_scale: 32.0 2024-09-25 01:27:04,029 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=7.05 vs. limit=8.0 2024-09-25 01:27:09,946 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=20.15 vs. limit=22.5 2024-09-25 01:27:10,942 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=632403.3333333334, ans=0.125 2024-09-25 01:27:43,536 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=632496.6666666666, ans=0.125 2024-09-25 01:27:47,635 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.59 vs. limit=6.0 2024-09-25 01:28:05,976 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.16 vs. limit=22.5 2024-09-25 01:28:13,919 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.23 vs. limit=15.0 2024-09-25 01:28:23,385 INFO [train.py:1198] (2/4) Epoch 35, batch 3100, loss[loss=0.2175, ctc_loss=0.1426, cr_loss=0.3745, over 15971.00 frames. ], tot_loss[loss=0.1957, ctc_loss=0.1269, cr_loss=0.3442, over 3340024.18 frames. ], batch size: 74, lr: 3.38e-03, grad_scale: 32.0 2024-09-25 01:28:48,996 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.27 vs. limit=15.0 2024-09-25 01:28:50,222 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=632683.3333333334, ans=0.07 2024-09-25 01:29:01,045 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.135e+02 1.287e+02 1.342e+02 1.471e+02 1.803e+02, threshold=2.684e+02, percent-clipped=0.0 2024-09-25 01:29:04,417 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=632730.0, ans=0.2 2024-09-25 01:29:04,567 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=632730.0, ans=0.0 2024-09-25 01:29:04,568 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=632730.0, ans=0.125 2024-09-25 01:29:14,344 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.06 vs. limit=15.0 2024-09-25 01:29:25,685 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=632776.6666666666, ans=0.025 2024-09-25 01:29:30,361 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=632823.3333333334, ans=0.125 2024-09-25 01:29:33,531 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=632823.3333333334, ans=0.1 2024-09-25 01:29:46,583 INFO [train.py:1198] (2/4) Epoch 35, batch 3150, loss[loss=0.2074, ctc_loss=0.1324, cr_loss=0.3749, over 17008.00 frames. ], tot_loss[loss=0.1963, ctc_loss=0.1272, cr_loss=0.3455, over 3342474.31 frames. ], batch size: 51, lr: 3.38e-03, grad_scale: 32.0 2024-09-25 01:30:10,351 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=632916.6666666666, ans=0.125 2024-09-25 01:31:04,920 INFO [train.py:1198] (2/4) Epoch 35, batch 3200, loss[loss=0.205, ctc_loss=0.1343, cr_loss=0.3538, over 16988.00 frames. ], tot_loss[loss=0.1956, ctc_loss=0.1267, cr_loss=0.3444, over 3343339.24 frames. ], batch size: 53, lr: 3.38e-03, grad_scale: 32.0 2024-09-25 01:31:20,750 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=633150.0, ans=0.1 2024-09-25 01:31:31,833 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=633150.0, ans=0.125 2024-09-25 01:31:41,217 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=633196.6666666666, ans=0.125 2024-09-25 01:31:42,383 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.282e+02 1.354e+02 1.513e+02 2.603e+02, threshold=2.707e+02, percent-clipped=0.0 2024-09-25 01:31:44,864 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.32 vs. limit=12.0 2024-09-25 01:31:58,414 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=633243.3333333334, ans=0.0 2024-09-25 01:32:03,046 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=633243.3333333334, ans=0.125 2024-09-25 01:32:23,311 INFO [train.py:1198] (2/4) Epoch 35, batch 3250, loss[loss=0.1892, ctc_loss=0.1207, cr_loss=0.343, over 15860.00 frames. ], tot_loss[loss=0.1947, ctc_loss=0.1262, cr_loss=0.3426, over 3330007.32 frames. ], batch size: 35, lr: 3.38e-03, grad_scale: 32.0 2024-09-25 01:32:44,859 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=633383.3333333334, ans=0.125 2024-09-25 01:32:44,860 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=633383.3333333334, ans=0.015 2024-09-25 01:32:46,886 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=9.08 vs. limit=15.0 2024-09-25 01:32:49,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=633383.3333333334, ans=0.2 2024-09-25 01:32:51,289 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=633383.3333333334, ans=0.125 2024-09-25 01:32:55,999 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=633430.0, ans=10.0 2024-09-25 01:32:59,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=633430.0, ans=0.125 2024-09-25 01:33:03,771 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=633430.0, ans=0.0 2024-09-25 01:33:23,113 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=9.23 vs. limit=15.0 2024-09-25 01:33:28,978 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=633523.3333333334, ans=0.125 2024-09-25 01:33:30,610 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.whiten.whitening_limit, batch_count=633523.3333333334, ans=12.0 2024-09-25 01:33:40,960 INFO [train.py:1198] (2/4) Epoch 35, batch 3300, loss[loss=0.1965, ctc_loss=0.1268, cr_loss=0.3482, over 17315.00 frames. ], tot_loss[loss=0.1948, ctc_loss=0.1262, cr_loss=0.3427, over 3340954.93 frames. ], batch size: 46, lr: 3.38e-03, grad_scale: 32.0 2024-09-25 01:33:57,191 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=633616.6666666666, ans=0.125 2024-09-25 01:33:57,329 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=633616.6666666666, ans=0.1 2024-09-25 01:34:00,401 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=633616.6666666666, ans=0.2 2024-09-25 01:34:17,831 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=633663.3333333334, ans=0.125 2024-09-25 01:34:17,880 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=633663.3333333334, ans=0.025 2024-09-25 01:34:19,045 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.144e+02 1.276e+02 1.335e+02 1.450e+02 3.347e+02, threshold=2.669e+02, percent-clipped=1.0 2024-09-25 01:34:31,704 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.max_positive, batch_count=633710.0, ans=0.95 2024-09-25 01:34:36,739 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=633710.0, ans=0.0 2024-09-25 01:34:40,361 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.43 vs. limit=15.0 2024-09-25 01:34:43,184 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=633756.6666666666, ans=0.0 2024-09-25 01:34:52,531 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=633756.6666666666, ans=0.125 2024-09-25 01:35:00,133 INFO [train.py:1198] (2/4) Epoch 35, batch 3350, loss[loss=0.2106, ctc_loss=0.1376, cr_loss=0.3647, over 16980.00 frames. ], tot_loss[loss=0.1952, ctc_loss=0.1265, cr_loss=0.3437, over 3349360.33 frames. ], batch size: 56, lr: 3.38e-03, grad_scale: 32.0 2024-09-25 01:35:42,861 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=633896.6666666666, ans=0.0 2024-09-25 01:36:11,374 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=9.07 vs. limit=15.0 2024-09-25 01:36:18,434 INFO [train.py:1198] (2/4) Epoch 35, batch 3400, loss[loss=0.2007, ctc_loss=0.1293, cr_loss=0.3571, over 17214.00 frames. ], tot_loss[loss=0.1946, ctc_loss=0.126, cr_loss=0.3431, over 3357660.93 frames. ], batch size: 47, lr: 3.38e-03, grad_scale: 16.0 2024-09-25 01:36:18,678 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=634036.6666666666, ans=0.125 2024-09-25 01:36:31,242 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=634036.6666666666, ans=0.1 2024-09-25 01:36:31,260 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=634036.6666666666, ans=0.1 2024-09-25 01:36:57,298 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.280e+02 1.352e+02 1.458e+02 2.510e+02, threshold=2.703e+02, percent-clipped=0.0 2024-09-25 01:37:02,450 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=11.61 vs. limit=15.0 2024-09-25 01:37:32,450 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=634223.3333333334, ans=0.2 2024-09-25 01:37:38,588 INFO [train.py:1198] (2/4) Epoch 35, batch 3450, loss[loss=0.1603, ctc_loss=0.1014, cr_loss=0.2945, over 16287.00 frames. ], tot_loss[loss=0.1943, ctc_loss=0.1257, cr_loss=0.3428, over 3357985.96 frames. ], batch size: 36, lr: 3.38e-03, grad_scale: 16.0 2024-09-25 01:37:40,450 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=634270.0, ans=0.0 2024-09-25 01:37:55,765 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=634316.6666666666, ans=0.125 2024-09-25 01:38:21,342 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=634363.3333333334, ans=0.125 2024-09-25 01:38:38,597 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=634410.0, ans=0.2 2024-09-25 01:38:40,824 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.85 vs. limit=15.0 2024-09-25 01:38:41,814 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=634456.6666666666, ans=0.025 2024-09-25 01:38:54,276 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=634456.6666666666, ans=0.0 2024-09-25 01:38:58,784 INFO [train.py:1198] (2/4) Epoch 35, batch 3500, loss[loss=0.168, ctc_loss=0.1064, cr_loss=0.3082, over 17034.00 frames. ], tot_loss[loss=0.1953, ctc_loss=0.1265, cr_loss=0.344, over 3349184.56 frames. ], batch size: 39, lr: 3.38e-03, grad_scale: 16.0 2024-09-25 01:39:40,109 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.089e+02 1.254e+02 1.331e+02 1.459e+02 2.382e+02, threshold=2.663e+02, percent-clipped=0.0 2024-09-25 01:39:56,287 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 01:40:04,985 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.81 vs. limit=15.0 2024-09-25 01:40:16,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=634690.0, ans=0.1 2024-09-25 01:40:22,748 INFO [train.py:1198] (2/4) Epoch 35, batch 3550, loss[loss=0.1487, ctc_loss=0.09234, cr_loss=0.282, over 17091.00 frames. ], tot_loss[loss=0.1948, ctc_loss=0.1261, cr_loss=0.3435, over 3360222.33 frames. ], batch size: 43, lr: 3.38e-03, grad_scale: 16.0 2024-09-25 01:40:43,284 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=634783.3333333334, ans=0.125 2024-09-25 01:40:50,040 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=5.61 vs. limit=12.0 2024-09-25 01:40:53,937 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=634830.0, ans=0.025 2024-09-25 01:41:33,299 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 01:41:40,836 INFO [train.py:1198] (2/4) Epoch 35, batch 3600, loss[loss=0.2397, ctc_loss=0.1669, cr_loss=0.3642, over 12134.00 frames. ], tot_loss[loss=0.1953, ctc_loss=0.1265, cr_loss=0.344, over 3355236.10 frames. ], batch size: 124, lr: 3.38e-03, grad_scale: 32.0 2024-09-25 01:41:44,124 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=634970.0, ans=0.1 2024-09-25 01:42:01,426 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=635016.6666666666, ans=0.0 2024-09-25 01:42:19,816 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.143e+02 1.292e+02 1.369e+02 1.496e+02 2.107e+02, threshold=2.739e+02, percent-clipped=0.0 2024-09-25 01:42:45,083 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=635156.6666666666, ans=0.125 2024-09-25 01:42:51,166 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=635156.6666666666, ans=0.0 2024-09-25 01:42:58,846 INFO [train.py:1198] (2/4) Epoch 35, batch 3650, loss[loss=0.1752, ctc_loss=0.1123, cr_loss=0.3145, over 17068.00 frames. ], tot_loss[loss=0.195, ctc_loss=0.1263, cr_loss=0.3435, over 3366302.15 frames. ], batch size: 46, lr: 3.38e-03, grad_scale: 32.0 2024-09-25 01:43:02,226 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=635203.3333333334, ans=0.0 2024-09-25 01:43:27,075 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=635250.0, ans=0.125 2024-09-25 01:43:48,010 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=6.29 vs. limit=15.0 2024-09-25 01:43:53,548 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=635343.3333333334, ans=0.125 2024-09-25 01:44:14,815 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=635390.0, ans=0.125 2024-09-25 01:44:17,733 INFO [train.py:1198] (2/4) Epoch 35, batch 3700, loss[loss=0.1639, ctc_loss=0.1029, cr_loss=0.3052, over 16742.00 frames. ], tot_loss[loss=0.1954, ctc_loss=0.1266, cr_loss=0.3439, over 3365691.46 frames. ], batch size: 37, lr: 3.38e-03, grad_scale: 32.0 2024-09-25 01:44:55,553 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=635530.0, ans=0.0 2024-09-25 01:44:56,873 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.290e+02 1.367e+02 1.473e+02 2.318e+02, threshold=2.733e+02, percent-clipped=0.0 2024-09-25 01:45:18,066 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=635576.6666666666, ans=0.125 2024-09-25 01:45:31,244 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=5.85 vs. limit=12.0 2024-09-25 01:45:35,289 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=635670.0, ans=0.125 2024-09-25 01:45:36,670 INFO [train.py:1198] (2/4) Epoch 35, batch 3750, loss[loss=0.2344, ctc_loss=0.1583, cr_loss=0.3806, over 15131.00 frames. ], tot_loss[loss=0.1965, ctc_loss=0.1275, cr_loss=0.3453, over 3353604.35 frames. ], batch size: 89, lr: 3.38e-03, grad_scale: 32.0 2024-09-25 01:45:40,147 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.min_positive, batch_count=635670.0, ans=0.05 2024-09-25 01:46:55,868 INFO [train.py:1198] (2/4) Epoch 35, batch 3800, loss[loss=0.1488, ctc_loss=0.09525, cr_loss=0.268, over 16281.00 frames. ], tot_loss[loss=0.1963, ctc_loss=0.1274, cr_loss=0.3446, over 3340025.98 frames. ], batch size: 36, lr: 3.38e-03, grad_scale: 32.0 2024-09-25 01:47:17,201 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=16.05 vs. limit=15.0 2024-09-25 01:47:34,349 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer_na.min_abs, batch_count=635996.6666666666, ans=0.02 2024-09-25 01:47:35,577 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.304e+02 1.388e+02 1.504e+02 2.196e+02, threshold=2.776e+02, percent-clipped=0.0 2024-09-25 01:47:40,769 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=635996.6666666666, ans=0.125 2024-09-25 01:47:42,419 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=636043.3333333334, ans=0.0 2024-09-25 01:47:47,069 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=636043.3333333334, ans=0.0 2024-09-25 01:48:01,964 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=636090.0, ans=0.07 2024-09-25 01:48:03,457 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=636090.0, ans=0.125 2024-09-25 01:48:14,974 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=636136.6666666666, ans=0.0 2024-09-25 01:48:16,223 INFO [train.py:1198] (2/4) Epoch 35, batch 3850, loss[loss=0.2257, ctc_loss=0.1499, cr_loss=0.3792, over 11976.00 frames. ], tot_loss[loss=0.1955, ctc_loss=0.1268, cr_loss=0.3431, over 3305853.44 frames. ], batch size: 126, lr: 3.38e-03, grad_scale: 32.0 2024-09-25 01:48:23,711 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=18.99 vs. limit=22.5 2024-09-25 01:48:36,789 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=636183.3333333334, ans=0.1 2024-09-25 01:48:41,373 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=636183.3333333334, ans=0.125 2024-09-25 01:48:57,646 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=8.42 vs. limit=12.0 2024-09-25 01:49:04,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=636276.6666666666, ans=0.125 2024-09-25 01:49:07,532 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=636276.6666666666, ans=0.125 2024-09-25 01:49:15,162 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=636276.6666666666, ans=0.0 2024-09-25 01:50:19,160 INFO [train.py:1198] (2/4) Epoch 36, batch 0, loss[loss=0.1741, ctc_loss=0.1105, cr_loss=0.3182, over 16946.00 frames. ], tot_loss[loss=0.1741, ctc_loss=0.1105, cr_loss=0.3182, over 16946.00 frames. ], batch size: 42, lr: 3.33e-03, grad_scale: 32.0 2024-09-25 01:50:19,161 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 01:50:34,827 INFO [train.py:1230] (2/4) Epoch 36, validation: loss=0.0356, ctc_loss=0.0356, cr_loss=9.615e-15, over 944034.00 frames. 2024-09-25 01:50:34,828 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 01:50:35,090 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=636351.3333333334, ans=0.125 2024-09-25 01:51:00,931 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=636398.0, ans=0.0 2024-09-25 01:51:21,248 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.070e+02 1.459e+02 1.589e+02 1.794e+02 2.988e+02, threshold=3.179e+02, percent-clipped=1.0 2024-09-25 01:51:50,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=636538.0, ans=0.025 2024-09-25 01:51:55,583 INFO [train.py:1198] (2/4) Epoch 36, batch 50, loss[loss=0.1958, ctc_loss=0.1289, cr_loss=0.3348, over 17311.00 frames. ], tot_loss[loss=0.1933, ctc_loss=0.1253, cr_loss=0.3402, over 751966.99 frames. ], batch size: 49, lr: 3.33e-03, grad_scale: 32.0 2024-09-25 01:52:53,884 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=5.07 vs. limit=15.0 2024-09-25 01:53:07,704 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=636771.3333333334, ans=0.125 2024-09-25 01:53:21,605 INFO [train.py:1198] (2/4) Epoch 36, batch 100, loss[loss=0.1943, ctc_loss=0.125, cr_loss=0.3465, over 16861.00 frames. ], tot_loss[loss=0.194, ctc_loss=0.1256, cr_loss=0.3418, over 1332322.51 frames. ], batch size: 58, lr: 3.33e-03, grad_scale: 32.0 2024-09-25 01:53:26,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=636818.0, ans=0.125 2024-09-25 01:53:29,867 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=636818.0, ans=0.2 2024-09-25 01:53:29,957 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.max_positive, batch_count=636818.0, ans=0.95 2024-09-25 01:53:41,384 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=636864.6666666666, ans=0.2 2024-09-25 01:53:42,867 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=636864.6666666666, ans=0.0 2024-09-25 01:53:52,640 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=636911.3333333334, ans=0.125 2024-09-25 01:54:05,417 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=636911.3333333334, ans=0.125 2024-09-25 01:54:09,699 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=636911.3333333334, ans=0.04949747468305833 2024-09-25 01:54:11,062 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.080e+02 1.276e+02 1.354e+02 1.441e+02 1.843e+02, threshold=2.708e+02, percent-clipped=0.0 2024-09-25 01:54:14,580 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=636958.0, ans=0.1 2024-09-25 01:54:30,803 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=637004.6666666666, ans=0.2 2024-09-25 01:54:44,745 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=637004.6666666666, ans=0.1 2024-09-25 01:54:47,637 INFO [train.py:1198] (2/4) Epoch 36, batch 150, loss[loss=0.1747, ctc_loss=0.1127, cr_loss=0.31, over 17019.00 frames. ], tot_loss[loss=0.196, ctc_loss=0.1269, cr_loss=0.3453, over 1778028.79 frames. ], batch size: 39, lr: 3.32e-03, grad_scale: 16.0 2024-09-25 01:55:16,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=637098.0, ans=0.1 2024-09-25 01:55:18,138 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=637144.6666666666, ans=0.125 2024-09-25 01:56:07,384 INFO [train.py:1198] (2/4) Epoch 36, batch 200, loss[loss=0.2295, ctc_loss=0.1517, cr_loss=0.3891, over 16951.00 frames. ], tot_loss[loss=0.1955, ctc_loss=0.1267, cr_loss=0.3437, over 2118550.17 frames. ], batch size: 58, lr: 3.32e-03, grad_scale: 16.0 2024-09-25 01:56:15,600 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 01:56:36,421 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=637331.3333333334, ans=0.125 2024-09-25 01:56:55,269 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.143e+02 1.279e+02 1.373e+02 1.478e+02 2.081e+02, threshold=2.747e+02, percent-clipped=0.0 2024-09-25 01:57:26,986 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.max_positive, batch_count=637471.3333333334, ans=0.95 2024-09-25 01:57:27,136 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=637471.3333333334, ans=0.2 2024-09-25 01:57:29,959 INFO [train.py:1198] (2/4) Epoch 36, batch 250, loss[loss=0.1851, ctc_loss=0.1173, cr_loss=0.3387, over 17295.00 frames. ], tot_loss[loss=0.1961, ctc_loss=0.1272, cr_loss=0.3443, over 2386419.58 frames. ], batch size: 46, lr: 3.32e-03, grad_scale: 16.0 2024-09-25 01:57:58,190 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=3.97 vs. limit=15.0 2024-09-25 01:58:13,801 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=14.53 vs. limit=22.5 2024-09-25 01:58:43,932 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=6.93 vs. limit=15.0 2024-09-25 01:58:52,444 INFO [train.py:1198] (2/4) Epoch 36, batch 300, loss[loss=0.1568, ctc_loss=0.1005, cr_loss=0.2812, over 17039.00 frames. ], tot_loss[loss=0.1961, ctc_loss=0.1272, cr_loss=0.3445, over 2598748.70 frames. ], batch size: 39, lr: 3.32e-03, grad_scale: 16.0 2024-09-25 01:59:46,603 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.037e+02 1.272e+02 1.363e+02 1.432e+02 1.912e+02, threshold=2.726e+02, percent-clipped=0.0 2024-09-25 01:59:46,979 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=637891.3333333334, ans=0.2 2024-09-25 02:00:06,108 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=637938.0, ans=0.125 2024-09-25 02:00:11,409 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=16.64 vs. limit=22.5 2024-09-25 02:00:18,409 INFO [train.py:1198] (2/4) Epoch 36, batch 350, loss[loss=0.2179, ctc_loss=0.1399, cr_loss=0.39, over 16499.00 frames. ], tot_loss[loss=0.1953, ctc_loss=0.1265, cr_loss=0.3441, over 2768120.86 frames. ], batch size: 66, lr: 3.32e-03, grad_scale: 16.0 2024-09-25 02:00:31,695 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=637984.6666666666, ans=0.2 2024-09-25 02:00:49,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=638078.0, ans=0.1 2024-09-25 02:01:32,772 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=638171.3333333334, ans=0.125 2024-09-25 02:01:36,641 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten.whitening_limit, batch_count=638171.3333333334, ans=15.0 2024-09-25 02:01:38,834 INFO [train.py:1198] (2/4) Epoch 36, batch 400, loss[loss=0.2211, ctc_loss=0.1503, cr_loss=0.3542, over 11979.00 frames. ], tot_loss[loss=0.1951, ctc_loss=0.1264, cr_loss=0.3437, over 2893854.90 frames. ], batch size: 123, lr: 3.32e-03, grad_scale: 32.0 2024-09-25 02:01:51,851 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=638218.0, ans=0.1 2024-09-25 02:02:03,057 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=638264.6666666666, ans=0.125 2024-09-25 02:02:09,267 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=638311.3333333334, ans=0.125 2024-09-25 02:02:18,262 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=638311.3333333334, ans=0.1 2024-09-25 02:02:29,164 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.130e+02 1.293e+02 1.367e+02 1.475e+02 2.656e+02, threshold=2.733e+02, percent-clipped=0.0 2024-09-25 02:02:42,215 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=638358.0, ans=0.1 2024-09-25 02:03:01,276 INFO [train.py:1198] (2/4) Epoch 36, batch 450, loss[loss=0.217, ctc_loss=0.1411, cr_loss=0.3794, over 16576.00 frames. ], tot_loss[loss=0.1959, ctc_loss=0.1269, cr_loss=0.3446, over 2987622.80 frames. ], batch size: 66, lr: 3.32e-03, grad_scale: 32.0 2024-09-25 02:03:14,762 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=3.58 vs. limit=6.0 2024-09-25 02:03:22,253 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=638498.0, ans=0.125 2024-09-25 02:03:30,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=638498.0, ans=0.125 2024-09-25 02:03:31,338 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module1.whiten, num_groups=1, num_channels=192, metric=9.86 vs. limit=15.0 2024-09-25 02:03:32,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=638498.0, ans=0.125 2024-09-25 02:04:03,714 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 02:04:14,932 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=638638.0, ans=0.04949747468305833 2024-09-25 02:04:27,072 INFO [train.py:1198] (2/4) Epoch 36, batch 500, loss[loss=0.2111, ctc_loss=0.1396, cr_loss=0.3574, over 17035.00 frames. ], tot_loss[loss=0.1961, ctc_loss=0.1271, cr_loss=0.3452, over 3073990.20 frames. ], batch size: 51, lr: 3.32e-03, grad_scale: 32.0 2024-09-25 02:04:38,322 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.16 vs. limit=15.0 2024-09-25 02:04:41,319 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=638684.6666666666, ans=0.125 2024-09-25 02:05:08,487 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=638778.0, ans=0.1 2024-09-25 02:05:11,728 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=638778.0, ans=0.0 2024-09-25 02:05:16,749 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.96 vs. limit=15.0 2024-09-25 02:05:19,275 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.049e+02 1.244e+02 1.315e+02 1.459e+02 2.015e+02, threshold=2.630e+02, percent-clipped=0.0 2024-09-25 02:05:49,776 INFO [train.py:1198] (2/4) Epoch 36, batch 550, loss[loss=0.1964, ctc_loss=0.1253, cr_loss=0.3555, over 17151.00 frames. ], tot_loss[loss=0.1965, ctc_loss=0.1272, cr_loss=0.3464, over 3144534.68 frames. ], batch size: 45, lr: 3.32e-03, grad_scale: 16.0 2024-09-25 02:06:02,769 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=638918.0, ans=0.1 2024-09-25 02:06:10,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.min_positive, batch_count=638964.6666666666, ans=0.05 2024-09-25 02:06:15,405 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=638964.6666666666, ans=0.1 2024-09-25 02:06:33,283 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.min_abs, batch_count=639011.3333333334, ans=0.5 2024-09-25 02:06:39,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=639058.0, ans=0.04949747468305833 2024-09-25 02:07:09,441 INFO [train.py:1198] (2/4) Epoch 36, batch 600, loss[loss=0.1623, ctc_loss=0.1034, cr_loss=0.2943, over 16959.00 frames. ], tot_loss[loss=0.1958, ctc_loss=0.1267, cr_loss=0.3454, over 3184534.35 frames. ], batch size: 42, lr: 3.32e-03, grad_scale: 16.0 2024-09-25 02:07:46,104 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=639244.6666666666, ans=0.0 2024-09-25 02:08:01,748 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.146e+02 1.285e+02 1.361e+02 1.464e+02 2.442e+02, threshold=2.722e+02, percent-clipped=0.0 2024-09-25 02:08:04,974 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=639291.3333333334, ans=0.0 2024-09-25 02:08:07,894 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=639291.3333333334, ans=0.125 2024-09-25 02:08:07,913 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=639291.3333333334, ans=0.025 2024-09-25 02:08:34,769 INFO [train.py:1198] (2/4) Epoch 36, batch 650, loss[loss=0.1567, ctc_loss=0.09785, cr_loss=0.2943, over 16322.00 frames. ], tot_loss[loss=0.1952, ctc_loss=0.1264, cr_loss=0.344, over 3222148.43 frames. ], batch size: 36, lr: 3.32e-03, grad_scale: 16.0 2024-09-25 02:08:44,544 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=639384.6666666666, ans=0.125 2024-09-25 02:08:47,892 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=639384.6666666666, ans=0.0 2024-09-25 02:09:11,312 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=639478.0, ans=0.04949747468305833 2024-09-25 02:09:50,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=639571.3333333334, ans=0.125 2024-09-25 02:09:52,108 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=639571.3333333334, ans=0.125 2024-09-25 02:09:59,836 INFO [train.py:1198] (2/4) Epoch 36, batch 700, loss[loss=0.2098, ctc_loss=0.1378, cr_loss=0.3601, over 16999.00 frames. ], tot_loss[loss=0.1943, ctc_loss=0.1258, cr_loss=0.3428, over 3259802.04 frames. ], batch size: 53, lr: 3.32e-03, grad_scale: 16.0 2024-09-25 02:10:01,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=639618.0, ans=0.0 2024-09-25 02:10:29,230 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=639664.6666666666, ans=0.125 2024-09-25 02:10:37,359 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.16 vs. limit=15.0 2024-09-25 02:10:49,615 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.257e+02 1.341e+02 1.454e+02 1.758e+02, threshold=2.682e+02, percent-clipped=0.0 2024-09-25 02:11:19,964 INFO [train.py:1198] (2/4) Epoch 36, batch 750, loss[loss=0.212, ctc_loss=0.1372, cr_loss=0.3739, over 17072.00 frames. ], tot_loss[loss=0.1938, ctc_loss=0.1254, cr_loss=0.3418, over 3282449.89 frames. ], batch size: 46, lr: 3.32e-03, grad_scale: 16.0 2024-09-25 02:11:33,315 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=639851.3333333334, ans=0.0 2024-09-25 02:11:34,852 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=639898.0, ans=0.1 2024-09-25 02:12:39,714 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=640038.0, ans=0.0 2024-09-25 02:12:42,670 INFO [train.py:1198] (2/4) Epoch 36, batch 800, loss[loss=0.1817, ctc_loss=0.1187, cr_loss=0.3153, over 17030.00 frames. ], tot_loss[loss=0.1946, ctc_loss=0.126, cr_loss=0.3431, over 3305239.78 frames. ], batch size: 51, lr: 3.32e-03, grad_scale: 32.0 2024-09-25 02:13:08,214 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=640131.3333333334, ans=0.0 2024-09-25 02:13:09,609 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=640131.3333333334, ans=0.1 2024-09-25 02:13:26,066 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=640178.0, ans=0.125 2024-09-25 02:13:33,877 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=640224.6666666666, ans=0.2 2024-09-25 02:13:33,991 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=640224.6666666666, ans=0.0 2024-09-25 02:13:35,135 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.147e+02 1.278e+02 1.351e+02 1.457e+02 1.942e+02, threshold=2.702e+02, percent-clipped=0.0 2024-09-25 02:13:39,599 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=10.67 vs. limit=15.0 2024-09-25 02:13:50,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=640271.3333333334, ans=0.0 2024-09-25 02:14:03,922 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=640271.3333333334, ans=0.0 2024-09-25 02:14:08,468 INFO [train.py:1198] (2/4) Epoch 36, batch 850, loss[loss=0.1905, ctc_loss=0.1248, cr_loss=0.3283, over 17144.00 frames. ], tot_loss[loss=0.195, ctc_loss=0.1262, cr_loss=0.3436, over 3320052.69 frames. ], batch size: 48, lr: 3.32e-03, grad_scale: 32.0 2024-09-25 02:14:20,744 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=11.15 vs. limit=15.0 2024-09-25 02:14:26,522 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=10.19 vs. limit=22.5 2024-09-25 02:14:49,567 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=640411.3333333334, ans=0.125 2024-09-25 02:15:13,424 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=640504.6666666666, ans=0.025 2024-09-25 02:15:30,656 INFO [train.py:1198] (2/4) Epoch 36, batch 900, loss[loss=0.1829, ctc_loss=0.117, cr_loss=0.3297, over 17016.00 frames. ], tot_loss[loss=0.1948, ctc_loss=0.126, cr_loss=0.3436, over 3334309.02 frames. ], batch size: 44, lr: 3.32e-03, grad_scale: 32.0 2024-09-25 02:15:39,468 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=11.88 vs. limit=22.5 2024-09-25 02:15:48,672 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=640598.0, ans=0.0 2024-09-25 02:15:50,224 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=640598.0, ans=0.0 2024-09-25 02:16:08,010 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=640644.6666666666, ans=0.1 2024-09-25 02:16:22,098 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.143e+02 1.267e+02 1.323e+02 1.416e+02 1.789e+02, threshold=2.647e+02, percent-clipped=0.0 2024-09-25 02:16:28,948 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=640691.3333333334, ans=0.125 2024-09-25 02:16:51,065 INFO [train.py:1198] (2/4) Epoch 36, batch 950, loss[loss=0.1916, ctc_loss=0.1235, cr_loss=0.3408, over 15984.00 frames. ], tot_loss[loss=0.1951, ctc_loss=0.1263, cr_loss=0.3444, over 3343964.84 frames. ], batch size: 74, lr: 3.32e-03, grad_scale: 16.0 2024-09-25 02:17:02,616 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=640784.6666666666, ans=0.125 2024-09-25 02:17:04,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=640784.6666666666, ans=0.125 2024-09-25 02:17:14,928 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=640831.3333333334, ans=0.0 2024-09-25 02:17:15,060 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=640831.3333333334, ans=0.025 2024-09-25 02:17:24,300 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.min_positive, batch_count=640878.0, ans=0.025 2024-09-25 02:17:32,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=640878.0, ans=0.125 2024-09-25 02:18:04,975 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=640971.3333333334, ans=0.0 2024-09-25 02:18:16,108 INFO [train.py:1198] (2/4) Epoch 36, batch 1000, loss[loss=0.1895, ctc_loss=0.122, cr_loss=0.3373, over 17141.00 frames. ], tot_loss[loss=0.1954, ctc_loss=0.1265, cr_loss=0.3444, over 3339048.81 frames. ], batch size: 45, lr: 3.31e-03, grad_scale: 16.0 2024-09-25 02:18:42,439 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 02:19:10,045 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.268e+02 1.360e+02 1.441e+02 1.989e+02, threshold=2.720e+02, percent-clipped=0.0 2024-09-25 02:19:41,361 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.24 vs. limit=5.0 2024-09-25 02:19:41,706 INFO [train.py:1198] (2/4) Epoch 36, batch 1050, loss[loss=0.1879, ctc_loss=0.123, cr_loss=0.3244, over 17074.00 frames. ], tot_loss[loss=0.1952, ctc_loss=0.1264, cr_loss=0.344, over 3349822.73 frames. ], batch size: 46, lr: 3.31e-03, grad_scale: 16.0 2024-09-25 02:19:45,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=641251.3333333334, ans=0.0 2024-09-25 02:19:53,124 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=641251.3333333334, ans=0.2 2024-09-25 02:20:02,694 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=641298.0, ans=0.0 2024-09-25 02:20:25,731 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=8.73 vs. limit=15.0 2024-09-25 02:21:01,838 INFO [train.py:1198] (2/4) Epoch 36, batch 1100, loss[loss=0.1503, ctc_loss=0.09468, cr_loss=0.2782, over 17112.00 frames. ], tot_loss[loss=0.1951, ctc_loss=0.1263, cr_loss=0.3438, over 3358173.13 frames. ], batch size: 40, lr: 3.31e-03, grad_scale: 16.0 2024-09-25 02:21:19,222 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=641531.3333333334, ans=0.125 2024-09-25 02:21:43,451 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=641578.0, ans=0.1 2024-09-25 02:21:51,761 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=641624.6666666666, ans=0.125 2024-09-25 02:21:52,920 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.099e+02 1.290e+02 1.394e+02 1.498e+02 2.022e+02, threshold=2.788e+02, percent-clipped=0.0 2024-09-25 02:22:23,356 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.02 vs. limit=6.0 2024-09-25 02:22:24,254 INFO [train.py:1198] (2/4) Epoch 36, batch 1150, loss[loss=0.183, ctc_loss=0.1181, cr_loss=0.3244, over 17305.00 frames. ], tot_loss[loss=0.1948, ctc_loss=0.1261, cr_loss=0.3434, over 3362515.86 frames. ], batch size: 46, lr: 3.31e-03, grad_scale: 16.0 2024-09-25 02:22:32,624 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=641718.0, ans=0.125 2024-09-25 02:23:14,381 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=6.58 vs. limit=15.0 2024-09-25 02:23:17,530 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=9.80 vs. limit=15.0 2024-09-25 02:23:18,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=641858.0, ans=0.1 2024-09-25 02:23:49,551 INFO [train.py:1198] (2/4) Epoch 36, batch 1200, loss[loss=0.2011, ctc_loss=0.1284, cr_loss=0.3637, over 17307.00 frames. ], tot_loss[loss=0.1957, ctc_loss=0.1268, cr_loss=0.3442, over 3358420.74 frames. ], batch size: 46, lr: 3.31e-03, grad_scale: 32.0 2024-09-25 02:24:07,286 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=641998.0, ans=0.125 2024-09-25 02:24:13,676 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=641998.0, ans=0.1 2024-09-25 02:24:22,659 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=642044.6666666666, ans=0.1 2024-09-25 02:24:34,054 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=642044.6666666666, ans=0.125 2024-09-25 02:24:40,286 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=642091.3333333334, ans=0.025 2024-09-25 02:24:42,468 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=6.89 vs. limit=15.0 2024-09-25 02:24:42,607 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.40 vs. limit=6.0 2024-09-25 02:24:43,286 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.116e+02 1.264e+02 1.352e+02 1.439e+02 3.839e+02, threshold=2.704e+02, percent-clipped=1.0 2024-09-25 02:24:43,829 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=17.64 vs. limit=22.5 2024-09-25 02:25:12,608 INFO [train.py:1198] (2/4) Epoch 36, batch 1250, loss[loss=0.1913, ctc_loss=0.1238, cr_loss=0.3374, over 17145.00 frames. ], tot_loss[loss=0.196, ctc_loss=0.1271, cr_loss=0.3445, over 3354033.46 frames. ], batch size: 48, lr: 3.31e-03, grad_scale: 32.0 2024-09-25 02:25:15,098 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=8.29 vs. limit=15.0 2024-09-25 02:25:19,246 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=642184.6666666666, ans=0.1 2024-09-25 02:25:25,596 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=642184.6666666666, ans=0.1 2024-09-25 02:25:33,884 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=642231.3333333334, ans=0.0 2024-09-25 02:26:00,498 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=642324.6666666666, ans=0.125 2024-09-25 02:26:32,183 INFO [train.py:1198] (2/4) Epoch 36, batch 1300, loss[loss=0.1872, ctc_loss=0.1196, cr_loss=0.338, over 15903.00 frames. ], tot_loss[loss=0.1969, ctc_loss=0.1278, cr_loss=0.3455, over 3335072.93 frames. ], batch size: 74, lr: 3.31e-03, grad_scale: 32.0 2024-09-25 02:26:37,403 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=642418.0, ans=0.2 2024-09-25 02:26:50,151 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=642464.6666666666, ans=0.0 2024-09-25 02:27:07,079 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=642511.3333333334, ans=0.0 2024-09-25 02:27:07,091 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=642511.3333333334, ans=0.125 2024-09-25 02:27:15,034 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=642511.3333333334, ans=0.025 2024-09-25 02:27:15,138 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 02:27:25,951 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.156e+02 1.281e+02 1.359e+02 1.510e+02 2.229e+02, threshold=2.717e+02, percent-clipped=0.0 2024-09-25 02:27:26,853 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.55 vs. limit=10.0 2024-09-25 02:27:57,335 INFO [train.py:1198] (2/4) Epoch 36, batch 1350, loss[loss=0.1791, ctc_loss=0.115, cr_loss=0.3207, over 17362.00 frames. ], tot_loss[loss=0.1968, ctc_loss=0.1277, cr_loss=0.3458, over 3338214.62 frames. ], batch size: 48, lr: 3.31e-03, grad_scale: 16.0 2024-09-25 02:28:24,684 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=642698.0, ans=0.125 2024-09-25 02:28:38,194 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=642744.6666666666, ans=0.125 2024-09-25 02:28:41,681 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=642744.6666666666, ans=0.2 2024-09-25 02:29:21,998 INFO [train.py:1198] (2/4) Epoch 36, batch 1400, loss[loss=0.2422, ctc_loss=0.1613, cr_loss=0.4046, over 15129.00 frames. ], tot_loss[loss=0.1961, ctc_loss=0.1271, cr_loss=0.345, over 3343038.53 frames. ], batch size: 89, lr: 3.31e-03, grad_scale: 16.0 2024-09-25 02:29:30,212 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 02:29:49,652 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=642931.3333333334, ans=0.0 2024-09-25 02:30:14,813 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.188e+02 1.323e+02 1.435e+02 1.580e+02 6.083e+02, threshold=2.870e+02, percent-clipped=1.0 2024-09-25 02:30:42,402 INFO [train.py:1198] (2/4) Epoch 36, batch 1450, loss[loss=0.1644, ctc_loss=0.1019, cr_loss=0.3122, over 17198.00 frames. ], tot_loss[loss=0.1953, ctc_loss=0.1265, cr_loss=0.3443, over 3344745.91 frames. ], batch size: 41, lr: 3.31e-03, grad_scale: 16.0 2024-09-25 02:31:00,258 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=643164.6666666666, ans=0.1 2024-09-25 02:31:00,281 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=643164.6666666666, ans=0.2 2024-09-25 02:31:01,988 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=643164.6666666666, ans=0.125 2024-09-25 02:31:03,692 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=12.32 vs. limit=22.5 2024-09-25 02:31:18,646 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=3.97 vs. limit=15.0 2024-09-25 02:31:27,655 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=643211.3333333334, ans=0.07 2024-09-25 02:31:33,714 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=643258.0, ans=0.0 2024-09-25 02:31:41,676 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=643258.0, ans=0.125 2024-09-25 02:31:49,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=643304.6666666666, ans=0.0 2024-09-25 02:31:58,854 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=643304.6666666666, ans=0.0 2024-09-25 02:32:04,901 INFO [train.py:1198] (2/4) Epoch 36, batch 1500, loss[loss=0.1979, ctc_loss=0.1273, cr_loss=0.3531, over 17234.00 frames. ], tot_loss[loss=0.1954, ctc_loss=0.1265, cr_loss=0.3445, over 3348772.70 frames. ], batch size: 50, lr: 3.31e-03, grad_scale: 16.0 2024-09-25 02:33:00,745 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.291e+02 1.396e+02 1.518e+02 2.099e+02, threshold=2.792e+02, percent-clipped=0.0 2024-09-25 02:33:07,362 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=643491.3333333334, ans=0.125 2024-09-25 02:33:07,504 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=643491.3333333334, ans=0.2 2024-09-25 02:33:18,673 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=643538.0, ans=0.125 2024-09-25 02:33:25,053 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=643538.0, ans=0.1 2024-09-25 02:33:30,538 INFO [train.py:1198] (2/4) Epoch 36, batch 1550, loss[loss=0.1727, ctc_loss=0.1082, cr_loss=0.3228, over 17203.00 frames. ], tot_loss[loss=0.1941, ctc_loss=0.1255, cr_loss=0.343, over 3349077.10 frames. ], batch size: 47, lr: 3.31e-03, grad_scale: 16.0 2024-09-25 02:33:51,798 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=643631.3333333334, ans=0.125 2024-09-25 02:33:56,502 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=643631.3333333334, ans=0.07 2024-09-25 02:34:20,362 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=9.55 vs. limit=22.5 2024-09-25 02:34:21,445 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=643724.6666666666, ans=0.125 2024-09-25 02:34:24,804 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=643724.6666666666, ans=0.125 2024-09-25 02:34:53,198 INFO [train.py:1198] (2/4) Epoch 36, batch 1600, loss[loss=0.1652, ctc_loss=0.105, cr_loss=0.301, over 17006.00 frames. ], tot_loss[loss=0.1939, ctc_loss=0.1254, cr_loss=0.3427, over 3352029.82 frames. ], batch size: 39, lr: 3.31e-03, grad_scale: 32.0 2024-09-25 02:35:05,466 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.07 vs. limit=15.0 2024-09-25 02:35:24,185 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=643911.3333333334, ans=0.125 2024-09-25 02:35:32,122 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=643911.3333333334, ans=0.04949747468305833 2024-09-25 02:35:46,524 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.122e+02 1.268e+02 1.352e+02 1.455e+02 2.374e+02, threshold=2.704e+02, percent-clipped=0.0 2024-09-25 02:35:56,881 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=644004.6666666666, ans=0.125 2024-09-25 02:35:56,922 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=644004.6666666666, ans=0.125 2024-09-25 02:36:06,987 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=4.65 vs. limit=12.0 2024-09-25 02:36:08,112 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=644004.6666666666, ans=0.2 2024-09-25 02:36:11,514 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.max_positive, batch_count=644004.6666666666, ans=0.95 2024-09-25 02:36:14,435 INFO [train.py:1198] (2/4) Epoch 36, batch 1650, loss[loss=0.1806, ctc_loss=0.1159, cr_loss=0.3233, over 17107.00 frames. ], tot_loss[loss=0.1944, ctc_loss=0.1258, cr_loss=0.3432, over 3354770.61 frames. ], batch size: 43, lr: 3.31e-03, grad_scale: 32.0 2024-09-25 02:36:24,107 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=644051.3333333334, ans=0.0 2024-09-25 02:36:32,233 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=644098.0, ans=0.09899494936611666 2024-09-25 02:36:51,390 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=644144.6666666666, ans=0.0 2024-09-25 02:37:16,512 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=644191.3333333334, ans=0.125 2024-09-25 02:37:18,091 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=644191.3333333334, ans=0.0 2024-09-25 02:37:37,127 INFO [train.py:1198] (2/4) Epoch 36, batch 1700, loss[loss=0.1915, ctc_loss=0.1226, cr_loss=0.3447, over 17059.00 frames. ], tot_loss[loss=0.1945, ctc_loss=0.1258, cr_loss=0.3433, over 3352075.79 frames. ], batch size: 46, lr: 3.31e-03, grad_scale: 16.0 2024-09-25 02:37:37,363 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=644284.6666666666, ans=0.2 2024-09-25 02:37:57,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=644331.3333333334, ans=0.09899494936611666 2024-09-25 02:38:13,802 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=644378.0, ans=0.2 2024-09-25 02:38:37,024 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.256e+02 1.337e+02 1.426e+02 1.735e+02, threshold=2.674e+02, percent-clipped=0.0 2024-09-25 02:38:56,725 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=644471.3333333334, ans=0.125 2024-09-25 02:38:56,837 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=644471.3333333334, ans=0.125 2024-09-25 02:39:02,768 INFO [train.py:1198] (2/4) Epoch 36, batch 1750, loss[loss=0.2398, ctc_loss=0.1623, cr_loss=0.3873, over 12146.00 frames. ], tot_loss[loss=0.1946, ctc_loss=0.1259, cr_loss=0.3434, over 3355646.66 frames. ], batch size: 124, lr: 3.31e-03, grad_scale: 16.0 2024-09-25 02:39:20,278 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=644564.6666666666, ans=0.2 2024-09-25 02:39:28,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=644564.6666666666, ans=0.0 2024-09-25 02:39:47,396 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=644611.3333333334, ans=0.125 2024-09-25 02:39:47,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.min_positive, batch_count=644611.3333333334, ans=0.025 2024-09-25 02:39:47,466 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=644611.3333333334, ans=0.0 2024-09-25 02:40:08,534 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=13.70 vs. limit=22.5 2024-09-25 02:40:17,608 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=644704.6666666666, ans=0.125 2024-09-25 02:40:25,299 INFO [train.py:1198] (2/4) Epoch 36, batch 1800, loss[loss=0.2223, ctc_loss=0.148, cr_loss=0.3716, over 17154.00 frames. ], tot_loss[loss=0.1947, ctc_loss=0.126, cr_loss=0.3433, over 3364933.16 frames. ], batch size: 48, lr: 3.30e-03, grad_scale: 16.0 2024-09-25 02:40:30,365 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=644751.3333333334, ans=0.125 2024-09-25 02:40:38,397 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=644751.3333333334, ans=0.0 2024-09-25 02:40:49,867 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.92 vs. limit=10.0 2024-09-25 02:41:14,177 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=13.14 vs. limit=15.0 2024-09-25 02:41:19,904 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.269e+02 1.345e+02 1.435e+02 2.144e+02, threshold=2.691e+02, percent-clipped=0.0 2024-09-25 02:41:36,360 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=644938.0, ans=0.125 2024-09-25 02:41:42,922 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=644938.0, ans=0.0 2024-09-25 02:41:45,750 INFO [train.py:1198] (2/4) Epoch 36, batch 1850, loss[loss=0.1985, ctc_loss=0.1275, cr_loss=0.355, over 16922.00 frames. ], tot_loss[loss=0.1942, ctc_loss=0.1257, cr_loss=0.3425, over 3353757.43 frames. ], batch size: 58, lr: 3.30e-03, grad_scale: 16.0 2024-09-25 02:41:47,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=644984.6666666666, ans=0.1 2024-09-25 02:41:49,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=644984.6666666666, ans=0.125 2024-09-25 02:42:24,880 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=645078.0, ans=0.125 2024-09-25 02:42:30,818 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=9.34 vs. limit=15.0 2024-09-25 02:43:04,464 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=645171.3333333334, ans=0.09899494936611666 2024-09-25 02:43:10,544 INFO [train.py:1198] (2/4) Epoch 36, batch 1900, loss[loss=0.2271, ctc_loss=0.1449, cr_loss=0.4111, over 16989.00 frames. ], tot_loss[loss=0.1946, ctc_loss=0.1261, cr_loss=0.3429, over 3348626.24 frames. ], batch size: 53, lr: 3.30e-03, grad_scale: 16.0 2024-09-25 02:43:32,939 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=645264.6666666666, ans=0.125 2024-09-25 02:43:44,124 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer_ff2.min_abs, batch_count=645311.3333333334, ans=0.1 2024-09-25 02:44:10,239 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.279e+02 1.358e+02 1.455e+02 1.986e+02, threshold=2.715e+02, percent-clipped=0.0 2024-09-25 02:44:15,403 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=645358.0, ans=0.0 2024-09-25 02:44:26,771 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=645404.6666666666, ans=0.1 2024-09-25 02:44:26,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=645404.6666666666, ans=0.0 2024-09-25 02:44:36,298 INFO [train.py:1198] (2/4) Epoch 36, batch 1950, loss[loss=0.2156, ctc_loss=0.138, cr_loss=0.388, over 17025.00 frames. ], tot_loss[loss=0.1941, ctc_loss=0.1257, cr_loss=0.3424, over 3354655.92 frames. ], batch size: 52, lr: 3.30e-03, grad_scale: 16.0 2024-09-25 02:44:52,633 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=645498.0, ans=0.0 2024-09-25 02:44:54,383 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.31 vs. limit=15.0 2024-09-25 02:45:03,665 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=645498.0, ans=0.0 2024-09-25 02:45:05,380 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 02:45:08,601 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 02:45:25,983 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=645591.3333333334, ans=0.125 2024-09-25 02:45:38,912 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 02:45:51,671 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=645638.0, ans=0.1 2024-09-25 02:45:56,173 INFO [train.py:1198] (2/4) Epoch 36, batch 2000, loss[loss=0.2097, ctc_loss=0.1363, cr_loss=0.3671, over 17101.00 frames. ], tot_loss[loss=0.1943, ctc_loss=0.1257, cr_loss=0.343, over 3361720.88 frames. ], batch size: 49, lr: 3.30e-03, grad_scale: 32.0 2024-09-25 02:46:01,276 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=645684.6666666666, ans=0.0 2024-09-25 02:46:27,615 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.whiten, num_groups=1, num_channels=192, metric=4.55 vs. limit=12.0 2024-09-25 02:46:28,423 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.max_positive, batch_count=645778.0, ans=0.95 2024-09-25 02:46:32,201 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=15.83 vs. limit=22.5 2024-09-25 02:46:50,156 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 1.284e+02 1.382e+02 1.479e+02 3.122e+02, threshold=2.765e+02, percent-clipped=1.0 2024-09-25 02:46:50,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=645824.6666666666, ans=0.125 2024-09-25 02:47:18,701 INFO [train.py:1198] (2/4) Epoch 36, batch 2050, loss[loss=0.2059, ctc_loss=0.1347, cr_loss=0.356, over 16004.00 frames. ], tot_loss[loss=0.1948, ctc_loss=0.126, cr_loss=0.3438, over 3361153.15 frames. ], batch size: 74, lr: 3.30e-03, grad_scale: 32.0 2024-09-25 02:47:35,789 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=645964.6666666666, ans=0.1 2024-09-25 02:47:40,683 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 02:47:48,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=645964.6666666666, ans=0.0 2024-09-25 02:48:08,461 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=15.10 vs. limit=22.5 2024-09-25 02:48:26,591 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=646104.6666666666, ans=0.125 2024-09-25 02:48:32,112 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.59 vs. limit=6.0 2024-09-25 02:48:33,474 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.81 vs. limit=15.0 2024-09-25 02:48:43,712 INFO [train.py:1198] (2/4) Epoch 36, batch 2100, loss[loss=0.2027, ctc_loss=0.1346, cr_loss=0.3403, over 17285.00 frames. ], tot_loss[loss=0.1947, ctc_loss=0.1259, cr_loss=0.3439, over 3358434.49 frames. ], batch size: 49, lr: 3.30e-03, grad_scale: 32.0 2024-09-25 02:48:49,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=646151.3333333334, ans=0.125 2024-09-25 02:49:37,472 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=646291.3333333334, ans=0.125 2024-09-25 02:49:40,606 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.163e+02 1.268e+02 1.352e+02 1.474e+02 3.330e+02, threshold=2.705e+02, percent-clipped=1.0 2024-09-25 02:49:40,981 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 02:49:52,109 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=646338.0, ans=0.125 2024-09-25 02:49:52,450 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=7.35 vs. limit=15.0 2024-09-25 02:49:53,733 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=646338.0, ans=0.125 2024-09-25 02:50:06,068 INFO [train.py:1198] (2/4) Epoch 36, batch 2150, loss[loss=0.2174, ctc_loss=0.1478, cr_loss=0.3481, over 11838.00 frames. ], tot_loss[loss=0.195, ctc_loss=0.1262, cr_loss=0.344, over 3362264.66 frames. ], batch size: 123, lr: 3.30e-03, grad_scale: 32.0 2024-09-25 02:50:12,664 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=646384.6666666666, ans=0.07 2024-09-25 02:50:17,502 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=646384.6666666666, ans=0.125 2024-09-25 02:50:36,552 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=646478.0, ans=0.1 2024-09-25 02:50:49,452 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 02:51:07,466 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.55 vs. limit=6.0 2024-09-25 02:51:13,382 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 02:51:25,842 INFO [train.py:1198] (2/4) Epoch 36, batch 2200, loss[loss=0.211, ctc_loss=0.1364, cr_loss=0.3732, over 16807.00 frames. ], tot_loss[loss=0.1951, ctc_loss=0.1263, cr_loss=0.3444, over 3356173.40 frames. ], batch size: 61, lr: 3.30e-03, grad_scale: 32.0 2024-09-25 02:51:34,099 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=646618.0, ans=0.1 2024-09-25 02:52:01,039 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=2.569e-03 2024-09-25 02:52:22,863 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.282e+02 1.343e+02 1.455e+02 1.826e+02, threshold=2.686e+02, percent-clipped=0.0 2024-09-25 02:52:24,743 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=646758.0, ans=0.0 2024-09-25 02:52:51,063 INFO [train.py:1198] (2/4) Epoch 36, batch 2250, loss[loss=0.1834, ctc_loss=0.117, cr_loss=0.3321, over 17023.00 frames. ], tot_loss[loss=0.1947, ctc_loss=0.1259, cr_loss=0.3444, over 3358807.77 frames. ], batch size: 39, lr: 3.30e-03, grad_scale: 32.0 2024-09-25 02:52:54,743 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=646851.3333333334, ans=0.125 2024-09-25 02:53:02,481 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=646851.3333333334, ans=0.125 2024-09-25 02:53:13,908 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=646898.0, ans=0.2 2024-09-25 02:53:40,574 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.05 vs. limit=6.0 2024-09-25 02:54:16,046 INFO [train.py:1198] (2/4) Epoch 36, batch 2300, loss[loss=0.2083, ctc_loss=0.1355, cr_loss=0.3642, over 16979.00 frames. ], tot_loss[loss=0.1964, ctc_loss=0.1272, cr_loss=0.3462, over 3354504.31 frames. ], batch size: 56, lr: 3.30e-03, grad_scale: 8.0 2024-09-25 02:54:31,994 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.78 vs. limit=6.0 2024-09-25 02:54:48,508 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=647178.0, ans=0.2 2024-09-25 02:54:56,725 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=647178.0, ans=0.0 2024-09-25 02:55:06,680 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=3.93 vs. limit=10.0 2024-09-25 02:55:14,011 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.282e+02 1.356e+02 1.478e+02 2.427e+02, threshold=2.712e+02, percent-clipped=0.0 2024-09-25 02:55:16,142 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten.whitening_limit, batch_count=647224.6666666666, ans=15.0 2024-09-25 02:55:28,500 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=647271.3333333334, ans=0.015 2024-09-25 02:55:36,334 INFO [train.py:1198] (2/4) Epoch 36, batch 2350, loss[loss=0.1744, ctc_loss=0.1092, cr_loss=0.326, over 17068.00 frames. ], tot_loss[loss=0.1954, ctc_loss=0.1265, cr_loss=0.3446, over 3354715.44 frames. ], batch size: 43, lr: 3.30e-03, grad_scale: 8.0 2024-09-25 02:55:44,607 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=647318.0, ans=0.125 2024-09-25 02:55:50,793 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=647364.6666666666, ans=0.125 2024-09-25 02:55:55,761 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=647364.6666666666, ans=0.0 2024-09-25 02:55:59,191 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=7.06 vs. limit=15.0 2024-09-25 02:56:27,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=647458.0, ans=0.125 2024-09-25 02:56:28,288 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.21 vs. limit=10.0 2024-09-25 02:56:42,236 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=9.51 vs. limit=15.0 2024-09-25 02:56:52,527 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=647504.6666666666, ans=0.0 2024-09-25 02:56:56,657 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=4.98 vs. limit=15.0 2024-09-25 02:56:58,845 INFO [train.py:1198] (2/4) Epoch 36, batch 2400, loss[loss=0.2258, ctc_loss=0.1532, cr_loss=0.3628, over 12091.00 frames. ], tot_loss[loss=0.1956, ctc_loss=0.1267, cr_loss=0.3446, over 3351583.12 frames. ], batch size: 124, lr: 3.30e-03, grad_scale: 16.0 2024-09-25 02:57:02,838 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.46 vs. limit=22.5 2024-09-25 02:57:07,306 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=647551.3333333334, ans=0.07 2024-09-25 02:57:16,766 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=647598.0, ans=0.125 2024-09-25 02:57:43,228 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=647644.6666666666, ans=0.125 2024-09-25 02:57:58,955 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.280e+02 1.384e+02 1.461e+02 3.295e+02, threshold=2.768e+02, percent-clipped=1.0 2024-09-25 02:57:59,344 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=647691.3333333334, ans=0.0 2024-09-25 02:58:05,996 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=647738.0, ans=0.125 2024-09-25 02:58:24,117 INFO [train.py:1198] (2/4) Epoch 36, batch 2450, loss[loss=0.2148, ctc_loss=0.1377, cr_loss=0.3859, over 17203.00 frames. ], tot_loss[loss=0.1956, ctc_loss=0.1267, cr_loss=0.3446, over 3348869.64 frames. ], batch size: 50, lr: 3.30e-03, grad_scale: 16.0 2024-09-25 02:58:24,451 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=647784.6666666666, ans=0.2 2024-09-25 02:59:08,709 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=647878.0, ans=0.035 2024-09-25 02:59:15,217 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=647924.6666666666, ans=0.09899494936611666 2024-09-25 02:59:46,718 INFO [train.py:1198] (2/4) Epoch 36, batch 2500, loss[loss=0.1654, ctc_loss=0.1064, cr_loss=0.2949, over 16700.00 frames. ], tot_loss[loss=0.1961, ctc_loss=0.1271, cr_loss=0.3452, over 3353649.39 frames. ], batch size: 37, lr: 3.30e-03, grad_scale: 16.0 2024-09-25 02:59:56,911 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=15.52 vs. limit=15.0 2024-09-25 03:00:06,452 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=648064.6666666666, ans=0.125 2024-09-25 03:00:25,882 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=648111.3333333334, ans=0.0 2024-09-25 03:00:44,515 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.168e+02 1.286e+02 1.353e+02 1.494e+02 3.015e+02, threshold=2.705e+02, percent-clipped=1.0 2024-09-25 03:00:52,820 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=648204.6666666666, ans=0.0 2024-09-25 03:01:06,740 INFO [train.py:1198] (2/4) Epoch 36, batch 2550, loss[loss=0.1871, ctc_loss=0.1181, cr_loss=0.3453, over 17176.00 frames. ], tot_loss[loss=0.1947, ctc_loss=0.1261, cr_loss=0.343, over 3361491.57 frames. ], batch size: 41, lr: 3.30e-03, grad_scale: 16.0 2024-09-25 03:02:00,852 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=648391.3333333334, ans=0.0 2024-09-25 03:02:02,407 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=648391.3333333334, ans=0.0 2024-09-25 03:02:31,700 INFO [train.py:1198] (2/4) Epoch 36, batch 2600, loss[loss=0.2072, ctc_loss=0.1379, cr_loss=0.3466, over 16064.00 frames. ], tot_loss[loss=0.1941, ctc_loss=0.1257, cr_loss=0.3423, over 3368720.95 frames. ], batch size: 74, lr: 3.30e-03, grad_scale: 16.0 2024-09-25 03:02:35,523 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=10.80 vs. limit=15.0 2024-09-25 03:02:36,714 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=648484.6666666666, ans=0.1 2024-09-25 03:02:44,480 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=648484.6666666666, ans=0.0 2024-09-25 03:02:53,029 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=10.73 vs. limit=15.0 2024-09-25 03:03:12,275 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=648578.0, ans=0.025 2024-09-25 03:03:31,213 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.121e+02 1.288e+02 1.366e+02 1.471e+02 2.094e+02, threshold=2.731e+02, percent-clipped=0.0 2024-09-25 03:03:48,734 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=3.84 vs. limit=10.0 2024-09-25 03:03:56,674 INFO [train.py:1198] (2/4) Epoch 36, batch 2650, loss[loss=0.2091, ctc_loss=0.1371, cr_loss=0.3601, over 15939.00 frames. ], tot_loss[loss=0.1949, ctc_loss=0.1263, cr_loss=0.343, over 3356112.88 frames. ], batch size: 74, lr: 3.29e-03, grad_scale: 16.0 2024-09-25 03:04:03,801 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.58 vs. limit=15.0 2024-09-25 03:04:16,160 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 03:04:21,501 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn1.whiten.whitening_limit, batch_count=648764.6666666666, ans=22.5 2024-09-25 03:05:02,563 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=648904.6666666666, ans=0.05 2024-09-25 03:05:09,342 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2.whitening_limit, batch_count=648904.6666666666, ans=15.0 2024-09-25 03:05:13,667 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=648904.6666666666, ans=0.0 2024-09-25 03:05:16,416 INFO [train.py:1198] (2/4) Epoch 36, batch 2700, loss[loss=0.1904, ctc_loss=0.1219, cr_loss=0.3426, over 17171.00 frames. ], tot_loss[loss=0.1944, ctc_loss=0.1259, cr_loss=0.3424, over 3362224.73 frames. ], batch size: 45, lr: 3.29e-03, grad_scale: 8.0 2024-09-25 03:05:18,304 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=648951.3333333334, ans=0.0 2024-09-25 03:05:20,459 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=17.89 vs. limit=22.5 2024-09-25 03:05:27,511 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=648951.3333333334, ans=0.95 2024-09-25 03:05:33,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=648998.0, ans=0.2 2024-09-25 03:06:15,209 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.266e+02 1.332e+02 1.436e+02 2.982e+02, threshold=2.664e+02, percent-clipped=1.0 2024-09-25 03:06:15,667 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=649091.3333333334, ans=0.2 2024-09-25 03:06:18,709 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=649138.0, ans=0.125 2024-09-25 03:06:23,501 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=649138.0, ans=0.1 2024-09-25 03:06:35,962 INFO [train.py:1198] (2/4) Epoch 36, batch 2750, loss[loss=0.2136, ctc_loss=0.1373, cr_loss=0.3817, over 16904.00 frames. ], tot_loss[loss=0.195, ctc_loss=0.1263, cr_loss=0.3433, over 3363971.39 frames. ], batch size: 58, lr: 3.29e-03, grad_scale: 8.0 2024-09-25 03:07:40,696 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=649324.6666666666, ans=0.5 2024-09-25 03:07:41,327 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=5.53 vs. limit=15.0 2024-09-25 03:07:53,453 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=649371.3333333334, ans=0.125 2024-09-25 03:08:01,227 INFO [train.py:1198] (2/4) Epoch 36, batch 2800, loss[loss=0.1587, ctc_loss=0.1009, cr_loss=0.289, over 17031.00 frames. ], tot_loss[loss=0.1936, ctc_loss=0.1253, cr_loss=0.3417, over 3360938.41 frames. ], batch size: 39, lr: 3.29e-03, grad_scale: 16.0 2024-09-25 03:08:17,175 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.64 vs. limit=15.0 2024-09-25 03:08:21,726 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=649464.6666666666, ans=0.125 2024-09-25 03:08:37,898 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=649511.3333333334, ans=0.2 2024-09-25 03:08:50,587 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=649558.0, ans=0.125 2024-09-25 03:09:05,753 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.121e+02 1.257e+02 1.332e+02 1.427e+02 2.151e+02, threshold=2.663e+02, percent-clipped=0.0 2024-09-25 03:09:19,355 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=649604.6666666666, ans=0.025 2024-09-25 03:09:19,396 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=649604.6666666666, ans=0.125 2024-09-25 03:09:22,494 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=649604.6666666666, ans=0.125 2024-09-25 03:09:27,092 INFO [train.py:1198] (2/4) Epoch 36, batch 2850, loss[loss=0.1976, ctc_loss=0.1283, cr_loss=0.3462, over 17305.00 frames. ], tot_loss[loss=0.192, ctc_loss=0.1241, cr_loss=0.3391, over 3355291.53 frames. ], batch size: 51, lr: 3.29e-03, grad_scale: 16.0 2024-09-25 03:09:35,499 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=649651.3333333334, ans=0.2 2024-09-25 03:09:35,596 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 03:09:48,285 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=649698.0, ans=0.2 2024-09-25 03:10:46,601 INFO [train.py:1198] (2/4) Epoch 36, batch 2900, loss[loss=0.2277, ctc_loss=0.1475, cr_loss=0.4013, over 17317.00 frames. ], tot_loss[loss=0.1927, ctc_loss=0.1245, cr_loss=0.3409, over 3357109.26 frames. ], batch size: 51, lr: 3.29e-03, grad_scale: 16.0 2024-09-25 03:10:48,479 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=649884.6666666666, ans=0.2 2024-09-25 03:10:55,223 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=3.61 vs. limit=6.0 2024-09-25 03:11:15,367 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=649931.3333333334, ans=0.0 2024-09-25 03:11:38,758 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=5.91 vs. limit=15.0 2024-09-25 03:11:41,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.whiten.whitening_limit, batch_count=650024.6666666666, ans=12.0 2024-09-25 03:11:45,703 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.147e+02 1.278e+02 1.363e+02 1.442e+02 1.816e+02, threshold=2.725e+02, percent-clipped=0.0 2024-09-25 03:11:58,153 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=650071.3333333334, ans=0.09899494936611666 2024-09-25 03:11:59,610 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=650071.3333333334, ans=0.0 2024-09-25 03:12:07,736 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.max_positive, batch_count=650118.0, ans=0.95 2024-09-25 03:12:08,898 INFO [train.py:1198] (2/4) Epoch 36, batch 2950, loss[loss=0.1919, ctc_loss=0.1231, cr_loss=0.3444, over 17298.00 frames. ], tot_loss[loss=0.1938, ctc_loss=0.1254, cr_loss=0.3424, over 3359332.39 frames. ], batch size: 46, lr: 3.29e-03, grad_scale: 16.0 2024-09-25 03:12:17,989 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=650118.0, ans=0.0 2024-09-25 03:12:37,156 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=650164.6666666666, ans=0.125 2024-09-25 03:13:32,846 INFO [train.py:1198] (2/4) Epoch 36, batch 3000, loss[loss=0.2103, ctc_loss=0.1366, cr_loss=0.3685, over 17152.00 frames. ], tot_loss[loss=0.1943, ctc_loss=0.1256, cr_loss=0.3432, over 3360041.58 frames. ], batch size: 48, lr: 3.29e-03, grad_scale: 16.0 2024-09-25 03:13:32,847 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 03:13:48,648 INFO [train.py:1230] (2/4) Epoch 36, validation: loss=0.03616, ctc_loss=0.03616, cr_loss=9.264e-15, over 944034.00 frames. 2024-09-25 03:13:48,649 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 03:14:06,396 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=650398.0, ans=0.1 2024-09-25 03:14:13,052 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=9.51 vs. limit=15.0 2024-09-25 03:14:20,484 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=650444.6666666666, ans=0.125 2024-09-25 03:14:26,516 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=650444.6666666666, ans=0.04949747468305833 2024-09-25 03:14:32,902 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 03:14:48,260 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.70 vs. limit=10.0 2024-09-25 03:14:48,920 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.136e+02 1.273e+02 1.377e+02 1.498e+02 1.977e+02, threshold=2.754e+02, percent-clipped=0.0 2024-09-25 03:15:02,310 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=650538.0, ans=0.125 2024-09-25 03:15:09,705 INFO [train.py:1198] (2/4) Epoch 36, batch 3050, loss[loss=0.1946, ctc_loss=0.1297, cr_loss=0.3242, over 16093.00 frames. ], tot_loss[loss=0.1939, ctc_loss=0.1255, cr_loss=0.3421, over 3339855.67 frames. ], batch size: 74, lr: 3.29e-03, grad_scale: 16.0 2024-09-25 03:15:18,019 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=650584.6666666666, ans=0.125 2024-09-25 03:15:19,843 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=8.56 vs. limit=15.0 2024-09-25 03:15:25,646 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=650631.3333333334, ans=0.125 2024-09-25 03:15:30,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=650631.3333333334, ans=0.125 2024-09-25 03:15:31,931 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=650631.3333333334, ans=0.05 2024-09-25 03:15:34,951 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=650631.3333333334, ans=0.125 2024-09-25 03:15:38,958 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=3.99 vs. limit=15.0 2024-09-25 03:15:39,745 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.max_abs, batch_count=650678.0, ans=10.0 2024-09-25 03:15:43,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=650678.0, ans=0.1 2024-09-25 03:15:49,228 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=650678.0, ans=0.125 2024-09-25 03:15:54,061 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=650678.0, ans=0.125 2024-09-25 03:16:05,057 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=650724.6666666666, ans=0.125 2024-09-25 03:16:24,131 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=650771.3333333334, ans=0.04949747468305833 2024-09-25 03:16:27,113 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=650818.0, ans=0.1 2024-09-25 03:16:28,398 INFO [train.py:1198] (2/4) Epoch 36, batch 3100, loss[loss=0.1966, ctc_loss=0.1225, cr_loss=0.3706, over 17177.00 frames. ], tot_loss[loss=0.1946, ctc_loss=0.1259, cr_loss=0.3435, over 3348407.99 frames. ], batch size: 45, lr: 3.29e-03, grad_scale: 16.0 2024-09-25 03:16:45,939 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 03:17:15,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=650958.0, ans=0.125 2024-09-25 03:17:26,168 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.257e+02 1.357e+02 1.451e+02 2.020e+02, threshold=2.713e+02, percent-clipped=0.0 2024-09-25 03:17:46,709 INFO [train.py:1198] (2/4) Epoch 36, batch 3150, loss[loss=0.1778, ctc_loss=0.1141, cr_loss=0.3187, over 17296.00 frames. ], tot_loss[loss=0.1939, ctc_loss=0.1254, cr_loss=0.3428, over 3337827.82 frames. ], batch size: 51, lr: 3.29e-03, grad_scale: 16.0 2024-09-25 03:18:05,528 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=651098.0, ans=0.2 2024-09-25 03:18:11,964 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=651098.0, ans=0.1 2024-09-25 03:18:52,806 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=651238.0, ans=0.1 2024-09-25 03:18:53,170 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten.whitening_limit, batch_count=651238.0, ans=15.0 2024-09-25 03:18:55,947 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=651238.0, ans=0.0 2024-09-25 03:19:05,093 INFO [train.py:1198] (2/4) Epoch 36, batch 3200, loss[loss=0.1736, ctc_loss=0.1108, cr_loss=0.3142, over 17183.00 frames. ], tot_loss[loss=0.1942, ctc_loss=0.1257, cr_loss=0.3427, over 3340095.86 frames. ], batch size: 41, lr: 3.29e-03, grad_scale: 32.0 2024-09-25 03:19:11,827 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.min_positive, batch_count=651284.6666666666, ans=0.025 2024-09-25 03:19:44,585 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=651378.0, ans=0.1 2024-09-25 03:19:46,228 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=651378.0, ans=0.0 2024-09-25 03:19:49,304 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=651378.0, ans=0.125 2024-09-25 03:20:02,960 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.112e+02 1.272e+02 1.367e+02 1.451e+02 1.708e+02, threshold=2.733e+02, percent-clipped=0.0 2024-09-25 03:20:23,374 INFO [train.py:1198] (2/4) Epoch 36, batch 3250, loss[loss=0.1638, ctc_loss=0.1051, cr_loss=0.2934, over 17043.00 frames. ], tot_loss[loss=0.1944, ctc_loss=0.1258, cr_loss=0.3429, over 3338188.02 frames. ], batch size: 39, lr: 3.29e-03, grad_scale: 32.0 2024-09-25 03:20:23,717 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=651518.0, ans=0.07 2024-09-25 03:20:29,991 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 03:20:31,408 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=651518.0, ans=0.125 2024-09-25 03:20:35,973 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=651518.0, ans=0.125 2024-09-25 03:21:02,555 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.75 vs. limit=6.0 2024-09-25 03:21:10,468 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=3.99 vs. limit=15.0 2024-09-25 03:21:26,991 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=651704.6666666666, ans=0.0 2024-09-25 03:21:30,263 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=651704.6666666666, ans=0.125 2024-09-25 03:21:43,850 INFO [train.py:1198] (2/4) Epoch 36, batch 3300, loss[loss=0.1755, ctc_loss=0.113, cr_loss=0.3121, over 17244.00 frames. ], tot_loss[loss=0.1943, ctc_loss=0.1257, cr_loss=0.3432, over 3347220.17 frames. ], batch size: 42, lr: 3.29e-03, grad_scale: 32.0 2024-09-25 03:21:47,837 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=651751.3333333334, ans=0.1 2024-09-25 03:22:35,693 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=651891.3333333334, ans=0.125 2024-09-25 03:22:44,634 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.051e+02 1.278e+02 1.344e+02 1.481e+02 2.113e+02, threshold=2.688e+02, percent-clipped=0.0 2024-09-25 03:22:48,015 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=651938.0, ans=0.125 2024-09-25 03:23:03,251 INFO [train.py:1198] (2/4) Epoch 36, batch 3350, loss[loss=0.2419, ctc_loss=0.162, cr_loss=0.3993, over 16996.00 frames. ], tot_loss[loss=0.1954, ctc_loss=0.1265, cr_loss=0.3449, over 3352243.97 frames. ], batch size: 53, lr: 3.29e-03, grad_scale: 16.0 2024-09-25 03:23:04,154 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.43 vs. limit=6.0 2024-09-25 03:23:13,014 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=651984.6666666666, ans=0.125 2024-09-25 03:23:13,025 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=651984.6666666666, ans=0.025 2024-09-25 03:23:14,550 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=651984.6666666666, ans=0.1 2024-09-25 03:23:21,075 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.93 vs. limit=6.0 2024-09-25 03:23:40,043 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=652078.0, ans=0.125 2024-09-25 03:24:04,935 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=652124.6666666666, ans=0.125 2024-09-25 03:24:06,477 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=652171.3333333334, ans=0.0 2024-09-25 03:24:09,703 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=652171.3333333334, ans=0.125 2024-09-25 03:24:17,519 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=652171.3333333334, ans=0.125 2024-09-25 03:24:20,707 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=652171.3333333334, ans=0.125 2024-09-25 03:24:23,598 INFO [train.py:1198] (2/4) Epoch 36, batch 3400, loss[loss=0.2221, ctc_loss=0.1425, cr_loss=0.3976, over 16074.00 frames. ], tot_loss[loss=0.197, ctc_loss=0.1276, cr_loss=0.3471, over 3345171.94 frames. ], batch size: 74, lr: 3.29e-03, grad_scale: 16.0 2024-09-25 03:25:07,751 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=652311.3333333334, ans=0.2 2024-09-25 03:25:22,899 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.287e+02 1.358e+02 1.469e+02 2.050e+02, threshold=2.715e+02, percent-clipped=0.0 2024-09-25 03:25:44,229 INFO [train.py:1198] (2/4) Epoch 36, batch 3450, loss[loss=0.2431, ctc_loss=0.1593, cr_loss=0.4189, over 16575.00 frames. ], tot_loss[loss=0.1964, ctc_loss=0.1272, cr_loss=0.346, over 3340082.11 frames. ], batch size: 66, lr: 3.29e-03, grad_scale: 16.0 2024-09-25 03:26:07,939 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=652498.0, ans=0.1 2024-09-25 03:26:11,005 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=652498.0, ans=0.125 2024-09-25 03:26:15,549 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=652544.6666666666, ans=0.125 2024-09-25 03:26:17,233 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=652544.6666666666, ans=0.07 2024-09-25 03:26:20,233 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=652544.6666666666, ans=0.125 2024-09-25 03:26:31,409 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=652591.3333333334, ans=0.0 2024-09-25 03:26:40,427 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=652591.3333333334, ans=0.125 2024-09-25 03:26:45,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.max_abs, batch_count=652638.0, ans=10.0 2024-09-25 03:26:56,045 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=652638.0, ans=0.125 2024-09-25 03:27:00,791 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=652684.6666666666, ans=0.125 2024-09-25 03:27:02,026 INFO [train.py:1198] (2/4) Epoch 36, batch 3500, loss[loss=0.2177, ctc_loss=0.1417, cr_loss=0.3798, over 17110.00 frames. ], tot_loss[loss=0.1973, ctc_loss=0.128, cr_loss=0.3468, over 3324837.13 frames. ], batch size: 49, lr: 3.28e-03, grad_scale: 16.0 2024-09-25 03:27:13,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=652684.6666666666, ans=0.2 2024-09-25 03:27:22,261 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=652731.3333333334, ans=0.125 2024-09-25 03:27:23,953 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=652731.3333333334, ans=0.1 2024-09-25 03:28:02,935 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.150e+02 1.297e+02 1.416e+02 1.575e+02 3.387e+02, threshold=2.833e+02, percent-clipped=1.0 2024-09-25 03:28:20,554 INFO [train.py:1198] (2/4) Epoch 36, batch 3550, loss[loss=0.1879, ctc_loss=0.1212, cr_loss=0.3333, over 17056.00 frames. ], tot_loss[loss=0.1961, ctc_loss=0.1271, cr_loss=0.345, over 3342084.98 frames. ], batch size: 39, lr: 3.28e-03, grad_scale: 8.0 2024-09-25 03:28:22,352 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=652918.0, ans=0.05 2024-09-25 03:29:17,820 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=11.54 vs. limit=22.5 2024-09-25 03:29:38,744 INFO [train.py:1198] (2/4) Epoch 36, batch 3600, loss[loss=0.1754, ctc_loss=0.1108, cr_loss=0.3227, over 17216.00 frames. ], tot_loss[loss=0.1953, ctc_loss=0.1265, cr_loss=0.3441, over 3351093.56 frames. ], batch size: 47, lr: 3.28e-03, grad_scale: 16.0 2024-09-25 03:29:49,727 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=653151.3333333334, ans=0.125 2024-09-25 03:30:02,650 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=14.66 vs. limit=22.5 2024-09-25 03:30:04,037 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=653198.0, ans=0.0 2024-09-25 03:30:31,522 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=13.34 vs. limit=15.0 2024-09-25 03:30:43,994 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.131e+02 1.277e+02 1.347e+02 1.445e+02 1.925e+02, threshold=2.694e+02, percent-clipped=0.0 2024-09-25 03:30:59,781 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=653384.6666666666, ans=0.2 2024-09-25 03:31:00,962 INFO [train.py:1198] (2/4) Epoch 36, batch 3650, loss[loss=0.1889, ctc_loss=0.1207, cr_loss=0.341, over 17359.00 frames. ], tot_loss[loss=0.1957, ctc_loss=0.1268, cr_loss=0.3446, over 3351238.54 frames. ], batch size: 48, lr: 3.28e-03, grad_scale: 16.0 2024-09-25 03:31:14,020 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=653384.6666666666, ans=0.015 2024-09-25 03:31:24,979 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=653431.3333333334, ans=0.125 2024-09-25 03:32:02,106 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=653524.6666666666, ans=0.025 2024-09-25 03:32:06,821 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 03:32:14,683 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=653571.3333333334, ans=0.2 2024-09-25 03:32:22,170 INFO [train.py:1198] (2/4) Epoch 36, batch 3700, loss[loss=0.1844, ctc_loss=0.1181, cr_loss=0.3311, over 17290.00 frames. ], tot_loss[loss=0.196, ctc_loss=0.1269, cr_loss=0.3454, over 3351487.14 frames. ], batch size: 51, lr: 3.28e-03, grad_scale: 16.0 2024-09-25 03:32:35,283 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=18.73 vs. limit=22.5 2024-09-25 03:32:36,776 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 03:32:42,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=653664.6666666666, ans=0.1 2024-09-25 03:32:44,352 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=653664.6666666666, ans=0.07 2024-09-25 03:32:51,570 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=653664.6666666666, ans=0.125 2024-09-25 03:33:21,293 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=653758.0, ans=0.125 2024-09-25 03:33:24,003 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.071e+02 1.267e+02 1.341e+02 1.465e+02 1.805e+02, threshold=2.682e+02, percent-clipped=0.0 2024-09-25 03:33:41,504 INFO [train.py:1198] (2/4) Epoch 36, batch 3750, loss[loss=0.1715, ctc_loss=0.1099, cr_loss=0.3082, over 17345.00 frames. ], tot_loss[loss=0.1954, ctc_loss=0.1265, cr_loss=0.3441, over 3342958.69 frames. ], batch size: 48, lr: 3.28e-03, grad_scale: 16.0 2024-09-25 03:33:47,045 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=16.99 vs. limit=22.5 2024-09-25 03:33:59,335 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.95 vs. limit=6.0 2024-09-25 03:34:10,557 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.71 vs. limit=6.0 2024-09-25 03:34:41,921 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=14.85 vs. limit=22.5 2024-09-25 03:35:01,333 INFO [train.py:1198] (2/4) Epoch 36, batch 3800, loss[loss=0.1893, ctc_loss=0.1207, cr_loss=0.3425, over 17181.00 frames. ], tot_loss[loss=0.1948, ctc_loss=0.1261, cr_loss=0.3431, over 3324308.75 frames. ], batch size: 45, lr: 3.28e-03, grad_scale: 16.0 2024-09-25 03:35:03,242 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=654084.6666666666, ans=0.125 2024-09-25 03:35:07,888 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=654084.6666666666, ans=0.0 2024-09-25 03:35:09,359 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=654084.6666666666, ans=0.1 2024-09-25 03:35:40,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=654178.0, ans=0.1 2024-09-25 03:35:50,076 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=654224.6666666666, ans=0.0 2024-09-25 03:36:02,509 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.295e+02 1.360e+02 1.464e+02 2.754e+02, threshold=2.720e+02, percent-clipped=1.0 2024-09-25 03:36:02,743 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=654271.3333333334, ans=0.2 2024-09-25 03:36:02,915 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-25 03:36:19,640 INFO [train.py:1198] (2/4) Epoch 36, batch 3850, loss[loss=0.1724, ctc_loss=0.1072, cr_loss=0.326, over 17025.00 frames. ], tot_loss[loss=0.1934, ctc_loss=0.1252, cr_loss=0.3414, over 3313627.38 frames. ], batch size: 39, lr: 3.28e-03, grad_scale: 16.0 2024-09-25 03:36:47,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=654364.6666666666, ans=0.125 2024-09-25 03:37:12,803 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=654458.0, ans=0.125 2024-09-25 03:37:21,886 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=654504.6666666666, ans=0.0 2024-09-25 03:37:25,037 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=654504.6666666666, ans=0.125 2024-09-25 03:38:18,969 INFO [train.py:1198] (2/4) Epoch 37, batch 0, loss[loss=0.2024, ctc_loss=0.1334, cr_loss=0.3451, over 12261.00 frames. ], tot_loss[loss=0.2024, ctc_loss=0.1334, cr_loss=0.3451, over 12261.00 frames. ], batch size: 123, lr: 3.23e-03, grad_scale: 16.0 2024-09-25 03:38:18,969 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 03:38:34,303 INFO [train.py:1230] (2/4) Epoch 37, validation: loss=0.03489, ctc_loss=0.03489, cr_loss=9.463e-15, over 944034.00 frames. 2024-09-25 03:38:34,304 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 03:38:37,851 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=654532.6666666666, ans=0.0 2024-09-25 03:38:39,409 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=654532.6666666666, ans=0.2 2024-09-25 03:38:52,342 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=654579.3333333334, ans=0.125 2024-09-25 03:39:04,985 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=654626.0, ans=0.2 2024-09-25 03:39:47,467 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.132e+02 1.358e+02 1.528e+02 1.726e+02 8.114e+02, threshold=3.057e+02, percent-clipped=1.0 2024-09-25 03:39:57,023 INFO [train.py:1198] (2/4) Epoch 37, batch 50, loss[loss=0.1996, ctc_loss=0.1261, cr_loss=0.3673, over 17101.00 frames. ], tot_loss[loss=0.1965, ctc_loss=0.1273, cr_loss=0.3456, over 754256.53 frames. ], batch size: 49, lr: 3.23e-03, grad_scale: 16.0 2024-09-25 03:40:03,609 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=654766.0, ans=0.0 2024-09-25 03:40:08,473 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=654766.0, ans=0.0 2024-09-25 03:40:54,006 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=654906.0, ans=0.0 2024-09-25 03:40:59,286 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.59 vs. limit=15.0 2024-09-25 03:41:19,001 INFO [train.py:1198] (2/4) Epoch 37, batch 100, loss[loss=0.2105, ctc_loss=0.1379, cr_loss=0.363, over 17158.00 frames. ], tot_loss[loss=0.1956, ctc_loss=0.1267, cr_loss=0.3445, over 1316049.93 frames. ], batch size: 45, lr: 3.23e-03, grad_scale: 16.0 2024-09-25 03:41:32,277 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=17.82 vs. limit=22.5 2024-09-25 03:42:21,542 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=655186.0, ans=0.0 2024-09-25 03:42:29,428 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.170e+02 1.259e+02 1.323e+02 1.391e+02 2.896e+02, threshold=2.646e+02, percent-clipped=0.0 2024-09-25 03:42:39,024 INFO [train.py:1198] (2/4) Epoch 37, batch 150, loss[loss=0.1501, ctc_loss=0.09276, cr_loss=0.2866, over 16269.00 frames. ], tot_loss[loss=0.1945, ctc_loss=0.1259, cr_loss=0.3429, over 1764527.01 frames. ], batch size: 36, lr: 3.23e-03, grad_scale: 16.0 2024-09-25 03:42:50,619 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=655232.6666666666, ans=0.125 2024-09-25 03:43:06,307 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=655279.3333333334, ans=0.2 2024-09-25 03:43:27,378 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.70 vs. limit=15.0 2024-09-25 03:44:06,076 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=655466.0, ans=0.125 2024-09-25 03:44:07,479 INFO [train.py:1198] (2/4) Epoch 37, batch 200, loss[loss=0.2232, ctc_loss=0.1472, cr_loss=0.3798, over 17043.00 frames. ], tot_loss[loss=0.1932, ctc_loss=0.1249, cr_loss=0.3417, over 2122658.74 frames. ], batch size: 52, lr: 3.23e-03, grad_scale: 16.0 2024-09-25 03:44:25,085 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=655512.6666666666, ans=0.125 2024-09-25 03:45:19,826 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.270e+02 1.344e+02 1.450e+02 1.913e+02, threshold=2.687e+02, percent-clipped=0.0 2024-09-25 03:45:29,186 INFO [train.py:1198] (2/4) Epoch 37, batch 250, loss[loss=0.1837, ctc_loss=0.1166, cr_loss=0.3354, over 17133.00 frames. ], tot_loss[loss=0.1935, ctc_loss=0.1251, cr_loss=0.342, over 2392389.14 frames. ], batch size: 48, lr: 3.23e-03, grad_scale: 16.0 2024-09-25 03:46:03,806 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=655792.6666666666, ans=0.125 2024-09-25 03:46:10,764 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.75 vs. limit=6.0 2024-09-25 03:46:26,500 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.min_positive, batch_count=655839.3333333334, ans=0.05 2024-09-25 03:46:47,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=655886.0, ans=0.0 2024-09-25 03:46:50,161 INFO [train.py:1198] (2/4) Epoch 37, batch 300, loss[loss=0.1673, ctc_loss=0.1061, cr_loss=0.3057, over 17019.00 frames. ], tot_loss[loss=0.1946, ctc_loss=0.1259, cr_loss=0.3434, over 2602183.59 frames. ], batch size: 39, lr: 3.23e-03, grad_scale: 16.0 2024-09-25 03:46:55,216 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=655932.6666666666, ans=0.125 2024-09-25 03:47:12,068 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=6.94 vs. limit=8.0 2024-09-25 03:47:17,735 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=655979.3333333334, ans=0.125 2024-09-25 03:47:27,317 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=656026.0, ans=0.125 2024-09-25 03:47:41,918 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=656072.6666666666, ans=0.0 2024-09-25 03:47:46,524 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=656072.6666666666, ans=0.0 2024-09-25 03:47:51,243 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=656072.6666666666, ans=0.2 2024-09-25 03:47:57,519 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=656119.3333333334, ans=0.0 2024-09-25 03:48:00,707 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.029e+02 1.285e+02 1.370e+02 1.458e+02 2.873e+02, threshold=2.740e+02, percent-clipped=1.0 2024-09-25 03:48:10,573 INFO [train.py:1198] (2/4) Epoch 37, batch 350, loss[loss=0.1834, ctc_loss=0.1166, cr_loss=0.3342, over 17099.00 frames. ], tot_loss[loss=0.1955, ctc_loss=0.1266, cr_loss=0.3447, over 2762707.18 frames. ], batch size: 49, lr: 3.23e-03, grad_scale: 16.0 2024-09-25 03:48:24,634 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=656166.0, ans=0.0 2024-09-25 03:48:35,576 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.min_positive, batch_count=656212.6666666666, ans=0.025 2024-09-25 03:48:46,614 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=656259.3333333334, ans=0.125 2024-09-25 03:49:11,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=656306.0, ans=0.0 2024-09-25 03:49:38,696 INFO [train.py:1198] (2/4) Epoch 37, batch 400, loss[loss=0.2074, ctc_loss=0.133, cr_loss=0.3723, over 16475.00 frames. ], tot_loss[loss=0.1946, ctc_loss=0.1261, cr_loss=0.3428, over 2877753.05 frames. ], batch size: 66, lr: 3.23e-03, grad_scale: 32.0 2024-09-25 03:50:01,151 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.min_positive, batch_count=656446.0, ans=0.025 2024-09-25 03:50:23,281 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=5.16 vs. limit=15.0 2024-09-25 03:50:37,330 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=656539.3333333334, ans=0.125 2024-09-25 03:50:50,159 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=656586.0, ans=0.125 2024-09-25 03:50:51,380 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.258e+02 1.358e+02 1.447e+02 2.750e+02, threshold=2.715e+02, percent-clipped=1.0 2024-09-25 03:51:01,055 INFO [train.py:1198] (2/4) Epoch 37, batch 450, loss[loss=0.2053, ctc_loss=0.1313, cr_loss=0.37, over 17345.00 frames. ], tot_loss[loss=0.1957, ctc_loss=0.1267, cr_loss=0.3451, over 2986014.39 frames. ], batch size: 48, lr: 3.23e-03, grad_scale: 32.0 2024-09-25 03:51:19,078 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=656679.3333333334, ans=0.125 2024-09-25 03:51:21,143 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.52 vs. limit=15.0 2024-09-25 03:51:22,186 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.min_positive, batch_count=656679.3333333334, ans=0.025 2024-09-25 03:51:24,480 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=16.61 vs. limit=22.5 2024-09-25 03:51:31,877 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=656726.0, ans=0.1 2024-09-25 03:51:31,937 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=656726.0, ans=0.2 2024-09-25 03:51:57,652 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=656772.6666666666, ans=0.0 2024-09-25 03:52:13,361 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=656819.3333333334, ans=0.125 2024-09-25 03:52:21,375 INFO [train.py:1198] (2/4) Epoch 37, batch 500, loss[loss=0.1846, ctc_loss=0.1192, cr_loss=0.3269, over 17085.00 frames. ], tot_loss[loss=0.1942, ctc_loss=0.1255, cr_loss=0.3435, over 3081028.41 frames. ], batch size: 49, lr: 3.23e-03, grad_scale: 32.0 2024-09-25 03:52:26,416 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=656866.0, ans=0.1 2024-09-25 03:52:43,946 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=656912.6666666666, ans=0.2 2024-09-25 03:53:00,393 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.81 vs. limit=6.0 2024-09-25 03:53:09,124 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=656959.3333333334, ans=0.025 2024-09-25 03:53:15,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=657006.0, ans=0.125 2024-09-25 03:53:37,041 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.149e+02 1.284e+02 1.360e+02 1.514e+02 2.432e+02, threshold=2.721e+02, percent-clipped=0.0 2024-09-25 03:53:46,707 INFO [train.py:1198] (2/4) Epoch 37, batch 550, loss[loss=0.1567, ctc_loss=0.09752, cr_loss=0.2957, over 17081.00 frames. ], tot_loss[loss=0.1935, ctc_loss=0.125, cr_loss=0.3426, over 3149368.68 frames. ], batch size: 43, lr: 3.23e-03, grad_scale: 32.0 2024-09-25 03:54:11,048 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=6.05 vs. limit=15.0 2024-09-25 03:54:15,021 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=657146.0, ans=0.0 2024-09-25 03:54:18,494 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.68 vs. limit=15.0 2024-09-25 03:55:03,369 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=657286.0, ans=0.0 2024-09-25 03:55:07,961 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=657286.0, ans=0.125 2024-09-25 03:55:12,200 INFO [train.py:1198] (2/4) Epoch 37, batch 600, loss[loss=0.2385, ctc_loss=0.1562, cr_loss=0.4116, over 15135.00 frames. ], tot_loss[loss=0.1945, ctc_loss=0.1257, cr_loss=0.3441, over 3195236.74 frames. ], batch size: 89, lr: 3.23e-03, grad_scale: 32.0 2024-09-25 03:55:19,038 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=657332.6666666666, ans=0.0 2024-09-25 03:55:54,046 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=657426.0, ans=0.0 2024-09-25 03:55:54,085 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=657426.0, ans=0.125 2024-09-25 03:56:00,368 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=657472.6666666666, ans=0.125 2024-09-25 03:56:02,086 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=657472.6666666666, ans=0.0 2024-09-25 03:56:13,261 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=657472.6666666666, ans=0.125 2024-09-25 03:56:22,506 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.149e+02 1.287e+02 1.356e+02 1.479e+02 3.442e+02, threshold=2.712e+02, percent-clipped=1.0 2024-09-25 03:56:32,225 INFO [train.py:1198] (2/4) Epoch 37, batch 650, loss[loss=0.2257, ctc_loss=0.1495, cr_loss=0.381, over 16860.00 frames. ], tot_loss[loss=0.1951, ctc_loss=0.1261, cr_loss=0.3448, over 3234486.27 frames. ], batch size: 58, lr: 3.23e-03, grad_scale: 32.0 2024-09-25 03:56:37,438 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=657566.0, ans=0.2 2024-09-25 03:56:37,440 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=657566.0, ans=0.1 2024-09-25 03:56:47,028 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=657612.6666666666, ans=0.125 2024-09-25 03:57:26,881 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.max_positive, batch_count=657706.0, ans=0.95 2024-09-25 03:57:27,531 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=10.07 vs. limit=15.0 2024-09-25 03:57:40,164 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=657752.6666666666, ans=0.0 2024-09-25 03:57:46,566 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=657752.6666666666, ans=0.2 2024-09-25 03:57:52,559 INFO [train.py:1198] (2/4) Epoch 37, batch 700, loss[loss=0.1873, ctc_loss=0.123, cr_loss=0.3212, over 17366.00 frames. ], tot_loss[loss=0.1945, ctc_loss=0.1258, cr_loss=0.3432, over 3264623.00 frames. ], batch size: 48, lr: 3.23e-03, grad_scale: 16.0 2024-09-25 03:58:05,997 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.16 vs. limit=15.0 2024-09-25 03:58:29,144 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=657892.6666666666, ans=0.2 2024-09-25 03:59:06,772 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=657986.0, ans=0.125 2024-09-25 03:59:13,008 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.135e+02 1.266e+02 1.364e+02 1.473e+02 1.883e+02, threshold=2.728e+02, percent-clipped=0.0 2024-09-25 03:59:13,239 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=657986.0, ans=0.0 2024-09-25 03:59:15,154 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=657986.0, ans=0.2 2024-09-25 03:59:16,788 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=657986.0, ans=0.0 2024-09-25 03:59:21,269 INFO [train.py:1198] (2/4) Epoch 37, batch 750, loss[loss=0.1591, ctc_loss=0.09894, cr_loss=0.3009, over 17157.00 frames. ], tot_loss[loss=0.193, ctc_loss=0.1248, cr_loss=0.3407, over 3276096.58 frames. ], batch size: 41, lr: 3.23e-03, grad_scale: 16.0 2024-09-25 03:59:33,549 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.61 vs. limit=6.0 2024-09-25 03:59:42,300 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=658079.3333333334, ans=0.2 2024-09-25 03:59:43,988 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=658079.3333333334, ans=0.125 2024-09-25 03:59:49,031 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.94 vs. limit=6.0 2024-09-25 04:00:15,106 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=658172.6666666666, ans=0.0 2024-09-25 04:00:19,912 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=658172.6666666666, ans=0.1 2024-09-25 04:00:24,836 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=658172.6666666666, ans=0.1 2024-09-25 04:00:26,695 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=10.56 vs. limit=22.5 2024-09-25 04:00:31,598 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.40 vs. limit=10.0 2024-09-25 04:00:38,308 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.85 vs. limit=6.0 2024-09-25 04:00:43,389 INFO [train.py:1198] (2/4) Epoch 37, batch 800, loss[loss=0.1623, ctc_loss=0.1031, cr_loss=0.296, over 17035.00 frames. ], tot_loss[loss=0.1935, ctc_loss=0.1252, cr_loss=0.3414, over 3290485.23 frames. ], batch size: 39, lr: 3.23e-03, grad_scale: 32.0 2024-09-25 04:01:15,362 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=658359.3333333334, ans=0.035 2024-09-25 04:01:29,116 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=4.57 vs. limit=15.0 2024-09-25 04:01:32,890 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=658406.0, ans=0.2 2024-09-25 04:01:54,704 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.144e+02 1.298e+02 1.352e+02 1.475e+02 2.154e+02, threshold=2.705e+02, percent-clipped=0.0 2024-09-25 04:01:55,122 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=658452.6666666666, ans=0.2 2024-09-25 04:02:02,796 INFO [train.py:1198] (2/4) Epoch 37, batch 850, loss[loss=0.204, ctc_loss=0.13, cr_loss=0.3702, over 16883.00 frames. ], tot_loss[loss=0.1929, ctc_loss=0.1248, cr_loss=0.3407, over 3309949.54 frames. ], batch size: 58, lr: 3.22e-03, grad_scale: 32.0 2024-09-25 04:02:07,818 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=658499.3333333334, ans=0.1 2024-09-25 04:02:36,922 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=658592.6666666666, ans=0.125 2024-09-25 04:03:00,583 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=658639.3333333334, ans=0.1 2024-09-25 04:03:16,948 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=16.45 vs. limit=22.5 2024-09-25 04:03:28,264 INFO [train.py:1198] (2/4) Epoch 37, batch 900, loss[loss=0.1571, ctc_loss=0.1011, cr_loss=0.2804, over 17094.00 frames. ], tot_loss[loss=0.1918, ctc_loss=0.1239, cr_loss=0.3392, over 3321381.90 frames. ], batch size: 40, lr: 3.22e-03, grad_scale: 32.0 2024-09-25 04:03:36,689 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=658732.6666666666, ans=0.125 2024-09-25 04:03:48,675 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 04:03:53,465 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=658779.3333333334, ans=0.0 2024-09-25 04:03:55,878 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=13.73 vs. limit=15.0 2024-09-25 04:04:28,714 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=658872.6666666666, ans=0.1 2024-09-25 04:04:35,254 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=658919.3333333334, ans=0.2 2024-09-25 04:04:43,135 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.276e+02 1.332e+02 1.387e+02 1.934e+02, threshold=2.664e+02, percent-clipped=0.0 2024-09-25 04:04:51,041 INFO [train.py:1198] (2/4) Epoch 37, batch 950, loss[loss=0.2411, ctc_loss=0.1582, cr_loss=0.4145, over 17188.00 frames. ], tot_loss[loss=0.1925, ctc_loss=0.1243, cr_loss=0.341, over 3338485.36 frames. ], batch size: 55, lr: 3.22e-03, grad_scale: 32.0 2024-09-25 04:04:53,262 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.26 vs. limit=10.0 2024-09-25 04:05:14,670 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.80 vs. limit=15.0 2024-09-25 04:05:15,918 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=659012.6666666666, ans=0.1 2024-09-25 04:05:37,144 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=6.28 vs. limit=15.0 2024-09-25 04:05:49,278 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=659106.0, ans=0.1 2024-09-25 04:06:11,832 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=659199.3333333334, ans=0.1 2024-09-25 04:06:13,085 INFO [train.py:1198] (2/4) Epoch 37, batch 1000, loss[loss=0.203, ctc_loss=0.131, cr_loss=0.36, over 17292.00 frames. ], tot_loss[loss=0.1926, ctc_loss=0.1244, cr_loss=0.341, over 3342154.93 frames. ], batch size: 49, lr: 3.22e-03, grad_scale: 32.0 2024-09-25 04:06:27,472 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=659246.0, ans=0.1 2024-09-25 04:06:32,772 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=4.81 vs. limit=15.0 2024-09-25 04:06:43,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=659292.6666666666, ans=0.125 2024-09-25 04:06:45,242 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=659292.6666666666, ans=0.0 2024-09-25 04:07:25,165 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.121e+02 1.283e+02 1.388e+02 1.495e+02 1.963e+02, threshold=2.776e+02, percent-clipped=0.0 2024-09-25 04:07:33,125 INFO [train.py:1198] (2/4) Epoch 37, batch 1050, loss[loss=0.1775, ctc_loss=0.1115, cr_loss=0.3302, over 17028.00 frames. ], tot_loss[loss=0.1934, ctc_loss=0.1248, cr_loss=0.3428, over 3349181.41 frames. ], batch size: 44, lr: 3.22e-03, grad_scale: 32.0 2024-09-25 04:07:44,549 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=659432.6666666666, ans=0.125 2024-09-25 04:08:28,355 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.52 vs. limit=15.0 2024-09-25 04:08:55,826 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer_ff2.min_abs, batch_count=659619.3333333334, ans=0.1 2024-09-25 04:09:00,280 INFO [train.py:1198] (2/4) Epoch 37, batch 1100, loss[loss=0.1524, ctc_loss=0.09773, cr_loss=0.2735, over 17274.00 frames. ], tot_loss[loss=0.1934, ctc_loss=0.1249, cr_loss=0.3425, over 3351387.32 frames. ], batch size: 42, lr: 3.22e-03, grad_scale: 32.0 2024-09-25 04:09:40,430 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=659759.3333333334, ans=0.07 2024-09-25 04:09:45,159 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=659759.3333333334, ans=0.125 2024-09-25 04:09:48,283 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=659806.0, ans=0.125 2024-09-25 04:10:14,845 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.153e+02 1.287e+02 1.378e+02 1.545e+02 2.459e+02, threshold=2.756e+02, percent-clipped=0.0 2024-09-25 04:10:20,559 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.33 vs. limit=15.0 2024-09-25 04:10:22,844 INFO [train.py:1198] (2/4) Epoch 37, batch 1150, loss[loss=0.177, ctc_loss=0.1125, cr_loss=0.3224, over 17188.00 frames. ], tot_loss[loss=0.194, ctc_loss=0.1254, cr_loss=0.3432, over 3339363.48 frames. ], batch size: 41, lr: 3.22e-03, grad_scale: 32.0 2024-09-25 04:11:35,646 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=660086.0, ans=0.0 2024-09-25 04:11:43,093 INFO [train.py:1198] (2/4) Epoch 37, batch 1200, loss[loss=0.2056, ctc_loss=0.1325, cr_loss=0.3657, over 17293.00 frames. ], tot_loss[loss=0.1936, ctc_loss=0.1251, cr_loss=0.3424, over 3348947.41 frames. ], batch size: 51, lr: 3.22e-03, grad_scale: 32.0 2024-09-25 04:11:44,306 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=13.11 vs. limit=22.5 2024-09-25 04:11:50,462 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.78 vs. limit=15.0 2024-09-25 04:12:48,954 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=660319.3333333334, ans=0.2 2024-09-25 04:12:57,285 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.131e+02 1.254e+02 1.340e+02 1.432e+02 2.120e+02, threshold=2.680e+02, percent-clipped=0.0 2024-09-25 04:13:05,188 INFO [train.py:1198] (2/4) Epoch 37, batch 1250, loss[loss=0.1575, ctc_loss=0.09952, cr_loss=0.2899, over 17028.00 frames. ], tot_loss[loss=0.1932, ctc_loss=0.1248, cr_loss=0.342, over 3356217.59 frames. ], batch size: 39, lr: 3.22e-03, grad_scale: 32.0 2024-09-25 04:13:18,043 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=660366.0, ans=0.2 2024-09-25 04:13:18,153 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.min_positive, batch_count=660366.0, ans=0.025 2024-09-25 04:13:30,094 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=660412.6666666666, ans=0.125 2024-09-25 04:14:05,485 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2.whitening_limit, batch_count=660506.0, ans=15.0 2024-09-25 04:14:19,271 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=660552.6666666666, ans=0.0 2024-09-25 04:14:30,158 INFO [train.py:1198] (2/4) Epoch 37, batch 1300, loss[loss=0.1569, ctc_loss=0.09974, cr_loss=0.286, over 16356.00 frames. ], tot_loss[loss=0.1945, ctc_loss=0.1259, cr_loss=0.3431, over 3336035.55 frames. ], batch size: 36, lr: 3.22e-03, grad_scale: 32.0 2024-09-25 04:14:34,500 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=6.05 vs. limit=15.0 2024-09-25 04:14:38,876 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=5.68 vs. limit=12.0 2024-09-25 04:14:45,765 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.75 vs. limit=6.0 2024-09-25 04:15:03,171 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=660692.6666666666, ans=0.125 2024-09-25 04:15:11,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=660692.6666666666, ans=0.2 2024-09-25 04:15:25,509 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=660739.3333333334, ans=0.125 2024-09-25 04:15:44,613 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.056e+02 1.301e+02 1.386e+02 1.515e+02 1.831e+02, threshold=2.772e+02, percent-clipped=0.0 2024-09-25 04:15:45,037 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=660786.0, ans=0.125 2024-09-25 04:15:52,775 INFO [train.py:1198] (2/4) Epoch 37, batch 1350, loss[loss=0.1907, ctc_loss=0.118, cr_loss=0.3633, over 17004.00 frames. ], tot_loss[loss=0.1938, ctc_loss=0.1254, cr_loss=0.3424, over 3342616.13 frames. ], batch size: 44, lr: 3.22e-03, grad_scale: 32.0 2024-09-25 04:16:07,468 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=660879.3333333334, ans=0.2 2024-09-25 04:16:09,916 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.81 vs. limit=6.0 2024-09-25 04:16:33,285 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=660926.0, ans=0.125 2024-09-25 04:16:33,328 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=660926.0, ans=0.2 2024-09-25 04:17:06,967 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=661019.3333333334, ans=0.125 2024-09-25 04:17:11,676 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=661066.0, ans=0.1 2024-09-25 04:17:13,042 INFO [train.py:1198] (2/4) Epoch 37, batch 1400, loss[loss=0.2187, ctc_loss=0.1424, cr_loss=0.3816, over 17241.00 frames. ], tot_loss[loss=0.1936, ctc_loss=0.1252, cr_loss=0.3423, over 3354931.62 frames. ], batch size: 55, lr: 3.22e-03, grad_scale: 16.0 2024-09-25 04:18:02,974 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=7.38 vs. limit=8.0 2024-09-25 04:18:09,931 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=661206.0, ans=10.0 2024-09-25 04:18:31,385 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.099e+02 1.311e+02 1.398e+02 1.519e+02 2.404e+02, threshold=2.797e+02, percent-clipped=0.0 2024-09-25 04:18:40,188 INFO [train.py:1198] (2/4) Epoch 37, batch 1450, loss[loss=0.1889, ctc_loss=0.1249, cr_loss=0.3201, over 17069.00 frames. ], tot_loss[loss=0.1947, ctc_loss=0.126, cr_loss=0.3433, over 3340958.02 frames. ], batch size: 46, lr: 3.22e-03, grad_scale: 16.0 2024-09-25 04:18:42,133 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=661299.3333333334, ans=0.1 2024-09-25 04:18:48,318 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=661299.3333333334, ans=0.1 2024-09-25 04:18:49,030 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.35 vs. limit=15.0 2024-09-25 04:18:49,884 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.min_abs, batch_count=661299.3333333334, ans=0.5 2024-09-25 04:18:53,289 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=661299.3333333334, ans=0.2 2024-09-25 04:18:54,733 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=661346.0, ans=0.1 2024-09-25 04:18:58,156 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=661346.0, ans=0.125 2024-09-25 04:19:02,996 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=661346.0, ans=0.0 2024-09-25 04:19:14,386 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=661392.6666666666, ans=0.125 2024-09-25 04:19:27,226 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=661439.3333333334, ans=0.0 2024-09-25 04:19:43,256 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=661486.0, ans=0.125 2024-09-25 04:19:52,671 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=661486.0, ans=0.125 2024-09-25 04:20:02,950 INFO [train.py:1198] (2/4) Epoch 37, batch 1500, loss[loss=0.1737, ctc_loss=0.111, cr_loss=0.3136, over 17103.00 frames. ], tot_loss[loss=0.1942, ctc_loss=0.1256, cr_loss=0.3429, over 3351590.11 frames. ], batch size: 43, lr: 3.22e-03, grad_scale: 16.0 2024-09-25 04:20:17,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.max_abs, batch_count=661579.3333333334, ans=10.0 2024-09-25 04:20:19,035 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=661579.3333333334, ans=0.0 2024-09-25 04:20:24,097 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.35 vs. limit=22.5 2024-09-25 04:20:41,216 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=661626.0, ans=0.1 2024-09-25 04:21:06,715 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=661719.3333333334, ans=0.1 2024-09-25 04:21:08,296 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=661719.3333333334, ans=0.04949747468305833 2024-09-25 04:21:16,231 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.237e+02 1.312e+02 1.425e+02 1.916e+02, threshold=2.625e+02, percent-clipped=0.0 2024-09-25 04:21:19,657 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=661719.3333333334, ans=0.125 2024-09-25 04:21:22,594 INFO [train.py:1198] (2/4) Epoch 37, batch 1550, loss[loss=0.2084, ctc_loss=0.137, cr_loss=0.3572, over 16757.00 frames. ], tot_loss[loss=0.1956, ctc_loss=0.1266, cr_loss=0.3449, over 3350534.51 frames. ], batch size: 61, lr: 3.22e-03, grad_scale: 16.0 2024-09-25 04:21:23,322 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.74 vs. limit=22.5 2024-09-25 04:21:30,759 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=661766.0, ans=0.0 2024-09-25 04:21:42,145 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=661812.6666666666, ans=0.125 2024-09-25 04:21:48,678 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=661812.6666666666, ans=0.0 2024-09-25 04:22:00,916 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.33 vs. limit=15.0 2024-09-25 04:22:22,844 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=661906.0, ans=0.1 2024-09-25 04:22:24,792 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=5.34 vs. limit=10.0 2024-09-25 04:22:29,133 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=661952.6666666666, ans=0.025 2024-09-25 04:22:33,819 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=661952.6666666666, ans=0.0 2024-09-25 04:22:45,505 INFO [train.py:1198] (2/4) Epoch 37, batch 1600, loss[loss=0.1615, ctc_loss=0.1031, cr_loss=0.2921, over 16996.00 frames. ], tot_loss[loss=0.1947, ctc_loss=0.1259, cr_loss=0.344, over 3350788.17 frames. ], batch size: 44, lr: 3.22e-03, grad_scale: 32.0 2024-09-25 04:23:00,573 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=5.42 vs. limit=15.0 2024-09-25 04:23:31,679 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=8.04 vs. limit=15.0 2024-09-25 04:23:50,028 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.min_positive, batch_count=662139.3333333334, ans=0.025 2024-09-25 04:24:04,105 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.112e+02 1.273e+02 1.358e+02 1.458e+02 2.530e+02, threshold=2.716e+02, percent-clipped=0.0 2024-09-25 04:24:10,512 INFO [train.py:1198] (2/4) Epoch 37, batch 1650, loss[loss=0.2052, ctc_loss=0.1324, cr_loss=0.3637, over 17034.00 frames. ], tot_loss[loss=0.1951, ctc_loss=0.1263, cr_loss=0.3443, over 3344103.44 frames. ], batch size: 51, lr: 3.22e-03, grad_scale: 32.0 2024-09-25 04:24:36,350 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=662279.3333333334, ans=0.025 2024-09-25 04:24:42,566 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=662326.0, ans=0.0 2024-09-25 04:25:23,699 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=662419.3333333334, ans=0.1 2024-09-25 04:25:28,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=662419.3333333334, ans=0.025 2024-09-25 04:25:32,851 INFO [train.py:1198] (2/4) Epoch 37, batch 1700, loss[loss=0.1759, ctc_loss=0.1121, cr_loss=0.3192, over 17093.00 frames. ], tot_loss[loss=0.1939, ctc_loss=0.1254, cr_loss=0.3424, over 3333781.07 frames. ], batch size: 40, lr: 3.22e-03, grad_scale: 32.0 2024-09-25 04:25:56,838 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=662512.6666666666, ans=0.125 2024-09-25 04:26:06,221 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=662559.3333333334, ans=0.125 2024-09-25 04:26:09,758 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.78 vs. limit=15.0 2024-09-25 04:26:14,014 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=662559.3333333334, ans=0.95 2024-09-25 04:26:14,452 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=15.44 vs. limit=15.0 2024-09-25 04:26:15,673 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=662559.3333333334, ans=0.125 2024-09-25 04:26:35,520 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=5.52 vs. limit=15.0 2024-09-25 04:26:38,424 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=10.37 vs. limit=22.5 2024-09-25 04:26:43,070 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.min_abs, batch_count=662652.6666666666, ans=0.5 2024-09-25 04:26:45,874 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.286e+02 1.377e+02 1.479e+02 2.270e+02, threshold=2.754e+02, percent-clipped=0.0 2024-09-25 04:26:52,404 INFO [train.py:1198] (2/4) Epoch 37, batch 1750, loss[loss=0.1897, ctc_loss=0.1214, cr_loss=0.341, over 17043.00 frames. ], tot_loss[loss=0.1941, ctc_loss=0.1256, cr_loss=0.343, over 3341996.28 frames. ], batch size: 44, lr: 3.21e-03, grad_scale: 32.0 2024-09-25 04:26:54,825 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.37 vs. limit=15.0 2024-09-25 04:27:16,615 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=662746.0, ans=0.1 2024-09-25 04:27:31,170 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=662792.6666666666, ans=0.5 2024-09-25 04:28:08,263 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 04:28:14,219 INFO [train.py:1198] (2/4) Epoch 37, batch 1800, loss[loss=0.2078, ctc_loss=0.1383, cr_loss=0.3477, over 14921.00 frames. ], tot_loss[loss=0.1938, ctc_loss=0.1254, cr_loss=0.3424, over 3334960.47 frames. ], batch size: 89, lr: 3.21e-03, grad_scale: 32.0 2024-09-25 04:28:27,269 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=14.34 vs. limit=22.5 2024-09-25 04:28:34,758 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=662979.3333333334, ans=0.125 2024-09-25 04:28:36,661 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=7.94 vs. limit=15.0 2024-09-25 04:28:56,828 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.min_positive, batch_count=663026.0, ans=0.025 2024-09-25 04:29:13,282 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.70 vs. limit=6.0 2024-09-25 04:29:30,047 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.280e+02 1.352e+02 1.451e+02 1.795e+02, threshold=2.704e+02, percent-clipped=0.0 2024-09-25 04:29:39,048 INFO [train.py:1198] (2/4) Epoch 37, batch 1850, loss[loss=0.1584, ctc_loss=0.1003, cr_loss=0.2906, over 17063.00 frames. ], tot_loss[loss=0.1943, ctc_loss=0.1257, cr_loss=0.3429, over 3321712.44 frames. ], batch size: 46, lr: 3.21e-03, grad_scale: 32.0 2024-09-25 04:30:14,228 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=663259.3333333334, ans=0.0 2024-09-25 04:30:32,069 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.85 vs. limit=15.0 2024-09-25 04:30:33,256 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=663306.0, ans=0.0 2024-09-25 04:30:58,766 INFO [train.py:1198] (2/4) Epoch 37, batch 1900, loss[loss=0.2254, ctc_loss=0.1478, cr_loss=0.388, over 16890.00 frames. ], tot_loss[loss=0.1948, ctc_loss=0.1261, cr_loss=0.3436, over 3322336.72 frames. ], batch size: 58, lr: 3.21e-03, grad_scale: 32.0 2024-09-25 04:31:16,786 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=663446.0, ans=0.0 2024-09-25 04:31:52,467 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.09 vs. limit=10.0 2024-09-25 04:32:03,423 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=663586.0, ans=0.125 2024-09-25 04:32:12,966 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.076e+02 1.273e+02 1.351e+02 1.492e+02 1.881e+02, threshold=2.703e+02, percent-clipped=0.0 2024-09-25 04:32:18,239 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.30 vs. limit=15.0 2024-09-25 04:32:19,249 INFO [train.py:1198] (2/4) Epoch 37, batch 1950, loss[loss=0.2002, ctc_loss=0.1308, cr_loss=0.3469, over 17052.00 frames. ], tot_loss[loss=0.1938, ctc_loss=0.1253, cr_loss=0.3422, over 3333997.41 frames. ], batch size: 46, lr: 3.21e-03, grad_scale: 32.0 2024-09-25 04:32:35,215 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.74 vs. limit=6.0 2024-09-25 04:33:46,763 INFO [train.py:1198] (2/4) Epoch 37, batch 2000, loss[loss=0.2109, ctc_loss=0.137, cr_loss=0.3696, over 15202.00 frames. ], tot_loss[loss=0.1937, ctc_loss=0.1252, cr_loss=0.3423, over 3344774.99 frames. ], batch size: 89, lr: 3.21e-03, grad_scale: 32.0 2024-09-25 04:33:59,971 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=663866.0, ans=0.125 2024-09-25 04:34:06,316 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=663912.6666666666, ans=0.125 2024-09-25 04:34:31,951 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=663959.3333333334, ans=0.0 2024-09-25 04:34:40,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.min_abs, batch_count=664006.0, ans=0.5 2024-09-25 04:35:02,488 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.153e+02 1.268e+02 1.366e+02 1.468e+02 1.745e+02, threshold=2.732e+02, percent-clipped=0.0 2024-09-25 04:35:08,884 INFO [train.py:1198] (2/4) Epoch 37, batch 2050, loss[loss=0.1788, ctc_loss=0.1142, cr_loss=0.3229, over 17160.00 frames. ], tot_loss[loss=0.1936, ctc_loss=0.1252, cr_loss=0.3423, over 3342556.42 frames. ], batch size: 41, lr: 3.21e-03, grad_scale: 32.0 2024-09-25 04:35:44,107 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=664192.6666666666, ans=0.2 2024-09-25 04:36:14,432 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=664286.0, ans=0.1 2024-09-25 04:36:16,194 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=664286.0, ans=0.125 2024-09-25 04:36:28,588 INFO [train.py:1198] (2/4) Epoch 37, batch 2100, loss[loss=0.234, ctc_loss=0.1575, cr_loss=0.3824, over 14831.00 frames. ], tot_loss[loss=0.194, ctc_loss=0.1253, cr_loss=0.3436, over 3351210.00 frames. ], batch size: 89, lr: 3.21e-03, grad_scale: 16.0 2024-09-25 04:36:34,986 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=664332.6666666666, ans=0.125 2024-09-25 04:36:44,498 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.max_abs, batch_count=664379.3333333334, ans=10.0 2024-09-25 04:36:50,780 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=664379.3333333334, ans=0.0 2024-09-25 04:37:01,892 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=664426.0, ans=0.1 2024-09-25 04:37:46,466 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.153e+02 1.254e+02 1.343e+02 1.449e+02 1.904e+02, threshold=2.685e+02, percent-clipped=0.0 2024-09-25 04:37:50,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=664566.0, ans=0.1 2024-09-25 04:37:51,356 INFO [train.py:1198] (2/4) Epoch 37, batch 2150, loss[loss=0.197, ctc_loss=0.1262, cr_loss=0.3541, over 17144.00 frames. ], tot_loss[loss=0.1936, ctc_loss=0.125, cr_loss=0.3427, over 3355733.45 frames. ], batch size: 48, lr: 3.21e-03, grad_scale: 16.0 2024-09-25 04:38:50,758 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=664706.0, ans=0.125 2024-09-25 04:38:53,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=664706.0, ans=0.04949747468305833 2024-09-25 04:39:00,282 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=664752.6666666666, ans=0.125 2024-09-25 04:39:13,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=664752.6666666666, ans=0.125 2024-09-25 04:39:15,888 INFO [train.py:1198] (2/4) Epoch 37, batch 2200, loss[loss=0.2101, ctc_loss=0.1372, cr_loss=0.3644, over 17139.00 frames. ], tot_loss[loss=0.1941, ctc_loss=0.1254, cr_loss=0.3438, over 3351140.84 frames. ], batch size: 48, lr: 3.21e-03, grad_scale: 16.0 2024-09-25 04:39:21,170 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=664799.3333333334, ans=0.025 2024-09-25 04:39:55,901 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=664892.6666666666, ans=0.125 2024-09-25 04:40:07,117 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=664939.3333333334, ans=0.125 2024-09-25 04:40:34,121 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.115e+02 1.296e+02 1.362e+02 1.438e+02 1.964e+02, threshold=2.724e+02, percent-clipped=0.0 2024-09-25 04:40:38,875 INFO [train.py:1198] (2/4) Epoch 37, batch 2250, loss[loss=0.1612, ctc_loss=0.1006, cr_loss=0.3028, over 17049.00 frames. ], tot_loss[loss=0.1944, ctc_loss=0.1256, cr_loss=0.3439, over 3355169.93 frames. ], batch size: 39, lr: 3.21e-03, grad_scale: 16.0 2024-09-25 04:40:57,261 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.30 vs. limit=10.0 2024-09-25 04:41:14,410 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=665126.0, ans=0.1 2024-09-25 04:41:24,225 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.28 vs. limit=15.0 2024-09-25 04:41:27,834 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.06 vs. limit=15.0 2024-09-25 04:41:45,182 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=6.58 vs. limit=15.0 2024-09-25 04:41:58,859 INFO [train.py:1198] (2/4) Epoch 37, batch 2300, loss[loss=0.2116, ctc_loss=0.1373, cr_loss=0.3718, over 15829.00 frames. ], tot_loss[loss=0.1942, ctc_loss=0.1255, cr_loss=0.3433, over 3360684.78 frames. ], batch size: 74, lr: 3.21e-03, grad_scale: 16.0 2024-09-25 04:41:59,073 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=665266.0, ans=0.125 2024-09-25 04:42:10,223 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=665266.0, ans=0.0 2024-09-25 04:42:32,049 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=665359.3333333334, ans=0.1 2024-09-25 04:42:43,160 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=665359.3333333334, ans=0.035 2024-09-25 04:43:07,588 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=665406.0, ans=0.1 2024-09-25 04:43:21,780 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.164e+02 1.297e+02 1.364e+02 1.435e+02 2.216e+02, threshold=2.727e+02, percent-clipped=0.0 2024-09-25 04:43:25,197 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=665499.3333333334, ans=0.1 2024-09-25 04:43:26,539 INFO [train.py:1198] (2/4) Epoch 37, batch 2350, loss[loss=0.2133, ctc_loss=0.1378, cr_loss=0.3775, over 17299.00 frames. ], tot_loss[loss=0.1941, ctc_loss=0.1255, cr_loss=0.3432, over 3365103.32 frames. ], batch size: 51, lr: 3.21e-03, grad_scale: 16.0 2024-09-25 04:43:33,269 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.min_abs, batch_count=665499.3333333334, ans=0.5 2024-09-25 04:43:47,792 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=665546.0, ans=0.0 2024-09-25 04:44:19,794 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=665639.3333333334, ans=0.0 2024-09-25 04:44:48,923 INFO [train.py:1198] (2/4) Epoch 37, batch 2400, loss[loss=0.1597, ctc_loss=0.1018, cr_loss=0.2893, over 17144.00 frames. ], tot_loss[loss=0.1921, ctc_loss=0.1241, cr_loss=0.3403, over 3370286.74 frames. ], batch size: 40, lr: 3.21e-03, grad_scale: 32.0 2024-09-25 04:44:51,241 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=4.49 vs. limit=12.0 2024-09-25 04:45:00,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=665732.6666666666, ans=0.125 2024-09-25 04:45:09,760 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=665779.3333333334, ans=0.125 2024-09-25 04:45:09,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=665779.3333333334, ans=0.0 2024-09-25 04:45:09,814 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=665779.3333333334, ans=0.0 2024-09-25 04:45:09,877 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=665779.3333333334, ans=0.025 2024-09-25 04:45:14,659 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=665779.3333333334, ans=0.2 2024-09-25 04:45:32,243 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=665826.0, ans=0.125 2024-09-25 04:46:03,760 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.110e+02 1.277e+02 1.353e+02 1.451e+02 1.749e+02, threshold=2.707e+02, percent-clipped=0.0 2024-09-25 04:46:07,220 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=665966.0, ans=0.025 2024-09-25 04:46:08,610 INFO [train.py:1198] (2/4) Epoch 37, batch 2450, loss[loss=0.2026, ctc_loss=0.1332, cr_loss=0.3473, over 16136.00 frames. ], tot_loss[loss=0.1932, ctc_loss=0.1248, cr_loss=0.3418, over 3358567.72 frames. ], batch size: 74, lr: 3.21e-03, grad_scale: 32.0 2024-09-25 04:46:23,375 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=666012.6666666666, ans=0.125 2024-09-25 04:46:28,077 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=666012.6666666666, ans=0.125 2024-09-25 04:46:37,726 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=666012.6666666666, ans=0.0 2024-09-25 04:47:01,704 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=666106.0, ans=0.0 2024-09-25 04:47:04,829 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=666106.0, ans=0.125 2024-09-25 04:47:30,969 INFO [train.py:1198] (2/4) Epoch 37, batch 2500, loss[loss=0.1887, ctc_loss=0.1232, cr_loss=0.3276, over 17171.00 frames. ], tot_loss[loss=0.1934, ctc_loss=0.1249, cr_loss=0.3425, over 3358376.31 frames. ], batch size: 45, lr: 3.21e-03, grad_scale: 32.0 2024-09-25 04:47:37,756 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=666199.3333333334, ans=0.125 2024-09-25 04:47:38,470 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=8.27 vs. limit=15.0 2024-09-25 04:47:49,076 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=666246.0, ans=0.125 2024-09-25 04:48:03,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=666246.0, ans=0.125 2024-09-25 04:48:20,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=666292.6666666666, ans=0.125 2024-09-25 04:48:23,259 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=666339.3333333334, ans=0.0 2024-09-25 04:48:34,317 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=666339.3333333334, ans=0.0 2024-09-25 04:48:43,818 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=666386.0, ans=0.0 2024-09-25 04:48:53,436 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.167e+02 1.305e+02 1.391e+02 1.498e+02 2.183e+02, threshold=2.782e+02, percent-clipped=0.0 2024-09-25 04:48:56,603 INFO [train.py:1198] (2/4) Epoch 37, batch 2550, loss[loss=0.202, ctc_loss=0.1375, cr_loss=0.3229, over 11632.00 frames. ], tot_loss[loss=0.1938, ctc_loss=0.1251, cr_loss=0.3435, over 3349221.65 frames. ], batch size: 125, lr: 3.21e-03, grad_scale: 16.0 2024-09-25 04:49:19,609 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=666479.3333333334, ans=0.0 2024-09-25 04:49:36,578 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=666526.0, ans=0.05 2024-09-25 04:50:00,511 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=666572.6666666666, ans=0.025 2024-09-25 04:50:19,600 INFO [train.py:1198] (2/4) Epoch 37, batch 2600, loss[loss=0.2237, ctc_loss=0.146, cr_loss=0.3886, over 17235.00 frames. ], tot_loss[loss=0.1945, ctc_loss=0.1256, cr_loss=0.345, over 3353930.08 frames. ], batch size: 55, lr: 3.21e-03, grad_scale: 16.0 2024-09-25 04:51:17,128 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=666806.0, ans=0.0 2024-09-25 04:51:20,326 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=666806.0, ans=0.0 2024-09-25 04:51:23,546 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=666852.6666666666, ans=0.1 2024-09-25 04:51:36,208 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.093e+02 1.272e+02 1.334e+02 1.506e+02 4.167e+02, threshold=2.667e+02, percent-clipped=1.0 2024-09-25 04:51:39,400 INFO [train.py:1198] (2/4) Epoch 37, batch 2650, loss[loss=0.1879, ctc_loss=0.1237, cr_loss=0.3211, over 17154.00 frames. ], tot_loss[loss=0.1943, ctc_loss=0.1254, cr_loss=0.3442, over 3347933.64 frames. ], batch size: 48, lr: 3.20e-03, grad_scale: 16.0 2024-09-25 04:51:41,230 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=666899.3333333334, ans=0.025 2024-09-25 04:51:57,395 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.min_abs, batch_count=666946.0, ans=0.5 2024-09-25 04:52:07,896 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=666946.0, ans=0.0 2024-09-25 04:52:19,447 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.max_abs, batch_count=666992.6666666666, ans=10.0 2024-09-25 04:52:19,494 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.min_positive, batch_count=666992.6666666666, ans=0.05 2024-09-25 04:52:21,097 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=666992.6666666666, ans=0.125 2024-09-25 04:52:24,076 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=666992.6666666666, ans=0.1 2024-09-25 04:52:27,421 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=666992.6666666666, ans=0.0 2024-09-25 04:53:07,638 INFO [train.py:1198] (2/4) Epoch 37, batch 2700, loss[loss=0.167, ctc_loss=0.1064, cr_loss=0.3033, over 17202.00 frames. ], tot_loss[loss=0.1931, ctc_loss=0.1246, cr_loss=0.3424, over 3354710.78 frames. ], batch size: 41, lr: 3.20e-03, grad_scale: 16.0 2024-09-25 04:53:22,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten.whitening_limit, batch_count=667179.3333333334, ans=15.0 2024-09-25 04:53:25,810 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.06 vs. limit=15.0 2024-09-25 04:53:35,123 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=667179.3333333334, ans=0.0 2024-09-25 04:54:02,184 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=667272.6666666666, ans=0.125 2024-09-25 04:54:03,026 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=6.27 vs. limit=15.0 2024-09-25 04:54:11,929 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=667319.3333333334, ans=0.125 2024-09-25 04:54:27,063 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.290e+02 1.345e+02 1.477e+02 2.693e+02, threshold=2.691e+02, percent-clipped=1.0 2024-09-25 04:54:30,303 INFO [train.py:1198] (2/4) Epoch 37, batch 2750, loss[loss=0.2238, ctc_loss=0.1439, cr_loss=0.3998, over 17020.00 frames. ], tot_loss[loss=0.194, ctc_loss=0.1253, cr_loss=0.3438, over 3350498.92 frames. ], batch size: 52, lr: 3.20e-03, grad_scale: 16.0 2024-09-25 04:54:49,685 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=667412.6666666666, ans=0.125 2024-09-25 04:55:42,638 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=667552.6666666666, ans=0.1 2024-09-25 04:55:50,321 INFO [train.py:1198] (2/4) Epoch 37, batch 2800, loss[loss=0.1747, ctc_loss=0.114, cr_loss=0.3035, over 17291.00 frames. ], tot_loss[loss=0.193, ctc_loss=0.1246, cr_loss=0.3421, over 3359565.24 frames. ], batch size: 46, lr: 3.20e-03, grad_scale: 32.0 2024-09-25 04:56:05,265 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 04:56:13,508 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=11.01 vs. limit=22.5 2024-09-25 04:56:14,747 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=667646.0, ans=0.125 2024-09-25 04:56:27,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=667692.6666666666, ans=0.125 2024-09-25 04:56:27,977 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.min_positive, batch_count=667692.6666666666, ans=0.05 2024-09-25 04:56:36,068 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=667692.6666666666, ans=0.0 2024-09-25 04:56:40,826 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=667739.3333333334, ans=0.0 2024-09-25 04:56:45,779 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=9.26 vs. limit=15.0 2024-09-25 04:57:10,009 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.075e+02 1.310e+02 1.386e+02 1.476e+02 1.796e+02, threshold=2.772e+02, percent-clipped=0.0 2024-09-25 04:57:13,305 INFO [train.py:1198] (2/4) Epoch 37, batch 2850, loss[loss=0.1664, ctc_loss=0.1058, cr_loss=0.3032, over 17258.00 frames. ], tot_loss[loss=0.1918, ctc_loss=0.1237, cr_loss=0.3405, over 3357226.24 frames. ], batch size: 44, lr: 3.20e-03, grad_scale: 32.0 2024-09-25 04:57:20,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=667832.6666666666, ans=0.125 2024-09-25 04:57:31,572 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.29 vs. limit=15.0 2024-09-25 04:58:19,420 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=667972.6666666666, ans=0.125 2024-09-25 04:58:32,030 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=668019.3333333334, ans=0.125 2024-09-25 04:58:32,069 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.min_positive, batch_count=668019.3333333334, ans=0.025 2024-09-25 04:58:33,691 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=668019.3333333334, ans=0.0 2024-09-25 04:58:38,162 INFO [train.py:1198] (2/4) Epoch 37, batch 2900, loss[loss=0.1941, ctc_loss=0.1264, cr_loss=0.3384, over 17221.00 frames. ], tot_loss[loss=0.1909, ctc_loss=0.1231, cr_loss=0.3391, over 3351608.49 frames. ], batch size: 50, lr: 3.20e-03, grad_scale: 32.0 2024-09-25 04:58:47,850 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=668066.0, ans=0.125 2024-09-25 04:59:07,230 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=668112.6666666666, ans=0.125 2024-09-25 04:59:08,699 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=668159.3333333334, ans=0.0 2024-09-25 04:59:33,697 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=668206.0, ans=0.025 2024-09-25 04:59:43,423 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=668252.6666666666, ans=0.0 2024-09-25 04:59:57,885 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.277e+02 1.411e+02 1.557e+02 2.099e+02, threshold=2.822e+02, percent-clipped=0.0 2024-09-25 05:00:01,063 INFO [train.py:1198] (2/4) Epoch 37, batch 2950, loss[loss=0.1836, ctc_loss=0.1166, cr_loss=0.3348, over 17078.00 frames. ], tot_loss[loss=0.1911, ctc_loss=0.1233, cr_loss=0.3391, over 3352589.18 frames. ], batch size: 43, lr: 3.20e-03, grad_scale: 32.0 2024-09-25 05:00:20,432 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=668346.0, ans=0.125 2024-09-25 05:00:25,305 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=668346.0, ans=0.0 2024-09-25 05:00:30,179 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=668346.0, ans=0.0 2024-09-25 05:00:36,424 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=668392.6666666666, ans=0.0 2024-09-25 05:00:43,011 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=668392.6666666666, ans=0.125 2024-09-25 05:01:17,908 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=8.16 vs. limit=15.0 2024-09-25 05:01:20,183 INFO [train.py:1198] (2/4) Epoch 37, batch 3000, loss[loss=0.2157, ctc_loss=0.1412, cr_loss=0.3728, over 17139.00 frames. ], tot_loss[loss=0.1921, ctc_loss=0.124, cr_loss=0.3403, over 3338831.06 frames. ], batch size: 48, lr: 3.20e-03, grad_scale: 16.0 2024-09-25 05:01:20,184 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 05:01:35,770 INFO [train.py:1230] (2/4) Epoch 37, validation: loss=0.03526, ctc_loss=0.03526, cr_loss=1.039e-14, over 944034.00 frames. 2024-09-25 05:01:35,771 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 05:01:42,335 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=668532.6666666666, ans=0.125 2024-09-25 05:01:46,954 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=668532.6666666666, ans=0.0 2024-09-25 05:02:05,723 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=668626.0, ans=0.125 2024-09-25 05:02:13,511 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=668626.0, ans=0.125 2024-09-25 05:02:13,632 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=668626.0, ans=0.04949747468305833 2024-09-25 05:02:47,940 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=668719.3333333334, ans=0.125 2024-09-25 05:02:51,382 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.21 vs. limit=15.0 2024-09-25 05:02:54,649 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.102e+02 1.290e+02 1.381e+02 1.487e+02 2.036e+02, threshold=2.762e+02, percent-clipped=0.0 2024-09-25 05:02:56,256 INFO [train.py:1198] (2/4) Epoch 37, batch 3050, loss[loss=0.1819, ctc_loss=0.1183, cr_loss=0.3178, over 17263.00 frames. ], tot_loss[loss=0.1933, ctc_loss=0.1249, cr_loss=0.3421, over 3332854.27 frames. ], batch size: 44, lr: 3.20e-03, grad_scale: 16.0 2024-09-25 05:03:29,265 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=668859.3333333334, ans=0.0 2024-09-25 05:03:39,419 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=5.61 vs. limit=15.0 2024-09-25 05:03:43,390 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 05:04:14,277 INFO [train.py:1198] (2/4) Epoch 37, batch 3100, loss[loss=0.1682, ctc_loss=0.1084, cr_loss=0.2986, over 17226.00 frames. ], tot_loss[loss=0.1941, ctc_loss=0.1255, cr_loss=0.3431, over 3334966.16 frames. ], batch size: 47, lr: 3.20e-03, grad_scale: 16.0 2024-09-25 05:04:26,287 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=8.29 vs. limit=15.0 2024-09-25 05:04:46,017 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=669092.6666666666, ans=0.125 2024-09-25 05:04:52,007 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=669092.6666666666, ans=0.125 2024-09-25 05:05:03,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=669139.3333333334, ans=0.05 2024-09-25 05:05:32,884 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=669186.0, ans=0.125 2024-09-25 05:05:35,823 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.145e+02 1.257e+02 1.346e+02 1.447e+02 1.997e+02, threshold=2.692e+02, percent-clipped=0.0 2024-09-25 05:05:36,135 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=669232.6666666666, ans=0.125 2024-09-25 05:05:37,417 INFO [train.py:1198] (2/4) Epoch 37, batch 3150, loss[loss=0.1556, ctc_loss=0.09918, cr_loss=0.2822, over 17319.00 frames. ], tot_loss[loss=0.1943, ctc_loss=0.1257, cr_loss=0.3428, over 3341784.79 frames. ], batch size: 46, lr: 3.20e-03, grad_scale: 16.0 2024-09-25 05:05:43,831 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=669232.6666666666, ans=0.0 2024-09-25 05:05:53,247 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=669279.3333333334, ans=0.0 2024-09-25 05:05:54,791 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=669279.3333333334, ans=0.1 2024-09-25 05:06:05,890 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=669279.3333333334, ans=0.0 2024-09-25 05:06:55,584 INFO [train.py:1198] (2/4) Epoch 37, batch 3200, loss[loss=0.1721, ctc_loss=0.1108, cr_loss=0.3063, over 17100.00 frames. ], tot_loss[loss=0.1927, ctc_loss=0.1246, cr_loss=0.3402, over 3344117.47 frames. ], batch size: 49, lr: 3.20e-03, grad_scale: 32.0 2024-09-25 05:06:58,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=669466.0, ans=0.1 2024-09-25 05:07:05,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=669466.0, ans=0.0 2024-09-25 05:07:20,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=669512.6666666666, ans=0.125 2024-09-25 05:07:26,966 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=669559.3333333334, ans=0.125 2024-09-25 05:07:53,528 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=669606.0, ans=0.125 2024-09-25 05:08:09,360 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=4.32 vs. limit=15.0 2024-09-25 05:08:15,376 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.135e+02 1.285e+02 1.355e+02 1.464e+02 2.821e+02, threshold=2.709e+02, percent-clipped=1.0 2024-09-25 05:08:15,400 INFO [train.py:1198] (2/4) Epoch 37, batch 3250, loss[loss=0.2116, ctc_loss=0.1383, cr_loss=0.3664, over 17023.00 frames. ], tot_loss[loss=0.1929, ctc_loss=0.1247, cr_loss=0.3411, over 3351595.31 frames. ], batch size: 52, lr: 3.20e-03, grad_scale: 16.0 2024-09-25 05:08:29,874 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=5.07 vs. limit=10.0 2024-09-25 05:08:34,807 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=4.85 vs. limit=12.0 2024-09-25 05:08:48,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=669792.6666666666, ans=0.2 2024-09-25 05:08:53,332 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=669792.6666666666, ans=0.0 2024-09-25 05:08:59,508 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=669792.6666666666, ans=0.0 2024-09-25 05:09:02,672 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 05:09:07,552 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=18.10 vs. limit=22.5 2024-09-25 05:09:14,901 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=669839.3333333334, ans=0.0 2024-09-25 05:09:33,486 INFO [train.py:1198] (2/4) Epoch 37, batch 3300, loss[loss=0.2226, ctc_loss=0.1426, cr_loss=0.4001, over 16618.00 frames. ], tot_loss[loss=0.1934, ctc_loss=0.1251, cr_loss=0.3418, over 3355644.44 frames. ], batch size: 66, lr: 3.20e-03, grad_scale: 16.0 2024-09-25 05:09:33,860 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=669932.6666666666, ans=0.0 2024-09-25 05:09:35,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=669932.6666666666, ans=0.0 2024-09-25 05:09:44,871 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=10.95 vs. limit=12.0 2024-09-25 05:09:57,891 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=12.44 vs. limit=15.0 2024-09-25 05:10:16,062 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=670026.0, ans=0.025 2024-09-25 05:10:22,431 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=670072.6666666666, ans=0.0 2024-09-25 05:10:51,977 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.300e+02 1.355e+02 1.430e+02 2.152e+02, threshold=2.710e+02, percent-clipped=0.0 2024-09-25 05:10:52,001 INFO [train.py:1198] (2/4) Epoch 37, batch 3350, loss[loss=0.1859, ctc_loss=0.1193, cr_loss=0.333, over 17177.00 frames. ], tot_loss[loss=0.1933, ctc_loss=0.125, cr_loss=0.3416, over 3362121.87 frames. ], batch size: 41, lr: 3.20e-03, grad_scale: 16.0 2024-09-25 05:10:58,358 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=670166.0, ans=0.0 2024-09-25 05:11:09,330 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=670212.6666666666, ans=0.0 2024-09-25 05:11:20,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=670212.6666666666, ans=0.0 2024-09-25 05:11:27,940 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=670259.3333333334, ans=0.125 2024-09-25 05:11:30,957 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=670259.3333333334, ans=0.07 2024-09-25 05:12:05,308 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=670352.6666666666, ans=0.125 2024-09-25 05:12:06,202 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.64 vs. limit=15.0 2024-09-25 05:12:09,881 INFO [train.py:1198] (2/4) Epoch 37, batch 3400, loss[loss=0.1643, ctc_loss=0.1046, cr_loss=0.2982, over 16683.00 frames. ], tot_loss[loss=0.1923, ctc_loss=0.1243, cr_loss=0.3398, over 3359546.19 frames. ], batch size: 37, lr: 3.20e-03, grad_scale: 16.0 2024-09-25 05:12:10,147 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=670399.3333333334, ans=0.0 2024-09-25 05:12:21,099 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=670399.3333333334, ans=0.2 2024-09-25 05:12:44,687 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=670492.6666666666, ans=0.0 2024-09-25 05:13:12,928 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=670586.0, ans=0.0 2024-09-25 05:13:20,818 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=670586.0, ans=0.125 2024-09-25 05:13:28,360 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.294e+02 1.379e+02 1.549e+02 2.225e+02, threshold=2.758e+02, percent-clipped=0.0 2024-09-25 05:13:28,384 INFO [train.py:1198] (2/4) Epoch 37, batch 3450, loss[loss=0.207, ctc_loss=0.1337, cr_loss=0.3662, over 17226.00 frames. ], tot_loss[loss=0.192, ctc_loss=0.1241, cr_loss=0.3398, over 3363544.80 frames. ], batch size: 55, lr: 3.20e-03, grad_scale: 16.0 2024-09-25 05:14:08,519 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=670726.0, ans=0.04949747468305833 2024-09-25 05:14:14,805 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=670726.0, ans=0.0 2024-09-25 05:14:49,368 INFO [train.py:1198] (2/4) Epoch 37, batch 3500, loss[loss=0.1934, ctc_loss=0.1255, cr_loss=0.3395, over 17239.00 frames. ], tot_loss[loss=0.1925, ctc_loss=0.1244, cr_loss=0.3407, over 3361997.88 frames. ], batch size: 47, lr: 3.20e-03, grad_scale: 16.0 2024-09-25 05:15:18,719 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=670912.6666666666, ans=0.125 2024-09-25 05:15:21,867 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=670912.6666666666, ans=0.125 2024-09-25 05:15:23,390 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=670959.3333333334, ans=0.0 2024-09-25 05:15:23,495 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=670959.3333333334, ans=0.1 2024-09-25 05:15:40,768 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=671006.0, ans=0.0 2024-09-25 05:16:09,646 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=671052.6666666666, ans=0.0 2024-09-25 05:16:12,281 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.083e+02 1.290e+02 1.375e+02 1.489e+02 2.034e+02, threshold=2.750e+02, percent-clipped=0.0 2024-09-25 05:16:12,305 INFO [train.py:1198] (2/4) Epoch 37, batch 3550, loss[loss=0.1872, ctc_loss=0.1213, cr_loss=0.3297, over 17083.00 frames. ], tot_loss[loss=0.1926, ctc_loss=0.1244, cr_loss=0.3408, over 3360201.02 frames. ], batch size: 49, lr: 3.19e-03, grad_scale: 16.0 2024-09-25 05:16:12,556 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=671099.3333333334, ans=0.0 2024-09-25 05:16:25,059 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.max_abs, batch_count=671099.3333333334, ans=10.0 2024-09-25 05:16:26,483 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=671146.0, ans=0.025 2024-09-25 05:16:35,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=671146.0, ans=0.035 2024-09-25 05:16:49,745 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=671192.6666666666, ans=0.0 2024-09-25 05:16:59,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=671239.3333333334, ans=0.125 2024-09-25 05:17:22,543 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=671286.0, ans=0.0 2024-09-25 05:17:30,027 INFO [train.py:1198] (2/4) Epoch 37, batch 3600, loss[loss=0.1729, ctc_loss=0.1098, cr_loss=0.3154, over 17268.00 frames. ], tot_loss[loss=0.1917, ctc_loss=0.1238, cr_loss=0.3393, over 3360056.49 frames. ], batch size: 44, lr: 3.19e-03, grad_scale: 32.0 2024-09-25 05:17:30,253 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=671332.6666666666, ans=0.025 2024-09-25 05:17:39,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=671332.6666666666, ans=0.125 2024-09-25 05:18:12,930 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=671426.0, ans=0.125 2024-09-25 05:18:17,629 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=671472.6666666666, ans=0.125 2024-09-25 05:18:50,877 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.287e+02 1.354e+02 1.480e+02 1.820e+02, threshold=2.708e+02, percent-clipped=0.0 2024-09-25 05:18:50,901 INFO [train.py:1198] (2/4) Epoch 37, batch 3650, loss[loss=0.1848, ctc_loss=0.1162, cr_loss=0.3429, over 17282.00 frames. ], tot_loss[loss=0.192, ctc_loss=0.1241, cr_loss=0.3396, over 3349138.56 frames. ], batch size: 51, lr: 3.19e-03, grad_scale: 32.0 2024-09-25 05:19:20,694 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=671659.3333333334, ans=0.04949747468305833 2024-09-25 05:19:52,990 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.95 vs. limit=6.0 2024-09-25 05:19:57,897 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.56 vs. limit=22.5 2024-09-25 05:20:09,451 INFO [train.py:1198] (2/4) Epoch 37, batch 3700, loss[loss=0.1788, ctc_loss=0.1144, cr_loss=0.3217, over 17144.00 frames. ], tot_loss[loss=0.1927, ctc_loss=0.1246, cr_loss=0.3406, over 3349747.03 frames. ], batch size: 48, lr: 3.19e-03, grad_scale: 32.0 2024-09-25 05:20:37,846 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 05:20:46,936 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=671892.6666666666, ans=0.0 2024-09-25 05:20:58,286 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=4.64 vs. limit=12.0 2024-09-25 05:21:24,726 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=671986.0, ans=0.125 2024-09-25 05:21:29,090 INFO [train.py:1198] (2/4) Epoch 37, batch 3750, loss[loss=0.1897, ctc_loss=0.1208, cr_loss=0.3448, over 17253.00 frames. ], tot_loss[loss=0.1922, ctc_loss=0.1242, cr_loss=0.34, over 3358189.27 frames. ], batch size: 42, lr: 3.19e-03, grad_scale: 16.0 2024-09-25 05:21:30,592 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.062e+02 1.295e+02 1.401e+02 1.510e+02 2.261e+02, threshold=2.801e+02, percent-clipped=0.0 2024-09-25 05:21:42,100 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=672032.6666666666, ans=0.07 2024-09-25 05:21:48,702 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 05:22:28,055 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten.whitening_limit, batch_count=672172.6666666666, ans=15.0 2024-09-25 05:22:47,832 INFO [train.py:1198] (2/4) Epoch 37, batch 3800, loss[loss=0.1916, ctc_loss=0.1251, cr_loss=0.3325, over 16632.00 frames. ], tot_loss[loss=0.1922, ctc_loss=0.1241, cr_loss=0.3401, over 3350556.81 frames. ], batch size: 66, lr: 3.19e-03, grad_scale: 16.0 2024-09-25 05:22:48,130 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 05:22:52,737 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 05:23:18,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=672359.3333333334, ans=0.0 2024-09-25 05:23:32,622 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=672359.3333333334, ans=0.1 2024-09-25 05:23:38,059 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=6.57 vs. limit=15.0 2024-09-25 05:23:42,387 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=9.91 vs. limit=15.0 2024-09-25 05:23:45,073 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=672406.0, ans=0.0 2024-09-25 05:23:51,878 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.53 vs. limit=15.0 2024-09-25 05:23:56,287 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=672452.6666666666, ans=0.125 2024-09-25 05:24:07,328 INFO [train.py:1198] (2/4) Epoch 37, batch 3850, loss[loss=0.2064, ctc_loss=0.1409, cr_loss=0.3274, over 12126.00 frames. ], tot_loss[loss=0.1925, ctc_loss=0.1246, cr_loss=0.3397, over 3302725.73 frames. ], batch size: 123, lr: 3.19e-03, grad_scale: 16.0 2024-09-25 05:24:08,856 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.077e+02 1.255e+02 1.356e+02 1.479e+02 3.474e+02, threshold=2.713e+02, percent-clipped=1.0 2024-09-25 05:24:16,596 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=8.79 vs. limit=10.0 2024-09-25 05:24:17,605 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=4.47 vs. limit=12.0 2024-09-25 05:24:21,760 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.max_positive, batch_count=672546.0, ans=0.95 2024-09-25 05:25:08,615 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=672686.0, ans=0.0 2024-09-25 05:26:04,403 INFO [train.py:1198] (2/4) Epoch 38, batch 0, loss[loss=0.2117, ctc_loss=0.1369, cr_loss=0.374, over 16099.00 frames. ], tot_loss[loss=0.2117, ctc_loss=0.1369, cr_loss=0.374, over 16099.00 frames. ], batch size: 74, lr: 3.15e-03, grad_scale: 32.0 2024-09-25 05:26:04,403 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 05:26:20,203 INFO [train.py:1230] (2/4) Epoch 38, validation: loss=0.03515, ctc_loss=0.03515, cr_loss=9.44e-15, over 944034.00 frames. 2024-09-25 05:26:20,204 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 05:26:29,176 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.35 vs. limit=15.0 2024-09-25 05:26:49,682 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=672760.6666666666, ans=0.1 2024-09-25 05:27:12,315 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=7.17 vs. limit=15.0 2024-09-25 05:27:15,063 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=672854.0, ans=0.025 2024-09-25 05:27:37,772 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=672900.6666666666, ans=0.125 2024-09-25 05:27:40,674 INFO [train.py:1198] (2/4) Epoch 38, batch 50, loss[loss=0.1878, ctc_loss=0.12, cr_loss=0.339, over 17012.00 frames. ], tot_loss[loss=0.1975, ctc_loss=0.128, cr_loss=0.3476, over 758023.44 frames. ], batch size: 51, lr: 3.15e-03, grad_scale: 16.0 2024-09-25 05:27:42,969 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=14.07 vs. limit=22.5 2024-09-25 05:27:44,265 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=672947.3333333334, ans=0.2 2024-09-25 05:27:50,497 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.374e+02 1.537e+02 1.726e+02 2.147e+02, threshold=3.075e+02, percent-clipped=0.0 2024-09-25 05:28:11,162 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=672994.0, ans=0.125 2024-09-25 05:28:20,409 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=673040.6666666666, ans=0.0 2024-09-25 05:28:57,447 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=673134.0, ans=0.125 2024-09-25 05:29:03,636 INFO [train.py:1198] (2/4) Epoch 38, batch 100, loss[loss=0.186, ctc_loss=0.1166, cr_loss=0.3468, over 17231.00 frames. ], tot_loss[loss=0.1936, ctc_loss=0.125, cr_loss=0.3429, over 1341228.50 frames. ], batch size: 50, lr: 3.15e-03, grad_scale: 8.0 2024-09-25 05:29:21,597 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=673227.3333333334, ans=0.125 2024-09-25 05:29:45,888 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=7.65 vs. limit=15.0 2024-09-25 05:29:55,303 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=673320.6666666666, ans=0.0 2024-09-25 05:30:20,014 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=673367.3333333334, ans=0.0 2024-09-25 05:30:24,903 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.min_positive, batch_count=673367.3333333334, ans=0.025 2024-09-25 05:30:31,017 INFO [train.py:1198] (2/4) Epoch 38, batch 150, loss[loss=0.1417, ctc_loss=0.08908, cr_loss=0.2631, over 17179.00 frames. ], tot_loss[loss=0.1918, ctc_loss=0.1236, cr_loss=0.3406, over 1791136.84 frames. ], batch size: 41, lr: 3.15e-03, grad_scale: 8.0 2024-09-25 05:30:31,390 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=673414.0, ans=0.125 2024-09-25 05:30:32,917 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=673414.0, ans=0.125 2024-09-25 05:30:42,270 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.262e+02 1.329e+02 1.427e+02 1.998e+02, threshold=2.657e+02, percent-clipped=0.0 2024-09-25 05:31:07,197 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.47 vs. limit=22.5 2024-09-25 05:31:12,484 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.92 vs. limit=6.0 2024-09-25 05:31:40,753 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.93 vs. limit=15.0 2024-09-25 05:31:44,946 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=673600.6666666666, ans=0.125 2024-09-25 05:31:51,203 INFO [train.py:1198] (2/4) Epoch 38, batch 200, loss[loss=0.2048, ctc_loss=0.1323, cr_loss=0.3623, over 17313.00 frames. ], tot_loss[loss=0.192, ctc_loss=0.1236, cr_loss=0.3416, over 2149898.21 frames. ], batch size: 51, lr: 3.15e-03, grad_scale: 8.0 2024-09-25 05:32:06,504 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.65 vs. limit=15.0 2024-09-25 05:32:34,828 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 05:32:38,053 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=673787.3333333334, ans=0.0 2024-09-25 05:32:44,319 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=673787.3333333334, ans=0.2 2024-09-25 05:32:45,899 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=673787.3333333334, ans=0.2 2024-09-25 05:32:51,162 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.20 vs. limit=10.0 2024-09-25 05:32:52,530 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=9.18 vs. limit=15.0 2024-09-25 05:33:00,657 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.60 vs. limit=12.0 2024-09-25 05:33:13,724 INFO [train.py:1198] (2/4) Epoch 38, batch 250, loss[loss=0.1996, ctc_loss=0.1278, cr_loss=0.3591, over 17145.00 frames. ], tot_loss[loss=0.1917, ctc_loss=0.1233, cr_loss=0.342, over 2425875.52 frames. ], batch size: 45, lr: 3.15e-03, grad_scale: 8.0 2024-09-25 05:33:20,293 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=673880.6666666666, ans=0.1 2024-09-25 05:33:24,710 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.142e+02 1.262e+02 1.325e+02 1.385e+02 3.377e+02, threshold=2.651e+02, percent-clipped=1.0 2024-09-25 05:33:40,955 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=673927.3333333334, ans=0.0 2024-09-25 05:34:15,744 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=674067.3333333334, ans=0.0 2024-09-25 05:34:35,576 INFO [train.py:1198] (2/4) Epoch 38, batch 300, loss[loss=0.1918, ctc_loss=0.1218, cr_loss=0.35, over 17152.00 frames. ], tot_loss[loss=0.1923, ctc_loss=0.1238, cr_loss=0.3427, over 2643665.71 frames. ], batch size: 45, lr: 3.14e-03, grad_scale: 8.0 2024-09-25 05:34:38,123 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=4.96 vs. limit=15.0 2024-09-25 05:34:41,554 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.72 vs. limit=6.0 2024-09-25 05:35:03,147 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2.whitening_limit, batch_count=674160.6666666666, ans=15.0 2024-09-25 05:35:15,256 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=674207.3333333334, ans=0.09899494936611666 2024-09-25 05:35:46,903 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=674300.6666666666, ans=0.04949747468305833 2024-09-25 05:35:50,269 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=674300.6666666666, ans=0.2 2024-09-25 05:35:54,968 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=674300.6666666666, ans=0.125 2024-09-25 05:35:55,275 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.61 vs. limit=15.0 2024-09-25 05:35:55,500 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.93 vs. limit=15.0 2024-09-25 05:35:58,219 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=674300.6666666666, ans=0.1 2024-09-25 05:36:01,135 INFO [train.py:1198] (2/4) Epoch 38, batch 350, loss[loss=0.1995, ctc_loss=0.1289, cr_loss=0.3531, over 17322.00 frames. ], tot_loss[loss=0.1921, ctc_loss=0.1237, cr_loss=0.3421, over 2796218.77 frames. ], batch size: 51, lr: 3.14e-03, grad_scale: 8.0 2024-09-25 05:36:01,363 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=674347.3333333334, ans=0.2 2024-09-25 05:36:09,724 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.89 vs. limit=6.0 2024-09-25 05:36:12,222 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.073e+02 1.293e+02 1.371e+02 1.501e+02 2.181e+02, threshold=2.742e+02, percent-clipped=0.0 2024-09-25 05:36:28,548 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=674394.0, ans=0.2 2024-09-25 05:36:29,992 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=674394.0, ans=0.125 2024-09-25 05:36:33,328 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=674440.6666666666, ans=0.125 2024-09-25 05:36:38,046 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=674440.6666666666, ans=0.0 2024-09-25 05:37:00,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=674487.3333333334, ans=0.125 2024-09-25 05:37:11,328 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=674534.0, ans=0.125 2024-09-25 05:37:20,498 INFO [train.py:1198] (2/4) Epoch 38, batch 400, loss[loss=0.1921, ctc_loss=0.1296, cr_loss=0.3124, over 12259.00 frames. ], tot_loss[loss=0.1921, ctc_loss=0.1238, cr_loss=0.3416, over 2911397.06 frames. ], batch size: 124, lr: 3.14e-03, grad_scale: 16.0 2024-09-25 05:37:20,778 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=674580.6666666666, ans=0.125 2024-09-25 05:37:54,123 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=674674.0, ans=0.09899494936611666 2024-09-25 05:38:20,332 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=674720.6666666666, ans=0.125 2024-09-25 05:38:26,850 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=674767.3333333334, ans=0.04949747468305833 2024-09-25 05:38:38,670 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.63 vs. limit=15.0 2024-09-25 05:38:41,654 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=674814.0, ans=0.2 2024-09-25 05:38:42,324 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=5.96 vs. limit=15.0 2024-09-25 05:38:42,901 INFO [train.py:1198] (2/4) Epoch 38, batch 450, loss[loss=0.2049, ctc_loss=0.1346, cr_loss=0.3516, over 17202.00 frames. ], tot_loss[loss=0.1915, ctc_loss=0.1235, cr_loss=0.3402, over 3000330.11 frames. ], batch size: 55, lr: 3.14e-03, grad_scale: 16.0 2024-09-25 05:38:46,347 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=674814.0, ans=0.0 2024-09-25 05:38:55,495 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.071e+02 1.292e+02 1.363e+02 1.447e+02 2.119e+02, threshold=2.727e+02, percent-clipped=0.0 2024-09-25 05:39:01,308 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=512, metric=6.06 vs. limit=15.0 2024-09-25 05:39:18,247 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=674907.3333333334, ans=0.125 2024-09-25 05:39:21,642 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=674907.3333333334, ans=0.125 2024-09-25 05:39:27,395 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=674907.3333333334, ans=0.1 2024-09-25 05:39:32,619 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=9.44 vs. limit=15.0 2024-09-25 05:40:11,290 INFO [train.py:1198] (2/4) Epoch 38, batch 500, loss[loss=0.2491, ctc_loss=0.1678, cr_loss=0.4065, over 14964.00 frames. ], tot_loss[loss=0.1918, ctc_loss=0.1237, cr_loss=0.3405, over 3072749.94 frames. ], batch size: 88, lr: 3.14e-03, grad_scale: 8.0 2024-09-25 05:40:24,378 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=675047.3333333334, ans=0.2 2024-09-25 05:40:33,880 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=675094.0, ans=0.025 2024-09-25 05:41:23,269 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.54 vs. limit=12.0 2024-09-25 05:41:29,324 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=675280.6666666666, ans=0.125 2024-09-25 05:41:30,503 INFO [train.py:1198] (2/4) Epoch 38, batch 550, loss[loss=0.1857, ctc_loss=0.1172, cr_loss=0.342, over 17098.00 frames. ], tot_loss[loss=0.1932, ctc_loss=0.1248, cr_loss=0.342, over 3126105.54 frames. ], batch size: 40, lr: 3.14e-03, grad_scale: 8.0 2024-09-25 05:41:43,253 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.284e+02 1.364e+02 1.434e+02 1.794e+02, threshold=2.728e+02, percent-clipped=0.0 2024-09-25 05:41:54,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=675327.3333333334, ans=0.0 2024-09-25 05:42:28,226 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=675420.6666666666, ans=0.125 2024-09-25 05:42:33,181 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=675467.3333333334, ans=0.125 2024-09-25 05:42:36,364 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 05:42:43,184 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.57 vs. limit=10.0 2024-09-25 05:42:44,156 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=675467.3333333334, ans=0.2 2024-09-25 05:42:50,160 INFO [train.py:1198] (2/4) Epoch 38, batch 600, loss[loss=0.2151, ctc_loss=0.14, cr_loss=0.3758, over 16827.00 frames. ], tot_loss[loss=0.1924, ctc_loss=0.1243, cr_loss=0.3405, over 3180146.05 frames. ], batch size: 61, lr: 3.14e-03, grad_scale: 8.0 2024-09-25 05:42:50,421 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=675514.0, ans=0.0 2024-09-25 05:44:12,412 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=7.89 vs. limit=15.0 2024-09-25 05:44:13,109 INFO [train.py:1198] (2/4) Epoch 38, batch 650, loss[loss=0.1645, ctc_loss=0.1039, cr_loss=0.3031, over 17171.00 frames. ], tot_loss[loss=0.1923, ctc_loss=0.1242, cr_loss=0.3402, over 3215570.17 frames. ], batch size: 41, lr: 3.14e-03, grad_scale: 8.0 2024-09-25 05:44:28,504 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.165e+02 1.283e+02 1.368e+02 1.490e+02 2.037e+02, threshold=2.736e+02, percent-clipped=0.0 2024-09-25 05:45:07,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=675887.3333333334, ans=0.1 2024-09-25 05:45:08,885 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=675887.3333333334, ans=0.035 2024-09-25 05:45:23,305 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=675934.0, ans=0.125 2024-09-25 05:45:40,414 INFO [train.py:1198] (2/4) Epoch 38, batch 700, loss[loss=0.239, ctc_loss=0.1611, cr_loss=0.3893, over 12032.00 frames. ], tot_loss[loss=0.193, ctc_loss=0.1247, cr_loss=0.3412, over 3240463.35 frames. ], batch size: 123, lr: 3.14e-03, grad_scale: 8.0 2024-09-25 05:45:51,913 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=675980.6666666666, ans=0.125 2024-09-25 05:45:53,535 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=675980.6666666666, ans=0.125 2024-09-25 05:45:56,734 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=676027.3333333334, ans=0.07 2024-09-25 05:46:01,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=676027.3333333334, ans=0.2 2024-09-25 05:46:06,258 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=676027.3333333334, ans=0.07 2024-09-25 05:46:06,403 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=676027.3333333334, ans=0.125 2024-09-25 05:46:09,652 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=3.183e-02 2024-09-25 05:46:47,790 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=676167.3333333334, ans=0.125 2024-09-25 05:47:00,009 INFO [train.py:1198] (2/4) Epoch 38, batch 750, loss[loss=0.2061, ctc_loss=0.1374, cr_loss=0.3434, over 17248.00 frames. ], tot_loss[loss=0.193, ctc_loss=0.1248, cr_loss=0.3413, over 3264093.73 frames. ], batch size: 44, lr: 3.14e-03, grad_scale: 8.0 2024-09-25 05:47:03,551 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=676214.0, ans=0.2 2024-09-25 05:47:10,172 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.91 vs. limit=22.5 2024-09-25 05:47:12,449 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.277e+02 1.363e+02 1.416e+02 2.105e+02, threshold=2.725e+02, percent-clipped=0.0 2024-09-25 05:47:17,789 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=5.57 vs. limit=12.0 2024-09-25 05:47:26,960 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=676260.6666666666, ans=0.07 2024-09-25 05:47:27,000 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=676260.6666666666, ans=0.125 2024-09-25 05:47:38,652 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.67 vs. limit=10.0 2024-09-25 05:47:44,407 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=676307.3333333334, ans=0.125 2024-09-25 05:48:14,214 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=7.03 vs. limit=15.0 2024-09-25 05:48:21,649 INFO [train.py:1198] (2/4) Epoch 38, batch 800, loss[loss=0.179, ctc_loss=0.1131, cr_loss=0.3293, over 17116.00 frames. ], tot_loss[loss=0.1928, ctc_loss=0.1246, cr_loss=0.3409, over 3278870.15 frames. ], batch size: 49, lr: 3.14e-03, grad_scale: 16.0 2024-09-25 05:48:21,969 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=676447.3333333334, ans=0.1 2024-09-25 05:48:46,829 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.83 vs. limit=6.0 2024-09-25 05:48:52,359 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=676540.6666666666, ans=0.2 2024-09-25 05:49:23,091 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=676587.3333333334, ans=0.1 2024-09-25 05:49:41,799 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer_na.min_abs, batch_count=676634.0, ans=0.02 2024-09-25 05:49:49,052 INFO [train.py:1198] (2/4) Epoch 38, batch 850, loss[loss=0.2059, ctc_loss=0.1345, cr_loss=0.3572, over 16871.00 frames. ], tot_loss[loss=0.1926, ctc_loss=0.1245, cr_loss=0.3406, over 3298600.74 frames. ], batch size: 58, lr: 3.14e-03, grad_scale: 16.0 2024-09-25 05:49:57,404 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=676680.6666666666, ans=0.0 2024-09-25 05:50:01,636 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.050e+02 1.281e+02 1.360e+02 1.434e+02 2.186e+02, threshold=2.720e+02, percent-clipped=0.0 2024-09-25 05:50:03,724 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=676727.3333333334, ans=0.125 2024-09-25 05:50:09,987 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=676727.3333333334, ans=0.1 2024-09-25 05:50:21,208 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=676774.0, ans=0.1 2024-09-25 05:50:40,483 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 05:50:41,923 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=676820.6666666666, ans=0.125 2024-09-25 05:50:48,321 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=676820.6666666666, ans=0.0 2024-09-25 05:50:53,027 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=676867.3333333334, ans=0.0 2024-09-25 05:51:08,594 INFO [train.py:1198] (2/4) Epoch 38, batch 900, loss[loss=0.2079, ctc_loss=0.1344, cr_loss=0.3676, over 15926.00 frames. ], tot_loss[loss=0.1936, ctc_loss=0.1252, cr_loss=0.3421, over 3303224.07 frames. ], batch size: 74, lr: 3.14e-03, grad_scale: 16.0 2024-09-25 05:51:13,679 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=676914.0, ans=0.0 2024-09-25 05:51:24,850 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=676960.6666666666, ans=0.035 2024-09-25 05:51:28,232 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 05:52:10,061 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=677054.0, ans=0.125 2024-09-25 05:52:29,101 INFO [train.py:1198] (2/4) Epoch 38, batch 950, loss[loss=0.1739, ctc_loss=0.1109, cr_loss=0.3149, over 16697.00 frames. ], tot_loss[loss=0.1947, ctc_loss=0.126, cr_loss=0.3435, over 3305477.50 frames. ], batch size: 37, lr: 3.14e-03, grad_scale: 16.0 2024-09-25 05:52:30,960 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=677147.3333333334, ans=0.125 2024-09-25 05:52:41,915 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.327e+02 1.410e+02 1.537e+02 3.330e+02, threshold=2.819e+02, percent-clipped=2.0 2024-09-25 05:52:59,163 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=677194.0, ans=0.1 2024-09-25 05:53:52,325 INFO [train.py:1198] (2/4) Epoch 38, batch 1000, loss[loss=0.1644, ctc_loss=0.1065, cr_loss=0.2893, over 17029.00 frames. ], tot_loss[loss=0.1942, ctc_loss=0.1256, cr_loss=0.3427, over 3315920.26 frames. ], batch size: 39, lr: 3.14e-03, grad_scale: 16.0 2024-09-25 05:54:36,098 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=677474.0, ans=0.2 2024-09-25 05:54:46,718 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=677520.6666666666, ans=0.0 2024-09-25 05:54:58,445 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.97 vs. limit=6.0 2024-09-25 05:55:20,559 INFO [train.py:1198] (2/4) Epoch 38, batch 1050, loss[loss=0.2254, ctc_loss=0.1496, cr_loss=0.3789, over 15131.00 frames. ], tot_loss[loss=0.1945, ctc_loss=0.1258, cr_loss=0.3431, over 3311986.21 frames. ], batch size: 89, lr: 3.14e-03, grad_scale: 16.0 2024-09-25 05:55:20,944 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=677614.0, ans=0.125 2024-09-25 05:55:30,617 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=677614.0, ans=0.0 2024-09-25 05:55:32,362 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=677614.0, ans=0.0 2024-09-25 05:55:33,504 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.113e+02 1.267e+02 1.356e+02 1.451e+02 1.928e+02, threshold=2.711e+02, percent-clipped=0.0 2024-09-25 05:55:48,999 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.whiten, num_groups=1, num_channels=192, metric=4.80 vs. limit=12.0 2024-09-25 05:55:49,587 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=677660.6666666666, ans=0.125 2024-09-25 05:56:07,957 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=18.69 vs. limit=22.5 2024-09-25 05:56:40,343 INFO [train.py:1198] (2/4) Epoch 38, batch 1100, loss[loss=0.1605, ctc_loss=0.1025, cr_loss=0.2899, over 16963.00 frames. ], tot_loss[loss=0.193, ctc_loss=0.1247, cr_loss=0.3414, over 3328303.71 frames. ], batch size: 42, lr: 3.14e-03, grad_scale: 16.0 2024-09-25 05:57:23,803 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=677940.6666666666, ans=0.07 2024-09-25 05:57:29,166 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=9.14 vs. limit=15.0 2024-09-25 05:57:49,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=678034.0, ans=0.0 2024-09-25 05:57:50,820 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=678034.0, ans=0.125 2024-09-25 05:57:58,083 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=678034.0, ans=0.1 2024-09-25 05:58:02,651 INFO [train.py:1198] (2/4) Epoch 38, batch 1150, loss[loss=0.1618, ctc_loss=0.1018, cr_loss=0.2995, over 17189.00 frames. ], tot_loss[loss=0.1929, ctc_loss=0.1246, cr_loss=0.3414, over 3343200.84 frames. ], batch size: 41, lr: 3.14e-03, grad_scale: 16.0 2024-09-25 05:58:07,965 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=6.62 vs. limit=15.0 2024-09-25 05:58:15,129 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.083e+02 1.254e+02 1.322e+02 1.438e+02 2.414e+02, threshold=2.644e+02, percent-clipped=0.0 2024-09-25 05:58:25,059 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.min_positive, batch_count=678127.3333333334, ans=0.05 2024-09-25 05:58:30,051 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=678127.3333333334, ans=0.0 2024-09-25 05:58:31,504 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=678127.3333333334, ans=0.0 2024-09-25 05:58:55,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=678220.6666666666, ans=0.125 2024-09-25 05:58:55,881 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=678220.6666666666, ans=0.025 2024-09-25 05:58:59,095 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=678220.6666666666, ans=0.025 2024-09-25 05:59:25,215 INFO [train.py:1198] (2/4) Epoch 38, batch 1200, loss[loss=0.1503, ctc_loss=0.09255, cr_loss=0.289, over 17122.00 frames. ], tot_loss[loss=0.191, ctc_loss=0.1233, cr_loss=0.3385, over 3352370.20 frames. ], batch size: 40, lr: 3.13e-03, grad_scale: 16.0 2024-09-25 05:59:44,116 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=678360.6666666666, ans=0.025 2024-09-25 05:59:44,472 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=10.73 vs. limit=12.0 2024-09-25 05:59:59,587 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 06:00:10,737 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 06:00:28,062 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.min_positive, batch_count=678454.0, ans=0.025 2024-09-25 06:00:50,407 INFO [train.py:1198] (2/4) Epoch 38, batch 1250, loss[loss=0.167, ctc_loss=0.1054, cr_loss=0.308, over 17286.00 frames. ], tot_loss[loss=0.1919, ctc_loss=0.124, cr_loss=0.3396, over 3334487.64 frames. ], batch size: 42, lr: 3.13e-03, grad_scale: 16.0 2024-09-25 06:00:52,316 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=678547.3333333334, ans=0.5 2024-09-25 06:00:57,553 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.98 vs. limit=22.5 2024-09-25 06:01:04,690 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.074e+02 1.283e+02 1.378e+02 1.489e+02 1.932e+02, threshold=2.757e+02, percent-clipped=0.0 2024-09-25 06:01:07,216 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.58 vs. limit=15.0 2024-09-25 06:01:22,762 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=678640.6666666666, ans=0.2 2024-09-25 06:01:22,832 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=678640.6666666666, ans=0.125 2024-09-25 06:01:24,348 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=678640.6666666666, ans=0.0 2024-09-25 06:01:32,952 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=11.74 vs. limit=22.5 2024-09-25 06:02:03,405 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=678734.0, ans=0.0 2024-09-25 06:02:09,790 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=678780.6666666666, ans=0.07 2024-09-25 06:02:11,135 INFO [train.py:1198] (2/4) Epoch 38, batch 1300, loss[loss=0.1878, ctc_loss=0.1194, cr_loss=0.3419, over 17012.00 frames. ], tot_loss[loss=0.1915, ctc_loss=0.1237, cr_loss=0.3387, over 3338940.48 frames. ], batch size: 44, lr: 3.13e-03, grad_scale: 16.0 2024-09-25 06:02:15,228 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.65 vs. limit=15.0 2024-09-25 06:02:21,185 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=678780.6666666666, ans=0.125 2024-09-25 06:02:32,795 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=3.79 vs. limit=10.0 2024-09-25 06:03:13,353 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=678920.6666666666, ans=0.1 2024-09-25 06:03:33,484 INFO [train.py:1198] (2/4) Epoch 38, batch 1350, loss[loss=0.2401, ctc_loss=0.1586, cr_loss=0.4076, over 15289.00 frames. ], tot_loss[loss=0.1916, ctc_loss=0.1238, cr_loss=0.339, over 3348111.55 frames. ], batch size: 89, lr: 3.13e-03, grad_scale: 16.0 2024-09-25 06:03:47,755 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.262e+02 1.346e+02 1.430e+02 2.071e+02, threshold=2.693e+02, percent-clipped=0.0 2024-09-25 06:04:25,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=679154.0, ans=0.125 2024-09-25 06:04:37,995 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2.whitening_limit, batch_count=679154.0, ans=15.0 2024-09-25 06:04:54,870 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 06:04:54,893 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=679200.6666666666, ans=0.0 2024-09-25 06:05:01,173 INFO [train.py:1198] (2/4) Epoch 38, batch 1400, loss[loss=0.2164, ctc_loss=0.1395, cr_loss=0.3847, over 17291.00 frames. ], tot_loss[loss=0.1918, ctc_loss=0.1239, cr_loss=0.3397, over 3352463.46 frames. ], batch size: 49, lr: 3.13e-03, grad_scale: 16.0 2024-09-25 06:05:03,195 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=679247.3333333334, ans=0.1 2024-09-25 06:05:20,578 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=679294.0, ans=0.125 2024-09-25 06:05:38,055 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=679340.6666666666, ans=0.0 2024-09-25 06:05:44,586 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=679340.6666666666, ans=0.125 2024-09-25 06:05:45,431 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=7.22 vs. limit=8.0 2024-09-25 06:05:54,507 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=12.91 vs. limit=22.5 2024-09-25 06:05:57,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=679387.3333333334, ans=0.07 2024-09-25 06:06:02,026 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=679387.3333333334, ans=0.125 2024-09-25 06:06:21,081 INFO [train.py:1198] (2/4) Epoch 38, batch 1450, loss[loss=0.1523, ctc_loss=0.09734, cr_loss=0.2748, over 17183.00 frames. ], tot_loss[loss=0.191, ctc_loss=0.1232, cr_loss=0.339, over 3363943.96 frames. ], batch size: 41, lr: 3.13e-03, grad_scale: 16.0 2024-09-25 06:06:29,499 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=679480.6666666666, ans=0.07 2024-09-25 06:06:35,618 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.086e+02 1.250e+02 1.326e+02 1.407e+02 2.354e+02, threshold=2.651e+02, percent-clipped=0.0 2024-09-25 06:06:57,179 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.52 vs. limit=12.0 2024-09-25 06:07:07,716 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=679620.6666666666, ans=0.0 2024-09-25 06:07:14,543 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=3.78 vs. limit=10.0 2024-09-25 06:07:21,853 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=679620.6666666666, ans=0.0 2024-09-25 06:07:24,141 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=5.06 vs. limit=12.0 2024-09-25 06:07:41,099 INFO [train.py:1198] (2/4) Epoch 38, batch 1500, loss[loss=0.1891, ctc_loss=0.121, cr_loss=0.3404, over 17102.00 frames. ], tot_loss[loss=0.1917, ctc_loss=0.1236, cr_loss=0.3403, over 3368644.84 frames. ], batch size: 49, lr: 3.13e-03, grad_scale: 16.0 2024-09-25 06:07:58,662 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=679760.6666666666, ans=0.0 2024-09-25 06:08:19,101 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=679807.3333333334, ans=0.125 2024-09-25 06:08:30,331 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=679854.0, ans=0.0 2024-09-25 06:08:36,668 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=679854.0, ans=0.125 2024-09-25 06:08:51,352 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=679900.6666666666, ans=0.125 2024-09-25 06:09:06,164 INFO [train.py:1198] (2/4) Epoch 38, batch 1550, loss[loss=0.1982, ctc_loss=0.1274, cr_loss=0.3545, over 17353.00 frames. ], tot_loss[loss=0.1915, ctc_loss=0.1235, cr_loss=0.3401, over 3366736.13 frames. ], batch size: 48, lr: 3.13e-03, grad_scale: 16.0 2024-09-25 06:09:20,508 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.155e+02 1.263e+02 1.343e+02 1.440e+02 2.044e+02, threshold=2.685e+02, percent-clipped=0.0 2024-09-25 06:10:02,199 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.whiten, num_groups=1, num_channels=192, metric=5.08 vs. limit=12.0 2024-09-25 06:10:17,352 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=680134.0, ans=0.0 2024-09-25 06:10:31,478 INFO [train.py:1198] (2/4) Epoch 38, batch 1600, loss[loss=0.1696, ctc_loss=0.1076, cr_loss=0.31, over 17310.00 frames. ], tot_loss[loss=0.1911, ctc_loss=0.1233, cr_loss=0.3391, over 3363407.47 frames. ], batch size: 46, lr: 3.13e-03, grad_scale: 32.0 2024-09-25 06:10:38,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=680180.6666666666, ans=0.125 2024-09-25 06:11:02,127 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=680274.0, ans=0.0 2024-09-25 06:11:41,746 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.max_abs, batch_count=680367.3333333334, ans=10.0 2024-09-25 06:11:51,471 INFO [train.py:1198] (2/4) Epoch 38, batch 1650, loss[loss=0.2153, ctc_loss=0.1363, cr_loss=0.3949, over 17251.00 frames. ], tot_loss[loss=0.1906, ctc_loss=0.1229, cr_loss=0.3389, over 3365958.00 frames. ], batch size: 44, lr: 3.13e-03, grad_scale: 32.0 2024-09-25 06:12:05,734 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.139e+02 1.273e+02 1.346e+02 1.505e+02 2.146e+02, threshold=2.692e+02, percent-clipped=0.0 2024-09-25 06:12:13,164 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=4.75 vs. limit=15.0 2024-09-25 06:12:29,883 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=680507.3333333334, ans=0.125 2024-09-25 06:12:38,346 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=12.27 vs. limit=22.5 2024-09-25 06:12:45,741 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=680554.0, ans=0.125 2024-09-25 06:12:51,624 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=680554.0, ans=0.0 2024-09-25 06:13:13,538 INFO [train.py:1198] (2/4) Epoch 38, batch 1700, loss[loss=0.1815, ctc_loss=0.1163, cr_loss=0.3264, over 17250.00 frames. ], tot_loss[loss=0.1918, ctc_loss=0.1237, cr_loss=0.3409, over 3357575.45 frames. ], batch size: 44, lr: 3.13e-03, grad_scale: 32.0 2024-09-25 06:13:31,291 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=680694.0, ans=0.0 2024-09-25 06:13:32,846 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=680694.0, ans=0.2 2024-09-25 06:13:45,642 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=680740.6666666666, ans=0.125 2024-09-25 06:13:48,864 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=680740.6666666666, ans=0.125 2024-09-25 06:14:38,370 INFO [train.py:1198] (2/4) Epoch 38, batch 1750, loss[loss=0.2031, ctc_loss=0.1292, cr_loss=0.3696, over 17029.00 frames. ], tot_loss[loss=0.1922, ctc_loss=0.1239, cr_loss=0.3415, over 3359887.76 frames. ], batch size: 52, lr: 3.13e-03, grad_scale: 32.0 2024-09-25 06:14:52,906 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.33 vs. limit=15.0 2024-09-25 06:14:55,388 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.278e+02 1.373e+02 1.469e+02 4.120e+02, threshold=2.745e+02, percent-clipped=1.0 2024-09-25 06:14:55,650 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=680927.3333333334, ans=0.125 2024-09-25 06:14:57,279 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=680927.3333333334, ans=0.025 2024-09-25 06:14:58,967 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=680927.3333333334, ans=0.125 2024-09-25 06:15:08,460 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=680927.3333333334, ans=0.125 2024-09-25 06:16:01,066 INFO [train.py:1198] (2/4) Epoch 38, batch 1800, loss[loss=0.2131, ctc_loss=0.1433, cr_loss=0.3489, over 12315.00 frames. ], tot_loss[loss=0.1928, ctc_loss=0.1243, cr_loss=0.3422, over 3361438.26 frames. ], batch size: 123, lr: 3.13e-03, grad_scale: 32.0 2024-09-25 06:16:14,298 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=681114.0, ans=0.1 2024-09-25 06:16:25,839 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=11.76 vs. limit=12.0 2024-09-25 06:16:51,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=681254.0, ans=0.0 2024-09-25 06:17:18,967 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=681300.6666666666, ans=0.0 2024-09-25 06:17:21,941 INFO [train.py:1198] (2/4) Epoch 38, batch 1850, loss[loss=0.2235, ctc_loss=0.1447, cr_loss=0.3937, over 17009.00 frames. ], tot_loss[loss=0.1926, ctc_loss=0.1242, cr_loss=0.3418, over 3369297.02 frames. ], batch size: 53, lr: 3.13e-03, grad_scale: 32.0 2024-09-25 06:17:30,272 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=681347.3333333334, ans=0.1 2024-09-25 06:17:36,436 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.079e+02 1.263e+02 1.331e+02 1.457e+02 2.352e+02, threshold=2.661e+02, percent-clipped=0.0 2024-09-25 06:18:03,476 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=681440.6666666666, ans=0.025 2024-09-25 06:18:30,407 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 06:18:32,169 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=681534.0, ans=0.125 2024-09-25 06:18:44,346 INFO [train.py:1198] (2/4) Epoch 38, batch 1900, loss[loss=0.1599, ctc_loss=0.1001, cr_loss=0.2989, over 17111.00 frames. ], tot_loss[loss=0.192, ctc_loss=0.1238, cr_loss=0.3408, over 3365071.76 frames. ], batch size: 40, lr: 3.13e-03, grad_scale: 32.0 2024-09-25 06:18:44,790 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 06:19:06,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=681627.3333333334, ans=0.125 2024-09-25 06:19:15,810 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=681627.3333333334, ans=0.04949747468305833 2024-09-25 06:19:19,087 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-25 06:19:37,652 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=681720.6666666666, ans=0.2 2024-09-25 06:19:42,509 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=681720.6666666666, ans=0.125 2024-09-25 06:19:59,467 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=681767.3333333334, ans=0.1 2024-09-25 06:20:07,377 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=681767.3333333334, ans=0.125 2024-09-25 06:20:11,785 INFO [train.py:1198] (2/4) Epoch 38, batch 1950, loss[loss=0.2215, ctc_loss=0.1467, cr_loss=0.3742, over 15991.00 frames. ], tot_loss[loss=0.1933, ctc_loss=0.1246, cr_loss=0.3431, over 3375465.83 frames. ], batch size: 74, lr: 3.13e-03, grad_scale: 32.0 2024-09-25 06:20:13,699 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=681814.0, ans=0.125 2024-09-25 06:20:27,500 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.069e+02 1.276e+02 1.376e+02 1.498e+02 2.117e+02, threshold=2.753e+02, percent-clipped=0.0 2024-09-25 06:20:42,499 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=681907.3333333334, ans=0.125 2024-09-25 06:20:45,561 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=681907.3333333334, ans=0.025 2024-09-25 06:20:47,605 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.17 vs. limit=15.0 2024-09-25 06:21:04,867 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=681954.0, ans=0.125 2024-09-25 06:21:06,222 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=681954.0, ans=10.0 2024-09-25 06:21:12,554 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=681954.0, ans=0.0 2024-09-25 06:21:19,059 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=682000.6666666666, ans=0.1 2024-09-25 06:21:21,835 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=682000.6666666666, ans=0.125 2024-09-25 06:21:31,234 INFO [train.py:1198] (2/4) Epoch 38, batch 2000, loss[loss=0.1866, ctc_loss=0.1198, cr_loss=0.334, over 17343.00 frames. ], tot_loss[loss=0.193, ctc_loss=0.1246, cr_loss=0.3423, over 3376036.27 frames. ], batch size: 52, lr: 3.13e-03, grad_scale: 32.0 2024-09-25 06:21:47,585 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=682094.0, ans=0.0 2024-09-25 06:21:50,743 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=682094.0, ans=0.07 2024-09-25 06:21:57,133 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=682094.0, ans=0.0 2024-09-25 06:22:10,036 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=682140.6666666666, ans=0.125 2024-09-25 06:22:42,676 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.14 vs. limit=6.0 2024-09-25 06:22:51,389 INFO [train.py:1198] (2/4) Epoch 38, batch 2050, loss[loss=0.1608, ctc_loss=0.1024, cr_loss=0.2923, over 16257.00 frames. ], tot_loss[loss=0.1939, ctc_loss=0.1252, cr_loss=0.3439, over 3369809.02 frames. ], batch size: 36, lr: 3.13e-03, grad_scale: 32.0 2024-09-25 06:22:53,291 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=682280.6666666666, ans=0.125 2024-09-25 06:23:05,646 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=682280.6666666666, ans=0.07 2024-09-25 06:23:09,956 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.278e+02 1.341e+02 1.480e+02 2.182e+02, threshold=2.683e+02, percent-clipped=0.0 2024-09-25 06:23:36,160 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=682374.0, ans=0.1 2024-09-25 06:23:39,112 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=682374.0, ans=0.125 2024-09-25 06:24:09,262 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=5.58 vs. limit=15.0 2024-09-25 06:24:16,374 INFO [train.py:1198] (2/4) Epoch 38, batch 2100, loss[loss=0.2096, ctc_loss=0.1347, cr_loss=0.3742, over 17199.00 frames. ], tot_loss[loss=0.1942, ctc_loss=0.1254, cr_loss=0.3441, over 3369272.04 frames. ], batch size: 55, lr: 3.13e-03, grad_scale: 32.0 2024-09-25 06:24:23,133 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=682514.0, ans=0.0 2024-09-25 06:24:23,146 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=682514.0, ans=0.2 2024-09-25 06:24:30,627 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=682514.0, ans=0.04949747468305833 2024-09-25 06:24:55,084 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=682607.3333333334, ans=0.125 2024-09-25 06:24:58,562 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=13.04 vs. limit=15.0 2024-09-25 06:24:59,726 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=682607.3333333334, ans=0.125 2024-09-25 06:24:59,970 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=682607.3333333334, ans=0.125 2024-09-25 06:25:13,017 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=8.51 vs. limit=12.0 2024-09-25 06:25:15,804 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=682654.0, ans=0.0 2024-09-25 06:25:26,087 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.40 vs. limit=15.0 2024-09-25 06:25:26,688 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=682700.6666666666, ans=0.125 2024-09-25 06:25:41,097 INFO [train.py:1198] (2/4) Epoch 38, batch 2150, loss[loss=0.2351, ctc_loss=0.1587, cr_loss=0.3819, over 12275.00 frames. ], tot_loss[loss=0.1945, ctc_loss=0.1257, cr_loss=0.3443, over 3362468.69 frames. ], batch size: 123, lr: 3.12e-03, grad_scale: 16.0 2024-09-25 06:25:59,160 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.107e+02 1.297e+02 1.376e+02 1.525e+02 2.502e+02, threshold=2.751e+02, percent-clipped=0.0 2024-09-25 06:26:13,895 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=682840.6666666666, ans=0.125 2024-09-25 06:26:14,577 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=11.46 vs. limit=22.5 2024-09-25 06:26:18,573 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=682840.6666666666, ans=0.125 2024-09-25 06:26:18,979 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.24 vs. limit=15.0 2024-09-25 06:26:36,127 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=682887.3333333334, ans=0.025 2024-09-25 06:26:50,842 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=4.02 vs. limit=15.0 2024-09-25 06:27:01,625 INFO [train.py:1198] (2/4) Epoch 38, batch 2200, loss[loss=0.1812, ctc_loss=0.1135, cr_loss=0.3384, over 17005.00 frames. ], tot_loss[loss=0.196, ctc_loss=0.1267, cr_loss=0.3463, over 3355169.62 frames. ], batch size: 44, lr: 3.12e-03, grad_scale: 16.0 2024-09-25 06:27:01,943 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=682980.6666666666, ans=0.1 2024-09-25 06:27:23,233 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module1.whiten, num_groups=1, num_channels=512, metric=6.54 vs. limit=15.0 2024-09-25 06:27:25,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=683027.3333333334, ans=0.125 2024-09-25 06:27:35,701 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=683074.0, ans=0.125 2024-09-25 06:27:41,858 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=683074.0, ans=0.125 2024-09-25 06:28:24,735 INFO [train.py:1198] (2/4) Epoch 38, batch 2250, loss[loss=0.1767, ctc_loss=0.1136, cr_loss=0.3156, over 17081.00 frames. ], tot_loss[loss=0.1949, ctc_loss=0.126, cr_loss=0.3446, over 3354812.97 frames. ], batch size: 43, lr: 3.12e-03, grad_scale: 16.0 2024-09-25 06:28:30,033 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=683214.0, ans=0.2 2024-09-25 06:28:31,509 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=683214.0, ans=0.1 2024-09-25 06:28:31,526 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=683214.0, ans=0.125 2024-09-25 06:28:33,246 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=683214.0, ans=0.025 2024-09-25 06:28:42,346 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.094e+02 1.283e+02 1.354e+02 1.470e+02 2.386e+02, threshold=2.708e+02, percent-clipped=0.0 2024-09-25 06:28:45,905 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=683260.6666666666, ans=0.0 2024-09-25 06:28:50,751 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=683260.6666666666, ans=0.2 2024-09-25 06:28:57,104 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=683307.3333333334, ans=0.2 2024-09-25 06:29:17,432 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=13.05 vs. limit=22.5 2024-09-25 06:29:20,262 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=683354.0, ans=0.1 2024-09-25 06:29:41,874 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=683400.6666666666, ans=0.2 2024-09-25 06:29:42,014 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=683400.6666666666, ans=0.125 2024-09-25 06:29:49,854 INFO [train.py:1198] (2/4) Epoch 38, batch 2300, loss[loss=0.1786, ctc_loss=0.1127, cr_loss=0.3296, over 17025.00 frames. ], tot_loss[loss=0.194, ctc_loss=0.1253, cr_loss=0.3435, over 3360808.39 frames. ], batch size: 44, lr: 3.12e-03, grad_scale: 16.0 2024-09-25 06:30:08,429 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=683494.0, ans=0.0 2024-09-25 06:31:12,144 INFO [train.py:1198] (2/4) Epoch 38, batch 2350, loss[loss=0.2067, ctc_loss=0.1357, cr_loss=0.3552, over 15232.00 frames. ], tot_loss[loss=0.1929, ctc_loss=0.1245, cr_loss=0.3423, over 3360190.67 frames. ], batch size: 89, lr: 3.12e-03, grad_scale: 16.0 2024-09-25 06:31:14,226 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=683680.6666666666, ans=0.125 2024-09-25 06:31:22,068 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=683680.6666666666, ans=0.0 2024-09-25 06:31:29,751 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.295e+02 1.360e+02 1.434e+02 2.304e+02, threshold=2.721e+02, percent-clipped=0.0 2024-09-25 06:31:41,503 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=683727.3333333334, ans=0.125 2024-09-25 06:31:50,979 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=683774.0, ans=0.125 2024-09-25 06:31:54,162 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=683774.0, ans=0.0 2024-09-25 06:31:57,337 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=683774.0, ans=0.025 2024-09-25 06:32:13,179 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=683820.6666666666, ans=0.0 2024-09-25 06:32:16,217 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=683867.3333333334, ans=0.125 2024-09-25 06:32:31,858 INFO [train.py:1198] (2/4) Epoch 38, batch 2400, loss[loss=0.2006, ctc_loss=0.1269, cr_loss=0.3683, over 17088.00 frames. ], tot_loss[loss=0.1923, ctc_loss=0.124, cr_loss=0.3414, over 3360994.36 frames. ], batch size: 49, lr: 3.12e-03, grad_scale: 32.0 2024-09-25 06:32:46,358 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=683960.6666666666, ans=0.125 2024-09-25 06:32:46,510 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=683960.6666666666, ans=0.0 2024-09-25 06:33:03,125 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=683960.6666666666, ans=0.1 2024-09-25 06:33:12,737 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=684007.3333333334, ans=0.125 2024-09-25 06:33:54,234 INFO [train.py:1198] (2/4) Epoch 38, batch 2450, loss[loss=0.1896, ctc_loss=0.1223, cr_loss=0.3365, over 17025.00 frames. ], tot_loss[loss=0.1926, ctc_loss=0.1244, cr_loss=0.3415, over 3361734.20 frames. ], batch size: 44, lr: 3.12e-03, grad_scale: 32.0 2024-09-25 06:34:02,092 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=684147.3333333334, ans=0.0 2024-09-25 06:34:03,841 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten.whitening_limit, batch_count=684147.3333333334, ans=15.0 2024-09-25 06:34:14,632 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.026e+02 1.306e+02 1.378e+02 1.472e+02 2.938e+02, threshold=2.756e+02, percent-clipped=1.0 2024-09-25 06:34:53,954 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=684287.3333333334, ans=0.125 2024-09-25 06:35:00,780 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2.whitening_limit, batch_count=684287.3333333334, ans=15.0 2024-09-25 06:35:22,453 INFO [train.py:1198] (2/4) Epoch 38, batch 2500, loss[loss=0.2014, ctc_loss=0.1302, cr_loss=0.3559, over 16758.00 frames. ], tot_loss[loss=0.1918, ctc_loss=0.1237, cr_loss=0.3404, over 3356868.01 frames. ], batch size: 61, lr: 3.12e-03, grad_scale: 32.0 2024-09-25 06:36:18,727 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 06:36:29,696 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=684567.3333333334, ans=0.04949747468305833 2024-09-25 06:36:32,943 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=684567.3333333334, ans=0.125 2024-09-25 06:36:42,199 INFO [train.py:1198] (2/4) Epoch 38, batch 2550, loss[loss=0.1461, ctc_loss=0.09116, cr_loss=0.2745, over 16682.00 frames. ], tot_loss[loss=0.1923, ctc_loss=0.124, cr_loss=0.3415, over 3358608.45 frames. ], batch size: 37, lr: 3.12e-03, grad_scale: 32.0 2024-09-25 06:36:45,738 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=684614.0, ans=0.125 2024-09-25 06:36:49,858 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.47 vs. limit=5.0 2024-09-25 06:36:55,481 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.min_positive, batch_count=684614.0, ans=0.05 2024-09-25 06:37:00,104 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.092e+02 1.274e+02 1.355e+02 1.436e+02 2.221e+02, threshold=2.709e+02, percent-clipped=0.0 2024-09-25 06:37:02,115 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=684660.6666666666, ans=0.125 2024-09-25 06:37:05,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=684660.6666666666, ans=0.09899494936611666 2024-09-25 06:37:10,232 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=684660.6666666666, ans=0.125 2024-09-25 06:37:15,236 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 06:37:59,908 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 06:38:05,958 INFO [train.py:1198] (2/4) Epoch 38, batch 2600, loss[loss=0.1678, ctc_loss=0.1045, cr_loss=0.3166, over 17257.00 frames. ], tot_loss[loss=0.1921, ctc_loss=0.1239, cr_loss=0.3412, over 3361115.77 frames. ], batch size: 42, lr: 3.12e-03, grad_scale: 32.0 2024-09-25 06:38:11,456 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.82 vs. limit=22.5 2024-09-25 06:38:21,502 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=3.78 vs. limit=15.0 2024-09-25 06:38:23,878 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=684894.0, ans=0.1 2024-09-25 06:38:32,058 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=684894.0, ans=0.125 2024-09-25 06:39:03,432 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=684987.3333333334, ans=0.125 2024-09-25 06:39:08,973 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.98 vs. limit=6.0 2024-09-25 06:39:31,217 INFO [train.py:1198] (2/4) Epoch 38, batch 2650, loss[loss=0.2339, ctc_loss=0.1549, cr_loss=0.3947, over 16581.00 frames. ], tot_loss[loss=0.1936, ctc_loss=0.1251, cr_loss=0.3427, over 3349231.66 frames. ], batch size: 66, lr: 3.12e-03, grad_scale: 32.0 2024-09-25 06:39:47,542 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=685127.3333333334, ans=0.125 2024-09-25 06:39:48,718 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.144e+02 1.311e+02 1.384e+02 1.483e+02 1.840e+02, threshold=2.769e+02, percent-clipped=0.0 2024-09-25 06:40:18,656 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=685174.0, ans=0.0 2024-09-25 06:40:53,182 INFO [train.py:1198] (2/4) Epoch 38, batch 2700, loss[loss=0.1999, ctc_loss=0.1276, cr_loss=0.3613, over 17291.00 frames. ], tot_loss[loss=0.1932, ctc_loss=0.1248, cr_loss=0.3418, over 3352179.75 frames. ], batch size: 46, lr: 3.12e-03, grad_scale: 32.0 2024-09-25 06:40:56,673 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.min_positive, batch_count=685314.0, ans=0.025 2024-09-25 06:41:30,422 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=685407.3333333334, ans=0.125 2024-09-25 06:41:53,056 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=9.13 vs. limit=15.0 2024-09-25 06:41:54,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=685454.0, ans=0.0 2024-09-25 06:42:01,304 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.32 vs. limit=10.0 2024-09-25 06:42:12,892 INFO [train.py:1198] (2/4) Epoch 38, batch 2750, loss[loss=0.171, ctc_loss=0.1076, cr_loss=0.3166, over 17126.00 frames. ], tot_loss[loss=0.1933, ctc_loss=0.1249, cr_loss=0.3419, over 3334450.17 frames. ], batch size: 40, lr: 3.12e-03, grad_scale: 16.0 2024-09-25 06:42:32,106 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.084e+02 1.260e+02 1.351e+02 1.429e+02 2.193e+02, threshold=2.703e+02, percent-clipped=0.0 2024-09-25 06:42:52,478 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=685640.6666666666, ans=0.025 2024-09-25 06:43:19,510 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=685734.0, ans=0.125 2024-09-25 06:43:35,230 INFO [train.py:1198] (2/4) Epoch 38, batch 2800, loss[loss=0.1833, ctc_loss=0.1171, cr_loss=0.3312, over 17158.00 frames. ], tot_loss[loss=0.1931, ctc_loss=0.1247, cr_loss=0.3419, over 3347353.56 frames. ], batch size: 45, lr: 3.12e-03, grad_scale: 32.0 2024-09-25 06:43:52,319 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.78 vs. limit=12.0 2024-09-25 06:44:11,795 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=685874.0, ans=0.0 2024-09-25 06:44:36,701 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 06:44:56,930 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=685967.3333333334, ans=0.0 2024-09-25 06:44:58,797 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=685967.3333333334, ans=0.125 2024-09-25 06:45:03,154 INFO [train.py:1198] (2/4) Epoch 38, batch 2850, loss[loss=0.2192, ctc_loss=0.1457, cr_loss=0.3679, over 16787.00 frames. ], tot_loss[loss=0.1928, ctc_loss=0.1246, cr_loss=0.3413, over 3345876.21 frames. ], batch size: 61, lr: 3.12e-03, grad_scale: 32.0 2024-09-25 06:45:22,309 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.153e+02 1.271e+02 1.362e+02 1.479e+02 2.279e+02, threshold=2.724e+02, percent-clipped=0.0 2024-09-25 06:45:27,968 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.72 vs. limit=6.0 2024-09-25 06:45:29,578 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.06 vs. limit=15.0 2024-09-25 06:45:32,636 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=11.90 vs. limit=15.0 2024-09-25 06:45:36,018 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.whiten.whitening_limit, batch_count=686107.3333333334, ans=15.0 2024-09-25 06:45:38,774 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten.whitening_limit, batch_count=686107.3333333334, ans=15.0 2024-09-25 06:45:46,518 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=686107.3333333334, ans=0.0 2024-09-25 06:46:13,704 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=686200.6666666666, ans=0.2 2024-09-25 06:46:16,898 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=686200.6666666666, ans=0.125 2024-09-25 06:46:23,066 INFO [train.py:1198] (2/4) Epoch 38, batch 2900, loss[loss=0.2222, ctc_loss=0.1459, cr_loss=0.3817, over 17213.00 frames. ], tot_loss[loss=0.1925, ctc_loss=0.1243, cr_loss=0.3409, over 3351209.52 frames. ], batch size: 55, lr: 3.12e-03, grad_scale: 32.0 2024-09-25 06:46:37,674 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=686294.0, ans=0.0 2024-09-25 06:46:52,852 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.81 vs. limit=6.0 2024-09-25 06:47:10,979 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=686387.3333333334, ans=0.125 2024-09-25 06:47:25,512 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=686434.0, ans=0.0 2024-09-25 06:47:25,549 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=686434.0, ans=0.0 2024-09-25 06:47:42,603 INFO [train.py:1198] (2/4) Epoch 38, batch 2950, loss[loss=0.1724, ctc_loss=0.1095, cr_loss=0.3145, over 17077.00 frames. ], tot_loss[loss=0.1927, ctc_loss=0.1244, cr_loss=0.3416, over 3339126.72 frames. ], batch size: 46, lr: 3.12e-03, grad_scale: 32.0 2024-09-25 06:47:58,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=686480.6666666666, ans=0.0 2024-09-25 06:48:04,470 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.143e+02 1.305e+02 1.376e+02 1.477e+02 2.268e+02, threshold=2.752e+02, percent-clipped=0.0 2024-09-25 06:48:12,646 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=686527.3333333334, ans=0.07 2024-09-25 06:48:38,148 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=686620.6666666666, ans=0.125 2024-09-25 06:48:44,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=686620.6666666666, ans=0.0 2024-09-25 06:48:46,690 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=8.55 vs. limit=15.0 2024-09-25 06:49:04,347 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=686667.3333333334, ans=0.0 2024-09-25 06:49:07,363 INFO [train.py:1198] (2/4) Epoch 38, batch 3000, loss[loss=0.197, ctc_loss=0.1273, cr_loss=0.3486, over 17228.00 frames. ], tot_loss[loss=0.1923, ctc_loss=0.1241, cr_loss=0.3413, over 3354198.05 frames. ], batch size: 50, lr: 3.12e-03, grad_scale: 16.0 2024-09-25 06:49:07,364 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 06:49:22,912 INFO [train.py:1230] (2/4) Epoch 38, validation: loss=0.03571, ctc_loss=0.03571, cr_loss=9.665e-15, over 944034.00 frames. 2024-09-25 06:49:22,913 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 06:49:38,363 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.73 vs. limit=6.0 2024-09-25 06:49:53,149 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=686807.3333333334, ans=0.125 2024-09-25 06:50:09,507 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=4.73 vs. limit=15.0 2024-09-25 06:50:17,880 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=686854.0, ans=0.0 2024-09-25 06:50:44,315 INFO [train.py:1198] (2/4) Epoch 38, batch 3050, loss[loss=0.1709, ctc_loss=0.1069, cr_loss=0.3201, over 16705.00 frames. ], tot_loss[loss=0.192, ctc_loss=0.1238, cr_loss=0.3407, over 3351419.83 frames. ], batch size: 37, lr: 3.12e-03, grad_scale: 16.0 2024-09-25 06:50:47,656 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=686947.3333333334, ans=0.1 2024-09-25 06:51:01,336 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=686994.0, ans=0.125 2024-09-25 06:51:04,197 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.171e+02 1.282e+02 1.358e+02 1.470e+02 1.835e+02, threshold=2.717e+02, percent-clipped=0.0 2024-09-25 06:51:13,827 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=687040.6666666666, ans=0.125 2024-09-25 06:51:32,294 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=687087.3333333334, ans=0.0 2024-09-25 06:52:01,989 INFO [train.py:1198] (2/4) Epoch 38, batch 3100, loss[loss=0.1623, ctc_loss=0.1028, cr_loss=0.2972, over 17098.00 frames. ], tot_loss[loss=0.1918, ctc_loss=0.1237, cr_loss=0.3405, over 3355771.83 frames. ], batch size: 43, lr: 3.11e-03, grad_scale: 16.0 2024-09-25 06:52:20,558 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.32 vs. limit=15.0 2024-09-25 06:52:29,592 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=12.83 vs. limit=15.0 2024-09-25 06:52:43,258 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=687274.0, ans=0.0 2024-09-25 06:53:00,447 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=687320.6666666666, ans=0.125 2024-09-25 06:53:01,817 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=687320.6666666666, ans=0.0 2024-09-25 06:53:03,979 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten.whitening_limit, batch_count=687367.3333333334, ans=22.5 2024-09-25 06:53:08,067 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=687367.3333333334, ans=0.125 2024-09-25 06:53:18,818 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=687414.0, ans=0.0 2024-09-25 06:53:20,024 INFO [train.py:1198] (2/4) Epoch 38, batch 3150, loss[loss=0.2319, ctc_loss=0.1539, cr_loss=0.3901, over 16717.00 frames. ], tot_loss[loss=0.192, ctc_loss=0.124, cr_loss=0.3398, over 3343491.75 frames. ], batch size: 61, lr: 3.11e-03, grad_scale: 16.0 2024-09-25 06:53:40,301 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.158e+02 1.270e+02 1.356e+02 1.474e+02 1.773e+02, threshold=2.711e+02, percent-clipped=0.0 2024-09-25 06:53:46,607 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=687460.6666666666, ans=0.125 2024-09-25 06:54:32,626 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.69 vs. limit=6.0 2024-09-25 06:54:37,787 INFO [train.py:1198] (2/4) Epoch 38, batch 3200, loss[loss=0.1988, ctc_loss=0.1269, cr_loss=0.3594, over 17024.00 frames. ], tot_loss[loss=0.1922, ctc_loss=0.1241, cr_loss=0.3403, over 3347427.06 frames. ], batch size: 44, lr: 3.11e-03, grad_scale: 32.0 2024-09-25 06:55:26,805 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.00 vs. limit=6.0 2024-09-25 06:55:56,022 INFO [train.py:1198] (2/4) Epoch 38, batch 3250, loss[loss=0.1514, ctc_loss=0.0946, cr_loss=0.284, over 17102.00 frames. ], tot_loss[loss=0.1916, ctc_loss=0.1236, cr_loss=0.3399, over 3355810.65 frames. ], batch size: 40, lr: 3.11e-03, grad_scale: 32.0 2024-09-25 06:56:02,805 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=687880.6666666666, ans=0.125 2024-09-25 06:56:17,749 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.130e+02 1.275e+02 1.365e+02 1.473e+02 2.154e+02, threshold=2.729e+02, percent-clipped=0.0 2024-09-25 06:56:25,956 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=687974.0, ans=0.125 2024-09-25 06:56:47,760 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=688020.6666666666, ans=0.125 2024-09-25 06:57:15,284 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.81 vs. limit=15.0 2024-09-25 06:57:16,214 INFO [train.py:1198] (2/4) Epoch 38, batch 3300, loss[loss=0.1621, ctc_loss=0.1052, cr_loss=0.2846, over 17178.00 frames. ], tot_loss[loss=0.1919, ctc_loss=0.1238, cr_loss=0.3406, over 3361328.30 frames. ], batch size: 41, lr: 3.11e-03, grad_scale: 16.0 2024-09-25 06:57:16,459 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=688114.0, ans=0.0 2024-09-25 06:57:31,127 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=12.60 vs. limit=22.5 2024-09-25 06:57:32,387 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=6.61 vs. limit=15.0 2024-09-25 06:57:41,215 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=688160.6666666666, ans=0.0 2024-09-25 06:57:49,048 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=688207.3333333334, ans=0.0 2024-09-25 06:57:49,080 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=688207.3333333334, ans=0.125 2024-09-25 06:57:59,798 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=688207.3333333334, ans=0.1 2024-09-25 06:58:09,210 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=688254.0, ans=0.125 2024-09-25 06:58:12,312 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=688254.0, ans=0.0 2024-09-25 06:58:14,014 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=688254.0, ans=0.125 2024-09-25 06:58:20,275 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=688300.6666666666, ans=0.125 2024-09-25 06:58:23,701 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=9.45 vs. limit=15.0 2024-09-25 06:58:28,401 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.40 vs. limit=15.0 2024-09-25 06:58:33,935 INFO [train.py:1198] (2/4) Epoch 38, batch 3350, loss[loss=0.1909, ctc_loss=0.1221, cr_loss=0.3438, over 17301.00 frames. ], tot_loss[loss=0.1917, ctc_loss=0.1236, cr_loss=0.3404, over 3364744.54 frames. ], batch size: 46, lr: 3.11e-03, grad_scale: 16.0 2024-09-25 06:58:43,581 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=688347.3333333334, ans=0.1 2024-09-25 06:58:55,604 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.109e+02 1.280e+02 1.359e+02 1.507e+02 2.410e+02, threshold=2.719e+02, percent-clipped=0.0 2024-09-25 06:59:03,206 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.14 vs. limit=10.0 2024-09-25 06:59:19,034 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.47 vs. limit=10.0 2024-09-25 06:59:48,613 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=688534.0, ans=0.1 2024-09-25 06:59:56,003 INFO [train.py:1198] (2/4) Epoch 38, batch 3400, loss[loss=0.1817, ctc_loss=0.1155, cr_loss=0.3311, over 17161.00 frames. ], tot_loss[loss=0.1918, ctc_loss=0.1237, cr_loss=0.3406, over 3366990.59 frames. ], batch size: 45, lr: 3.11e-03, grad_scale: 16.0 2024-09-25 07:00:10,486 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=688627.3333333334, ans=0.0 2024-09-25 07:00:10,486 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=688627.3333333334, ans=0.1 2024-09-25 07:00:26,248 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=688674.0, ans=0.07 2024-09-25 07:00:42,190 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.95 vs. limit=6.0 2024-09-25 07:01:16,601 INFO [train.py:1198] (2/4) Epoch 38, batch 3450, loss[loss=0.2331, ctc_loss=0.1529, cr_loss=0.4008, over 17070.00 frames. ], tot_loss[loss=0.1927, ctc_loss=0.1243, cr_loss=0.3419, over 3353767.13 frames. ], batch size: 52, lr: 3.11e-03, grad_scale: 16.0 2024-09-25 07:01:23,116 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=688814.0, ans=0.0 2024-09-25 07:01:28,707 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=6.02 vs. limit=15.0 2024-09-25 07:01:29,276 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=688814.0, ans=0.1 2024-09-25 07:01:35,710 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=688860.6666666666, ans=0.1 2024-09-25 07:01:37,712 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=14.25 vs. limit=22.5 2024-09-25 07:01:40,121 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.144e+02 1.283e+02 1.390e+02 1.486e+02 2.473e+02, threshold=2.781e+02, percent-clipped=0.0 2024-09-25 07:01:45,664 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=4.88 vs. limit=12.0 2024-09-25 07:01:53,033 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=688907.3333333334, ans=0.04949747468305833 2024-09-25 07:02:02,434 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=688954.0, ans=0.125 2024-09-25 07:02:35,150 INFO [train.py:1198] (2/4) Epoch 38, batch 3500, loss[loss=0.2323, ctc_loss=0.1522, cr_loss=0.4006, over 16691.00 frames. ], tot_loss[loss=0.1917, ctc_loss=0.1237, cr_loss=0.3401, over 3354986.17 frames. ], batch size: 66, lr: 3.11e-03, grad_scale: 8.0 2024-09-25 07:02:37,335 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.79 vs. limit=10.0 2024-09-25 07:02:45,590 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn2.whiten, num_groups=1, num_channels=192, metric=8.53 vs. limit=22.5 2024-09-25 07:02:49,423 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=689094.0, ans=0.125 2024-09-25 07:02:52,617 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=689094.0, ans=0.0 2024-09-25 07:02:59,613 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.whiten, num_groups=1, num_channels=192, metric=4.99 vs. limit=12.0 2024-09-25 07:03:06,929 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.whiten.whitening_limit, batch_count=689140.6666666666, ans=15.0 2024-09-25 07:03:30,159 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=689187.3333333334, ans=0.125 2024-09-25 07:03:53,030 INFO [train.py:1198] (2/4) Epoch 38, batch 3550, loss[loss=0.2156, ctc_loss=0.1403, cr_loss=0.3765, over 16409.00 frames. ], tot_loss[loss=0.1915, ctc_loss=0.1236, cr_loss=0.3398, over 3365808.03 frames. ], batch size: 66, lr: 3.11e-03, grad_scale: 8.0 2024-09-25 07:04:16,679 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.115e+02 1.303e+02 1.383e+02 1.456e+02 2.390e+02, threshold=2.765e+02, percent-clipped=0.0 2024-09-25 07:04:31,841 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=8.19 vs. limit=15.0 2024-09-25 07:04:57,084 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=689467.3333333334, ans=0.1 2024-09-25 07:05:02,433 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.38 vs. limit=22.5 2024-09-25 07:05:07,263 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.06 vs. limit=15.0 2024-09-25 07:05:11,163 INFO [train.py:1198] (2/4) Epoch 38, batch 3600, loss[loss=0.1856, ctc_loss=0.1216, cr_loss=0.3199, over 17011.00 frames. ], tot_loss[loss=0.1916, ctc_loss=0.1236, cr_loss=0.3401, over 3361910.86 frames. ], batch size: 51, lr: 3.11e-03, grad_scale: 16.0 2024-09-25 07:05:30,427 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.16 vs. limit=15.0 2024-09-25 07:05:36,313 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=689560.6666666666, ans=0.0 2024-09-25 07:06:00,046 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=689654.0, ans=0.125 2024-09-25 07:06:11,072 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=689654.0, ans=0.025 2024-09-25 07:06:23,582 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=689700.6666666666, ans=0.125 2024-09-25 07:06:29,867 INFO [train.py:1198] (2/4) Epoch 38, batch 3650, loss[loss=0.1759, ctc_loss=0.1121, cr_loss=0.3191, over 17302.00 frames. ], tot_loss[loss=0.1916, ctc_loss=0.1237, cr_loss=0.3396, over 3359279.21 frames. ], batch size: 42, lr: 3.11e-03, grad_scale: 16.0 2024-09-25 07:06:30,163 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=689747.3333333334, ans=0.125 2024-09-25 07:06:52,496 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=689794.0, ans=0.0 2024-09-25 07:06:55,349 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.248e+02 1.330e+02 1.434e+02 2.019e+02, threshold=2.659e+02, percent-clipped=0.0 2024-09-25 07:07:19,481 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=689887.3333333334, ans=0.125 2024-09-25 07:07:50,475 INFO [train.py:1198] (2/4) Epoch 38, batch 3700, loss[loss=0.1659, ctc_loss=0.1067, cr_loss=0.2961, over 16381.00 frames. ], tot_loss[loss=0.1922, ctc_loss=0.1241, cr_loss=0.3403, over 3359054.30 frames. ], batch size: 36, lr: 3.11e-03, grad_scale: 16.0 2024-09-25 07:07:50,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=689980.6666666666, ans=0.0 2024-09-25 07:08:03,178 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=3.85 vs. limit=10.0 2024-09-25 07:08:21,317 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.max_abs, batch_count=690074.0, ans=10.0 2024-09-25 07:08:40,072 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.26 vs. limit=10.0 2024-09-25 07:08:55,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=690167.3333333334, ans=0.125 2024-09-25 07:09:10,773 INFO [train.py:1198] (2/4) Epoch 38, batch 3750, loss[loss=0.1721, ctc_loss=0.1085, cr_loss=0.3181, over 17251.00 frames. ], tot_loss[loss=0.1921, ctc_loss=0.124, cr_loss=0.3407, over 3363872.69 frames. ], batch size: 44, lr: 3.11e-03, grad_scale: 16.0 2024-09-25 07:09:13,256 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=7.79 vs. limit=15.0 2024-09-25 07:09:34,562 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.279e+02 1.362e+02 1.449e+02 2.293e+02, threshold=2.725e+02, percent-clipped=0.0 2024-09-25 07:09:34,781 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=690260.6666666666, ans=10.0 2024-09-25 07:09:36,443 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=690260.6666666666, ans=0.125 2024-09-25 07:09:45,793 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=7.67 vs. limit=15.0 2024-09-25 07:09:49,969 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=690307.3333333334, ans=0.125 2024-09-25 07:09:57,931 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=690354.0, ans=0.1 2024-09-25 07:10:12,159 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=690354.0, ans=0.0 2024-09-25 07:10:12,189 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=690354.0, ans=0.1 2024-09-25 07:10:30,980 INFO [train.py:1198] (2/4) Epoch 38, batch 3800, loss[loss=0.1877, ctc_loss=0.1213, cr_loss=0.3316, over 16929.00 frames. ], tot_loss[loss=0.1932, ctc_loss=0.1248, cr_loss=0.342, over 3339736.97 frames. ], batch size: 58, lr: 3.11e-03, grad_scale: 16.0 2024-09-25 07:10:36,349 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=9.47 vs. limit=12.0 2024-09-25 07:10:40,851 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=690447.3333333334, ans=0.07 2024-09-25 07:10:51,401 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=690494.0, ans=0.0 2024-09-25 07:10:52,787 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=690494.0, ans=0.0 2024-09-25 07:10:59,095 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.max_abs, batch_count=690494.0, ans=10.0 2024-09-25 07:11:00,698 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=690494.0, ans=0.125 2024-09-25 07:11:21,811 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=5.05 vs. limit=10.0 2024-09-25 07:11:26,614 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=5.22 vs. limit=10.0 2024-09-25 07:11:33,954 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=15.49 vs. limit=15.0 2024-09-25 07:11:35,106 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=690634.0, ans=0.0 2024-09-25 07:11:39,771 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=690634.0, ans=0.125 2024-09-25 07:11:52,446 INFO [train.py:1198] (2/4) Epoch 38, batch 3850, loss[loss=0.2099, ctc_loss=0.137, cr_loss=0.3645, over 15172.00 frames. ], tot_loss[loss=0.1952, ctc_loss=0.1265, cr_loss=0.3434, over 3256177.29 frames. ], batch size: 89, lr: 3.11e-03, grad_scale: 16.0 2024-09-25 07:11:54,295 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=690680.6666666666, ans=0.2 2024-09-25 07:12:11,051 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=690727.3333333334, ans=0.125 2024-09-25 07:12:15,185 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.166e+02 1.374e+02 1.486e+02 1.640e+02 2.118e+02, threshold=2.972e+02, percent-clipped=0.0 2024-09-25 07:12:17,001 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=690727.3333333334, ans=0.125 2024-09-25 07:12:33,717 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.max_abs, batch_count=690774.0, ans=10.0 2024-09-25 07:12:37,311 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=13.67 vs. limit=15.0 2024-09-25 07:12:43,264 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=690820.6666666666, ans=0.125 2024-09-25 07:12:46,152 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=690820.6666666666, ans=0.5 2024-09-25 07:12:47,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=690820.6666666666, ans=0.0 2024-09-25 07:12:49,243 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=690820.6666666666, ans=0.0 2024-09-25 07:12:55,226 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=690867.3333333334, ans=0.125 2024-09-25 07:13:50,137 INFO [train.py:1198] (2/4) Epoch 39, batch 0, loss[loss=0.1908, ctc_loss=0.1223, cr_loss=0.3425, over 17108.00 frames. ], tot_loss[loss=0.1908, ctc_loss=0.1223, cr_loss=0.3425, over 17108.00 frames. ], batch size: 49, lr: 3.07e-03, grad_scale: 32.0 2024-09-25 07:13:50,138 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 07:13:57,691 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.3.encoder.layers.2.self_attn_weights, attn_weights_entropy = tensor([2.1231, 2.3952, 2.3994, 2.2987, 2.2654, 2.2191, 2.3649, 2.4109], device='cuda:2') 2024-09-25 07:14:06,120 INFO [train.py:1230] (2/4) Epoch 39, validation: loss=0.03529, ctc_loss=0.03529, cr_loss=1.033e-14, over 944034.00 frames. 2024-09-25 07:14:06,121 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 07:14:25,164 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=690942.0, ans=0.125 2024-09-25 07:14:28,782 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.47 vs. limit=12.0 2024-09-25 07:14:49,140 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=690988.6666666666, ans=0.2 2024-09-25 07:14:54,067 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=690988.6666666666, ans=0.125 2024-09-25 07:15:00,563 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=691035.3333333334, ans=0.125 2024-09-25 07:15:28,965 INFO [train.py:1198] (2/4) Epoch 39, batch 50, loss[loss=0.2212, ctc_loss=0.1447, cr_loss=0.3825, over 17011.00 frames. ], tot_loss[loss=0.1949, ctc_loss=0.1263, cr_loss=0.3429, over 752104.35 frames. ], batch size: 51, lr: 3.06e-03, grad_scale: 32.0 2024-09-25 07:15:54,240 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.67 vs. limit=15.0 2024-09-25 07:15:59,580 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.081e+02 1.279e+02 1.417e+02 1.630e+02 3.403e+02, threshold=2.834e+02, percent-clipped=1.0 2024-09-25 07:16:01,529 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=691222.0, ans=0.0 2024-09-25 07:16:22,138 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=691268.6666666666, ans=0.125 2024-09-25 07:16:22,874 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=4.00 vs. limit=15.0 2024-09-25 07:16:44,487 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=691315.3333333334, ans=0.1 2024-09-25 07:16:52,217 INFO [train.py:1198] (2/4) Epoch 39, batch 100, loss[loss=0.2095, ctc_loss=0.1355, cr_loss=0.3699, over 17308.00 frames. ], tot_loss[loss=0.194, ctc_loss=0.1254, cr_loss=0.3428, over 1328127.47 frames. ], batch size: 46, lr: 3.06e-03, grad_scale: 32.0 2024-09-25 07:16:52,512 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=691362.0, ans=0.0 2024-09-25 07:16:58,910 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=691362.0, ans=0.1 2024-09-25 07:17:34,407 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=691455.3333333334, ans=0.0 2024-09-25 07:17:59,720 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=691548.6666666666, ans=0.125 2024-09-25 07:18:09,392 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=691548.6666666666, ans=0.0 2024-09-25 07:18:12,289 INFO [train.py:1198] (2/4) Epoch 39, batch 150, loss[loss=0.1507, ctc_loss=0.09651, cr_loss=0.2709, over 17077.00 frames. ], tot_loss[loss=0.1957, ctc_loss=0.1266, cr_loss=0.3456, over 1762193.52 frames. ], batch size: 43, lr: 3.06e-03, grad_scale: 32.0 2024-09-25 07:18:45,367 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.121e+02 1.277e+02 1.390e+02 1.504e+02 2.454e+02, threshold=2.780e+02, percent-clipped=0.0 2024-09-25 07:19:01,393 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=691688.6666666666, ans=0.0 2024-09-25 07:19:12,392 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 07:19:40,576 INFO [train.py:1198] (2/4) Epoch 39, batch 200, loss[loss=0.1796, ctc_loss=0.1159, cr_loss=0.3183, over 17005.00 frames. ], tot_loss[loss=0.1933, ctc_loss=0.1249, cr_loss=0.3422, over 2115224.21 frames. ], batch size: 56, lr: 3.06e-03, grad_scale: 32.0 2024-09-25 07:20:09,654 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=691875.3333333334, ans=0.125 2024-09-25 07:20:13,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=691922.0, ans=0.05 2024-09-25 07:20:32,239 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=691968.6666666666, ans=0.2 2024-09-25 07:20:43,176 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=692015.3333333334, ans=0.2 2024-09-25 07:20:52,714 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.max_abs, batch_count=692015.3333333334, ans=10.0 2024-09-25 07:21:00,250 INFO [train.py:1198] (2/4) Epoch 39, batch 250, loss[loss=0.2163, ctc_loss=0.1414, cr_loss=0.3743, over 17307.00 frames. ], tot_loss[loss=0.1945, ctc_loss=0.1258, cr_loss=0.3434, over 2379810.37 frames. ], batch size: 49, lr: 3.06e-03, grad_scale: 32.0 2024-09-25 07:21:17,978 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=692108.6666666666, ans=0.125 2024-09-25 07:21:33,649 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.089e+02 1.285e+02 1.349e+02 1.463e+02 2.685e+02, threshold=2.698e+02, percent-clipped=0.0 2024-09-25 07:21:40,494 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=692155.3333333334, ans=0.125 2024-09-25 07:21:41,960 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=692155.3333333334, ans=0.125 2024-09-25 07:21:43,525 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=692155.3333333334, ans=0.2 2024-09-25 07:21:51,514 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=692202.0, ans=0.0 2024-09-25 07:21:59,524 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=692202.0, ans=0.0 2024-09-25 07:22:03,092 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=9.05 vs. limit=15.0 2024-09-25 07:22:04,198 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=692202.0, ans=0.125 2024-09-25 07:22:05,876 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=692248.6666666666, ans=0.125 2024-09-25 07:22:22,920 INFO [train.py:1198] (2/4) Epoch 39, batch 300, loss[loss=0.1696, ctc_loss=0.1073, cr_loss=0.3115, over 17231.00 frames. ], tot_loss[loss=0.1927, ctc_loss=0.1244, cr_loss=0.3412, over 2606640.55 frames. ], batch size: 47, lr: 3.06e-03, grad_scale: 32.0 2024-09-25 07:22:33,010 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=692295.3333333334, ans=0.125 2024-09-25 07:22:34,570 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=692295.3333333334, ans=0.2 2024-09-25 07:22:43,981 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=692342.0, ans=0.2 2024-09-25 07:22:58,369 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.min_positive, batch_count=692388.6666666666, ans=0.05 2024-09-25 07:22:59,854 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=692388.6666666666, ans=0.125 2024-09-25 07:23:45,872 INFO [train.py:1198] (2/4) Epoch 39, batch 350, loss[loss=0.224, ctc_loss=0.1487, cr_loss=0.3765, over 16601.00 frames. ], tot_loss[loss=0.193, ctc_loss=0.1247, cr_loss=0.342, over 2780616.80 frames. ], batch size: 66, lr: 3.06e-03, grad_scale: 32.0 2024-09-25 07:24:10,742 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=692575.3333333334, ans=0.125 2024-09-25 07:24:21,675 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.269e+02 1.341e+02 1.429e+02 1.987e+02, threshold=2.681e+02, percent-clipped=0.0 2024-09-25 07:24:31,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=692622.0, ans=0.025 2024-09-25 07:24:42,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=692668.6666666666, ans=0.025 2024-09-25 07:25:10,803 INFO [train.py:1198] (2/4) Epoch 39, batch 400, loss[loss=0.1763, ctc_loss=0.1127, cr_loss=0.3178, over 17103.00 frames. ], tot_loss[loss=0.1928, ctc_loss=0.1244, cr_loss=0.3423, over 2913624.91 frames. ], batch size: 40, lr: 3.06e-03, grad_scale: 32.0 2024-09-25 07:25:58,136 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=15.98 vs. limit=22.5 2024-09-25 07:26:32,429 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=692995.3333333334, ans=0.2 2024-09-25 07:26:33,628 INFO [train.py:1198] (2/4) Epoch 39, batch 450, loss[loss=0.1782, ctc_loss=0.1143, cr_loss=0.3195, over 17032.00 frames. ], tot_loss[loss=0.1938, ctc_loss=0.125, cr_loss=0.3443, over 3013167.10 frames. ], batch size: 44, lr: 3.06e-03, grad_scale: 32.0 2024-09-25 07:26:35,662 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=692995.3333333334, ans=0.0 2024-09-25 07:26:48,222 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=693042.0, ans=0.0 2024-09-25 07:26:51,546 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=693042.0, ans=0.07 2024-09-25 07:27:03,898 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.286e+02 1.365e+02 1.440e+02 1.919e+02, threshold=2.731e+02, percent-clipped=0.0 2024-09-25 07:27:17,433 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.25 vs. limit=15.0 2024-09-25 07:27:33,020 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=693135.3333333334, ans=0.2 2024-09-25 07:27:53,306 INFO [train.py:1198] (2/4) Epoch 39, batch 500, loss[loss=0.2392, ctc_loss=0.1595, cr_loss=0.3987, over 15132.00 frames. ], tot_loss[loss=0.1934, ctc_loss=0.1247, cr_loss=0.3437, over 3087376.40 frames. ], batch size: 89, lr: 3.06e-03, grad_scale: 32.0 2024-09-25 07:28:03,425 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=693228.6666666666, ans=0.1 2024-09-25 07:28:13,128 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=693275.3333333334, ans=0.1 2024-09-25 07:28:19,462 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=693275.3333333334, ans=0.125 2024-09-25 07:29:03,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=693415.3333333334, ans=0.1 2024-09-25 07:29:17,164 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=693415.3333333334, ans=0.2 2024-09-25 07:29:21,520 INFO [train.py:1198] (2/4) Epoch 39, batch 550, loss[loss=0.1782, ctc_loss=0.1144, cr_loss=0.3187, over 17022.00 frames. ], tot_loss[loss=0.1931, ctc_loss=0.1245, cr_loss=0.3427, over 3150976.59 frames. ], batch size: 44, lr: 3.06e-03, grad_scale: 32.0 2024-09-25 07:29:23,697 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=693462.0, ans=0.125 2024-09-25 07:29:38,169 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=693508.6666666666, ans=10.0 2024-09-25 07:29:52,388 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.262e+02 1.347e+02 1.440e+02 2.072e+02, threshold=2.694e+02, percent-clipped=0.0 2024-09-25 07:30:05,502 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=693555.3333333334, ans=0.2 2024-09-25 07:30:10,449 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=693602.0, ans=0.2 2024-09-25 07:30:12,048 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=693602.0, ans=0.0 2024-09-25 07:30:42,145 INFO [train.py:1198] (2/4) Epoch 39, batch 600, loss[loss=0.1986, ctc_loss=0.1289, cr_loss=0.3486, over 17230.00 frames. ], tot_loss[loss=0.192, ctc_loss=0.1237, cr_loss=0.3414, over 3208409.84 frames. ], batch size: 50, lr: 3.06e-03, grad_scale: 32.0 2024-09-25 07:30:42,382 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=693695.3333333334, ans=0.2 2024-09-25 07:30:44,426 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=9.04 vs. limit=12.0 2024-09-25 07:30:58,719 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=693742.0, ans=0.125 2024-09-25 07:31:06,026 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=4.23 vs. limit=15.0 2024-09-25 07:31:06,780 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=693742.0, ans=0.125 2024-09-25 07:31:10,021 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=693742.0, ans=0.125 2024-09-25 07:31:14,789 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=693788.6666666666, ans=0.125 2024-09-25 07:31:25,124 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=693788.6666666666, ans=0.025 2024-09-25 07:31:34,722 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=693835.3333333334, ans=0.125 2024-09-25 07:32:04,861 INFO [train.py:1198] (2/4) Epoch 39, batch 650, loss[loss=0.1959, ctc_loss=0.1288, cr_loss=0.3357, over 17024.00 frames. ], tot_loss[loss=0.1919, ctc_loss=0.1237, cr_loss=0.3411, over 3243078.73 frames. ], batch size: 52, lr: 3.06e-03, grad_scale: 32.0 2024-09-25 07:32:35,162 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.084e+02 1.253e+02 1.371e+02 1.476e+02 2.121e+02, threshold=2.742e+02, percent-clipped=0.0 2024-09-25 07:32:46,651 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=694022.0, ans=0.1 2024-09-25 07:33:10,517 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=694115.3333333334, ans=0.0 2024-09-25 07:33:15,573 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=694115.3333333334, ans=0.2 2024-09-25 07:33:24,352 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.21 vs. limit=15.0 2024-09-25 07:33:24,769 INFO [train.py:1198] (2/4) Epoch 39, batch 700, loss[loss=0.2306, ctc_loss=0.1563, cr_loss=0.3715, over 11934.00 frames. ], tot_loss[loss=0.1908, ctc_loss=0.123, cr_loss=0.3391, over 3257546.74 frames. ], batch size: 123, lr: 3.06e-03, grad_scale: 32.0 2024-09-25 07:33:47,055 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=694208.6666666666, ans=0.1 2024-09-25 07:34:40,189 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=694348.6666666666, ans=0.2 2024-09-25 07:34:49,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=694348.6666666666, ans=0.125 2024-09-25 07:34:52,576 INFO [train.py:1198] (2/4) Epoch 39, batch 750, loss[loss=0.2313, ctc_loss=0.1548, cr_loss=0.3828, over 14935.00 frames. ], tot_loss[loss=0.192, ctc_loss=0.1238, cr_loss=0.3409, over 3283590.72 frames. ], batch size: 89, lr: 3.06e-03, grad_scale: 32.0 2024-09-25 07:35:05,893 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=694395.3333333334, ans=0.125 2024-09-25 07:35:23,324 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.271e+02 1.368e+02 1.469e+02 1.814e+02, threshold=2.736e+02, percent-clipped=0.0 2024-09-25 07:35:28,614 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=694488.6666666666, ans=0.025 2024-09-25 07:36:13,107 INFO [train.py:1198] (2/4) Epoch 39, batch 800, loss[loss=0.1953, ctc_loss=0.1289, cr_loss=0.3319, over 16993.00 frames. ], tot_loss[loss=0.1925, ctc_loss=0.1242, cr_loss=0.3413, over 3301469.15 frames. ], batch size: 51, lr: 3.06e-03, grad_scale: 32.0 2024-09-25 07:36:14,961 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=694628.6666666666, ans=0.125 2024-09-25 07:36:24,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=694628.6666666666, ans=0.125 2024-09-25 07:36:27,374 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=694628.6666666666, ans=0.125 2024-09-25 07:36:43,569 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=694675.3333333334, ans=0.125 2024-09-25 07:36:48,218 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=694722.0, ans=0.2 2024-09-25 07:36:51,577 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=10.24 vs. limit=12.0 2024-09-25 07:37:07,930 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=13.27 vs. limit=15.0 2024-09-25 07:37:10,576 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=694768.6666666666, ans=0.2 2024-09-25 07:37:15,330 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=694768.6666666666, ans=0.1 2024-09-25 07:37:17,323 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.83 vs. limit=10.0 2024-09-25 07:37:35,916 INFO [train.py:1198] (2/4) Epoch 39, batch 850, loss[loss=0.1739, ctc_loss=0.1107, cr_loss=0.3159, over 17068.00 frames. ], tot_loss[loss=0.1915, ctc_loss=0.1235, cr_loss=0.3399, over 3316304.01 frames. ], batch size: 49, lr: 3.06e-03, grad_scale: 32.0 2024-09-25 07:38:02,072 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=694908.6666666666, ans=0.0 2024-09-25 07:38:06,360 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.150e+02 1.258e+02 1.327e+02 1.456e+02 1.924e+02, threshold=2.653e+02, percent-clipped=0.0 2024-09-25 07:38:16,394 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=694955.3333333334, ans=0.1 2024-09-25 07:38:29,245 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=695002.0, ans=0.125 2024-09-25 07:38:30,068 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=6.60 vs. limit=15.0 2024-09-25 07:38:47,650 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=695048.6666666666, ans=0.125 2024-09-25 07:38:52,789 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=9.14 vs. limit=15.0 2024-09-25 07:39:01,739 INFO [train.py:1198] (2/4) Epoch 39, batch 900, loss[loss=0.2166, ctc_loss=0.1431, cr_loss=0.3671, over 16991.00 frames. ], tot_loss[loss=0.1911, ctc_loss=0.1233, cr_loss=0.3393, over 3316618.70 frames. ], batch size: 53, lr: 3.06e-03, grad_scale: 32.0 2024-09-25 07:39:10,946 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=695095.3333333334, ans=0.025 2024-09-25 07:39:15,782 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=695095.3333333334, ans=0.0 2024-09-25 07:39:18,938 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=695142.0, ans=0.125 2024-09-25 07:39:33,264 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=695142.0, ans=0.125 2024-09-25 07:39:34,831 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=695188.6666666666, ans=0.1 2024-09-25 07:39:57,636 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=695235.3333333334, ans=0.0 2024-09-25 07:39:57,680 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=695235.3333333334, ans=0.025 2024-09-25 07:39:59,953 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=16.31 vs. limit=22.5 2024-09-25 07:40:24,418 INFO [train.py:1198] (2/4) Epoch 39, batch 950, loss[loss=0.1671, ctc_loss=0.1056, cr_loss=0.3077, over 17099.00 frames. ], tot_loss[loss=0.1918, ctc_loss=0.1238, cr_loss=0.3402, over 3329709.83 frames. ], batch size: 43, lr: 3.06e-03, grad_scale: 32.0 2024-09-25 07:40:32,955 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=695328.6666666666, ans=0.0 2024-09-25 07:40:41,057 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=695375.3333333334, ans=0.2 2024-09-25 07:40:55,205 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.152e+02 1.302e+02 1.382e+02 1.480e+02 1.852e+02, threshold=2.764e+02, percent-clipped=0.0 2024-09-25 07:41:14,874 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=695468.6666666666, ans=0.125 2024-09-25 07:41:43,045 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=695515.3333333334, ans=0.125 2024-09-25 07:41:47,516 INFO [train.py:1198] (2/4) Epoch 39, batch 1000, loss[loss=0.1929, ctc_loss=0.1239, cr_loss=0.3453, over 17005.00 frames. ], tot_loss[loss=0.1912, ctc_loss=0.1233, cr_loss=0.3394, over 3327873.81 frames. ], batch size: 56, lr: 3.06e-03, grad_scale: 32.0 2024-09-25 07:42:20,625 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.41 vs. limit=6.0 2024-09-25 07:42:29,426 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=695655.3333333334, ans=0.1 2024-09-25 07:43:07,516 INFO [train.py:1198] (2/4) Epoch 39, batch 1050, loss[loss=0.219, ctc_loss=0.1445, cr_loss=0.3723, over 17020.00 frames. ], tot_loss[loss=0.1911, ctc_loss=0.1232, cr_loss=0.3397, over 3339919.47 frames. ], batch size: 52, lr: 3.05e-03, grad_scale: 32.0 2024-09-25 07:43:07,846 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=695795.3333333334, ans=0.0 2024-09-25 07:43:10,884 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=695795.3333333334, ans=0.025 2024-09-25 07:43:32,161 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.99 vs. limit=15.0 2024-09-25 07:43:40,129 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.305e+02 1.388e+02 1.496e+02 1.693e+02, threshold=2.775e+02, percent-clipped=0.0 2024-09-25 07:44:16,470 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=512, metric=3.21 vs. limit=15.0 2024-09-25 07:44:20,795 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=695982.0, ans=0.0 2024-09-25 07:44:34,381 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.79 vs. limit=6.0 2024-09-25 07:44:34,843 INFO [train.py:1198] (2/4) Epoch 39, batch 1100, loss[loss=0.1863, ctc_loss=0.1185, cr_loss=0.339, over 17203.00 frames. ], tot_loss[loss=0.1909, ctc_loss=0.123, cr_loss=0.3398, over 3340998.00 frames. ], batch size: 47, lr: 3.05e-03, grad_scale: 16.0 2024-09-25 07:44:52,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=696075.3333333334, ans=0.2 2024-09-25 07:45:12,022 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=8.68 vs. limit=15.0 2024-09-25 07:45:26,395 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=696168.6666666666, ans=0.125 2024-09-25 07:45:54,571 INFO [train.py:1198] (2/4) Epoch 39, batch 1150, loss[loss=0.1906, ctc_loss=0.1225, cr_loss=0.3402, over 17061.00 frames. ], tot_loss[loss=0.1929, ctc_loss=0.1244, cr_loss=0.3425, over 3334807.62 frames. ], batch size: 46, lr: 3.05e-03, grad_scale: 8.0 2024-09-25 07:45:56,623 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=696262.0, ans=0.0 2024-09-25 07:46:13,513 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.85 vs. limit=6.0 2024-09-25 07:46:30,825 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.295e+02 1.353e+02 1.453e+02 1.736e+02, threshold=2.706e+02, percent-clipped=0.0 2024-09-25 07:46:44,279 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.80 vs. limit=10.0 2024-09-25 07:47:07,503 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=696448.6666666666, ans=0.2 2024-09-25 07:47:16,819 INFO [train.py:1198] (2/4) Epoch 39, batch 1200, loss[loss=0.1813, ctc_loss=0.1156, cr_loss=0.3288, over 17027.00 frames. ], tot_loss[loss=0.1925, ctc_loss=0.1242, cr_loss=0.3416, over 3337668.50 frames. ], batch size: 39, lr: 3.05e-03, grad_scale: 16.0 2024-09-25 07:48:39,497 INFO [train.py:1198] (2/4) Epoch 39, batch 1250, loss[loss=0.1699, ctc_loss=0.1093, cr_loss=0.3032, over 16948.00 frames. ], tot_loss[loss=0.1923, ctc_loss=0.124, cr_loss=0.3413, over 3336971.77 frames. ], batch size: 42, lr: 3.05e-03, grad_scale: 16.0 2024-09-25 07:49:14,765 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=696822.0, ans=0.1 2024-09-25 07:49:17,452 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.275e+02 1.351e+02 1.497e+02 2.020e+02, threshold=2.702e+02, percent-clipped=0.0 2024-09-25 07:49:19,414 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=696822.0, ans=0.2 2024-09-25 07:49:39,837 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=696868.6666666666, ans=0.2 2024-09-25 07:50:03,593 INFO [train.py:1198] (2/4) Epoch 39, batch 1300, loss[loss=0.1783, ctc_loss=0.1107, cr_loss=0.3382, over 17100.00 frames. ], tot_loss[loss=0.1916, ctc_loss=0.1234, cr_loss=0.3406, over 3341307.79 frames. ], batch size: 49, lr: 3.05e-03, grad_scale: 16.0 2024-09-25 07:50:21,596 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=697008.6666666666, ans=0.0 2024-09-25 07:50:45,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=697055.3333333334, ans=0.2 2024-09-25 07:50:53,457 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=697102.0, ans=0.2 2024-09-25 07:51:26,343 INFO [train.py:1198] (2/4) Epoch 39, batch 1350, loss[loss=0.1906, ctc_loss=0.1218, cr_loss=0.3443, over 17213.00 frames. ], tot_loss[loss=0.1912, ctc_loss=0.1232, cr_loss=0.34, over 3341979.26 frames. ], batch size: 50, lr: 3.05e-03, grad_scale: 16.0 2024-09-25 07:51:31,833 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=697195.3333333334, ans=0.2 2024-09-25 07:51:48,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=697242.0, ans=0.0 2024-09-25 07:51:57,944 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.00 vs. limit=6.0 2024-09-25 07:52:00,575 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.264e+02 1.325e+02 1.430e+02 2.601e+02, threshold=2.650e+02, percent-clipped=0.0 2024-09-25 07:52:13,621 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=697335.3333333334, ans=0.2 2024-09-25 07:52:47,367 INFO [train.py:1198] (2/4) Epoch 39, batch 1400, loss[loss=0.1848, ctc_loss=0.1225, cr_loss=0.3117, over 16884.00 frames. ], tot_loss[loss=0.1913, ctc_loss=0.1233, cr_loss=0.34, over 3344843.16 frames. ], batch size: 58, lr: 3.05e-03, grad_scale: 16.0 2024-09-25 07:52:53,123 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn2.whiten, num_groups=1, num_channels=192, metric=8.59 vs. limit=22.5 2024-09-25 07:53:00,491 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer_ff2.min_abs, batch_count=697428.6666666666, ans=0.1 2024-09-25 07:53:26,528 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=697522.0, ans=0.025 2024-09-25 07:53:47,873 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=8.52 vs. limit=15.0 2024-09-25 07:53:59,317 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=697615.3333333334, ans=0.1 2024-09-25 07:54:09,931 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.13 vs. limit=15.0 2024-09-25 07:54:15,110 INFO [train.py:1198] (2/4) Epoch 39, batch 1450, loss[loss=0.2298, ctc_loss=0.1489, cr_loss=0.4045, over 16998.00 frames. ], tot_loss[loss=0.1907, ctc_loss=0.1229, cr_loss=0.3391, over 3349741.26 frames. ], batch size: 53, lr: 3.05e-03, grad_scale: 16.0 2024-09-25 07:54:17,675 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten.whitening_limit, batch_count=697662.0, ans=15.0 2024-09-25 07:54:28,771 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=8.78 vs. limit=15.0 2024-09-25 07:54:48,362 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.159e+02 1.285e+02 1.356e+02 1.493e+02 2.927e+02, threshold=2.712e+02, percent-clipped=2.0 2024-09-25 07:54:58,372 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=697755.3333333334, ans=0.125 2024-09-25 07:55:25,417 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=697848.6666666666, ans=0.035 2024-09-25 07:55:31,726 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=697848.6666666666, ans=0.125 2024-09-25 07:55:34,690 INFO [train.py:1198] (2/4) Epoch 39, batch 1500, loss[loss=0.1788, ctc_loss=0.1145, cr_loss=0.3213, over 17104.00 frames. ], tot_loss[loss=0.1916, ctc_loss=0.1235, cr_loss=0.3405, over 3357550.22 frames. ], batch size: 43, lr: 3.05e-03, grad_scale: 16.0 2024-09-25 07:56:02,612 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.45 vs. limit=22.5 2024-09-25 07:56:03,626 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=697942.0, ans=0.0 2024-09-25 07:56:05,163 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=697988.6666666666, ans=0.0 2024-09-25 07:56:15,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=697988.6666666666, ans=0.0 2024-09-25 07:56:33,147 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=698035.3333333334, ans=0.0 2024-09-25 07:56:57,034 INFO [train.py:1198] (2/4) Epoch 39, batch 1550, loss[loss=0.1805, ctc_loss=0.1147, cr_loss=0.329, over 17314.00 frames. ], tot_loss[loss=0.1911, ctc_loss=0.1232, cr_loss=0.3395, over 3358519.90 frames. ], batch size: 46, lr: 3.05e-03, grad_scale: 16.0 2024-09-25 07:57:04,040 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=698128.6666666666, ans=0.1 2024-09-25 07:57:30,847 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.098e+02 1.291e+02 1.368e+02 1.486e+02 2.492e+02, threshold=2.736e+02, percent-clipped=0.0 2024-09-25 07:57:54,644 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=698268.6666666666, ans=0.0 2024-09-25 07:58:06,218 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=698315.3333333334, ans=0.125 2024-09-25 07:58:14,508 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.60 vs. limit=15.0 2024-09-25 07:58:17,121 INFO [train.py:1198] (2/4) Epoch 39, batch 1600, loss[loss=0.2512, ctc_loss=0.1673, cr_loss=0.4197, over 16509.00 frames. ], tot_loss[loss=0.1923, ctc_loss=0.1241, cr_loss=0.3409, over 3354234.29 frames. ], batch size: 66, lr: 3.05e-03, grad_scale: 32.0 2024-09-25 07:58:49,656 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=698408.6666666666, ans=0.025 2024-09-25 07:59:44,295 INFO [train.py:1198] (2/4) Epoch 39, batch 1650, loss[loss=0.2093, ctc_loss=0.1366, cr_loss=0.3635, over 17019.00 frames. ], tot_loss[loss=0.1926, ctc_loss=0.1243, cr_loss=0.3416, over 3349139.78 frames. ], batch size: 52, lr: 3.05e-03, grad_scale: 32.0 2024-09-25 07:59:54,004 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=698595.3333333334, ans=0.2 2024-09-25 08:00:07,081 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=698642.0, ans=0.125 2024-09-25 08:00:08,702 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=698642.0, ans=0.125 2024-09-25 08:00:17,963 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.058e+02 1.287e+02 1.365e+02 1.434e+02 2.064e+02, threshold=2.730e+02, percent-clipped=0.0 2024-09-25 08:00:21,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=698688.6666666666, ans=0.125 2024-09-25 08:00:35,841 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=698735.3333333334, ans=0.125 2024-09-25 08:00:42,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=698735.3333333334, ans=0.2 2024-09-25 08:01:04,261 INFO [train.py:1198] (2/4) Epoch 39, batch 1700, loss[loss=0.217, ctc_loss=0.1403, cr_loss=0.3833, over 16893.00 frames. ], tot_loss[loss=0.1924, ctc_loss=0.1241, cr_loss=0.3413, over 3355009.52 frames. ], batch size: 58, lr: 3.05e-03, grad_scale: 32.0 2024-09-25 08:01:48,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=698922.0, ans=0.125 2024-09-25 08:01:58,027 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=698968.6666666666, ans=0.07 2024-09-25 08:02:03,282 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.61 vs. limit=15.0 2024-09-25 08:02:18,325 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=699015.3333333334, ans=0.0 2024-09-25 08:02:18,718 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.24 vs. limit=15.0 2024-09-25 08:02:25,987 INFO [train.py:1198] (2/4) Epoch 39, batch 1750, loss[loss=0.1863, ctc_loss=0.1196, cr_loss=0.3337, over 17096.00 frames. ], tot_loss[loss=0.1926, ctc_loss=0.1242, cr_loss=0.3417, over 3360530.08 frames. ], batch size: 49, lr: 3.05e-03, grad_scale: 32.0 2024-09-25 08:02:35,860 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=699062.0, ans=0.125 2024-09-25 08:02:43,772 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=699108.6666666666, ans=0.1 2024-09-25 08:03:01,042 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.161e+02 1.281e+02 1.369e+02 1.457e+02 2.012e+02, threshold=2.738e+02, percent-clipped=0.0 2024-09-25 08:03:16,152 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=10.83 vs. limit=12.0 2024-09-25 08:03:30,924 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=699248.6666666666, ans=0.0 2024-09-25 08:03:47,801 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=699248.6666666666, ans=0.0 2024-09-25 08:03:53,686 INFO [train.py:1198] (2/4) Epoch 39, batch 1800, loss[loss=0.1979, ctc_loss=0.1276, cr_loss=0.3514, over 17205.00 frames. ], tot_loss[loss=0.192, ctc_loss=0.1238, cr_loss=0.3408, over 3356939.44 frames. ], batch size: 47, lr: 3.05e-03, grad_scale: 16.0 2024-09-25 08:03:58,886 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=699295.3333333334, ans=0.1 2024-09-25 08:04:06,679 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=699295.3333333334, ans=0.0 2024-09-25 08:04:38,743 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer_na.min_abs, batch_count=699388.6666666666, ans=0.02 2024-09-25 08:04:40,355 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=699435.3333333334, ans=0.2 2024-09-25 08:04:51,596 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=699435.3333333334, ans=0.025 2024-09-25 08:05:02,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=699482.0, ans=0.0 2024-09-25 08:05:09,476 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=699482.0, ans=0.0 2024-09-25 08:05:13,794 INFO [train.py:1198] (2/4) Epoch 39, batch 1850, loss[loss=0.2042, ctc_loss=0.1353, cr_loss=0.3444, over 17021.00 frames. ], tot_loss[loss=0.1915, ctc_loss=0.1234, cr_loss=0.3405, over 3366323.93 frames. ], batch size: 56, lr: 3.05e-03, grad_scale: 16.0 2024-09-25 08:05:36,439 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=699575.3333333334, ans=0.0 2024-09-25 08:05:46,416 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=9.99 vs. limit=22.5 2024-09-25 08:05:48,762 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.097e+02 1.264e+02 1.367e+02 1.511e+02 2.303e+02, threshold=2.733e+02, percent-clipped=0.0 2024-09-25 08:06:01,145 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=12.43 vs. limit=15.0 2024-09-25 08:06:17,112 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=699668.6666666666, ans=0.0 2024-09-25 08:06:18,787 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=699715.3333333334, ans=0.125 2024-09-25 08:06:22,089 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=699715.3333333334, ans=0.125 2024-09-25 08:06:36,625 INFO [train.py:1198] (2/4) Epoch 39, batch 1900, loss[loss=0.1865, ctc_loss=0.1189, cr_loss=0.3379, over 17299.00 frames. ], tot_loss[loss=0.1923, ctc_loss=0.1239, cr_loss=0.3419, over 3354829.48 frames. ], batch size: 46, lr: 3.05e-03, grad_scale: 16.0 2024-09-25 08:06:36,879 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=699762.0, ans=0.125 2024-09-25 08:06:43,213 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=699762.0, ans=0.0 2024-09-25 08:06:44,799 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=699762.0, ans=0.1 2024-09-25 08:06:59,192 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=699808.6666666666, ans=0.125 2024-09-25 08:07:02,439 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer_na.min_abs, batch_count=699808.6666666666, ans=0.02 2024-09-25 08:07:07,268 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=699855.3333333334, ans=0.04949747468305833 2024-09-25 08:07:12,744 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=6.30 vs. limit=15.0 2024-09-25 08:07:18,212 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=699855.3333333334, ans=0.125 2024-09-25 08:07:38,609 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.29 vs. limit=5.0 2024-09-25 08:07:49,030 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=699948.6666666666, ans=0.125 2024-09-25 08:07:56,980 INFO [train.py:1198] (2/4) Epoch 39, batch 1950, loss[loss=0.2291, ctc_loss=0.1538, cr_loss=0.3766, over 16606.00 frames. ], tot_loss[loss=0.193, ctc_loss=0.1244, cr_loss=0.343, over 3350698.70 frames. ], batch size: 66, lr: 3.05e-03, grad_scale: 16.0 2024-09-25 08:08:21,911 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=700042.0, ans=0.125 2024-09-25 08:08:37,869 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.289e+02 1.367e+02 1.538e+02 1.984e+02, threshold=2.734e+02, percent-clipped=0.0 2024-09-25 08:08:38,271 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=700088.6666666666, ans=0.125 2024-09-25 08:08:39,829 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=700088.6666666666, ans=0.025 2024-09-25 08:09:20,922 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=700182.0, ans=0.125 2024-09-25 08:09:25,330 INFO [train.py:1198] (2/4) Epoch 39, batch 2000, loss[loss=0.155, ctc_loss=0.0974, cr_loss=0.2881, over 16677.00 frames. ], tot_loss[loss=0.1914, ctc_loss=0.1232, cr_loss=0.3409, over 3360305.09 frames. ], batch size: 37, lr: 3.04e-03, grad_scale: 32.0 2024-09-25 08:09:30,532 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=700228.6666666666, ans=0.125 2024-09-25 08:09:56,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=700322.0, ans=0.0 2024-09-25 08:09:57,711 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=700322.0, ans=0.125 2024-09-25 08:10:01,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=700322.0, ans=0.0 2024-09-25 08:10:12,181 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=700368.6666666666, ans=0.1 2024-09-25 08:10:13,578 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=700368.6666666666, ans=0.1 2024-09-25 08:10:21,643 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=700368.6666666666, ans=0.0 2024-09-25 08:10:28,011 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=700415.3333333334, ans=0.125 2024-09-25 08:10:45,493 INFO [train.py:1198] (2/4) Epoch 39, batch 2050, loss[loss=0.2132, ctc_loss=0.1389, cr_loss=0.3717, over 17040.00 frames. ], tot_loss[loss=0.192, ctc_loss=0.1237, cr_loss=0.3418, over 3366313.49 frames. ], batch size: 52, lr: 3.04e-03, grad_scale: 16.0 2024-09-25 08:10:57,926 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn2.whiten, num_groups=1, num_channels=192, metric=8.63 vs. limit=22.5 2024-09-25 08:11:25,132 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.300e+02 1.377e+02 1.457e+02 2.836e+02, threshold=2.753e+02, percent-clipped=1.0 2024-09-25 08:11:30,075 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=700555.3333333334, ans=0.0 2024-09-25 08:11:34,971 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=700602.0, ans=0.0 2024-09-25 08:11:34,977 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=700602.0, ans=0.125 2024-09-25 08:11:38,132 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=700602.0, ans=0.0 2024-09-25 08:11:48,125 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=13.56 vs. limit=22.5 2024-09-25 08:12:05,270 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=700648.6666666666, ans=0.0 2024-09-25 08:12:07,301 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.25 vs. limit=15.0 2024-09-25 08:12:08,227 INFO [train.py:1198] (2/4) Epoch 39, batch 2100, loss[loss=0.1973, ctc_loss=0.1281, cr_loss=0.3462, over 17079.00 frames. ], tot_loss[loss=0.1913, ctc_loss=0.1231, cr_loss=0.3407, over 3372697.94 frames. ], batch size: 43, lr: 3.04e-03, grad_scale: 16.0 2024-09-25 08:12:14,883 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=700695.3333333334, ans=0.125 2024-09-25 08:12:16,542 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=700695.3333333334, ans=0.125 2024-09-25 08:12:24,511 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.min_abs, batch_count=700742.0, ans=0.5 2024-09-25 08:12:35,491 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=700742.0, ans=0.125 2024-09-25 08:12:40,385 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=700788.6666666666, ans=0.0 2024-09-25 08:12:42,342 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.61 vs. limit=15.0 2024-09-25 08:12:50,109 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=700788.6666666666, ans=0.125 2024-09-25 08:12:54,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=700835.3333333334, ans=0.125 2024-09-25 08:13:00,100 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.46 vs. limit=10.0 2024-09-25 08:13:05,975 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=700835.3333333334, ans=0.2 2024-09-25 08:13:30,506 INFO [train.py:1198] (2/4) Epoch 39, batch 2150, loss[loss=0.2412, ctc_loss=0.1581, cr_loss=0.4153, over 15195.00 frames. ], tot_loss[loss=0.192, ctc_loss=0.1237, cr_loss=0.3416, over 3366803.96 frames. ], batch size: 89, lr: 3.04e-03, grad_scale: 16.0 2024-09-25 08:13:30,911 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=700928.6666666666, ans=0.0 2024-09-25 08:13:33,292 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.32 vs. limit=15.0 2024-09-25 08:13:37,399 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=700928.6666666666, ans=0.125 2024-09-25 08:13:50,754 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=700975.3333333334, ans=0.125 2024-09-25 08:14:10,019 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.289e+02 1.363e+02 1.447e+02 2.047e+02, threshold=2.727e+02, percent-clipped=0.0 2024-09-25 08:14:11,936 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=701022.0, ans=0.0 2024-09-25 08:14:16,706 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=701022.0, ans=0.1 2024-09-25 08:14:20,129 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=11.04 vs. limit=12.0 2024-09-25 08:14:40,686 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 08:14:42,181 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=701115.3333333334, ans=0.0 2024-09-25 08:14:53,283 INFO [train.py:1198] (2/4) Epoch 39, batch 2200, loss[loss=0.2251, ctc_loss=0.1483, cr_loss=0.3839, over 14979.00 frames. ], tot_loss[loss=0.192, ctc_loss=0.1238, cr_loss=0.3412, over 3366106.95 frames. ], batch size: 89, lr: 3.04e-03, grad_scale: 16.0 2024-09-25 08:14:57,373 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=3.56 vs. limit=6.0 2024-09-25 08:15:09,723 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=701208.6666666666, ans=0.2 2024-09-25 08:15:32,244 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=701255.3333333334, ans=0.1 2024-09-25 08:15:44,972 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=701302.0, ans=0.125 2024-09-25 08:15:46,553 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=701302.0, ans=0.125 2024-09-25 08:16:16,046 INFO [train.py:1198] (2/4) Epoch 39, batch 2250, loss[loss=0.2411, ctc_loss=0.1623, cr_loss=0.3943, over 14900.00 frames. ], tot_loss[loss=0.1917, ctc_loss=0.1237, cr_loss=0.3404, over 3355413.89 frames. ], batch size: 89, lr: 3.04e-03, grad_scale: 16.0 2024-09-25 08:16:19,617 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=701395.3333333334, ans=0.125 2024-09-25 08:16:29,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=701395.3333333334, ans=0.125 2024-09-25 08:16:32,370 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=701442.0, ans=0.0 2024-09-25 08:16:32,405 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=701442.0, ans=0.5 2024-09-25 08:16:33,007 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.41 vs. limit=10.0 2024-09-25 08:16:52,314 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.99 vs. limit=12.0 2024-09-25 08:16:52,926 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.277e+02 1.348e+02 1.481e+02 2.538e+02, threshold=2.695e+02, percent-clipped=0.0 2024-09-25 08:16:59,751 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys.whitening_limit, batch_count=701488.6666666666, ans=6.0 2024-09-25 08:17:21,731 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=701582.0, ans=0.0 2024-09-25 08:17:35,950 INFO [train.py:1198] (2/4) Epoch 39, batch 2300, loss[loss=0.2221, ctc_loss=0.1484, cr_loss=0.3686, over 12238.00 frames. ], tot_loss[loss=0.1929, ctc_loss=0.1246, cr_loss=0.3416, over 3317043.96 frames. ], batch size: 123, lr: 3.04e-03, grad_scale: 16.0 2024-09-25 08:17:38,184 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.25 vs. limit=22.5 2024-09-25 08:17:47,576 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=701628.6666666666, ans=0.125 2024-09-25 08:17:49,233 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=701628.6666666666, ans=0.1 2024-09-25 08:17:55,564 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=701675.3333333334, ans=0.035 2024-09-25 08:18:33,541 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.74 vs. limit=6.0 2024-09-25 08:18:45,569 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=4.47 vs. limit=12.0 2024-09-25 08:18:46,737 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=701815.3333333334, ans=0.1 2024-09-25 08:19:03,993 INFO [train.py:1198] (2/4) Epoch 39, batch 2350, loss[loss=0.1942, ctc_loss=0.1268, cr_loss=0.337, over 16157.00 frames. ], tot_loss[loss=0.1929, ctc_loss=0.1244, cr_loss=0.3421, over 3331514.86 frames. ], batch size: 74, lr: 3.04e-03, grad_scale: 16.0 2024-09-25 08:19:07,739 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=701862.0, ans=0.125 2024-09-25 08:19:12,388 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=701862.0, ans=0.2 2024-09-25 08:19:15,486 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=701862.0, ans=0.2 2024-09-25 08:19:33,087 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 08:19:40,703 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.145e+02 1.317e+02 1.394e+02 1.473e+02 1.777e+02, threshold=2.787e+02, percent-clipped=0.0 2024-09-25 08:19:45,753 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=701955.3333333334, ans=0.0 2024-09-25 08:19:47,370 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=701955.3333333334, ans=0.125 2024-09-25 08:20:11,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=702048.6666666666, ans=0.0 2024-09-25 08:20:23,819 INFO [train.py:1198] (2/4) Epoch 39, batch 2400, loss[loss=0.2122, ctc_loss=0.137, cr_loss=0.3762, over 16996.00 frames. ], tot_loss[loss=0.1928, ctc_loss=0.1244, cr_loss=0.3418, over 3325757.66 frames. ], batch size: 53, lr: 3.04e-03, grad_scale: 32.0 2024-09-25 08:20:46,324 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=702142.0, ans=0.125 2024-09-25 08:21:17,594 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=702235.3333333334, ans=0.125 2024-09-25 08:21:19,347 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.max_abs, batch_count=702235.3333333334, ans=10.0 2024-09-25 08:21:28,908 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=702282.0, ans=0.04949747468305833 2024-09-25 08:21:31,946 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=702282.0, ans=0.0 2024-09-25 08:21:45,938 INFO [train.py:1198] (2/4) Epoch 39, batch 2450, loss[loss=0.1988, ctc_loss=0.1288, cr_loss=0.3499, over 17019.00 frames. ], tot_loss[loss=0.1926, ctc_loss=0.1243, cr_loss=0.3416, over 3340006.27 frames. ], batch size: 56, lr: 3.04e-03, grad_scale: 32.0 2024-09-25 08:21:58,923 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 08:22:19,749 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=702422.0, ans=0.125 2024-09-25 08:22:21,292 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=702422.0, ans=0.2 2024-09-25 08:22:24,052 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.006e+02 1.314e+02 1.404e+02 1.496e+02 1.831e+02, threshold=2.808e+02, percent-clipped=0.0 2024-09-25 08:22:43,438 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=702468.6666666666, ans=0.125 2024-09-25 08:23:08,340 INFO [train.py:1198] (2/4) Epoch 39, batch 2500, loss[loss=0.1737, ctc_loss=0.1093, cr_loss=0.3219, over 16939.00 frames. ], tot_loss[loss=0.1927, ctc_loss=0.1243, cr_loss=0.342, over 3345808.07 frames. ], batch size: 42, lr: 3.04e-03, grad_scale: 16.0 2024-09-25 08:23:58,501 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=702655.3333333334, ans=0.125 2024-09-25 08:24:19,449 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=702748.6666666666, ans=0.1 2024-09-25 08:24:19,620 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=10.86 vs. limit=15.0 2024-09-25 08:24:24,446 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.30 vs. limit=15.0 2024-09-25 08:24:33,558 INFO [train.py:1198] (2/4) Epoch 39, batch 2550, loss[loss=0.2203, ctc_loss=0.1455, cr_loss=0.3743, over 15951.00 frames. ], tot_loss[loss=0.1925, ctc_loss=0.1241, cr_loss=0.3422, over 3342264.38 frames. ], batch size: 74, lr: 3.04e-03, grad_scale: 16.0 2024-09-25 08:24:33,829 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=702795.3333333334, ans=0.125 2024-09-25 08:24:43,745 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=702795.3333333334, ans=0.1 2024-09-25 08:24:57,906 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=702842.0, ans=0.2 2024-09-25 08:25:07,599 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=702888.6666666666, ans=0.0 2024-09-25 08:25:12,023 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.167e+02 1.300e+02 1.397e+02 1.510e+02 1.872e+02, threshold=2.794e+02, percent-clipped=0.0 2024-09-25 08:25:15,665 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=702888.6666666666, ans=0.125 2024-09-25 08:25:22,140 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=702935.3333333334, ans=0.125 2024-09-25 08:25:23,902 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=702935.3333333334, ans=0.0 2024-09-25 08:25:26,991 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=702935.3333333334, ans=0.125 2024-09-25 08:25:46,131 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=702982.0, ans=0.0 2024-09-25 08:25:56,179 INFO [train.py:1198] (2/4) Epoch 39, batch 2600, loss[loss=0.2071, ctc_loss=0.1353, cr_loss=0.3592, over 16868.00 frames. ], tot_loss[loss=0.1927, ctc_loss=0.1242, cr_loss=0.3422, over 3346184.86 frames. ], batch size: 58, lr: 3.04e-03, grad_scale: 16.0 2024-09-25 08:25:56,561 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=703028.6666666666, ans=0.1 2024-09-25 08:26:50,680 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=703168.6666666666, ans=0.1 2024-09-25 08:27:15,879 INFO [train.py:1198] (2/4) Epoch 39, batch 2650, loss[loss=0.1917, ctc_loss=0.1211, cr_loss=0.3528, over 17057.00 frames. ], tot_loss[loss=0.1933, ctc_loss=0.1247, cr_loss=0.3432, over 3352272.23 frames. ], batch size: 39, lr: 3.04e-03, grad_scale: 16.0 2024-09-25 08:27:17,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=703262.0, ans=0.0 2024-09-25 08:27:31,042 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=4.78 vs. limit=15.0 2024-09-25 08:27:53,518 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.290e+02 1.367e+02 1.451e+02 1.893e+02, threshold=2.734e+02, percent-clipped=0.0 2024-09-25 08:28:15,401 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=703402.0, ans=0.125 2024-09-25 08:28:23,253 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.29 vs. limit=15.0 2024-09-25 08:28:43,413 INFO [train.py:1198] (2/4) Epoch 39, batch 2700, loss[loss=0.2293, ctc_loss=0.1497, cr_loss=0.3984, over 16527.00 frames. ], tot_loss[loss=0.1935, ctc_loss=0.1248, cr_loss=0.3433, over 3348348.25 frames. ], batch size: 66, lr: 3.04e-03, grad_scale: 16.0 2024-09-25 08:28:45,076 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=703495.3333333334, ans=0.1 2024-09-25 08:28:50,057 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=703495.3333333334, ans=0.125 2024-09-25 08:28:56,443 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=703495.3333333334, ans=0.125 2024-09-25 08:29:10,620 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=703542.0, ans=0.125 2024-09-25 08:29:24,870 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=703588.6666666666, ans=0.125 2024-09-25 08:29:44,049 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=703635.3333333334, ans=0.04949747468305833 2024-09-25 08:29:50,217 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=703682.0, ans=0.1 2024-09-25 08:30:02,891 INFO [train.py:1198] (2/4) Epoch 39, batch 2750, loss[loss=0.2088, ctc_loss=0.1368, cr_loss=0.3596, over 16797.00 frames. ], tot_loss[loss=0.1943, ctc_loss=0.1255, cr_loss=0.3443, over 3357511.98 frames. ], batch size: 61, lr: 3.04e-03, grad_scale: 16.0 2024-09-25 08:30:03,141 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=703728.6666666666, ans=0.0 2024-09-25 08:30:24,325 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=703775.3333333334, ans=0.125 2024-09-25 08:30:25,852 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=703775.3333333334, ans=0.1 2024-09-25 08:30:27,654 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=703775.3333333334, ans=0.1 2024-09-25 08:30:39,214 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.85 vs. limit=15.0 2024-09-25 08:30:41,402 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.135e+02 1.292e+02 1.404e+02 1.488e+02 2.567e+02, threshold=2.807e+02, percent-clipped=0.0 2024-09-25 08:30:41,719 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=703822.0, ans=0.0 2024-09-25 08:30:57,003 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=703868.6666666666, ans=0.0 2024-09-25 08:30:59,312 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=13.60 vs. limit=22.5 2024-09-25 08:31:01,944 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=703868.6666666666, ans=0.1 2024-09-25 08:31:11,432 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=703915.3333333334, ans=0.0 2024-09-25 08:31:17,122 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.07 vs. limit=15.0 2024-09-25 08:31:25,414 INFO [train.py:1198] (2/4) Epoch 39, batch 2800, loss[loss=0.1785, ctc_loss=0.1145, cr_loss=0.3199, over 17029.00 frames. ], tot_loss[loss=0.1936, ctc_loss=0.1249, cr_loss=0.3437, over 3365884.77 frames. ], batch size: 39, lr: 3.04e-03, grad_scale: 32.0 2024-09-25 08:31:30,654 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=703962.0, ans=0.125 2024-09-25 08:31:40,464 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 08:31:46,698 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=704008.6666666666, ans=0.1 2024-09-25 08:32:25,124 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=704102.0, ans=0.125 2024-09-25 08:32:32,272 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=7.42 vs. limit=15.0 2024-09-25 08:32:44,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=704195.3333333334, ans=0.125 2024-09-25 08:32:45,469 INFO [train.py:1198] (2/4) Epoch 39, batch 2850, loss[loss=0.2054, ctc_loss=0.1336, cr_loss=0.3591, over 16995.00 frames. ], tot_loss[loss=0.194, ctc_loss=0.1252, cr_loss=0.3442, over 3366490.86 frames. ], batch size: 53, lr: 3.04e-03, grad_scale: 32.0 2024-09-25 08:33:07,813 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=11.21 vs. limit=15.0 2024-09-25 08:33:28,933 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=704288.6666666666, ans=0.1 2024-09-25 08:33:31,755 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.164e+02 1.313e+02 1.374e+02 1.493e+02 1.951e+02, threshold=2.749e+02, percent-clipped=0.0 2024-09-25 08:33:43,546 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 08:33:58,039 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=704382.0, ans=0.0 2024-09-25 08:33:59,058 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=9.12 vs. limit=15.0 2024-09-25 08:34:07,545 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=704382.0, ans=0.125 2024-09-25 08:34:13,710 INFO [train.py:1198] (2/4) Epoch 39, batch 2900, loss[loss=0.2131, ctc_loss=0.1396, cr_loss=0.3675, over 17313.00 frames. ], tot_loss[loss=0.1921, ctc_loss=0.1238, cr_loss=0.3418, over 3366014.27 frames. ], batch size: 51, lr: 3.04e-03, grad_scale: 32.0 2024-09-25 08:34:20,439 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=704428.6666666666, ans=0.125 2024-09-25 08:34:23,689 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=704428.6666666666, ans=0.05 2024-09-25 08:34:27,078 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=11.85 vs. limit=15.0 2024-09-25 08:34:44,544 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=704522.0, ans=0.1 2024-09-25 08:35:00,660 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=704568.6666666666, ans=0.0 2024-09-25 08:35:33,809 INFO [train.py:1198] (2/4) Epoch 39, batch 2950, loss[loss=0.1802, ctc_loss=0.1141, cr_loss=0.3304, over 17049.00 frames. ], tot_loss[loss=0.1919, ctc_loss=0.1236, cr_loss=0.3415, over 3365501.85 frames. ], batch size: 44, lr: 3.04e-03, grad_scale: 32.0 2024-09-25 08:35:49,708 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=704662.0, ans=0.09899494936611666 2024-09-25 08:36:15,121 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.295e+02 1.375e+02 1.468e+02 1.743e+02, threshold=2.750e+02, percent-clipped=0.0 2024-09-25 08:36:18,909 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.39 vs. limit=22.5 2024-09-25 08:36:19,169 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.83 vs. limit=22.5 2024-09-25 08:36:48,366 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=704848.6666666666, ans=0.1 2024-09-25 08:36:49,714 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=704848.6666666666, ans=0.1 2024-09-25 08:36:55,829 INFO [train.py:1198] (2/4) Epoch 39, batch 3000, loss[loss=0.1832, ctc_loss=0.1168, cr_loss=0.3318, over 17074.00 frames. ], tot_loss[loss=0.1923, ctc_loss=0.1238, cr_loss=0.3423, over 3366615.14 frames. ], batch size: 46, lr: 3.03e-03, grad_scale: 32.0 2024-09-25 08:36:55,830 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 08:37:11,148 INFO [train.py:1230] (2/4) Epoch 39, validation: loss=0.03549, ctc_loss=0.03549, cr_loss=9.367e-15, over 944034.00 frames. 2024-09-25 08:37:11,149 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 08:37:11,418 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=704895.3333333334, ans=0.125 2024-09-25 08:37:50,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=704988.6666666666, ans=0.1 2024-09-25 08:37:52,698 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.72 vs. limit=6.0 2024-09-25 08:37:55,130 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=704988.6666666666, ans=0.1 2024-09-25 08:38:07,824 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=7.59 vs. limit=15.0 2024-09-25 08:38:15,107 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=705082.0, ans=0.1 2024-09-25 08:38:29,013 INFO [train.py:1198] (2/4) Epoch 39, batch 3050, loss[loss=0.2139, ctc_loss=0.1377, cr_loss=0.3808, over 16865.00 frames. ], tot_loss[loss=0.1926, ctc_loss=0.124, cr_loss=0.343, over 3367057.90 frames. ], batch size: 58, lr: 3.03e-03, grad_scale: 32.0 2024-09-25 08:39:08,845 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.272e+02 1.353e+02 1.453e+02 2.990e+02, threshold=2.707e+02, percent-clipped=1.0 2024-09-25 08:39:25,494 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=705268.6666666666, ans=0.0 2024-09-25 08:39:52,867 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=705362.0, ans=0.2 2024-09-25 08:39:54,025 INFO [train.py:1198] (2/4) Epoch 39, batch 3100, loss[loss=0.1583, ctc_loss=0.09899, cr_loss=0.2966, over 17013.00 frames. ], tot_loss[loss=0.1924, ctc_loss=0.124, cr_loss=0.3422, over 3363677.88 frames. ], batch size: 51, lr: 3.03e-03, grad_scale: 32.0 2024-09-25 08:40:00,661 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=705362.0, ans=0.2 2024-09-25 08:40:05,338 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=705362.0, ans=0.025 2024-09-25 08:40:19,754 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.06 vs. limit=15.0 2024-09-25 08:40:20,971 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=705408.6666666666, ans=0.0 2024-09-25 08:40:30,420 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=705455.3333333334, ans=0.025 2024-09-25 08:40:33,650 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 08:40:59,288 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=7.94 vs. limit=15.0 2024-09-25 08:41:11,987 INFO [train.py:1198] (2/4) Epoch 39, batch 3150, loss[loss=0.185, ctc_loss=0.1226, cr_loss=0.3115, over 17239.00 frames. ], tot_loss[loss=0.1932, ctc_loss=0.1246, cr_loss=0.3429, over 3353751.58 frames. ], batch size: 50, lr: 3.03e-03, grad_scale: 16.0 2024-09-25 08:41:16,115 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.02 vs. limit=15.0 2024-09-25 08:41:21,745 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=705595.3333333334, ans=0.1 2024-09-25 08:41:44,463 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn1.whiten, num_groups=1, num_channels=192, metric=8.79 vs. limit=22.5 2024-09-25 08:41:50,876 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.157e+02 1.273e+02 1.360e+02 1.459e+02 1.912e+02, threshold=2.719e+02, percent-clipped=0.0 2024-09-25 08:42:03,535 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=705735.3333333334, ans=0.125 2024-09-25 08:42:29,860 INFO [train.py:1198] (2/4) Epoch 39, batch 3200, loss[loss=0.2082, ctc_loss=0.1332, cr_loss=0.3747, over 17274.00 frames. ], tot_loss[loss=0.1929, ctc_loss=0.1243, cr_loss=0.3429, over 3358323.46 frames. ], batch size: 46, lr: 3.03e-03, grad_scale: 32.0 2024-09-25 08:42:47,382 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=705875.3333333334, ans=0.0 2024-09-25 08:42:55,328 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=705875.3333333334, ans=0.125 2024-09-25 08:43:00,972 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.44 vs. limit=5.0 2024-09-25 08:43:09,570 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=705922.0, ans=0.2 2024-09-25 08:43:16,258 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.93 vs. limit=10.0 2024-09-25 08:43:18,924 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=705968.6666666666, ans=0.0 2024-09-25 08:43:48,013 INFO [train.py:1198] (2/4) Epoch 39, batch 3250, loss[loss=0.1932, ctc_loss=0.1241, cr_loss=0.3451, over 17030.00 frames. ], tot_loss[loss=0.1927, ctc_loss=0.1242, cr_loss=0.3426, over 3356697.36 frames. ], batch size: 52, lr: 3.03e-03, grad_scale: 32.0 2024-09-25 08:44:16,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=706108.6666666666, ans=0.125 2024-09-25 08:44:26,959 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.281e+02 1.355e+02 1.471e+02 2.202e+02, threshold=2.709e+02, percent-clipped=0.0 2024-09-25 08:44:45,924 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=706202.0, ans=0.1 2024-09-25 08:45:05,828 INFO [train.py:1198] (2/4) Epoch 39, batch 3300, loss[loss=0.1776, ctc_loss=0.1122, cr_loss=0.3269, over 16955.00 frames. ], tot_loss[loss=0.1917, ctc_loss=0.1235, cr_loss=0.3407, over 3348690.61 frames. ], batch size: 42, lr: 3.03e-03, grad_scale: 32.0 2024-09-25 08:45:18,538 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=706295.3333333334, ans=0.0 2024-09-25 08:45:30,413 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.59 vs. limit=15.0 2024-09-25 08:45:30,434 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.whiten, num_groups=1, num_channels=192, metric=3.92 vs. limit=12.0 2024-09-25 08:45:33,574 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=706342.0, ans=0.125 2024-09-25 08:45:46,255 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=7.58 vs. limit=15.0 2024-09-25 08:45:49,045 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=706388.6666666666, ans=0.1 2024-09-25 08:46:14,487 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.58 vs. limit=10.0 2024-09-25 08:46:24,908 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=706528.6666666666, ans=0.1 2024-09-25 08:46:26,217 INFO [train.py:1198] (2/4) Epoch 39, batch 3350, loss[loss=0.146, ctc_loss=0.08932, cr_loss=0.2836, over 17028.00 frames. ], tot_loss[loss=0.1906, ctc_loss=0.1228, cr_loss=0.3392, over 3343857.19 frames. ], batch size: 39, lr: 3.03e-03, grad_scale: 32.0 2024-09-25 08:46:35,268 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module1.whiten, num_groups=1, num_channels=192, metric=7.54 vs. limit=15.0 2024-09-25 08:47:05,601 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.259e+02 1.360e+02 1.476e+02 1.978e+02, threshold=2.721e+02, percent-clipped=0.0 2024-09-25 08:47:44,980 INFO [train.py:1198] (2/4) Epoch 39, batch 3400, loss[loss=0.1912, ctc_loss=0.1254, cr_loss=0.329, over 17312.00 frames. ], tot_loss[loss=0.191, ctc_loss=0.1231, cr_loss=0.3396, over 3336976.20 frames. ], batch size: 49, lr: 3.03e-03, grad_scale: 32.0 2024-09-25 08:47:46,972 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=706762.0, ans=0.1 2024-09-25 08:47:50,534 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=9.94 vs. limit=15.0 2024-09-25 08:47:50,630 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=16.12 vs. limit=22.5 2024-09-25 08:47:58,564 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn1.whiten, num_groups=1, num_channels=192, metric=10.10 vs. limit=22.5 2024-09-25 08:48:07,065 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=706808.6666666666, ans=0.125 2024-09-25 08:48:08,482 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=706808.6666666666, ans=0.1 2024-09-25 08:48:13,101 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=706808.6666666666, ans=0.05 2024-09-25 08:48:25,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=706855.3333333334, ans=0.125 2024-09-25 08:48:30,619 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=706902.0, ans=0.2 2024-09-25 08:48:30,711 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=706902.0, ans=0.1 2024-09-25 08:48:33,561 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=706902.0, ans=0.125 2024-09-25 08:48:49,461 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=706948.6666666666, ans=0.125 2024-09-25 08:48:52,910 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=8.36 vs. limit=15.0 2024-09-25 08:49:03,008 INFO [train.py:1198] (2/4) Epoch 39, batch 3450, loss[loss=0.1833, ctc_loss=0.1153, cr_loss=0.3401, over 17243.00 frames. ], tot_loss[loss=0.1914, ctc_loss=0.1233, cr_loss=0.3403, over 3343266.19 frames. ], batch size: 44, lr: 3.03e-03, grad_scale: 32.0 2024-09-25 08:49:48,211 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.321e+02 1.412e+02 1.498e+02 2.585e+02, threshold=2.824e+02, percent-clipped=0.0 2024-09-25 08:49:52,954 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=707088.6666666666, ans=0.2 2024-09-25 08:50:26,704 INFO [train.py:1198] (2/4) Epoch 39, batch 3500, loss[loss=0.1892, ctc_loss=0.1229, cr_loss=0.3313, over 17167.00 frames. ], tot_loss[loss=0.193, ctc_loss=0.1245, cr_loss=0.3425, over 3343358.44 frames. ], batch size: 45, lr: 3.03e-03, grad_scale: 32.0 2024-09-25 08:50:31,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=707228.6666666666, ans=0.0 2024-09-25 08:50:34,581 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=707228.6666666666, ans=0.125 2024-09-25 08:50:45,658 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-25 08:51:26,303 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-25 08:51:29,346 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=707415.3333333334, ans=0.04949747468305833 2024-09-25 08:51:44,609 INFO [train.py:1198] (2/4) Epoch 39, batch 3550, loss[loss=0.1949, ctc_loss=0.1262, cr_loss=0.3434, over 17025.00 frames. ], tot_loss[loss=0.1925, ctc_loss=0.1241, cr_loss=0.3418, over 3351426.94 frames. ], batch size: 44, lr: 3.03e-03, grad_scale: 32.0 2024-09-25 08:51:53,108 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=707462.0, ans=0.0 2024-09-25 08:51:58,107 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.17 vs. limit=15.0 2024-09-25 08:52:23,764 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.109e+02 1.264e+02 1.345e+02 1.429e+02 1.997e+02, threshold=2.690e+02, percent-clipped=0.0 2024-09-25 08:52:27,840 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.58 vs. limit=10.0 2024-09-25 08:53:00,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=707648.6666666666, ans=0.0 2024-09-25 08:53:02,932 INFO [train.py:1198] (2/4) Epoch 39, batch 3600, loss[loss=0.1851, ctc_loss=0.1194, cr_loss=0.3282, over 16899.00 frames. ], tot_loss[loss=0.1933, ctc_loss=0.1247, cr_loss=0.343, over 3352117.02 frames. ], batch size: 58, lr: 3.03e-03, grad_scale: 32.0 2024-09-25 08:53:15,506 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=707695.3333333334, ans=0.025 2024-09-25 08:53:15,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=707695.3333333334, ans=0.125 2024-09-25 08:53:35,616 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=707788.6666666666, ans=0.125 2024-09-25 08:53:43,439 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=707788.6666666666, ans=0.2 2024-09-25 08:54:06,611 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 08:54:08,215 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=707882.0, ans=0.125 2024-09-25 08:54:20,462 INFO [train.py:1198] (2/4) Epoch 39, batch 3650, loss[loss=0.1614, ctc_loss=0.1017, cr_loss=0.2984, over 16943.00 frames. ], tot_loss[loss=0.1934, ctc_loss=0.1248, cr_loss=0.3429, over 3346289.03 frames. ], batch size: 42, lr: 3.03e-03, grad_scale: 32.0 2024-09-25 08:54:25,498 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=707928.6666666666, ans=0.05 2024-09-25 08:54:59,909 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.155e+02 1.284e+02 1.368e+02 1.461e+02 2.127e+02, threshold=2.736e+02, percent-clipped=0.0 2024-09-25 08:55:10,896 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=3.97 vs. limit=15.0 2024-09-25 08:55:14,477 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=708068.6666666666, ans=0.1 2024-09-25 08:55:28,570 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=708115.3333333334, ans=0.0 2024-09-25 08:55:40,513 INFO [train.py:1198] (2/4) Epoch 39, batch 3700, loss[loss=0.2266, ctc_loss=0.1484, cr_loss=0.3906, over 15113.00 frames. ], tot_loss[loss=0.1935, ctc_loss=0.125, cr_loss=0.3428, over 3345073.41 frames. ], batch size: 89, lr: 3.03e-03, grad_scale: 32.0 2024-09-25 08:56:28,024 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=708302.0, ans=0.1 2024-09-25 08:56:35,989 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=708302.0, ans=0.125 2024-09-25 08:56:35,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=708302.0, ans=0.2 2024-09-25 08:56:43,723 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=708348.6666666666, ans=0.1 2024-09-25 08:56:59,403 INFO [train.py:1198] (2/4) Epoch 39, batch 3750, loss[loss=0.2301, ctc_loss=0.1528, cr_loss=0.3867, over 15018.00 frames. ], tot_loss[loss=0.1931, ctc_loss=0.1246, cr_loss=0.3426, over 3346388.63 frames. ], batch size: 89, lr: 3.03e-03, grad_scale: 32.0 2024-09-25 08:57:15,634 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=708442.0, ans=0.125 2024-09-25 08:57:26,517 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=708442.0, ans=0.125 2024-09-25 08:57:38,679 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.270e+02 1.361e+02 1.496e+02 2.109e+02, threshold=2.722e+02, percent-clipped=0.0 2024-09-25 08:57:40,585 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=708488.6666666666, ans=0.125 2024-09-25 08:58:18,291 INFO [train.py:1198] (2/4) Epoch 39, batch 3800, loss[loss=0.1754, ctc_loss=0.1104, cr_loss=0.325, over 16956.00 frames. ], tot_loss[loss=0.1909, ctc_loss=0.1231, cr_loss=0.3392, over 3342339.76 frames. ], batch size: 42, lr: 3.03e-03, grad_scale: 32.0 2024-09-25 08:58:20,286 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=708628.6666666666, ans=0.2 2024-09-25 08:58:27,319 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.24 vs. limit=15.0 2024-09-25 08:58:43,440 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=708675.3333333334, ans=0.015 2024-09-25 08:59:28,246 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=708815.3333333334, ans=0.125 2024-09-25 08:59:33,642 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=9.18 vs. limit=15.0 2024-09-25 08:59:36,465 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=708815.3333333334, ans=0.125 2024-09-25 08:59:39,439 INFO [train.py:1198] (2/4) Epoch 39, batch 3850, loss[loss=0.1766, ctc_loss=0.1112, cr_loss=0.3273, over 17179.00 frames. ], tot_loss[loss=0.1899, ctc_loss=0.1224, cr_loss=0.3373, over 3311692.63 frames. ], batch size: 41, lr: 3.03e-03, grad_scale: 32.0 2024-09-25 08:59:41,450 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-25 08:59:46,094 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=708862.0, ans=0.0 2024-09-25 09:00:00,343 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=708908.6666666666, ans=0.0 2024-09-25 09:00:02,468 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=11.54 vs. limit=12.0 2024-09-25 09:00:18,551 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.173e+02 1.306e+02 1.417e+02 1.619e+02 3.168e+02, threshold=2.835e+02, percent-clipped=2.0 2024-09-25 09:00:26,925 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=709002.0, ans=0.0 2024-09-25 09:00:33,057 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=709002.0, ans=0.1 2024-09-25 09:01:40,861 INFO [train.py:1198] (2/4) Epoch 40, batch 0, loss[loss=0.2008, ctc_loss=0.1289, cr_loss=0.3597, over 17226.00 frames. ], tot_loss[loss=0.2008, ctc_loss=0.1289, cr_loss=0.3597, over 17226.00 frames. ], batch size: 50, lr: 2.99e-03, grad_scale: 32.0 2024-09-25 09:01:40,862 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 09:01:56,595 INFO [train.py:1230] (2/4) Epoch 40, validation: loss=0.03491, ctc_loss=0.03491, cr_loss=1.007e-14, over 944034.00 frames. 2024-09-25 09:01:56,596 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 09:02:00,164 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=709076.6666666666, ans=0.2 2024-09-25 09:02:04,020 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.98 vs. limit=15.0 2024-09-25 09:03:00,995 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.59 vs. limit=22.5 2024-09-25 09:03:06,907 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=709263.3333333334, ans=0.025 2024-09-25 09:03:14,551 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.min_abs, batch_count=709310.0, ans=0.5 2024-09-25 09:03:15,889 INFO [train.py:1198] (2/4) Epoch 40, batch 50, loss[loss=0.2055, ctc_loss=0.1305, cr_loss=0.3746, over 17004.00 frames. ], tot_loss[loss=0.1906, ctc_loss=0.1226, cr_loss=0.3398, over 755133.44 frames. ], batch size: 56, lr: 2.99e-03, grad_scale: 32.0 2024-09-25 09:03:16,268 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 09:03:38,958 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=7.57 vs. limit=15.0 2024-09-25 09:03:39,908 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=709356.6666666666, ans=0.025 2024-09-25 09:04:10,882 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.135e+02 1.285e+02 1.412e+02 1.570e+02 2.190e+02, threshold=2.824e+02, percent-clipped=0.0 2024-09-25 09:04:31,936 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=709496.6666666666, ans=10.0 2024-09-25 09:04:39,017 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=5.00 vs. limit=15.0 2024-09-25 09:04:43,260 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=709543.3333333334, ans=0.125 2024-09-25 09:04:44,462 INFO [train.py:1198] (2/4) Epoch 40, batch 100, loss[loss=0.1881, ctc_loss=0.1204, cr_loss=0.3386, over 17318.00 frames. ], tot_loss[loss=0.1903, ctc_loss=0.1226, cr_loss=0.3386, over 1323887.13 frames. ], batch size: 46, lr: 2.99e-03, grad_scale: 32.0 2024-09-25 09:04:55,917 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.71 vs. limit=22.5 2024-09-25 09:05:17,817 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=709636.6666666666, ans=0.1 2024-09-25 09:05:20,938 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=709636.6666666666, ans=0.125 2024-09-25 09:05:21,041 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=709636.6666666666, ans=0.125 2024-09-25 09:05:22,832 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten.whitening_limit, batch_count=709636.6666666666, ans=15.0 2024-09-25 09:05:24,150 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=709636.6666666666, ans=0.125 2024-09-25 09:05:36,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=709683.3333333334, ans=0.0 2024-09-25 09:05:53,776 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=5.24 vs. limit=15.0 2024-09-25 09:06:02,589 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=709730.0, ans=0.125 2024-09-25 09:06:06,998 INFO [train.py:1198] (2/4) Epoch 40, batch 150, loss[loss=0.1574, ctc_loss=0.09992, cr_loss=0.2872, over 17024.00 frames. ], tot_loss[loss=0.1906, ctc_loss=0.1228, cr_loss=0.339, over 1763714.26 frames. ], batch size: 39, lr: 2.99e-03, grad_scale: 32.0 2024-09-25 09:06:35,938 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=709823.3333333334, ans=0.125 2024-09-25 09:06:43,795 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=709870.0, ans=0.07 2024-09-25 09:06:53,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=709870.0, ans=0.125 2024-09-25 09:06:56,159 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.270e+02 1.336e+02 1.420e+02 2.555e+02, threshold=2.672e+02, percent-clipped=0.0 2024-09-25 09:07:04,341 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=709916.6666666666, ans=0.125 2024-09-25 09:07:09,141 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=709916.6666666666, ans=0.125 2024-09-25 09:07:16,099 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.41 vs. limit=15.0 2024-09-25 09:07:29,562 INFO [train.py:1198] (2/4) Epoch 40, batch 200, loss[loss=0.2044, ctc_loss=0.1347, cr_loss=0.3488, over 16888.00 frames. ], tot_loss[loss=0.1923, ctc_loss=0.124, cr_loss=0.3413, over 2113948.94 frames. ], batch size: 58, lr: 2.99e-03, grad_scale: 32.0 2024-09-25 09:07:33,063 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=710010.0, ans=0.125 2024-09-25 09:07:44,127 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=710056.6666666666, ans=0.0 2024-09-25 09:07:44,258 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=710056.6666666666, ans=0.07 2024-09-25 09:08:00,189 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=710103.3333333334, ans=10.0 2024-09-25 09:08:01,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=710103.3333333334, ans=0.125 2024-09-25 09:08:19,347 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=710150.0, ans=0.125 2024-09-25 09:08:50,009 INFO [train.py:1198] (2/4) Epoch 40, batch 250, loss[loss=0.2232, ctc_loss=0.1469, cr_loss=0.3815, over 16103.00 frames. ], tot_loss[loss=0.1939, ctc_loss=0.1252, cr_loss=0.3437, over 2378953.61 frames. ], batch size: 74, lr: 2.98e-03, grad_scale: 32.0 2024-09-25 09:08:53,576 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 09:08:55,139 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=710243.3333333334, ans=0.0 2024-09-25 09:09:11,733 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.91 vs. limit=6.0 2024-09-25 09:09:31,100 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=11.88 vs. limit=15.0 2024-09-25 09:09:39,757 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.275e+02 1.359e+02 1.486e+02 2.398e+02, threshold=2.718e+02, percent-clipped=0.0 2024-09-25 09:10:10,429 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 09:10:16,459 INFO [train.py:1198] (2/4) Epoch 40, batch 300, loss[loss=0.1651, ctc_loss=0.1006, cr_loss=0.323, over 16640.00 frames. ], tot_loss[loss=0.1931, ctc_loss=0.1245, cr_loss=0.3428, over 2600737.34 frames. ], batch size: 37, lr: 2.98e-03, grad_scale: 32.0 2024-09-25 09:10:34,773 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=15.40 vs. limit=22.5 2024-09-25 09:10:35,850 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=710523.3333333334, ans=0.025 2024-09-25 09:10:37,363 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=710523.3333333334, ans=0.0 2024-09-25 09:11:08,039 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=710616.6666666666, ans=0.125 2024-09-25 09:11:36,229 INFO [train.py:1198] (2/4) Epoch 40, batch 350, loss[loss=0.1999, ctc_loss=0.1288, cr_loss=0.3555, over 17333.00 frames. ], tot_loss[loss=0.1927, ctc_loss=0.1242, cr_loss=0.3427, over 2775456.62 frames. ], batch size: 51, lr: 2.98e-03, grad_scale: 32.0 2024-09-25 09:11:42,997 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 09:11:58,190 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=710756.6666666666, ans=0.2 2024-09-25 09:12:27,035 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.301e+02 1.389e+02 1.518e+02 2.541e+02, threshold=2.778e+02, percent-clipped=0.0 2024-09-25 09:12:46,482 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=710896.6666666666, ans=0.05 2024-09-25 09:12:48,049 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=710896.6666666666, ans=0.125 2024-09-25 09:12:51,335 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=710896.6666666666, ans=0.0 2024-09-25 09:12:59,094 INFO [train.py:1198] (2/4) Epoch 40, batch 400, loss[loss=0.2217, ctc_loss=0.1525, cr_loss=0.3463, over 12157.00 frames. ], tot_loss[loss=0.1933, ctc_loss=0.1247, cr_loss=0.3435, over 2912021.66 frames. ], batch size: 123, lr: 2.98e-03, grad_scale: 32.0 2024-09-25 09:13:15,222 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=710990.0, ans=0.1 2024-09-25 09:13:26,313 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=710990.0, ans=0.0 2024-09-25 09:13:29,764 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=711036.6666666666, ans=0.125 2024-09-25 09:13:42,731 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=711036.6666666666, ans=0.125 2024-09-25 09:14:01,601 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=711083.3333333334, ans=0.0 2024-09-25 09:14:05,524 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=8.65 vs. limit=15.0 2024-09-25 09:14:09,685 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=711130.0, ans=0.125 2024-09-25 09:14:14,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=711130.0, ans=0.0 2024-09-25 09:14:22,072 INFO [train.py:1198] (2/4) Epoch 40, batch 450, loss[loss=0.1705, ctc_loss=0.1067, cr_loss=0.319, over 17261.00 frames. ], tot_loss[loss=0.1938, ctc_loss=0.125, cr_loss=0.344, over 3006485.58 frames. ], batch size: 44, lr: 2.98e-03, grad_scale: 32.0 2024-09-25 09:14:30,356 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=711176.6666666666, ans=0.125 2024-09-25 09:15:05,097 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.min_positive, batch_count=711270.0, ans=0.05 2024-09-25 09:15:11,211 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=711316.6666666666, ans=0.0 2024-09-25 09:15:12,492 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.098e+02 1.283e+02 1.337e+02 1.424e+02 2.250e+02, threshold=2.674e+02, percent-clipped=0.0 2024-09-25 09:15:20,706 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=711316.6666666666, ans=0.125 2024-09-25 09:15:22,396 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=711316.6666666666, ans=0.2 2024-09-25 09:15:44,620 INFO [train.py:1198] (2/4) Epoch 40, batch 500, loss[loss=0.1799, ctc_loss=0.1133, cr_loss=0.3332, over 16971.00 frames. ], tot_loss[loss=0.194, ctc_loss=0.1251, cr_loss=0.3446, over 3079322.96 frames. ], batch size: 42, lr: 2.98e-03, grad_scale: 32.0 2024-09-25 09:15:52,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=711410.0, ans=0.0 2024-09-25 09:16:03,854 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=711456.6666666666, ans=0.125 2024-09-25 09:16:12,405 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=9.80 vs. limit=15.0 2024-09-25 09:16:40,909 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=711550.0, ans=0.025 2024-09-25 09:16:48,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=711550.0, ans=0.05 2024-09-25 09:16:53,791 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=4.86 vs. limit=15.0 2024-09-25 09:16:56,436 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=711596.6666666666, ans=0.125 2024-09-25 09:17:07,178 INFO [train.py:1198] (2/4) Epoch 40, batch 550, loss[loss=0.185, ctc_loss=0.1193, cr_loss=0.3284, over 17254.00 frames. ], tot_loss[loss=0.1928, ctc_loss=0.1241, cr_loss=0.3432, over 3152754.17 frames. ], batch size: 44, lr: 2.98e-03, grad_scale: 32.0 2024-09-25 09:17:07,428 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=711643.3333333334, ans=0.2 2024-09-25 09:17:28,517 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=711690.0, ans=0.025 2024-09-25 09:17:36,318 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=711690.0, ans=0.035 2024-09-25 09:17:57,147 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.279e+02 1.358e+02 1.490e+02 2.059e+02, threshold=2.716e+02, percent-clipped=0.0 2024-09-25 09:18:20,043 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=711830.0, ans=0.025 2024-09-25 09:18:28,028 INFO [train.py:1198] (2/4) Epoch 40, batch 600, loss[loss=0.1831, ctc_loss=0.1185, cr_loss=0.323, over 17223.00 frames. ], tot_loss[loss=0.192, ctc_loss=0.1236, cr_loss=0.342, over 3193349.82 frames. ], batch size: 50, lr: 2.98e-03, grad_scale: 16.0 2024-09-25 09:18:36,839 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.73 vs. limit=6.0 2024-09-25 09:18:41,055 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=711876.6666666666, ans=0.1 2024-09-25 09:19:21,701 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=712016.6666666666, ans=0.2 2024-09-25 09:19:51,405 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=712063.3333333334, ans=0.125 2024-09-25 09:19:55,898 INFO [train.py:1198] (2/4) Epoch 40, batch 650, loss[loss=0.1858, ctc_loss=0.1209, cr_loss=0.3243, over 17020.00 frames. ], tot_loss[loss=0.1933, ctc_loss=0.1246, cr_loss=0.3435, over 3228461.41 frames. ], batch size: 51, lr: 2.98e-03, grad_scale: 16.0 2024-09-25 09:20:12,499 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=712156.6666666666, ans=0.125 2024-09-25 09:20:36,416 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=712203.3333333334, ans=0.125 2024-09-25 09:20:36,907 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=9.03 vs. limit=15.0 2024-09-25 09:20:45,845 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.262e+02 1.364e+02 1.466e+02 1.763e+02, threshold=2.728e+02, percent-clipped=0.0 2024-09-25 09:20:46,117 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=712250.0, ans=0.0 2024-09-25 09:20:53,299 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=8.71 vs. limit=15.0 2024-09-25 09:21:12,513 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.84 vs. limit=15.0 2024-09-25 09:21:16,165 INFO [train.py:1198] (2/4) Epoch 40, batch 700, loss[loss=0.166, ctc_loss=0.1046, cr_loss=0.3071, over 17244.00 frames. ], tot_loss[loss=0.1935, ctc_loss=0.1248, cr_loss=0.3432, over 3252199.49 frames. ], batch size: 44, lr: 2.98e-03, grad_scale: 16.0 2024-09-25 09:21:16,379 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=712343.3333333334, ans=0.0 2024-09-25 09:21:34,038 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=712390.0, ans=0.125 2024-09-25 09:21:35,694 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=712390.0, ans=0.0 2024-09-25 09:22:21,186 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=712530.0, ans=0.0 2024-09-25 09:22:33,740 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=712530.0, ans=0.1 2024-09-25 09:22:38,081 INFO [train.py:1198] (2/4) Epoch 40, batch 750, loss[loss=0.1777, ctc_loss=0.1106, cr_loss=0.3356, over 17008.00 frames. ], tot_loss[loss=0.1945, ctc_loss=0.1255, cr_loss=0.3447, over 3263006.15 frames. ], batch size: 44, lr: 2.98e-03, grad_scale: 16.0 2024-09-25 09:22:40,438 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.96 vs. limit=10.0 2024-09-25 09:22:41,586 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=712576.6666666666, ans=0.125 2024-09-25 09:22:53,198 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.65 vs. limit=10.0 2024-09-25 09:23:27,401 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.065e+02 1.294e+02 1.361e+02 1.444e+02 2.754e+02, threshold=2.722e+02, percent-clipped=1.0 2024-09-25 09:23:36,165 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=7.58 vs. limit=15.0 2024-09-25 09:23:43,078 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=712763.3333333334, ans=0.125 2024-09-25 09:24:03,443 INFO [train.py:1198] (2/4) Epoch 40, batch 800, loss[loss=0.1721, ctc_loss=0.1103, cr_loss=0.3091, over 17078.00 frames. ], tot_loss[loss=0.1938, ctc_loss=0.1252, cr_loss=0.3431, over 3280611.31 frames. ], batch size: 43, lr: 2.98e-03, grad_scale: 32.0 2024-09-25 09:24:19,608 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=712856.6666666666, ans=0.2 2024-09-25 09:24:58,994 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.34 vs. limit=15.0 2024-09-25 09:25:26,601 INFO [train.py:1198] (2/4) Epoch 40, batch 850, loss[loss=0.2132, ctc_loss=0.1484, cr_loss=0.3242, over 11501.00 frames. ], tot_loss[loss=0.1924, ctc_loss=0.1242, cr_loss=0.3412, over 3287750.55 frames. ], batch size: 123, lr: 2.98e-03, grad_scale: 32.0 2024-09-25 09:26:01,156 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten.whitening_limit, batch_count=713136.6666666666, ans=15.0 2024-09-25 09:26:11,837 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=713136.6666666666, ans=0.0 2024-09-25 09:26:16,238 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.145e+02 1.263e+02 1.352e+02 1.405e+02 2.259e+02, threshold=2.704e+02, percent-clipped=0.0 2024-09-25 09:26:22,778 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=713183.3333333334, ans=0.05 2024-09-25 09:26:25,909 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=713183.3333333334, ans=0.0 2024-09-25 09:26:49,140 INFO [train.py:1198] (2/4) Epoch 40, batch 900, loss[loss=0.1964, ctc_loss=0.1247, cr_loss=0.3585, over 17306.00 frames. ], tot_loss[loss=0.1923, ctc_loss=0.124, cr_loss=0.3416, over 3304088.79 frames. ], batch size: 51, lr: 2.98e-03, grad_scale: 32.0 2024-09-25 09:27:00,682 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=713276.6666666666, ans=0.1 2024-09-25 09:27:29,702 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=713370.0, ans=0.125 2024-09-25 09:27:42,409 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=713416.6666666666, ans=0.0 2024-09-25 09:27:56,597 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=713463.3333333334, ans=0.5 2024-09-25 09:27:58,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.min_positive, batch_count=713463.3333333334, ans=0.05 2024-09-25 09:27:58,342 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 09:28:09,204 INFO [train.py:1198] (2/4) Epoch 40, batch 950, loss[loss=0.1834, ctc_loss=0.1153, cr_loss=0.3404, over 17077.00 frames. ], tot_loss[loss=0.1916, ctc_loss=0.1235, cr_loss=0.3408, over 3318678.36 frames. ], batch size: 46, lr: 2.98e-03, grad_scale: 32.0 2024-09-25 09:28:20,895 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=713510.0, ans=0.2 2024-09-25 09:28:49,954 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=713603.3333333334, ans=0.0 2024-09-25 09:29:04,047 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.318e+02 1.382e+02 1.485e+02 2.096e+02, threshold=2.764e+02, percent-clipped=0.0 2024-09-25 09:29:14,096 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=713650.0, ans=0.125 2024-09-25 09:29:37,543 INFO [train.py:1198] (2/4) Epoch 40, batch 1000, loss[loss=0.2157, ctc_loss=0.1399, cr_loss=0.379, over 17311.00 frames. ], tot_loss[loss=0.1921, ctc_loss=0.1238, cr_loss=0.3413, over 3315916.29 frames. ], batch size: 51, lr: 2.98e-03, grad_scale: 32.0 2024-09-25 09:29:44,137 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=713743.3333333334, ans=0.025 2024-09-25 09:29:47,310 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=713743.3333333334, ans=0.125 2024-09-25 09:29:49,060 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=713743.3333333334, ans=0.1 2024-09-25 09:30:04,970 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=713790.0, ans=0.125 2024-09-25 09:30:08,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=713836.6666666666, ans=0.125 2024-09-25 09:30:19,431 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=713836.6666666666, ans=0.125 2024-09-25 09:30:33,652 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=713883.3333333334, ans=0.0 2024-09-25 09:30:40,759 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.76 vs. limit=10.0 2024-09-25 09:30:50,570 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.07 vs. limit=15.0 2024-09-25 09:30:57,910 INFO [train.py:1198] (2/4) Epoch 40, batch 1050, loss[loss=0.1924, ctc_loss=0.1227, cr_loss=0.3487, over 17093.00 frames. ], tot_loss[loss=0.191, ctc_loss=0.123, cr_loss=0.3399, over 3327741.61 frames. ], batch size: 49, lr: 2.98e-03, grad_scale: 32.0 2024-09-25 09:31:08,280 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=713976.6666666666, ans=0.0 2024-09-25 09:31:29,006 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=714070.0, ans=0.1 2024-09-25 09:31:29,154 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=714070.0, ans=0.125 2024-09-25 09:31:36,453 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 09:31:48,955 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=714116.6666666666, ans=0.125 2024-09-25 09:31:51,006 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.04 vs. limit=10.0 2024-09-25 09:31:51,952 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.281e+02 1.362e+02 1.480e+02 1.897e+02, threshold=2.724e+02, percent-clipped=0.0 2024-09-25 09:32:01,864 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=714116.6666666666, ans=0.0 2024-09-25 09:32:20,606 INFO [train.py:1198] (2/4) Epoch 40, batch 1100, loss[loss=0.2261, ctc_loss=0.147, cr_loss=0.3957, over 16768.00 frames. ], tot_loss[loss=0.1906, ctc_loss=0.1228, cr_loss=0.3394, over 3332351.25 frames. ], batch size: 61, lr: 2.98e-03, grad_scale: 16.0 2024-09-25 09:32:36,820 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=714256.6666666666, ans=0.125 2024-09-25 09:32:53,644 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.whiten, num_groups=1, num_channels=192, metric=4.63 vs. limit=12.0 2024-09-25 09:33:43,633 INFO [train.py:1198] (2/4) Epoch 40, batch 1150, loss[loss=0.1927, ctc_loss=0.1336, cr_loss=0.2951, over 11791.00 frames. ], tot_loss[loss=0.1906, ctc_loss=0.1227, cr_loss=0.3397, over 3338321.93 frames. ], batch size: 123, lr: 2.98e-03, grad_scale: 16.0 2024-09-25 09:34:30,011 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.09 vs. limit=15.0 2024-09-25 09:34:38,907 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=714583.3333333334, ans=0.1 2024-09-25 09:34:40,142 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.266e+02 1.369e+02 1.564e+02 2.152e+02, threshold=2.737e+02, percent-clipped=0.0 2024-09-25 09:34:48,419 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=714583.3333333334, ans=0.0 2024-09-25 09:34:53,186 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=714630.0, ans=0.2 2024-09-25 09:35:08,859 INFO [train.py:1198] (2/4) Epoch 40, batch 1200, loss[loss=0.1745, ctc_loss=0.1115, cr_loss=0.3154, over 16974.00 frames. ], tot_loss[loss=0.1917, ctc_loss=0.1234, cr_loss=0.3413, over 3333551.68 frames. ], batch size: 42, lr: 2.98e-03, grad_scale: 32.0 2024-09-25 09:35:17,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=714676.6666666666, ans=0.125 2024-09-25 09:35:32,347 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=13.12 vs. limit=22.5 2024-09-25 09:35:42,828 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=714770.0, ans=0.0 2024-09-25 09:35:52,501 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=714770.0, ans=0.0 2024-09-25 09:36:10,635 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.whiten.whitening_limit, batch_count=714816.6666666666, ans=12.0 2024-09-25 09:36:21,452 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=714863.3333333334, ans=0.125 2024-09-25 09:36:21,481 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=714863.3333333334, ans=0.07 2024-09-25 09:36:29,057 INFO [train.py:1198] (2/4) Epoch 40, batch 1250, loss[loss=0.1869, ctc_loss=0.1211, cr_loss=0.3286, over 16743.00 frames. ], tot_loss[loss=0.1915, ctc_loss=0.1234, cr_loss=0.3408, over 3343630.89 frames. ], batch size: 61, lr: 2.98e-03, grad_scale: 32.0 2024-09-25 09:36:29,459 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.max_abs, batch_count=714910.0, ans=10.0 2024-09-25 09:36:36,586 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=714910.0, ans=0.125 2024-09-25 09:36:49,696 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=714956.6666666666, ans=0.125 2024-09-25 09:37:05,498 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer_ff2.min_abs, batch_count=715003.3333333334, ans=0.1 2024-09-25 09:37:07,829 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.90 vs. limit=12.0 2024-09-25 09:37:08,650 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=715003.3333333334, ans=0.2 2024-09-25 09:37:10,267 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=715003.3333333334, ans=0.125 2024-09-25 09:37:22,756 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.149e+02 1.288e+02 1.344e+02 1.458e+02 2.122e+02, threshold=2.687e+02, percent-clipped=0.0 2024-09-25 09:37:42,137 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=715096.6666666666, ans=0.0 2024-09-25 09:37:51,161 INFO [train.py:1198] (2/4) Epoch 40, batch 1300, loss[loss=0.1818, ctc_loss=0.1179, cr_loss=0.3198, over 17237.00 frames. ], tot_loss[loss=0.1921, ctc_loss=0.1239, cr_loss=0.3411, over 3341529.64 frames. ], batch size: 44, lr: 2.97e-03, grad_scale: 32.0 2024-09-25 09:37:51,401 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=715143.3333333334, ans=0.0 2024-09-25 09:38:00,978 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=715143.3333333334, ans=0.025 2024-09-25 09:38:12,041 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=715190.0, ans=0.125 2024-09-25 09:38:29,189 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=715236.6666666666, ans=0.2 2024-09-25 09:38:30,832 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=715236.6666666666, ans=0.0 2024-09-25 09:38:32,380 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=715236.6666666666, ans=0.2 2024-09-25 09:39:03,942 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=715330.0, ans=0.125 2024-09-25 09:39:07,140 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=715330.0, ans=0.125 2024-09-25 09:39:11,938 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=715330.0, ans=0.0 2024-09-25 09:39:19,071 INFO [train.py:1198] (2/4) Epoch 40, batch 1350, loss[loss=0.2128, ctc_loss=0.1402, cr_loss=0.3631, over 16913.00 frames. ], tot_loss[loss=0.1927, ctc_loss=0.1245, cr_loss=0.3414, over 3341266.73 frames. ], batch size: 58, lr: 2.97e-03, grad_scale: 32.0 2024-09-25 09:39:19,308 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer_na.min_abs, batch_count=715376.6666666666, ans=0.02 2024-09-25 09:39:28,936 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=715376.6666666666, ans=0.0 2024-09-25 09:39:35,529 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=715423.3333333334, ans=0.09899494936611666 2024-09-25 09:40:10,085 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.078e+02 1.269e+02 1.368e+02 1.487e+02 1.942e+02, threshold=2.736e+02, percent-clipped=0.0 2024-09-25 09:40:16,711 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=715516.6666666666, ans=0.125 2024-09-25 09:40:31,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=715563.3333333334, ans=0.2 2024-09-25 09:40:34,320 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.max_abs, batch_count=715563.3333333334, ans=10.0 2024-09-25 09:40:39,122 INFO [train.py:1198] (2/4) Epoch 40, batch 1400, loss[loss=0.1766, ctc_loss=0.1125, cr_loss=0.3203, over 17193.00 frames. ], tot_loss[loss=0.1914, ctc_loss=0.1234, cr_loss=0.3403, over 3356890.27 frames. ], batch size: 41, lr: 2.97e-03, grad_scale: 32.0 2024-09-25 09:40:41,179 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=715610.0, ans=0.04949747468305833 2024-09-25 09:40:49,302 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=715610.0, ans=0.125 2024-09-25 09:40:55,567 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=715656.6666666666, ans=0.0 2024-09-25 09:41:28,006 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=20.20 vs. limit=22.5 2024-09-25 09:41:36,702 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=715750.0, ans=0.125 2024-09-25 09:41:55,313 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.87 vs. limit=12.0 2024-09-25 09:42:02,342 INFO [train.py:1198] (2/4) Epoch 40, batch 1450, loss[loss=0.1728, ctc_loss=0.108, cr_loss=0.324, over 17248.00 frames. ], tot_loss[loss=0.1906, ctc_loss=0.1228, cr_loss=0.3393, over 3358278.53 frames. ], batch size: 44, lr: 2.97e-03, grad_scale: 32.0 2024-09-25 09:42:13,761 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=715843.3333333334, ans=0.125 2024-09-25 09:42:15,829 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys.whitening_limit, batch_count=715843.3333333334, ans=6.0 2024-09-25 09:42:21,744 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=715890.0, ans=0.0 2024-09-25 09:42:28,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=715890.0, ans=0.125 2024-09-25 09:42:29,793 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=715890.0, ans=0.04949747468305833 2024-09-25 09:42:32,911 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=715936.6666666666, ans=0.0 2024-09-25 09:42:53,419 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.248e+02 1.311e+02 1.404e+02 2.777e+02, threshold=2.622e+02, percent-clipped=1.0 2024-09-25 09:43:07,995 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=716030.0, ans=0.07 2024-09-25 09:43:19,072 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=716030.0, ans=0.2 2024-09-25 09:43:21,941 INFO [train.py:1198] (2/4) Epoch 40, batch 1500, loss[loss=0.2304, ctc_loss=0.1573, cr_loss=0.3655, over 12022.00 frames. ], tot_loss[loss=0.1909, ctc_loss=0.123, cr_loss=0.3399, over 3348732.89 frames. ], batch size: 123, lr: 2.97e-03, grad_scale: 32.0 2024-09-25 09:43:33,215 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.46 vs. limit=15.0 2024-09-25 09:44:49,469 INFO [train.py:1198] (2/4) Epoch 40, batch 1550, loss[loss=0.1929, ctc_loss=0.125, cr_loss=0.3394, over 17025.00 frames. ], tot_loss[loss=0.1912, ctc_loss=0.1232, cr_loss=0.3399, over 3341583.73 frames. ], batch size: 56, lr: 2.97e-03, grad_scale: 32.0 2024-09-25 09:44:49,738 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=716310.0, ans=0.125 2024-09-25 09:44:53,069 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=716310.0, ans=0.2 2024-09-25 09:45:21,893 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=716403.3333333334, ans=0.125 2024-09-25 09:45:34,761 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=716403.3333333334, ans=0.2 2024-09-25 09:45:37,748 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=716450.0, ans=0.0 2024-09-25 09:45:42,231 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.120e+02 1.286e+02 1.352e+02 1.488e+02 2.645e+02, threshold=2.703e+02, percent-clipped=1.0 2024-09-25 09:45:49,036 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=716450.0, ans=0.125 2024-09-25 09:45:50,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=716450.0, ans=0.1 2024-09-25 09:45:58,557 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=716496.6666666666, ans=0.0 2024-09-25 09:46:01,660 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=716496.6666666666, ans=0.0 2024-09-25 09:46:09,314 INFO [train.py:1198] (2/4) Epoch 40, batch 1600, loss[loss=0.2076, ctc_loss=0.1341, cr_loss=0.3674, over 17022.00 frames. ], tot_loss[loss=0.19, ctc_loss=0.1223, cr_loss=0.3382, over 3347023.88 frames. ], batch size: 44, lr: 2.97e-03, grad_scale: 32.0 2024-09-25 09:46:54,366 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=11.31 vs. limit=12.0 2024-09-25 09:47:19,370 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=716730.0, ans=0.125 2024-09-25 09:47:22,731 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=716730.0, ans=0.125 2024-09-25 09:47:32,133 INFO [train.py:1198] (2/4) Epoch 40, batch 1650, loss[loss=0.2166, ctc_loss=0.1392, cr_loss=0.3871, over 17212.00 frames. ], tot_loss[loss=0.1901, ctc_loss=0.1224, cr_loss=0.3381, over 3345588.69 frames. ], batch size: 55, lr: 2.97e-03, grad_scale: 32.0 2024-09-25 09:47:58,012 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=716823.3333333334, ans=0.125 2024-09-25 09:48:12,207 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=716870.0, ans=0.125 2024-09-25 09:48:15,358 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=716870.0, ans=0.2 2024-09-25 09:48:27,241 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.303e+02 1.369e+02 1.457e+02 1.993e+02, threshold=2.739e+02, percent-clipped=0.0 2024-09-25 09:48:56,877 INFO [train.py:1198] (2/4) Epoch 40, batch 1700, loss[loss=0.1965, ctc_loss=0.1262, cr_loss=0.3513, over 16740.00 frames. ], tot_loss[loss=0.1902, ctc_loss=0.1226, cr_loss=0.3378, over 3348136.09 frames. ], batch size: 61, lr: 2.97e-03, grad_scale: 32.0 2024-09-25 09:48:59,464 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=18.76 vs. limit=22.5 2024-09-25 09:49:06,441 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=717010.0, ans=0.0 2024-09-25 09:49:42,778 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=717103.3333333334, ans=0.125 2024-09-25 09:49:44,355 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=717103.3333333334, ans=0.125 2024-09-25 09:49:51,996 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=717150.0, ans=0.125 2024-09-25 09:50:18,902 INFO [train.py:1198] (2/4) Epoch 40, batch 1750, loss[loss=0.1992, ctc_loss=0.1259, cr_loss=0.3662, over 17281.00 frames. ], tot_loss[loss=0.1896, ctc_loss=0.1222, cr_loss=0.3374, over 3338269.93 frames. ], batch size: 46, lr: 2.97e-03, grad_scale: 32.0 2024-09-25 09:50:33,662 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=717290.0, ans=0.0 2024-09-25 09:50:51,148 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=717336.6666666666, ans=0.125 2024-09-25 09:51:04,068 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=717336.6666666666, ans=0.125 2024-09-25 09:51:10,625 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=717383.3333333334, ans=0.0 2024-09-25 09:51:11,756 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.139e+02 1.261e+02 1.332e+02 1.446e+02 2.217e+02, threshold=2.663e+02, percent-clipped=0.0 2024-09-25 09:51:12,177 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=717383.3333333334, ans=0.125 2024-09-25 09:51:16,955 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=717383.3333333334, ans=0.05 2024-09-25 09:51:41,753 INFO [train.py:1198] (2/4) Epoch 40, batch 1800, loss[loss=0.1825, ctc_loss=0.1198, cr_loss=0.3138, over 17301.00 frames. ], tot_loss[loss=0.19, ctc_loss=0.1223, cr_loss=0.3382, over 3342575.05 frames. ], batch size: 49, lr: 2.97e-03, grad_scale: 32.0 2024-09-25 09:52:39,053 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=4.82 vs. limit=10.0 2024-09-25 09:52:45,344 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.whiten, num_groups=1, num_channels=192, metric=4.87 vs. limit=12.0 2024-09-25 09:53:01,964 INFO [train.py:1198] (2/4) Epoch 40, batch 1850, loss[loss=0.1837, ctc_loss=0.1168, cr_loss=0.3347, over 17172.00 frames. ], tot_loss[loss=0.1895, ctc_loss=0.122, cr_loss=0.3378, over 3343457.70 frames. ], batch size: 45, lr: 2.97e-03, grad_scale: 32.0 2024-09-25 09:53:02,259 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=717710.0, ans=0.0 2024-09-25 09:53:13,592 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.max_abs, batch_count=717710.0, ans=10.0 2024-09-25 09:53:27,184 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=717756.6666666666, ans=0.2 2024-09-25 09:53:49,237 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten.whitening_limit, batch_count=717803.3333333334, ans=22.5 2024-09-25 09:53:55,655 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 09:54:00,036 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.246e+02 1.343e+02 1.431e+02 1.891e+02, threshold=2.686e+02, percent-clipped=0.0 2024-09-25 09:54:20,590 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=717896.6666666666, ans=0.0 2024-09-25 09:54:23,695 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=717896.6666666666, ans=0.1 2024-09-25 09:54:29,719 INFO [train.py:1198] (2/4) Epoch 40, batch 1900, loss[loss=0.1913, ctc_loss=0.1227, cr_loss=0.3431, over 17276.00 frames. ], tot_loss[loss=0.1906, ctc_loss=0.1226, cr_loss=0.3399, over 3356132.20 frames. ], batch size: 42, lr: 2.97e-03, grad_scale: 32.0 2024-09-25 09:54:32,120 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=12.40 vs. limit=15.0 2024-09-25 09:54:33,232 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=717943.3333333334, ans=0.125 2024-09-25 09:54:36,569 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=717943.3333333334, ans=0.125 2024-09-25 09:54:58,730 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=717990.0, ans=0.125 2024-09-25 09:55:19,512 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=718083.3333333334, ans=0.025 2024-09-25 09:55:19,555 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=718083.3333333334, ans=0.1 2024-09-25 09:55:40,370 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=718130.0, ans=0.125 2024-09-25 09:55:49,711 INFO [train.py:1198] (2/4) Epoch 40, batch 1950, loss[loss=0.1713, ctc_loss=0.1086, cr_loss=0.3135, over 17263.00 frames. ], tot_loss[loss=0.1905, ctc_loss=0.1225, cr_loss=0.3398, over 3358344.12 frames. ], batch size: 42, lr: 2.97e-03, grad_scale: 32.0 2024-09-25 09:55:56,356 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=718176.6666666666, ans=0.0 2024-09-25 09:55:59,578 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=718176.6666666666, ans=0.125 2024-09-25 09:56:01,347 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=718176.6666666666, ans=0.2 2024-09-25 09:56:29,550 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=718270.0, ans=0.0 2024-09-25 09:56:44,047 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=718316.6666666666, ans=0.1 2024-09-25 09:56:45,247 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.303e+02 1.365e+02 1.463e+02 2.159e+02, threshold=2.730e+02, percent-clipped=0.0 2024-09-25 09:56:45,634 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=718316.6666666666, ans=0.125 2024-09-25 09:56:51,986 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=718316.6666666666, ans=0.2 2024-09-25 09:56:53,971 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.62 vs. limit=15.0 2024-09-25 09:57:12,465 INFO [train.py:1198] (2/4) Epoch 40, batch 2000, loss[loss=0.2103, ctc_loss=0.1367, cr_loss=0.3677, over 16851.00 frames. ], tot_loss[loss=0.1902, ctc_loss=0.1223, cr_loss=0.3392, over 3368890.18 frames. ], batch size: 58, lr: 2.97e-03, grad_scale: 32.0 2024-09-25 09:57:47,942 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=718503.3333333334, ans=0.05 2024-09-25 09:58:00,788 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=718550.0, ans=0.0 2024-09-25 09:58:25,607 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=718596.6666666666, ans=0.0 2024-09-25 09:58:37,830 INFO [train.py:1198] (2/4) Epoch 40, batch 2050, loss[loss=0.203, ctc_loss=0.1314, cr_loss=0.3582, over 16748.00 frames. ], tot_loss[loss=0.1901, ctc_loss=0.1223, cr_loss=0.3394, over 3366873.99 frames. ], batch size: 61, lr: 2.97e-03, grad_scale: 32.0 2024-09-25 09:58:40,677 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=11.29 vs. limit=22.5 2024-09-25 09:58:51,125 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=718643.3333333334, ans=0.1 2024-09-25 09:58:52,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=718690.0, ans=0.035 2024-09-25 09:58:52,747 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=718690.0, ans=0.125 2024-09-25 09:59:08,903 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.15 vs. limit=15.0 2024-09-25 09:59:26,845 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=718783.3333333334, ans=0.0 2024-09-25 09:59:33,096 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.276e+02 1.355e+02 1.490e+02 2.224e+02, threshold=2.711e+02, percent-clipped=0.0 2024-09-25 09:59:52,528 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=718830.0, ans=0.125 2024-09-25 10:00:00,330 INFO [train.py:1198] (2/4) Epoch 40, batch 2100, loss[loss=0.1907, ctc_loss=0.1243, cr_loss=0.3318, over 17192.00 frames. ], tot_loss[loss=0.1902, ctc_loss=0.1223, cr_loss=0.3396, over 3370364.82 frames. ], batch size: 45, lr: 2.97e-03, grad_scale: 32.0 2024-09-25 10:00:04,018 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=718876.6666666666, ans=0.125 2024-09-25 10:00:08,695 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=718876.6666666666, ans=0.2 2024-09-25 10:00:13,605 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=718876.6666666666, ans=0.125 2024-09-25 10:00:18,485 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=718923.3333333334, ans=0.0 2024-09-25 10:00:28,087 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=718923.3333333334, ans=0.125 2024-09-25 10:01:01,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=719016.6666666666, ans=0.1 2024-09-25 10:01:15,201 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.94 vs. limit=15.0 2024-09-25 10:01:17,574 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=719063.3333333334, ans=0.0 2024-09-25 10:01:20,394 INFO [train.py:1198] (2/4) Epoch 40, batch 2150, loss[loss=0.1677, ctc_loss=0.1065, cr_loss=0.306, over 17073.00 frames. ], tot_loss[loss=0.1901, ctc_loss=0.1222, cr_loss=0.3394, over 3373370.28 frames. ], batch size: 39, lr: 2.97e-03, grad_scale: 32.0 2024-09-25 10:01:47,286 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=719156.6666666666, ans=0.125 2024-09-25 10:01:47,394 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=719156.6666666666, ans=0.0 2024-09-25 10:02:16,124 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.106e+02 1.284e+02 1.361e+02 1.483e+02 2.210e+02, threshold=2.722e+02, percent-clipped=0.0 2024-09-25 10:02:23,074 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 10:02:33,355 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=4.07 vs. limit=15.0 2024-09-25 10:02:43,177 INFO [train.py:1198] (2/4) Epoch 40, batch 2200, loss[loss=0.2197, ctc_loss=0.144, cr_loss=0.3784, over 17012.00 frames. ], tot_loss[loss=0.1914, ctc_loss=0.1232, cr_loss=0.3411, over 3361191.56 frames. ], batch size: 52, lr: 2.97e-03, grad_scale: 32.0 2024-09-25 10:03:32,264 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=719483.3333333334, ans=0.125 2024-09-25 10:03:54,643 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.whiten.whitening_limit, batch_count=719530.0, ans=12.0 2024-09-25 10:04:04,500 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.63 vs. limit=6.0 2024-09-25 10:04:08,273 INFO [train.py:1198] (2/4) Epoch 40, batch 2250, loss[loss=0.223, ctc_loss=0.151, cr_loss=0.3599, over 12094.00 frames. ], tot_loss[loss=0.1918, ctc_loss=0.1235, cr_loss=0.3418, over 3362705.37 frames. ], batch size: 123, lr: 2.97e-03, grad_scale: 32.0 2024-09-25 10:04:43,738 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=719670.0, ans=0.1 2024-09-25 10:04:57,958 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=719716.6666666666, ans=0.2 2024-09-25 10:04:59,856 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=19.42 vs. limit=22.5 2024-09-25 10:05:04,026 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.272e+02 1.370e+02 1.440e+02 2.779e+02, threshold=2.741e+02, percent-clipped=1.0 2024-09-25 10:05:09,176 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=719716.6666666666, ans=0.125 2024-09-25 10:05:13,991 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=719763.3333333334, ans=0.125 2024-09-25 10:05:18,821 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=719763.3333333334, ans=0.125 2024-09-25 10:05:31,318 INFO [train.py:1198] (2/4) Epoch 40, batch 2300, loss[loss=0.1717, ctc_loss=0.1078, cr_loss=0.3195, over 16951.00 frames. ], tot_loss[loss=0.1921, ctc_loss=0.1236, cr_loss=0.3421, over 3359940.91 frames. ], batch size: 42, lr: 2.96e-03, grad_scale: 32.0 2024-09-25 10:05:34,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=719810.0, ans=0.0 2024-09-25 10:05:59,102 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=719856.6666666666, ans=0.2 2024-09-25 10:06:54,043 INFO [train.py:1198] (2/4) Epoch 40, batch 2350, loss[loss=0.213, ctc_loss=0.1395, cr_loss=0.3675, over 17307.00 frames. ], tot_loss[loss=0.193, ctc_loss=0.1244, cr_loss=0.3431, over 3359302.23 frames. ], batch size: 51, lr: 2.96e-03, grad_scale: 32.0 2024-09-25 10:07:15,362 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=720090.0, ans=0.125 2024-09-25 10:07:17,467 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=15.06 vs. limit=22.5 2024-09-25 10:07:46,591 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.135e+02 1.272e+02 1.331e+02 1.456e+02 1.927e+02, threshold=2.662e+02, percent-clipped=0.0 2024-09-25 10:08:16,522 INFO [train.py:1198] (2/4) Epoch 40, batch 2400, loss[loss=0.2183, ctc_loss=0.1437, cr_loss=0.3731, over 17031.00 frames. ], tot_loss[loss=0.1938, ctc_loss=0.1251, cr_loss=0.344, over 3349417.44 frames. ], batch size: 52, lr: 2.96e-03, grad_scale: 32.0 2024-09-25 10:08:43,248 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=720323.3333333334, ans=0.1 2024-09-25 10:09:16,606 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=720416.6666666666, ans=0.0 2024-09-25 10:09:39,422 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.whiten, num_groups=1, num_channels=192, metric=3.99 vs. limit=12.0 2024-09-25 10:09:41,532 INFO [train.py:1198] (2/4) Epoch 40, batch 2450, loss[loss=0.199, ctc_loss=0.1284, cr_loss=0.3526, over 16975.00 frames. ], tot_loss[loss=0.1931, ctc_loss=0.1245, cr_loss=0.3432, over 3347356.44 frames. ], batch size: 58, lr: 2.96e-03, grad_scale: 32.0 2024-09-25 10:09:48,810 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.31 vs. limit=6.0 2024-09-25 10:09:50,102 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=720510.0, ans=0.0 2024-09-25 10:10:04,670 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=720556.6666666666, ans=0.2 2024-09-25 10:10:07,798 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=720556.6666666666, ans=0.125 2024-09-25 10:10:22,343 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=720603.3333333334, ans=0.125 2024-09-25 10:10:22,998 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=9.12 vs. limit=15.0 2024-09-25 10:10:30,450 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=720650.0, ans=0.1 2024-09-25 10:10:34,963 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.126e+02 1.276e+02 1.362e+02 1.464e+02 1.936e+02, threshold=2.724e+02, percent-clipped=0.0 2024-09-25 10:10:40,605 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=4.77 vs. limit=15.0 2024-09-25 10:10:41,732 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=720650.0, ans=0.125 2024-09-25 10:10:42,183 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=4.83 vs. limit=15.0 2024-09-25 10:10:54,347 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=720696.6666666666, ans=0.1 2024-09-25 10:11:02,046 INFO [train.py:1198] (2/4) Epoch 40, batch 2500, loss[loss=0.1791, ctc_loss=0.1145, cr_loss=0.3229, over 17060.00 frames. ], tot_loss[loss=0.1924, ctc_loss=0.124, cr_loss=0.3419, over 3351640.71 frames. ], batch size: 46, lr: 2.96e-03, grad_scale: 32.0 2024-09-25 10:11:30,925 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=11.67 vs. limit=22.5 2024-09-25 10:11:33,820 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=720790.0, ans=0.125 2024-09-25 10:11:56,661 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.90 vs. limit=6.0 2024-09-25 10:12:18,806 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=720930.0, ans=0.125 2024-09-25 10:12:24,938 INFO [train.py:1198] (2/4) Epoch 40, batch 2550, loss[loss=0.1537, ctc_loss=0.09768, cr_loss=0.2802, over 17177.00 frames. ], tot_loss[loss=0.1913, ctc_loss=0.1232, cr_loss=0.3403, over 3360265.81 frames. ], batch size: 41, lr: 2.96e-03, grad_scale: 32.0 2024-09-25 10:12:25,286 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=720976.6666666666, ans=0.125 2024-09-25 10:13:20,274 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.160e+02 1.334e+02 1.455e+02 1.593e+02 2.101e+02, threshold=2.910e+02, percent-clipped=0.0 2024-09-25 10:13:27,535 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=8.94 vs. limit=15.0 2024-09-25 10:13:29,272 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=19.48 vs. limit=22.5 2024-09-25 10:13:41,240 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.97 vs. limit=6.0 2024-09-25 10:13:47,325 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=721163.3333333334, ans=0.125 2024-09-25 10:13:50,327 INFO [train.py:1198] (2/4) Epoch 40, batch 2600, loss[loss=0.1845, ctc_loss=0.1178, cr_loss=0.3337, over 17010.00 frames. ], tot_loss[loss=0.192, ctc_loss=0.1238, cr_loss=0.3411, over 3341758.82 frames. ], batch size: 51, lr: 2.96e-03, grad_scale: 32.0 2024-09-25 10:14:31,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=721303.3333333334, ans=0.2 2024-09-25 10:14:36,961 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 10:14:55,934 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=721396.6666666666, ans=0.125 2024-09-25 10:15:13,246 INFO [train.py:1198] (2/4) Epoch 40, batch 2650, loss[loss=0.1946, ctc_loss=0.125, cr_loss=0.3481, over 17237.00 frames. ], tot_loss[loss=0.193, ctc_loss=0.1244, cr_loss=0.3428, over 3346710.10 frames. ], batch size: 50, lr: 2.96e-03, grad_scale: 32.0 2024-09-25 10:15:27,207 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=13.06 vs. limit=15.0 2024-09-25 10:15:29,939 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=721490.0, ans=0.125 2024-09-25 10:15:33,126 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=721490.0, ans=0.0 2024-09-25 10:15:54,059 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=721536.6666666666, ans=0.125 2024-09-25 10:15:54,886 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.58 vs. limit=6.0 2024-09-25 10:16:06,673 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.132e+02 1.318e+02 1.393e+02 1.493e+02 1.834e+02, threshold=2.785e+02, percent-clipped=0.0 2024-09-25 10:16:36,435 INFO [train.py:1198] (2/4) Epoch 40, batch 2700, loss[loss=0.1796, ctc_loss=0.1142, cr_loss=0.3272, over 17214.00 frames. ], tot_loss[loss=0.1914, ctc_loss=0.1233, cr_loss=0.3408, over 3350223.39 frames. ], batch size: 47, lr: 2.96e-03, grad_scale: 32.0 2024-09-25 10:16:39,887 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=721676.6666666666, ans=0.0 2024-09-25 10:17:05,729 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=721723.3333333334, ans=0.125 2024-09-25 10:17:12,381 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=721770.0, ans=0.1 2024-09-25 10:17:37,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=721816.6666666666, ans=0.0 2024-09-25 10:17:56,552 INFO [train.py:1198] (2/4) Epoch 40, batch 2750, loss[loss=0.2181, ctc_loss=0.1438, cr_loss=0.3715, over 14930.00 frames. ], tot_loss[loss=0.1925, ctc_loss=0.124, cr_loss=0.3424, over 3352906.86 frames. ], batch size: 89, lr: 2.96e-03, grad_scale: 32.0 2024-09-25 10:18:15,343 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=721956.6666666666, ans=0.0 2024-09-25 10:18:18,661 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=721956.6666666666, ans=0.2 2024-09-25 10:18:35,530 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=722003.3333333334, ans=0.0 2024-09-25 10:18:41,757 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=722003.3333333334, ans=0.1 2024-09-25 10:18:44,914 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=722003.3333333334, ans=0.0 2024-09-25 10:18:54,278 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.078e+02 1.293e+02 1.370e+02 1.533e+02 1.979e+02, threshold=2.741e+02, percent-clipped=0.0 2024-09-25 10:19:21,537 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=2.44 vs. limit=6.0 2024-09-25 10:19:24,054 INFO [train.py:1198] (2/4) Epoch 40, batch 2800, loss[loss=0.1692, ctc_loss=0.1081, cr_loss=0.3059, over 17036.00 frames. ], tot_loss[loss=0.1902, ctc_loss=0.1225, cr_loss=0.3388, over 3364867.04 frames. ], batch size: 39, lr: 2.96e-03, grad_scale: 32.0 2024-09-25 10:19:34,031 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=722143.3333333334, ans=0.125 2024-09-25 10:20:09,320 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=722236.6666666666, ans=0.125 2024-09-25 10:20:44,359 INFO [train.py:1198] (2/4) Epoch 40, batch 2850, loss[loss=0.1712, ctc_loss=0.1093, cr_loss=0.3092, over 17266.00 frames. ], tot_loss[loss=0.1911, ctc_loss=0.1231, cr_loss=0.34, over 3360848.27 frames. ], batch size: 46, lr: 2.96e-03, grad_scale: 32.0 2024-09-25 10:21:40,028 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.145e+02 1.304e+02 1.350e+02 1.486e+02 2.860e+02, threshold=2.699e+02, percent-clipped=1.0 2024-09-25 10:22:01,380 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=722563.3333333334, ans=0.125 2024-09-25 10:22:03,010 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=722563.3333333334, ans=0.0 2024-09-25 10:22:07,585 INFO [train.py:1198] (2/4) Epoch 40, batch 2900, loss[loss=0.1657, ctc_loss=0.1022, cr_loss=0.3173, over 17037.00 frames. ], tot_loss[loss=0.1907, ctc_loss=0.1228, cr_loss=0.3394, over 3362699.49 frames. ], batch size: 44, lr: 2.96e-03, grad_scale: 32.0 2024-09-25 10:22:31,760 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 10:22:32,485 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.16 vs. limit=15.0 2024-09-25 10:22:33,556 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 10:22:38,318 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=722703.3333333334, ans=0.125 2024-09-25 10:23:14,707 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=722796.6666666666, ans=0.0 2024-09-25 10:23:20,905 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=722796.6666666666, ans=0.0 2024-09-25 10:23:31,489 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=722843.3333333334, ans=0.125 2024-09-25 10:23:32,768 INFO [train.py:1198] (2/4) Epoch 40, batch 2950, loss[loss=0.2175, ctc_loss=0.1448, cr_loss=0.3634, over 17354.00 frames. ], tot_loss[loss=0.1904, ctc_loss=0.1226, cr_loss=0.3388, over 3353189.15 frames. ], batch size: 48, lr: 2.96e-03, grad_scale: 32.0 2024-09-25 10:23:37,671 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=722843.3333333334, ans=0.125 2024-09-25 10:23:38,374 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.63 vs. limit=15.0 2024-09-25 10:23:45,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=722843.3333333334, ans=0.0 2024-09-25 10:23:52,380 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=722890.0, ans=0.125 2024-09-25 10:23:58,162 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=722890.0, ans=0.0 2024-09-25 10:24:25,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten.whitening_limit, batch_count=722983.3333333334, ans=15.0 2024-09-25 10:24:27,665 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.163e+02 1.302e+02 1.375e+02 1.472e+02 2.387e+02, threshold=2.750e+02, percent-clipped=0.0 2024-09-25 10:24:49,710 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=723030.0, ans=0.025 2024-09-25 10:24:53,998 INFO [train.py:1198] (2/4) Epoch 40, batch 3000, loss[loss=0.2039, ctc_loss=0.1321, cr_loss=0.3592, over 15786.00 frames. ], tot_loss[loss=0.19, ctc_loss=0.1224, cr_loss=0.3379, over 3351991.59 frames. ], batch size: 74, lr: 2.96e-03, grad_scale: 32.0 2024-09-25 10:24:53,999 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 10:25:09,254 INFO [train.py:1230] (2/4) Epoch 40, validation: loss=0.03571, ctc_loss=0.03571, cr_loss=9.785e-15, over 944034.00 frames. 2024-09-25 10:25:09,254 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 10:25:14,251 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=723076.6666666666, ans=0.0 2024-09-25 10:25:19,280 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=2.93 vs. limit=15.0 2024-09-25 10:25:33,419 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=8.79 vs. limit=15.0 2024-09-25 10:25:46,724 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=723170.0, ans=0.125 2024-09-25 10:25:53,082 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=723170.0, ans=0.1 2024-09-25 10:26:13,619 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=723263.3333333334, ans=0.0 2024-09-25 10:26:19,784 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=723263.3333333334, ans=0.0 2024-09-25 10:26:27,243 INFO [train.py:1198] (2/4) Epoch 40, batch 3050, loss[loss=0.1891, ctc_loss=0.1208, cr_loss=0.3416, over 17352.00 frames. ], tot_loss[loss=0.1893, ctc_loss=0.1219, cr_loss=0.3371, over 3359113.06 frames. ], batch size: 48, lr: 2.96e-03, grad_scale: 16.0 2024-09-25 10:26:46,513 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=723356.6666666666, ans=0.125 2024-09-25 10:26:51,115 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=723356.6666666666, ans=0.1 2024-09-25 10:26:58,984 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=723403.3333333334, ans=0.0 2024-09-25 10:26:59,002 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=723403.3333333334, ans=0.125 2024-09-25 10:27:20,628 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.084e+02 1.265e+02 1.358e+02 1.442e+02 1.711e+02, threshold=2.717e+02, percent-clipped=0.0 2024-09-25 10:27:24,420 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=11.96 vs. limit=22.5 2024-09-25 10:27:27,102 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=723450.0, ans=0.1 2024-09-25 10:27:28,691 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=723496.6666666666, ans=0.125 2024-09-25 10:27:45,677 INFO [train.py:1198] (2/4) Epoch 40, batch 3100, loss[loss=0.1776, ctc_loss=0.114, cr_loss=0.3181, over 16846.00 frames. ], tot_loss[loss=0.1893, ctc_loss=0.1218, cr_loss=0.3375, over 3370286.98 frames. ], batch size: 58, lr: 2.96e-03, grad_scale: 16.0 2024-09-25 10:27:46,023 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=723543.3333333334, ans=0.2 2024-09-25 10:27:57,024 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=723543.3333333334, ans=0.0 2024-09-25 10:27:59,202 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.17 vs. limit=22.5 2024-09-25 10:28:05,135 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.04 vs. limit=15.0 2024-09-25 10:28:47,753 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 10:29:06,037 INFO [train.py:1198] (2/4) Epoch 40, batch 3150, loss[loss=0.1667, ctc_loss=0.1049, cr_loss=0.3093, over 17315.00 frames. ], tot_loss[loss=0.1899, ctc_loss=0.1222, cr_loss=0.3381, over 3372562.81 frames. ], batch size: 46, lr: 2.96e-03, grad_scale: 16.0 2024-09-25 10:29:26,674 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=723823.3333333334, ans=0.05 2024-09-25 10:29:27,205 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.18 vs. limit=15.0 2024-09-25 10:29:34,282 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=723823.3333333334, ans=0.1 2024-09-25 10:29:35,993 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=723870.0, ans=0.125 2024-09-25 10:29:36,213 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=12.13 vs. limit=12.0 2024-09-25 10:30:00,972 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.278e+02 1.373e+02 1.497e+02 1.845e+02, threshold=2.745e+02, percent-clipped=0.0 2024-09-25 10:30:01,258 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=723916.6666666666, ans=0.125 2024-09-25 10:30:24,823 INFO [train.py:1198] (2/4) Epoch 40, batch 3200, loss[loss=0.1931, ctc_loss=0.1231, cr_loss=0.35, over 17208.00 frames. ], tot_loss[loss=0.1902, ctc_loss=0.1225, cr_loss=0.3384, over 3357850.42 frames. ], batch size: 50, lr: 2.96e-03, grad_scale: 16.0 2024-09-25 10:30:26,613 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.max_abs, batch_count=724010.0, ans=10.0 2024-09-25 10:31:10,415 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.max_positive, batch_count=724150.0, ans=0.95 2024-09-25 10:31:43,139 INFO [train.py:1198] (2/4) Epoch 40, batch 3250, loss[loss=0.188, ctc_loss=0.1208, cr_loss=0.3361, over 17236.00 frames. ], tot_loss[loss=0.1904, ctc_loss=0.1227, cr_loss=0.3386, over 3344474.34 frames. ], batch size: 50, lr: 2.96e-03, grad_scale: 16.0 2024-09-25 10:32:15,674 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=724336.6666666666, ans=0.125 2024-09-25 10:32:31,700 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.18 vs. limit=15.0 2024-09-25 10:32:42,129 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.132e+02 1.319e+02 1.415e+02 1.534e+02 3.607e+02, threshold=2.830e+02, percent-clipped=1.0 2024-09-25 10:32:45,380 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=724383.3333333334, ans=0.125 2024-09-25 10:32:53,082 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=724430.0, ans=0.125 2024-09-25 10:33:05,491 INFO [train.py:1198] (2/4) Epoch 40, batch 3300, loss[loss=0.164, ctc_loss=0.1023, cr_loss=0.3082, over 17195.00 frames. ], tot_loss[loss=0.1907, ctc_loss=0.1229, cr_loss=0.3389, over 3348914.84 frames. ], batch size: 41, lr: 2.96e-03, grad_scale: 16.0 2024-09-25 10:33:15,219 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=724476.6666666666, ans=0.1 2024-09-25 10:33:28,006 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.36 vs. limit=15.0 2024-09-25 10:33:59,448 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=5.10 vs. limit=12.0 2024-09-25 10:34:20,920 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=724663.3333333334, ans=0.125 2024-09-25 10:34:22,797 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=5.21 vs. limit=10.0 2024-09-25 10:34:25,460 INFO [train.py:1198] (2/4) Epoch 40, batch 3350, loss[loss=0.2062, ctc_loss=0.1353, cr_loss=0.3546, over 16943.00 frames. ], tot_loss[loss=0.1911, ctc_loss=0.1232, cr_loss=0.3396, over 3343339.29 frames. ], batch size: 58, lr: 2.95e-03, grad_scale: 16.0 2024-09-25 10:34:33,593 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=724710.0, ans=0.125 2024-09-25 10:34:33,977 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.whiten, num_groups=1, num_channels=512, metric=9.90 vs. limit=12.0 2024-09-25 10:34:58,434 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=724803.3333333334, ans=0.1 2024-09-25 10:35:19,910 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.291e+02 1.364e+02 1.477e+02 1.997e+02, threshold=2.727e+02, percent-clipped=0.0 2024-09-25 10:35:43,179 INFO [train.py:1198] (2/4) Epoch 40, batch 3400, loss[loss=0.1955, ctc_loss=0.1275, cr_loss=0.3398, over 16871.00 frames. ], tot_loss[loss=0.1896, ctc_loss=0.1221, cr_loss=0.3377, over 3352939.90 frames. ], batch size: 58, lr: 2.95e-03, grad_scale: 16.0 2024-09-25 10:35:52,765 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=724943.3333333334, ans=0.0 2024-09-25 10:36:16,281 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=725036.6666666666, ans=0.125 2024-09-25 10:36:33,508 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=725083.3333333334, ans=0.0 2024-09-25 10:36:39,719 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=725083.3333333334, ans=0.125 2024-09-25 10:36:56,932 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=725130.0, ans=0.0 2024-09-25 10:37:01,219 INFO [train.py:1198] (2/4) Epoch 40, batch 3450, loss[loss=0.2118, ctc_loss=0.1397, cr_loss=0.3607, over 16517.00 frames. ], tot_loss[loss=0.1905, ctc_loss=0.1227, cr_loss=0.339, over 3353359.09 frames. ], batch size: 66, lr: 2.95e-03, grad_scale: 16.0 2024-09-25 10:37:25,394 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=725223.3333333334, ans=0.125 2024-09-25 10:37:56,649 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.302e+02 1.377e+02 1.496e+02 2.659e+02, threshold=2.753e+02, percent-clipped=0.0 2024-09-25 10:38:04,866 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=725363.3333333334, ans=0.125 2024-09-25 10:38:11,244 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.min_positive, batch_count=725363.3333333334, ans=0.025 2024-09-25 10:38:22,231 INFO [train.py:1198] (2/4) Epoch 40, batch 3500, loss[loss=0.2223, ctc_loss=0.1475, cr_loss=0.3742, over 12037.00 frames. ], tot_loss[loss=0.1918, ctc_loss=0.1237, cr_loss=0.3404, over 3343662.02 frames. ], batch size: 123, lr: 2.95e-03, grad_scale: 16.0 2024-09-25 10:38:31,846 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=725410.0, ans=0.125 2024-09-25 10:38:47,950 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=5.50 vs. limit=15.0 2024-09-25 10:39:16,863 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=725550.0, ans=0.0 2024-09-25 10:39:26,036 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=725596.6666666666, ans=0.1 2024-09-25 10:39:27,894 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=725596.6666666666, ans=0.125 2024-09-25 10:39:31,002 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=725596.6666666666, ans=0.0 2024-09-25 10:39:32,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=725596.6666666666, ans=0.0 2024-09-25 10:39:39,941 INFO [train.py:1198] (2/4) Epoch 40, batch 3550, loss[loss=0.181, ctc_loss=0.1153, cr_loss=0.3284, over 17108.00 frames. ], tot_loss[loss=0.1913, ctc_loss=0.1233, cr_loss=0.3396, over 3348899.55 frames. ], batch size: 40, lr: 2.95e-03, grad_scale: 16.0 2024-09-25 10:39:40,461 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=8.99 vs. limit=15.0 2024-09-25 10:39:44,920 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=725643.3333333334, ans=0.125 2024-09-25 10:40:34,532 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.284e+02 1.346e+02 1.448e+02 2.075e+02, threshold=2.692e+02, percent-clipped=0.0 2024-09-25 10:40:58,147 INFO [train.py:1198] (2/4) Epoch 40, batch 3600, loss[loss=0.2215, ctc_loss=0.1459, cr_loss=0.3779, over 14821.00 frames. ], tot_loss[loss=0.1906, ctc_loss=0.1228, cr_loss=0.3387, over 3347036.53 frames. ], batch size: 89, lr: 2.95e-03, grad_scale: 32.0 2024-09-25 10:40:58,449 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=725876.6666666666, ans=0.125 2024-09-25 10:41:33,344 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=7.09 vs. limit=15.0 2024-09-25 10:42:11,902 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=9.32 vs. limit=10.0 2024-09-25 10:42:20,143 INFO [train.py:1198] (2/4) Epoch 40, batch 3650, loss[loss=0.1913, ctc_loss=0.1249, cr_loss=0.3323, over 17349.00 frames. ], tot_loss[loss=0.1904, ctc_loss=0.1227, cr_loss=0.3383, over 3336641.10 frames. ], batch size: 48, lr: 2.95e-03, grad_scale: 16.0 2024-09-25 10:42:30,103 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=726110.0, ans=0.125 2024-09-25 10:42:33,054 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=726110.0, ans=0.125 2024-09-25 10:42:41,201 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=726156.6666666666, ans=0.125 2024-09-25 10:43:06,010 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.82 vs. limit=10.0 2024-09-25 10:43:12,405 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=10.35 vs. limit=15.0 2024-09-25 10:43:19,107 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.099e+02 1.298e+02 1.359e+02 1.454e+02 1.962e+02, threshold=2.718e+02, percent-clipped=0.0 2024-09-25 10:43:35,216 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=726296.6666666666, ans=0.125 2024-09-25 10:43:38,383 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=726296.6666666666, ans=0.1 2024-09-25 10:43:41,335 INFO [train.py:1198] (2/4) Epoch 40, batch 3700, loss[loss=0.2091, ctc_loss=0.1359, cr_loss=0.3657, over 17041.00 frames. ], tot_loss[loss=0.1914, ctc_loss=0.1234, cr_loss=0.34, over 3333990.72 frames. ], batch size: 52, lr: 2.95e-03, grad_scale: 16.0 2024-09-25 10:43:41,642 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=726343.3333333334, ans=0.1 2024-09-25 10:44:28,720 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=726483.3333333334, ans=0.0 2024-09-25 10:44:41,376 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=726483.3333333334, ans=0.125 2024-09-25 10:44:59,749 INFO [train.py:1198] (2/4) Epoch 40, batch 3750, loss[loss=0.198, ctc_loss=0.128, cr_loss=0.3501, over 17033.00 frames. ], tot_loss[loss=0.1921, ctc_loss=0.1239, cr_loss=0.341, over 3340496.73 frames. ], batch size: 51, lr: 2.95e-03, grad_scale: 16.0 2024-09-25 10:45:18,810 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=726623.3333333334, ans=0.125 2024-09-25 10:45:23,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=726623.3333333334, ans=0.125 2024-09-25 10:45:27,838 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=726623.3333333334, ans=0.025 2024-09-25 10:45:40,334 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=726670.0, ans=0.09899494936611666 2024-09-25 10:45:55,148 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.071e+02 1.300e+02 1.354e+02 1.476e+02 2.861e+02, threshold=2.708e+02, percent-clipped=2.0 2024-09-25 10:46:12,170 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=726763.3333333334, ans=0.125 2024-09-25 10:46:16,509 INFO [train.py:1198] (2/4) Epoch 40, batch 3800, loss[loss=0.1892, ctc_loss=0.1242, cr_loss=0.3251, over 17036.00 frames. ], tot_loss[loss=0.1915, ctc_loss=0.1234, cr_loss=0.3404, over 3341101.03 frames. ], batch size: 44, lr: 2.95e-03, grad_scale: 16.0 2024-09-25 10:46:19,975 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=726810.0, ans=0.2 2024-09-25 10:46:49,796 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=9.13 vs. limit=15.0 2024-09-25 10:46:54,540 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten.whitening_limit, batch_count=726903.3333333334, ans=15.0 2024-09-25 10:47:06,481 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=726950.0, ans=0.125 2024-09-25 10:47:14,148 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=726950.0, ans=0.125 2024-09-25 10:47:18,606 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=726996.6666666666, ans=0.0 2024-09-25 10:47:24,258 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=5.28 vs. limit=15.0 2024-09-25 10:47:28,124 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=726996.6666666666, ans=0.0 2024-09-25 10:47:34,473 INFO [train.py:1198] (2/4) Epoch 40, batch 3850, loss[loss=0.2394, ctc_loss=0.1582, cr_loss=0.4059, over 15130.00 frames. ], tot_loss[loss=0.1919, ctc_loss=0.1239, cr_loss=0.3403, over 3300646.64 frames. ], batch size: 89, lr: 2.95e-03, grad_scale: 16.0 2024-09-25 10:47:49,387 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.whiten.whitening_limit, batch_count=727090.0, ans=15.0 2024-09-25 10:47:53,847 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=16.04 vs. limit=22.5 2024-09-25 10:48:23,023 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=727183.3333333334, ans=0.125 2024-09-25 10:48:30,295 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.106e+02 1.319e+02 1.427e+02 1.593e+02 2.504e+02, threshold=2.853e+02, percent-clipped=0.0 2024-09-25 10:49:36,294 INFO [train.py:1198] (2/4) Epoch 41, batch 0, loss[loss=0.1575, ctc_loss=0.09766, cr_loss=0.2994, over 17030.00 frames. ], tot_loss[loss=0.1575, ctc_loss=0.09766, cr_loss=0.2994, over 17030.00 frames. ], batch size: 39, lr: 2.91e-03, grad_scale: 32.0 2024-09-25 10:49:36,295 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 10:49:51,731 INFO [train.py:1230] (2/4) Epoch 41, validation: loss=0.03537, ctc_loss=0.03537, cr_loss=1.035e-14, over 944034.00 frames. 2024-09-25 10:49:51,732 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 10:49:52,070 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=727258.0, ans=0.125 2024-09-25 10:50:32,367 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.min_positive, batch_count=727351.3333333334, ans=0.025 2024-09-25 10:50:34,544 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.51 vs. limit=15.0 2024-09-25 10:50:58,384 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.92 vs. limit=6.0 2024-09-25 10:51:15,276 INFO [train.py:1198] (2/4) Epoch 41, batch 50, loss[loss=0.1792, ctc_loss=0.111, cr_loss=0.3409, over 17251.00 frames. ], tot_loss[loss=0.1875, ctc_loss=0.1202, cr_loss=0.3364, over 763975.18 frames. ], batch size: 44, lr: 2.91e-03, grad_scale: 32.0 2024-09-25 10:51:28,401 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 10:51:53,877 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=727584.6666666666, ans=0.125 2024-09-25 10:52:00,531 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=727584.6666666666, ans=0.125 2024-09-25 10:52:19,248 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.297e+02 1.379e+02 1.480e+02 1.921e+02, threshold=2.758e+02, percent-clipped=0.0 2024-09-25 10:52:27,996 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=5.05 vs. limit=12.0 2024-09-25 10:52:35,283 INFO [train.py:1198] (2/4) Epoch 41, batch 100, loss[loss=0.1817, ctc_loss=0.1182, cr_loss=0.3174, over 17143.00 frames. ], tot_loss[loss=0.189, ctc_loss=0.1215, cr_loss=0.3378, over 1325395.50 frames. ], batch size: 48, lr: 2.91e-03, grad_scale: 32.0 2024-09-25 10:52:37,152 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=727724.6666666666, ans=0.0 2024-09-25 10:52:38,700 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=727724.6666666666, ans=0.1 2024-09-25 10:52:38,788 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=727724.6666666666, ans=0.125 2024-09-25 10:52:44,996 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=727724.6666666666, ans=0.125 2024-09-25 10:53:13,773 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=727818.0, ans=0.2 2024-09-25 10:53:48,761 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=727911.3333333334, ans=0.1 2024-09-25 10:54:00,638 INFO [train.py:1198] (2/4) Epoch 41, batch 150, loss[loss=0.2079, ctc_loss=0.1379, cr_loss=0.35, over 17003.00 frames. ], tot_loss[loss=0.1892, ctc_loss=0.1217, cr_loss=0.3377, over 1784098.51 frames. ], batch size: 53, lr: 2.91e-03, grad_scale: 32.0 2024-09-25 10:54:01,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=727958.0, ans=0.0 2024-09-25 10:54:32,563 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.39 vs. limit=6.0 2024-09-25 10:54:56,465 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=728098.0, ans=0.0 2024-09-25 10:55:01,195 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=728098.0, ans=0.1 2024-09-25 10:55:09,320 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 10:55:10,424 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.169e+02 1.287e+02 1.355e+02 1.476e+02 1.968e+02, threshold=2.710e+02, percent-clipped=0.0 2024-09-25 10:55:27,698 INFO [train.py:1198] (2/4) Epoch 41, batch 200, loss[loss=0.1787, ctc_loss=0.1119, cr_loss=0.3341, over 17086.00 frames. ], tot_loss[loss=0.1885, ctc_loss=0.1211, cr_loss=0.3368, over 2135810.14 frames. ], batch size: 49, lr: 2.91e-03, grad_scale: 16.0 2024-09-25 10:56:08,408 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 10:56:46,321 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=728424.6666666666, ans=0.0 2024-09-25 10:56:47,729 INFO [train.py:1198] (2/4) Epoch 41, batch 250, loss[loss=0.1588, ctc_loss=0.1022, cr_loss=0.2833, over 17128.00 frames. ], tot_loss[loss=0.1894, ctc_loss=0.1217, cr_loss=0.3384, over 2403312.61 frames. ], batch size: 40, lr: 2.91e-03, grad_scale: 16.0 2024-09-25 10:56:53,168 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=7.63 vs. limit=15.0 2024-09-25 10:56:53,181 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.08 vs. limit=15.0 2024-09-25 10:57:07,748 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=3.98 vs. limit=15.0 2024-09-25 10:57:34,406 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.18 vs. limit=22.5 2024-09-25 10:57:39,081 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=728564.6666666666, ans=0.125 2024-09-25 10:57:40,768 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=728564.6666666666, ans=0.125 2024-09-25 10:57:53,022 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.120e+02 1.260e+02 1.343e+02 1.433e+02 1.845e+02, threshold=2.685e+02, percent-clipped=0.0 2024-09-25 10:58:07,360 INFO [train.py:1198] (2/4) Epoch 41, batch 300, loss[loss=0.2069, ctc_loss=0.1344, cr_loss=0.3628, over 16674.00 frames. ], tot_loss[loss=0.1893, ctc_loss=0.1215, cr_loss=0.3388, over 2609002.12 frames. ], batch size: 61, lr: 2.91e-03, grad_scale: 16.0 2024-09-25 10:58:11,230 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=7.27 vs. limit=15.0 2024-09-25 10:58:25,497 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=11.94 vs. limit=15.0 2024-09-25 10:58:29,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=728704.6666666666, ans=0.0 2024-09-25 10:58:56,844 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer2.prob, batch_count=728751.3333333334, ans=0.125 2024-09-25 10:59:01,847 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=728798.0, ans=0.125 2024-09-25 10:59:03,742 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.55 vs. limit=10.0 2024-09-25 10:59:10,075 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=9.40 vs. limit=15.0 2024-09-25 10:59:26,835 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=728844.6666666666, ans=0.025 2024-09-25 10:59:31,707 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=728844.6666666666, ans=0.125 2024-09-25 10:59:36,173 INFO [train.py:1198] (2/4) Epoch 41, batch 350, loss[loss=0.1943, ctc_loss=0.1252, cr_loss=0.3455, over 17002.00 frames. ], tot_loss[loss=0.1894, ctc_loss=0.1217, cr_loss=0.3384, over 2775777.51 frames. ], batch size: 56, lr: 2.91e-03, grad_scale: 16.0 2024-09-25 11:00:44,742 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.120e+02 1.270e+02 1.343e+02 1.483e+02 2.666e+02, threshold=2.685e+02, percent-clipped=0.0 2024-09-25 11:00:59,199 INFO [train.py:1198] (2/4) Epoch 41, batch 400, loss[loss=0.1976, ctc_loss=0.1265, cr_loss=0.3558, over 17302.00 frames. ], tot_loss[loss=0.1904, ctc_loss=0.1224, cr_loss=0.3402, over 2889362.32 frames. ], batch size: 49, lr: 2.91e-03, grad_scale: 32.0 2024-09-25 11:01:18,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=729171.3333333334, ans=0.0 2024-09-25 11:01:33,066 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=729218.0, ans=0.125 2024-09-25 11:01:38,388 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=9.15 vs. limit=15.0 2024-09-25 11:01:52,264 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=729264.6666666666, ans=0.0 2024-09-25 11:02:03,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=729311.3333333334, ans=0.125 2024-09-25 11:02:07,428 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=12.34 vs. limit=15.0 2024-09-25 11:02:07,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=729311.3333333334, ans=0.125 2024-09-25 11:02:16,051 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=729311.3333333334, ans=0.125 2024-09-25 11:02:18,860 INFO [train.py:1198] (2/4) Epoch 41, batch 450, loss[loss=0.213, ctc_loss=0.1398, cr_loss=0.3657, over 16988.00 frames. ], tot_loss[loss=0.1908, ctc_loss=0.1228, cr_loss=0.3399, over 2987414.85 frames. ], batch size: 56, lr: 2.91e-03, grad_scale: 32.0 2024-09-25 11:02:34,434 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=15.53 vs. limit=22.5 2024-09-25 11:02:38,710 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=729404.6666666666, ans=0.0 2024-09-25 11:02:40,143 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=729404.6666666666, ans=0.025 2024-09-25 11:02:47,232 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.50 vs. limit=6.0 2024-09-25 11:02:49,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=729451.3333333334, ans=0.125 2024-09-25 11:02:54,702 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 11:03:20,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=729498.0, ans=0.0 2024-09-25 11:03:21,920 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=729544.6666666666, ans=0.125 2024-09-25 11:03:26,324 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.138e+02 1.254e+02 1.327e+02 1.463e+02 1.935e+02, threshold=2.654e+02, percent-clipped=0.0 2024-09-25 11:03:41,630 INFO [train.py:1198] (2/4) Epoch 41, batch 500, loss[loss=0.2214, ctc_loss=0.1452, cr_loss=0.3809, over 16591.00 frames. ], tot_loss[loss=0.1901, ctc_loss=0.1223, cr_loss=0.3388, over 3066694.30 frames. ], batch size: 66, lr: 2.91e-03, grad_scale: 16.0 2024-09-25 11:03:47,999 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=729591.3333333334, ans=0.125 2024-09-25 11:04:23,028 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=729684.6666666666, ans=0.09899494936611666 2024-09-25 11:04:38,913 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=729731.3333333334, ans=0.07 2024-09-25 11:04:42,059 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=729731.3333333334, ans=0.0 2024-09-25 11:05:09,614 INFO [train.py:1198] (2/4) Epoch 41, batch 550, loss[loss=0.1969, ctc_loss=0.1308, cr_loss=0.3305, over 17357.00 frames. ], tot_loss[loss=0.1908, ctc_loss=0.1228, cr_loss=0.3397, over 3133665.77 frames. ], batch size: 48, lr: 2.91e-03, grad_scale: 16.0 2024-09-25 11:05:40,770 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.43 vs. limit=15.0 2024-09-25 11:05:58,852 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.38 vs. limit=15.0 2024-09-25 11:06:18,377 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.102e+02 1.304e+02 1.377e+02 1.509e+02 2.527e+02, threshold=2.754e+02, percent-clipped=0.0 2024-09-25 11:06:20,343 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=730011.3333333334, ans=0.04949747468305833 2024-09-25 11:06:29,788 INFO [train.py:1198] (2/4) Epoch 41, batch 600, loss[loss=0.1874, ctc_loss=0.1229, cr_loss=0.3227, over 17302.00 frames. ], tot_loss[loss=0.191, ctc_loss=0.1231, cr_loss=0.3398, over 3183632.18 frames. ], batch size: 49, lr: 2.91e-03, grad_scale: 8.0 2024-09-25 11:06:30,147 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer_na.min_abs, batch_count=730058.0, ans=0.02 2024-09-25 11:06:54,299 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=12.40 vs. limit=12.0 2024-09-25 11:07:02,033 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=730151.3333333334, ans=0.0 2024-09-25 11:07:16,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=730198.0, ans=0.0 2024-09-25 11:07:30,835 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=730198.0, ans=0.2 2024-09-25 11:07:44,315 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.89 vs. limit=22.5 2024-09-25 11:07:49,972 INFO [train.py:1198] (2/4) Epoch 41, batch 650, loss[loss=0.1624, ctc_loss=0.1034, cr_loss=0.2949, over 17127.00 frames. ], tot_loss[loss=0.1899, ctc_loss=0.1222, cr_loss=0.3384, over 3223560.69 frames. ], batch size: 40, lr: 2.91e-03, grad_scale: 8.0 2024-09-25 11:08:14,667 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=730338.0, ans=0.0 2024-09-25 11:08:17,840 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=730338.0, ans=0.2 2024-09-25 11:09:06,925 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.080e+02 1.289e+02 1.341e+02 1.440e+02 1.849e+02, threshold=2.683e+02, percent-clipped=0.0 2024-09-25 11:09:07,950 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=10.70 vs. limit=15.0 2024-09-25 11:09:18,185 INFO [train.py:1198] (2/4) Epoch 41, batch 700, loss[loss=0.1732, ctc_loss=0.1094, cr_loss=0.3189, over 17107.00 frames. ], tot_loss[loss=0.1896, ctc_loss=0.1219, cr_loss=0.3387, over 3262106.72 frames. ], batch size: 40, lr: 2.91e-03, grad_scale: 8.0 2024-09-25 11:09:28,273 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=11.41 vs. limit=22.5 2024-09-25 11:09:44,196 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=730571.3333333334, ans=0.0 2024-09-25 11:10:03,184 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=730618.0, ans=0.125 2024-09-25 11:10:21,174 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=8.26 vs. limit=15.0 2024-09-25 11:10:25,192 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=730711.3333333334, ans=0.1 2024-09-25 11:10:25,654 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.24 vs. limit=15.0 2024-09-25 11:10:38,021 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=730711.3333333334, ans=0.1 2024-09-25 11:10:39,580 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=730758.0, ans=0.0 2024-09-25 11:10:40,884 INFO [train.py:1198] (2/4) Epoch 41, batch 750, loss[loss=0.1866, ctc_loss=0.1181, cr_loss=0.3428, over 17237.00 frames. ], tot_loss[loss=0.1896, ctc_loss=0.1219, cr_loss=0.3384, over 3277653.42 frames. ], batch size: 50, lr: 2.91e-03, grad_scale: 8.0 2024-09-25 11:10:41,281 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=730758.0, ans=0.0 2024-09-25 11:11:01,873 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=730804.6666666666, ans=0.09899494936611666 2024-09-25 11:11:48,154 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=730944.6666666666, ans=0.125 2024-09-25 11:11:49,459 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.142e+02 1.291e+02 1.357e+02 1.455e+02 2.788e+02, threshold=2.714e+02, percent-clipped=1.0 2024-09-25 11:11:56,225 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=730944.6666666666, ans=0.0 2024-09-25 11:11:56,407 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=730944.6666666666, ans=0.2 2024-09-25 11:12:00,722 INFO [train.py:1198] (2/4) Epoch 41, batch 800, loss[loss=0.1714, ctc_loss=0.107, cr_loss=0.3222, over 17056.00 frames. ], tot_loss[loss=0.1892, ctc_loss=0.1216, cr_loss=0.3381, over 3297119.50 frames. ], batch size: 39, lr: 2.91e-03, grad_scale: 16.0 2024-09-25 11:12:26,748 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=731038.0, ans=0.125 2024-09-25 11:13:11,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=731178.0, ans=0.0 2024-09-25 11:13:17,275 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn2.whiten, num_groups=1, num_channels=192, metric=8.86 vs. limit=22.5 2024-09-25 11:13:21,001 INFO [train.py:1198] (2/4) Epoch 41, batch 850, loss[loss=0.1795, ctc_loss=0.1165, cr_loss=0.3152, over 17328.00 frames. ], tot_loss[loss=0.1887, ctc_loss=0.1212, cr_loss=0.3375, over 3319504.55 frames. ], batch size: 51, lr: 2.91e-03, grad_scale: 16.0 2024-09-25 11:14:00,297 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.75 vs. limit=6.0 2024-09-25 11:14:27,932 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=731364.6666666666, ans=0.125 2024-09-25 11:14:29,421 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=731364.6666666666, ans=0.0 2024-09-25 11:14:37,169 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.274e+02 1.366e+02 1.473e+02 2.977e+02, threshold=2.732e+02, percent-clipped=1.0 2024-09-25 11:14:48,456 INFO [train.py:1198] (2/4) Epoch 41, batch 900, loss[loss=0.2018, ctc_loss=0.1314, cr_loss=0.3523, over 17101.00 frames. ], tot_loss[loss=0.1881, ctc_loss=0.1208, cr_loss=0.3366, over 3330395.74 frames. ], batch size: 49, lr: 2.90e-03, grad_scale: 16.0 2024-09-25 11:14:56,540 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=731458.0, ans=0.2 2024-09-25 11:15:04,711 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=731504.6666666666, ans=0.125 2024-09-25 11:15:21,745 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=731551.3333333334, ans=0.125 2024-09-25 11:15:23,383 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=731551.3333333334, ans=0.125 2024-09-25 11:15:30,253 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=8.27 vs. limit=15.0 2024-09-25 11:15:36,347 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=731551.3333333334, ans=0.125 2024-09-25 11:15:38,054 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=731598.0, ans=0.07 2024-09-25 11:15:39,550 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=731598.0, ans=0.07 2024-09-25 11:16:10,879 INFO [train.py:1198] (2/4) Epoch 41, batch 950, loss[loss=0.1649, ctc_loss=0.1019, cr_loss=0.3147, over 17098.00 frames. ], tot_loss[loss=0.1873, ctc_loss=0.1202, cr_loss=0.3356, over 3348067.35 frames. ], batch size: 43, lr: 2.90e-03, grad_scale: 16.0 2024-09-25 11:16:16,283 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=731691.3333333334, ans=0.2 2024-09-25 11:17:15,925 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=20.43 vs. limit=22.5 2024-09-25 11:17:19,942 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.177e+02 1.310e+02 1.391e+02 1.460e+02 2.998e+02, threshold=2.782e+02, percent-clipped=1.0 2024-09-25 11:17:22,137 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=10.41 vs. limit=12.0 2024-09-25 11:17:31,152 INFO [train.py:1198] (2/4) Epoch 41, batch 1000, loss[loss=0.1722, ctc_loss=0.1098, cr_loss=0.3117, over 16961.00 frames. ], tot_loss[loss=0.1889, ctc_loss=0.1214, cr_loss=0.3374, over 3343809.10 frames. ], batch size: 42, lr: 2.90e-03, grad_scale: 16.0 2024-09-25 11:17:50,707 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=731971.3333333334, ans=0.125 2024-09-25 11:18:02,131 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=13.05 vs. limit=22.5 2024-09-25 11:18:16,043 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=732018.0, ans=0.2 2024-09-25 11:18:16,820 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=6.55 vs. limit=15.0 2024-09-25 11:18:29,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=732064.6666666666, ans=0.0 2024-09-25 11:18:52,043 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=732111.3333333334, ans=0.2 2024-09-25 11:18:56,467 INFO [train.py:1198] (2/4) Epoch 41, batch 1050, loss[loss=0.2199, ctc_loss=0.151, cr_loss=0.3442, over 11755.00 frames. ], tot_loss[loss=0.1897, ctc_loss=0.122, cr_loss=0.3382, over 3333230.30 frames. ], batch size: 123, lr: 2.90e-03, grad_scale: 16.0 2024-09-25 11:19:07,233 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=732158.0, ans=0.2 2024-09-25 11:19:08,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=732158.0, ans=0.0 2024-09-25 11:19:13,632 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=732204.6666666666, ans=0.125 2024-09-25 11:19:37,809 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=732251.3333333334, ans=0.125 2024-09-25 11:20:10,107 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.269e+02 1.332e+02 1.413e+02 1.822e+02, threshold=2.664e+02, percent-clipped=0.0 2024-09-25 11:20:21,421 INFO [train.py:1198] (2/4) Epoch 41, batch 1100, loss[loss=0.2358, ctc_loss=0.1575, cr_loss=0.3918, over 15066.00 frames. ], tot_loss[loss=0.19, ctc_loss=0.1222, cr_loss=0.3388, over 3341994.61 frames. ], batch size: 89, lr: 2.90e-03, grad_scale: 16.0 2024-09-25 11:20:23,479 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=732391.3333333334, ans=0.025 2024-09-25 11:20:39,572 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=732438.0, ans=0.125 2024-09-25 11:20:56,927 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 11:20:57,443 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.75 vs. limit=6.0 2024-09-25 11:21:05,671 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=14.58 vs. limit=15.0 2024-09-25 11:21:10,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=732531.3333333334, ans=0.125 2024-09-25 11:21:10,773 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=5.93 vs. limit=10.0 2024-09-25 11:21:41,759 INFO [train.py:1198] (2/4) Epoch 41, batch 1150, loss[loss=0.1616, ctc_loss=0.0999, cr_loss=0.3083, over 17288.00 frames. ], tot_loss[loss=0.1903, ctc_loss=0.1224, cr_loss=0.3393, over 3337445.08 frames. ], batch size: 46, lr: 2.90e-03, grad_scale: 16.0 2024-09-25 11:21:43,529 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=732624.6666666666, ans=0.1 2024-09-25 11:21:46,788 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=732624.6666666666, ans=0.125 2024-09-25 11:21:56,625 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=732671.3333333334, ans=0.125 2024-09-25 11:22:17,316 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=732718.0, ans=0.1 2024-09-25 11:22:18,051 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=3.88 vs. limit=15.0 2024-09-25 11:22:30,081 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=732764.6666666666, ans=0.0 2024-09-25 11:22:32,884 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=732764.6666666666, ans=0.1 2024-09-25 11:22:33,059 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 11:22:50,423 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.312e+02 1.375e+02 1.453e+02 2.138e+02, threshold=2.750e+02, percent-clipped=0.0 2024-09-25 11:22:55,527 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=732811.3333333334, ans=0.0 2024-09-25 11:22:57,654 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.60 vs. limit=15.0 2024-09-25 11:23:01,755 INFO [train.py:1198] (2/4) Epoch 41, batch 1200, loss[loss=0.1807, ctc_loss=0.116, cr_loss=0.3235, over 17304.00 frames. ], tot_loss[loss=0.1908, ctc_loss=0.1229, cr_loss=0.3395, over 3335449.06 frames. ], batch size: 51, lr: 2.90e-03, grad_scale: 32.0 2024-09-25 11:23:02,145 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=732858.0, ans=0.125 2024-09-25 11:23:16,119 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=732904.6666666666, ans=0.0 2024-09-25 11:23:16,454 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.08 vs. limit=10.0 2024-09-25 11:23:24,327 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=732904.6666666666, ans=0.0 2024-09-25 11:23:44,913 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=4.66 vs. limit=15.0 2024-09-25 11:23:50,647 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=732951.3333333334, ans=0.1 2024-09-25 11:23:58,568 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=732998.0, ans=0.125 2024-09-25 11:24:23,499 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=733044.6666666666, ans=0.0 2024-09-25 11:24:29,677 INFO [train.py:1198] (2/4) Epoch 41, batch 1250, loss[loss=0.2083, ctc_loss=0.1339, cr_loss=0.3722, over 16755.00 frames. ], tot_loss[loss=0.1908, ctc_loss=0.1229, cr_loss=0.3395, over 3351100.15 frames. ], batch size: 61, lr: 2.90e-03, grad_scale: 32.0 2024-09-25 11:24:54,278 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.89 vs. limit=6.0 2024-09-25 11:25:37,854 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=733278.0, ans=0.025 2024-09-25 11:25:39,276 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=733278.0, ans=0.1 2024-09-25 11:25:40,678 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.293e+02 1.355e+02 1.454e+02 2.054e+02, threshold=2.709e+02, percent-clipped=0.0 2024-09-25 11:25:47,543 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=733278.0, ans=0.125 2024-09-25 11:25:52,136 INFO [train.py:1198] (2/4) Epoch 41, batch 1300, loss[loss=0.1978, ctc_loss=0.1266, cr_loss=0.3563, over 17293.00 frames. ], tot_loss[loss=0.1912, ctc_loss=0.1231, cr_loss=0.3402, over 3352840.70 frames. ], batch size: 46, lr: 2.90e-03, grad_scale: 32.0 2024-09-25 11:25:53,021 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.41 vs. limit=10.0 2024-09-25 11:26:00,532 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=733324.6666666666, ans=0.125 2024-09-25 11:26:14,757 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=733371.3333333334, ans=0.125 2024-09-25 11:27:09,945 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2.whitening_limit, batch_count=733511.3333333334, ans=15.0 2024-09-25 11:27:12,354 INFO [train.py:1198] (2/4) Epoch 41, batch 1350, loss[loss=0.162, ctc_loss=0.1005, cr_loss=0.3075, over 17039.00 frames. ], tot_loss[loss=0.1908, ctc_loss=0.1229, cr_loss=0.3396, over 3356175.17 frames. ], batch size: 39, lr: 2.90e-03, grad_scale: 32.0 2024-09-25 11:27:25,568 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=733558.0, ans=0.125 2024-09-25 11:27:27,454 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=733604.6666666666, ans=0.125 2024-09-25 11:27:36,923 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=733604.6666666666, ans=0.0 2024-09-25 11:27:52,941 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=10.09 vs. limit=22.5 2024-09-25 11:28:20,787 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.169e+02 1.331e+02 1.402e+02 1.512e+02 1.864e+02, threshold=2.804e+02, percent-clipped=0.0 2024-09-25 11:28:21,635 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=6.00 vs. limit=15.0 2024-09-25 11:28:37,187 INFO [train.py:1198] (2/4) Epoch 41, batch 1400, loss[loss=0.2184, ctc_loss=0.1435, cr_loss=0.3744, over 17255.00 frames. ], tot_loss[loss=0.1921, ctc_loss=0.1238, cr_loss=0.3417, over 3347354.86 frames. ], batch size: 55, lr: 2.90e-03, grad_scale: 32.0 2024-09-25 11:28:39,091 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=733791.3333333334, ans=0.125 2024-09-25 11:28:51,880 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=733838.0, ans=0.125 2024-09-25 11:29:35,588 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=733931.3333333334, ans=0.125 2024-09-25 11:29:54,542 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=733978.0, ans=0.1 2024-09-25 11:30:01,594 INFO [train.py:1198] (2/4) Epoch 41, batch 1450, loss[loss=0.2091, ctc_loss=0.1384, cr_loss=0.3533, over 15216.00 frames. ], tot_loss[loss=0.1912, ctc_loss=0.1231, cr_loss=0.3405, over 3346055.61 frames. ], batch size: 89, lr: 2.90e-03, grad_scale: 32.0 2024-09-25 11:30:01,891 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=734024.6666666666, ans=0.0 2024-09-25 11:30:28,852 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=734071.3333333334, ans=0.125 2024-09-25 11:30:32,157 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.max_abs, batch_count=734118.0, ans=10.0 2024-09-25 11:30:39,869 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=734118.0, ans=0.125 2024-09-25 11:30:39,917 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=734118.0, ans=0.125 2024-09-25 11:30:44,733 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=734118.0, ans=0.125 2024-09-25 11:30:51,212 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=734164.6666666666, ans=0.07 2024-09-25 11:31:10,081 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.115e+02 1.281e+02 1.374e+02 1.455e+02 2.816e+02, threshold=2.747e+02, percent-clipped=1.0 2024-09-25 11:31:13,576 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=734211.3333333334, ans=0.125 2024-09-25 11:31:15,575 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=11.34 vs. limit=15.0 2024-09-25 11:31:21,396 INFO [train.py:1198] (2/4) Epoch 41, batch 1500, loss[loss=0.1954, ctc_loss=0.1251, cr_loss=0.3515, over 17255.00 frames. ], tot_loss[loss=0.1911, ctc_loss=0.1231, cr_loss=0.3403, over 3339466.15 frames. ], batch size: 44, lr: 2.90e-03, grad_scale: 32.0 2024-09-25 11:31:42,569 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=734304.6666666666, ans=0.125 2024-09-25 11:32:00,445 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=13.25 vs. limit=15.0 2024-09-25 11:32:02,143 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=13.13 vs. limit=22.5 2024-09-25 11:32:33,501 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=734444.6666666666, ans=0.125 2024-09-25 11:32:38,175 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 11:32:41,150 INFO [train.py:1198] (2/4) Epoch 41, batch 1550, loss[loss=0.2154, ctc_loss=0.1391, cr_loss=0.3815, over 17206.00 frames. ], tot_loss[loss=0.1914, ctc_loss=0.1232, cr_loss=0.3407, over 3340239.28 frames. ], batch size: 55, lr: 2.90e-03, grad_scale: 16.0 2024-09-25 11:32:49,641 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=734491.3333333334, ans=0.0 2024-09-25 11:32:52,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=734491.3333333334, ans=0.0 2024-09-25 11:32:59,314 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=734538.0, ans=0.125 2024-09-25 11:33:31,838 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=734584.6666666666, ans=0.125 2024-09-25 11:33:33,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=734631.3333333334, ans=0.025 2024-09-25 11:33:59,384 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.099e+02 1.278e+02 1.349e+02 1.424e+02 1.781e+02, threshold=2.698e+02, percent-clipped=0.0 2024-09-25 11:33:59,741 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=734678.0, ans=0.125 2024-09-25 11:34:06,105 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=734678.0, ans=10.0 2024-09-25 11:34:09,063 INFO [train.py:1198] (2/4) Epoch 41, batch 1600, loss[loss=0.2043, ctc_loss=0.1311, cr_loss=0.3663, over 17212.00 frames. ], tot_loss[loss=0.191, ctc_loss=0.123, cr_loss=0.3402, over 3345605.54 frames. ], batch size: 55, lr: 2.90e-03, grad_scale: 32.0 2024-09-25 11:34:09,731 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.06 vs. limit=6.0 2024-09-25 11:34:41,339 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=734818.0, ans=0.125 2024-09-25 11:34:58,450 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=734864.6666666666, ans=0.0 2024-09-25 11:34:58,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=734864.6666666666, ans=0.1 2024-09-25 11:35:25,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.min_positive, batch_count=734911.3333333334, ans=0.025 2024-09-25 11:35:31,973 INFO [train.py:1198] (2/4) Epoch 41, batch 1650, loss[loss=0.2093, ctc_loss=0.1354, cr_loss=0.3695, over 16957.00 frames. ], tot_loss[loss=0.1921, ctc_loss=0.1238, cr_loss=0.3417, over 3342720.16 frames. ], batch size: 56, lr: 2.90e-03, grad_scale: 32.0 2024-09-25 11:35:33,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=734958.0, ans=0.2 2024-09-25 11:36:10,632 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=735051.3333333334, ans=0.125 2024-09-25 11:36:42,043 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.158e+02 1.308e+02 1.358e+02 1.444e+02 2.176e+02, threshold=2.717e+02, percent-clipped=0.0 2024-09-25 11:36:51,845 INFO [train.py:1198] (2/4) Epoch 41, batch 1700, loss[loss=0.1884, ctc_loss=0.1213, cr_loss=0.3357, over 17142.00 frames. ], tot_loss[loss=0.1915, ctc_loss=0.1232, cr_loss=0.3411, over 3355952.61 frames. ], batch size: 48, lr: 2.90e-03, grad_scale: 32.0 2024-09-25 11:36:52,852 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=4.76 vs. limit=15.0 2024-09-25 11:37:06,506 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=735238.0, ans=0.125 2024-09-25 11:37:39,502 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.86 vs. limit=15.0 2024-09-25 11:38:12,453 INFO [train.py:1198] (2/4) Epoch 41, batch 1750, loss[loss=0.2028, ctc_loss=0.1331, cr_loss=0.3484, over 16720.00 frames. ], tot_loss[loss=0.1912, ctc_loss=0.1231, cr_loss=0.3406, over 3358312.07 frames. ], batch size: 61, lr: 2.90e-03, grad_scale: 32.0 2024-09-25 11:38:35,528 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff3_skip_rate, batch_count=735471.3333333334, ans=0.0 2024-09-25 11:39:19,636 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=735564.6666666666, ans=0.0 2024-09-25 11:39:21,581 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=10.46 vs. limit=22.5 2024-09-25 11:39:30,219 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.135e+02 1.294e+02 1.365e+02 1.443e+02 2.325e+02, threshold=2.730e+02, percent-clipped=0.0 2024-09-25 11:39:39,718 INFO [train.py:1198] (2/4) Epoch 41, batch 1800, loss[loss=0.2308, ctc_loss=0.1491, cr_loss=0.4082, over 16522.00 frames. ], tot_loss[loss=0.1907, ctc_loss=0.1227, cr_loss=0.34, over 3358803.02 frames. ], batch size: 66, lr: 2.90e-03, grad_scale: 32.0 2024-09-25 11:39:46,579 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=735658.0, ans=0.0 2024-09-25 11:39:55,415 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=735658.0, ans=0.04949747468305833 2024-09-25 11:40:13,719 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=11.36 vs. limit=22.5 2024-09-25 11:41:02,239 INFO [train.py:1198] (2/4) Epoch 41, batch 1850, loss[loss=0.1838, ctc_loss=0.1152, cr_loss=0.343, over 17267.00 frames. ], tot_loss[loss=0.1896, ctc_loss=0.1219, cr_loss=0.3387, over 3362423.75 frames. ], batch size: 42, lr: 2.90e-03, grad_scale: 32.0 2024-09-25 11:41:28,032 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=735938.0, ans=0.125 2024-09-25 11:41:37,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.min_positive, batch_count=735984.6666666666, ans=0.05 2024-09-25 11:41:44,584 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.83 vs. limit=15.0 2024-09-25 11:41:45,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=735984.6666666666, ans=0.125 2024-09-25 11:42:09,871 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=8.94 vs. limit=15.0 2024-09-25 11:42:13,779 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.264e+02 1.364e+02 1.473e+02 1.819e+02, threshold=2.729e+02, percent-clipped=0.0 2024-09-25 11:42:21,749 INFO [train.py:1198] (2/4) Epoch 41, batch 1900, loss[loss=0.2051, ctc_loss=0.1322, cr_loss=0.3643, over 17207.00 frames. ], tot_loss[loss=0.1909, ctc_loss=0.1228, cr_loss=0.3405, over 3357720.07 frames. ], batch size: 47, lr: 2.90e-03, grad_scale: 16.0 2024-09-25 11:42:24,036 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=10.21 vs. limit=22.5 2024-09-25 11:43:02,610 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=736218.0, ans=0.035 2024-09-25 11:43:20,587 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=736264.6666666666, ans=0.0 2024-09-25 11:43:23,836 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=736264.6666666666, ans=0.0 2024-09-25 11:43:50,146 INFO [train.py:1198] (2/4) Epoch 41, batch 1950, loss[loss=0.2059, ctc_loss=0.1312, cr_loss=0.3736, over 17281.00 frames. ], tot_loss[loss=0.1907, ctc_loss=0.1227, cr_loss=0.34, over 3350843.09 frames. ], batch size: 44, lr: 2.89e-03, grad_scale: 16.0 2024-09-25 11:44:02,149 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=4.87 vs. limit=15.0 2024-09-25 11:44:13,327 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.46 vs. limit=15.0 2024-09-25 11:44:14,588 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=736404.6666666666, ans=0.125 2024-09-25 11:44:22,879 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=736451.3333333334, ans=0.05 2024-09-25 11:44:30,863 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=736451.3333333334, ans=0.125 2024-09-25 11:44:38,970 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=736498.0, ans=0.125 2024-09-25 11:45:01,369 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=13.56 vs. limit=15.0 2024-09-25 11:45:04,284 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.29 vs. limit=10.0 2024-09-25 11:45:05,168 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.168e+02 1.301e+02 1.403e+02 1.536e+02 5.268e+02, threshold=2.807e+02, percent-clipped=2.0 2024-09-25 11:45:13,264 INFO [train.py:1198] (2/4) Epoch 41, batch 2000, loss[loss=0.1932, ctc_loss=0.1243, cr_loss=0.3449, over 17260.00 frames. ], tot_loss[loss=0.19, ctc_loss=0.1222, cr_loss=0.3387, over 3354702.47 frames. ], batch size: 44, lr: 2.89e-03, grad_scale: 32.0 2024-09-25 11:45:45,820 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=736684.6666666666, ans=0.1 2024-09-25 11:45:47,421 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=736684.6666666666, ans=0.04949747468305833 2024-09-25 11:46:22,881 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=736778.0, ans=0.125 2024-09-25 11:46:26,698 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=6.47 vs. limit=15.0 2024-09-25 11:46:29,510 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.min_positive, batch_count=736778.0, ans=0.05 2024-09-25 11:46:34,055 INFO [train.py:1198] (2/4) Epoch 41, batch 2050, loss[loss=0.2061, ctc_loss=0.1341, cr_loss=0.3604, over 17035.00 frames. ], tot_loss[loss=0.1893, ctc_loss=0.1217, cr_loss=0.3378, over 3357871.10 frames. ], batch size: 52, lr: 2.89e-03, grad_scale: 16.0 2024-09-25 11:46:38,131 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.87 vs. limit=6.0 2024-09-25 11:47:01,808 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=736871.3333333334, ans=0.1 2024-09-25 11:47:12,936 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=736918.0, ans=0.07 2024-09-25 11:47:30,919 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=736964.6666666666, ans=0.2 2024-09-25 11:47:46,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=737011.3333333334, ans=0.025 2024-09-25 11:47:47,792 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.110e+02 1.289e+02 1.380e+02 1.485e+02 2.059e+02, threshold=2.761e+02, percent-clipped=0.0 2024-09-25 11:47:51,445 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.min_abs, batch_count=737011.3333333334, ans=0.5 2024-09-25 11:47:54,263 INFO [train.py:1198] (2/4) Epoch 41, batch 2100, loss[loss=0.2077, ctc_loss=0.1352, cr_loss=0.3625, over 17044.00 frames. ], tot_loss[loss=0.1905, ctc_loss=0.1227, cr_loss=0.3393, over 3354024.25 frames. ], batch size: 52, lr: 2.89e-03, grad_scale: 16.0 2024-09-25 11:48:06,256 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.61 vs. limit=15.0 2024-09-25 11:49:21,659 INFO [train.py:1198] (2/4) Epoch 41, batch 2150, loss[loss=0.1713, ctc_loss=0.1105, cr_loss=0.3038, over 17265.00 frames. ], tot_loss[loss=0.1902, ctc_loss=0.1224, cr_loss=0.3388, over 3356471.54 frames. ], batch size: 42, lr: 2.89e-03, grad_scale: 16.0 2024-09-25 11:49:39,710 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=737338.0, ans=0.125 2024-09-25 11:49:51,676 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=737338.0, ans=0.125 2024-09-25 11:50:10,863 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=737431.3333333334, ans=0.1 2024-09-25 11:50:34,935 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=737478.0, ans=0.125 2024-09-25 11:50:37,767 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.160e+02 1.274e+02 1.356e+02 1.503e+02 3.384e+02, threshold=2.711e+02, percent-clipped=1.0 2024-09-25 11:50:38,190 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=737478.0, ans=0.2 2024-09-25 11:50:44,082 INFO [train.py:1198] (2/4) Epoch 41, batch 2200, loss[loss=0.2238, ctc_loss=0.1434, cr_loss=0.4021, over 16998.00 frames. ], tot_loss[loss=0.1917, ctc_loss=0.1234, cr_loss=0.3413, over 3342633.56 frames. ], batch size: 53, lr: 2.89e-03, grad_scale: 16.0 2024-09-25 11:51:22,938 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=737618.0, ans=0.125 2024-09-25 11:51:40,677 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=737664.6666666666, ans=0.125 2024-09-25 11:51:41,523 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=13.03 vs. limit=22.5 2024-09-25 11:51:58,939 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.45 vs. limit=6.0 2024-09-25 11:52:04,267 INFO [train.py:1198] (2/4) Epoch 41, batch 2250, loss[loss=0.1828, ctc_loss=0.1143, cr_loss=0.3423, over 17081.00 frames. ], tot_loss[loss=0.1905, ctc_loss=0.1226, cr_loss=0.3399, over 3350690.62 frames. ], batch size: 46, lr: 2.89e-03, grad_scale: 16.0 2024-09-25 11:52:30,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=737804.6666666666, ans=0.0 2024-09-25 11:52:42,661 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=737851.3333333334, ans=0.0 2024-09-25 11:52:42,692 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=737851.3333333334, ans=0.1 2024-09-25 11:52:43,320 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.96 vs. limit=15.0 2024-09-25 11:52:49,311 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.82 vs. limit=15.0 2024-09-25 11:52:59,193 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.03 vs. limit=22.5 2024-09-25 11:53:18,792 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.16 vs. limit=22.5 2024-09-25 11:53:22,626 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.150e+02 1.265e+02 1.339e+02 1.413e+02 1.733e+02, threshold=2.679e+02, percent-clipped=0.0 2024-09-25 11:53:24,512 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=737944.6666666666, ans=0.0 2024-09-25 11:53:29,038 INFO [train.py:1198] (2/4) Epoch 41, batch 2300, loss[loss=0.1867, ctc_loss=0.12, cr_loss=0.3337, over 17288.00 frames. ], tot_loss[loss=0.1906, ctc_loss=0.1226, cr_loss=0.34, over 3354286.76 frames. ], batch size: 46, lr: 2.89e-03, grad_scale: 16.0 2024-09-25 11:53:29,401 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.max_abs, batch_count=737991.3333333334, ans=10.0 2024-09-25 11:54:53,982 INFO [train.py:1198] (2/4) Epoch 41, batch 2350, loss[loss=0.1464, ctc_loss=0.09231, cr_loss=0.2702, over 16229.00 frames. ], tot_loss[loss=0.1901, ctc_loss=0.1223, cr_loss=0.3389, over 3363865.19 frames. ], batch size: 36, lr: 2.89e-03, grad_scale: 16.0 2024-09-25 11:54:56,041 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=738224.6666666666, ans=0.0 2024-09-25 11:55:06,985 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=738224.6666666666, ans=0.125 2024-09-25 11:55:16,848 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=738271.3333333334, ans=0.125 2024-09-25 11:55:18,367 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=738271.3333333334, ans=0.0 2024-09-25 11:55:18,468 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=738271.3333333334, ans=0.04949747468305833 2024-09-25 11:55:40,805 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.min_abs, batch_count=738364.6666666666, ans=0.5 2024-09-25 11:55:43,828 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=738364.6666666666, ans=0.0 2024-09-25 11:55:45,491 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=738364.6666666666, ans=0.025 2024-09-25 11:55:53,385 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=738364.6666666666, ans=0.0 2024-09-25 11:56:07,455 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.141e+02 1.297e+02 1.370e+02 1.455e+02 1.687e+02, threshold=2.740e+02, percent-clipped=0.0 2024-09-25 11:56:13,956 INFO [train.py:1198] (2/4) Epoch 41, batch 2400, loss[loss=0.2188, ctc_loss=0.1439, cr_loss=0.3746, over 14785.00 frames. ], tot_loss[loss=0.1903, ctc_loss=0.1224, cr_loss=0.3394, over 3361402.35 frames. ], batch size: 90, lr: 2.89e-03, grad_scale: 32.0 2024-09-25 11:56:45,027 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.41 vs. limit=15.0 2024-09-25 11:56:46,817 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=18.56 vs. limit=22.5 2024-09-25 11:57:11,616 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=738598.0, ans=0.125 2024-09-25 11:57:23,315 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten.whitening_limit, batch_count=738644.6666666666, ans=15.0 2024-09-25 11:57:33,332 INFO [train.py:1198] (2/4) Epoch 41, batch 2450, loss[loss=0.2058, ctc_loss=0.1316, cr_loss=0.3709, over 17276.00 frames. ], tot_loss[loss=0.1906, ctc_loss=0.1226, cr_loss=0.3399, over 3358770.74 frames. ], batch size: 51, lr: 2.89e-03, grad_scale: 16.0 2024-09-25 11:57:35,259 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=738691.3333333334, ans=0.1 2024-09-25 11:58:43,312 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=738878.0, ans=0.125 2024-09-25 11:58:46,450 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=738878.0, ans=0.5 2024-09-25 11:58:55,657 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.283e+02 1.406e+02 1.500e+02 1.911e+02, threshold=2.813e+02, percent-clipped=0.0 2024-09-25 11:59:00,465 INFO [train.py:1198] (2/4) Epoch 41, batch 2500, loss[loss=0.1421, ctc_loss=0.08965, cr_loss=0.2624, over 16294.00 frames. ], tot_loss[loss=0.1914, ctc_loss=0.1232, cr_loss=0.3408, over 3353177.53 frames. ], batch size: 36, lr: 2.89e-03, grad_scale: 16.0 2024-09-25 11:59:13,805 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=738924.6666666666, ans=0.0 2024-09-25 12:00:03,983 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=739064.6666666666, ans=0.125 2024-09-25 12:00:23,049 INFO [train.py:1198] (2/4) Epoch 41, batch 2550, loss[loss=0.2347, ctc_loss=0.1514, cr_loss=0.4166, over 17040.00 frames. ], tot_loss[loss=0.1902, ctc_loss=0.1223, cr_loss=0.3391, over 3356319.03 frames. ], batch size: 52, lr: 2.89e-03, grad_scale: 16.0 2024-09-25 12:00:28,560 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=4.63 vs. limit=12.0 2024-09-25 12:00:36,687 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=4.71 vs. limit=15.0 2024-09-25 12:00:47,478 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=739204.6666666666, ans=0.04949747468305833 2024-09-25 12:01:14,574 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=739298.0, ans=0.2 2024-09-25 12:01:38,422 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.126e+02 1.313e+02 1.392e+02 1.468e+02 1.882e+02, threshold=2.785e+02, percent-clipped=0.0 2024-09-25 12:01:43,156 INFO [train.py:1198] (2/4) Epoch 41, batch 2600, loss[loss=0.2324, ctc_loss=0.1527, cr_loss=0.3988, over 16991.00 frames. ], tot_loss[loss=0.1898, ctc_loss=0.122, cr_loss=0.3388, over 3363390.04 frames. ], batch size: 53, lr: 2.89e-03, grad_scale: 16.0 2024-09-25 12:01:49,724 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=739391.3333333334, ans=0.1 2024-09-25 12:02:00,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=739438.0, ans=0.125 2024-09-25 12:02:18,311 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.min_positive, batch_count=739484.6666666666, ans=0.025 2024-09-25 12:02:42,197 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=739531.3333333334, ans=0.125 2024-09-25 12:02:48,986 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=11.28 vs. limit=22.5 2024-09-25 12:03:07,751 INFO [train.py:1198] (2/4) Epoch 41, batch 2650, loss[loss=0.2175, ctc_loss=0.1389, cr_loss=0.3929, over 16987.00 frames. ], tot_loss[loss=0.1909, ctc_loss=0.1228, cr_loss=0.3405, over 3366880.47 frames. ], batch size: 53, lr: 2.89e-03, grad_scale: 16.0 2024-09-25 12:03:15,035 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.66 vs. limit=10.0 2024-09-25 12:03:24,200 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=739671.3333333334, ans=0.2 2024-09-25 12:04:03,961 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=739764.6666666666, ans=0.125 2024-09-25 12:04:26,077 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.313e+02 1.405e+02 1.499e+02 1.845e+02, threshold=2.809e+02, percent-clipped=0.0 2024-09-25 12:04:26,373 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=739811.3333333334, ans=0.125 2024-09-25 12:04:30,838 INFO [train.py:1198] (2/4) Epoch 41, batch 2700, loss[loss=0.2, ctc_loss=0.1256, cr_loss=0.3717, over 17216.00 frames. ], tot_loss[loss=0.1908, ctc_loss=0.1226, cr_loss=0.3408, over 3375506.57 frames. ], batch size: 50, lr: 2.89e-03, grad_scale: 16.0 2024-09-25 12:04:46,619 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=739858.0, ans=0.025 2024-09-25 12:04:51,590 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=739904.6666666666, ans=0.025 2024-09-25 12:05:00,127 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=5.02 vs. limit=10.0 2024-09-25 12:05:03,201 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.77 vs. limit=12.0 2024-09-25 12:05:22,910 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=739998.0, ans=0.125 2024-09-25 12:05:27,655 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=739998.0, ans=0.125 2024-09-25 12:05:34,050 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=739998.0, ans=0.125 2024-09-25 12:05:53,474 INFO [train.py:1198] (2/4) Epoch 41, batch 2750, loss[loss=0.246, ctc_loss=0.1654, cr_loss=0.4033, over 14878.00 frames. ], tot_loss[loss=0.1904, ctc_loss=0.1224, cr_loss=0.3402, over 3378191.30 frames. ], batch size: 89, lr: 2.89e-03, grad_scale: 16.0 2024-09-25 12:07:03,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=740278.0, ans=0.125 2024-09-25 12:07:09,101 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.274e+02 1.386e+02 1.486e+02 2.179e+02, threshold=2.772e+02, percent-clipped=0.0 2024-09-25 12:07:14,038 INFO [train.py:1198] (2/4) Epoch 41, batch 2800, loss[loss=0.1865, ctc_loss=0.1172, cr_loss=0.3463, over 17225.00 frames. ], tot_loss[loss=0.1896, ctc_loss=0.1219, cr_loss=0.3385, over 3363627.25 frames. ], batch size: 55, lr: 2.89e-03, grad_scale: 32.0 2024-09-25 12:07:20,835 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=740324.6666666666, ans=0.125 2024-09-25 12:07:24,787 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.23 vs. limit=22.5 2024-09-25 12:07:41,464 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.70 vs. limit=6.0 2024-09-25 12:07:45,525 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=740418.0, ans=0.1 2024-09-25 12:08:20,892 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=740464.6666666666, ans=0.2 2024-09-25 12:08:38,005 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=740511.3333333334, ans=0.125 2024-09-25 12:08:39,625 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=740511.3333333334, ans=0.125 2024-09-25 12:08:42,543 INFO [train.py:1198] (2/4) Epoch 41, batch 2850, loss[loss=0.1507, ctc_loss=0.09346, cr_loss=0.286, over 17103.00 frames. ], tot_loss[loss=0.1907, ctc_loss=0.1227, cr_loss=0.34, over 3348985.82 frames. ], batch size: 40, lr: 2.89e-03, grad_scale: 32.0 2024-09-25 12:09:14,874 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=740651.3333333334, ans=0.125 2024-09-25 12:09:24,515 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=740651.3333333334, ans=0.125 2024-09-25 12:09:35,326 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 12:10:00,303 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.136e+02 1.293e+02 1.358e+02 1.450e+02 1.925e+02, threshold=2.717e+02, percent-clipped=0.0 2024-09-25 12:10:05,283 INFO [train.py:1198] (2/4) Epoch 41, batch 2900, loss[loss=0.1747, ctc_loss=0.1105, cr_loss=0.3207, over 17151.00 frames. ], tot_loss[loss=0.1907, ctc_loss=0.1227, cr_loss=0.3401, over 3360794.87 frames. ], batch size: 45, lr: 2.89e-03, grad_scale: 32.0 2024-09-25 12:10:26,657 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=740838.0, ans=0.125 2024-09-25 12:10:29,912 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=740838.0, ans=0.125 2024-09-25 12:10:42,612 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=740884.6666666666, ans=0.125 2024-09-25 12:10:52,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=740931.3333333334, ans=0.09899494936611666 2024-09-25 12:11:13,639 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=20.11 vs. limit=22.5 2024-09-25 12:11:25,601 INFO [train.py:1198] (2/4) Epoch 41, batch 2950, loss[loss=0.1684, ctc_loss=0.1044, cr_loss=0.3198, over 17117.00 frames. ], tot_loss[loss=0.1905, ctc_loss=0.1224, cr_loss=0.3404, over 3363527.51 frames. ], batch size: 40, lr: 2.89e-03, grad_scale: 32.0 2024-09-25 12:11:51,447 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=741071.3333333334, ans=0.125 2024-09-25 12:12:02,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=741118.0, ans=0.1 2024-09-25 12:12:20,192 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=741164.6666666666, ans=0.125 2024-09-25 12:12:40,613 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.116e+02 1.313e+02 1.388e+02 1.485e+02 2.724e+02, threshold=2.776e+02, percent-clipped=1.0 2024-09-25 12:12:43,837 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=741258.0, ans=0.2 2024-09-25 12:12:45,291 INFO [train.py:1198] (2/4) Epoch 41, batch 3000, loss[loss=0.1577, ctc_loss=0.1006, cr_loss=0.2855, over 16951.00 frames. ], tot_loss[loss=0.1893, ctc_loss=0.1215, cr_loss=0.3387, over 3371651.49 frames. ], batch size: 42, lr: 2.89e-03, grad_scale: 32.0 2024-09-25 12:12:45,292 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 12:13:00,793 INFO [train.py:1230] (2/4) Epoch 41, validation: loss=0.03575, ctc_loss=0.03575, cr_loss=9.81e-15, over 944034.00 frames. 2024-09-25 12:13:00,794 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 12:13:02,591 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=741258.0, ans=0.125 2024-09-25 12:13:28,979 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.50 vs. limit=15.0 2024-09-25 12:13:42,452 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=741351.3333333334, ans=0.125 2024-09-25 12:13:48,791 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=741398.0, ans=0.2 2024-09-25 12:14:12,219 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.78 vs. limit=6.0 2024-09-25 12:14:26,511 INFO [train.py:1198] (2/4) Epoch 41, batch 3050, loss[loss=0.2108, ctc_loss=0.1423, cr_loss=0.3427, over 15078.00 frames. ], tot_loss[loss=0.1901, ctc_loss=0.1222, cr_loss=0.3396, over 3352207.70 frames. ], batch size: 89, lr: 2.88e-03, grad_scale: 32.0 2024-09-25 12:14:30,102 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.min_positive, batch_count=741491.3333333334, ans=0.05 2024-09-25 12:14:33,125 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=741491.3333333334, ans=0.125 2024-09-25 12:14:40,192 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=9.99 vs. limit=15.0 2024-09-25 12:14:42,335 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 12:14:48,589 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=741538.0, ans=0.125 2024-09-25 12:14:54,677 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=741538.0, ans=0.2 2024-09-25 12:15:03,904 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=741584.6666666666, ans=0.025 2024-09-25 12:15:07,149 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=741584.6666666666, ans=0.0 2024-09-25 12:15:10,125 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=741584.6666666666, ans=0.0 2024-09-25 12:15:39,715 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.278e+02 1.352e+02 1.472e+02 2.246e+02, threshold=2.705e+02, percent-clipped=0.0 2024-09-25 12:15:44,427 INFO [train.py:1198] (2/4) Epoch 41, batch 3100, loss[loss=0.2024, ctc_loss=0.1317, cr_loss=0.3534, over 16488.00 frames. ], tot_loss[loss=0.1905, ctc_loss=0.1225, cr_loss=0.34, over 3336697.97 frames. ], batch size: 66, lr: 2.88e-03, grad_scale: 32.0 2024-09-25 12:16:09,753 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=741771.3333333334, ans=0.125 2024-09-25 12:16:09,759 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=741771.3333333334, ans=0.125 2024-09-25 12:16:28,304 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=741818.0, ans=0.2 2024-09-25 12:16:50,971 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=741911.3333333334, ans=0.125 2024-09-25 12:17:04,754 INFO [train.py:1198] (2/4) Epoch 41, batch 3150, loss[loss=0.1604, ctc_loss=0.1013, cr_loss=0.2957, over 16663.00 frames. ], tot_loss[loss=0.1911, ctc_loss=0.123, cr_loss=0.3405, over 3335441.11 frames. ], batch size: 37, lr: 2.88e-03, grad_scale: 32.0 2024-09-25 12:17:37,157 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.29 vs. limit=15.0 2024-09-25 12:17:58,887 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=742098.0, ans=0.2 2024-09-25 12:18:07,654 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=7.03 vs. limit=15.0 2024-09-25 12:18:19,185 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.096e+02 1.276e+02 1.346e+02 1.473e+02 3.100e+02, threshold=2.691e+02, percent-clipped=1.0 2024-09-25 12:18:23,855 INFO [train.py:1198] (2/4) Epoch 41, batch 3200, loss[loss=0.2101, ctc_loss=0.1409, cr_loss=0.3459, over 11718.00 frames. ], tot_loss[loss=0.1911, ctc_loss=0.123, cr_loss=0.3401, over 3344658.26 frames. ], batch size: 123, lr: 2.88e-03, grad_scale: 32.0 2024-09-25 12:18:24,236 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=742191.3333333334, ans=0.07 2024-09-25 12:18:25,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=742191.3333333334, ans=0.0 2024-09-25 12:18:30,290 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=742191.3333333334, ans=0.2 2024-09-25 12:18:31,988 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=742191.3333333334, ans=0.05 2024-09-25 12:18:54,085 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 12:19:01,774 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=742284.6666666666, ans=0.125 2024-09-25 12:19:14,305 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=742331.3333333334, ans=0.125 2024-09-25 12:19:25,327 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=742378.0, ans=0.0 2024-09-25 12:19:37,866 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=742378.0, ans=0.125 2024-09-25 12:19:42,197 INFO [train.py:1198] (2/4) Epoch 41, batch 3250, loss[loss=0.1761, ctc_loss=0.1134, cr_loss=0.3134, over 17066.00 frames. ], tot_loss[loss=0.1919, ctc_loss=0.1237, cr_loss=0.3414, over 3343274.43 frames. ], batch size: 46, lr: 2.88e-03, grad_scale: 8.0 2024-09-25 12:19:44,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=742424.6666666666, ans=0.1 2024-09-25 12:19:47,201 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=742424.6666666666, ans=0.1 2024-09-25 12:19:50,932 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=7.01 vs. limit=15.0 2024-09-25 12:19:57,910 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=742471.3333333334, ans=0.2 2024-09-25 12:20:20,624 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=3.75 vs. limit=10.0 2024-09-25 12:20:27,971 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=742564.6666666666, ans=0.125 2024-09-25 12:20:28,579 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.50 vs. limit=22.5 2024-09-25 12:20:38,864 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=742564.6666666666, ans=0.0 2024-09-25 12:20:45,187 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=742611.3333333334, ans=0.025 2024-09-25 12:20:55,969 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=742611.3333333334, ans=0.0 2024-09-25 12:20:58,869 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.131e+02 1.297e+02 1.390e+02 1.461e+02 1.762e+02, threshold=2.780e+02, percent-clipped=0.0 2024-09-25 12:21:00,472 INFO [train.py:1198] (2/4) Epoch 41, batch 3300, loss[loss=0.1679, ctc_loss=0.1073, cr_loss=0.3027, over 17089.00 frames. ], tot_loss[loss=0.1917, ctc_loss=0.1235, cr_loss=0.3412, over 3343630.80 frames. ], batch size: 40, lr: 2.88e-03, grad_scale: 8.0 2024-09-25 12:21:05,964 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=16.71 vs. limit=22.5 2024-09-25 12:21:16,546 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=742704.6666666666, ans=0.125 2024-09-25 12:21:19,818 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=742704.6666666666, ans=0.0 2024-09-25 12:21:29,161 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=742704.6666666666, ans=0.0 2024-09-25 12:21:41,519 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=742751.3333333334, ans=0.0 2024-09-25 12:22:04,806 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=742844.6666666666, ans=0.125 2024-09-25 12:22:18,676 INFO [train.py:1198] (2/4) Epoch 41, batch 3350, loss[loss=0.1795, ctc_loss=0.1148, cr_loss=0.3238, over 17302.00 frames. ], tot_loss[loss=0.1924, ctc_loss=0.124, cr_loss=0.3421, over 3354002.85 frames. ], batch size: 49, lr: 2.88e-03, grad_scale: 8.0 2024-09-25 12:22:19,074 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 12:22:19,081 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=742891.3333333334, ans=0.125 2024-09-25 12:23:13,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=743031.3333333334, ans=0.1 2024-09-25 12:23:31,248 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=20.05 vs. limit=22.5 2024-09-25 12:23:35,162 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.142e+02 1.288e+02 1.390e+02 1.522e+02 3.340e+02, threshold=2.781e+02, percent-clipped=2.0 2024-09-25 12:23:36,767 INFO [train.py:1198] (2/4) Epoch 41, batch 3400, loss[loss=0.1481, ctc_loss=0.0936, cr_loss=0.2723, over 17297.00 frames. ], tot_loss[loss=0.1917, ctc_loss=0.1234, cr_loss=0.3412, over 3354457.30 frames. ], batch size: 42, lr: 2.88e-03, grad_scale: 8.0 2024-09-25 12:23:57,520 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=743171.3333333334, ans=0.0 2024-09-25 12:24:31,403 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=743264.6666666666, ans=0.1 2024-09-25 12:24:40,939 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=743264.6666666666, ans=0.2 2024-09-25 12:25:00,982 INFO [train.py:1198] (2/4) Epoch 41, batch 3450, loss[loss=0.2041, ctc_loss=0.1309, cr_loss=0.3659, over 16728.00 frames. ], tot_loss[loss=0.1915, ctc_loss=0.1234, cr_loss=0.3406, over 3345411.97 frames. ], batch size: 61, lr: 2.88e-03, grad_scale: 8.0 2024-09-25 12:25:01,280 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=743358.0, ans=0.09899494936611666 2024-09-25 12:25:26,213 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=743404.6666666666, ans=0.0 2024-09-25 12:25:32,677 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=12.66 vs. limit=15.0 2024-09-25 12:25:43,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=743451.3333333334, ans=0.125 2024-09-25 12:25:58,132 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=8.74 vs. limit=10.0 2024-09-25 12:25:58,849 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=743498.0, ans=0.2 2024-09-25 12:26:12,908 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=743544.6666666666, ans=0.2 2024-09-25 12:26:17,203 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.301e+02 1.363e+02 1.481e+02 3.003e+02, threshold=2.726e+02, percent-clipped=1.0 2024-09-25 12:26:18,732 INFO [train.py:1198] (2/4) Epoch 41, batch 3500, loss[loss=0.1769, ctc_loss=0.1091, cr_loss=0.3392, over 16952.00 frames. ], tot_loss[loss=0.1913, ctc_loss=0.1232, cr_loss=0.3405, over 3347799.38 frames. ], batch size: 42, lr: 2.88e-03, grad_scale: 8.0 2024-09-25 12:26:23,860 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=743591.3333333334, ans=0.0 2024-09-25 12:26:54,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=743684.6666666666, ans=0.125 2024-09-25 12:26:55,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=743684.6666666666, ans=0.05 2024-09-25 12:27:09,788 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=743731.3333333334, ans=10.0 2024-09-25 12:27:19,091 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=743731.3333333334, ans=0.125 2024-09-25 12:27:22,570 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=16.92 vs. limit=22.5 2024-09-25 12:27:28,342 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=743778.0, ans=0.125 2024-09-25 12:27:33,083 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=743778.0, ans=0.025 2024-09-25 12:27:37,675 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=743824.6666666666, ans=0.1 2024-09-25 12:27:39,044 INFO [train.py:1198] (2/4) Epoch 41, batch 3550, loss[loss=0.1847, ctc_loss=0.1232, cr_loss=0.3075, over 15885.00 frames. ], tot_loss[loss=0.1909, ctc_loss=0.123, cr_loss=0.3398, over 3345349.71 frames. ], batch size: 74, lr: 2.88e-03, grad_scale: 8.0 2024-09-25 12:27:53,522 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=743871.3333333334, ans=0.125 2024-09-25 12:28:00,138 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=512, metric=11.39 vs. limit=22.5 2024-09-25 12:28:55,696 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.288e+02 1.370e+02 1.460e+02 2.719e+02, threshold=2.741e+02, percent-clipped=0.0 2024-09-25 12:28:57,317 INFO [train.py:1198] (2/4) Epoch 41, batch 3600, loss[loss=0.2015, ctc_loss=0.1287, cr_loss=0.3642, over 17289.00 frames. ], tot_loss[loss=0.1899, ctc_loss=0.1223, cr_loss=0.3382, over 3340663.56 frames. ], batch size: 49, lr: 2.88e-03, grad_scale: 16.0 2024-09-25 12:29:33,648 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=744151.3333333334, ans=0.2 2024-09-25 12:29:52,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.min_positive, batch_count=744198.0, ans=0.025 2024-09-25 12:30:15,424 INFO [train.py:1198] (2/4) Epoch 41, batch 3650, loss[loss=0.1976, ctc_loss=0.1278, cr_loss=0.3492, over 17370.00 frames. ], tot_loss[loss=0.1894, ctc_loss=0.1219, cr_loss=0.3376, over 3352199.37 frames. ], batch size: 48, lr: 2.88e-03, grad_scale: 16.0 2024-09-25 12:30:23,368 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=744291.3333333334, ans=0.2 2024-09-25 12:30:30,281 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=13.25 vs. limit=22.5 2024-09-25 12:30:45,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=744384.6666666666, ans=0.125 2024-09-25 12:31:07,636 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=744431.3333333334, ans=0.125 2024-09-25 12:31:17,274 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=744478.0, ans=0.125 2024-09-25 12:31:21,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=744478.0, ans=0.125 2024-09-25 12:31:23,232 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=744478.0, ans=0.0 2024-09-25 12:31:27,992 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=744478.0, ans=0.025 2024-09-25 12:31:31,212 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=744478.0, ans=0.125 2024-09-25 12:31:32,488 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.167e+02 1.315e+02 1.407e+02 1.494e+02 1.743e+02, threshold=2.814e+02, percent-clipped=0.0 2024-09-25 12:31:34,120 INFO [train.py:1198] (2/4) Epoch 41, batch 3700, loss[loss=0.2238, ctc_loss=0.1461, cr_loss=0.3882, over 16602.00 frames. ], tot_loss[loss=0.1907, ctc_loss=0.1228, cr_loss=0.3396, over 3350777.82 frames. ], batch size: 66, lr: 2.88e-03, grad_scale: 16.0 2024-09-25 12:32:13,881 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=744618.0, ans=0.2 2024-09-25 12:32:14,213 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.36 vs. limit=10.0 2024-09-25 12:32:18,958 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=512, metric=6.05 vs. limit=15.0 2024-09-25 12:32:37,378 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=744711.3333333334, ans=0.1 2024-09-25 12:32:37,395 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=744711.3333333334, ans=0.0 2024-09-25 12:32:53,077 INFO [train.py:1198] (2/4) Epoch 41, batch 3750, loss[loss=0.1691, ctc_loss=0.1062, cr_loss=0.3148, over 17026.00 frames. ], tot_loss[loss=0.1911, ctc_loss=0.123, cr_loss=0.3402, over 3347372.92 frames. ], batch size: 39, lr: 2.88e-03, grad_scale: 16.0 2024-09-25 12:32:57,940 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=744758.0, ans=0.125 2024-09-25 12:33:09,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=744804.6666666666, ans=0.2 2024-09-25 12:33:41,688 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 12:33:43,099 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=744898.0, ans=0.125 2024-09-25 12:33:46,995 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.31 vs. limit=22.5 2024-09-25 12:33:50,288 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=4.57 vs. limit=15.0 2024-09-25 12:34:05,756 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=744944.6666666666, ans=0.025 2024-09-25 12:34:11,724 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.165e+02 1.288e+02 1.369e+02 1.453e+02 1.957e+02, threshold=2.738e+02, percent-clipped=0.0 2024-09-25 12:34:14,053 INFO [train.py:1198] (2/4) Epoch 41, batch 3800, loss[loss=0.1906, ctc_loss=0.122, cr_loss=0.343, over 17272.00 frames. ], tot_loss[loss=0.191, ctc_loss=0.123, cr_loss=0.3398, over 3343503.41 frames. ], batch size: 44, lr: 2.88e-03, grad_scale: 16.0 2024-09-25 12:34:26,670 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=744991.3333333334, ans=0.125 2024-09-25 12:34:37,609 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=745038.0, ans=0.05 2024-09-25 12:34:42,219 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=745038.0, ans=0.125 2024-09-25 12:34:50,085 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=745084.6666666666, ans=0.05 2024-09-25 12:35:04,249 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=745131.3333333334, ans=0.125 2024-09-25 12:35:09,106 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=12.58 vs. limit=22.5 2024-09-25 12:35:15,400 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=745178.0, ans=0.125 2024-09-25 12:35:32,596 INFO [train.py:1198] (2/4) Epoch 41, batch 3850, loss[loss=0.2475, ctc_loss=0.1679, cr_loss=0.3978, over 11655.00 frames. ], tot_loss[loss=0.1916, ctc_loss=0.1235, cr_loss=0.3401, over 3317468.80 frames. ], batch size: 123, lr: 2.88e-03, grad_scale: 16.0 2024-09-25 12:35:51,971 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 12:36:36,999 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=745411.3333333334, ans=0.125 2024-09-25 12:37:33,472 INFO [train.py:1198] (2/4) Epoch 42, batch 0, loss[loss=0.2048, ctc_loss=0.1314, cr_loss=0.3667, over 17237.00 frames. ], tot_loss[loss=0.2048, ctc_loss=0.1314, cr_loss=0.3667, over 17237.00 frames. ], batch size: 55, lr: 2.84e-03, grad_scale: 32.0 2024-09-25 12:37:33,473 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 12:37:48,883 INFO [train.py:1230] (2/4) Epoch 42, validation: loss=0.03453, ctc_loss=0.03453, cr_loss=1.019e-14, over 944034.00 frames. 2024-09-25 12:37:48,884 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 12:37:53,634 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.354e+02 1.491e+02 1.700e+02 3.066e+02, threshold=2.981e+02, percent-clipped=1.0 2024-09-25 12:39:09,842 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=745672.6666666666, ans=0.0 2024-09-25 12:39:11,033 INFO [train.py:1198] (2/4) Epoch 42, batch 50, loss[loss=0.2066, ctc_loss=0.137, cr_loss=0.3479, over 17025.00 frames. ], tot_loss[loss=0.192, ctc_loss=0.1236, cr_loss=0.3419, over 751629.90 frames. ], batch size: 56, lr: 2.84e-03, grad_scale: 32.0 2024-09-25 12:39:54,996 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=4.84 vs. limit=15.0 2024-09-25 12:40:15,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=745812.6666666666, ans=0.0 2024-09-25 12:40:37,093 INFO [train.py:1198] (2/4) Epoch 42, batch 100, loss[loss=0.2059, ctc_loss=0.1359, cr_loss=0.3504, over 15004.00 frames. ], tot_loss[loss=0.1884, ctc_loss=0.121, cr_loss=0.3373, over 1332402.47 frames. ], batch size: 89, lr: 2.84e-03, grad_scale: 32.0 2024-09-25 12:40:41,769 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.296e+02 1.378e+02 1.504e+02 1.895e+02, threshold=2.756e+02, percent-clipped=0.0 2024-09-25 12:40:42,310 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 12:41:01,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=745952.6666666666, ans=0.1 2024-09-25 12:41:31,043 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=746046.0, ans=0.125 2024-09-25 12:41:39,054 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=746046.0, ans=0.125 2024-09-25 12:41:41,073 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=3.95 vs. limit=10.0 2024-09-25 12:41:50,508 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=746092.6666666666, ans=0.125 2024-09-25 12:41:59,754 INFO [train.py:1198] (2/4) Epoch 42, batch 150, loss[loss=0.1908, ctc_loss=0.1274, cr_loss=0.3169, over 17021.00 frames. ], tot_loss[loss=0.187, ctc_loss=0.12, cr_loss=0.3351, over 1789745.13 frames. ], batch size: 51, lr: 2.84e-03, grad_scale: 32.0 2024-09-25 12:42:01,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=746139.3333333334, ans=0.07 2024-09-25 12:42:04,849 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=746139.3333333334, ans=0.125 2024-09-25 12:42:12,567 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=746139.3333333334, ans=0.2 2024-09-25 12:42:39,798 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=746232.6666666666, ans=0.125 2024-09-25 12:42:43,051 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=746232.6666666666, ans=0.0 2024-09-25 12:43:02,476 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=746326.0, ans=0.1 2024-09-25 12:43:19,865 INFO [train.py:1198] (2/4) Epoch 42, batch 200, loss[loss=0.2027, ctc_loss=0.1275, cr_loss=0.3761, over 17283.00 frames. ], tot_loss[loss=0.1876, ctc_loss=0.1205, cr_loss=0.3354, over 2139339.75 frames. ], batch size: 46, lr: 2.84e-03, grad_scale: 16.0 2024-09-25 12:43:23,407 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=746372.6666666666, ans=0.0 2024-09-25 12:43:26,428 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.149e+02 1.308e+02 1.400e+02 1.491e+02 1.853e+02, threshold=2.800e+02, percent-clipped=0.0 2024-09-25 12:43:57,185 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=746466.0, ans=0.1 2024-09-25 12:44:22,727 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=4.60 vs. limit=12.0 2024-09-25 12:44:45,843 INFO [train.py:1198] (2/4) Epoch 42, batch 250, loss[loss=0.2276, ctc_loss=0.1546, cr_loss=0.3648, over 11750.00 frames. ], tot_loss[loss=0.1896, ctc_loss=0.122, cr_loss=0.3379, over 2404728.57 frames. ], batch size: 125, lr: 2.84e-03, grad_scale: 16.0 2024-09-25 12:44:58,958 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=746606.0, ans=0.125 2024-09-25 12:45:14,395 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 12:45:55,091 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=746792.6666666666, ans=0.125 2024-09-25 12:46:09,469 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=746839.3333333334, ans=0.5 2024-09-25 12:46:13,665 INFO [train.py:1198] (2/4) Epoch 42, batch 300, loss[loss=0.2065, ctc_loss=0.1331, cr_loss=0.3669, over 17014.00 frames. ], tot_loss[loss=0.1888, ctc_loss=0.1214, cr_loss=0.3369, over 2608641.42 frames. ], batch size: 56, lr: 2.84e-03, grad_scale: 16.0 2024-09-25 12:46:17,102 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=746839.3333333334, ans=0.125 2024-09-25 12:46:20,048 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.150e+02 1.307e+02 1.390e+02 1.481e+02 2.059e+02, threshold=2.780e+02, percent-clipped=0.0 2024-09-25 12:47:01,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=746979.3333333334, ans=0.0 2024-09-25 12:47:33,913 INFO [train.py:1198] (2/4) Epoch 42, batch 350, loss[loss=0.2067, ctc_loss=0.1348, cr_loss=0.3591, over 16056.00 frames. ], tot_loss[loss=0.1886, ctc_loss=0.1211, cr_loss=0.3377, over 2785387.42 frames. ], batch size: 74, lr: 2.84e-03, grad_scale: 16.0 2024-09-25 12:47:34,153 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=747072.6666666666, ans=0.125 2024-09-25 12:47:58,178 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=747119.3333333334, ans=0.035 2024-09-25 12:48:02,896 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=747119.3333333334, ans=0.0 2024-09-25 12:48:15,645 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.min_abs, batch_count=747166.0, ans=0.5 2024-09-25 12:48:20,669 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=747212.6666666666, ans=0.125 2024-09-25 12:48:55,118 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=747306.0, ans=0.0 2024-09-25 12:48:56,550 INFO [train.py:1198] (2/4) Epoch 42, batch 400, loss[loss=0.2167, ctc_loss=0.1405, cr_loss=0.3811, over 17211.00 frames. ], tot_loss[loss=0.1898, ctc_loss=0.122, cr_loss=0.3393, over 2907924.74 frames. ], batch size: 47, lr: 2.84e-03, grad_scale: 32.0 2024-09-25 12:48:56,943 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=747306.0, ans=0.125 2024-09-25 12:49:02,843 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.157e+02 1.303e+02 1.377e+02 1.468e+02 2.064e+02, threshold=2.754e+02, percent-clipped=0.0 2024-09-25 12:49:43,815 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=747399.3333333334, ans=0.125 2024-09-25 12:50:03,179 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=747492.6666666666, ans=0.2 2024-09-25 12:50:04,730 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=747492.6666666666, ans=0.125 2024-09-25 12:50:06,909 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2.whitening_limit, batch_count=747492.6666666666, ans=15.0 2024-09-25 12:50:07,950 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=747492.6666666666, ans=0.0 2024-09-25 12:50:08,088 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=747492.6666666666, ans=0.5 2024-09-25 12:50:09,588 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=747492.6666666666, ans=0.125 2024-09-25 12:50:21,629 INFO [train.py:1198] (2/4) Epoch 42, batch 450, loss[loss=0.1375, ctc_loss=0.08492, cr_loss=0.2628, over 16951.00 frames. ], tot_loss[loss=0.1898, ctc_loss=0.122, cr_loss=0.3393, over 3016806.54 frames. ], batch size: 42, lr: 2.84e-03, grad_scale: 32.0 2024-09-25 12:50:28,281 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=747539.3333333334, ans=0.0 2024-09-25 12:51:44,413 INFO [train.py:1198] (2/4) Epoch 42, batch 500, loss[loss=0.1859, ctc_loss=0.1179, cr_loss=0.3403, over 17326.00 frames. ], tot_loss[loss=0.1906, ctc_loss=0.1226, cr_loss=0.3401, over 3096036.84 frames. ], batch size: 51, lr: 2.84e-03, grad_scale: 32.0 2024-09-25 12:51:50,866 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 1.301e+02 1.386e+02 1.488e+02 2.359e+02, threshold=2.773e+02, percent-clipped=0.0 2024-09-25 12:52:18,262 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=747866.0, ans=0.1 2024-09-25 12:52:25,205 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.80 vs. limit=6.0 2024-09-25 12:52:31,439 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module1.whiten, num_groups=1, num_channels=512, metric=8.36 vs. limit=15.0 2024-09-25 12:52:38,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=747912.6666666666, ans=0.0 2024-09-25 12:52:38,832 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=747912.6666666666, ans=0.0 2024-09-25 12:52:56,493 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=747959.3333333334, ans=0.125 2024-09-25 12:53:04,112 INFO [train.py:1198] (2/4) Epoch 42, batch 550, loss[loss=0.2311, ctc_loss=0.1529, cr_loss=0.3909, over 15043.00 frames. ], tot_loss[loss=0.1905, ctc_loss=0.1226, cr_loss=0.3397, over 3158333.68 frames. ], batch size: 89, lr: 2.84e-03, grad_scale: 32.0 2024-09-25 12:54:24,690 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=748192.6666666666, ans=0.0 2024-09-25 12:54:29,010 INFO [train.py:1198] (2/4) Epoch 42, batch 600, loss[loss=0.1684, ctc_loss=0.1081, cr_loss=0.3013, over 17212.00 frames. ], tot_loss[loss=0.1914, ctc_loss=0.1232, cr_loss=0.3411, over 3200025.22 frames. ], batch size: 47, lr: 2.84e-03, grad_scale: 32.0 2024-09-25 12:54:35,339 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.189e+02 1.284e+02 1.358e+02 1.453e+02 2.155e+02, threshold=2.716e+02, percent-clipped=0.0 2024-09-25 12:55:09,378 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=748332.6666666666, ans=0.0 2024-09-25 12:55:45,817 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 12:55:51,814 INFO [train.py:1198] (2/4) Epoch 42, batch 650, loss[loss=0.2041, ctc_loss=0.1321, cr_loss=0.36, over 15777.00 frames. ], tot_loss[loss=0.1913, ctc_loss=0.1231, cr_loss=0.341, over 3226615.31 frames. ], batch size: 74, lr: 2.84e-03, grad_scale: 32.0 2024-09-25 12:56:09,238 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=748519.3333333334, ans=0.0 2024-09-25 12:56:20,808 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 12:56:28,644 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=748566.0, ans=0.0 2024-09-25 12:57:14,806 INFO [train.py:1198] (2/4) Epoch 42, batch 700, loss[loss=0.2125, ctc_loss=0.1372, cr_loss=0.3765, over 17201.00 frames. ], tot_loss[loss=0.191, ctc_loss=0.1229, cr_loss=0.3407, over 3259245.90 frames. ], batch size: 55, lr: 2.84e-03, grad_scale: 32.0 2024-09-25 12:57:20,031 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=748706.0, ans=0.025 2024-09-25 12:57:21,242 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.310e+02 1.390e+02 1.481e+02 1.937e+02, threshold=2.780e+02, percent-clipped=0.0 2024-09-25 12:57:25,683 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=4.35 vs. limit=15.0 2024-09-25 12:57:31,279 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=748752.6666666666, ans=0.2 2024-09-25 12:57:48,789 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=748799.3333333334, ans=0.0 2024-09-25 12:58:06,322 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=748846.0, ans=0.125 2024-09-25 12:58:14,353 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=748846.0, ans=0.0 2024-09-25 12:58:14,377 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=748846.0, ans=0.125 2024-09-25 12:58:34,916 INFO [train.py:1198] (2/4) Epoch 42, batch 750, loss[loss=0.2115, ctc_loss=0.1359, cr_loss=0.3779, over 17029.00 frames. ], tot_loss[loss=0.1906, ctc_loss=0.1226, cr_loss=0.34, over 3283101.85 frames. ], batch size: 56, lr: 2.84e-03, grad_scale: 32.0 2024-09-25 12:58:57,667 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=13.10 vs. limit=22.5 2024-09-25 12:59:25,569 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.90 vs. limit=10.0 2024-09-25 12:59:36,536 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=749079.3333333334, ans=0.1 2024-09-25 13:00:00,473 INFO [train.py:1198] (2/4) Epoch 42, batch 800, loss[loss=0.1598, ctc_loss=0.09971, cr_loss=0.3003, over 17277.00 frames. ], tot_loss[loss=0.1903, ctc_loss=0.1225, cr_loss=0.3392, over 3297908.67 frames. ], batch size: 42, lr: 2.84e-03, grad_scale: 32.0 2024-09-25 13:00:03,019 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=7.38 vs. limit=15.0 2024-09-25 13:00:06,821 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.092e+02 1.303e+02 1.405e+02 1.532e+02 1.999e+02, threshold=2.810e+02, percent-clipped=0.0 2024-09-25 13:00:19,791 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.00 vs. limit=6.0 2024-09-25 13:00:41,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=749266.0, ans=0.125 2024-09-25 13:00:53,983 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=7.36 vs. limit=15.0 2024-09-25 13:00:59,950 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=749312.6666666666, ans=0.125 2024-09-25 13:01:20,820 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=8.83 vs. limit=12.0 2024-09-25 13:01:22,171 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=749359.3333333334, ans=0.125 2024-09-25 13:01:26,598 INFO [train.py:1198] (2/4) Epoch 42, batch 850, loss[loss=0.1702, ctc_loss=0.1071, cr_loss=0.3154, over 17067.00 frames. ], tot_loss[loss=0.1897, ctc_loss=0.1221, cr_loss=0.338, over 3310459.92 frames. ], batch size: 46, lr: 2.83e-03, grad_scale: 32.0 2024-09-25 13:01:27,015 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=749406.0, ans=0.125 2024-09-25 13:01:38,059 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=749406.0, ans=0.0 2024-09-25 13:01:56,442 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys.whitening_limit, batch_count=749452.6666666666, ans=6.0 2024-09-25 13:02:07,202 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=749499.3333333334, ans=0.0 2024-09-25 13:02:11,978 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=749499.3333333334, ans=0.2 2024-09-25 13:02:12,117 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=749499.3333333334, ans=0.125 2024-09-25 13:02:27,073 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=9.62 vs. limit=12.0 2024-09-25 13:02:31,722 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=749592.6666666666, ans=0.125 2024-09-25 13:02:43,130 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=749592.6666666666, ans=0.125 2024-09-25 13:02:47,616 INFO [train.py:1198] (2/4) Epoch 42, batch 900, loss[loss=0.1962, ctc_loss=0.126, cr_loss=0.351, over 17066.00 frames. ], tot_loss[loss=0.1897, ctc_loss=0.1221, cr_loss=0.338, over 3317440.29 frames. ], batch size: 46, lr: 2.83e-03, grad_scale: 32.0 2024-09-25 13:02:53,984 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.278e+02 1.357e+02 1.447e+02 3.889e+02, threshold=2.715e+02, percent-clipped=1.0 2024-09-25 13:02:57,540 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=749639.3333333334, ans=0.125 2024-09-25 13:03:01,082 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=12.98 vs. limit=22.5 2024-09-25 13:03:04,167 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 13:03:08,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=749686.0, ans=0.1 2024-09-25 13:03:15,491 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=749686.0, ans=0.125 2024-09-25 13:03:36,988 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=749779.3333333334, ans=0.0 2024-09-25 13:03:47,695 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=749779.3333333334, ans=0.125 2024-09-25 13:03:56,297 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=5.70 vs. limit=12.0 2024-09-25 13:04:07,385 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=8.87 vs. limit=15.0 2024-09-25 13:04:08,506 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=749826.0, ans=0.125 2024-09-25 13:04:11,573 INFO [train.py:1198] (2/4) Epoch 42, batch 950, loss[loss=0.196, ctc_loss=0.1243, cr_loss=0.3586, over 17015.00 frames. ], tot_loss[loss=0.1902, ctc_loss=0.1225, cr_loss=0.3387, over 3331219.81 frames. ], batch size: 39, lr: 2.83e-03, grad_scale: 32.0 2024-09-25 13:05:17,734 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=750059.3333333334, ans=0.0 2024-09-25 13:05:37,406 INFO [train.py:1198] (2/4) Epoch 42, batch 1000, loss[loss=0.2348, ctc_loss=0.1505, cr_loss=0.4214, over 17230.00 frames. ], tot_loss[loss=0.1906, ctc_loss=0.1226, cr_loss=0.3397, over 3339452.81 frames. ], batch size: 55, lr: 2.83e-03, grad_scale: 32.0 2024-09-25 13:05:42,909 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.55 vs. limit=12.0 2024-09-25 13:05:43,608 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.041e+02 1.283e+02 1.358e+02 1.462e+02 1.840e+02, threshold=2.715e+02, percent-clipped=0.0 2024-09-25 13:05:59,480 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=750152.6666666666, ans=0.125 2024-09-25 13:06:03,697 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 13:06:47,472 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=4.94 vs. limit=15.0 2024-09-25 13:06:55,147 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=750292.6666666666, ans=0.125 2024-09-25 13:06:59,676 INFO [train.py:1198] (2/4) Epoch 42, batch 1050, loss[loss=0.1491, ctc_loss=0.09531, cr_loss=0.2692, over 17312.00 frames. ], tot_loss[loss=0.1901, ctc_loss=0.1223, cr_loss=0.3386, over 3342772.02 frames. ], batch size: 42, lr: 2.83e-03, grad_scale: 32.0 2024-09-25 13:07:12,815 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=750339.3333333334, ans=0.09899494936611666 2024-09-25 13:07:22,713 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=750386.0, ans=0.025 2024-09-25 13:07:32,381 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=750432.6666666666, ans=0.125 2024-09-25 13:07:38,933 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=750432.6666666666, ans=0.125 2024-09-25 13:07:42,240 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=12.43 vs. limit=12.0 2024-09-25 13:08:09,296 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=750526.0, ans=0.0 2024-09-25 13:08:20,229 INFO [train.py:1198] (2/4) Epoch 42, batch 1100, loss[loss=0.1623, ctc_loss=0.1034, cr_loss=0.2947, over 17103.00 frames. ], tot_loss[loss=0.1897, ctc_loss=0.122, cr_loss=0.3386, over 3351694.35 frames. ], batch size: 43, lr: 2.83e-03, grad_scale: 16.0 2024-09-25 13:08:25,340 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=750572.6666666666, ans=0.125 2024-09-25 13:08:28,296 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.170e+02 1.338e+02 1.422e+02 1.527e+02 1.799e+02, threshold=2.844e+02, percent-clipped=0.0 2024-09-25 13:08:47,088 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=750619.3333333334, ans=0.125 2024-09-25 13:09:01,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=750666.0, ans=0.1 2024-09-25 13:09:14,443 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten.whitening_limit, batch_count=750712.6666666666, ans=15.0 2024-09-25 13:09:22,713 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=750712.6666666666, ans=0.0 2024-09-25 13:09:24,192 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=750712.6666666666, ans=0.125 2024-09-25 13:09:40,165 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=750759.3333333334, ans=0.2 2024-09-25 13:09:44,703 INFO [train.py:1198] (2/4) Epoch 42, batch 1150, loss[loss=0.2202, ctc_loss=0.1433, cr_loss=0.3845, over 14989.00 frames. ], tot_loss[loss=0.1891, ctc_loss=0.1215, cr_loss=0.3381, over 3361064.07 frames. ], batch size: 89, lr: 2.83e-03, grad_scale: 16.0 2024-09-25 13:09:59,546 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=750852.6666666666, ans=0.1 2024-09-25 13:10:01,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=750852.6666666666, ans=0.125 2024-09-25 13:10:21,192 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=750899.3333333334, ans=0.05 2024-09-25 13:10:26,030 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=750899.3333333334, ans=0.04949747468305833 2024-09-25 13:10:29,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=750899.3333333334, ans=0.125 2024-09-25 13:10:32,462 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=750899.3333333334, ans=0.0 2024-09-25 13:10:34,184 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=750946.0, ans=0.1 2024-09-25 13:10:44,353 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.10 vs. limit=6.0 2024-09-25 13:10:54,641 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.16 vs. limit=22.5 2024-09-25 13:11:00,380 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=750992.6666666666, ans=0.0 2024-09-25 13:11:05,143 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=750992.6666666666, ans=0.125 2024-09-25 13:11:09,660 INFO [train.py:1198] (2/4) Epoch 42, batch 1200, loss[loss=0.213, ctc_loss=0.1384, cr_loss=0.373, over 17238.00 frames. ], tot_loss[loss=0.1888, ctc_loss=0.1214, cr_loss=0.3371, over 3361497.66 frames. ], batch size: 55, lr: 2.83e-03, grad_scale: 32.0 2024-09-25 13:11:16,321 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=751039.3333333334, ans=0.0 2024-09-25 13:11:17,515 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.163e+02 1.296e+02 1.388e+02 1.485e+02 1.813e+02, threshold=2.776e+02, percent-clipped=0.0 2024-09-25 13:11:48,168 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=751132.6666666666, ans=0.125 2024-09-25 13:11:56,228 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=751179.3333333334, ans=0.125 2024-09-25 13:12:29,346 INFO [train.py:1198] (2/4) Epoch 42, batch 1250, loss[loss=0.1982, ctc_loss=0.125, cr_loss=0.3658, over 17319.00 frames. ], tot_loss[loss=0.1897, ctc_loss=0.1219, cr_loss=0.3388, over 3359842.29 frames. ], batch size: 49, lr: 2.83e-03, grad_scale: 32.0 2024-09-25 13:12:59,907 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=751366.0, ans=0.0 2024-09-25 13:13:11,656 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.18 vs. limit=15.0 2024-09-25 13:13:14,099 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=751366.0, ans=0.125 2024-09-25 13:13:14,245 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=751366.0, ans=0.1 2024-09-25 13:13:24,306 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=13.13 vs. limit=15.0 2024-09-25 13:13:46,979 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=751459.3333333334, ans=0.0 2024-09-25 13:13:49,756 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=751506.0, ans=0.125 2024-09-25 13:13:51,190 INFO [train.py:1198] (2/4) Epoch 42, batch 1300, loss[loss=0.1845, ctc_loss=0.1162, cr_loss=0.3413, over 17154.00 frames. ], tot_loss[loss=0.1891, ctc_loss=0.1215, cr_loss=0.3378, over 3351815.86 frames. ], batch size: 45, lr: 2.83e-03, grad_scale: 16.0 2024-09-25 13:14:00,129 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=15.74 vs. limit=22.5 2024-09-25 13:14:00,855 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.307e+02 1.370e+02 1.468e+02 2.127e+02, threshold=2.740e+02, percent-clipped=0.0 2024-09-25 13:14:02,677 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=751506.0, ans=0.0 2024-09-25 13:14:16,592 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.max_abs, batch_count=751552.6666666666, ans=10.0 2024-09-25 13:14:23,219 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=751552.6666666666, ans=0.0 2024-09-25 13:14:31,206 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=751599.3333333334, ans=0.1 2024-09-25 13:14:34,357 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=751599.3333333334, ans=0.025 2024-09-25 13:14:44,571 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=7.79 vs. limit=15.0 2024-09-25 13:15:00,280 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=751692.6666666666, ans=0.2 2024-09-25 13:15:00,318 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=751692.6666666666, ans=0.1 2024-09-25 13:15:08,152 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=751692.6666666666, ans=0.0 2024-09-25 13:15:16,745 INFO [train.py:1198] (2/4) Epoch 42, batch 1350, loss[loss=0.2118, ctc_loss=0.1365, cr_loss=0.3764, over 17002.00 frames. ], tot_loss[loss=0.1895, ctc_loss=0.1217, cr_loss=0.3388, over 3358170.07 frames. ], batch size: 56, lr: 2.83e-03, grad_scale: 16.0 2024-09-25 13:15:41,005 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=751786.0, ans=0.125 2024-09-25 13:15:56,731 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.35 vs. limit=6.0 2024-09-25 13:16:11,222 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.63 vs. limit=10.0 2024-09-25 13:16:18,262 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=751879.3333333334, ans=0.05 2024-09-25 13:16:38,718 INFO [train.py:1198] (2/4) Epoch 42, batch 1400, loss[loss=0.1976, ctc_loss=0.1267, cr_loss=0.3542, over 17223.00 frames. ], tot_loss[loss=0.1896, ctc_loss=0.1218, cr_loss=0.3391, over 3358554.65 frames. ], batch size: 50, lr: 2.83e-03, grad_scale: 16.0 2024-09-25 13:16:41,099 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.out_whiten.whitening_limit, batch_count=751972.6666666666, ans=15.0 2024-09-25 13:16:48,176 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.146e+02 1.272e+02 1.355e+02 1.432e+02 2.085e+02, threshold=2.710e+02, percent-clipped=0.0 2024-09-25 13:17:47,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=752159.3333333334, ans=0.125 2024-09-25 13:17:47,481 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=752159.3333333334, ans=0.125 2024-09-25 13:17:58,206 INFO [train.py:1198] (2/4) Epoch 42, batch 1450, loss[loss=0.1688, ctc_loss=0.1053, cr_loss=0.3177, over 17309.00 frames. ], tot_loss[loss=0.1897, ctc_loss=0.1219, cr_loss=0.3391, over 3362661.05 frames. ], batch size: 46, lr: 2.83e-03, grad_scale: 16.0 2024-09-25 13:18:09,859 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=752206.0, ans=0.1 2024-09-25 13:18:21,607 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=752252.6666666666, ans=0.0 2024-09-25 13:18:26,994 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=11.17 vs. limit=22.5 2024-09-25 13:18:40,253 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=752299.3333333334, ans=0.95 2024-09-25 13:18:49,500 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=752346.0, ans=0.0 2024-09-25 13:18:51,569 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=10.91 vs. limit=22.5 2024-09-25 13:18:52,578 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=752346.0, ans=0.025 2024-09-25 13:18:54,794 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.75 vs. limit=15.0 2024-09-25 13:18:56,289 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=8.84 vs. limit=15.0 2024-09-25 13:19:16,483 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn2.whiten, num_groups=1, num_channels=512, metric=9.82 vs. limit=22.5 2024-09-25 13:19:23,690 INFO [train.py:1198] (2/4) Epoch 42, batch 1500, loss[loss=0.1874, ctc_loss=0.1231, cr_loss=0.3212, over 16891.00 frames. ], tot_loss[loss=0.191, ctc_loss=0.1229, cr_loss=0.3407, over 3352797.30 frames. ], batch size: 58, lr: 2.83e-03, grad_scale: 16.0 2024-09-25 13:19:33,186 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.208e+02 1.278e+02 1.340e+02 1.432e+02 2.576e+02, threshold=2.680e+02, percent-clipped=0.0 2024-09-25 13:19:33,578 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=752439.3333333334, ans=0.125 2024-09-25 13:19:54,739 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.79 vs. limit=6.0 2024-09-25 13:19:55,914 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=752532.6666666666, ans=0.0 2024-09-25 13:20:01,137 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=2.69 vs. limit=6.0 2024-09-25 13:20:02,259 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=752532.6666666666, ans=0.125 2024-09-25 13:20:03,846 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=752532.6666666666, ans=0.125 2024-09-25 13:20:21,004 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=752579.3333333334, ans=0.2 2024-09-25 13:20:22,548 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=752579.3333333334, ans=0.0 2024-09-25 13:20:48,640 INFO [train.py:1198] (2/4) Epoch 42, batch 1550, loss[loss=0.2042, ctc_loss=0.1319, cr_loss=0.3612, over 17316.00 frames. ], tot_loss[loss=0.1921, ctc_loss=0.1236, cr_loss=0.3424, over 3351085.05 frames. ], batch size: 51, lr: 2.83e-03, grad_scale: 16.0 2024-09-25 13:20:57,138 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=752672.6666666666, ans=0.04949747468305833 2024-09-25 13:21:00,273 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=752672.6666666666, ans=0.025 2024-09-25 13:21:05,581 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.52 vs. limit=10.0 2024-09-25 13:21:06,702 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=752719.3333333334, ans=0.125 2024-09-25 13:21:11,424 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=752719.3333333334, ans=0.0 2024-09-25 13:21:33,951 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=752766.0, ans=0.0 2024-09-25 13:21:48,472 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=752812.6666666666, ans=0.125 2024-09-25 13:22:06,076 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=752859.3333333334, ans=0.1 2024-09-25 13:22:09,035 INFO [train.py:1198] (2/4) Epoch 42, batch 1600, loss[loss=0.1799, ctc_loss=0.1155, cr_loss=0.322, over 17152.00 frames. ], tot_loss[loss=0.1922, ctc_loss=0.1237, cr_loss=0.3424, over 3353007.91 frames. ], batch size: 48, lr: 2.83e-03, grad_scale: 16.0 2024-09-25 13:22:11,178 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=11.59 vs. limit=12.0 2024-09-25 13:22:18,184 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=3.79 vs. limit=15.0 2024-09-25 13:22:18,974 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=752906.0, ans=0.2 2024-09-25 13:22:20,350 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.308e+02 1.394e+02 1.520e+02 2.214e+02, threshold=2.789e+02, percent-clipped=0.0 2024-09-25 13:22:20,643 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=752906.0, ans=0.0 2024-09-25 13:22:21,179 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=13.87 vs. limit=22.5 2024-09-25 13:22:22,715 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.82 vs. limit=15.0 2024-09-25 13:22:32,698 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module1.whiten, num_groups=1, num_channels=192, metric=9.23 vs. limit=15.0 2024-09-25 13:22:49,625 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=752999.3333333334, ans=0.0 2024-09-25 13:22:49,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=752999.3333333334, ans=0.0 2024-09-25 13:23:31,967 INFO [train.py:1198] (2/4) Epoch 42, batch 1650, loss[loss=0.2099, ctc_loss=0.1388, cr_loss=0.3556, over 17041.00 frames. ], tot_loss[loss=0.1917, ctc_loss=0.1235, cr_loss=0.3411, over 3342272.20 frames. ], batch size: 52, lr: 2.83e-03, grad_scale: 8.0 2024-09-25 13:23:40,257 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 13:24:07,294 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=753232.6666666666, ans=0.0 2024-09-25 13:24:17,032 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=753232.6666666666, ans=0.125 2024-09-25 13:24:37,866 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.max_abs, batch_count=753326.0, ans=10.0 2024-09-25 13:24:41,021 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=753326.0, ans=0.125 2024-09-25 13:24:46,163 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.30 vs. limit=15.0 2024-09-25 13:24:55,037 INFO [train.py:1198] (2/4) Epoch 42, batch 1700, loss[loss=0.1659, ctc_loss=0.1029, cr_loss=0.3151, over 16938.00 frames. ], tot_loss[loss=0.1916, ctc_loss=0.1234, cr_loss=0.341, over 3333430.33 frames. ], batch size: 42, lr: 2.83e-03, grad_scale: 8.0 2024-09-25 13:25:10,263 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.072e+02 1.274e+02 1.365e+02 1.477e+02 2.615e+02, threshold=2.729e+02, percent-clipped=0.0 2024-09-25 13:25:16,894 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=753419.3333333334, ans=0.1 2024-09-25 13:25:20,554 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.whiten, num_groups=1, num_channels=512, metric=4.63 vs. limit=12.0 2024-09-25 13:25:38,931 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=14.13 vs. limit=22.5 2024-09-25 13:26:11,904 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=753559.3333333334, ans=0.125 2024-09-25 13:26:18,186 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=753606.0, ans=0.125 2024-09-25 13:26:19,504 INFO [train.py:1198] (2/4) Epoch 42, batch 1750, loss[loss=0.1893, ctc_loss=0.1226, cr_loss=0.3336, over 17065.00 frames. ], tot_loss[loss=0.1905, ctc_loss=0.1226, cr_loss=0.3395, over 3338306.46 frames. ], batch size: 46, lr: 2.83e-03, grad_scale: 8.0 2024-09-25 13:26:27,974 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=753606.0, ans=0.125 2024-09-25 13:26:31,322 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=753606.0, ans=0.1 2024-09-25 13:26:49,106 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=753652.6666666666, ans=0.125 2024-09-25 13:27:11,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=753746.0, ans=0.1 2024-09-25 13:27:36,502 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.65 vs. limit=5.0 2024-09-25 13:27:40,014 INFO [train.py:1198] (2/4) Epoch 42, batch 1800, loss[loss=0.1949, ctc_loss=0.1238, cr_loss=0.3556, over 17219.00 frames. ], tot_loss[loss=0.192, ctc_loss=0.1236, cr_loss=0.3418, over 3342422.87 frames. ], batch size: 55, lr: 2.83e-03, grad_scale: 8.0 2024-09-25 13:27:43,634 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=753839.3333333334, ans=0.125 2024-09-25 13:27:52,869 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.293e+02 1.332e+02 1.454e+02 1.867e+02, threshold=2.665e+02, percent-clipped=0.0 2024-09-25 13:28:33,301 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=753979.3333333334, ans=0.1 2024-09-25 13:28:51,131 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=754026.0, ans=0.1 2024-09-25 13:28:54,239 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=754026.0, ans=0.125 2024-09-25 13:29:04,500 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=754072.6666666666, ans=0.2 2024-09-25 13:29:05,784 INFO [train.py:1198] (2/4) Epoch 42, batch 1850, loss[loss=0.1957, ctc_loss=0.1279, cr_loss=0.339, over 17038.00 frames. ], tot_loss[loss=0.192, ctc_loss=0.1237, cr_loss=0.3418, over 3351228.46 frames. ], batch size: 56, lr: 2.83e-03, grad_scale: 8.0 2024-09-25 13:29:09,386 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=754072.6666666666, ans=0.125 2024-09-25 13:29:23,660 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=754119.3333333334, ans=0.2 2024-09-25 13:29:42,292 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.50 vs. limit=12.0 2024-09-25 13:29:43,905 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.04 vs. limit=15.0 2024-09-25 13:29:52,843 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=754212.6666666666, ans=0.125 2024-09-25 13:30:06,576 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.min_abs, batch_count=754212.6666666666, ans=0.5 2024-09-25 13:30:31,298 INFO [train.py:1198] (2/4) Epoch 42, batch 1900, loss[loss=0.1828, ctc_loss=0.1165, cr_loss=0.3318, over 17090.00 frames. ], tot_loss[loss=0.1913, ctc_loss=0.1232, cr_loss=0.3405, over 3347384.09 frames. ], batch size: 43, lr: 2.83e-03, grad_scale: 8.0 2024-09-25 13:30:33,125 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=754306.0, ans=0.125 2024-09-25 13:30:36,360 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=754306.0, ans=0.0 2024-09-25 13:30:36,816 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=3.52 vs. limit=6.0 2024-09-25 13:30:42,985 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=754306.0, ans=0.0 2024-09-25 13:30:44,083 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.145e+02 1.292e+02 1.376e+02 1.481e+02 2.312e+02, threshold=2.752e+02, percent-clipped=0.0 2024-09-25 13:31:50,779 INFO [train.py:1198] (2/4) Epoch 42, batch 1950, loss[loss=0.2044, ctc_loss=0.1287, cr_loss=0.3782, over 17018.00 frames. ], tot_loss[loss=0.1907, ctc_loss=0.1227, cr_loss=0.3398, over 3347470.19 frames. ], batch size: 52, lr: 2.83e-03, grad_scale: 8.0 2024-09-25 13:32:02,302 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=754539.3333333334, ans=0.125 2024-09-25 13:32:03,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=754539.3333333334, ans=0.1 2024-09-25 13:32:12,092 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=754586.0, ans=0.0 2024-09-25 13:32:17,018 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=754586.0, ans=0.125 2024-09-25 13:32:27,937 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=754632.6666666666, ans=0.025 2024-09-25 13:32:29,459 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=754632.6666666666, ans=0.0 2024-09-25 13:32:32,850 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=754632.6666666666, ans=0.2 2024-09-25 13:32:37,724 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=754679.3333333334, ans=0.2 2024-09-25 13:32:57,520 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.30 vs. limit=10.0 2024-09-25 13:33:13,475 INFO [train.py:1198] (2/4) Epoch 42, batch 2000, loss[loss=0.2094, ctc_loss=0.1366, cr_loss=0.364, over 17146.00 frames. ], tot_loss[loss=0.1906, ctc_loss=0.1226, cr_loss=0.3398, over 3353578.81 frames. ], batch size: 48, lr: 2.82e-03, grad_scale: 16.0 2024-09-25 13:33:21,659 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=754772.6666666666, ans=0.0 2024-09-25 13:33:25,998 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.295e+02 1.343e+02 1.422e+02 2.059e+02, threshold=2.687e+02, percent-clipped=0.0 2024-09-25 13:34:01,302 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 13:34:02,887 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=754912.6666666666, ans=0.125 2024-09-25 13:34:07,618 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=754912.6666666666, ans=0.0 2024-09-25 13:34:33,206 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=754959.3333333334, ans=0.0 2024-09-25 13:34:36,177 INFO [train.py:1198] (2/4) Epoch 42, batch 2050, loss[loss=0.1567, ctc_loss=0.09914, cr_loss=0.2877, over 17199.00 frames. ], tot_loss[loss=0.1906, ctc_loss=0.1226, cr_loss=0.34, over 3362607.72 frames. ], batch size: 41, lr: 2.82e-03, grad_scale: 16.0 2024-09-25 13:35:11,409 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=755099.3333333334, ans=0.0 2024-09-25 13:35:21,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=755099.3333333334, ans=0.015 2024-09-25 13:36:00,178 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=755239.3333333334, ans=0.05 2024-09-25 13:36:01,516 INFO [train.py:1198] (2/4) Epoch 42, batch 2100, loss[loss=0.1848, ctc_loss=0.1209, cr_loss=0.3194, over 17347.00 frames. ], tot_loss[loss=0.1909, ctc_loss=0.1228, cr_loss=0.3403, over 3362289.32 frames. ], batch size: 48, lr: 2.82e-03, grad_scale: 16.0 2024-09-25 13:36:14,532 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.293e+02 1.359e+02 1.448e+02 2.316e+02, threshold=2.717e+02, percent-clipped=0.0 2024-09-25 13:36:16,506 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=755286.0, ans=0.125 2024-09-25 13:36:31,069 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=755286.0, ans=0.0 2024-09-25 13:36:34,519 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=11.83 vs. limit=12.0 2024-09-25 13:37:21,952 INFO [train.py:1198] (2/4) Epoch 42, batch 2150, loss[loss=0.2267, ctc_loss=0.1498, cr_loss=0.3844, over 16896.00 frames. ], tot_loss[loss=0.1915, ctc_loss=0.1233, cr_loss=0.3409, over 3358435.13 frames. ], batch size: 58, lr: 2.82e-03, grad_scale: 16.0 2024-09-25 13:37:42,471 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module1.whiten, num_groups=1, num_channels=192, metric=8.04 vs. limit=15.0 2024-09-25 13:38:44,292 INFO [train.py:1198] (2/4) Epoch 42, batch 2200, loss[loss=0.2079, ctc_loss=0.1354, cr_loss=0.3629, over 17091.00 frames. ], tot_loss[loss=0.1911, ctc_loss=0.1229, cr_loss=0.3412, over 3364966.47 frames. ], batch size: 49, lr: 2.82e-03, grad_scale: 16.0 2024-09-25 13:38:55,401 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=755706.0, ans=0.125 2024-09-25 13:38:59,636 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.309e+02 1.387e+02 1.496e+02 2.285e+02, threshold=2.773e+02, percent-clipped=0.0 2024-09-25 13:39:30,361 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=755799.3333333334, ans=0.125 2024-09-25 13:39:57,710 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=755892.6666666666, ans=0.2 2024-09-25 13:40:06,794 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=755892.6666666666, ans=0.0 2024-09-25 13:40:09,777 INFO [train.py:1198] (2/4) Epoch 42, batch 2250, loss[loss=0.1606, ctc_loss=0.09996, cr_loss=0.3033, over 17106.00 frames. ], tot_loss[loss=0.1906, ctc_loss=0.1225, cr_loss=0.3407, over 3367206.03 frames. ], batch size: 40, lr: 2.82e-03, grad_scale: 16.0 2024-09-25 13:40:21,241 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=5.09 vs. limit=6.0 2024-09-25 13:40:22,955 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=4.24 vs. limit=15.0 2024-09-25 13:40:28,602 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 13:40:30,350 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=755986.0, ans=0.2 2024-09-25 13:41:01,016 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=756079.3333333334, ans=0.125 2024-09-25 13:41:23,618 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=756126.0, ans=0.125 2024-09-25 13:41:33,095 INFO [train.py:1198] (2/4) Epoch 42, batch 2300, loss[loss=0.1409, ctc_loss=0.08662, cr_loss=0.2716, over 17197.00 frames. ], tot_loss[loss=0.1899, ctc_loss=0.122, cr_loss=0.3395, over 3368191.98 frames. ], batch size: 41, lr: 2.82e-03, grad_scale: 16.0 2024-09-25 13:41:42,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=756172.6666666666, ans=0.125 2024-09-25 13:41:45,867 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.265e+02 1.352e+02 1.460e+02 1.967e+02, threshold=2.704e+02, percent-clipped=0.0 2024-09-25 13:41:52,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=756219.3333333334, ans=0.1 2024-09-25 13:42:04,046 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=756266.0, ans=0.125 2024-09-25 13:42:12,283 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 13:42:16,954 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=756266.0, ans=0.0 2024-09-25 13:42:21,776 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=756312.6666666666, ans=0.125 2024-09-25 13:42:39,753 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=9.26 vs. limit=15.0 2024-09-25 13:42:51,959 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=756406.0, ans=0.09899494936611666 2024-09-25 13:42:53,052 INFO [train.py:1198] (2/4) Epoch 42, batch 2350, loss[loss=0.2097, ctc_loss=0.1407, cr_loss=0.3449, over 11680.00 frames. ], tot_loss[loss=0.19, ctc_loss=0.1222, cr_loss=0.3393, over 3356564.32 frames. ], batch size: 123, lr: 2.82e-03, grad_scale: 16.0 2024-09-25 13:43:02,944 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.80 vs. limit=15.0 2024-09-25 13:43:08,841 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=756406.0, ans=0.0 2024-09-25 13:43:24,805 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=756452.6666666666, ans=0.0 2024-09-25 13:43:34,351 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=756499.3333333334, ans=0.0 2024-09-25 13:43:57,910 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=756546.0, ans=0.125 2024-09-25 13:44:14,287 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.49 vs. limit=15.0 2024-09-25 13:44:18,217 INFO [train.py:1198] (2/4) Epoch 42, batch 2400, loss[loss=0.1942, ctc_loss=0.1237, cr_loss=0.3524, over 17202.00 frames. ], tot_loss[loss=0.1908, ctc_loss=0.1227, cr_loss=0.3405, over 3361002.60 frames. ], batch size: 55, lr: 2.82e-03, grad_scale: 32.0 2024-09-25 13:44:30,873 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.304e+02 1.412e+02 1.522e+02 4.391e+02, threshold=2.825e+02, percent-clipped=1.0 2024-09-25 13:44:32,801 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=756686.0, ans=0.1 2024-09-25 13:44:52,011 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=756732.6666666666, ans=0.0 2024-09-25 13:45:12,500 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=756779.3333333334, ans=0.125 2024-09-25 13:45:15,915 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=11.76 vs. limit=22.5 2024-09-25 13:45:43,397 INFO [train.py:1198] (2/4) Epoch 42, batch 2450, loss[loss=0.1612, ctc_loss=0.09996, cr_loss=0.3062, over 17013.00 frames. ], tot_loss[loss=0.1913, ctc_loss=0.123, cr_loss=0.3412, over 3359229.11 frames. ], batch size: 39, lr: 2.82e-03, grad_scale: 32.0 2024-09-25 13:45:47,666 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.75 vs. limit=12.0 2024-09-25 13:46:29,521 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.49 vs. limit=22.5 2024-09-25 13:46:43,528 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=757012.6666666666, ans=0.125 2024-09-25 13:46:47,314 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=4.80 vs. limit=15.0 2024-09-25 13:46:48,825 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=3.70 vs. limit=10.0 2024-09-25 13:47:03,850 INFO [train.py:1198] (2/4) Epoch 42, batch 2500, loss[loss=0.1485, ctc_loss=0.09391, cr_loss=0.2728, over 17259.00 frames. ], tot_loss[loss=0.1916, ctc_loss=0.1233, cr_loss=0.3415, over 3358339.23 frames. ], batch size: 42, lr: 2.82e-03, grad_scale: 16.0 2024-09-25 13:47:18,220 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.143e+02 1.315e+02 1.378e+02 1.471e+02 2.327e+02, threshold=2.756e+02, percent-clipped=0.0 2024-09-25 13:48:06,198 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=757246.0, ans=0.2 2024-09-25 13:48:14,484 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=4.03 vs. limit=15.0 2024-09-25 13:48:24,551 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.30 vs. limit=15.0 2024-09-25 13:48:26,432 INFO [train.py:1198] (2/4) Epoch 42, batch 2550, loss[loss=0.2292, ctc_loss=0.1508, cr_loss=0.392, over 15288.00 frames. ], tot_loss[loss=0.1918, ctc_loss=0.1235, cr_loss=0.3413, over 3355487.15 frames. ], batch size: 89, lr: 2.82e-03, grad_scale: 16.0 2024-09-25 13:48:53,583 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.66 vs. limit=15.0 2024-09-25 13:49:09,004 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=757432.6666666666, ans=0.07 2024-09-25 13:49:12,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=757432.6666666666, ans=0.0 2024-09-25 13:49:19,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=757479.3333333334, ans=0.1 2024-09-25 13:49:20,173 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=757479.3333333334, ans=0.0 2024-09-25 13:49:48,976 INFO [train.py:1198] (2/4) Epoch 42, batch 2600, loss[loss=0.1685, ctc_loss=0.1069, cr_loss=0.3083, over 17323.00 frames. ], tot_loss[loss=0.1905, ctc_loss=0.1226, cr_loss=0.3394, over 3360274.34 frames. ], batch size: 51, lr: 2.82e-03, grad_scale: 16.0 2024-09-25 13:50:03,003 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=757572.6666666666, ans=0.125 2024-09-25 13:50:05,736 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.054e+02 1.272e+02 1.357e+02 1.427e+02 2.038e+02, threshold=2.713e+02, percent-clipped=0.0 2024-09-25 13:50:36,485 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.68 vs. limit=12.0 2024-09-25 13:50:39,446 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=8.04 vs. limit=15.0 2024-09-25 13:51:00,846 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=757759.3333333334, ans=0.125 2024-09-25 13:51:13,646 INFO [train.py:1198] (2/4) Epoch 42, batch 2650, loss[loss=0.2067, ctc_loss=0.1362, cr_loss=0.3525, over 15258.00 frames. ], tot_loss[loss=0.1901, ctc_loss=0.1223, cr_loss=0.3391, over 3360412.27 frames. ], batch size: 89, lr: 2.82e-03, grad_scale: 16.0 2024-09-25 13:51:26,954 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=757806.0, ans=0.2 2024-09-25 13:51:34,302 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=9.95 vs. limit=10.0 2024-09-25 13:52:00,747 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=757946.0, ans=0.2 2024-09-25 13:52:07,767 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.68 vs. limit=15.0 2024-09-25 13:52:08,645 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=757946.0, ans=0.0 2024-09-25 13:52:24,969 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=757992.6666666666, ans=0.2 2024-09-25 13:52:34,234 INFO [train.py:1198] (2/4) Epoch 42, batch 2700, loss[loss=0.173, ctc_loss=0.1105, cr_loss=0.3125, over 17196.00 frames. ], tot_loss[loss=0.1901, ctc_loss=0.1223, cr_loss=0.3393, over 3350920.18 frames. ], batch size: 47, lr: 2.82e-03, grad_scale: 8.0 2024-09-25 13:52:44,952 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=10.47 vs. limit=22.5 2024-09-25 13:52:45,550 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=758039.3333333334, ans=0.0 2024-09-25 13:52:47,841 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=8.80 vs. limit=15.0 2024-09-25 13:52:50,044 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.301e+02 1.411e+02 1.546e+02 2.916e+02, threshold=2.822e+02, percent-clipped=1.0 2024-09-25 13:53:16,890 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=758132.6666666666, ans=0.125 2024-09-25 13:53:59,181 INFO [train.py:1198] (2/4) Epoch 42, batch 2750, loss[loss=0.2325, ctc_loss=0.1585, cr_loss=0.37, over 14928.00 frames. ], tot_loss[loss=0.1895, ctc_loss=0.1217, cr_loss=0.3389, over 3357947.92 frames. ], batch size: 89, lr: 2.82e-03, grad_scale: 8.0 2024-09-25 13:54:11,981 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=758272.6666666666, ans=0.0 2024-09-25 13:54:31,371 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=758366.0, ans=0.125 2024-09-25 13:54:36,012 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=758366.0, ans=0.1 2024-09-25 13:55:05,062 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.63 vs. limit=6.0 2024-09-25 13:55:23,620 INFO [train.py:1198] (2/4) Epoch 42, batch 2800, loss[loss=0.1552, ctc_loss=0.09749, cr_loss=0.2885, over 17060.00 frames. ], tot_loss[loss=0.1897, ctc_loss=0.1219, cr_loss=0.3393, over 3361386.62 frames. ], batch size: 39, lr: 2.82e-03, grad_scale: 16.0 2024-09-25 13:55:26,974 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=758506.0, ans=0.125 2024-09-25 13:55:39,669 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.160e+02 1.295e+02 1.378e+02 1.487e+02 2.331e+02, threshold=2.756e+02, percent-clipped=0.0 2024-09-25 13:55:41,662 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=758552.6666666666, ans=0.125 2024-09-25 13:56:15,583 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.17 vs. limit=15.0 2024-09-25 13:56:18,736 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=11.85 vs. limit=15.0 2024-09-25 13:56:34,456 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=758692.6666666666, ans=0.0 2024-09-25 13:56:43,681 INFO [train.py:1198] (2/4) Epoch 42, batch 2850, loss[loss=0.1883, ctc_loss=0.1243, cr_loss=0.3196, over 16935.00 frames. ], tot_loss[loss=0.1895, ctc_loss=0.1217, cr_loss=0.3391, over 3364338.04 frames. ], batch size: 42, lr: 2.82e-03, grad_scale: 16.0 2024-09-25 13:56:50,611 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 13:56:50,909 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.91 vs. limit=15.0 2024-09-25 13:56:55,362 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=758739.3333333334, ans=0.2 2024-09-25 13:57:06,622 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=758786.0, ans=0.125 2024-09-25 13:57:32,262 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=758879.3333333334, ans=0.1 2024-09-25 13:57:43,579 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=758879.3333333334, ans=0.125 2024-09-25 13:57:48,449 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=758926.0, ans=0.125 2024-09-25 13:58:06,778 INFO [train.py:1198] (2/4) Epoch 42, batch 2900, loss[loss=0.2014, ctc_loss=0.1311, cr_loss=0.3519, over 17310.00 frames. ], tot_loss[loss=0.1884, ctc_loss=0.121, cr_loss=0.3371, over 3362297.47 frames. ], batch size: 51, lr: 2.82e-03, grad_scale: 16.0 2024-09-25 13:58:14,953 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=758972.6666666666, ans=0.125 2024-09-25 13:58:22,533 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.147e+02 1.263e+02 1.349e+02 1.466e+02 2.283e+02, threshold=2.697e+02, percent-clipped=0.0 2024-09-25 13:58:48,113 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=759066.0, ans=0.025 2024-09-25 13:58:50,123 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=5.75 vs. limit=12.0 2024-09-25 13:59:11,491 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=9.15 vs. limit=10.0 2024-09-25 13:59:24,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.min_positive, batch_count=759159.3333333334, ans=0.025 2024-09-25 13:59:29,384 INFO [train.py:1198] (2/4) Epoch 42, batch 2950, loss[loss=0.1738, ctc_loss=0.1084, cr_loss=0.3272, over 17249.00 frames. ], tot_loss[loss=0.1882, ctc_loss=0.1208, cr_loss=0.337, over 3366318.91 frames. ], batch size: 44, lr: 2.82e-03, grad_scale: 16.0 2024-09-25 13:59:44,138 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=759252.6666666666, ans=0.125 2024-09-25 14:00:07,759 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=5.05 vs. limit=15.0 2024-09-25 14:00:18,509 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten.whitening_limit, batch_count=759299.3333333334, ans=22.5 2024-09-25 14:00:27,351 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=759346.0, ans=0.125 2024-09-25 14:00:53,412 INFO [train.py:1198] (2/4) Epoch 42, batch 3000, loss[loss=0.1963, ctc_loss=0.1249, cr_loss=0.3569, over 17024.00 frames. ], tot_loss[loss=0.1894, ctc_loss=0.1217, cr_loss=0.3386, over 3361740.69 frames. ], batch size: 51, lr: 2.82e-03, grad_scale: 16.0 2024-09-25 14:00:53,412 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 14:01:08,867 INFO [train.py:1230] (2/4) Epoch 42, validation: loss=0.03543, ctc_loss=0.03543, cr_loss=1.019e-14, over 944034.00 frames. 2024-09-25 14:01:08,868 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 14:01:15,677 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=759439.3333333334, ans=0.125 2024-09-25 14:01:24,612 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.139e+02 1.302e+02 1.378e+02 1.459e+02 2.338e+02, threshold=2.756e+02, percent-clipped=0.0 2024-09-25 14:02:01,345 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=3.94 vs. limit=15.0 2024-09-25 14:02:20,521 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=759626.0, ans=0.1 2024-09-25 14:02:26,593 INFO [train.py:1198] (2/4) Epoch 42, batch 3050, loss[loss=0.2159, ctc_loss=0.142, cr_loss=0.3694, over 17036.00 frames. ], tot_loss[loss=0.1891, ctc_loss=0.1215, cr_loss=0.3381, over 3362444.46 frames. ], batch size: 56, lr: 2.82e-03, grad_scale: 16.0 2024-09-25 14:02:42,611 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=759719.3333333334, ans=0.0 2024-09-25 14:03:08,929 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.30 vs. limit=15.0 2024-09-25 14:03:31,661 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=759859.3333333334, ans=0.125 2024-09-25 14:03:45,316 INFO [train.py:1198] (2/4) Epoch 42, batch 3100, loss[loss=0.2223, ctc_loss=0.1512, cr_loss=0.3554, over 11446.00 frames. ], tot_loss[loss=0.1896, ctc_loss=0.122, cr_loss=0.3382, over 3359293.14 frames. ], batch size: 124, lr: 2.82e-03, grad_scale: 16.0 2024-09-25 14:03:59,679 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=759952.6666666666, ans=0.2 2024-09-25 14:04:00,900 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.153e+02 1.293e+02 1.378e+02 1.463e+02 2.447e+02, threshold=2.756e+02, percent-clipped=0.0 2024-09-25 14:04:13,460 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=759952.6666666666, ans=0.125 2024-09-25 14:04:18,137 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=759999.3333333334, ans=0.125 2024-09-25 14:04:23,446 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.12 vs. limit=10.0 2024-09-25 14:05:03,355 INFO [train.py:1198] (2/4) Epoch 42, batch 3150, loss[loss=0.1803, ctc_loss=0.1153, cr_loss=0.3251, over 16947.00 frames. ], tot_loss[loss=0.1894, ctc_loss=0.1218, cr_loss=0.3377, over 3362532.19 frames. ], batch size: 42, lr: 2.81e-03, grad_scale: 16.0 2024-09-25 14:05:11,564 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=760139.3333333334, ans=0.1 2024-09-25 14:05:11,606 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=760139.3333333334, ans=0.2 2024-09-25 14:05:13,724 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=6.76 vs. limit=15.0 2024-09-25 14:05:23,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=760186.0, ans=0.0 2024-09-25 14:05:45,050 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=760232.6666666666, ans=0.1 2024-09-25 14:05:57,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=760279.3333333334, ans=0.0 2024-09-25 14:06:02,132 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 14:06:11,408 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=760326.0, ans=0.2 2024-09-25 14:06:23,519 INFO [train.py:1198] (2/4) Epoch 42, batch 3200, loss[loss=0.1428, ctc_loss=0.08786, cr_loss=0.2749, over 17104.00 frames. ], tot_loss[loss=0.1892, ctc_loss=0.1217, cr_loss=0.3372, over 3362032.02 frames. ], batch size: 40, lr: 2.81e-03, grad_scale: 32.0 2024-09-25 14:06:31,675 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=760372.6666666666, ans=0.0 2024-09-25 14:06:36,452 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=760372.6666666666, ans=0.0 2024-09-25 14:06:39,248 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.120e+02 1.277e+02 1.364e+02 1.425e+02 3.566e+02, threshold=2.728e+02, percent-clipped=1.0 2024-09-25 14:06:59,162 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.whiten, num_groups=1, num_channels=192, metric=4.04 vs. limit=12.0 2024-09-25 14:07:13,983 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=760512.6666666666, ans=0.1 2024-09-25 14:07:32,041 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 14:07:39,624 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=760559.3333333334, ans=0.5 2024-09-25 14:07:41,280 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=760559.3333333334, ans=0.125 2024-09-25 14:07:43,979 INFO [train.py:1198] (2/4) Epoch 42, batch 3250, loss[loss=0.2323, ctc_loss=0.1507, cr_loss=0.4082, over 17142.00 frames. ], tot_loss[loss=0.1901, ctc_loss=0.1223, cr_loss=0.339, over 3352588.36 frames. ], batch size: 48, lr: 2.81e-03, grad_scale: 32.0 2024-09-25 14:07:59,717 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=760652.6666666666, ans=0.125 2024-09-25 14:08:12,484 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=760652.6666666666, ans=0.125 2024-09-25 14:08:17,017 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=760699.3333333334, ans=0.0 2024-09-25 14:08:29,985 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.27 vs. limit=10.0 2024-09-25 14:08:38,285 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=10.26 vs. limit=15.0 2024-09-25 14:09:02,228 INFO [train.py:1198] (2/4) Epoch 42, batch 3300, loss[loss=0.2088, ctc_loss=0.1373, cr_loss=0.3572, over 16041.00 frames. ], tot_loss[loss=0.1899, ctc_loss=0.1223, cr_loss=0.3382, over 3346528.35 frames. ], batch size: 74, lr: 2.81e-03, grad_scale: 16.0 2024-09-25 14:09:10,245 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=760839.3333333334, ans=0.1 2024-09-25 14:09:19,405 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.292e+02 1.397e+02 1.503e+02 2.274e+02, threshold=2.794e+02, percent-clipped=0.0 2024-09-25 14:09:25,974 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=760886.0, ans=0.125 2024-09-25 14:09:29,040 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.max_positive, batch_count=760886.0, ans=0.95 2024-09-25 14:09:38,318 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=760932.6666666666, ans=0.0 2024-09-25 14:09:47,736 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=760979.3333333334, ans=0.1 2024-09-25 14:09:58,456 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=760979.3333333334, ans=0.125 2024-09-25 14:10:22,240 INFO [train.py:1198] (2/4) Epoch 42, batch 3350, loss[loss=0.1944, ctc_loss=0.1221, cr_loss=0.3616, over 17070.00 frames. ], tot_loss[loss=0.1908, ctc_loss=0.1229, cr_loss=0.3397, over 3345105.39 frames. ], batch size: 46, lr: 2.81e-03, grad_scale: 16.0 2024-09-25 14:10:36,103 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.45 vs. limit=10.0 2024-09-25 14:10:45,184 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=761119.3333333334, ans=0.125 2024-09-25 14:10:45,300 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=761119.3333333334, ans=0.125 2024-09-25 14:10:48,281 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=761119.3333333334, ans=0.125 2024-09-25 14:11:00,912 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=761166.0, ans=0.0 2024-09-25 14:11:05,862 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.07 vs. limit=15.0 2024-09-25 14:11:42,862 INFO [train.py:1198] (2/4) Epoch 42, batch 3400, loss[loss=0.1884, ctc_loss=0.119, cr_loss=0.3471, over 17143.00 frames. ], tot_loss[loss=0.191, ctc_loss=0.123, cr_loss=0.34, over 3351567.31 frames. ], batch size: 48, lr: 2.81e-03, grad_scale: 16.0 2024-09-25 14:11:49,498 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=761306.0, ans=0.1 2024-09-25 14:12:00,049 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.177e+02 1.312e+02 1.394e+02 1.492e+02 2.078e+02, threshold=2.788e+02, percent-clipped=0.0 2024-09-25 14:12:06,603 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=761352.6666666666, ans=0.0 2024-09-25 14:12:22,539 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=761399.3333333334, ans=0.0 2024-09-25 14:12:30,392 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=761446.0, ans=0.09899494936611666 2024-09-25 14:12:58,483 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=761492.6666666666, ans=0.04949747468305833 2024-09-25 14:13:01,388 INFO [train.py:1198] (2/4) Epoch 42, batch 3450, loss[loss=0.2018, ctc_loss=0.13, cr_loss=0.3589, over 16958.00 frames. ], tot_loss[loss=0.1908, ctc_loss=0.1229, cr_loss=0.34, over 3353266.58 frames. ], batch size: 58, lr: 2.81e-03, grad_scale: 16.0 2024-09-25 14:13:38,344 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=7.60 vs. limit=15.0 2024-09-25 14:13:40,928 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=761632.6666666666, ans=0.2 2024-09-25 14:13:47,192 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=761679.3333333334, ans=0.04949747468305833 2024-09-25 14:14:19,996 INFO [train.py:1198] (2/4) Epoch 42, batch 3500, loss[loss=0.1678, ctc_loss=0.1068, cr_loss=0.3053, over 17256.00 frames. ], tot_loss[loss=0.19, ctc_loss=0.1223, cr_loss=0.3386, over 3360327.72 frames. ], batch size: 42, lr: 2.81e-03, grad_scale: 16.0 2024-09-25 14:14:29,503 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=761772.6666666666, ans=0.1 2024-09-25 14:14:34,308 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=761819.3333333334, ans=0.125 2024-09-25 14:14:35,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=761819.3333333334, ans=0.2 2024-09-25 14:14:35,940 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=761819.3333333334, ans=0.1 2024-09-25 14:14:37,159 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.158e+02 1.281e+02 1.384e+02 1.487e+02 1.767e+02, threshold=2.768e+02, percent-clipped=0.0 2024-09-25 14:14:43,920 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=761819.3333333334, ans=0.0 2024-09-25 14:15:25,129 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=9.53 vs. limit=15.0 2024-09-25 14:15:31,876 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.22 vs. limit=15.0 2024-09-25 14:15:32,644 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=761959.3333333334, ans=0.125 2024-09-25 14:15:37,327 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=761959.3333333334, ans=0.0 2024-09-25 14:15:38,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=762006.0, ans=0.125 2024-09-25 14:15:40,264 INFO [train.py:1198] (2/4) Epoch 42, batch 3550, loss[loss=0.2302, ctc_loss=0.1502, cr_loss=0.4001, over 17033.00 frames. ], tot_loss[loss=0.1897, ctc_loss=0.122, cr_loss=0.3384, over 3363880.51 frames. ], batch size: 52, lr: 2.81e-03, grad_scale: 16.0 2024-09-25 14:15:42,294 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=762006.0, ans=0.0 2024-09-25 14:15:46,794 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 14:15:56,015 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=762052.6666666666, ans=0.2 2024-09-25 14:16:05,486 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=762052.6666666666, ans=0.0 2024-09-25 14:16:16,825 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=9.94 vs. limit=15.0 2024-09-25 14:16:21,377 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=762099.3333333334, ans=0.09899494936611666 2024-09-25 14:16:56,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=762239.3333333334, ans=0.1 2024-09-25 14:16:56,942 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=762239.3333333334, ans=0.0 2024-09-25 14:16:58,303 INFO [train.py:1198] (2/4) Epoch 42, batch 3600, loss[loss=0.2026, ctc_loss=0.1332, cr_loss=0.3468, over 17236.00 frames. ], tot_loss[loss=0.1896, ctc_loss=0.1219, cr_loss=0.3383, over 3369588.44 frames. ], batch size: 50, lr: 2.81e-03, grad_scale: 32.0 2024-09-25 14:17:15,186 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.282e+02 1.377e+02 1.493e+02 1.761e+02, threshold=2.753e+02, percent-clipped=0.0 2024-09-25 14:17:19,624 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.74 vs. limit=22.5 2024-09-25 14:17:26,981 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=762286.0, ans=0.1 2024-09-25 14:17:54,851 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=762379.3333333334, ans=0.125 2024-09-25 14:18:01,016 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=762426.0, ans=0.0 2024-09-25 14:18:12,643 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=20.39 vs. limit=22.5 2024-09-25 14:18:13,682 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=762426.0, ans=0.125 2024-09-25 14:18:17,993 INFO [train.py:1198] (2/4) Epoch 42, batch 3650, loss[loss=0.1921, ctc_loss=0.1204, cr_loss=0.3587, over 17010.00 frames. ], tot_loss[loss=0.1898, ctc_loss=0.122, cr_loss=0.3391, over 3368177.93 frames. ], batch size: 51, lr: 2.81e-03, grad_scale: 32.0 2024-09-25 14:18:22,991 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=762472.6666666666, ans=0.125 2024-09-25 14:18:48,114 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=762566.0, ans=0.1 2024-09-25 14:18:49,827 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=762566.0, ans=0.1 2024-09-25 14:19:07,925 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=762612.6666666666, ans=0.2 2024-09-25 14:19:29,233 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=762659.3333333334, ans=0.0 2024-09-25 14:19:34,239 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.29 vs. limit=10.0 2024-09-25 14:19:38,301 INFO [train.py:1198] (2/4) Epoch 42, batch 3700, loss[loss=0.1859, ctc_loss=0.1181, cr_loss=0.3389, over 17024.00 frames. ], tot_loss[loss=0.1907, ctc_loss=0.1225, cr_loss=0.3407, over 3368988.21 frames. ], batch size: 44, lr: 2.81e-03, grad_scale: 32.0 2024-09-25 14:19:51,277 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=762706.0, ans=0.1 2024-09-25 14:19:55,540 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.181e+02 1.273e+02 1.387e+02 1.507e+02 1.911e+02, threshold=2.774e+02, percent-clipped=0.0 2024-09-25 14:20:23,313 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=762799.3333333334, ans=0.0 2024-09-25 14:20:35,807 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 14:20:48,526 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=762892.6666666666, ans=0.125 2024-09-25 14:20:57,613 INFO [train.py:1198] (2/4) Epoch 42, batch 3750, loss[loss=0.1577, ctc_loss=0.09872, cr_loss=0.2947, over 17041.00 frames. ], tot_loss[loss=0.1912, ctc_loss=0.123, cr_loss=0.341, over 3359801.39 frames. ], batch size: 39, lr: 2.81e-03, grad_scale: 32.0 2024-09-25 14:20:59,379 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=762939.3333333334, ans=0.125 2024-09-25 14:21:14,965 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=762986.0, ans=0.0 2024-09-25 14:22:08,072 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=763126.0, ans=0.125 2024-09-25 14:22:08,138 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=763126.0, ans=0.125 2024-09-25 14:22:15,973 INFO [train.py:1198] (2/4) Epoch 42, batch 3800, loss[loss=0.2075, ctc_loss=0.1326, cr_loss=0.3745, over 16953.00 frames. ], tot_loss[loss=0.1926, ctc_loss=0.124, cr_loss=0.343, over 3330069.51 frames. ], batch size: 58, lr: 2.81e-03, grad_scale: 32.0 2024-09-25 14:22:33,279 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.315e+02 1.408e+02 1.487e+02 2.777e+02, threshold=2.816e+02, percent-clipped=1.0 2024-09-25 14:22:37,215 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=5.19 vs. limit=15.0 2024-09-25 14:22:41,367 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=763219.3333333334, ans=0.2 2024-09-25 14:22:47,554 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=763266.0, ans=0.125 2024-09-25 14:23:05,072 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 14:23:17,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.max_abs, batch_count=763359.3333333334, ans=10.0 2024-09-25 14:23:17,885 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=9.94 vs. limit=15.0 2024-09-25 14:23:34,529 INFO [train.py:1198] (2/4) Epoch 42, batch 3850, loss[loss=0.1592, ctc_loss=0.1005, cr_loss=0.2937, over 16704.00 frames. ], tot_loss[loss=0.1909, ctc_loss=0.1227, cr_loss=0.341, over 3296307.58 frames. ], batch size: 37, lr: 2.81e-03, grad_scale: 32.0 2024-09-25 14:23:51,090 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn1.whiten, num_groups=1, num_channels=192, metric=8.03 vs. limit=22.5 2024-09-25 14:23:59,966 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=763452.6666666666, ans=0.0 2024-09-25 14:24:08,987 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=763499.3333333334, ans=0.1 2024-09-25 14:24:26,218 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=5.60 vs. limit=10.0 2024-09-25 14:24:33,864 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=763546.0, ans=0.125 2024-09-25 14:25:36,253 INFO [train.py:1198] (2/4) Epoch 43, batch 0, loss[loss=0.2054, ctc_loss=0.1388, cr_loss=0.3331, over 11614.00 frames. ], tot_loss[loss=0.2054, ctc_loss=0.1388, cr_loss=0.3331, over 11614.00 frames. ], batch size: 123, lr: 2.77e-03, grad_scale: 32.0 2024-09-25 14:25:36,253 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 14:25:44,713 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.3.encoder.layers.3.self_attn_weights, attn_weights_entropy = tensor([2.4802, 2.7173, 3.1132, 2.9111, 3.5021, 3.3297, 3.4497, 2.5546], device='cuda:2') 2024-09-25 14:25:51,487 INFO [train.py:1230] (2/4) Epoch 43, validation: loss=0.03486, ctc_loss=0.03486, cr_loss=1.051e-14, over 944034.00 frames. 2024-09-25 14:25:51,487 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 14:25:51,748 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=763620.6666666666, ans=0.2 2024-09-25 14:26:02,678 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=763620.6666666666, ans=0.0 2024-09-25 14:26:15,237 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.147e+02 1.334e+02 1.498e+02 1.673e+02 2.107e+02, threshold=2.995e+02, percent-clipped=0.0 2024-09-25 14:26:19,232 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=9.45 vs. limit=15.0 2024-09-25 14:26:39,337 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=763760.6666666666, ans=0.125 2024-09-25 14:27:10,606 INFO [train.py:1198] (2/4) Epoch 43, batch 50, loss[loss=0.2345, ctc_loss=0.1547, cr_loss=0.3992, over 16526.00 frames. ], tot_loss[loss=0.1911, ctc_loss=0.1229, cr_loss=0.3414, over 748734.80 frames. ], batch size: 66, lr: 2.77e-03, grad_scale: 32.0 2024-09-25 14:27:28,635 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=763900.6666666666, ans=0.2 2024-09-25 14:27:51,224 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.04 vs. limit=6.0 2024-09-25 14:28:27,299 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=764040.6666666666, ans=0.5 2024-09-25 14:28:30,178 INFO [train.py:1198] (2/4) Epoch 43, batch 100, loss[loss=0.2155, ctc_loss=0.139, cr_loss=0.3826, over 17041.00 frames. ], tot_loss[loss=0.1903, ctc_loss=0.1222, cr_loss=0.3403, over 1323130.94 frames. ], batch size: 52, lr: 2.77e-03, grad_scale: 32.0 2024-09-25 14:28:37,063 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=10.55 vs. limit=15.0 2024-09-25 14:28:38,434 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=764087.3333333334, ans=0.0 2024-09-25 14:28:41,638 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=764087.3333333334, ans=0.125 2024-09-25 14:28:54,052 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.158e+02 1.307e+02 1.401e+02 1.473e+02 2.012e+02, threshold=2.802e+02, percent-clipped=0.0 2024-09-25 14:28:57,572 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.max_abs, batch_count=764134.0, ans=10.0 2024-09-25 14:29:13,428 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=764180.6666666666, ans=0.125 2024-09-25 14:29:22,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=764227.3333333334, ans=0.0 2024-09-25 14:29:54,510 INFO [train.py:1198] (2/4) Epoch 43, batch 150, loss[loss=0.2112, ctc_loss=0.1353, cr_loss=0.3796, over 17205.00 frames. ], tot_loss[loss=0.1899, ctc_loss=0.1219, cr_loss=0.34, over 1775600.48 frames. ], batch size: 55, lr: 2.77e-03, grad_scale: 32.0 2024-09-25 14:30:17,223 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=764367.3333333334, ans=0.1 2024-09-25 14:30:29,006 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=764414.0, ans=0.2 2024-09-25 14:31:02,690 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=764507.3333333334, ans=0.125 2024-09-25 14:31:07,539 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=764507.3333333334, ans=0.125 2024-09-25 14:31:09,835 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.whiten, num_groups=1, num_channels=192, metric=4.02 vs. limit=12.0 2024-09-25 14:31:19,818 INFO [train.py:1198] (2/4) Epoch 43, batch 200, loss[loss=0.164, ctc_loss=0.1054, cr_loss=0.293, over 17254.00 frames. ], tot_loss[loss=0.1883, ctc_loss=0.1206, cr_loss=0.3385, over 2129247.22 frames. ], batch size: 44, lr: 2.77e-03, grad_scale: 32.0 2024-09-25 14:31:28,118 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=764554.0, ans=0.0 2024-09-25 14:31:32,808 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=764554.0, ans=0.0 2024-09-25 14:31:41,277 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=7.19 vs. limit=15.0 2024-09-25 14:31:43,639 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.159e+02 1.287e+02 1.384e+02 1.479e+02 1.740e+02, threshold=2.769e+02, percent-clipped=0.0 2024-09-25 14:31:45,927 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.24 vs. limit=10.0 2024-09-25 14:31:47,670 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=5.87 vs. limit=12.0 2024-09-25 14:32:01,768 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=2.54 vs. limit=6.0 2024-09-25 14:32:30,251 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=764740.6666666666, ans=0.125 2024-09-25 14:32:33,331 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=764740.6666666666, ans=0.125 2024-09-25 14:32:36,358 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=764740.6666666666, ans=0.0 2024-09-25 14:32:39,303 INFO [train.py:1198] (2/4) Epoch 43, batch 250, loss[loss=0.2025, ctc_loss=0.1306, cr_loss=0.3598, over 16976.00 frames. ], tot_loss[loss=0.1871, ctc_loss=0.1197, cr_loss=0.3368, over 2406359.71 frames. ], batch size: 56, lr: 2.77e-03, grad_scale: 32.0 2024-09-25 14:32:41,183 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=764787.3333333334, ans=0.125 2024-09-25 14:33:12,017 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 14:33:17,784 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=5.87 vs. limit=15.0 2024-09-25 14:33:31,352 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=764927.3333333334, ans=0.125 2024-09-25 14:33:40,641 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=764927.3333333334, ans=0.125 2024-09-25 14:33:43,956 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=764974.0, ans=0.125 2024-09-25 14:33:59,384 INFO [train.py:1198] (2/4) Epoch 43, batch 300, loss[loss=0.2233, ctc_loss=0.1419, cr_loss=0.4068, over 16986.00 frames. ], tot_loss[loss=0.1871, ctc_loss=0.1198, cr_loss=0.3364, over 2617983.96 frames. ], batch size: 53, lr: 2.77e-03, grad_scale: 32.0 2024-09-25 14:34:18,323 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=765067.3333333334, ans=0.0 2024-09-25 14:34:25,900 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.161e+02 1.299e+02 1.370e+02 1.451e+02 2.001e+02, threshold=2.741e+02, percent-clipped=0.0 2024-09-25 14:35:24,625 INFO [train.py:1198] (2/4) Epoch 43, batch 350, loss[loss=0.1727, ctc_loss=0.1064, cr_loss=0.3318, over 17218.00 frames. ], tot_loss[loss=0.1881, ctc_loss=0.1207, cr_loss=0.3371, over 2784829.96 frames. ], batch size: 47, lr: 2.77e-03, grad_scale: 32.0 2024-09-25 14:35:45,208 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 14:36:38,015 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=765440.6666666666, ans=0.025 2024-09-25 14:36:48,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=765440.6666666666, ans=0.0 2024-09-25 14:36:51,771 INFO [train.py:1198] (2/4) Epoch 43, batch 400, loss[loss=0.2077, ctc_loss=0.1358, cr_loss=0.3594, over 17031.00 frames. ], tot_loss[loss=0.1873, ctc_loss=0.1201, cr_loss=0.3362, over 2922133.66 frames. ], batch size: 52, lr: 2.77e-03, grad_scale: 32.0 2024-09-25 14:36:53,517 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=765487.3333333334, ans=0.035 2024-09-25 14:37:15,438 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.144e+02 1.295e+02 1.341e+02 1.449e+02 2.336e+02, threshold=2.681e+02, percent-clipped=0.0 2024-09-25 14:37:31,935 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=765580.6666666666, ans=0.125 2024-09-25 14:37:33,492 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=765580.6666666666, ans=0.0 2024-09-25 14:38:05,309 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=765674.0, ans=0.2 2024-09-25 14:38:09,888 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=765720.6666666666, ans=0.125 2024-09-25 14:38:11,327 INFO [train.py:1198] (2/4) Epoch 43, batch 450, loss[loss=0.1867, ctc_loss=0.1207, cr_loss=0.3303, over 17158.00 frames. ], tot_loss[loss=0.187, ctc_loss=0.1199, cr_loss=0.3357, over 3019713.28 frames. ], batch size: 45, lr: 2.77e-03, grad_scale: 32.0 2024-09-25 14:38:21,905 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn1.whiten, num_groups=1, num_channels=192, metric=9.20 vs. limit=22.5 2024-09-25 14:38:24,680 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.27 vs. limit=10.0 2024-09-25 14:38:35,901 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.72 vs. limit=22.5 2024-09-25 14:38:57,783 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=765860.6666666666, ans=0.2 2024-09-25 14:39:24,246 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=765907.3333333334, ans=0.125 2024-09-25 14:39:24,373 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=765907.3333333334, ans=0.07 2024-09-25 14:39:33,565 INFO [train.py:1198] (2/4) Epoch 43, batch 500, loss[loss=0.2216, ctc_loss=0.1429, cr_loss=0.3934, over 16591.00 frames. ], tot_loss[loss=0.1883, ctc_loss=0.1207, cr_loss=0.3376, over 3093195.25 frames. ], batch size: 66, lr: 2.77e-03, grad_scale: 32.0 2024-09-25 14:39:46,301 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=765954.0, ans=0.125 2024-09-25 14:40:00,058 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.174e+02 1.354e+02 1.449e+02 1.544e+02 3.228e+02, threshold=2.898e+02, percent-clipped=1.0 2024-09-25 14:40:06,585 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=766047.3333333334, ans=0.95 2024-09-25 14:40:59,430 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=766187.3333333334, ans=0.125 2024-09-25 14:41:00,787 INFO [train.py:1198] (2/4) Epoch 43, batch 550, loss[loss=0.1997, ctc_loss=0.1295, cr_loss=0.3513, over 17212.00 frames. ], tot_loss[loss=0.1887, ctc_loss=0.121, cr_loss=0.3384, over 3158078.58 frames. ], batch size: 47, lr: 2.77e-03, grad_scale: 32.0 2024-09-25 14:41:07,376 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=766187.3333333334, ans=0.025 2024-09-25 14:41:09,016 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=766187.3333333334, ans=0.0 2024-09-25 14:42:14,420 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=766374.0, ans=0.1 2024-09-25 14:42:20,463 INFO [train.py:1198] (2/4) Epoch 43, batch 600, loss[loss=0.2096, ctc_loss=0.1351, cr_loss=0.3724, over 17043.00 frames. ], tot_loss[loss=0.1878, ctc_loss=0.1204, cr_loss=0.3374, over 3216534.64 frames. ], batch size: 52, lr: 2.77e-03, grad_scale: 32.0 2024-09-25 14:42:33,537 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=766420.6666666666, ans=0.0 2024-09-25 14:42:44,283 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.128e+02 1.302e+02 1.423e+02 1.505e+02 3.409e+02, threshold=2.845e+02, percent-clipped=1.0 2024-09-25 14:43:18,431 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.15 vs. limit=10.0 2024-09-25 14:43:20,625 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.31 vs. limit=15.0 2024-09-25 14:43:40,675 INFO [train.py:1198] (2/4) Epoch 43, batch 650, loss[loss=0.2069, ctc_loss=0.1347, cr_loss=0.3613, over 17296.00 frames. ], tot_loss[loss=0.188, ctc_loss=0.1204, cr_loss=0.3378, over 3252633.61 frames. ], batch size: 46, lr: 2.77e-03, grad_scale: 32.0 2024-09-25 14:44:00,449 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.min_positive, batch_count=766700.6666666666, ans=0.05 2024-09-25 14:44:09,890 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=766700.6666666666, ans=0.1 2024-09-25 14:44:32,110 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=766794.0, ans=0.0 2024-09-25 14:44:38,234 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=766794.0, ans=0.0 2024-09-25 14:44:38,367 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=766794.0, ans=0.0 2024-09-25 14:44:47,447 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=766794.0, ans=0.0 2024-09-25 14:44:50,821 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=766840.6666666666, ans=0.1 2024-09-25 14:45:06,410 INFO [train.py:1198] (2/4) Epoch 43, batch 700, loss[loss=0.1996, ctc_loss=0.127, cr_loss=0.3632, over 17221.00 frames. ], tot_loss[loss=0.1886, ctc_loss=0.121, cr_loss=0.3381, over 3270103.51 frames. ], batch size: 50, lr: 2.77e-03, grad_scale: 32.0 2024-09-25 14:45:12,987 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.min_abs, batch_count=766887.3333333334, ans=0.5 2024-09-25 14:45:30,412 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.142e+02 1.308e+02 1.372e+02 1.490e+02 1.932e+02, threshold=2.743e+02, percent-clipped=0.0 2024-09-25 14:45:34,019 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=766934.0, ans=0.1 2024-09-25 14:45:54,310 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=766980.6666666666, ans=0.0 2024-09-25 14:46:17,156 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=11.88 vs. limit=15.0 2024-09-25 14:46:29,240 INFO [train.py:1198] (2/4) Epoch 43, batch 750, loss[loss=0.1985, ctc_loss=0.1347, cr_loss=0.319, over 11756.00 frames. ], tot_loss[loss=0.189, ctc_loss=0.1213, cr_loss=0.3382, over 3277050.14 frames. ], batch size: 123, lr: 2.77e-03, grad_scale: 32.0 2024-09-25 14:46:31,057 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=767120.6666666666, ans=0.125 2024-09-25 14:46:48,569 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=767167.3333333334, ans=0.0 2024-09-25 14:46:48,570 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=767167.3333333334, ans=0.125 2024-09-25 14:46:50,335 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=767167.3333333334, ans=0.0 2024-09-25 14:47:16,360 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=9.67 vs. limit=15.0 2024-09-25 14:47:35,654 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=9.64 vs. limit=15.0 2024-09-25 14:47:49,078 INFO [train.py:1198] (2/4) Epoch 43, batch 800, loss[loss=0.2125, ctc_loss=0.1383, cr_loss=0.3711, over 16517.00 frames. ], tot_loss[loss=0.1892, ctc_loss=0.1215, cr_loss=0.3386, over 3297011.88 frames. ], batch size: 66, lr: 2.77e-03, grad_scale: 32.0 2024-09-25 14:47:54,030 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=767354.0, ans=0.025 2024-09-25 14:48:12,578 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.117e+02 1.303e+02 1.387e+02 1.475e+02 2.331e+02, threshold=2.774e+02, percent-clipped=0.0 2024-09-25 14:48:27,308 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.max_abs, batch_count=767447.3333333334, ans=10.0 2024-09-25 14:48:38,465 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 14:49:00,823 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=767540.6666666666, ans=0.125 2024-09-25 14:49:08,214 INFO [train.py:1198] (2/4) Epoch 43, batch 850, loss[loss=0.2339, ctc_loss=0.1536, cr_loss=0.4014, over 15077.00 frames. ], tot_loss[loss=0.1887, ctc_loss=0.1212, cr_loss=0.3376, over 3316779.30 frames. ], batch size: 89, lr: 2.77e-03, grad_scale: 32.0 2024-09-25 14:49:29,007 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=767634.0, ans=0.2 2024-09-25 14:49:37,050 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=767634.0, ans=0.0 2024-09-25 14:49:49,140 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=767680.6666666666, ans=0.1 2024-09-25 14:49:50,609 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=767680.6666666666, ans=0.0 2024-09-25 14:49:50,730 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=767680.6666666666, ans=0.0 2024-09-25 14:49:58,716 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=767680.6666666666, ans=0.0 2024-09-25 14:49:58,752 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=767680.6666666666, ans=0.125 2024-09-25 14:50:02,496 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.46 vs. limit=15.0 2024-09-25 14:50:33,650 INFO [train.py:1198] (2/4) Epoch 43, batch 900, loss[loss=0.2081, ctc_loss=0.1354, cr_loss=0.3634, over 17009.00 frames. ], tot_loss[loss=0.188, ctc_loss=0.1206, cr_loss=0.3367, over 3332666.33 frames. ], batch size: 53, lr: 2.77e-03, grad_scale: 16.0 2024-09-25 14:50:40,328 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=767820.6666666666, ans=0.0 2024-09-25 14:50:56,635 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=767867.3333333334, ans=0.025 2024-09-25 14:50:58,178 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=767867.3333333334, ans=0.125 2024-09-25 14:51:01,358 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=767867.3333333334, ans=0.025 2024-09-25 14:51:04,267 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.180e+02 1.318e+02 1.401e+02 1.527e+02 2.167e+02, threshold=2.803e+02, percent-clipped=0.0 2024-09-25 14:51:06,055 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=767867.3333333334, ans=0.0 2024-09-25 14:51:14,232 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=767914.0, ans=0.125 2024-09-25 14:51:26,837 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=767960.6666666666, ans=0.1 2024-09-25 14:51:37,957 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=767960.6666666666, ans=0.125 2024-09-25 14:51:58,509 INFO [train.py:1198] (2/4) Epoch 43, batch 950, loss[loss=0.1896, ctc_loss=0.1211, cr_loss=0.3426, over 17080.00 frames. ], tot_loss[loss=0.1881, ctc_loss=0.1207, cr_loss=0.3368, over 3341973.51 frames. ], batch size: 43, lr: 2.77e-03, grad_scale: 16.0 2024-09-25 14:52:03,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=768054.0, ans=0.125 2024-09-25 14:52:45,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=768194.0, ans=0.125 2024-09-25 14:53:18,816 INFO [train.py:1198] (2/4) Epoch 43, batch 1000, loss[loss=0.1848, ctc_loss=0.118, cr_loss=0.334, over 17263.00 frames. ], tot_loss[loss=0.1882, ctc_loss=0.1209, cr_loss=0.3367, over 3344766.51 frames. ], batch size: 44, lr: 2.77e-03, grad_scale: 16.0 2024-09-25 14:53:44,360 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.294e+02 1.416e+02 1.489e+02 2.363e+02, threshold=2.832e+02, percent-clipped=0.0 2024-09-25 14:53:49,512 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=768380.6666666666, ans=0.0 2024-09-25 14:53:52,625 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=768380.6666666666, ans=0.2 2024-09-25 14:54:00,740 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=768380.6666666666, ans=0.05 2024-09-25 14:54:44,014 INFO [train.py:1198] (2/4) Epoch 43, batch 1050, loss[loss=0.2176, ctc_loss=0.1398, cr_loss=0.3892, over 17099.00 frames. ], tot_loss[loss=0.1894, ctc_loss=0.1217, cr_loss=0.3385, over 3341915.85 frames. ], batch size: 49, lr: 2.77e-03, grad_scale: 16.0 2024-09-25 14:54:50,749 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=768520.6666666666, ans=0.125 2024-09-25 14:55:06,620 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=768567.3333333334, ans=0.125 2024-09-25 14:55:47,621 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=768660.6666666666, ans=0.125 2024-09-25 14:55:58,867 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=10.18 vs. limit=22.5 2024-09-25 14:56:01,234 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=768707.3333333334, ans=0.0 2024-09-25 14:56:07,535 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=768754.0, ans=0.125 2024-09-25 14:56:08,918 INFO [train.py:1198] (2/4) Epoch 43, batch 1100, loss[loss=0.1523, ctc_loss=0.09694, cr_loss=0.2767, over 17200.00 frames. ], tot_loss[loss=0.1889, ctc_loss=0.1213, cr_loss=0.3382, over 3350213.64 frames. ], batch size: 41, lr: 2.77e-03, grad_scale: 16.0 2024-09-25 14:56:34,424 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.288e+02 1.354e+02 1.479e+02 2.223e+02, threshold=2.708e+02, percent-clipped=0.0 2024-09-25 14:56:39,618 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=768847.3333333334, ans=0.125 2024-09-25 14:56:43,209 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.17 vs. limit=15.0 2024-09-25 14:56:47,510 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=768847.3333333334, ans=0.1 2024-09-25 14:56:47,607 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=768847.3333333334, ans=0.0 2024-09-25 14:56:52,397 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=768847.3333333334, ans=0.0 2024-09-25 14:57:00,232 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=768894.0, ans=0.2 2024-09-25 14:57:01,609 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=768894.0, ans=0.125 2024-09-25 14:57:14,478 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=768940.6666666666, ans=0.125 2024-09-25 14:57:28,430 INFO [train.py:1198] (2/4) Epoch 43, batch 1150, loss[loss=0.1989, ctc_loss=0.128, cr_loss=0.3546, over 17275.00 frames. ], tot_loss[loss=0.1885, ctc_loss=0.121, cr_loss=0.3375, over 3356274.74 frames. ], batch size: 51, lr: 2.77e-03, grad_scale: 16.0 2024-09-25 14:57:43,169 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=769034.0, ans=0.125 2024-09-25 14:57:47,205 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.12 vs. limit=15.0 2024-09-25 14:57:49,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=769034.0, ans=0.09899494936611666 2024-09-25 14:58:04,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=769080.6666666666, ans=0.0 2024-09-25 14:58:31,674 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=769174.0, ans=0.0 2024-09-25 14:58:33,991 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn2.whiten, num_groups=1, num_channels=192, metric=8.98 vs. limit=22.5 2024-09-25 14:58:47,391 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=769220.6666666666, ans=0.125 2024-09-25 14:58:48,005 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten.whitening_limit, batch_count=769220.6666666666, ans=15.0 2024-09-25 14:58:48,005 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=7.81 vs. limit=15.0 2024-09-25 14:58:48,786 INFO [train.py:1198] (2/4) Epoch 43, batch 1200, loss[loss=0.174, ctc_loss=0.1088, cr_loss=0.3258, over 17191.00 frames. ], tot_loss[loss=0.1881, ctc_loss=0.1208, cr_loss=0.3368, over 3360922.19 frames. ], batch size: 41, lr: 2.76e-03, grad_scale: 32.0 2024-09-25 14:58:56,962 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=769220.6666666666, ans=0.1 2024-09-25 14:59:03,423 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=769267.3333333334, ans=0.07 2024-09-25 14:59:14,325 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.056e+02 1.304e+02 1.385e+02 1.497e+02 1.972e+02, threshold=2.770e+02, percent-clipped=0.0 2024-09-25 14:59:20,274 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=769267.3333333334, ans=0.2 2024-09-25 14:59:42,130 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=769360.6666666666, ans=0.125 2024-09-25 15:00:13,438 INFO [train.py:1198] (2/4) Epoch 43, batch 1250, loss[loss=0.203, ctc_loss=0.1302, cr_loss=0.364, over 16869.00 frames. ], tot_loss[loss=0.1891, ctc_loss=0.1213, cr_loss=0.3386, over 3360813.18 frames. ], batch size: 58, lr: 2.76e-03, grad_scale: 32.0 2024-09-25 15:00:18,730 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer_na.min_abs, batch_count=769454.0, ans=0.02 2024-09-25 15:01:13,225 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=769594.0, ans=0.2 2024-09-25 15:01:38,703 INFO [train.py:1198] (2/4) Epoch 43, batch 1300, loss[loss=0.1929, ctc_loss=0.1247, cr_loss=0.341, over 17305.00 frames. ], tot_loss[loss=0.1894, ctc_loss=0.1215, cr_loss=0.3394, over 3367124.24 frames. ], batch size: 51, lr: 2.76e-03, grad_scale: 32.0 2024-09-25 15:02:04,020 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.165e+02 1.298e+02 1.391e+02 1.472e+02 1.717e+02, threshold=2.781e+02, percent-clipped=0.0 2024-09-25 15:02:09,155 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=769780.6666666666, ans=0.035 2024-09-25 15:02:18,637 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=769780.6666666666, ans=0.025 2024-09-25 15:02:25,655 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=512, metric=11.96 vs. limit=22.5 2024-09-25 15:02:55,611 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=769874.0, ans=0.125 2024-09-25 15:02:58,578 INFO [train.py:1198] (2/4) Epoch 43, batch 1350, loss[loss=0.1595, ctc_loss=0.1028, cr_loss=0.2834, over 16962.00 frames. ], tot_loss[loss=0.189, ctc_loss=0.1213, cr_loss=0.3384, over 3366917.70 frames. ], batch size: 42, lr: 2.76e-03, grad_scale: 32.0 2024-09-25 15:03:03,722 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=769920.6666666666, ans=0.0 2024-09-25 15:03:56,957 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=770060.6666666666, ans=0.125 2024-09-25 15:04:01,041 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.70 vs. limit=5.0 2024-09-25 15:04:21,710 INFO [train.py:1198] (2/4) Epoch 43, batch 1400, loss[loss=0.189, ctc_loss=0.1184, cr_loss=0.3529, over 17309.00 frames. ], tot_loss[loss=0.1893, ctc_loss=0.1215, cr_loss=0.3389, over 3362977.17 frames. ], batch size: 51, lr: 2.76e-03, grad_scale: 32.0 2024-09-25 15:04:27,091 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=770154.0, ans=0.125 2024-09-25 15:04:42,481 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=770200.6666666666, ans=0.125 2024-09-25 15:04:50,140 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.179e+02 1.309e+02 1.380e+02 1.484e+02 2.530e+02, threshold=2.759e+02, percent-clipped=0.0 2024-09-25 15:05:01,615 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=770247.3333333334, ans=0.0 2024-09-25 15:05:24,101 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=770294.0, ans=0.0 2024-09-25 15:05:33,034 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=4.93 vs. limit=15.0 2024-09-25 15:05:46,977 INFO [train.py:1198] (2/4) Epoch 43, batch 1450, loss[loss=0.234, ctc_loss=0.1544, cr_loss=0.3983, over 16389.00 frames. ], tot_loss[loss=0.189, ctc_loss=0.1214, cr_loss=0.3379, over 3357434.49 frames. ], batch size: 66, lr: 2.76e-03, grad_scale: 32.0 2024-09-25 15:06:09,150 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=770434.0, ans=0.2 2024-09-25 15:06:26,715 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=770480.6666666666, ans=0.125 2024-09-25 15:06:41,704 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.17 vs. limit=6.0 2024-09-25 15:06:42,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=770527.3333333334, ans=0.1 2024-09-25 15:06:55,992 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.72 vs. limit=6.0 2024-09-25 15:06:58,730 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=770574.0, ans=0.2 2024-09-25 15:07:03,670 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=770574.0, ans=0.0 2024-09-25 15:07:09,744 INFO [train.py:1198] (2/4) Epoch 43, batch 1500, loss[loss=0.1648, ctc_loss=0.1034, cr_loss=0.307, over 17304.00 frames. ], tot_loss[loss=0.1893, ctc_loss=0.1216, cr_loss=0.3381, over 3356689.42 frames. ], batch size: 46, lr: 2.76e-03, grad_scale: 32.0 2024-09-25 15:07:29,513 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.43 vs. limit=15.0 2024-09-25 15:07:35,088 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.121e+02 1.284e+02 1.350e+02 1.419e+02 1.777e+02, threshold=2.700e+02, percent-clipped=0.0 2024-09-25 15:07:54,636 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=770714.0, ans=0.2 2024-09-25 15:08:08,987 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=770760.6666666666, ans=0.125 2024-09-25 15:08:29,383 INFO [train.py:1198] (2/4) Epoch 43, batch 1550, loss[loss=0.1842, ctc_loss=0.1176, cr_loss=0.3331, over 17355.00 frames. ], tot_loss[loss=0.1885, ctc_loss=0.1212, cr_loss=0.3369, over 3358764.40 frames. ], batch size: 48, lr: 2.76e-03, grad_scale: 32.0 2024-09-25 15:08:45,572 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=770900.6666666666, ans=0.125 2024-09-25 15:09:08,954 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=770947.3333333334, ans=0.125 2024-09-25 15:09:54,492 INFO [train.py:1198] (2/4) Epoch 43, batch 1600, loss[loss=0.1575, ctc_loss=0.09917, cr_loss=0.2915, over 17108.00 frames. ], tot_loss[loss=0.1885, ctc_loss=0.1211, cr_loss=0.3372, over 3351523.82 frames. ], batch size: 40, lr: 2.76e-03, grad_scale: 32.0 2024-09-25 15:10:14,816 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.42 vs. limit=15.0 2024-09-25 15:10:17,211 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=771134.0, ans=0.0 2024-09-25 15:10:18,935 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=771134.0, ans=0.0 2024-09-25 15:10:19,309 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=5.22 vs. limit=15.0 2024-09-25 15:10:20,125 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.278e+02 1.396e+02 1.510e+02 2.224e+02, threshold=2.791e+02, percent-clipped=0.0 2024-09-25 15:10:29,925 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer1.prob, batch_count=771180.6666666666, ans=0.125 2024-09-25 15:10:40,514 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=771180.6666666666, ans=0.1 2024-09-25 15:11:04,146 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=771274.0, ans=0.0 2024-09-25 15:11:17,645 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.65 vs. limit=6.0 2024-09-25 15:11:19,900 INFO [train.py:1198] (2/4) Epoch 43, batch 1650, loss[loss=0.2013, ctc_loss=0.1289, cr_loss=0.3621, over 17265.00 frames. ], tot_loss[loss=0.1879, ctc_loss=0.1205, cr_loss=0.337, over 3358172.64 frames. ], batch size: 44, lr: 2.76e-03, grad_scale: 32.0 2024-09-25 15:11:23,521 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 15:11:38,069 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.91 vs. limit=6.0 2024-09-25 15:11:45,840 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=771367.3333333334, ans=0.0 2024-09-25 15:11:53,804 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 15:11:57,069 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=771414.0, ans=0.2 2024-09-25 15:12:17,527 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.max_abs, batch_count=771460.6666666666, ans=10.0 2024-09-25 15:12:26,881 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=771507.3333333334, ans=0.1 2024-09-25 15:12:31,677 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=771507.3333333334, ans=0.125 2024-09-25 15:12:31,755 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=771507.3333333334, ans=0.1 2024-09-25 15:12:39,450 INFO [train.py:1198] (2/4) Epoch 43, batch 1700, loss[loss=0.1845, ctc_loss=0.1149, cr_loss=0.3477, over 17287.00 frames. ], tot_loss[loss=0.1876, ctc_loss=0.1202, cr_loss=0.3367, over 3365179.94 frames. ], batch size: 46, lr: 2.76e-03, grad_scale: 32.0 2024-09-25 15:13:05,042 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.157e+02 1.300e+02 1.382e+02 1.486e+02 1.912e+02, threshold=2.763e+02, percent-clipped=0.0 2024-09-25 15:13:09,320 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=6.41 vs. limit=15.0 2024-09-25 15:13:11,823 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=771647.3333333334, ans=0.125 2024-09-25 15:13:23,487 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.03 vs. limit=15.0 2024-09-25 15:13:29,612 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=771694.0, ans=0.125 2024-09-25 15:14:00,169 INFO [train.py:1198] (2/4) Epoch 43, batch 1750, loss[loss=0.188, ctc_loss=0.1209, cr_loss=0.336, over 17221.00 frames. ], tot_loss[loss=0.1864, ctc_loss=0.1194, cr_loss=0.3353, over 3376015.15 frames. ], batch size: 50, lr: 2.76e-03, grad_scale: 32.0 2024-09-25 15:14:32,863 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=771834.0, ans=0.1 2024-09-25 15:14:42,396 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=771880.6666666666, ans=0.125 2024-09-25 15:15:13,034 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.29 vs. limit=10.0 2024-09-25 15:15:23,434 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=772020.6666666666, ans=0.0 2024-09-25 15:15:24,845 INFO [train.py:1198] (2/4) Epoch 43, batch 1800, loss[loss=0.165, ctc_loss=0.1076, cr_loss=0.2873, over 17177.00 frames. ], tot_loss[loss=0.1861, ctc_loss=0.1192, cr_loss=0.3345, over 3377669.22 frames. ], batch size: 41, lr: 2.76e-03, grad_scale: 32.0 2024-09-25 15:15:29,947 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=772020.6666666666, ans=0.2 2024-09-25 15:15:43,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=772067.3333333334, ans=0.125 2024-09-25 15:15:55,550 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.211e+02 1.295e+02 1.383e+02 1.459e+02 2.589e+02, threshold=2.767e+02, percent-clipped=0.0 2024-09-25 15:16:06,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=772114.0, ans=0.125 2024-09-25 15:16:08,783 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=772114.0, ans=0.125 2024-09-25 15:16:26,181 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=772160.6666666666, ans=0.1 2024-09-25 15:16:36,489 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=4.06 vs. limit=15.0 2024-09-25 15:16:49,384 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=10.75 vs. limit=22.5 2024-09-25 15:16:49,940 INFO [train.py:1198] (2/4) Epoch 43, batch 1850, loss[loss=0.1636, ctc_loss=0.1049, cr_loss=0.2938, over 17160.00 frames. ], tot_loss[loss=0.1861, ctc_loss=0.1192, cr_loss=0.3344, over 3378231.31 frames. ], batch size: 41, lr: 2.76e-03, grad_scale: 32.0 2024-09-25 15:17:12,717 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=772300.6666666666, ans=0.025 2024-09-25 15:17:14,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=772300.6666666666, ans=0.0 2024-09-25 15:17:15,748 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=772300.6666666666, ans=0.0 2024-09-25 15:17:28,831 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=772347.3333333334, ans=0.125 2024-09-25 15:17:43,325 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=772394.0, ans=0.025 2024-09-25 15:17:51,975 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=9.56 vs. limit=15.0 2024-09-25 15:17:56,306 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=772440.6666666666, ans=0.0 2024-09-25 15:18:10,614 INFO [train.py:1198] (2/4) Epoch 43, batch 1900, loss[loss=0.1819, ctc_loss=0.1165, cr_loss=0.3271, over 17223.00 frames. ], tot_loss[loss=0.1865, ctc_loss=0.1196, cr_loss=0.3345, over 3371634.13 frames. ], batch size: 50, lr: 2.76e-03, grad_scale: 32.0 2024-09-25 15:18:10,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=772487.3333333334, ans=0.5 2024-09-25 15:18:25,692 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.92 vs. limit=15.0 2024-09-25 15:18:28,966 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=3.65 vs. limit=10.0 2024-09-25 15:18:36,191 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.070e+02 1.296e+02 1.386e+02 1.467e+02 1.936e+02, threshold=2.772e+02, percent-clipped=0.0 2024-09-25 15:18:39,870 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=772534.0, ans=0.2 2024-09-25 15:19:12,594 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=772627.3333333334, ans=0.0 2024-09-25 15:19:17,592 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=12.04 vs. limit=15.0 2024-09-25 15:19:26,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=772674.0, ans=0.125 2024-09-25 15:19:35,529 INFO [train.py:1198] (2/4) Epoch 43, batch 1950, loss[loss=0.1727, ctc_loss=0.11, cr_loss=0.3138, over 17092.00 frames. ], tot_loss[loss=0.1869, ctc_loss=0.1199, cr_loss=0.3352, over 3382796.99 frames. ], batch size: 49, lr: 2.76e-03, grad_scale: 32.0 2024-09-25 15:19:37,360 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=772720.6666666666, ans=0.0 2024-09-25 15:19:37,470 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=772720.6666666666, ans=0.125 2024-09-25 15:20:05,119 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.70 vs. limit=15.0 2024-09-25 15:20:14,432 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=772814.0, ans=0.1 2024-09-25 15:20:28,582 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=772860.6666666666, ans=0.125 2024-09-25 15:20:49,002 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=772907.3333333334, ans=0.1 2024-09-25 15:21:00,967 INFO [train.py:1198] (2/4) Epoch 43, batch 2000, loss[loss=0.1479, ctc_loss=0.09183, cr_loss=0.2802, over 17269.00 frames. ], tot_loss[loss=0.1869, ctc_loss=0.1199, cr_loss=0.3351, over 3391226.27 frames. ], batch size: 42, lr: 2.76e-03, grad_scale: 32.0 2024-09-25 15:21:10,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=772954.0, ans=0.125 2024-09-25 15:21:20,371 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=773000.6666666666, ans=0.0 2024-09-25 15:21:22,526 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.10 vs. limit=15.0 2024-09-25 15:21:26,606 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.130e+02 1.269e+02 1.346e+02 1.449e+02 2.025e+02, threshold=2.692e+02, percent-clipped=0.0 2024-09-25 15:22:20,774 INFO [train.py:1198] (2/4) Epoch 43, batch 2050, loss[loss=0.2036, ctc_loss=0.1301, cr_loss=0.3674, over 16735.00 frames. ], tot_loss[loss=0.1881, ctc_loss=0.1207, cr_loss=0.3368, over 3389498.13 frames. ], batch size: 61, lr: 2.76e-03, grad_scale: 32.0 2024-09-25 15:22:48,290 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=773234.0, ans=0.125 2024-09-25 15:22:59,992 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.74 vs. limit=10.0 2024-09-25 15:23:13,151 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module1.whiten, num_groups=1, num_channels=192, metric=9.12 vs. limit=15.0 2024-09-25 15:23:13,908 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=773327.3333333334, ans=0.1 2024-09-25 15:23:15,473 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=773327.3333333334, ans=0.125 2024-09-25 15:23:17,281 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=773327.3333333334, ans=0.125 2024-09-25 15:23:31,680 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=773374.0, ans=0.0 2024-09-25 15:23:40,843 INFO [train.py:1198] (2/4) Epoch 43, batch 2100, loss[loss=0.1861, ctc_loss=0.1201, cr_loss=0.3302, over 17069.00 frames. ], tot_loss[loss=0.1887, ctc_loss=0.1212, cr_loss=0.3373, over 3382344.13 frames. ], batch size: 39, lr: 2.76e-03, grad_scale: 16.0 2024-09-25 15:24:10,504 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.139e+02 1.257e+02 1.347e+02 1.435e+02 1.926e+02, threshold=2.695e+02, percent-clipped=0.0 2024-09-25 15:24:12,449 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=773467.3333333334, ans=0.0 2024-09-25 15:24:40,633 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=773560.6666666666, ans=0.1 2024-09-25 15:24:53,410 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=773607.3333333334, ans=0.125 2024-09-25 15:25:05,722 INFO [train.py:1198] (2/4) Epoch 43, batch 2150, loss[loss=0.1539, ctc_loss=0.09726, cr_loss=0.2833, over 17284.00 frames. ], tot_loss[loss=0.1883, ctc_loss=0.1209, cr_loss=0.3368, over 3374569.14 frames. ], batch size: 46, lr: 2.76e-03, grad_scale: 16.0 2024-09-25 15:25:09,385 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer_ff2.min_abs, batch_count=773654.0, ans=0.1 2024-09-25 15:25:23,710 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.attention_skip_rate, batch_count=773700.6666666666, ans=0.0 2024-09-25 15:25:39,601 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=773747.3333333334, ans=0.125 2024-09-25 15:25:59,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=773794.0, ans=0.0 2024-09-25 15:26:00,491 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=5.05 vs. limit=10.0 2024-09-25 15:26:31,353 INFO [train.py:1198] (2/4) Epoch 43, batch 2200, loss[loss=0.1943, ctc_loss=0.124, cr_loss=0.3514, over 17168.00 frames. ], tot_loss[loss=0.1897, ctc_loss=0.122, cr_loss=0.3386, over 3361730.92 frames. ], batch size: 45, lr: 2.76e-03, grad_scale: 16.0 2024-09-25 15:26:33,337 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 15:26:39,428 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=773887.3333333334, ans=0.0 2024-09-25 15:26:42,648 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=773887.3333333334, ans=0.125 2024-09-25 15:26:47,515 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=773934.0, ans=0.0 2024-09-25 15:26:49,525 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.86 vs. limit=10.0 2024-09-25 15:26:58,329 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.167e+02 1.324e+02 1.413e+02 1.558e+02 2.550e+02, threshold=2.826e+02, percent-clipped=0.0 2024-09-25 15:27:17,815 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 15:27:21,170 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=3.58 vs. limit=6.0 2024-09-25 15:27:33,603 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=774074.0, ans=0.2 2024-09-25 15:27:40,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.max_abs, batch_count=774074.0, ans=10.0 2024-09-25 15:27:50,922 INFO [train.py:1198] (2/4) Epoch 43, batch 2250, loss[loss=0.16, ctc_loss=0.09987, cr_loss=0.3009, over 17031.00 frames. ], tot_loss[loss=0.1888, ctc_loss=0.1213, cr_loss=0.3374, over 3359318.19 frames. ], batch size: 39, lr: 2.76e-03, grad_scale: 16.0 2024-09-25 15:28:03,097 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=11.77 vs. limit=22.5 2024-09-25 15:28:34,491 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=774214.0, ans=0.125 2024-09-25 15:28:37,697 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=774260.6666666666, ans=0.0 2024-09-25 15:28:57,092 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 15:29:05,859 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=774307.3333333334, ans=0.125 2024-09-25 15:29:10,552 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=774307.3333333334, ans=0.125 2024-09-25 15:29:13,564 INFO [train.py:1198] (2/4) Epoch 43, batch 2300, loss[loss=0.2075, ctc_loss=0.1349, cr_loss=0.3633, over 16914.00 frames. ], tot_loss[loss=0.1896, ctc_loss=0.1218, cr_loss=0.3387, over 3356321.08 frames. ], batch size: 58, lr: 2.76e-03, grad_scale: 16.0 2024-09-25 15:29:21,797 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=774354.0, ans=0.125 2024-09-25 15:29:25,085 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 15:29:29,371 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=774354.0, ans=0.0 2024-09-25 15:29:33,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=774400.6666666666, ans=0.0 2024-09-25 15:29:40,084 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=774400.6666666666, ans=0.1 2024-09-25 15:29:40,122 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=774400.6666666666, ans=0.125 2024-09-25 15:29:42,935 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.176e+02 1.301e+02 1.372e+02 1.481e+02 2.158e+02, threshold=2.745e+02, percent-clipped=0.0 2024-09-25 15:29:43,763 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=8.49 vs. limit=15.0 2024-09-25 15:29:56,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=774447.3333333334, ans=0.0 2024-09-25 15:30:10,238 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=774494.0, ans=0.025 2024-09-25 15:30:35,539 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.45 vs. limit=10.0 2024-09-25 15:30:37,779 INFO [train.py:1198] (2/4) Epoch 43, batch 2350, loss[loss=0.1478, ctc_loss=0.09068, cr_loss=0.2854, over 16284.00 frames. ], tot_loss[loss=0.1894, ctc_loss=0.1215, cr_loss=0.3392, over 3356779.50 frames. ], batch size: 36, lr: 2.76e-03, grad_scale: 16.0 2024-09-25 15:30:39,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.min_abs, batch_count=774587.3333333334, ans=0.5 2024-09-25 15:30:50,863 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=19.79 vs. limit=22.5 2024-09-25 15:31:49,586 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=774774.0, ans=0.0 2024-09-25 15:32:00,368 INFO [train.py:1198] (2/4) Epoch 43, batch 2400, loss[loss=0.1502, ctc_loss=0.0945, cr_loss=0.2784, over 17031.00 frames. ], tot_loss[loss=0.1888, ctc_loss=0.1211, cr_loss=0.3383, over 3351548.91 frames. ], batch size: 39, lr: 2.75e-03, grad_scale: 32.0 2024-09-25 15:32:00,675 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=774820.6666666666, ans=0.025 2024-09-25 15:32:11,798 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=774820.6666666666, ans=0.125 2024-09-25 15:32:16,681 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=774867.3333333334, ans=0.125 2024-09-25 15:32:24,954 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=774867.3333333334, ans=0.125 2024-09-25 15:32:27,693 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.130e+02 1.292e+02 1.363e+02 1.432e+02 2.291e+02, threshold=2.726e+02, percent-clipped=0.0 2024-09-25 15:32:44,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=774914.0, ans=0.125 2024-09-25 15:32:45,756 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=774914.0, ans=0.0 2024-09-25 15:33:05,238 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=775007.3333333334, ans=0.0 2024-09-25 15:33:05,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=775007.3333333334, ans=0.0 2024-09-25 15:33:11,575 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=775007.3333333334, ans=0.0 2024-09-25 15:33:20,860 INFO [train.py:1198] (2/4) Epoch 43, batch 2450, loss[loss=0.1918, ctc_loss=0.1221, cr_loss=0.3487, over 17100.00 frames. ], tot_loss[loss=0.1888, ctc_loss=0.1211, cr_loss=0.3383, over 3356866.64 frames. ], batch size: 43, lr: 2.75e-03, grad_scale: 32.0 2024-09-25 15:33:30,695 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=775054.0, ans=0.125 2024-09-25 15:33:30,817 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.min_positive, batch_count=775054.0, ans=0.025 2024-09-25 15:33:44,919 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=775100.6666666666, ans=0.0 2024-09-25 15:34:04,836 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=775147.3333333334, ans=0.0 2024-09-25 15:34:34,505 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=775240.6666666666, ans=0.0 2024-09-25 15:34:37,817 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=775240.6666666666, ans=0.125 2024-09-25 15:34:42,606 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=775240.6666666666, ans=0.125 2024-09-25 15:34:45,575 INFO [train.py:1198] (2/4) Epoch 43, batch 2500, loss[loss=0.1937, ctc_loss=0.1256, cr_loss=0.3403, over 16996.00 frames. ], tot_loss[loss=0.1894, ctc_loss=0.1215, cr_loss=0.3392, over 3362327.08 frames. ], batch size: 51, lr: 2.75e-03, grad_scale: 32.0 2024-09-25 15:34:50,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=775287.3333333334, ans=0.125 2024-09-25 15:34:59,344 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=4.85 vs. limit=12.0 2024-09-25 15:35:05,250 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=775334.0, ans=0.035 2024-09-25 15:35:13,017 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.056e+02 1.284e+02 1.374e+02 1.483e+02 2.994e+02, threshold=2.747e+02, percent-clipped=1.0 2024-09-25 15:35:15,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.min_positive, batch_count=775334.0, ans=0.025 2024-09-25 15:35:24,780 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=775380.6666666666, ans=0.0 2024-09-25 15:35:35,975 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=13.12 vs. limit=22.5 2024-09-25 15:36:10,320 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=775520.6666666666, ans=0.025 2024-09-25 15:36:11,635 INFO [train.py:1198] (2/4) Epoch 43, batch 2550, loss[loss=0.195, ctc_loss=0.1207, cr_loss=0.3715, over 17008.00 frames. ], tot_loss[loss=0.1894, ctc_loss=0.1215, cr_loss=0.3396, over 3356808.07 frames. ], batch size: 52, lr: 2.75e-03, grad_scale: 32.0 2024-09-25 15:36:11,913 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=775520.6666666666, ans=0.125 2024-09-25 15:36:20,441 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=8.36 vs. limit=15.0 2024-09-25 15:36:36,145 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=775567.3333333334, ans=0.0 2024-09-25 15:36:36,698 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=16.40 vs. limit=22.5 2024-09-25 15:36:37,643 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=775567.3333333334, ans=0.125 2024-09-25 15:36:40,069 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=10.46 vs. limit=15.0 2024-09-25 15:37:03,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=775660.6666666666, ans=0.1 2024-09-25 15:37:08,585 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=775660.6666666666, ans=0.0 2024-09-25 15:37:16,937 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=11.00 vs. limit=15.0 2024-09-25 15:37:22,952 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 15:37:32,112 INFO [train.py:1198] (2/4) Epoch 43, batch 2600, loss[loss=0.2138, ctc_loss=0.1404, cr_loss=0.3668, over 17264.00 frames. ], tot_loss[loss=0.19, ctc_loss=0.1219, cr_loss=0.3404, over 3355008.94 frames. ], batch size: 55, lr: 2.75e-03, grad_scale: 32.0 2024-09-25 15:37:43,499 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=775754.0, ans=0.2 2024-09-25 15:37:57,680 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=775800.6666666666, ans=0.1 2024-09-25 15:37:58,960 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.142e+02 1.271e+02 1.387e+02 1.465e+02 2.023e+02, threshold=2.774e+02, percent-clipped=0.0 2024-09-25 15:38:12,154 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=775847.3333333334, ans=0.125 2024-09-25 15:38:30,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=775894.0, ans=0.125 2024-09-25 15:38:44,809 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.46 vs. limit=15.0 2024-09-25 15:38:51,970 INFO [train.py:1198] (2/4) Epoch 43, batch 2650, loss[loss=0.1975, ctc_loss=0.1277, cr_loss=0.3492, over 16679.00 frames. ], tot_loss[loss=0.1903, ctc_loss=0.1222, cr_loss=0.3406, over 3354083.04 frames. ], batch size: 61, lr: 2.75e-03, grad_scale: 32.0 2024-09-25 15:38:52,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=775987.3333333334, ans=0.0 2024-09-25 15:38:53,799 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=775987.3333333334, ans=0.125 2024-09-25 15:39:06,157 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=775987.3333333334, ans=0.07 2024-09-25 15:39:13,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=776034.0, ans=0.05 2024-09-25 15:39:26,272 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=776034.0, ans=0.125 2024-09-25 15:39:42,215 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=776080.6666666666, ans=0.0 2024-09-25 15:39:45,956 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=8.81 vs. limit=15.0 2024-09-25 15:39:53,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=776127.3333333334, ans=0.0 2024-09-25 15:39:58,370 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=14.95 vs. limit=15.0 2024-09-25 15:40:05,996 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=776174.0, ans=0.125 2024-09-25 15:40:16,626 INFO [train.py:1198] (2/4) Epoch 43, batch 2700, loss[loss=0.2, ctc_loss=0.1275, cr_loss=0.3623, over 16798.00 frames. ], tot_loss[loss=0.1898, ctc_loss=0.1218, cr_loss=0.3399, over 3346463.24 frames. ], batch size: 61, lr: 2.75e-03, grad_scale: 32.0 2024-09-25 15:40:27,578 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=776220.6666666666, ans=0.2 2024-09-25 15:40:49,149 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.311e+02 1.386e+02 1.489e+02 1.705e+02, threshold=2.772e+02, percent-clipped=0.0 2024-09-25 15:41:18,600 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.98 vs. limit=15.0 2024-09-25 15:41:23,462 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.64 vs. limit=15.0 2024-09-25 15:41:24,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=776407.3333333334, ans=0.0 2024-09-25 15:41:41,505 INFO [train.py:1198] (2/4) Epoch 43, batch 2750, loss[loss=0.1952, ctc_loss=0.1232, cr_loss=0.3603, over 17310.00 frames. ], tot_loss[loss=0.1896, ctc_loss=0.1217, cr_loss=0.3394, over 3350518.68 frames. ], batch size: 49, lr: 2.75e-03, grad_scale: 32.0 2024-09-25 15:41:44,996 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=776454.0, ans=0.125 2024-09-25 15:41:59,338 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=776500.6666666666, ans=0.1 2024-09-25 15:42:09,791 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=11.39 vs. limit=22.5 2024-09-25 15:42:09,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys.whitening_limit, batch_count=776500.6666666666, ans=6.0 2024-09-25 15:42:51,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=776640.6666666666, ans=0.125 2024-09-25 15:42:54,903 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=9.80 vs. limit=22.5 2024-09-25 15:43:02,155 INFO [train.py:1198] (2/4) Epoch 43, batch 2800, loss[loss=0.1999, ctc_loss=0.1283, cr_loss=0.3579, over 17292.00 frames. ], tot_loss[loss=0.1891, ctc_loss=0.1213, cr_loss=0.3387, over 3349009.78 frames. ], batch size: 51, lr: 2.75e-03, grad_scale: 32.0 2024-09-25 15:43:10,328 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=776687.3333333334, ans=0.025 2024-09-25 15:43:27,993 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=776734.0, ans=0.125 2024-09-25 15:43:29,176 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.146e+02 1.317e+02 1.406e+02 1.487e+02 1.823e+02, threshold=2.813e+02, percent-clipped=0.0 2024-09-25 15:43:40,743 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=776780.6666666666, ans=0.125 2024-09-25 15:43:44,062 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=776780.6666666666, ans=0.125 2024-09-25 15:43:51,945 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=776827.3333333334, ans=0.125 2024-09-25 15:44:27,271 INFO [train.py:1198] (2/4) Epoch 43, batch 2850, loss[loss=0.1484, ctc_loss=0.092, cr_loss=0.2821, over 17099.00 frames. ], tot_loss[loss=0.1885, ctc_loss=0.121, cr_loss=0.3377, over 3358004.53 frames. ], batch size: 40, lr: 2.75e-03, grad_scale: 32.0 2024-09-25 15:44:30,682 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=776920.6666666666, ans=0.125 2024-09-25 15:44:45,573 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.77 vs. limit=10.0 2024-09-25 15:44:50,126 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=776967.3333333334, ans=0.125 2024-09-25 15:44:54,806 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=776967.3333333334, ans=0.125 2024-09-25 15:45:01,326 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=7.42 vs. limit=15.0 2024-09-25 15:45:38,600 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=777107.3333333334, ans=0.0 2024-09-25 15:45:52,191 INFO [train.py:1198] (2/4) Epoch 43, batch 2900, loss[loss=0.1778, ctc_loss=0.111, cr_loss=0.334, over 17081.00 frames. ], tot_loss[loss=0.1884, ctc_loss=0.1208, cr_loss=0.3378, over 3366349.43 frames. ], batch size: 43, lr: 2.75e-03, grad_scale: 32.0 2024-09-25 15:46:19,411 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.188e+02 1.318e+02 1.398e+02 1.451e+02 2.249e+02, threshold=2.795e+02, percent-clipped=0.0 2024-09-25 15:46:42,308 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=777294.0, ans=0.125 2024-09-25 15:47:03,307 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=777340.6666666666, ans=0.0 2024-09-25 15:47:12,741 INFO [train.py:1198] (2/4) Epoch 43, batch 2950, loss[loss=0.2239, ctc_loss=0.1451, cr_loss=0.3945, over 17021.00 frames. ], tot_loss[loss=0.1881, ctc_loss=0.1206, cr_loss=0.3377, over 3375307.66 frames. ], batch size: 52, lr: 2.75e-03, grad_scale: 32.0 2024-09-25 15:47:37,229 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=777434.0, ans=0.0 2024-09-25 15:47:39,369 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=5.37 vs. limit=12.0 2024-09-25 15:47:42,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=777434.0, ans=0.1 2024-09-25 15:47:56,490 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=777480.6666666666, ans=0.0 2024-09-25 15:48:33,057 INFO [train.py:1198] (2/4) Epoch 43, batch 3000, loss[loss=0.1991, ctc_loss=0.1302, cr_loss=0.3443, over 14933.00 frames. ], tot_loss[loss=0.1874, ctc_loss=0.1201, cr_loss=0.3366, over 3370330.55 frames. ], batch size: 89, lr: 2.75e-03, grad_scale: 32.0 2024-09-25 15:48:33,058 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 15:48:50,092 INFO [train.py:1230] (2/4) Epoch 43, validation: loss=0.03539, ctc_loss=0.03539, cr_loss=1.015e-14, over 944034.00 frames. 2024-09-25 15:48:50,093 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 15:48:55,021 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=777620.6666666666, ans=0.2 2024-09-25 15:49:16,945 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.046e+02 1.300e+02 1.373e+02 1.452e+02 1.992e+02, threshold=2.746e+02, percent-clipped=0.0 2024-09-25 15:49:39,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=777760.6666666666, ans=0.125 2024-09-25 15:49:57,620 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.31 vs. limit=15.0 2024-09-25 15:50:10,664 INFO [train.py:1198] (2/4) Epoch 43, batch 3050, loss[loss=0.1861, ctc_loss=0.1199, cr_loss=0.3311, over 17108.00 frames. ], tot_loss[loss=0.1874, ctc_loss=0.1201, cr_loss=0.3364, over 3367913.08 frames. ], batch size: 49, lr: 2.75e-03, grad_scale: 16.0 2024-09-25 15:50:35,729 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=777900.6666666666, ans=0.125 2024-09-25 15:51:06,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=777994.0, ans=0.05 2024-09-25 15:51:18,081 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=3.91 vs. limit=15.0 2024-09-25 15:51:19,260 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=778040.6666666666, ans=0.1 2024-09-25 15:51:28,333 INFO [train.py:1198] (2/4) Epoch 43, batch 3100, loss[loss=0.1958, ctc_loss=0.1272, cr_loss=0.3429, over 16964.00 frames. ], tot_loss[loss=0.1878, ctc_loss=0.1204, cr_loss=0.3366, over 3360924.27 frames. ], batch size: 58, lr: 2.75e-03, grad_scale: 16.0 2024-09-25 15:51:28,672 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=778087.3333333334, ans=0.125 2024-09-25 15:51:58,668 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.089e+02 1.269e+02 1.353e+02 1.457e+02 2.895e+02, threshold=2.707e+02, percent-clipped=1.0 2024-09-25 15:52:02,340 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=778180.6666666666, ans=0.125 2024-09-25 15:52:42,662 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.59 vs. limit=15.0 2024-09-25 15:52:51,552 INFO [train.py:1198] (2/4) Epoch 43, batch 3150, loss[loss=0.2006, ctc_loss=0.1299, cr_loss=0.3534, over 17141.00 frames. ], tot_loss[loss=0.1872, ctc_loss=0.1201, cr_loss=0.3356, over 3368021.97 frames. ], batch size: 48, lr: 2.75e-03, grad_scale: 16.0 2024-09-25 15:53:23,572 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=778414.0, ans=0.025 2024-09-25 15:53:32,945 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=778414.0, ans=0.1 2024-09-25 15:53:45,520 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=778460.6666666666, ans=0.125 2024-09-25 15:53:56,543 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=778507.3333333334, ans=0.2 2024-09-25 15:54:10,165 INFO [train.py:1198] (2/4) Epoch 43, batch 3200, loss[loss=0.21, ctc_loss=0.1371, cr_loss=0.3648, over 17091.00 frames. ], tot_loss[loss=0.1883, ctc_loss=0.1208, cr_loss=0.3372, over 3360942.50 frames. ], batch size: 49, lr: 2.75e-03, grad_scale: 32.0 2024-09-25 15:54:21,912 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=7.41 vs. limit=15.0 2024-09-25 15:54:38,168 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.191e+02 1.301e+02 1.369e+02 1.499e+02 2.028e+02, threshold=2.738e+02, percent-clipped=0.0 2024-09-25 15:55:18,019 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=778740.6666666666, ans=0.125 2024-09-25 15:55:28,972 INFO [train.py:1198] (2/4) Epoch 43, batch 3250, loss[loss=0.2181, ctc_loss=0.1395, cr_loss=0.3928, over 15902.00 frames. ], tot_loss[loss=0.1874, ctc_loss=0.1202, cr_loss=0.3356, over 3365028.77 frames. ], batch size: 74, lr: 2.75e-03, grad_scale: 32.0 2024-09-25 15:55:35,671 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=778787.3333333334, ans=0.1 2024-09-25 15:55:52,611 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=778834.0, ans=0.125 2024-09-25 15:55:54,355 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=9.30 vs. limit=15.0 2024-09-25 15:56:16,184 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=778927.3333333334, ans=0.1 2024-09-25 15:56:46,768 INFO [train.py:1198] (2/4) Epoch 43, batch 3300, loss[loss=0.1528, ctc_loss=0.09891, cr_loss=0.2695, over 17046.00 frames. ], tot_loss[loss=0.1881, ctc_loss=0.1207, cr_loss=0.3368, over 3364705.91 frames. ], batch size: 39, lr: 2.75e-03, grad_scale: 32.0 2024-09-25 15:56:54,808 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=779020.6666666666, ans=0.0 2024-09-25 15:57:01,159 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=779067.3333333334, ans=0.2 2024-09-25 15:57:07,399 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=779067.3333333334, ans=0.125 2024-09-25 15:57:10,316 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=779067.3333333334, ans=0.125 2024-09-25 15:57:14,795 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.168e+02 1.314e+02 1.383e+02 1.471e+02 2.059e+02, threshold=2.766e+02, percent-clipped=0.0 2024-09-25 15:57:33,790 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=779160.6666666666, ans=0.1 2024-09-25 15:57:50,659 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=779207.3333333334, ans=0.0 2024-09-25 15:58:04,605 INFO [train.py:1198] (2/4) Epoch 43, batch 3350, loss[loss=0.1904, ctc_loss=0.1229, cr_loss=0.338, over 17005.00 frames. ], tot_loss[loss=0.1875, ctc_loss=0.1203, cr_loss=0.336, over 3355526.73 frames. ], batch size: 51, lr: 2.75e-03, grad_scale: 32.0 2024-09-25 15:58:15,901 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=779254.0, ans=0.125 2024-09-25 15:58:17,306 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=779254.0, ans=0.0 2024-09-25 15:58:24,982 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=779300.6666666666, ans=0.0 2024-09-25 15:58:35,007 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=779347.3333333334, ans=0.0 2024-09-25 15:58:39,598 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=779347.3333333334, ans=0.125 2024-09-25 15:58:44,369 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=779347.3333333334, ans=0.125 2024-09-25 15:58:44,478 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=779347.3333333334, ans=0.2 2024-09-25 15:58:58,546 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=779394.0, ans=0.125 2024-09-25 15:59:23,219 INFO [train.py:1198] (2/4) Epoch 43, batch 3400, loss[loss=0.1667, ctc_loss=0.1034, cr_loss=0.3162, over 16996.00 frames. ], tot_loss[loss=0.1877, ctc_loss=0.1203, cr_loss=0.3372, over 3361660.44 frames. ], batch size: 39, lr: 2.75e-03, grad_scale: 32.0 2024-09-25 15:59:25,091 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=779487.3333333334, ans=0.125 2024-09-25 15:59:55,330 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.136e+02 1.291e+02 1.360e+02 1.462e+02 4.308e+02, threshold=2.720e+02, percent-clipped=1.0 2024-09-25 16:00:42,192 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=779674.0, ans=0.125 2024-09-25 16:00:45,033 INFO [train.py:1198] (2/4) Epoch 43, batch 3450, loss[loss=0.1868, ctc_loss=0.1184, cr_loss=0.3421, over 16949.00 frames. ], tot_loss[loss=0.1875, ctc_loss=0.1202, cr_loss=0.3369, over 3358149.20 frames. ], batch size: 42, lr: 2.75e-03, grad_scale: 32.0 2024-09-25 16:00:53,096 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=779720.6666666666, ans=0.1 2024-09-25 16:01:13,619 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=779767.3333333334, ans=0.05 2024-09-25 16:01:24,470 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=779814.0, ans=0.025 2024-09-25 16:01:34,357 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.48 vs. limit=12.0 2024-09-25 16:01:39,952 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=779860.6666666666, ans=0.0 2024-09-25 16:01:41,835 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.74 vs. limit=15.0 2024-09-25 16:01:49,505 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=779907.3333333334, ans=0.125 2024-09-25 16:02:00,645 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 16:02:03,506 INFO [train.py:1198] (2/4) Epoch 43, batch 3500, loss[loss=0.1816, ctc_loss=0.1171, cr_loss=0.3223, over 17312.00 frames. ], tot_loss[loss=0.188, ctc_loss=0.1205, cr_loss=0.3375, over 3370595.73 frames. ], batch size: 49, lr: 2.75e-03, grad_scale: 16.0 2024-09-25 16:02:14,595 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=779954.0, ans=0.015 2024-09-25 16:02:24,716 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=780000.6666666666, ans=0.125 2024-09-25 16:02:32,978 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=8.35 vs. limit=15.0 2024-09-25 16:02:35,356 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.093e+02 1.311e+02 1.389e+02 1.472e+02 1.940e+02, threshold=2.779e+02, percent-clipped=0.0 2024-09-25 16:02:56,338 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=780094.0, ans=0.125 2024-09-25 16:03:13,786 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=780140.6666666666, ans=0.125 2024-09-25 16:03:18,998 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=15.23 vs. limit=22.5 2024-09-25 16:03:24,741 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=780187.3333333334, ans=0.2 2024-09-25 16:03:26,020 INFO [train.py:1198] (2/4) Epoch 43, batch 3550, loss[loss=0.2116, ctc_loss=0.137, cr_loss=0.373, over 16668.00 frames. ], tot_loss[loss=0.1888, ctc_loss=0.1211, cr_loss=0.3387, over 3366203.96 frames. ], batch size: 61, lr: 2.75e-03, grad_scale: 16.0 2024-09-25 16:03:55,116 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=780234.0, ans=0.0 2024-09-25 16:04:04,599 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=780280.6666666666, ans=0.125 2024-09-25 16:04:09,481 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=780280.6666666666, ans=0.1 2024-09-25 16:04:44,772 INFO [train.py:1198] (2/4) Epoch 43, batch 3600, loss[loss=0.1644, ctc_loss=0.1048, cr_loss=0.2982, over 17096.00 frames. ], tot_loss[loss=0.1878, ctc_loss=0.1204, cr_loss=0.3369, over 3358317.89 frames. ], batch size: 43, lr: 2.74e-03, grad_scale: 32.0 2024-09-25 16:04:54,731 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=13.08 vs. limit=22.5 2024-09-25 16:05:05,421 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=780467.3333333334, ans=0.0 2024-09-25 16:05:05,513 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=780467.3333333334, ans=0.2 2024-09-25 16:05:06,151 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1.whitening_limit, batch_count=780467.3333333334, ans=10.0 2024-09-25 16:05:07,010 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=780467.3333333334, ans=0.0 2024-09-25 16:05:14,401 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.305e+02 1.407e+02 1.507e+02 1.948e+02, threshold=2.813e+02, percent-clipped=0.0 2024-09-25 16:05:14,656 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=780514.0, ans=0.1 2024-09-25 16:05:36,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=780560.6666666666, ans=0.035 2024-09-25 16:05:47,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=780607.3333333334, ans=0.125 2024-09-25 16:06:03,263 INFO [train.py:1198] (2/4) Epoch 43, batch 3650, loss[loss=0.1866, ctc_loss=0.1194, cr_loss=0.3361, over 17011.00 frames. ], tot_loss[loss=0.1886, ctc_loss=0.121, cr_loss=0.3381, over 3357729.63 frames. ], batch size: 44, lr: 2.74e-03, grad_scale: 32.0 2024-09-25 16:06:05,180 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=780654.0, ans=0.1 2024-09-25 16:06:08,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=780654.0, ans=0.125 2024-09-25 16:06:38,233 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=780747.3333333334, ans=0.1 2024-09-25 16:06:49,934 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=4.87 vs. limit=15.0 2024-09-25 16:06:52,254 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=780794.0, ans=0.025 2024-09-25 16:06:53,742 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=780794.0, ans=0.1 2024-09-25 16:07:20,645 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.max_positive, batch_count=780887.3333333334, ans=0.95 2024-09-25 16:07:21,989 INFO [train.py:1198] (2/4) Epoch 43, batch 3700, loss[loss=0.1702, ctc_loss=0.1074, cr_loss=0.314, over 17273.00 frames. ], tot_loss[loss=0.1878, ctc_loss=0.1204, cr_loss=0.3372, over 3368614.58 frames. ], batch size: 42, lr: 2.74e-03, grad_scale: 32.0 2024-09-25 16:07:49,709 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=780934.0, ans=0.04949747468305833 2024-09-25 16:07:52,634 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.155e+02 1.295e+02 1.384e+02 1.500e+02 2.354e+02, threshold=2.769e+02, percent-clipped=0.0 2024-09-25 16:08:00,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=780980.6666666666, ans=0.2 2024-09-25 16:08:13,005 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=781027.3333333334, ans=0.025 2024-09-25 16:08:24,216 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=781074.0, ans=0.1 2024-09-25 16:08:24,933 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=12.53 vs. limit=22.5 2024-09-25 16:08:30,641 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=781074.0, ans=0.0 2024-09-25 16:08:41,139 INFO [train.py:1198] (2/4) Epoch 43, batch 3750, loss[loss=0.1708, ctc_loss=0.1049, cr_loss=0.3296, over 16809.00 frames. ], tot_loss[loss=0.1887, ctc_loss=0.121, cr_loss=0.3386, over 3359015.01 frames. ], batch size: 37, lr: 2.74e-03, grad_scale: 16.0 2024-09-25 16:08:43,023 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=781120.6666666666, ans=0.1 2024-09-25 16:09:18,530 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=781214.0, ans=0.0 2024-09-25 16:09:36,254 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=781260.6666666666, ans=0.0 2024-09-25 16:10:01,216 INFO [train.py:1198] (2/4) Epoch 43, batch 3800, loss[loss=0.2252, ctc_loss=0.1504, cr_loss=0.3741, over 11179.00 frames. ], tot_loss[loss=0.1893, ctc_loss=0.1215, cr_loss=0.3391, over 3320126.66 frames. ], batch size: 124, lr: 2.74e-03, grad_scale: 16.0 2024-09-25 16:10:14,016 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=781354.0, ans=0.125 2024-09-25 16:10:17,390 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.min_positive, batch_count=781400.6666666666, ans=0.025 2024-09-25 16:10:26,661 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=781400.6666666666, ans=0.125 2024-09-25 16:10:26,738 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=781400.6666666666, ans=0.1 2024-09-25 16:10:32,811 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.147e+02 1.355e+02 1.451e+02 1.569e+02 1.796e+02, threshold=2.903e+02, percent-clipped=0.0 2024-09-25 16:10:40,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=781447.3333333334, ans=0.125 2024-09-25 16:11:14,382 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=781540.6666666666, ans=0.1 2024-09-25 16:11:20,436 INFO [train.py:1198] (2/4) Epoch 43, batch 3850, loss[loss=0.2059, ctc_loss=0.1362, cr_loss=0.3484, over 11750.00 frames. ], tot_loss[loss=0.1924, ctc_loss=0.1241, cr_loss=0.3415, over 3239047.17 frames. ], batch size: 124, lr: 2.74e-03, grad_scale: 16.0 2024-09-25 16:11:29,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=781587.3333333334, ans=0.125 2024-09-25 16:11:53,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=781680.6666666666, ans=0.2 2024-09-25 16:12:24,038 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=11.35 vs. limit=22.5 2024-09-25 16:13:18,296 INFO [train.py:1198] (2/4) Epoch 44, batch 0, loss[loss=0.1781, ctc_loss=0.1114, cr_loss=0.3336, over 17250.00 frames. ], tot_loss[loss=0.1781, ctc_loss=0.1114, cr_loss=0.3336, over 17250.00 frames. ], batch size: 44, lr: 2.71e-03, grad_scale: 32.0 2024-09-25 16:13:18,296 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 16:13:33,580 INFO [train.py:1230] (2/4) Epoch 44, validation: loss=0.03507, ctc_loss=0.03507, cr_loss=1.053e-14, over 944034.00 frames. 2024-09-25 16:13:33,581 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 16:13:54,991 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=781848.6666666666, ans=0.0 2024-09-25 16:13:54,995 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=781848.6666666666, ans=10.0 2024-09-25 16:14:14,723 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 1.434e+02 1.591e+02 1.721e+02 2.734e+02, threshold=3.183e+02, percent-clipped=0.0 2024-09-25 16:14:15,075 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=781895.3333333334, ans=0.125 2024-09-25 16:14:21,364 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=781895.3333333334, ans=0.125 2024-09-25 16:14:31,391 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=13.38 vs. limit=22.5 2024-09-25 16:14:39,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=781988.6666666666, ans=0.125 2024-09-25 16:14:59,046 INFO [train.py:1198] (2/4) Epoch 44, batch 50, loss[loss=0.1878, ctc_loss=0.1194, cr_loss=0.3419, over 17026.00 frames. ], tot_loss[loss=0.1879, ctc_loss=0.1204, cr_loss=0.3376, over 753006.29 frames. ], batch size: 52, lr: 2.71e-03, grad_scale: 32.0 2024-09-25 16:15:00,936 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=782035.3333333334, ans=0.1 2024-09-25 16:15:16,016 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=7.67 vs. limit=15.0 2024-09-25 16:15:28,047 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=782082.0, ans=0.07 2024-09-25 16:15:36,192 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=782128.6666666666, ans=0.125 2024-09-25 16:15:38,756 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.31 vs. limit=15.0 2024-09-25 16:15:41,048 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=782128.6666666666, ans=0.125 2024-09-25 16:15:53,779 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=782175.3333333334, ans=0.125 2024-09-25 16:16:02,342 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=6.28 vs. limit=15.0 2024-09-25 16:16:03,186 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=782222.0, ans=0.125 2024-09-25 16:16:18,758 INFO [train.py:1198] (2/4) Epoch 44, batch 100, loss[loss=0.194, ctc_loss=0.1261, cr_loss=0.3393, over 17367.00 frames. ], tot_loss[loss=0.1884, ctc_loss=0.1207, cr_loss=0.3382, over 1337200.95 frames. ], batch size: 48, lr: 2.71e-03, grad_scale: 32.0 2024-09-25 16:16:19,094 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=782268.6666666666, ans=0.2 2024-09-25 16:16:30,925 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten.whitening_limit, batch_count=782268.6666666666, ans=15.0 2024-09-25 16:16:33,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=782315.3333333334, ans=0.0 2024-09-25 16:16:41,695 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=782315.3333333334, ans=0.04949747468305833 2024-09-25 16:16:57,191 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=782362.0, ans=0.2 2024-09-25 16:17:00,119 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.102e+02 1.309e+02 1.396e+02 1.536e+02 2.062e+02, threshold=2.792e+02, percent-clipped=0.0 2024-09-25 16:17:00,497 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=782362.0, ans=0.125 2024-09-25 16:17:14,464 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=782408.6666666666, ans=0.125 2024-09-25 16:17:41,586 INFO [train.py:1198] (2/4) Epoch 44, batch 150, loss[loss=0.1872, ctc_loss=0.1217, cr_loss=0.3275, over 16532.00 frames. ], tot_loss[loss=0.1895, ctc_loss=0.1216, cr_loss=0.3394, over 1783229.89 frames. ], batch size: 66, lr: 2.71e-03, grad_scale: 32.0 2024-09-25 16:17:58,424 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.52 vs. limit=15.0 2024-09-25 16:18:32,650 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=782642.0, ans=0.125 2024-09-25 16:18:48,540 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=782688.6666666666, ans=0.025 2024-09-25 16:19:07,046 INFO [train.py:1198] (2/4) Epoch 44, batch 200, loss[loss=0.2047, ctc_loss=0.1302, cr_loss=0.3725, over 17260.00 frames. ], tot_loss[loss=0.1897, ctc_loss=0.1218, cr_loss=0.3396, over 2121789.68 frames. ], batch size: 46, lr: 2.71e-03, grad_scale: 32.0 2024-09-25 16:19:20,304 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=782735.3333333334, ans=0.125 2024-09-25 16:19:25,840 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.15 vs. limit=15.0 2024-09-25 16:19:48,492 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.333e+02 1.408e+02 1.534e+02 2.430e+02, threshold=2.816e+02, percent-clipped=0.0 2024-09-25 16:20:14,958 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=782922.0, ans=0.05 2024-09-25 16:20:30,830 INFO [train.py:1198] (2/4) Epoch 44, batch 250, loss[loss=0.1907, ctc_loss=0.1216, cr_loss=0.3456, over 17233.00 frames. ], tot_loss[loss=0.1895, ctc_loss=0.1216, cr_loss=0.3395, over 2404514.60 frames. ], batch size: 50, lr: 2.71e-03, grad_scale: 32.0 2024-09-25 16:20:41,120 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2.whitening_limit, batch_count=782968.6666666666, ans=15.0 2024-09-25 16:20:46,859 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=783015.3333333334, ans=0.125 2024-09-25 16:21:03,041 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=783062.0, ans=0.125 2024-09-25 16:21:06,308 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=783062.0, ans=0.2 2024-09-25 16:21:54,065 INFO [train.py:1198] (2/4) Epoch 44, batch 300, loss[loss=0.1641, ctc_loss=0.1038, cr_loss=0.3017, over 17073.00 frames. ], tot_loss[loss=0.1896, ctc_loss=0.1216, cr_loss=0.3398, over 2612207.91 frames. ], batch size: 43, lr: 2.71e-03, grad_scale: 32.0 2024-09-25 16:21:54,464 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=783202.0, ans=0.125 2024-09-25 16:21:59,239 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=783202.0, ans=0.125 2024-09-25 16:22:10,461 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=783248.6666666666, ans=0.0 2024-09-25 16:22:32,113 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.188e+02 1.326e+02 1.428e+02 1.549e+02 2.699e+02, threshold=2.856e+02, percent-clipped=0.0 2024-09-25 16:23:16,538 INFO [train.py:1198] (2/4) Epoch 44, batch 350, loss[loss=0.2312, ctc_loss=0.1491, cr_loss=0.4103, over 16511.00 frames. ], tot_loss[loss=0.1899, ctc_loss=0.1218, cr_loss=0.3403, over 2771355.94 frames. ], batch size: 66, lr: 2.71e-03, grad_scale: 32.0 2024-09-25 16:23:50,146 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=783528.6666666666, ans=0.125 2024-09-25 16:24:42,412 INFO [train.py:1198] (2/4) Epoch 44, batch 400, loss[loss=0.2014, ctc_loss=0.1302, cr_loss=0.3559, over 17006.00 frames. ], tot_loss[loss=0.1916, ctc_loss=0.1232, cr_loss=0.3423, over 2891797.30 frames. ], batch size: 51, lr: 2.71e-03, grad_scale: 32.0 2024-09-25 16:24:55,884 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=9.51 vs. limit=15.0 2024-09-25 16:24:57,752 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=8.12 vs. limit=15.0 2024-09-25 16:25:11,685 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=783715.3333333334, ans=0.0 2024-09-25 16:25:16,693 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 16:25:20,975 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.141e+02 1.320e+02 1.420e+02 1.550e+02 2.069e+02, threshold=2.840e+02, percent-clipped=0.0 2024-09-25 16:25:26,132 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=783762.0, ans=0.125 2024-09-25 16:25:30,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=783808.6666666666, ans=0.2 2024-09-25 16:25:38,168 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=4.29 vs. limit=15.0 2024-09-25 16:26:02,566 INFO [train.py:1198] (2/4) Epoch 44, batch 450, loss[loss=0.2, ctc_loss=0.1271, cr_loss=0.3643, over 17152.00 frames. ], tot_loss[loss=0.1903, ctc_loss=0.1221, cr_loss=0.3408, over 3005190.40 frames. ], batch size: 45, lr: 2.71e-03, grad_scale: 32.0 2024-09-25 16:26:02,946 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=783902.0, ans=0.1 2024-09-25 16:26:07,592 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=783902.0, ans=0.2 2024-09-25 16:26:09,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=783902.0, ans=0.125 2024-09-25 16:26:10,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=783902.0, ans=0.0 2024-09-25 16:26:54,013 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=4.83 vs. limit=10.0 2024-09-25 16:27:28,174 INFO [train.py:1198] (2/4) Epoch 44, batch 500, loss[loss=0.2198, ctc_loss=0.1422, cr_loss=0.3878, over 16446.00 frames. ], tot_loss[loss=0.189, ctc_loss=0.1212, cr_loss=0.3389, over 3084687.85 frames. ], batch size: 66, lr: 2.71e-03, grad_scale: 16.0 2024-09-25 16:27:52,221 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=784182.0, ans=0.1 2024-09-25 16:28:10,662 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.152e+02 1.312e+02 1.366e+02 1.455e+02 1.775e+02, threshold=2.732e+02, percent-clipped=0.0 2024-09-25 16:28:50,233 INFO [train.py:1198] (2/4) Epoch 44, batch 550, loss[loss=0.1818, ctc_loss=0.1139, cr_loss=0.3396, over 17016.00 frames. ], tot_loss[loss=0.1909, ctc_loss=0.1226, cr_loss=0.3418, over 3149515.44 frames. ], batch size: 44, lr: 2.71e-03, grad_scale: 16.0 2024-09-25 16:29:09,028 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=784415.3333333334, ans=0.125 2024-09-25 16:29:20,213 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=784415.3333333334, ans=0.0 2024-09-25 16:29:38,862 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=784462.0, ans=0.125 2024-09-25 16:29:46,547 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=784508.6666666666, ans=0.125 2024-09-25 16:30:15,130 INFO [train.py:1198] (2/4) Epoch 44, batch 600, loss[loss=0.2092, ctc_loss=0.1355, cr_loss=0.3689, over 16522.00 frames. ], tot_loss[loss=0.1897, ctc_loss=0.1217, cr_loss=0.34, over 3206238.62 frames. ], batch size: 66, lr: 2.71e-03, grad_scale: 16.0 2024-09-25 16:30:23,522 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=784602.0, ans=0.125 2024-09-25 16:30:55,328 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.161e+02 1.308e+02 1.386e+02 1.485e+02 2.110e+02, threshold=2.772e+02, percent-clipped=0.0 2024-09-25 16:31:03,531 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer1.prob, batch_count=784742.0, ans=0.125 2024-09-25 16:31:06,882 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=784742.0, ans=0.125 2024-09-25 16:31:28,658 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=10.12 vs. limit=15.0 2024-09-25 16:31:29,612 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 16:31:35,711 INFO [train.py:1198] (2/4) Epoch 44, batch 650, loss[loss=0.1928, ctc_loss=0.1223, cr_loss=0.3529, over 16875.00 frames. ], tot_loss[loss=0.1893, ctc_loss=0.1213, cr_loss=0.3397, over 3242508.71 frames. ], batch size: 58, lr: 2.71e-03, grad_scale: 16.0 2024-09-25 16:31:39,183 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 16:31:49,954 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=784835.3333333334, ans=0.1 2024-09-25 16:31:59,333 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=784882.0, ans=0.0 2024-09-25 16:32:29,375 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=12.09 vs. limit=15.0 2024-09-25 16:32:40,227 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=784975.3333333334, ans=0.0 2024-09-25 16:32:51,395 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=785022.0, ans=0.025 2024-09-25 16:32:53,066 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 16:32:59,116 INFO [train.py:1198] (2/4) Epoch 44, batch 700, loss[loss=0.163, ctc_loss=0.1004, cr_loss=0.3128, over 15820.00 frames. ], tot_loss[loss=0.1897, ctc_loss=0.1218, cr_loss=0.3396, over 3256960.61 frames. ], batch size: 35, lr: 2.70e-03, grad_scale: 16.0 2024-09-25 16:33:04,152 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=785068.6666666666, ans=0.1 2024-09-25 16:33:11,597 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=785068.6666666666, ans=0.0 2024-09-25 16:33:35,461 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=785162.0, ans=0.2 2024-09-25 16:33:41,733 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.292e+02 1.388e+02 1.458e+02 2.267e+02, threshold=2.776e+02, percent-clipped=0.0 2024-09-25 16:33:56,452 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=785208.6666666666, ans=0.125 2024-09-25 16:34:17,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=785255.3333333334, ans=0.025 2024-09-25 16:34:24,043 INFO [train.py:1198] (2/4) Epoch 44, batch 750, loss[loss=0.2038, ctc_loss=0.1323, cr_loss=0.357, over 17025.00 frames. ], tot_loss[loss=0.1896, ctc_loss=0.1217, cr_loss=0.3392, over 3279978.61 frames. ], batch size: 56, lr: 2.70e-03, grad_scale: 16.0 2024-09-25 16:34:24,455 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=785302.0, ans=0.125 2024-09-25 16:35:10,481 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.74 vs. limit=6.0 2024-09-25 16:35:42,144 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=785488.6666666666, ans=0.5 2024-09-25 16:35:43,714 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=785488.6666666666, ans=0.125 2024-09-25 16:35:46,667 INFO [train.py:1198] (2/4) Epoch 44, batch 800, loss[loss=0.2111, ctc_loss=0.1374, cr_loss=0.3682, over 17051.00 frames. ], tot_loss[loss=0.19, ctc_loss=0.122, cr_loss=0.34, over 3295778.49 frames. ], batch size: 52, lr: 2.70e-03, grad_scale: 32.0 2024-09-25 16:35:49,091 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.41 vs. limit=12.0 2024-09-25 16:35:52,310 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=3.98 vs. limit=15.0 2024-09-25 16:36:00,070 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=785535.3333333334, ans=0.1 2024-09-25 16:36:17,131 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=785628.6666666666, ans=0.0 2024-09-25 16:36:17,275 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=785628.6666666666, ans=0.125 2024-09-25 16:36:26,438 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.158e+02 1.304e+02 1.360e+02 1.506e+02 2.198e+02, threshold=2.721e+02, percent-clipped=0.0 2024-09-25 16:36:41,408 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=785675.3333333334, ans=0.0 2024-09-25 16:36:42,897 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=785675.3333333334, ans=0.1 2024-09-25 16:36:54,985 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.max_abs, batch_count=785722.0, ans=10.0 2024-09-25 16:36:57,107 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.69 vs. limit=15.0 2024-09-25 16:36:59,627 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=785722.0, ans=0.0 2024-09-25 16:37:04,529 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=785722.0, ans=0.125 2024-09-25 16:37:09,076 INFO [train.py:1198] (2/4) Epoch 44, batch 850, loss[loss=0.1735, ctc_loss=0.1077, cr_loss=0.3293, over 17013.00 frames. ], tot_loss[loss=0.1903, ctc_loss=0.1222, cr_loss=0.3401, over 3317929.48 frames. ], batch size: 39, lr: 2.70e-03, grad_scale: 32.0 2024-09-25 16:37:17,412 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer_na.min_abs, batch_count=785768.6666666666, ans=0.02 2024-09-25 16:37:28,756 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=785815.3333333334, ans=0.125 2024-09-25 16:37:30,857 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=12.89 vs. limit=15.0 2024-09-25 16:37:36,648 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=785815.3333333334, ans=0.125 2024-09-25 16:37:40,750 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=5.75 vs. limit=15.0 2024-09-25 16:38:02,791 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=785908.6666666666, ans=0.2 2024-09-25 16:38:04,362 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=785908.6666666666, ans=0.07 2024-09-25 16:38:12,411 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.75 vs. limit=15.0 2024-09-25 16:38:15,203 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.max_abs, batch_count=785955.3333333334, ans=10.0 2024-09-25 16:38:23,708 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=2.45 vs. limit=6.0 2024-09-25 16:38:32,616 INFO [train.py:1198] (2/4) Epoch 44, batch 900, loss[loss=0.1837, ctc_loss=0.1155, cr_loss=0.3408, over 17350.00 frames. ], tot_loss[loss=0.1902, ctc_loss=0.1222, cr_loss=0.3402, over 3330966.53 frames. ], batch size: 48, lr: 2.70e-03, grad_scale: 32.0 2024-09-25 16:38:36,099 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=786002.0, ans=0.125 2024-09-25 16:38:42,573 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=786002.0, ans=0.1 2024-09-25 16:39:15,048 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.169e+02 1.314e+02 1.378e+02 1.470e+02 1.852e+02, threshold=2.756e+02, percent-clipped=0.0 2024-09-25 16:39:16,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.min_positive, batch_count=786095.3333333334, ans=0.05 2024-09-25 16:39:42,553 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=8.20 vs. limit=15.0 2024-09-25 16:39:46,937 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=786188.6666666666, ans=0.125 2024-09-25 16:39:47,057 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=786188.6666666666, ans=0.0 2024-09-25 16:39:53,451 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=786188.6666666666, ans=0.2 2024-09-25 16:39:58,020 INFO [train.py:1198] (2/4) Epoch 44, batch 950, loss[loss=0.1719, ctc_loss=0.1084, cr_loss=0.3171, over 16268.00 frames. ], tot_loss[loss=0.1906, ctc_loss=0.1225, cr_loss=0.3404, over 3340926.00 frames. ], batch size: 36, lr: 2.70e-03, grad_scale: 32.0 2024-09-25 16:39:59,971 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=786235.3333333334, ans=0.0 2024-09-25 16:40:12,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=786282.0, ans=0.07 2024-09-25 16:40:25,783 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.80 vs. limit=12.0 2024-09-25 16:40:34,929 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=786328.6666666666, ans=0.125 2024-09-25 16:40:57,385 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=786375.3333333334, ans=0.1 2024-09-25 16:41:13,679 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=786422.0, ans=0.125 2024-09-25 16:41:18,074 INFO [train.py:1198] (2/4) Epoch 44, batch 1000, loss[loss=0.1988, ctc_loss=0.1279, cr_loss=0.3543, over 16994.00 frames. ], tot_loss[loss=0.1914, ctc_loss=0.123, cr_loss=0.3419, over 3349641.89 frames. ], batch size: 53, lr: 2.70e-03, grad_scale: 16.0 2024-09-25 16:41:27,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=786468.6666666666, ans=0.125 2024-09-25 16:41:33,324 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.63 vs. limit=10.0 2024-09-25 16:41:34,227 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=786515.3333333334, ans=0.1 2024-09-25 16:42:02,019 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.147e+02 1.298e+02 1.360e+02 1.466e+02 2.434e+02, threshold=2.721e+02, percent-clipped=0.0 2024-09-25 16:42:40,314 INFO [train.py:1198] (2/4) Epoch 44, batch 1050, loss[loss=0.158, ctc_loss=0.09928, cr_loss=0.2936, over 16663.00 frames. ], tot_loss[loss=0.1896, ctc_loss=0.1216, cr_loss=0.3399, over 3361858.33 frames. ], batch size: 37, lr: 2.70e-03, grad_scale: 16.0 2024-09-25 16:43:19,988 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=16.61 vs. limit=22.5 2024-09-25 16:43:56,127 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=786888.6666666666, ans=0.125 2024-09-25 16:43:56,711 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=16.13 vs. limit=22.5 2024-09-25 16:43:57,679 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=786888.6666666666, ans=0.125 2024-09-25 16:44:02,232 INFO [train.py:1198] (2/4) Epoch 44, batch 1100, loss[loss=0.1715, ctc_loss=0.1088, cr_loss=0.3138, over 17084.00 frames. ], tot_loss[loss=0.189, ctc_loss=0.1211, cr_loss=0.3394, over 3367078.99 frames. ], batch size: 43, lr: 2.70e-03, grad_scale: 16.0 2024-09-25 16:44:03,430 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=9.89 vs. limit=15.0 2024-09-25 16:44:09,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=786935.3333333334, ans=0.07 2024-09-25 16:44:33,110 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=786982.0, ans=0.125 2024-09-25 16:44:36,195 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=786982.0, ans=0.0 2024-09-25 16:44:48,722 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.176e+02 1.281e+02 1.343e+02 1.422e+02 3.447e+02, threshold=2.685e+02, percent-clipped=1.0 2024-09-25 16:44:58,668 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=787075.3333333334, ans=0.2 2024-09-25 16:45:11,627 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=787122.0, ans=0.0 2024-09-25 16:45:12,294 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=10.90 vs. limit=22.5 2024-09-25 16:45:27,408 INFO [train.py:1198] (2/4) Epoch 44, batch 1150, loss[loss=0.189, ctc_loss=0.1214, cr_loss=0.3379, over 17314.00 frames. ], tot_loss[loss=0.1891, ctc_loss=0.1212, cr_loss=0.3392, over 3366408.27 frames. ], batch size: 49, lr: 2.70e-03, grad_scale: 16.0 2024-09-25 16:46:13,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=787308.6666666666, ans=0.0 2024-09-25 16:46:14,493 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.88 vs. limit=6.0 2024-09-25 16:46:37,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=787355.3333333334, ans=0.125 2024-09-25 16:46:47,229 INFO [train.py:1198] (2/4) Epoch 44, batch 1200, loss[loss=0.1895, ctc_loss=0.1242, cr_loss=0.3266, over 16781.00 frames. ], tot_loss[loss=0.1897, ctc_loss=0.1217, cr_loss=0.3401, over 3366644.74 frames. ], batch size: 61, lr: 2.70e-03, grad_scale: 32.0 2024-09-25 16:47:18,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=787448.6666666666, ans=0.125 2024-09-25 16:47:26,631 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=787495.3333333334, ans=0.0 2024-09-25 16:47:31,121 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.306e+02 1.377e+02 1.476e+02 2.006e+02, threshold=2.753e+02, percent-clipped=0.0 2024-09-25 16:47:33,060 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 16:47:38,069 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=787542.0, ans=0.125 2024-09-25 16:47:42,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=787542.0, ans=0.125 2024-09-25 16:48:01,537 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=787588.6666666666, ans=0.2 2024-09-25 16:48:05,086 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=8.84 vs. limit=15.0 2024-09-25 16:48:11,736 INFO [train.py:1198] (2/4) Epoch 44, batch 1250, loss[loss=0.1973, ctc_loss=0.1265, cr_loss=0.3543, over 17147.00 frames. ], tot_loss[loss=0.1883, ctc_loss=0.1207, cr_loss=0.3383, over 3368704.47 frames. ], batch size: 45, lr: 2.70e-03, grad_scale: 32.0 2024-09-25 16:48:19,133 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=3.91 vs. limit=15.0 2024-09-25 16:48:52,767 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 16:48:58,995 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=787775.3333333334, ans=0.0 2024-09-25 16:49:04,749 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=787775.3333333334, ans=0.125 2024-09-25 16:49:06,835 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=11.47 vs. limit=22.5 2024-09-25 16:49:09,460 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=787775.3333333334, ans=0.04949747468305833 2024-09-25 16:49:30,308 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=5.22 vs. limit=12.0 2024-09-25 16:49:37,341 INFO [train.py:1198] (2/4) Epoch 44, batch 1300, loss[loss=0.2219, ctc_loss=0.1433, cr_loss=0.3929, over 17214.00 frames. ], tot_loss[loss=0.1879, ctc_loss=0.1204, cr_loss=0.3373, over 3369804.13 frames. ], batch size: 47, lr: 2.70e-03, grad_scale: 32.0 2024-09-25 16:49:47,253 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=787868.6666666666, ans=0.0 2024-09-25 16:50:06,485 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=787915.3333333334, ans=0.1 2024-09-25 16:50:14,504 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=787962.0, ans=0.125 2024-09-25 16:50:18,914 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.156e+02 1.318e+02 1.377e+02 1.473e+02 1.934e+02, threshold=2.754e+02, percent-clipped=0.0 2024-09-25 16:50:44,465 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=788055.3333333334, ans=0.1 2024-09-25 16:50:57,001 INFO [train.py:1198] (2/4) Epoch 44, batch 1350, loss[loss=0.1755, ctc_loss=0.1115, cr_loss=0.3196, over 16757.00 frames. ], tot_loss[loss=0.1878, ctc_loss=0.1205, cr_loss=0.3367, over 3366176.84 frames. ], batch size: 37, lr: 2.70e-03, grad_scale: 32.0 2024-09-25 16:51:11,511 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=788148.6666666666, ans=0.2 2024-09-25 16:51:16,483 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=788148.6666666666, ans=0.125 2024-09-25 16:51:59,260 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=13.53 vs. limit=22.5 2024-09-25 16:52:08,487 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.min_positive, batch_count=788288.6666666666, ans=0.025 2024-09-25 16:52:14,852 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer_ff3.min_abs, batch_count=788288.6666666666, ans=0.2 2024-09-25 16:52:19,417 INFO [train.py:1198] (2/4) Epoch 44, batch 1400, loss[loss=0.1813, ctc_loss=0.1155, cr_loss=0.329, over 17308.00 frames. ], tot_loss[loss=0.1877, ctc_loss=0.1205, cr_loss=0.3364, over 3366510.46 frames. ], batch size: 49, lr: 2.70e-03, grad_scale: 32.0 2024-09-25 16:52:21,372 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=788335.3333333334, ans=0.1 2024-09-25 16:52:27,814 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=788335.3333333334, ans=0.5 2024-09-25 16:53:01,170 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.288e+02 1.388e+02 1.492e+02 2.105e+02, threshold=2.776e+02, percent-clipped=0.0 2024-09-25 16:53:18,453 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=788475.3333333334, ans=0.1 2024-09-25 16:53:31,408 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-25 16:53:42,245 INFO [train.py:1198] (2/4) Epoch 44, batch 1450, loss[loss=0.1968, ctc_loss=0.1275, cr_loss=0.3463, over 16725.00 frames. ], tot_loss[loss=0.1873, ctc_loss=0.12, cr_loss=0.3362, over 3369032.36 frames. ], batch size: 61, lr: 2.70e-03, grad_scale: 32.0 2024-09-25 16:53:45,919 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=788568.6666666666, ans=0.0 2024-09-25 16:54:31,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=788662.0, ans=0.0 2024-09-25 16:55:07,624 INFO [train.py:1198] (2/4) Epoch 44, batch 1500, loss[loss=0.1955, ctc_loss=0.1307, cr_loss=0.3237, over 16941.00 frames. ], tot_loss[loss=0.1877, ctc_loss=0.1204, cr_loss=0.3364, over 3366992.93 frames. ], batch size: 58, lr: 2.70e-03, grad_scale: 16.0 2024-09-25 16:55:27,165 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=788848.6666666666, ans=0.125 2024-09-25 16:55:40,052 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=788895.3333333334, ans=0.2 2024-09-25 16:55:49,582 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=788895.3333333334, ans=0.125 2024-09-25 16:55:50,910 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.157e+02 1.297e+02 1.379e+02 1.448e+02 1.999e+02, threshold=2.757e+02, percent-clipped=0.0 2024-09-25 16:56:06,926 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=788942.0, ans=0.125 2024-09-25 16:56:08,706 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=788942.0, ans=0.95 2024-09-25 16:56:10,713 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=11.99 vs. limit=15.0 2024-09-25 16:56:27,695 INFO [train.py:1198] (2/4) Epoch 44, batch 1550, loss[loss=0.1552, ctc_loss=0.09699, cr_loss=0.2911, over 16712.00 frames. ], tot_loss[loss=0.1887, ctc_loss=0.121, cr_loss=0.3384, over 3371040.09 frames. ], batch size: 37, lr: 2.70e-03, grad_scale: 16.0 2024-09-25 16:56:37,545 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=789035.3333333334, ans=0.125 2024-09-25 16:56:40,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=789035.3333333334, ans=0.0 2024-09-25 16:57:11,549 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.max_abs, batch_count=789128.6666666666, ans=10.0 2024-09-25 16:57:18,094 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=11.05 vs. limit=12.0 2024-09-25 16:57:40,135 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=789222.0, ans=0.125 2024-09-25 16:57:49,473 INFO [train.py:1198] (2/4) Epoch 44, batch 1600, loss[loss=0.1924, ctc_loss=0.1229, cr_loss=0.3474, over 17365.00 frames. ], tot_loss[loss=0.189, ctc_loss=0.1213, cr_loss=0.3385, over 3375793.00 frames. ], batch size: 48, lr: 2.70e-03, grad_scale: 32.0 2024-09-25 16:57:57,815 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=789268.6666666666, ans=0.025 2024-09-25 16:58:02,394 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=789268.6666666666, ans=0.2 2024-09-25 16:58:26,108 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.65 vs. limit=15.0 2024-09-25 16:58:34,987 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.295e+02 1.391e+02 1.482e+02 2.401e+02, threshold=2.782e+02, percent-clipped=0.0 2024-09-25 16:58:40,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=789408.6666666666, ans=0.125 2024-09-25 16:58:51,312 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=789408.6666666666, ans=0.125 2024-09-25 16:58:59,250 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=789455.3333333334, ans=0.125 2024-09-25 16:59:14,302 INFO [train.py:1198] (2/4) Epoch 44, batch 1650, loss[loss=0.1871, ctc_loss=0.1243, cr_loss=0.314, over 17306.00 frames. ], tot_loss[loss=0.1895, ctc_loss=0.1217, cr_loss=0.3391, over 3363019.00 frames. ], batch size: 51, lr: 2.70e-03, grad_scale: 32.0 2024-09-25 16:59:15,216 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=16.91 vs. limit=22.5 2024-09-25 16:59:28,276 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=789502.0, ans=0.0 2024-09-25 17:00:09,944 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=789642.0, ans=0.2 2024-09-25 17:00:22,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=789688.6666666666, ans=0.1 2024-09-25 17:00:34,772 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1.whitening_limit, batch_count=789688.6666666666, ans=10.0 2024-09-25 17:00:36,855 INFO [train.py:1198] (2/4) Epoch 44, batch 1700, loss[loss=0.1872, ctc_loss=0.1196, cr_loss=0.3382, over 17251.00 frames. ], tot_loss[loss=0.1897, ctc_loss=0.1218, cr_loss=0.3395, over 3357630.88 frames. ], batch size: 44, lr: 2.70e-03, grad_scale: 32.0 2024-09-25 17:00:38,713 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=789735.3333333334, ans=0.0 2024-09-25 17:00:56,489 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=789782.0, ans=0.1 2024-09-25 17:01:20,178 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.198e+02 1.326e+02 1.402e+02 1.495e+02 1.823e+02, threshold=2.804e+02, percent-clipped=0.0 2024-09-25 17:01:22,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=789828.6666666666, ans=0.2 2024-09-25 17:01:28,923 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=512, metric=6.05 vs. limit=15.0 2024-09-25 17:01:38,025 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=789875.3333333334, ans=0.125 2024-09-25 17:01:41,316 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=789922.0, ans=0.125 2024-09-25 17:01:56,517 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=789922.0, ans=0.0 2024-09-25 17:01:59,534 INFO [train.py:1198] (2/4) Epoch 44, batch 1750, loss[loss=0.1847, ctc_loss=0.1188, cr_loss=0.3296, over 17010.00 frames. ], tot_loss[loss=0.1895, ctc_loss=0.1217, cr_loss=0.3392, over 3365705.97 frames. ], batch size: 44, lr: 2.70e-03, grad_scale: 32.0 2024-09-25 17:02:20,708 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=790015.3333333334, ans=0.2 2024-09-25 17:03:14,393 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=790155.3333333334, ans=0.125 2024-09-25 17:03:22,076 INFO [train.py:1198] (2/4) Epoch 44, batch 1800, loss[loss=0.2067, ctc_loss=0.1335, cr_loss=0.3662, over 16923.00 frames. ], tot_loss[loss=0.1885, ctc_loss=0.1209, cr_loss=0.3378, over 3368963.31 frames. ], batch size: 58, lr: 2.70e-03, grad_scale: 32.0 2024-09-25 17:03:49,207 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=8.51 vs. limit=15.0 2024-09-25 17:03:54,586 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=790295.3333333334, ans=0.2 2024-09-25 17:04:08,009 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.167e+02 1.325e+02 1.389e+02 1.497e+02 2.037e+02, threshold=2.779e+02, percent-clipped=0.0 2024-09-25 17:04:16,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=790342.0, ans=0.125 2024-09-25 17:04:20,642 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=790342.0, ans=0.125 2024-09-25 17:04:23,732 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=790342.0, ans=0.125 2024-09-25 17:04:32,674 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.32 vs. limit=15.0 2024-09-25 17:04:33,355 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=790388.6666666666, ans=0.125 2024-09-25 17:04:36,778 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=790388.6666666666, ans=0.025 2024-09-25 17:04:44,731 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=790388.6666666666, ans=0.125 2024-09-25 17:04:47,703 INFO [train.py:1198] (2/4) Epoch 44, batch 1850, loss[loss=0.2092, ctc_loss=0.1352, cr_loss=0.37, over 17167.00 frames. ], tot_loss[loss=0.1878, ctc_loss=0.1205, cr_loss=0.3366, over 3365893.62 frames. ], batch size: 45, lr: 2.70e-03, grad_scale: 32.0 2024-09-25 17:05:00,770 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=790435.3333333334, ans=0.0 2024-09-25 17:05:03,083 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=9.38 vs. limit=10.0 2024-09-25 17:05:04,354 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=8.93 vs. limit=15.0 2024-09-25 17:05:18,349 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=790528.6666666666, ans=0.125 2024-09-25 17:06:05,356 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=790622.0, ans=0.025 2024-09-25 17:06:08,120 INFO [train.py:1198] (2/4) Epoch 44, batch 1900, loss[loss=0.1829, ctc_loss=0.1147, cr_loss=0.3407, over 16720.00 frames. ], tot_loss[loss=0.1873, ctc_loss=0.1201, cr_loss=0.336, over 3360730.38 frames. ], batch size: 61, lr: 2.70e-03, grad_scale: 32.0 2024-09-25 17:06:14,726 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=790668.6666666666, ans=0.125 2024-09-25 17:06:19,748 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=790668.6666666666, ans=0.0 2024-09-25 17:06:39,210 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 17:06:40,555 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=790762.0, ans=0.125 2024-09-25 17:06:54,117 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.146e+02 1.292e+02 1.379e+02 1.443e+02 2.422e+02, threshold=2.758e+02, percent-clipped=0.0 2024-09-25 17:06:56,073 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=790762.0, ans=0.1 2024-09-25 17:07:04,201 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=790808.6666666666, ans=0.2 2024-09-25 17:07:31,338 INFO [train.py:1198] (2/4) Epoch 44, batch 1950, loss[loss=0.1941, ctc_loss=0.1259, cr_loss=0.3408, over 17135.00 frames. ], tot_loss[loss=0.1882, ctc_loss=0.1208, cr_loss=0.337, over 3348490.02 frames. ], batch size: 48, lr: 2.69e-03, grad_scale: 32.0 2024-09-25 17:07:48,452 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=7.58 vs. limit=15.0 2024-09-25 17:08:01,333 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=9.07 vs. limit=10.0 2024-09-25 17:08:56,404 INFO [train.py:1198] (2/4) Epoch 44, batch 2000, loss[loss=0.1712, ctc_loss=0.1104, cr_loss=0.3039, over 17073.00 frames. ], tot_loss[loss=0.1879, ctc_loss=0.1206, cr_loss=0.3366, over 3348639.14 frames. ], batch size: 46, lr: 2.69e-03, grad_scale: 32.0 2024-09-25 17:08:58,837 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.42 vs. limit=15.0 2024-09-25 17:09:20,035 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=791182.0, ans=0.125 2024-09-25 17:09:27,613 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=791182.0, ans=0.025 2024-09-25 17:09:39,392 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=512, metric=11.92 vs. limit=22.5 2024-09-25 17:09:42,757 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=9.21 vs. limit=15.0 2024-09-25 17:09:43,343 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.169e+02 1.332e+02 1.440e+02 1.511e+02 2.187e+02, threshold=2.879e+02, percent-clipped=0.0 2024-09-25 17:09:50,464 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.85 vs. limit=15.0 2024-09-25 17:10:18,720 INFO [train.py:1198] (2/4) Epoch 44, batch 2050, loss[loss=0.1933, ctc_loss=0.1231, cr_loss=0.3507, over 17292.00 frames. ], tot_loss[loss=0.1893, ctc_loss=0.1215, cr_loss=0.3389, over 3352101.58 frames. ], batch size: 49, lr: 2.69e-03, grad_scale: 32.0 2024-09-25 17:10:22,309 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=791368.6666666666, ans=0.125 2024-09-25 17:10:23,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=791368.6666666666, ans=0.1 2024-09-25 17:10:28,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=791368.6666666666, ans=0.0 2024-09-25 17:10:33,433 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=791415.3333333334, ans=10.0 2024-09-25 17:10:38,598 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.79 vs. limit=10.0 2024-09-25 17:11:38,300 INFO [train.py:1198] (2/4) Epoch 44, batch 2100, loss[loss=0.192, ctc_loss=0.1241, cr_loss=0.3396, over 17102.00 frames. ], tot_loss[loss=0.1885, ctc_loss=0.1208, cr_loss=0.338, over 3362104.79 frames. ], batch size: 49, lr: 2.69e-03, grad_scale: 32.0 2024-09-25 17:11:51,406 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.30 vs. limit=10.0 2024-09-25 17:11:52,890 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.51 vs. limit=15.0 2024-09-25 17:12:00,090 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=791648.6666666666, ans=0.0 2024-09-25 17:12:14,864 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=9.25 vs. limit=12.0 2024-09-25 17:12:25,394 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.180e+02 1.332e+02 1.398e+02 1.467e+02 2.660e+02, threshold=2.796e+02, percent-clipped=0.0 2024-09-25 17:12:28,783 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=791742.0, ans=0.125 2024-09-25 17:12:30,672 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=791742.0, ans=0.125 2024-09-25 17:13:00,853 INFO [train.py:1198] (2/4) Epoch 44, batch 2150, loss[loss=0.193, ctc_loss=0.1222, cr_loss=0.3539, over 17204.00 frames. ], tot_loss[loss=0.1889, ctc_loss=0.1212, cr_loss=0.339, over 3364460.16 frames. ], batch size: 47, lr: 2.69e-03, grad_scale: 32.0 2024-09-25 17:13:37,601 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=791928.6666666666, ans=0.125 2024-09-25 17:13:39,259 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=791928.6666666666, ans=0.0 2024-09-25 17:13:42,474 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=791928.6666666666, ans=0.0 2024-09-25 17:13:51,883 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=791975.3333333334, ans=0.07 2024-09-25 17:14:07,930 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.87 vs. limit=6.0 2024-09-25 17:14:28,853 INFO [train.py:1198] (2/4) Epoch 44, batch 2200, loss[loss=0.143, ctc_loss=0.08758, cr_loss=0.2769, over 16256.00 frames. ], tot_loss[loss=0.1884, ctc_loss=0.1207, cr_loss=0.3383, over 3366478.67 frames. ], batch size: 36, lr: 2.69e-03, grad_scale: 32.0 2024-09-25 17:14:33,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=792068.6666666666, ans=0.125 2024-09-25 17:14:40,792 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.70 vs. limit=10.0 2024-09-25 17:14:41,962 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=792068.6666666666, ans=0.0 2024-09-25 17:14:47,269 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=10.63 vs. limit=22.5 2024-09-25 17:14:51,569 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=792115.3333333334, ans=0.2 2024-09-25 17:15:10,944 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=792162.0, ans=0.125 2024-09-25 17:15:13,651 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.155e+02 1.270e+02 1.367e+02 1.447e+02 1.926e+02, threshold=2.734e+02, percent-clipped=0.0 2024-09-25 17:15:34,927 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=792255.3333333334, ans=0.0 2024-09-25 17:15:47,706 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=792302.0, ans=0.1 2024-09-25 17:15:48,885 INFO [train.py:1198] (2/4) Epoch 44, batch 2250, loss[loss=0.2042, ctc_loss=0.1288, cr_loss=0.3771, over 17229.00 frames. ], tot_loss[loss=0.1876, ctc_loss=0.1201, cr_loss=0.3373, over 3374984.47 frames. ], batch size: 55, lr: 2.69e-03, grad_scale: 32.0 2024-09-25 17:15:51,144 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.30 vs. limit=15.0 2024-09-25 17:16:10,105 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=792348.6666666666, ans=0.125 2024-09-25 17:16:21,062 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=792395.3333333334, ans=0.125 2024-09-25 17:16:26,075 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.max_abs, batch_count=792395.3333333334, ans=10.0 2024-09-25 17:16:36,540 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=16.15 vs. limit=22.5 2024-09-25 17:16:39,166 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.33 vs. limit=15.0 2024-09-25 17:17:11,218 INFO [train.py:1198] (2/4) Epoch 44, batch 2300, loss[loss=0.1904, ctc_loss=0.1258, cr_loss=0.3227, over 16049.00 frames. ], tot_loss[loss=0.1877, ctc_loss=0.1201, cr_loss=0.3378, over 3374588.44 frames. ], batch size: 74, lr: 2.69e-03, grad_scale: 32.0 2024-09-25 17:17:19,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=792535.3333333334, ans=0.0 2024-09-25 17:17:51,397 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=792628.6666666666, ans=0.125 2024-09-25 17:17:55,695 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.062e+02 1.302e+02 1.387e+02 1.511e+02 2.811e+02, threshold=2.774e+02, percent-clipped=1.0 2024-09-25 17:18:33,646 INFO [train.py:1198] (2/4) Epoch 44, batch 2350, loss[loss=0.1878, ctc_loss=0.1181, cr_loss=0.3484, over 17276.00 frames. ], tot_loss[loss=0.1868, ctc_loss=0.1194, cr_loss=0.337, over 3375080.70 frames. ], batch size: 46, lr: 2.69e-03, grad_scale: 32.0 2024-09-25 17:18:34,041 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=792768.6666666666, ans=0.1 2024-09-25 17:18:48,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=792815.3333333334, ans=0.125 2024-09-25 17:19:16,591 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=792862.0, ans=0.05 2024-09-25 17:19:37,206 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=792908.6666666666, ans=0.0 2024-09-25 17:19:48,349 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=792955.3333333334, ans=0.125 2024-09-25 17:19:50,021 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=792955.3333333334, ans=0.125 2024-09-25 17:19:53,245 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=792955.3333333334, ans=0.125 2024-09-25 17:19:59,163 INFO [train.py:1198] (2/4) Epoch 44, batch 2400, loss[loss=0.18, ctc_loss=0.1118, cr_loss=0.341, over 17287.00 frames. ], tot_loss[loss=0.1872, ctc_loss=0.1196, cr_loss=0.3378, over 3381217.02 frames. ], batch size: 46, lr: 2.69e-03, grad_scale: 32.0 2024-09-25 17:20:01,152 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=793002.0, ans=0.2 2024-09-25 17:20:15,458 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=793048.6666666666, ans=0.0 2024-09-25 17:20:45,389 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.312e+02 1.417e+02 1.545e+02 2.964e+02, threshold=2.835e+02, percent-clipped=1.0 2024-09-25 17:21:12,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=793188.6666666666, ans=0.125 2024-09-25 17:21:19,082 INFO [train.py:1198] (2/4) Epoch 44, batch 2450, loss[loss=0.187, ctc_loss=0.12, cr_loss=0.335, over 17016.00 frames. ], tot_loss[loss=0.1868, ctc_loss=0.1195, cr_loss=0.3369, over 3373953.88 frames. ], batch size: 51, lr: 2.69e-03, grad_scale: 32.0 2024-09-25 17:21:39,399 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.40 vs. limit=15.0 2024-09-25 17:21:46,875 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=5.00 vs. limit=10.0 2024-09-25 17:22:27,098 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=793422.0, ans=0.125 2024-09-25 17:22:27,100 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=793422.0, ans=0.0 2024-09-25 17:22:42,508 INFO [train.py:1198] (2/4) Epoch 44, batch 2500, loss[loss=0.2072, ctc_loss=0.1355, cr_loss=0.3586, over 17038.00 frames. ], tot_loss[loss=0.1878, ctc_loss=0.1202, cr_loss=0.3379, over 3363712.72 frames. ], batch size: 52, lr: 2.69e-03, grad_scale: 32.0 2024-09-25 17:23:17,688 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=793562.0, ans=0.0 2024-09-25 17:23:31,902 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 1.312e+02 1.368e+02 1.453e+02 1.982e+02, threshold=2.736e+02, percent-clipped=0.0 2024-09-25 17:23:48,596 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.08 vs. limit=6.0 2024-09-25 17:24:02,273 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=793655.3333333334, ans=0.0 2024-09-25 17:24:05,466 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=793655.3333333334, ans=0.0 2024-09-25 17:24:08,164 INFO [train.py:1198] (2/4) Epoch 44, batch 2550, loss[loss=0.1969, ctc_loss=0.1259, cr_loss=0.355, over 17246.00 frames. ], tot_loss[loss=0.1879, ctc_loss=0.1204, cr_loss=0.338, over 3357871.85 frames. ], batch size: 44, lr: 2.69e-03, grad_scale: 16.0 2024-09-25 17:24:48,098 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=793795.3333333334, ans=0.0 2024-09-25 17:24:51,479 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 17:25:00,748 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=793842.0, ans=0.0 2024-09-25 17:25:12,132 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=793842.0, ans=0.0 2024-09-25 17:25:31,104 INFO [train.py:1198] (2/4) Epoch 44, batch 2600, loss[loss=0.2375, ctc_loss=0.1602, cr_loss=0.3868, over 15115.00 frames. ], tot_loss[loss=0.1892, ctc_loss=0.1213, cr_loss=0.3393, over 3349922.86 frames. ], batch size: 89, lr: 2.69e-03, grad_scale: 16.0 2024-09-25 17:25:31,496 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=793935.3333333334, ans=0.125 2024-09-25 17:25:43,162 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.40 vs. limit=6.0 2024-09-25 17:25:44,320 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 17:25:57,138 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=793982.0, ans=0.0 2024-09-25 17:26:19,163 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.146e+02 1.289e+02 1.404e+02 1.511e+02 2.276e+02, threshold=2.809e+02, percent-clipped=0.0 2024-09-25 17:26:19,759 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=13.91 vs. limit=22.5 2024-09-25 17:26:54,090 INFO [train.py:1198] (2/4) Epoch 44, batch 2650, loss[loss=0.2026, ctc_loss=0.1299, cr_loss=0.3632, over 17033.00 frames. ], tot_loss[loss=0.1895, ctc_loss=0.1215, cr_loss=0.3397, over 3353491.37 frames. ], batch size: 44, lr: 2.69e-03, grad_scale: 16.0 2024-09-25 17:27:05,938 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=794168.6666666666, ans=0.0 2024-09-25 17:27:17,114 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=794215.3333333334, ans=0.0 2024-09-25 17:27:21,673 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=794215.3333333334, ans=0.2 2024-09-25 17:27:31,661 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=794262.0, ans=0.125 2024-09-25 17:27:50,927 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=794308.6666666666, ans=0.125 2024-09-25 17:27:54,020 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=794308.6666666666, ans=0.0 2024-09-25 17:28:02,303 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=794355.3333333334, ans=0.125 2024-09-25 17:28:17,364 INFO [train.py:1198] (2/4) Epoch 44, batch 2700, loss[loss=0.2017, ctc_loss=0.1324, cr_loss=0.3464, over 16894.00 frames. ], tot_loss[loss=0.1899, ctc_loss=0.1219, cr_loss=0.3399, over 3353040.87 frames. ], batch size: 58, lr: 2.69e-03, grad_scale: 16.0 2024-09-25 17:28:27,275 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=794402.0, ans=0.125 2024-09-25 17:28:46,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=794448.6666666666, ans=0.09899494936611666 2024-09-25 17:28:51,858 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.20 vs. limit=12.0 2024-09-25 17:29:07,857 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.291e+02 1.350e+02 1.441e+02 1.690e+02, threshold=2.700e+02, percent-clipped=0.0 2024-09-25 17:29:32,364 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=13.55 vs. limit=22.5 2024-09-25 17:29:43,111 INFO [train.py:1198] (2/4) Epoch 44, batch 2750, loss[loss=0.2296, ctc_loss=0.1483, cr_loss=0.4061, over 15024.00 frames. ], tot_loss[loss=0.1885, ctc_loss=0.121, cr_loss=0.3379, over 3355331.68 frames. ], batch size: 89, lr: 2.69e-03, grad_scale: 16.0 2024-09-25 17:30:20,041 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=794728.6666666666, ans=0.0 2024-09-25 17:30:34,317 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=794775.3333333334, ans=0.125 2024-09-25 17:30:42,243 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=794775.3333333334, ans=0.1 2024-09-25 17:30:43,810 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=794775.3333333334, ans=0.125 2024-09-25 17:31:02,586 INFO [train.py:1198] (2/4) Epoch 44, batch 2800, loss[loss=0.2203, ctc_loss=0.1435, cr_loss=0.384, over 16590.00 frames. ], tot_loss[loss=0.1881, ctc_loss=0.1206, cr_loss=0.3376, over 3364741.75 frames. ], batch size: 66, lr: 2.69e-03, grad_scale: 32.0 2024-09-25 17:31:06,699 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten.whitening_limit, batch_count=794868.6666666666, ans=15.0 2024-09-25 17:31:11,142 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=13.98 vs. limit=15.0 2024-09-25 17:31:36,105 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=794962.0, ans=0.125 2024-09-25 17:31:36,380 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=794962.0, ans=0.125 2024-09-25 17:31:52,768 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.131e+02 1.331e+02 1.405e+02 1.538e+02 1.952e+02, threshold=2.809e+02, percent-clipped=0.0 2024-09-25 17:31:54,852 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=795008.6666666666, ans=0.0 2024-09-25 17:32:04,225 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=795008.6666666666, ans=0.0 2024-09-25 17:32:20,277 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=795055.3333333334, ans=0.04949747468305833 2024-09-25 17:32:20,404 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=795055.3333333334, ans=0.125 2024-09-25 17:32:24,904 INFO [train.py:1198] (2/4) Epoch 44, batch 2850, loss[loss=0.1986, ctc_loss=0.1273, cr_loss=0.3563, over 17098.00 frames. ], tot_loss[loss=0.1878, ctc_loss=0.1203, cr_loss=0.3375, over 3366517.21 frames. ], batch size: 49, lr: 2.69e-03, grad_scale: 32.0 2024-09-25 17:32:55,620 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=795195.3333333334, ans=0.0 2024-09-25 17:33:11,087 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=795195.3333333334, ans=0.0 2024-09-25 17:33:47,967 INFO [train.py:1198] (2/4) Epoch 44, batch 2900, loss[loss=0.2139, ctc_loss=0.1397, cr_loss=0.3707, over 15243.00 frames. ], tot_loss[loss=0.188, ctc_loss=0.1204, cr_loss=0.3379, over 3364129.99 frames. ], batch size: 89, lr: 2.69e-03, grad_scale: 32.0 2024-09-25 17:33:57,154 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=795335.3333333334, ans=0.125 2024-09-25 17:34:32,020 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=795428.6666666666, ans=0.0 2024-09-25 17:34:37,388 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=9.67 vs. limit=22.5 2024-09-25 17:34:41,409 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.183e+02 1.299e+02 1.362e+02 1.425e+02 2.572e+02, threshold=2.724e+02, percent-clipped=0.0 2024-09-25 17:34:54,453 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=795475.3333333334, ans=0.1 2024-09-25 17:35:13,453 INFO [train.py:1198] (2/4) Epoch 44, batch 2950, loss[loss=0.1983, ctc_loss=0.1275, cr_loss=0.3538, over 17303.00 frames. ], tot_loss[loss=0.1888, ctc_loss=0.121, cr_loss=0.3389, over 3367024.67 frames. ], batch size: 49, lr: 2.69e-03, grad_scale: 32.0 2024-09-25 17:35:34,948 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=795615.3333333334, ans=0.0 2024-09-25 17:36:32,585 INFO [train.py:1198] (2/4) Epoch 44, batch 3000, loss[loss=0.1831, ctc_loss=0.1147, cr_loss=0.342, over 17028.00 frames. ], tot_loss[loss=0.1901, ctc_loss=0.122, cr_loss=0.3403, over 3347407.76 frames. ], batch size: 39, lr: 2.69e-03, grad_scale: 32.0 2024-09-25 17:36:32,586 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 17:36:47,858 INFO [train.py:1230] (2/4) Epoch 44, validation: loss=0.03521, ctc_loss=0.03521, cr_loss=1.022e-14, over 944034.00 frames. 2024-09-25 17:36:47,859 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 17:37:17,996 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 17:37:22,604 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=795895.3333333334, ans=0.125 2024-09-25 17:37:22,640 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=795895.3333333334, ans=0.1 2024-09-25 17:37:24,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.max_positive, batch_count=795895.3333333334, ans=0.95 2024-09-25 17:37:34,778 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.305e+02 1.379e+02 1.486e+02 1.924e+02, threshold=2.758e+02, percent-clipped=0.0 2024-09-25 17:38:06,190 INFO [train.py:1198] (2/4) Epoch 44, batch 3050, loss[loss=0.1664, ctc_loss=0.107, cr_loss=0.2968, over 16950.00 frames. ], tot_loss[loss=0.1897, ctc_loss=0.1218, cr_loss=0.3397, over 3357692.22 frames. ], batch size: 42, lr: 2.69e-03, grad_scale: 32.0 2024-09-25 17:38:48,930 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=796128.6666666666, ans=0.1 2024-09-25 17:39:26,747 INFO [train.py:1198] (2/4) Epoch 44, batch 3100, loss[loss=0.1807, ctc_loss=0.115, cr_loss=0.3287, over 17083.00 frames. ], tot_loss[loss=0.1899, ctc_loss=0.1219, cr_loss=0.3401, over 3353704.73 frames. ], batch size: 49, lr: 2.69e-03, grad_scale: 32.0 2024-09-25 17:39:35,456 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.33 vs. limit=6.0 2024-09-25 17:40:13,596 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.166e+02 1.310e+02 1.386e+02 1.469e+02 1.981e+02, threshold=2.772e+02, percent-clipped=0.0 2024-09-25 17:40:18,582 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=796408.6666666666, ans=0.125 2024-09-25 17:40:22,345 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=11.49 vs. limit=22.5 2024-09-25 17:40:28,150 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.max_abs, batch_count=796455.3333333334, ans=10.0 2024-09-25 17:40:44,923 INFO [train.py:1198] (2/4) Epoch 44, batch 3150, loss[loss=0.1822, ctc_loss=0.1154, cr_loss=0.3342, over 17283.00 frames. ], tot_loss[loss=0.1887, ctc_loss=0.1209, cr_loss=0.3387, over 3361733.32 frames. ], batch size: 46, lr: 2.69e-03, grad_scale: 32.0 2024-09-25 17:40:56,249 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=796502.0, ans=0.125 2024-09-25 17:41:17,965 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.25 vs. limit=15.0 2024-09-25 17:41:34,763 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=796642.0, ans=0.125 2024-09-25 17:42:08,191 INFO [train.py:1198] (2/4) Epoch 44, batch 3200, loss[loss=0.1883, ctc_loss=0.1171, cr_loss=0.3559, over 17207.00 frames. ], tot_loss[loss=0.1891, ctc_loss=0.1212, cr_loss=0.3396, over 3355886.39 frames. ], batch size: 47, lr: 2.69e-03, grad_scale: 32.0 2024-09-25 17:42:08,579 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=796735.3333333334, ans=0.0 2024-09-25 17:42:10,384 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=11.02 vs. limit=12.0 2024-09-25 17:42:32,112 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=796782.0, ans=0.2 2024-09-25 17:42:56,543 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=796875.3333333334, ans=0.0 2024-09-25 17:42:57,898 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.113e+02 1.350e+02 1.424e+02 1.561e+02 1.915e+02, threshold=2.848e+02, percent-clipped=0.0 2024-09-25 17:43:04,512 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=796875.3333333334, ans=0.2 2024-09-25 17:43:21,909 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=796922.0, ans=0.125 2024-09-25 17:43:26,123 INFO [train.py:1198] (2/4) Epoch 44, batch 3250, loss[loss=0.2015, ctc_loss=0.1309, cr_loss=0.3531, over 17142.00 frames. ], tot_loss[loss=0.1887, ctc_loss=0.121, cr_loss=0.3388, over 3359035.18 frames. ], batch size: 48, lr: 2.68e-03, grad_scale: 16.0 2024-09-25 17:43:26,329 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=796968.6666666666, ans=0.125 2024-09-25 17:43:43,490 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=797015.3333333334, ans=0.025 2024-09-25 17:44:01,728 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=9.09 vs. limit=15.0 2024-09-25 17:44:08,221 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.46 vs. limit=15.0 2024-09-25 17:44:45,004 INFO [train.py:1198] (2/4) Epoch 44, batch 3300, loss[loss=0.1735, ctc_loss=0.1089, cr_loss=0.3228, over 17305.00 frames. ], tot_loss[loss=0.1882, ctc_loss=0.1207, cr_loss=0.3377, over 3359657.27 frames. ], batch size: 51, lr: 2.68e-03, grad_scale: 16.0 2024-09-25 17:44:56,149 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=797202.0, ans=0.1 2024-09-25 17:45:22,899 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=7.16 vs. limit=15.0 2024-09-25 17:45:23,870 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=797295.3333333334, ans=0.125 2024-09-25 17:45:25,513 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=797295.3333333334, ans=0.1 2024-09-25 17:45:34,580 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.138e+02 1.296e+02 1.381e+02 1.496e+02 2.395e+02, threshold=2.762e+02, percent-clipped=0.0 2024-09-25 17:45:41,028 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=797342.0, ans=0.1 2024-09-25 17:46:02,723 INFO [train.py:1198] (2/4) Epoch 44, batch 3350, loss[loss=0.1887, ctc_loss=0.1209, cr_loss=0.3392, over 17087.00 frames. ], tot_loss[loss=0.1875, ctc_loss=0.1202, cr_loss=0.3365, over 3358209.01 frames. ], batch size: 49, lr: 2.68e-03, grad_scale: 16.0 2024-09-25 17:46:18,385 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.min_positive, batch_count=797482.0, ans=0.05 2024-09-25 17:46:25,387 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.21 vs. limit=15.0 2024-09-25 17:46:27,972 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=797482.0, ans=0.125 2024-09-25 17:46:29,580 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=797482.0, ans=0.0 2024-09-25 17:47:21,459 INFO [train.py:1198] (2/4) Epoch 44, batch 3400, loss[loss=0.1741, ctc_loss=0.1101, cr_loss=0.32, over 17032.00 frames. ], tot_loss[loss=0.188, ctc_loss=0.1207, cr_loss=0.3367, over 3350280.88 frames. ], batch size: 39, lr: 2.68e-03, grad_scale: 16.0 2024-09-25 17:47:26,949 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=797668.6666666666, ans=0.1 2024-09-25 17:47:27,024 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=797668.6666666666, ans=0.125 2024-09-25 17:47:29,005 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=5.03 vs. limit=15.0 2024-09-25 17:47:46,269 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=8.50 vs. limit=12.0 2024-09-25 17:48:13,774 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.159e+02 1.291e+02 1.376e+02 1.452e+02 2.263e+02, threshold=2.752e+02, percent-clipped=0.0 2024-09-25 17:48:36,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=797855.3333333334, ans=0.0 2024-09-25 17:48:42,146 INFO [train.py:1198] (2/4) Epoch 44, batch 3450, loss[loss=0.2295, ctc_loss=0.1479, cr_loss=0.4079, over 16994.00 frames. ], tot_loss[loss=0.1896, ctc_loss=0.1218, cr_loss=0.3391, over 3340089.02 frames. ], batch size: 51, lr: 2.68e-03, grad_scale: 16.0 2024-09-25 17:49:07,309 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=797948.6666666666, ans=0.125 2024-09-25 17:49:20,105 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=797995.3333333334, ans=0.125 2024-09-25 17:49:27,735 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=798042.0, ans=0.125 2024-09-25 17:49:37,140 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=798042.0, ans=0.125 2024-09-25 17:50:02,278 INFO [train.py:1198] (2/4) Epoch 44, batch 3500, loss[loss=0.2349, ctc_loss=0.1629, cr_loss=0.36, over 11302.00 frames. ], tot_loss[loss=0.1882, ctc_loss=0.1208, cr_loss=0.337, over 3338181.43 frames. ], batch size: 123, lr: 2.68e-03, grad_scale: 16.0 2024-09-25 17:50:07,691 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=17.35 vs. limit=22.5 2024-09-25 17:50:18,659 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=798182.0, ans=0.2 2024-09-25 17:50:29,827 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.40 vs. limit=15.0 2024-09-25 17:50:39,053 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=798228.6666666666, ans=0.125 2024-09-25 17:50:51,684 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 17:50:52,795 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.122e+02 1.299e+02 1.358e+02 1.430e+02 3.438e+02, threshold=2.715e+02, percent-clipped=1.0 2024-09-25 17:51:00,350 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=4.88 vs. limit=15.0 2024-09-25 17:51:17,062 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=798322.0, ans=0.125 2024-09-25 17:51:22,862 INFO [train.py:1198] (2/4) Epoch 44, batch 3550, loss[loss=0.1592, ctc_loss=0.09943, cr_loss=0.2988, over 16301.00 frames. ], tot_loss[loss=0.1881, ctc_loss=0.1207, cr_loss=0.3372, over 3345529.22 frames. ], batch size: 36, lr: 2.68e-03, grad_scale: 16.0 2024-09-25 17:51:25,196 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=2.95 vs. limit=15.0 2024-09-25 17:51:43,534 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=798415.3333333334, ans=0.125 2024-09-25 17:51:43,554 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=798415.3333333334, ans=0.0 2024-09-25 17:51:54,338 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=4.56 vs. limit=15.0 2024-09-25 17:52:05,116 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=6.09 vs. limit=15.0 2024-09-25 17:52:14,482 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=7.64 vs. limit=15.0 2024-09-25 17:52:15,449 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=798508.6666666666, ans=0.1 2024-09-25 17:52:33,069 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=798555.3333333334, ans=0.0 2024-09-25 17:52:43,565 INFO [train.py:1198] (2/4) Epoch 44, batch 3600, loss[loss=0.1837, ctc_loss=0.1204, cr_loss=0.3163, over 17063.00 frames. ], tot_loss[loss=0.1885, ctc_loss=0.121, cr_loss=0.3377, over 3356732.20 frames. ], batch size: 46, lr: 2.68e-03, grad_scale: 16.0 2024-09-25 17:53:08,387 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=798648.6666666666, ans=0.025 2024-09-25 17:53:34,700 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.315e+02 1.375e+02 1.458e+02 2.973e+02, threshold=2.750e+02, percent-clipped=1.0 2024-09-25 17:53:35,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=798742.0, ans=0.0 2024-09-25 17:53:42,852 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=798742.0, ans=0.125 2024-09-25 17:54:01,419 INFO [train.py:1198] (2/4) Epoch 44, batch 3650, loss[loss=0.1837, ctc_loss=0.1161, cr_loss=0.338, over 17143.00 frames. ], tot_loss[loss=0.1891, ctc_loss=0.1213, cr_loss=0.3387, over 3355059.16 frames. ], batch size: 48, lr: 2.68e-03, grad_scale: 16.0 2024-09-25 17:54:41,536 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=13.25 vs. limit=22.5 2024-09-25 17:55:09,957 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=799022.0, ans=0.2 2024-09-25 17:55:20,636 INFO [train.py:1198] (2/4) Epoch 44, batch 3700, loss[loss=0.2021, ctc_loss=0.1364, cr_loss=0.3286, over 15966.00 frames. ], tot_loss[loss=0.1898, ctc_loss=0.1219, cr_loss=0.3395, over 3354689.85 frames. ], batch size: 74, lr: 2.68e-03, grad_scale: 16.0 2024-09-25 17:55:27,498 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.94 vs. limit=15.0 2024-09-25 17:55:28,546 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=799068.6666666666, ans=0.125 2024-09-25 17:55:41,090 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=799115.3333333334, ans=0.125 2024-09-25 17:55:41,473 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=9.89 vs. limit=22.5 2024-09-25 17:55:55,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=799162.0, ans=0.0 2024-09-25 17:56:10,684 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=799208.6666666666, ans=0.2 2024-09-25 17:56:10,733 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=799208.6666666666, ans=0.2 2024-09-25 17:56:11,869 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.103e+02 1.294e+02 1.350e+02 1.442e+02 2.627e+02, threshold=2.701e+02, percent-clipped=0.0 2024-09-25 17:56:15,849 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.31 vs. limit=15.0 2024-09-25 17:56:22,074 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=8.11 vs. limit=15.0 2024-09-25 17:56:38,698 INFO [train.py:1198] (2/4) Epoch 44, batch 3750, loss[loss=0.1567, ctc_loss=0.09723, cr_loss=0.2975, over 17175.00 frames. ], tot_loss[loss=0.19, ctc_loss=0.1221, cr_loss=0.3396, over 3345934.52 frames. ], batch size: 41, lr: 2.68e-03, grad_scale: 16.0 2024-09-25 17:56:59,483 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=799348.6666666666, ans=0.05 2024-09-25 17:57:24,817 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=799442.0, ans=0.0 2024-09-25 17:57:54,333 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=799488.6666666666, ans=0.2 2024-09-25 17:57:55,953 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=799535.3333333334, ans=0.125 2024-09-25 17:57:57,077 INFO [train.py:1198] (2/4) Epoch 44, batch 3800, loss[loss=0.1594, ctc_loss=0.09825, cr_loss=0.306, over 16243.00 frames. ], tot_loss[loss=0.1898, ctc_loss=0.1218, cr_loss=0.3397, over 3330973.49 frames. ], batch size: 36, lr: 2.68e-03, grad_scale: 16.0 2024-09-25 17:58:09,437 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=799535.3333333334, ans=0.125 2024-09-25 17:58:45,490 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=14.46 vs. limit=15.0 2024-09-25 17:58:47,998 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.185e+02 1.317e+02 1.409e+02 1.513e+02 1.849e+02, threshold=2.818e+02, percent-clipped=0.0 2024-09-25 17:59:14,342 INFO [train.py:1198] (2/4) Epoch 44, batch 3850, loss[loss=0.2192, ctc_loss=0.1459, cr_loss=0.3667, over 11589.00 frames. ], tot_loss[loss=0.1903, ctc_loss=0.1225, cr_loss=0.3394, over 3283141.46 frames. ], batch size: 123, lr: 2.68e-03, grad_scale: 16.0 2024-09-25 17:59:22,329 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=799768.6666666666, ans=0.125 2024-09-25 17:59:26,769 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=799768.6666666666, ans=0.0 2024-09-25 17:59:40,744 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=15.59 vs. limit=15.0 2024-09-25 17:59:59,383 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.99 vs. limit=15.0 2024-09-25 18:00:00,525 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=799908.6666666666, ans=0.04949747468305833 2024-09-25 18:01:14,663 INFO [train.py:1198] (2/4) Epoch 45, batch 0, loss[loss=0.1922, ctc_loss=0.1212, cr_loss=0.3547, over 17207.00 frames. ], tot_loss[loss=0.1922, ctc_loss=0.1212, cr_loss=0.3547, over 17207.00 frames. ], batch size: 47, lr: 2.65e-03, grad_scale: 32.0 2024-09-25 18:01:14,664 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 18:01:29,792 INFO [train.py:1230] (2/4) Epoch 45, validation: loss=0.03539, ctc_loss=0.03539, cr_loss=1.113e-14, over 944034.00 frames. 2024-09-25 18:01:29,793 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 18:01:38,858 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.06 vs. limit=6.0 2024-09-25 18:01:41,220 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=799983.3333333334, ans=0.2 2024-09-25 18:02:17,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=800076.6666666666, ans=0.125 2024-09-25 18:02:19,627 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=800123.3333333334, ans=0.025 2024-09-25 18:02:20,332 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.66 vs. limit=10.0 2024-09-25 18:02:22,090 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=6.55 vs. limit=15.0 2024-09-25 18:02:31,990 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.406e+02 1.547e+02 1.686e+02 2.322e+02, threshold=3.093e+02, percent-clipped=0.0 2024-09-25 18:02:49,881 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=800170.0, ans=0.0 2024-09-25 18:02:52,919 INFO [train.py:1198] (2/4) Epoch 45, batch 50, loss[loss=0.1899, ctc_loss=0.117, cr_loss=0.3646, over 16954.00 frames. ], tot_loss[loss=0.1867, ctc_loss=0.1198, cr_loss=0.3344, over 749226.09 frames. ], batch size: 42, lr: 2.65e-03, grad_scale: 32.0 2024-09-25 18:02:59,614 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=800216.6666666666, ans=0.2 2024-09-25 18:03:09,651 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten.whitening_limit, batch_count=800263.3333333334, ans=15.0 2024-09-25 18:03:10,774 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=800263.3333333334, ans=0.95 2024-09-25 18:03:16,369 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.38 vs. limit=10.0 2024-09-25 18:03:26,637 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=800310.0, ans=0.125 2024-09-25 18:03:35,120 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=5.40 vs. limit=10.0 2024-09-25 18:03:42,948 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=800356.6666666666, ans=0.07 2024-09-25 18:03:46,019 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=800356.6666666666, ans=0.0 2024-09-25 18:04:05,660 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=800403.3333333334, ans=0.125 2024-09-25 18:04:13,066 INFO [train.py:1198] (2/4) Epoch 45, batch 100, loss[loss=0.1929, ctc_loss=0.1221, cr_loss=0.354, over 17245.00 frames. ], tot_loss[loss=0.1878, ctc_loss=0.1205, cr_loss=0.3365, over 1326464.76 frames. ], batch size: 44, lr: 2.65e-03, grad_scale: 32.0 2024-09-25 18:04:27,858 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=800496.6666666666, ans=0.1 2024-09-25 18:04:27,996 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=800496.6666666666, ans=0.1 2024-09-25 18:04:43,680 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.prob, batch_count=800543.3333333334, ans=0.125 2024-09-25 18:05:12,263 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.147e+02 1.307e+02 1.359e+02 1.469e+02 2.520e+02, threshold=2.719e+02, percent-clipped=0.0 2024-09-25 18:05:35,982 INFO [train.py:1198] (2/4) Epoch 45, batch 150, loss[loss=0.1747, ctc_loss=0.1109, cr_loss=0.3187, over 17139.00 frames. ], tot_loss[loss=0.1872, ctc_loss=0.1202, cr_loss=0.3352, over 1770234.70 frames. ], batch size: 48, lr: 2.65e-03, grad_scale: 32.0 2024-09-25 18:06:01,700 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=800730.0, ans=0.125 2024-09-25 18:06:14,436 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=800776.6666666666, ans=0.125 2024-09-25 18:06:42,120 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=800823.3333333334, ans=0.0 2024-09-25 18:06:53,504 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=800870.0, ans=0.0 2024-09-25 18:07:02,957 INFO [train.py:1198] (2/4) Epoch 45, batch 200, loss[loss=0.1815, ctc_loss=0.1161, cr_loss=0.327, over 16662.00 frames. ], tot_loss[loss=0.1893, ctc_loss=0.1216, cr_loss=0.3387, over 2118510.14 frames. ], batch size: 61, lr: 2.65e-03, grad_scale: 32.0 2024-09-25 18:07:14,562 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=800916.6666666666, ans=0.125 2024-09-25 18:07:19,189 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=800963.3333333334, ans=0.0 2024-09-25 18:07:30,350 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=800963.3333333334, ans=0.125 2024-09-25 18:07:39,908 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=801010.0, ans=0.1 2024-09-25 18:08:02,077 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.132e+02 1.278e+02 1.376e+02 1.483e+02 1.957e+02, threshold=2.752e+02, percent-clipped=0.0 2024-09-25 18:08:12,361 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=801103.3333333334, ans=0.07 2024-09-25 18:08:23,394 INFO [train.py:1198] (2/4) Epoch 45, batch 250, loss[loss=0.1655, ctc_loss=0.1023, cr_loss=0.3163, over 17037.00 frames. ], tot_loss[loss=0.1896, ctc_loss=0.1217, cr_loss=0.3395, over 2384523.86 frames. ], batch size: 39, lr: 2.65e-03, grad_scale: 32.0 2024-09-25 18:08:28,729 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.95 vs. limit=6.0 2024-09-25 18:09:29,148 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=801336.6666666666, ans=0.0 2024-09-25 18:09:41,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=801383.3333333334, ans=0.125 2024-09-25 18:09:43,106 INFO [train.py:1198] (2/4) Epoch 45, batch 300, loss[loss=0.1791, ctc_loss=0.1167, cr_loss=0.3122, over 17023.00 frames. ], tot_loss[loss=0.1875, ctc_loss=0.1202, cr_loss=0.3368, over 2609757.46 frames. ], batch size: 51, lr: 2.65e-03, grad_scale: 32.0 2024-09-25 18:10:04,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=801430.0, ans=0.2 2024-09-25 18:10:08,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=801430.0, ans=0.125 2024-09-25 18:10:27,968 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=801476.6666666666, ans=0.1 2024-09-25 18:10:27,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=801476.6666666666, ans=0.0 2024-09-25 18:10:48,157 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.282e+02 1.376e+02 1.463e+02 2.041e+02, threshold=2.752e+02, percent-clipped=0.0 2024-09-25 18:11:09,114 INFO [train.py:1198] (2/4) Epoch 45, batch 350, loss[loss=0.1886, ctc_loss=0.1204, cr_loss=0.3414, over 16994.00 frames. ], tot_loss[loss=0.1873, ctc_loss=0.12, cr_loss=0.3366, over 2773897.02 frames. ], batch size: 53, lr: 2.65e-03, grad_scale: 32.0 2024-09-25 18:11:46,846 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.max_abs, batch_count=801710.0, ans=10.0 2024-09-25 18:12:33,382 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module1.whiten, num_groups=1, num_channels=512, metric=6.80 vs. limit=15.0 2024-09-25 18:12:34,306 INFO [train.py:1198] (2/4) Epoch 45, batch 400, loss[loss=0.1989, ctc_loss=0.1277, cr_loss=0.3557, over 17015.00 frames. ], tot_loss[loss=0.1889, ctc_loss=0.121, cr_loss=0.3391, over 2907623.43 frames. ], batch size: 44, lr: 2.65e-03, grad_scale: 32.0 2024-09-25 18:12:42,258 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=801850.0, ans=0.125 2024-09-25 18:12:53,757 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=13.63 vs. limit=15.0 2024-09-25 18:13:11,228 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.31 vs. limit=10.0 2024-09-25 18:13:34,523 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.278e+02 1.369e+02 1.473e+02 1.980e+02, threshold=2.739e+02, percent-clipped=0.0 2024-09-25 18:13:38,844 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.43 vs. limit=15.0 2024-09-25 18:13:51,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=802036.6666666666, ans=0.125 2024-09-25 18:13:53,920 INFO [train.py:1198] (2/4) Epoch 45, batch 450, loss[loss=0.171, ctc_loss=0.1078, cr_loss=0.3159, over 17062.00 frames. ], tot_loss[loss=0.1888, ctc_loss=0.121, cr_loss=0.339, over 3004482.10 frames. ], batch size: 46, lr: 2.65e-03, grad_scale: 16.0 2024-09-25 18:14:08,590 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=802130.0, ans=0.1 2024-09-25 18:14:18,095 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=802130.0, ans=0.125 2024-09-25 18:14:32,718 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=802176.6666666666, ans=0.125 2024-09-25 18:14:45,303 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=802223.3333333334, ans=0.0 2024-09-25 18:14:56,638 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=9.36 vs. limit=12.0 2024-09-25 18:14:59,496 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=802270.0, ans=0.0 2024-09-25 18:15:00,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=802270.0, ans=0.0 2024-09-25 18:15:16,371 INFO [train.py:1198] (2/4) Epoch 45, batch 500, loss[loss=0.1529, ctc_loss=0.09699, cr_loss=0.2794, over 17341.00 frames. ], tot_loss[loss=0.1872, ctc_loss=0.1199, cr_loss=0.3366, over 3085381.50 frames. ], batch size: 48, lr: 2.65e-03, grad_scale: 16.0 2024-09-25 18:15:19,920 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=802316.6666666666, ans=0.1 2024-09-25 18:15:21,606 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=802316.6666666666, ans=0.2 2024-09-25 18:15:51,884 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 18:15:56,564 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=802410.0, ans=0.125 2024-09-25 18:16:22,014 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.73 vs. limit=15.0 2024-09-25 18:16:22,896 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.155e+02 1.274e+02 1.342e+02 1.426e+02 1.767e+02, threshold=2.684e+02, percent-clipped=0.0 2024-09-25 18:16:37,609 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=802503.3333333334, ans=0.125 2024-09-25 18:16:42,131 INFO [train.py:1198] (2/4) Epoch 45, batch 550, loss[loss=0.1857, ctc_loss=0.1185, cr_loss=0.3358, over 17085.00 frames. ], tot_loss[loss=0.1867, ctc_loss=0.1195, cr_loss=0.3364, over 3144385.32 frames. ], batch size: 49, lr: 2.65e-03, grad_scale: 16.0 2024-09-25 18:16:44,113 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=802550.0, ans=0.125 2024-09-25 18:16:59,567 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=802596.6666666666, ans=0.2 2024-09-25 18:18:06,930 INFO [train.py:1198] (2/4) Epoch 45, batch 600, loss[loss=0.2023, ctc_loss=0.132, cr_loss=0.3515, over 17092.00 frames. ], tot_loss[loss=0.1866, ctc_loss=0.1194, cr_loss=0.3358, over 3189124.25 frames. ], batch size: 49, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:18:10,422 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=802783.3333333334, ans=0.0 2024-09-25 18:18:12,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=802783.3333333334, ans=0.0 2024-09-25 18:18:19,958 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=802783.3333333334, ans=0.125 2024-09-25 18:18:21,512 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=802830.0, ans=0.04949747468305833 2024-09-25 18:18:26,175 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=802830.0, ans=0.0 2024-09-25 18:18:27,930 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=802830.0, ans=0.1 2024-09-25 18:19:07,371 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.202e+02 1.313e+02 1.405e+02 1.513e+02 2.621e+02, threshold=2.810e+02, percent-clipped=0.0 2024-09-25 18:19:18,238 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=7.45 vs. limit=8.0 2024-09-25 18:19:26,370 INFO [train.py:1198] (2/4) Epoch 45, batch 650, loss[loss=0.1686, ctc_loss=0.1064, cr_loss=0.3108, over 17302.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1186, cr_loss=0.334, over 3236083.93 frames. ], batch size: 51, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:19:45,759 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=803063.3333333334, ans=0.125 2024-09-25 18:19:54,502 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=5.34 vs. limit=15.0 2024-09-25 18:20:00,212 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=803110.0, ans=0.125 2024-09-25 18:20:04,943 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=803110.0, ans=0.125 2024-09-25 18:20:20,579 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 18:20:52,180 INFO [train.py:1198] (2/4) Epoch 45, batch 700, loss[loss=0.1639, ctc_loss=0.1058, cr_loss=0.2905, over 17035.00 frames. ], tot_loss[loss=0.1865, ctc_loss=0.1195, cr_loss=0.335, over 3265148.88 frames. ], batch size: 39, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:21:15,153 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=803296.6666666666, ans=0.125 2024-09-25 18:21:26,092 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=803343.3333333334, ans=0.1 2024-09-25 18:21:35,489 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=803343.3333333334, ans=0.1 2024-09-25 18:21:55,975 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.139e+02 1.305e+02 1.379e+02 1.482e+02 2.055e+02, threshold=2.759e+02, percent-clipped=0.0 2024-09-25 18:22:02,224 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=803436.6666666666, ans=0.125 2024-09-25 18:22:02,227 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=803436.6666666666, ans=0.0 2024-09-25 18:22:17,190 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.25 vs. limit=22.5 2024-09-25 18:22:18,091 INFO [train.py:1198] (2/4) Epoch 45, batch 750, loss[loss=0.2417, ctc_loss=0.1592, cr_loss=0.4126, over 15205.00 frames. ], tot_loss[loss=0.1867, ctc_loss=0.1197, cr_loss=0.3354, over 3291667.46 frames. ], batch size: 89, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:22:26,196 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=803483.3333333334, ans=0.015 2024-09-25 18:22:54,344 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=10.69 vs. limit=15.0 2024-09-25 18:23:32,508 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 18:23:37,513 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=803716.6666666666, ans=0.1 2024-09-25 18:23:38,652 INFO [train.py:1198] (2/4) Epoch 45, batch 800, loss[loss=0.1801, ctc_loss=0.113, cr_loss=0.3356, over 17015.00 frames. ], tot_loss[loss=0.1873, ctc_loss=0.12, cr_loss=0.3362, over 3307712.37 frames. ], batch size: 44, lr: 2.64e-03, grad_scale: 32.0 2024-09-25 18:24:15,521 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=803810.0, ans=0.07 2024-09-25 18:24:24,192 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=7.28 vs. limit=15.0 2024-09-25 18:24:29,601 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module1.whiten, num_groups=1, num_channels=192, metric=7.65 vs. limit=15.0 2024-09-25 18:24:33,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=803856.6666666666, ans=0.0 2024-09-25 18:24:39,375 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.162e+02 1.279e+02 1.371e+02 1.474e+02 1.917e+02, threshold=2.742e+02, percent-clipped=0.0 2024-09-25 18:24:41,393 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=803903.3333333334, ans=0.025 2024-09-25 18:24:47,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=803903.3333333334, ans=0.1 2024-09-25 18:24:47,874 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=803903.3333333334, ans=0.1 2024-09-25 18:24:58,298 INFO [train.py:1198] (2/4) Epoch 45, batch 850, loss[loss=0.2102, ctc_loss=0.1367, cr_loss=0.3678, over 17303.00 frames. ], tot_loss[loss=0.1871, ctc_loss=0.12, cr_loss=0.3358, over 3314275.54 frames. ], batch size: 51, lr: 2.64e-03, grad_scale: 32.0 2024-09-25 18:25:06,580 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=803950.0, ans=0.125 2024-09-25 18:25:09,067 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=7.22 vs. limit=8.0 2024-09-25 18:25:27,245 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.42 vs. limit=15.0 2024-09-25 18:25:40,715 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=804043.3333333334, ans=0.125 2024-09-25 18:25:40,718 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=804043.3333333334, ans=0.125 2024-09-25 18:26:00,144 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=804090.0, ans=0.125 2024-09-25 18:26:02,398 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=4.91 vs. limit=15.0 2024-09-25 18:26:26,529 INFO [train.py:1198] (2/4) Epoch 45, batch 900, loss[loss=0.2147, ctc_loss=0.1386, cr_loss=0.3807, over 17042.00 frames. ], tot_loss[loss=0.1873, ctc_loss=0.12, cr_loss=0.3364, over 3328710.37 frames. ], batch size: 52, lr: 2.64e-03, grad_scale: 32.0 2024-09-25 18:26:26,775 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=804183.3333333334, ans=0.1 2024-09-25 18:26:35,475 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.21 vs. limit=12.0 2024-09-25 18:26:38,020 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=804183.3333333334, ans=0.125 2024-09-25 18:27:16,417 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=9.98 vs. limit=15.0 2024-09-25 18:27:18,909 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=804323.3333333334, ans=0.0 2024-09-25 18:27:20,510 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=804323.3333333334, ans=0.125 2024-09-25 18:27:29,591 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.180e+02 1.299e+02 1.367e+02 1.436e+02 2.740e+02, threshold=2.735e+02, percent-clipped=0.0 2024-09-25 18:27:36,415 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=804370.0, ans=0.125 2024-09-25 18:27:39,557 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=804370.0, ans=0.1 2024-09-25 18:27:41,137 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=804370.0, ans=0.0 2024-09-25 18:27:41,151 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=804370.0, ans=0.1 2024-09-25 18:27:42,553 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=804370.0, ans=0.1 2024-09-25 18:27:47,236 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=804416.6666666666, ans=0.125 2024-09-25 18:27:48,683 INFO [train.py:1198] (2/4) Epoch 45, batch 950, loss[loss=0.1719, ctc_loss=0.1053, cr_loss=0.3329, over 17319.00 frames. ], tot_loss[loss=0.1885, ctc_loss=0.1209, cr_loss=0.3381, over 3325331.51 frames. ], batch size: 46, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:27:50,642 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=804416.6666666666, ans=0.0 2024-09-25 18:27:58,493 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=804416.6666666666, ans=0.125 2024-09-25 18:27:59,999 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=804416.6666666666, ans=0.0 2024-09-25 18:28:15,064 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=10.28 vs. limit=22.5 2024-09-25 18:28:15,608 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=804463.3333333334, ans=0.125 2024-09-25 18:28:18,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=804510.0, ans=0.09899494936611666 2024-09-25 18:28:20,387 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=804510.0, ans=0.125 2024-09-25 18:28:33,259 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=804510.0, ans=0.125 2024-09-25 18:28:52,040 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.min_positive, batch_count=804603.3333333334, ans=0.025 2024-09-25 18:29:04,945 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=804603.3333333334, ans=0.125 2024-09-25 18:29:07,945 INFO [train.py:1198] (2/4) Epoch 45, batch 1000, loss[loss=0.1739, ctc_loss=0.1082, cr_loss=0.3281, over 17007.00 frames. ], tot_loss[loss=0.1879, ctc_loss=0.1204, cr_loss=0.3378, over 3337111.95 frames. ], batch size: 44, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:29:17,985 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=804650.0, ans=0.125 2024-09-25 18:29:21,207 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=804650.0, ans=0.125 2024-09-25 18:29:25,931 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=804696.6666666666, ans=0.125 2024-09-25 18:29:51,487 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=804743.3333333334, ans=0.0 2024-09-25 18:29:59,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=804790.0, ans=0.2 2024-09-25 18:30:01,057 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=804790.0, ans=0.0 2024-09-25 18:30:13,227 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.285e+02 1.391e+02 1.486e+02 1.776e+02, threshold=2.782e+02, percent-clipped=0.0 2024-09-25 18:30:24,439 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=804836.6666666666, ans=0.125 2024-09-25 18:30:28,416 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.83 vs. limit=10.0 2024-09-25 18:30:33,440 INFO [train.py:1198] (2/4) Epoch 45, batch 1050, loss[loss=0.2088, ctc_loss=0.1368, cr_loss=0.3601, over 16044.00 frames. ], tot_loss[loss=0.188, ctc_loss=0.1205, cr_loss=0.3378, over 3336145.56 frames. ], batch size: 74, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:31:19,315 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=804976.6666666666, ans=0.07 2024-09-25 18:31:24,047 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=805023.3333333334, ans=0.2 2024-09-25 18:31:28,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=805023.3333333334, ans=0.2 2024-09-25 18:31:52,260 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=805070.0, ans=0.0 2024-09-25 18:31:54,394 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1.whitening_limit, batch_count=805070.0, ans=10.0 2024-09-25 18:31:55,765 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=805070.0, ans=0.0 2024-09-25 18:31:58,555 INFO [train.py:1198] (2/4) Epoch 45, batch 1100, loss[loss=0.1761, ctc_loss=0.1136, cr_loss=0.3127, over 17273.00 frames. ], tot_loss[loss=0.1874, ctc_loss=0.12, cr_loss=0.3368, over 3348181.29 frames. ], batch size: 42, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:32:08,893 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=4.91 vs. limit=12.0 2024-09-25 18:32:09,899 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=805116.6666666666, ans=0.1 2024-09-25 18:32:44,972 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=805256.6666666666, ans=0.0 2024-09-25 18:33:00,831 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.150e+02 1.299e+02 1.379e+02 1.493e+02 1.866e+02, threshold=2.759e+02, percent-clipped=0.0 2024-09-25 18:33:04,249 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=805303.3333333334, ans=0.2 2024-09-25 18:33:18,415 INFO [train.py:1198] (2/4) Epoch 45, batch 1150, loss[loss=0.1769, ctc_loss=0.112, cr_loss=0.3242, over 17295.00 frames. ], tot_loss[loss=0.1876, ctc_loss=0.1202, cr_loss=0.3369, over 3349320.44 frames. ], batch size: 42, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:33:42,768 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=805396.6666666666, ans=0.0 2024-09-25 18:33:53,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=805443.3333333334, ans=0.0 2024-09-25 18:34:05,764 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=5.85 vs. limit=12.0 2024-09-25 18:34:14,123 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=4.75 vs. limit=15.0 2024-09-25 18:34:27,955 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=805536.6666666666, ans=0.0 2024-09-25 18:34:38,826 INFO [train.py:1198] (2/4) Epoch 45, batch 1200, loss[loss=0.1963, ctc_loss=0.1253, cr_loss=0.3546, over 17361.00 frames. ], tot_loss[loss=0.1884, ctc_loss=0.1207, cr_loss=0.3381, over 3349740.64 frames. ], batch size: 48, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:35:29,018 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=805676.6666666666, ans=0.2 2024-09-25 18:35:29,798 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn1.whiten, num_groups=1, num_channels=192, metric=9.60 vs. limit=22.5 2024-09-25 18:35:35,349 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=805723.3333333334, ans=0.0 2024-09-25 18:35:43,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.min_positive, batch_count=805723.3333333334, ans=0.05 2024-09-25 18:35:47,624 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.146e+02 1.307e+02 1.404e+02 1.482e+02 2.313e+02, threshold=2.807e+02, percent-clipped=0.0 2024-09-25 18:35:47,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=805770.0, ans=0.125 2024-09-25 18:36:05,113 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=805816.6666666666, ans=0.2 2024-09-25 18:36:06,487 INFO [train.py:1198] (2/4) Epoch 45, batch 1250, loss[loss=0.1714, ctc_loss=0.1093, cr_loss=0.3107, over 17071.00 frames. ], tot_loss[loss=0.1887, ctc_loss=0.1209, cr_loss=0.3387, over 3340344.44 frames. ], batch size: 39, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:36:10,012 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.min_abs, batch_count=805816.6666666666, ans=0.5 2024-09-25 18:36:21,428 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=9.11 vs. limit=15.0 2024-09-25 18:36:34,935 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=8.61 vs. limit=15.0 2024-09-25 18:37:13,412 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=806003.3333333334, ans=0.0 2024-09-25 18:37:15,117 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=806003.3333333334, ans=0.125 2024-09-25 18:37:19,737 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=806003.3333333334, ans=0.05 2024-09-25 18:37:27,833 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=806050.0, ans=0.125 2024-09-25 18:37:29,080 INFO [train.py:1198] (2/4) Epoch 45, batch 1300, loss[loss=0.2012, ctc_loss=0.1296, cr_loss=0.3582, over 16991.00 frames. ], tot_loss[loss=0.188, ctc_loss=0.1205, cr_loss=0.3377, over 3335290.04 frames. ], batch size: 56, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:37:43,847 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=806096.6666666666, ans=0.1 2024-09-25 18:37:50,053 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=806096.6666666666, ans=0.0 2024-09-25 18:37:51,628 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=806096.6666666666, ans=0.0 2024-09-25 18:38:32,668 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.116e+02 1.317e+02 1.369e+02 1.447e+02 1.900e+02, threshold=2.738e+02, percent-clipped=0.0 2024-09-25 18:38:34,502 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=806236.6666666666, ans=0.5 2024-09-25 18:38:39,399 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=806236.6666666666, ans=0.0 2024-09-25 18:38:48,676 INFO [train.py:1198] (2/4) Epoch 45, batch 1350, loss[loss=0.2149, ctc_loss=0.1447, cr_loss=0.3508, over 14747.00 frames. ], tot_loss[loss=0.1881, ctc_loss=0.1206, cr_loss=0.3374, over 3339043.78 frames. ], batch size: 89, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:38:49,323 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten.whitening_limit, batch_count=806283.3333333334, ans=15.0 2024-09-25 18:38:57,014 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=806283.3333333334, ans=0.0 2024-09-25 18:38:59,948 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=806283.3333333334, ans=0.025 2024-09-25 18:39:08,310 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=806330.0, ans=0.125 2024-09-25 18:39:16,233 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=806330.0, ans=0.125 2024-09-25 18:39:18,482 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=7.24 vs. limit=15.0 2024-09-25 18:39:34,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=806376.6666666666, ans=0.125 2024-09-25 18:40:14,349 INFO [train.py:1198] (2/4) Epoch 45, batch 1400, loss[loss=0.18, ctc_loss=0.1149, cr_loss=0.3255, over 17147.00 frames. ], tot_loss[loss=0.1884, ctc_loss=0.1208, cr_loss=0.3382, over 3351344.42 frames. ], batch size: 48, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:40:43,527 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=806563.3333333334, ans=0.2 2024-09-25 18:41:05,737 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=11.01 vs. limit=12.0 2024-09-25 18:41:21,112 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.028e+02 1.293e+02 1.370e+02 1.466e+02 2.131e+02, threshold=2.739e+02, percent-clipped=0.0 2024-09-25 18:41:37,441 INFO [train.py:1198] (2/4) Epoch 45, batch 1450, loss[loss=0.2027, ctc_loss=0.1279, cr_loss=0.3739, over 17047.00 frames. ], tot_loss[loss=0.1882, ctc_loss=0.1206, cr_loss=0.3382, over 3359542.61 frames. ], batch size: 52, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:41:51,510 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=806750.0, ans=0.0 2024-09-25 18:42:14,028 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=806843.3333333334, ans=0.025 2024-09-25 18:42:39,852 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=806890.0, ans=0.125 2024-09-25 18:42:44,734 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=806936.6666666666, ans=0.0 2024-09-25 18:42:46,790 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.63 vs. limit=15.0 2024-09-25 18:42:47,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=806936.6666666666, ans=0.025 2024-09-25 18:43:00,438 INFO [train.py:1198] (2/4) Epoch 45, batch 1500, loss[loss=0.2283, ctc_loss=0.1582, cr_loss=0.3508, over 11627.00 frames. ], tot_loss[loss=0.1874, ctc_loss=0.12, cr_loss=0.3369, over 3360129.86 frames. ], batch size: 123, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:43:05,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=806983.3333333334, ans=0.2 2024-09-25 18:43:10,205 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=806983.3333333334, ans=0.0 2024-09-25 18:43:18,354 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=807030.0, ans=0.2 2024-09-25 18:43:23,018 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=807030.0, ans=0.025 2024-09-25 18:43:45,585 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=807076.6666666666, ans=0.125 2024-09-25 18:44:04,286 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.164e+02 1.307e+02 1.374e+02 1.496e+02 1.969e+02, threshold=2.747e+02, percent-clipped=0.0 2024-09-25 18:44:08,583 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.79 vs. limit=15.0 2024-09-25 18:44:19,334 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=807216.6666666666, ans=0.1 2024-09-25 18:44:20,506 INFO [train.py:1198] (2/4) Epoch 45, batch 1550, loss[loss=0.1964, ctc_loss=0.1253, cr_loss=0.3554, over 17001.00 frames. ], tot_loss[loss=0.1866, ctc_loss=0.1194, cr_loss=0.336, over 3366761.32 frames. ], batch size: 53, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:44:46,270 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=807263.3333333334, ans=0.125 2024-09-25 18:45:13,982 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 18:45:37,844 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=807403.3333333334, ans=0.125 2024-09-25 18:45:45,592 INFO [train.py:1198] (2/4) Epoch 45, batch 1600, loss[loss=0.2042, ctc_loss=0.1336, cr_loss=0.353, over 17006.00 frames. ], tot_loss[loss=0.1871, ctc_loss=0.1198, cr_loss=0.3365, over 3367245.40 frames. ], batch size: 53, lr: 2.64e-03, grad_scale: 32.0 2024-09-25 18:45:57,218 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.02 vs. limit=15.0 2024-09-25 18:46:50,216 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=807590.0, ans=0.125 2024-09-25 18:46:56,353 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.165e+02 1.296e+02 1.378e+02 1.483e+02 2.434e+02, threshold=2.757e+02, percent-clipped=0.0 2024-09-25 18:47:04,681 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=807636.6666666666, ans=0.125 2024-09-25 18:47:10,853 INFO [train.py:1198] (2/4) Epoch 45, batch 1650, loss[loss=0.2105, ctc_loss=0.1383, cr_loss=0.3608, over 16397.00 frames. ], tot_loss[loss=0.1864, ctc_loss=0.1193, cr_loss=0.3357, over 3373662.74 frames. ], batch size: 66, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:47:28,835 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=807730.0, ans=0.125 2024-09-25 18:47:36,918 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=807730.0, ans=0.125 2024-09-25 18:48:12,005 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=807823.3333333334, ans=0.0 2024-09-25 18:48:31,083 INFO [train.py:1198] (2/4) Epoch 45, batch 1700, loss[loss=0.2025, ctc_loss=0.1339, cr_loss=0.3431, over 17025.00 frames. ], tot_loss[loss=0.1866, ctc_loss=0.1195, cr_loss=0.3354, over 3378688.57 frames. ], batch size: 52, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:49:01,768 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=808010.0, ans=0.125 2024-09-25 18:49:08,132 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=808010.0, ans=0.09899494936611666 2024-09-25 18:49:15,884 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=808010.0, ans=0.035 2024-09-25 18:49:22,153 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=808056.6666666666, ans=0.125 2024-09-25 18:49:34,899 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=808103.3333333334, ans=0.125 2024-09-25 18:49:36,102 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.156e+02 1.276e+02 1.351e+02 1.443e+02 2.943e+02, threshold=2.702e+02, percent-clipped=1.0 2024-09-25 18:49:50,471 INFO [train.py:1198] (2/4) Epoch 45, batch 1750, loss[loss=0.1898, ctc_loss=0.1208, cr_loss=0.345, over 17247.00 frames. ], tot_loss[loss=0.1876, ctc_loss=0.1202, cr_loss=0.3369, over 3371066.71 frames. ], batch size: 44, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:49:59,619 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.min_positive, batch_count=808150.0, ans=0.025 2024-09-25 18:50:01,202 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=808150.0, ans=0.0 2024-09-25 18:50:22,821 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=808196.6666666666, ans=0.1 2024-09-25 18:50:41,391 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.48 vs. limit=15.0 2024-09-25 18:51:01,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=808336.6666666666, ans=0.0 2024-09-25 18:51:18,075 INFO [train.py:1198] (2/4) Epoch 45, batch 1800, loss[loss=0.1767, ctc_loss=0.1131, cr_loss=0.3179, over 17276.00 frames. ], tot_loss[loss=0.1871, ctc_loss=0.1198, cr_loss=0.3361, over 3366725.73 frames. ], batch size: 44, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:51:18,334 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=808383.3333333334, ans=0.125 2024-09-25 18:51:25,270 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.04 vs. limit=6.0 2024-09-25 18:51:38,854 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=808430.0, ans=0.125 2024-09-25 18:51:42,188 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=12.79 vs. limit=15.0 2024-09-25 18:52:18,552 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=808523.3333333334, ans=0.125 2024-09-25 18:52:21,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=808523.3333333334, ans=0.125 2024-09-25 18:52:24,841 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=808570.0, ans=0.0 2024-09-25 18:52:26,148 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.286e+02 1.376e+02 1.494e+02 2.508e+02, threshold=2.752e+02, percent-clipped=0.0 2024-09-25 18:52:26,408 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=808570.0, ans=0.0 2024-09-25 18:52:36,106 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=808570.0, ans=0.125 2024-09-25 18:52:40,619 INFO [train.py:1198] (2/4) Epoch 45, batch 1850, loss[loss=0.1631, ctc_loss=0.1034, cr_loss=0.2987, over 17355.00 frames. ], tot_loss[loss=0.1877, ctc_loss=0.1203, cr_loss=0.337, over 3365868.86 frames. ], batch size: 48, lr: 2.64e-03, grad_scale: 16.0 2024-09-25 18:52:44,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=808616.6666666666, ans=0.0 2024-09-25 18:52:48,911 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=808616.6666666666, ans=0.2 2024-09-25 18:52:48,996 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=808616.6666666666, ans=0.0 2024-09-25 18:52:50,569 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=808616.6666666666, ans=0.0 2024-09-25 18:53:00,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=808663.3333333334, ans=0.05 2024-09-25 18:53:06,567 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=808663.3333333334, ans=0.125 2024-09-25 18:53:08,712 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=14.45 vs. limit=22.5 2024-09-25 18:53:27,948 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.72 vs. limit=15.0 2024-09-25 18:53:49,766 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=808803.3333333334, ans=0.1 2024-09-25 18:53:49,892 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=808803.3333333334, ans=0.0 2024-09-25 18:54:00,631 INFO [train.py:1198] (2/4) Epoch 45, batch 1900, loss[loss=0.2104, ctc_loss=0.137, cr_loss=0.3669, over 17220.00 frames. ], tot_loss[loss=0.1883, ctc_loss=0.1208, cr_loss=0.3376, over 3357630.87 frames. ], batch size: 55, lr: 2.63e-03, grad_scale: 16.0 2024-09-25 18:54:01,034 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=808850.0, ans=0.125 2024-09-25 18:54:08,879 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=808850.0, ans=0.125 2024-09-25 18:54:10,483 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=808850.0, ans=0.09899494936611666 2024-09-25 18:54:20,178 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=808896.6666666666, ans=0.125 2024-09-25 18:54:36,210 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=808943.3333333334, ans=0.125 2024-09-25 18:55:12,117 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.309e+02 1.390e+02 1.516e+02 1.973e+02, threshold=2.779e+02, percent-clipped=0.0 2024-09-25 18:55:15,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=809036.6666666666, ans=0.125 2024-09-25 18:55:23,620 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=809036.6666666666, ans=0.125 2024-09-25 18:55:26,420 INFO [train.py:1198] (2/4) Epoch 45, batch 1950, loss[loss=0.2156, ctc_loss=0.141, cr_loss=0.3732, over 16601.00 frames. ], tot_loss[loss=0.1895, ctc_loss=0.1216, cr_loss=0.3394, over 3356091.91 frames. ], batch size: 66, lr: 2.63e-03, grad_scale: 16.0 2024-09-25 18:56:19,540 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=809223.3333333334, ans=0.0 2024-09-25 18:56:41,019 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=809270.0, ans=0.2 2024-09-25 18:56:52,061 INFO [train.py:1198] (2/4) Epoch 45, batch 2000, loss[loss=0.2044, ctc_loss=0.1324, cr_loss=0.3597, over 17016.00 frames. ], tot_loss[loss=0.1881, ctc_loss=0.1206, cr_loss=0.3373, over 3357312.55 frames. ], batch size: 56, lr: 2.63e-03, grad_scale: 32.0 2024-09-25 18:57:03,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=809316.6666666666, ans=0.125 2024-09-25 18:57:03,493 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=809316.6666666666, ans=0.125 2024-09-25 18:57:09,936 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=809363.3333333334, ans=0.2 2024-09-25 18:57:46,814 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=809456.6666666666, ans=0.125 2024-09-25 18:57:49,897 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=809456.6666666666, ans=0.0 2024-09-25 18:57:54,650 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=809503.3333333334, ans=0.0 2024-09-25 18:57:57,598 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.152e+02 1.313e+02 1.395e+02 1.516e+02 2.306e+02, threshold=2.789e+02, percent-clipped=0.0 2024-09-25 18:57:58,462 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.07 vs. limit=6.0 2024-09-25 18:58:04,775 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.58 vs. limit=15.0 2024-09-25 18:58:05,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=809503.3333333334, ans=0.0 2024-09-25 18:58:10,842 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=809550.0, ans=0.125 2024-09-25 18:58:12,071 INFO [train.py:1198] (2/4) Epoch 45, batch 2050, loss[loss=0.2108, ctc_loss=0.1377, cr_loss=0.365, over 16012.00 frames. ], tot_loss[loss=0.1881, ctc_loss=0.1206, cr_loss=0.3373, over 3358155.28 frames. ], batch size: 74, lr: 2.63e-03, grad_scale: 32.0 2024-09-25 18:58:30,028 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=809596.6666666666, ans=0.0 2024-09-25 18:58:58,675 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=809690.0, ans=0.0 2024-09-25 18:59:14,821 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=809736.6666666666, ans=0.0 2024-09-25 18:59:32,218 INFO [train.py:1198] (2/4) Epoch 45, batch 2100, loss[loss=0.1588, ctc_loss=0.09894, cr_loss=0.2994, over 16951.00 frames. ], tot_loss[loss=0.1877, ctc_loss=0.1203, cr_loss=0.3369, over 3347719.70 frames. ], batch size: 42, lr: 2.63e-03, grad_scale: 32.0 2024-09-25 18:59:35,633 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=809783.3333333334, ans=0.1 2024-09-25 18:59:55,673 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=809830.0, ans=0.04949747468305833 2024-09-25 19:00:07,107 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=809876.6666666666, ans=0.0 2024-09-25 19:00:07,129 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=809876.6666666666, ans=0.125 2024-09-25 19:00:11,774 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=809876.6666666666, ans=0.125 2024-09-25 19:00:26,218 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=809923.3333333334, ans=0.025 2024-09-25 19:00:40,213 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.255e+02 1.338e+02 1.412e+02 1.754e+02, threshold=2.676e+02, percent-clipped=0.0 2024-09-25 19:00:42,157 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=809970.0, ans=0.0 2024-09-25 19:00:42,163 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=809970.0, ans=0.125 2024-09-25 19:00:43,555 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=809970.0, ans=0.0 2024-09-25 19:00:45,355 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=809970.0, ans=0.04949747468305833 2024-09-25 19:00:57,179 INFO [train.py:1198] (2/4) Epoch 45, batch 2150, loss[loss=0.1945, ctc_loss=0.1246, cr_loss=0.3494, over 16074.00 frames. ], tot_loss[loss=0.1888, ctc_loss=0.1211, cr_loss=0.3382, over 3341755.06 frames. ], batch size: 74, lr: 2.63e-03, grad_scale: 16.0 2024-09-25 19:01:31,911 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=810110.0, ans=0.125 2024-09-25 19:01:35,198 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=810110.0, ans=0.1 2024-09-25 19:02:18,754 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=810250.0, ans=0.1 2024-09-25 19:02:20,041 INFO [train.py:1198] (2/4) Epoch 45, batch 2200, loss[loss=0.1584, ctc_loss=0.09896, cr_loss=0.2974, over 17192.00 frames. ], tot_loss[loss=0.188, ctc_loss=0.1205, cr_loss=0.3376, over 3357436.46 frames. ], batch size: 41, lr: 2.63e-03, grad_scale: 16.0 2024-09-25 19:02:25,142 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=810250.0, ans=0.125 2024-09-25 19:02:26,793 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=810250.0, ans=0.0 2024-09-25 19:02:56,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=810343.3333333334, ans=0.1 2024-09-25 19:03:27,149 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.170e+02 1.297e+02 1.368e+02 1.487e+02 2.152e+02, threshold=2.736e+02, percent-clipped=0.0 2024-09-25 19:03:29,020 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=810436.6666666666, ans=0.09899494936611666 2024-09-25 19:03:40,105 INFO [train.py:1198] (2/4) Epoch 45, batch 2250, loss[loss=0.1586, ctc_loss=0.09963, cr_loss=0.2949, over 17203.00 frames. ], tot_loss[loss=0.1874, ctc_loss=0.12, cr_loss=0.3368, over 3363283.09 frames. ], batch size: 41, lr: 2.63e-03, grad_scale: 16.0 2024-09-25 19:03:56,415 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=810530.0, ans=0.125 2024-09-25 19:04:22,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=810576.6666666666, ans=0.1 2024-09-25 19:05:01,582 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=810716.6666666666, ans=0.1 2024-09-25 19:05:02,934 INFO [train.py:1198] (2/4) Epoch 45, batch 2300, loss[loss=0.194, ctc_loss=0.1222, cr_loss=0.3587, over 17299.00 frames. ], tot_loss[loss=0.1882, ctc_loss=0.1207, cr_loss=0.3375, over 3369188.49 frames. ], batch size: 51, lr: 2.63e-03, grad_scale: 16.0 2024-09-25 19:05:06,394 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 19:05:18,275 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=13.44 vs. limit=15.0 2024-09-25 19:05:23,612 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=810763.3333333334, ans=0.1 2024-09-25 19:05:30,089 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=810763.3333333334, ans=0.125 2024-09-25 19:05:33,232 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=810810.0, ans=0.125 2024-09-25 19:05:33,705 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.38 vs. limit=10.0 2024-09-25 19:05:34,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=810810.0, ans=0.2 2024-09-25 19:05:51,758 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=810856.6666666666, ans=0.125 2024-09-25 19:05:51,854 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=810856.6666666666, ans=0.125 2024-09-25 19:06:05,015 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=5.65 vs. limit=15.0 2024-09-25 19:06:12,003 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.109e+02 1.306e+02 1.375e+02 1.455e+02 1.967e+02, threshold=2.751e+02, percent-clipped=0.0 2024-09-25 19:06:24,355 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=810903.3333333334, ans=0.05 2024-09-25 19:06:24,500 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=810903.3333333334, ans=0.125 2024-09-25 19:06:27,348 INFO [train.py:1198] (2/4) Epoch 45, batch 2350, loss[loss=0.2102, ctc_loss=0.1362, cr_loss=0.3702, over 17002.00 frames. ], tot_loss[loss=0.1888, ctc_loss=0.1211, cr_loss=0.3386, over 3354853.62 frames. ], batch size: 51, lr: 2.63e-03, grad_scale: 16.0 2024-09-25 19:06:29,149 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=810950.0, ans=0.0 2024-09-25 19:06:43,425 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=810996.6666666666, ans=0.125 2024-09-25 19:06:44,974 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=810996.6666666666, ans=0.035 2024-09-25 19:06:45,159 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=810996.6666666666, ans=0.125 2024-09-25 19:06:54,700 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=810996.6666666666, ans=0.2 2024-09-25 19:07:42,461 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=811136.6666666666, ans=0.125 2024-09-25 19:07:46,977 INFO [train.py:1198] (2/4) Epoch 45, batch 2400, loss[loss=0.1609, ctc_loss=0.09848, cr_loss=0.3119, over 17060.00 frames. ], tot_loss[loss=0.1872, ctc_loss=0.1199, cr_loss=0.3362, over 3355593.35 frames. ], batch size: 39, lr: 2.63e-03, grad_scale: 16.0 2024-09-25 19:07:49,021 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=811183.3333333334, ans=0.1 2024-09-25 19:07:55,380 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=811183.3333333334, ans=0.125 2024-09-25 19:08:06,612 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.min_positive, batch_count=811230.0, ans=0.025 2024-09-25 19:08:16,316 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=811230.0, ans=0.07 2024-09-25 19:08:19,410 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=811276.6666666666, ans=0.125 2024-09-25 19:08:19,494 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=811276.6666666666, ans=0.0 2024-09-25 19:08:21,579 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=15.54 vs. limit=22.5 2024-09-25 19:08:26,614 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.82 vs. limit=12.0 2024-09-25 19:08:51,161 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=811370.0, ans=0.0 2024-09-25 19:08:55,592 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.054e+02 1.292e+02 1.391e+02 1.494e+02 2.070e+02, threshold=2.781e+02, percent-clipped=0.0 2024-09-25 19:09:07,045 INFO [train.py:1198] (2/4) Epoch 45, batch 2450, loss[loss=0.191, ctc_loss=0.1218, cr_loss=0.3462, over 17012.00 frames. ], tot_loss[loss=0.1873, ctc_loss=0.12, cr_loss=0.3367, over 3358760.63 frames. ], batch size: 51, lr: 2.63e-03, grad_scale: 16.0 2024-09-25 19:09:13,579 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=811416.6666666666, ans=0.2 2024-09-25 19:09:23,287 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=811463.3333333334, ans=0.125 2024-09-25 19:09:25,623 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=12.29 vs. limit=22.5 2024-09-25 19:09:54,776 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=5.58 vs. limit=12.0 2024-09-25 19:10:02,159 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=811556.6666666666, ans=0.0 2024-09-25 19:10:08,612 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=811556.6666666666, ans=0.0 2024-09-25 19:10:18,249 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=811603.3333333334, ans=0.2 2024-09-25 19:10:29,487 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=811603.3333333334, ans=0.125 2024-09-25 19:10:32,451 INFO [train.py:1198] (2/4) Epoch 45, batch 2500, loss[loss=0.2079, ctc_loss=0.1375, cr_loss=0.3519, over 17296.00 frames. ], tot_loss[loss=0.1878, ctc_loss=0.1203, cr_loss=0.3371, over 3363955.02 frames. ], batch size: 51, lr: 2.63e-03, grad_scale: 16.0 2024-09-25 19:10:37,904 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=18.22 vs. limit=22.5 2024-09-25 19:11:09,689 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.25 vs. limit=6.0 2024-09-25 19:11:15,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=811743.3333333334, ans=0.125 2024-09-25 19:11:22,840 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=811743.3333333334, ans=0.2 2024-09-25 19:11:23,333 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module1.whiten, num_groups=1, num_channels=512, metric=7.16 vs. limit=15.0 2024-09-25 19:11:46,431 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.195e+02 1.307e+02 1.405e+02 1.520e+02 3.997e+02, threshold=2.811e+02, percent-clipped=1.0 2024-09-25 19:11:48,207 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=811836.6666666666, ans=0.125 2024-09-25 19:11:54,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=811836.6666666666, ans=0.0 2024-09-25 19:11:57,470 INFO [train.py:1198] (2/4) Epoch 45, batch 2550, loss[loss=0.1768, ctc_loss=0.112, cr_loss=0.324, over 16320.00 frames. ], tot_loss[loss=0.1884, ctc_loss=0.1208, cr_loss=0.338, over 3359142.71 frames. ], batch size: 36, lr: 2.63e-03, grad_scale: 16.0 2024-09-25 19:12:02,299 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=811883.3333333334, ans=0.125 2024-09-25 19:12:11,849 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=811930.0, ans=0.125 2024-09-25 19:12:31,958 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.51 vs. limit=15.0 2024-09-25 19:12:39,894 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=811976.6666666666, ans=0.125 2024-09-25 19:13:18,002 INFO [train.py:1198] (2/4) Epoch 45, batch 2600, loss[loss=0.2084, ctc_loss=0.1384, cr_loss=0.3501, over 17043.00 frames. ], tot_loss[loss=0.1883, ctc_loss=0.1208, cr_loss=0.3377, over 3364002.44 frames. ], batch size: 52, lr: 2.63e-03, grad_scale: 16.0 2024-09-25 19:13:21,801 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=812116.6666666666, ans=0.1 2024-09-25 19:13:23,486 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=12.22 vs. limit=12.0 2024-09-25 19:13:58,373 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=812210.0, ans=0.125 2024-09-25 19:14:26,619 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.139e+02 1.298e+02 1.365e+02 1.508e+02 3.211e+02, threshold=2.730e+02, percent-clipped=1.0 2024-09-25 19:14:37,603 INFO [train.py:1198] (2/4) Epoch 45, batch 2650, loss[loss=0.1931, ctc_loss=0.1239, cr_loss=0.3456, over 17237.00 frames. ], tot_loss[loss=0.1886, ctc_loss=0.121, cr_loss=0.3379, over 3359089.58 frames. ], batch size: 55, lr: 2.63e-03, grad_scale: 16.0 2024-09-25 19:15:02,369 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=812396.6666666666, ans=0.1 2024-09-25 19:15:32,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=812490.0, ans=0.125 2024-09-25 19:15:44,918 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=812490.0, ans=0.1 2024-09-25 19:16:01,520 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.48 vs. limit=6.0 2024-09-25 19:16:05,228 INFO [train.py:1198] (2/4) Epoch 45, batch 2700, loss[loss=0.2076, ctc_loss=0.1322, cr_loss=0.377, over 17228.00 frames. ], tot_loss[loss=0.1885, ctc_loss=0.1209, cr_loss=0.3379, over 3351824.45 frames. ], batch size: 50, lr: 2.63e-03, grad_scale: 16.0 2024-09-25 19:16:07,735 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys.whitening_limit, batch_count=812583.3333333334, ans=6.0 2024-09-25 19:16:32,214 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=812630.0, ans=0.1 2024-09-25 19:17:12,501 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=12.35 vs. limit=15.0 2024-09-25 19:17:16,506 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.145e+02 1.332e+02 1.395e+02 1.480e+02 1.969e+02, threshold=2.789e+02, percent-clipped=0.0 2024-09-25 19:17:17,771 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.76 vs. limit=5.0 2024-09-25 19:17:23,222 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=812770.0, ans=0.0 2024-09-25 19:17:27,064 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=3.84 vs. limit=15.0 2024-09-25 19:17:27,699 INFO [train.py:1198] (2/4) Epoch 45, batch 2750, loss[loss=0.1803, ctc_loss=0.1148, cr_loss=0.3275, over 17165.00 frames. ], tot_loss[loss=0.1881, ctc_loss=0.1205, cr_loss=0.3379, over 3358460.86 frames. ], batch size: 45, lr: 2.63e-03, grad_scale: 16.0 2024-09-25 19:17:40,707 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=812816.6666666666, ans=0.09899494936611666 2024-09-25 19:17:50,162 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=812863.3333333334, ans=0.125 2024-09-25 19:17:55,013 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=812863.3333333334, ans=0.125 2024-09-25 19:18:23,998 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=812956.6666666666, ans=0.125 2024-09-25 19:18:38,567 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=813003.3333333334, ans=0.125 2024-09-25 19:18:47,783 INFO [train.py:1198] (2/4) Epoch 45, batch 2800, loss[loss=0.2205, ctc_loss=0.1438, cr_loss=0.3834, over 17205.00 frames. ], tot_loss[loss=0.1885, ctc_loss=0.1208, cr_loss=0.3387, over 3351143.19 frames. ], batch size: 55, lr: 2.63e-03, grad_scale: 32.0 2024-09-25 19:18:59,495 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=813050.0, ans=0.125 2024-09-25 19:19:04,186 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=813096.6666666666, ans=0.125 2024-09-25 19:19:38,611 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=813190.0, ans=0.1 2024-09-25 19:19:41,759 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=813190.0, ans=0.125 2024-09-25 19:19:58,788 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=4.24 vs. limit=15.0 2024-09-25 19:20:01,231 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.187e+02 1.322e+02 1.415e+02 1.536e+02 2.357e+02, threshold=2.830e+02, percent-clipped=0.0 2024-09-25 19:20:12,479 INFO [train.py:1198] (2/4) Epoch 45, batch 2850, loss[loss=0.1877, ctc_loss=0.1224, cr_loss=0.3266, over 17019.00 frames. ], tot_loss[loss=0.1879, ctc_loss=0.1204, cr_loss=0.3377, over 3358764.77 frames. ], batch size: 51, lr: 2.63e-03, grad_scale: 32.0 2024-09-25 19:20:33,058 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 19:21:14,612 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.62 vs. limit=12.0 2024-09-25 19:21:27,237 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=8.65 vs. limit=12.0 2024-09-25 19:21:37,529 INFO [train.py:1198] (2/4) Epoch 45, batch 2900, loss[loss=0.1829, ctc_loss=0.1189, cr_loss=0.3202, over 17165.00 frames. ], tot_loss[loss=0.1882, ctc_loss=0.1206, cr_loss=0.3384, over 3368695.50 frames. ], batch size: 45, lr: 2.63e-03, grad_scale: 32.0 2024-09-25 19:21:39,449 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=813516.6666666666, ans=0.125 2024-09-25 19:21:57,040 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=813563.3333333334, ans=0.2 2024-09-25 19:22:04,991 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.min_positive, batch_count=813563.3333333334, ans=0.05 2024-09-25 19:22:07,256 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=3.99 vs. limit=15.0 2024-09-25 19:22:17,717 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=813610.0, ans=0.025 2024-09-25 19:22:46,286 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.314e+02 1.430e+02 1.564e+02 2.364e+02, threshold=2.859e+02, percent-clipped=0.0 2024-09-25 19:22:53,113 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=813703.3333333334, ans=0.125 2024-09-25 19:22:57,460 INFO [train.py:1198] (2/4) Epoch 45, batch 2950, loss[loss=0.1536, ctc_loss=0.09534, cr_loss=0.2912, over 17039.00 frames. ], tot_loss[loss=0.1876, ctc_loss=0.1201, cr_loss=0.3375, over 3363451.43 frames. ], batch size: 39, lr: 2.63e-03, grad_scale: 32.0 2024-09-25 19:23:09,535 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2.whitening_limit, batch_count=813750.0, ans=15.0 2024-09-25 19:23:40,862 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=813843.3333333334, ans=0.125 2024-09-25 19:23:44,310 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=813890.0, ans=0.125 2024-09-25 19:23:55,393 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.00 vs. limit=15.0 2024-09-25 19:24:06,641 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=8.29 vs. limit=15.0 2024-09-25 19:24:07,620 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=813936.6666666666, ans=0.125 2024-09-25 19:24:09,308 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=813936.6666666666, ans=0.125 2024-09-25 19:24:16,618 INFO [train.py:1198] (2/4) Epoch 45, batch 3000, loss[loss=0.1916, ctc_loss=0.1222, cr_loss=0.3471, over 16780.00 frames. ], tot_loss[loss=0.1882, ctc_loss=0.1206, cr_loss=0.3382, over 3361521.28 frames. ], batch size: 61, lr: 2.63e-03, grad_scale: 32.0 2024-09-25 19:24:16,618 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 19:24:32,370 INFO [train.py:1230] (2/4) Epoch 45, validation: loss=0.03541, ctc_loss=0.03541, cr_loss=1.054e-14, over 944034.00 frames. 2024-09-25 19:24:32,370 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 19:24:57,894 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=814030.0, ans=0.2 2024-09-25 19:25:02,490 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=814076.6666666666, ans=0.0 2024-09-25 19:25:04,547 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=8.88 vs. limit=15.0 2024-09-25 19:25:09,032 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=4.94 vs. limit=15.0 2024-09-25 19:25:44,657 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.170e+02 1.310e+02 1.382e+02 1.500e+02 2.246e+02, threshold=2.763e+02, percent-clipped=0.0 2024-09-25 19:25:48,115 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=814170.0, ans=0.2 2024-09-25 19:25:55,843 INFO [train.py:1198] (2/4) Epoch 45, batch 3050, loss[loss=0.2042, ctc_loss=0.1311, cr_loss=0.3658, over 17301.00 frames. ], tot_loss[loss=0.1888, ctc_loss=0.121, cr_loss=0.339, over 3347155.12 frames. ], batch size: 51, lr: 2.63e-03, grad_scale: 32.0 2024-09-25 19:26:10,135 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=814263.3333333334, ans=0.125 2024-09-25 19:26:22,589 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=814263.3333333334, ans=0.0 2024-09-25 19:26:33,311 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=814310.0, ans=0.0 2024-09-25 19:27:09,437 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=814403.3333333334, ans=0.0 2024-09-25 19:27:13,876 INFO [train.py:1198] (2/4) Epoch 45, batch 3100, loss[loss=0.182, ctc_loss=0.1161, cr_loss=0.3294, over 17016.00 frames. ], tot_loss[loss=0.1885, ctc_loss=0.1208, cr_loss=0.3385, over 3350584.75 frames. ], batch size: 51, lr: 2.63e-03, grad_scale: 32.0 2024-09-25 19:27:38,447 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=814496.6666666666, ans=0.125 2024-09-25 19:27:58,981 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=814543.3333333334, ans=0.0 2024-09-25 19:28:10,176 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=814590.0, ans=0.1 2024-09-25 19:28:15,229 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=7.13 vs. limit=15.0 2024-09-25 19:28:23,833 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.168e+02 1.327e+02 1.418e+02 1.524e+02 1.856e+02, threshold=2.835e+02, percent-clipped=0.0 2024-09-25 19:28:36,919 INFO [train.py:1198] (2/4) Epoch 45, batch 3150, loss[loss=0.2035, ctc_loss=0.1304, cr_loss=0.3655, over 16645.00 frames. ], tot_loss[loss=0.1884, ctc_loss=0.1207, cr_loss=0.3383, over 3357672.83 frames. ], batch size: 66, lr: 2.63e-03, grad_scale: 32.0 2024-09-25 19:28:45,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=814683.3333333334, ans=0.0 2024-09-25 19:28:59,127 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=814730.0, ans=0.0 2024-09-25 19:29:06,091 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.74 vs. limit=15.0 2024-09-25 19:29:44,166 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.11 vs. limit=15.0 2024-09-25 19:29:55,900 INFO [train.py:1198] (2/4) Epoch 45, batch 3200, loss[loss=0.2202, ctc_loss=0.142, cr_loss=0.3912, over 16571.00 frames. ], tot_loss[loss=0.1891, ctc_loss=0.1212, cr_loss=0.3396, over 3352533.04 frames. ], batch size: 66, lr: 2.62e-03, grad_scale: 32.0 2024-09-25 19:30:31,099 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.46 vs. limit=15.0 2024-09-25 19:30:32,268 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=815010.0, ans=0.0 2024-09-25 19:31:03,191 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.085e+02 1.319e+02 1.396e+02 1.488e+02 2.041e+02, threshold=2.793e+02, percent-clipped=0.0 2024-09-25 19:31:08,272 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=815103.3333333334, ans=0.1 2024-09-25 19:31:14,383 INFO [train.py:1198] (2/4) Epoch 45, batch 3250, loss[loss=0.1776, ctc_loss=0.1121, cr_loss=0.3275, over 17167.00 frames. ], tot_loss[loss=0.188, ctc_loss=0.1204, cr_loss=0.338, over 3347293.11 frames. ], batch size: 41, lr: 2.62e-03, grad_scale: 32.0 2024-09-25 19:31:32,256 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 19:32:05,339 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=815290.0, ans=0.125 2024-09-25 19:32:27,101 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=815336.6666666666, ans=0.1 2024-09-25 19:32:33,038 INFO [train.py:1198] (2/4) Epoch 45, batch 3300, loss[loss=0.187, ctc_loss=0.1173, cr_loss=0.3486, over 17006.00 frames. ], tot_loss[loss=0.1884, ctc_loss=0.1206, cr_loss=0.3386, over 3350084.97 frames. ], batch size: 51, lr: 2.62e-03, grad_scale: 32.0 2024-09-25 19:32:56,650 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=815430.0, ans=0.125 2024-09-25 19:33:07,559 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=815476.6666666666, ans=0.0 2024-09-25 19:33:11,202 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=4.88 vs. limit=15.0 2024-09-25 19:33:40,403 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.116e+02 1.311e+02 1.407e+02 1.511e+02 1.886e+02, threshold=2.815e+02, percent-clipped=0.0 2024-09-25 19:33:51,417 INFO [train.py:1198] (2/4) Epoch 45, batch 3350, loss[loss=0.166, ctc_loss=0.1039, cr_loss=0.3102, over 17102.00 frames. ], tot_loss[loss=0.1885, ctc_loss=0.1207, cr_loss=0.3389, over 3358067.48 frames. ], batch size: 40, lr: 2.62e-03, grad_scale: 32.0 2024-09-25 19:34:04,179 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=815616.6666666666, ans=0.125 2024-09-25 19:34:18,459 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=815663.3333333334, ans=0.125 2024-09-25 19:34:21,677 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=815710.0, ans=0.04949747468305833 2024-09-25 19:34:23,076 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=815710.0, ans=0.125 2024-09-25 19:34:23,392 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=21.43 vs. limit=22.5 2024-09-25 19:34:30,128 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.61 vs. limit=10.0 2024-09-25 19:34:50,007 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=815756.6666666666, ans=0.125 2024-09-25 19:35:05,961 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=815803.3333333334, ans=0.2 2024-09-25 19:35:07,531 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=815803.3333333334, ans=0.125 2024-09-25 19:35:10,353 INFO [train.py:1198] (2/4) Epoch 45, batch 3400, loss[loss=0.2009, ctc_loss=0.128, cr_loss=0.3645, over 17109.00 frames. ], tot_loss[loss=0.1888, ctc_loss=0.121, cr_loss=0.3391, over 3361034.24 frames. ], batch size: 49, lr: 2.62e-03, grad_scale: 32.0 2024-09-25 19:35:13,752 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=815850.0, ans=0.0 2024-09-25 19:35:26,694 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.whiten.whitening_limit, batch_count=815896.6666666666, ans=12.0 2024-09-25 19:35:35,541 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=815896.6666666666, ans=0.125 2024-09-25 19:35:53,005 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 19:35:53,363 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=13.43 vs. limit=22.5 2024-09-25 19:36:17,105 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass_mid.scale_min, batch_count=816036.6666666666, ans=0.2 2024-09-25 19:36:19,819 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.104e+02 1.333e+02 1.429e+02 1.523e+02 2.294e+02, threshold=2.857e+02, percent-clipped=0.0 2024-09-25 19:36:30,574 INFO [train.py:1198] (2/4) Epoch 45, batch 3450, loss[loss=0.2089, ctc_loss=0.1319, cr_loss=0.3852, over 17206.00 frames. ], tot_loss[loss=0.1886, ctc_loss=0.1209, cr_loss=0.3385, over 3366051.14 frames. ], batch size: 55, lr: 2.62e-03, grad_scale: 32.0 2024-09-25 19:36:43,413 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=816083.3333333334, ans=0.125 2024-09-25 19:37:33,384 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=816270.0, ans=0.125 2024-09-25 19:37:38,276 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=816270.0, ans=0.0 2024-09-25 19:37:50,524 INFO [train.py:1198] (2/4) Epoch 45, batch 3500, loss[loss=0.1958, ctc_loss=0.1239, cr_loss=0.3598, over 17327.00 frames. ], tot_loss[loss=0.1884, ctc_loss=0.1207, cr_loss=0.3386, over 3360546.79 frames. ], batch size: 49, lr: 2.62e-03, grad_scale: 32.0 2024-09-25 19:37:50,825 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=816316.6666666666, ans=0.0 2024-09-25 19:37:53,875 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=816316.6666666666, ans=0.125 2024-09-25 19:38:05,698 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=4.81 vs. limit=15.0 2024-09-25 19:38:09,592 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=816363.3333333334, ans=0.025 2024-09-25 19:38:47,249 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=816456.6666666666, ans=0.0 2024-09-25 19:38:56,626 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=816503.3333333334, ans=0.0 2024-09-25 19:39:01,034 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.083e+02 1.307e+02 1.374e+02 1.461e+02 1.986e+02, threshold=2.748e+02, percent-clipped=0.0 2024-09-25 19:39:10,423 INFO [train.py:1198] (2/4) Epoch 45, batch 3550, loss[loss=0.2229, ctc_loss=0.1464, cr_loss=0.3828, over 14740.00 frames. ], tot_loss[loss=0.1884, ctc_loss=0.1207, cr_loss=0.3385, over 3362129.89 frames. ], batch size: 89, lr: 2.62e-03, grad_scale: 16.0 2024-09-25 19:39:26,950 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten1.whitening_limit, batch_count=816596.6666666666, ans=10.0 2024-09-25 19:39:31,137 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=816596.6666666666, ans=0.125 2024-09-25 19:39:56,797 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=11.99 vs. limit=22.5 2024-09-25 19:40:21,399 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=816736.6666666666, ans=0.125 2024-09-25 19:40:28,818 INFO [train.py:1198] (2/4) Epoch 45, batch 3600, loss[loss=0.1706, ctc_loss=0.1057, cr_loss=0.3244, over 17116.00 frames. ], tot_loss[loss=0.1886, ctc_loss=0.1208, cr_loss=0.339, over 3360036.44 frames. ], batch size: 40, lr: 2.62e-03, grad_scale: 32.0 2024-09-25 19:40:44,746 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=816830.0, ans=0.0 2024-09-25 19:40:48,176 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=816830.0, ans=0.0 2024-09-25 19:40:51,305 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=816830.0, ans=0.125 2024-09-25 19:41:09,851 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=816876.6666666666, ans=0.0 2024-09-25 19:41:15,999 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=816923.3333333334, ans=0.125 2024-09-25 19:41:23,036 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.38 vs. limit=15.0 2024-09-25 19:41:37,495 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.193e+02 1.298e+02 1.374e+02 1.488e+02 2.136e+02, threshold=2.747e+02, percent-clipped=0.0 2024-09-25 19:41:46,855 INFO [train.py:1198] (2/4) Epoch 45, batch 3650, loss[loss=0.1886, ctc_loss=0.122, cr_loss=0.333, over 17136.00 frames. ], tot_loss[loss=0.1883, ctc_loss=0.1205, cr_loss=0.3387, over 3361974.00 frames. ], batch size: 48, lr: 2.62e-03, grad_scale: 32.0 2024-09-25 19:42:12,050 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=817063.3333333334, ans=0.0 2024-09-25 19:42:56,503 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=817203.3333333334, ans=0.125 2024-09-25 19:43:05,643 INFO [train.py:1198] (2/4) Epoch 45, batch 3700, loss[loss=0.1812, ctc_loss=0.1148, cr_loss=0.3324, over 17024.00 frames. ], tot_loss[loss=0.1888, ctc_loss=0.121, cr_loss=0.3392, over 3357211.48 frames. ], batch size: 44, lr: 2.62e-03, grad_scale: 32.0 2024-09-25 19:43:09,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=817250.0, ans=0.025 2024-09-25 19:44:14,696 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.153e+02 1.280e+02 1.340e+02 1.437e+02 1.807e+02, threshold=2.680e+02, percent-clipped=0.0 2024-09-25 19:44:18,167 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=817436.6666666666, ans=0.125 2024-09-25 19:44:24,087 INFO [train.py:1198] (2/4) Epoch 45, batch 3750, loss[loss=0.2174, ctc_loss=0.1413, cr_loss=0.3806, over 15056.00 frames. ], tot_loss[loss=0.1897, ctc_loss=0.1217, cr_loss=0.3401, over 3343276.45 frames. ], batch size: 89, lr: 2.62e-03, grad_scale: 32.0 2024-09-25 19:45:21,538 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=817623.3333333334, ans=0.0 2024-09-25 19:45:44,856 INFO [train.py:1198] (2/4) Epoch 45, batch 3800, loss[loss=0.1615, ctc_loss=0.1009, cr_loss=0.3031, over 16952.00 frames. ], tot_loss[loss=0.189, ctc_loss=0.1213, cr_loss=0.3385, over 3329950.62 frames. ], batch size: 42, lr: 2.62e-03, grad_scale: 32.0 2024-09-25 19:45:51,337 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=817716.6666666666, ans=0.125 2024-09-25 19:45:59,625 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=3.55 vs. limit=15.0 2024-09-25 19:46:07,140 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=817763.3333333334, ans=0.125 2024-09-25 19:46:48,208 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=817903.3333333334, ans=0.125 2024-09-25 19:46:54,350 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.160e+02 1.363e+02 1.462e+02 1.575e+02 2.339e+02, threshold=2.925e+02, percent-clipped=0.0 2024-09-25 19:47:03,899 INFO [train.py:1198] (2/4) Epoch 45, batch 3850, loss[loss=0.2173, ctc_loss=0.1401, cr_loss=0.386, over 16737.00 frames. ], tot_loss[loss=0.1908, ctc_loss=0.1227, cr_loss=0.3403, over 3285518.17 frames. ], batch size: 61, lr: 2.62e-03, grad_scale: 32.0 2024-09-25 19:47:06,502 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.52 vs. limit=15.0 2024-09-25 19:47:15,723 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=817950.0, ans=0.125 2024-09-25 19:47:44,605 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=818043.3333333334, ans=0.125 2024-09-25 19:47:49,656 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=10.32 vs. limit=15.0 2024-09-25 19:49:07,309 INFO [train.py:1198] (2/4) Epoch 46, batch 0, loss[loss=0.1857, ctc_loss=0.1198, cr_loss=0.3292, over 17142.00 frames. ], tot_loss[loss=0.1857, ctc_loss=0.1198, cr_loss=0.3292, over 17142.00 frames. ], batch size: 48, lr: 2.59e-03, grad_scale: 32.0 2024-09-25 19:49:07,310 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 19:49:21,489 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.3.encoder.layers.0.self_attn_weights, attn_weights_entropy = tensor([1.8676, 2.9908, 2.9681, 3.1401, 3.0904, 2.6868, 2.9144, 1.9731], device='cuda:2') 2024-09-25 19:49:22,408 INFO [train.py:1230] (2/4) Epoch 46, validation: loss=0.03502, ctc_loss=0.03502, cr_loss=1.054e-14, over 944034.00 frames. 2024-09-25 19:49:22,409 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 19:49:35,267 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=818164.6666666666, ans=0.1 2024-09-25 19:49:51,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=818211.3333333334, ans=0.1 2024-09-25 19:50:04,476 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=818258.0, ans=0.125 2024-09-25 19:50:40,885 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.337e+02 1.469e+02 1.679e+02 2.405e+02, threshold=2.939e+02, percent-clipped=0.0 2024-09-25 19:50:44,076 INFO [train.py:1198] (2/4) Epoch 46, batch 50, loss[loss=0.2058, ctc_loss=0.1325, cr_loss=0.3665, over 17244.00 frames. ], tot_loss[loss=0.1862, ctc_loss=0.1194, cr_loss=0.3342, over 755421.21 frames. ], batch size: 55, lr: 2.59e-03, grad_scale: 32.0 2024-09-25 19:51:13,811 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=7.46 vs. limit=15.0 2024-09-25 19:51:30,307 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=818491.3333333334, ans=0.125 2024-09-25 19:51:48,612 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.56 vs. limit=15.0 2024-09-25 19:51:56,617 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=818584.6666666666, ans=0.1 2024-09-25 19:52:08,976 INFO [train.py:1198] (2/4) Epoch 46, batch 100, loss[loss=0.1752, ctc_loss=0.1079, cr_loss=0.3365, over 17263.00 frames. ], tot_loss[loss=0.1859, ctc_loss=0.119, cr_loss=0.3347, over 1336975.32 frames. ], batch size: 44, lr: 2.59e-03, grad_scale: 32.0 2024-09-25 19:52:28,675 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=818678.0, ans=0.2 2024-09-25 19:52:39,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=818724.6666666666, ans=0.2 2024-09-25 19:52:53,459 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.26 vs. limit=15.0 2024-09-25 19:53:05,076 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=818771.3333333334, ans=0.035 2024-09-25 19:53:16,633 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=818818.0, ans=0.2 2024-09-25 19:53:24,401 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=818818.0, ans=0.2 2024-09-25 19:53:28,956 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.141e+02 1.282e+02 1.339e+02 1.432e+02 3.555e+02, threshold=2.678e+02, percent-clipped=2.0 2024-09-25 19:53:29,224 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=818818.0, ans=0.2 2024-09-25 19:53:32,230 INFO [train.py:1198] (2/4) Epoch 46, batch 150, loss[loss=0.1827, ctc_loss=0.1179, cr_loss=0.3243, over 16915.00 frames. ], tot_loss[loss=0.1857, ctc_loss=0.1187, cr_loss=0.335, over 1784599.52 frames. ], batch size: 58, lr: 2.59e-03, grad_scale: 32.0 2024-09-25 19:53:44,386 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=12.27 vs. limit=15.0 2024-09-25 19:53:51,627 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=818911.3333333334, ans=0.1 2024-09-25 19:53:53,128 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=818911.3333333334, ans=0.125 2024-09-25 19:54:51,561 INFO [train.py:1198] (2/4) Epoch 46, batch 200, loss[loss=0.2108, ctc_loss=0.1377, cr_loss=0.3657, over 16075.00 frames. ], tot_loss[loss=0.1842, ctc_loss=0.1178, cr_loss=0.3324, over 2134235.35 frames. ], batch size: 74, lr: 2.59e-03, grad_scale: 32.0 2024-09-25 19:55:02,979 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=819098.0, ans=0.2 2024-09-25 19:55:23,955 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=819191.3333333334, ans=0.0 2024-09-25 19:55:28,168 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=819191.3333333334, ans=0.0 2024-09-25 19:55:48,747 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=819238.0, ans=0.0 2024-09-25 19:55:54,081 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=7.64 vs. limit=15.0 2024-09-25 19:55:58,356 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=819284.6666666666, ans=0.125 2024-09-25 19:56:10,854 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=819284.6666666666, ans=0.2 2024-09-25 19:56:13,857 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.146e+02 1.302e+02 1.353e+02 1.428e+02 2.054e+02, threshold=2.706e+02, percent-clipped=0.0 2024-09-25 19:56:17,269 INFO [train.py:1198] (2/4) Epoch 46, batch 250, loss[loss=0.1792, ctc_loss=0.1147, cr_loss=0.3225, over 16979.00 frames. ], tot_loss[loss=0.1843, ctc_loss=0.1178, cr_loss=0.3326, over 2408756.81 frames. ], batch size: 53, lr: 2.59e-03, grad_scale: 32.0 2024-09-25 19:56:19,143 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=819331.3333333334, ans=0.125 2024-09-25 19:56:26,460 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.64 vs. limit=6.0 2024-09-25 19:56:54,519 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=819424.6666666666, ans=0.125 2024-09-25 19:56:56,592 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=7.24 vs. limit=15.0 2024-09-25 19:57:17,210 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=819471.3333333334, ans=0.0 2024-09-25 19:57:23,763 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=819518.0, ans=0.0 2024-09-25 19:57:40,839 INFO [train.py:1198] (2/4) Epoch 46, batch 300, loss[loss=0.2113, ctc_loss=0.1367, cr_loss=0.3733, over 17009.00 frames. ], tot_loss[loss=0.1852, ctc_loss=0.1185, cr_loss=0.3336, over 2612947.62 frames. ], batch size: 53, lr: 2.59e-03, grad_scale: 16.0 2024-09-25 19:58:01,039 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=12.41 vs. limit=15.0 2024-09-25 19:58:17,529 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=819658.0, ans=0.125 2024-09-25 19:58:22,449 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=819658.0, ans=0.1 2024-09-25 19:58:25,718 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=819658.0, ans=0.125 2024-09-25 19:59:02,222 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.177e+02 1.300e+02 1.374e+02 1.446e+02 2.011e+02, threshold=2.748e+02, percent-clipped=0.0 2024-09-25 19:59:03,734 INFO [train.py:1198] (2/4) Epoch 46, batch 350, loss[loss=0.1657, ctc_loss=0.1047, cr_loss=0.3048, over 17096.00 frames. ], tot_loss[loss=0.1862, ctc_loss=0.1192, cr_loss=0.3349, over 2777281.82 frames. ], batch size: 40, lr: 2.59e-03, grad_scale: 16.0 2024-09-25 19:59:16,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=819798.0, ans=0.1 2024-09-25 19:59:24,660 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=819844.6666666666, ans=0.025 2024-09-25 20:00:01,208 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=819938.0, ans=0.125 2024-09-25 20:00:14,404 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.26 vs. limit=15.0 2024-09-25 20:00:18,504 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=819984.6666666666, ans=0.125 2024-09-25 20:00:22,257 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=4.94 vs. limit=15.0 2024-09-25 20:00:23,196 INFO [train.py:1198] (2/4) Epoch 46, batch 400, loss[loss=0.201, ctc_loss=0.1291, cr_loss=0.3595, over 17225.00 frames. ], tot_loss[loss=0.1872, ctc_loss=0.1199, cr_loss=0.3366, over 2902283.22 frames. ], batch size: 50, lr: 2.59e-03, grad_scale: 32.0 2024-09-25 20:00:36,218 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=820031.3333333334, ans=0.125 2024-09-25 20:00:39,354 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=820031.3333333334, ans=0.125 2024-09-25 20:01:34,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=820218.0, ans=0.125 2024-09-25 20:01:39,990 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.19 vs. limit=15.0 2024-09-25 20:01:50,262 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.173e+02 1.300e+02 1.386e+02 1.492e+02 1.969e+02, threshold=2.772e+02, percent-clipped=0.0 2024-09-25 20:01:52,002 INFO [train.py:1198] (2/4) Epoch 46, batch 450, loss[loss=0.1712, ctc_loss=0.1072, cr_loss=0.3197, over 17168.00 frames. ], tot_loss[loss=0.1878, ctc_loss=0.1202, cr_loss=0.338, over 3010676.96 frames. ], batch size: 45, lr: 2.59e-03, grad_scale: 32.0 2024-09-25 20:02:24,421 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=820358.0, ans=0.015 2024-09-25 20:02:25,401 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.21 vs. limit=15.0 2024-09-25 20:02:48,524 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=820404.6666666666, ans=0.0 2024-09-25 20:02:51,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=820404.6666666666, ans=0.125 2024-09-25 20:03:14,978 INFO [train.py:1198] (2/4) Epoch 46, batch 500, loss[loss=0.1538, ctc_loss=0.09584, cr_loss=0.2896, over 17186.00 frames. ], tot_loss[loss=0.1867, ctc_loss=0.1194, cr_loss=0.3365, over 3094264.33 frames. ], batch size: 41, lr: 2.59e-03, grad_scale: 32.0 2024-09-25 20:03:16,670 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=820498.0, ans=0.125 2024-09-25 20:03:26,365 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=820498.0, ans=0.0 2024-09-25 20:03:41,488 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=14.45 vs. limit=22.5 2024-09-25 20:04:03,761 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=820638.0, ans=0.2 2024-09-25 20:04:29,267 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=820684.6666666666, ans=0.0 2024-09-25 20:04:33,504 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.017e+02 1.336e+02 1.423e+02 1.544e+02 3.291e+02, threshold=2.846e+02, percent-clipped=2.0 2024-09-25 20:04:35,165 INFO [train.py:1198] (2/4) Epoch 46, batch 550, loss[loss=0.2297, ctc_loss=0.1501, cr_loss=0.3979, over 15110.00 frames. ], tot_loss[loss=0.1881, ctc_loss=0.1203, cr_loss=0.3388, over 3149063.35 frames. ], batch size: 89, lr: 2.59e-03, grad_scale: 32.0 2024-09-25 20:05:29,232 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=820871.3333333334, ans=0.0 2024-09-25 20:05:40,579 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=820918.0, ans=0.0 2024-09-25 20:05:49,958 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=820918.0, ans=0.125 2024-09-25 20:05:52,146 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.84 vs. limit=15.0 2024-09-25 20:06:00,561 INFO [train.py:1198] (2/4) Epoch 46, batch 600, loss[loss=0.1695, ctc_loss=0.1074, cr_loss=0.3101, over 16686.00 frames. ], tot_loss[loss=0.1876, ctc_loss=0.12, cr_loss=0.3381, over 3192787.43 frames. ], batch size: 37, lr: 2.59e-03, grad_scale: 32.0 2024-09-25 20:06:08,874 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=820964.6666666666, ans=0.0 2024-09-25 20:06:15,248 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=821011.3333333334, ans=0.0 2024-09-25 20:07:01,964 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=4.19 vs. limit=15.0 2024-09-25 20:07:21,441 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.282e+02 1.361e+02 1.501e+02 2.285e+02, threshold=2.723e+02, percent-clipped=0.0 2024-09-25 20:07:23,100 INFO [train.py:1198] (2/4) Epoch 46, batch 650, loss[loss=0.1927, ctc_loss=0.1218, cr_loss=0.3546, over 17297.00 frames. ], tot_loss[loss=0.1871, ctc_loss=0.1196, cr_loss=0.3376, over 3237157.44 frames. ], batch size: 49, lr: 2.59e-03, grad_scale: 32.0 2024-09-25 20:07:30,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten.whitening_limit, batch_count=821198.0, ans=22.5 2024-09-25 20:07:44,236 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=821244.6666666666, ans=0.2 2024-09-25 20:07:52,461 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=821244.6666666666, ans=0.2 2024-09-25 20:07:58,156 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=821291.3333333334, ans=0.125 2024-09-25 20:08:34,374 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=821384.6666666666, ans=0.1 2024-09-25 20:08:34,514 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=821384.6666666666, ans=0.0 2024-09-25 20:08:34,519 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=821384.6666666666, ans=0.1 2024-09-25 20:08:45,584 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=821384.6666666666, ans=0.125 2024-09-25 20:08:48,495 INFO [train.py:1198] (2/4) Epoch 46, batch 700, loss[loss=0.1953, ctc_loss=0.1233, cr_loss=0.3602, over 17025.00 frames. ], tot_loss[loss=0.1869, ctc_loss=0.1195, cr_loss=0.337, over 3268783.95 frames. ], batch size: 56, lr: 2.59e-03, grad_scale: 32.0 2024-09-25 20:09:14,466 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=821478.0, ans=0.125 2024-09-25 20:09:14,560 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-25 20:09:31,794 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=821524.6666666666, ans=0.025 2024-09-25 20:09:44,738 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=821571.3333333334, ans=0.125 2024-09-25 20:09:52,799 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=821618.0, ans=0.125 2024-09-25 20:09:56,704 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=8.13 vs. limit=15.0 2024-09-25 20:10:06,893 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.162e+02 1.305e+02 1.403e+02 1.487e+02 1.713e+02, threshold=2.806e+02, percent-clipped=0.0 2024-09-25 20:10:08,541 INFO [train.py:1198] (2/4) Epoch 46, batch 750, loss[loss=0.1677, ctc_loss=0.107, cr_loss=0.3037, over 17258.00 frames. ], tot_loss[loss=0.1869, ctc_loss=0.1195, cr_loss=0.3371, over 3285669.61 frames. ], batch size: 44, lr: 2.59e-03, grad_scale: 32.0 2024-09-25 20:10:24,753 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=13.69 vs. limit=15.0 2024-09-25 20:10:29,220 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=821711.3333333334, ans=0.0 2024-09-25 20:10:35,497 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.min_abs, batch_count=821711.3333333334, ans=0.5 2024-09-25 20:11:01,951 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=821804.6666666666, ans=0.125 2024-09-25 20:11:13,215 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=821804.6666666666, ans=0.125 2024-09-25 20:11:27,085 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=821851.3333333334, ans=0.025 2024-09-25 20:11:28,884 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=12.47 vs. limit=15.0 2024-09-25 20:11:36,191 INFO [train.py:1198] (2/4) Epoch 46, batch 800, loss[loss=0.2017, ctc_loss=0.13, cr_loss=0.3585, over 17043.00 frames. ], tot_loss[loss=0.1864, ctc_loss=0.1192, cr_loss=0.336, over 3303803.12 frames. ], batch size: 52, lr: 2.58e-03, grad_scale: 32.0 2024-09-25 20:11:47,691 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=821898.0, ans=0.0 2024-09-25 20:12:46,753 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=822084.6666666666, ans=0.125 2024-09-25 20:12:55,694 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=822084.6666666666, ans=0.035 2024-09-25 20:12:58,584 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.090e+02 1.283e+02 1.367e+02 1.457e+02 2.182e+02, threshold=2.735e+02, percent-clipped=0.0 2024-09-25 20:12:58,608 INFO [train.py:1198] (2/4) Epoch 46, batch 850, loss[loss=0.1989, ctc_loss=0.1277, cr_loss=0.3559, over 17290.00 frames. ], tot_loss[loss=0.1869, ctc_loss=0.1195, cr_loss=0.3366, over 3321920.25 frames. ], batch size: 51, lr: 2.58e-03, grad_scale: 16.0 2024-09-25 20:13:05,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=822131.3333333334, ans=0.0 2024-09-25 20:13:21,234 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=822178.0, ans=0.1 2024-09-25 20:13:21,264 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=822178.0, ans=0.0 2024-09-25 20:13:34,112 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=12.09 vs. limit=15.0 2024-09-25 20:13:45,486 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=822271.3333333334, ans=0.1 2024-09-25 20:13:48,770 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=822271.3333333334, ans=0.125 2024-09-25 20:14:18,696 INFO [train.py:1198] (2/4) Epoch 46, batch 900, loss[loss=0.1689, ctc_loss=0.1079, cr_loss=0.3048, over 17297.00 frames. ], tot_loss[loss=0.1874, ctc_loss=0.12, cr_loss=0.337, over 3329090.39 frames. ], batch size: 46, lr: 2.58e-03, grad_scale: 16.0 2024-09-25 20:14:19,341 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.47 vs. limit=22.5 2024-09-25 20:14:36,564 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=822411.3333333334, ans=0.125 2024-09-25 20:14:38,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=822411.3333333334, ans=0.0 2024-09-25 20:14:57,301 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=822458.0, ans=0.125 2024-09-25 20:15:19,251 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=822504.6666666666, ans=0.0 2024-09-25 20:15:41,536 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.142e+02 1.288e+02 1.364e+02 1.452e+02 2.497e+02, threshold=2.728e+02, percent-clipped=0.0 2024-09-25 20:15:41,561 INFO [train.py:1198] (2/4) Epoch 46, batch 950, loss[loss=0.1818, ctc_loss=0.1167, cr_loss=0.3255, over 17027.00 frames. ], tot_loss[loss=0.1871, ctc_loss=0.1197, cr_loss=0.3369, over 3344991.67 frames. ], batch size: 44, lr: 2.58e-03, grad_scale: 16.0 2024-09-25 20:16:00,754 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=822644.6666666666, ans=0.125 2024-09-25 20:16:12,167 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=822644.6666666666, ans=0.0 2024-09-25 20:16:18,310 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=822691.3333333334, ans=0.07 2024-09-25 20:16:40,425 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=822738.0, ans=0.125 2024-09-25 20:16:43,613 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=822738.0, ans=0.125 2024-09-25 20:16:45,284 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=822738.0, ans=0.125 2024-09-25 20:16:55,599 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=13.58 vs. limit=22.5 2024-09-25 20:17:07,490 INFO [train.py:1198] (2/4) Epoch 46, batch 1000, loss[loss=0.1651, ctc_loss=0.1023, cr_loss=0.3139, over 17162.00 frames. ], tot_loss[loss=0.1869, ctc_loss=0.1195, cr_loss=0.3369, over 3350832.98 frames. ], batch size: 41, lr: 2.58e-03, grad_scale: 8.0 2024-09-25 20:17:19,157 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=822831.3333333334, ans=0.125 2024-09-25 20:17:28,909 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=822878.0, ans=0.2 2024-09-25 20:17:33,938 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=822878.0, ans=0.0 2024-09-25 20:17:41,823 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=822924.6666666666, ans=0.035 2024-09-25 20:17:48,224 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=822924.6666666666, ans=0.125 2024-09-25 20:18:02,633 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.47 vs. limit=10.0 2024-09-25 20:18:18,055 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=823018.0, ans=0.0 2024-09-25 20:18:21,437 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=823018.0, ans=0.125 2024-09-25 20:18:30,593 INFO [train.py:1198] (2/4) Epoch 46, batch 1050, loss[loss=0.1905, ctc_loss=0.1219, cr_loss=0.3434, over 17072.00 frames. ], tot_loss[loss=0.1868, ctc_loss=0.1195, cr_loss=0.3363, over 3334989.88 frames. ], batch size: 46, lr: 2.58e-03, grad_scale: 8.0 2024-09-25 20:18:30,915 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=823064.6666666666, ans=0.1 2024-09-25 20:18:32,115 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.054e+02 1.303e+02 1.373e+02 1.497e+02 1.983e+02, threshold=2.746e+02, percent-clipped=0.0 2024-09-25 20:18:49,992 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=823111.3333333334, ans=0.1 2024-09-25 20:19:01,753 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.51 vs. limit=6.0 2024-09-25 20:19:36,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=823251.3333333334, ans=0.2 2024-09-25 20:19:45,879 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=823251.3333333334, ans=0.0 2024-09-25 20:19:50,553 INFO [train.py:1198] (2/4) Epoch 46, batch 1100, loss[loss=0.1734, ctc_loss=0.1116, cr_loss=0.3087, over 17300.00 frames. ], tot_loss[loss=0.1872, ctc_loss=0.1198, cr_loss=0.3368, over 3346818.44 frames. ], batch size: 49, lr: 2.58e-03, grad_scale: 8.0 2024-09-25 20:19:55,547 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=823298.0, ans=0.125 2024-09-25 20:19:57,657 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=7.47 vs. limit=15.0 2024-09-25 20:20:01,229 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.60 vs. limit=6.0 2024-09-25 20:20:08,598 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=823344.6666666666, ans=0.125 2024-09-25 20:20:17,420 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=823344.6666666666, ans=0.0 2024-09-25 20:20:20,440 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=823344.6666666666, ans=0.125 2024-09-25 20:20:30,949 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=14.92 vs. limit=22.5 2024-09-25 20:20:38,484 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=823391.3333333334, ans=0.125 2024-09-25 20:20:41,466 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=823438.0, ans=0.0 2024-09-25 20:20:47,841 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=823438.0, ans=0.125 2024-09-25 20:20:54,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=823438.0, ans=0.0 2024-09-25 20:20:57,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=823484.6666666666, ans=0.125 2024-09-25 20:20:58,145 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.97 vs. limit=22.5 2024-09-25 20:21:15,712 INFO [train.py:1198] (2/4) Epoch 46, batch 1150, loss[loss=0.1969, ctc_loss=0.1256, cr_loss=0.3565, over 16876.00 frames. ], tot_loss[loss=0.1866, ctc_loss=0.1193, cr_loss=0.3364, over 3357917.94 frames. ], batch size: 58, lr: 2.58e-03, grad_scale: 8.0 2024-09-25 20:21:17,339 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.324e+02 1.372e+02 1.463e+02 5.655e+02, threshold=2.745e+02, percent-clipped=1.0 2024-09-25 20:21:46,861 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=5.99 vs. limit=15.0 2024-09-25 20:22:00,506 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=823624.6666666666, ans=0.2 2024-09-25 20:22:05,956 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.58 vs. limit=6.0 2024-09-25 20:22:27,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=823718.0, ans=0.125 2024-09-25 20:22:34,180 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=823718.0, ans=0.0 2024-09-25 20:22:38,928 INFO [train.py:1198] (2/4) Epoch 46, batch 1200, loss[loss=0.1745, ctc_loss=0.1083, cr_loss=0.3312, over 17206.00 frames. ], tot_loss[loss=0.1862, ctc_loss=0.1191, cr_loss=0.3357, over 3356644.35 frames. ], batch size: 41, lr: 2.58e-03, grad_scale: 16.0 2024-09-25 20:22:45,687 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=823764.6666666666, ans=0.0 2024-09-25 20:23:05,733 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer1.prob, batch_count=823811.3333333334, ans=0.125 2024-09-25 20:23:24,711 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=823858.0, ans=0.125 2024-09-25 20:23:27,714 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.min_abs, batch_count=823904.6666666666, ans=0.5 2024-09-25 20:23:42,289 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=823904.6666666666, ans=0.125 2024-09-25 20:23:42,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=823904.6666666666, ans=0.0 2024-09-25 20:23:43,854 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=823951.3333333334, ans=0.125 2024-09-25 20:24:00,927 INFO [train.py:1198] (2/4) Epoch 46, batch 1250, loss[loss=0.1634, ctc_loss=0.1007, cr_loss=0.3134, over 17274.00 frames. ], tot_loss[loss=0.1857, ctc_loss=0.1187, cr_loss=0.3351, over 3363352.95 frames. ], batch size: 42, lr: 2.58e-03, grad_scale: 16.0 2024-09-25 20:24:02,522 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.285e+02 1.375e+02 1.486e+02 1.915e+02, threshold=2.751e+02, percent-clipped=0.0 2024-09-25 20:24:18,702 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=824044.6666666666, ans=0.035 2024-09-25 20:24:21,264 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.77 vs. limit=6.0 2024-09-25 20:24:38,319 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.82 vs. limit=15.0 2024-09-25 20:24:41,232 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=11.11 vs. limit=12.0 2024-09-25 20:24:47,891 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=824138.0, ans=0.0 2024-09-25 20:25:23,678 INFO [train.py:1198] (2/4) Epoch 46, batch 1300, loss[loss=0.1804, ctc_loss=0.1145, cr_loss=0.3294, over 17146.00 frames. ], tot_loss[loss=0.1864, ctc_loss=0.1192, cr_loss=0.3358, over 3356342.87 frames. ], batch size: 48, lr: 2.58e-03, grad_scale: 16.0 2024-09-25 20:25:51,201 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=824278.0, ans=0.125 2024-09-25 20:26:24,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=824371.3333333334, ans=0.0 2024-09-25 20:26:39,321 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=824418.0, ans=0.125 2024-09-25 20:26:40,025 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module1.whiten, num_groups=1, num_channels=192, metric=7.62 vs. limit=15.0 2024-09-25 20:26:40,817 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=824418.0, ans=0.125 2024-09-25 20:26:48,556 INFO [train.py:1198] (2/4) Epoch 46, batch 1350, loss[loss=0.2278, ctc_loss=0.1474, cr_loss=0.4016, over 17145.00 frames. ], tot_loss[loss=0.1873, ctc_loss=0.1198, cr_loss=0.3372, over 3348317.82 frames. ], batch size: 48, lr: 2.58e-03, grad_scale: 8.0 2024-09-25 20:26:51,683 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.142e+02 1.295e+02 1.386e+02 1.472e+02 1.767e+02, threshold=2.771e+02, percent-clipped=0.0 2024-09-25 20:26:58,627 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.05 vs. limit=15.0 2024-09-25 20:27:59,262 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.21 vs. limit=15.0 2024-09-25 20:28:00,320 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=824651.3333333334, ans=0.125 2024-09-25 20:28:11,415 INFO [train.py:1198] (2/4) Epoch 46, batch 1400, loss[loss=0.1805, ctc_loss=0.1163, cr_loss=0.3212, over 17298.00 frames. ], tot_loss[loss=0.1869, ctc_loss=0.1195, cr_loss=0.3367, over 3353070.51 frames. ], batch size: 49, lr: 2.58e-03, grad_scale: 8.0 2024-09-25 20:28:11,796 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.min_abs, batch_count=824698.0, ans=0.5 2024-09-25 20:28:14,874 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=824698.0, ans=0.0 2024-09-25 20:28:23,154 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=824698.0, ans=0.1 2024-09-25 20:28:28,776 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten.whitening_limit, batch_count=824744.6666666666, ans=15.0 2024-09-25 20:28:39,169 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=824744.6666666666, ans=0.125 2024-09-25 20:28:43,977 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=824791.3333333334, ans=0.05 2024-09-25 20:28:50,411 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=824791.3333333334, ans=0.125 2024-09-25 20:29:03,201 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=824838.0, ans=0.0 2024-09-25 20:29:22,593 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=824884.6666666666, ans=0.09899494936611666 2024-09-25 20:29:31,892 INFO [train.py:1198] (2/4) Epoch 46, batch 1450, loss[loss=0.1923, ctc_loss=0.1251, cr_loss=0.3357, over 17083.00 frames. ], tot_loss[loss=0.1879, ctc_loss=0.1203, cr_loss=0.338, over 3354074.95 frames. ], batch size: 46, lr: 2.58e-03, grad_scale: 8.0 2024-09-25 20:29:35,104 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.196e+02 1.313e+02 1.376e+02 1.492e+02 1.838e+02, threshold=2.751e+02, percent-clipped=0.0 2024-09-25 20:29:36,985 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=824931.3333333334, ans=0.025 2024-09-25 20:29:39,360 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.14 vs. limit=15.0 2024-09-25 20:30:09,352 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=825024.6666666666, ans=0.07 2024-09-25 20:30:39,420 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=825118.0, ans=0.125 2024-09-25 20:30:44,070 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=825118.0, ans=0.2 2024-09-25 20:30:50,414 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=825118.0, ans=0.2 2024-09-25 20:30:54,859 INFO [train.py:1198] (2/4) Epoch 46, batch 1500, loss[loss=0.2212, ctc_loss=0.1423, cr_loss=0.3945, over 17143.00 frames. ], tot_loss[loss=0.1876, ctc_loss=0.1201, cr_loss=0.3376, over 3354374.22 frames. ], batch size: 48, lr: 2.58e-03, grad_scale: 8.0 2024-09-25 20:31:12,250 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=825211.3333333334, ans=0.125 2024-09-25 20:31:50,183 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=825304.6666666666, ans=0.1 2024-09-25 20:32:09,573 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=825351.3333333334, ans=0.125 2024-09-25 20:32:20,343 INFO [train.py:1198] (2/4) Epoch 46, batch 1550, loss[loss=0.1791, ctc_loss=0.1116, cr_loss=0.3371, over 17003.00 frames. ], tot_loss[loss=0.1878, ctc_loss=0.1202, cr_loss=0.3379, over 3351829.01 frames. ], batch size: 51, lr: 2.58e-03, grad_scale: 8.0 2024-09-25 20:32:23,492 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.171e+02 1.290e+02 1.399e+02 1.518e+02 4.516e+02, threshold=2.798e+02, percent-clipped=1.0 2024-09-25 20:32:25,526 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=825398.0, ans=0.125 2024-09-25 20:32:36,087 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.03 vs. limit=15.0 2024-09-25 20:32:36,860 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=825444.6666666666, ans=0.1 2024-09-25 20:32:40,451 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.48 vs. limit=15.0 2024-09-25 20:33:18,202 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.max_positive, batch_count=825538.0, ans=0.95 2024-09-25 20:33:31,536 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=8.54 vs. limit=15.0 2024-09-25 20:33:36,114 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=11.13 vs. limit=22.5 2024-09-25 20:33:43,584 INFO [train.py:1198] (2/4) Epoch 46, batch 1600, loss[loss=0.1483, ctc_loss=0.09139, cr_loss=0.2846, over 16955.00 frames. ], tot_loss[loss=0.1875, ctc_loss=0.12, cr_loss=0.3374, over 3357031.99 frames. ], batch size: 42, lr: 2.58e-03, grad_scale: 16.0 2024-09-25 20:33:45,368 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=825631.3333333334, ans=0.1 2024-09-25 20:33:48,683 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=825631.3333333334, ans=0.0 2024-09-25 20:34:03,007 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=825678.0, ans=0.2 2024-09-25 20:34:21,524 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.whiten.whitening_limit, batch_count=825724.6666666666, ans=15.0 2024-09-25 20:34:54,579 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=825818.0, ans=0.1 2024-09-25 20:35:04,110 INFO [train.py:1198] (2/4) Epoch 46, batch 1650, loss[loss=0.1372, ctc_loss=0.08603, cr_loss=0.2561, over 17274.00 frames. ], tot_loss[loss=0.1871, ctc_loss=0.1197, cr_loss=0.3369, over 3353189.44 frames. ], batch size: 42, lr: 2.58e-03, grad_scale: 16.0 2024-09-25 20:35:06,144 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=825864.6666666666, ans=0.2 2024-09-25 20:35:07,356 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.199e+02 1.315e+02 1.405e+02 1.548e+02 2.408e+02, threshold=2.810e+02, percent-clipped=0.0 2024-09-25 20:35:21,861 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 20:35:26,619 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=825911.3333333334, ans=0.125 2024-09-25 20:35:31,489 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=825911.3333333334, ans=0.1 2024-09-25 20:35:37,800 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=825958.0, ans=0.125 2024-09-25 20:35:54,231 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=826004.6666666666, ans=0.1 2024-09-25 20:36:04,890 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=826004.6666666666, ans=0.09899494936611666 2024-09-25 20:36:12,873 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=826051.3333333334, ans=0.125 2024-09-25 20:36:14,513 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=826051.3333333334, ans=0.2 2024-09-25 20:36:32,697 INFO [train.py:1198] (2/4) Epoch 46, batch 1700, loss[loss=0.214, ctc_loss=0.1371, cr_loss=0.3841, over 16780.00 frames. ], tot_loss[loss=0.1863, ctc_loss=0.1191, cr_loss=0.3361, over 3363081.05 frames. ], batch size: 61, lr: 2.58e-03, grad_scale: 16.0 2024-09-25 20:36:50,568 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=826144.6666666666, ans=0.0 2024-09-25 20:37:00,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=826144.6666666666, ans=0.2 2024-09-25 20:37:11,348 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=826191.3333333334, ans=0.0 2024-09-25 20:37:37,551 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=5.70 vs. limit=12.0 2024-09-25 20:37:55,294 INFO [train.py:1198] (2/4) Epoch 46, batch 1750, loss[loss=0.2192, ctc_loss=0.1457, cr_loss=0.3674, over 14842.00 frames. ], tot_loss[loss=0.1869, ctc_loss=0.1195, cr_loss=0.3368, over 3354242.40 frames. ], batch size: 89, lr: 2.58e-03, grad_scale: 8.0 2024-09-25 20:37:59,044 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=826331.3333333334, ans=0.0 2024-09-25 20:38:00,304 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.166e+02 1.299e+02 1.380e+02 1.488e+02 2.427e+02, threshold=2.759e+02, percent-clipped=0.0 2024-09-25 20:38:05,418 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=826331.3333333334, ans=0.2 2024-09-25 20:38:10,851 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.42 vs. limit=22.5 2024-09-25 20:38:14,025 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.20 vs. limit=15.0 2024-09-25 20:38:27,881 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=826424.6666666666, ans=0.125 2024-09-25 20:38:44,354 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.72 vs. limit=6.0 2024-09-25 20:39:15,274 INFO [train.py:1198] (2/4) Epoch 46, batch 1800, loss[loss=0.2087, ctc_loss=0.1331, cr_loss=0.3777, over 17239.00 frames. ], tot_loss[loss=0.1858, ctc_loss=0.1189, cr_loss=0.3349, over 3355478.73 frames. ], batch size: 50, lr: 2.58e-03, grad_scale: 8.0 2024-09-25 20:39:17,194 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=826564.6666666666, ans=0.125 2024-09-25 20:39:21,946 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=826564.6666666666, ans=0.0 2024-09-25 20:39:24,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=826564.6666666666, ans=0.0 2024-09-25 20:39:39,540 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=826611.3333333334, ans=0.125 2024-09-25 20:39:40,084 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.84 vs. limit=10.0 2024-09-25 20:39:49,750 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=17.79 vs. limit=22.5 2024-09-25 20:40:27,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=826751.3333333334, ans=0.2 2024-09-25 20:40:34,021 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=826751.3333333334, ans=0.125 2024-09-25 20:40:37,496 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=826798.0, ans=0.0 2024-09-25 20:40:38,768 INFO [train.py:1198] (2/4) Epoch 46, batch 1850, loss[loss=0.2121, ctc_loss=0.1388, cr_loss=0.3666, over 17002.00 frames. ], tot_loss[loss=0.1865, ctc_loss=0.1193, cr_loss=0.3362, over 3361417.45 frames. ], batch size: 53, lr: 2.58e-03, grad_scale: 8.0 2024-09-25 20:40:43,541 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.312e+02 1.384e+02 1.504e+02 2.324e+02, threshold=2.767e+02, percent-clipped=0.0 2024-09-25 20:40:59,413 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=826844.6666666666, ans=0.125 2024-09-25 20:41:16,771 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=826891.3333333334, ans=0.125 2024-09-25 20:41:24,238 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=826891.3333333334, ans=0.125 2024-09-25 20:42:03,933 INFO [train.py:1198] (2/4) Epoch 46, batch 1900, loss[loss=0.1922, ctc_loss=0.123, cr_loss=0.346, over 17212.00 frames. ], tot_loss[loss=0.186, ctc_loss=0.119, cr_loss=0.3354, over 3364941.35 frames. ], batch size: 47, lr: 2.58e-03, grad_scale: 8.0 2024-09-25 20:42:05,829 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 20:42:50,202 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=827124.6666666666, ans=0.0 2024-09-25 20:43:04,891 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=827171.3333333334, ans=0.125 2024-09-25 20:43:08,116 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=827171.3333333334, ans=0.0 2024-09-25 20:43:26,921 INFO [train.py:1198] (2/4) Epoch 46, batch 1950, loss[loss=0.174, ctc_loss=0.1091, cr_loss=0.3248, over 17224.00 frames. ], tot_loss[loss=0.1853, ctc_loss=0.1183, cr_loss=0.3347, over 3372112.87 frames. ], batch size: 50, lr: 2.58e-03, grad_scale: 8.0 2024-09-25 20:43:30,555 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=827264.6666666666, ans=0.0 2024-09-25 20:43:31,731 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.139e+02 1.295e+02 1.354e+02 1.509e+02 3.056e+02, threshold=2.707e+02, percent-clipped=1.0 2024-09-25 20:43:47,597 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=827311.3333333334, ans=0.125 2024-09-25 20:43:47,853 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=827311.3333333334, ans=0.125 2024-09-25 20:43:57,308 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=827358.0, ans=0.125 2024-09-25 20:44:23,539 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=5.02 vs. limit=12.0 2024-09-25 20:44:46,531 INFO [train.py:1198] (2/4) Epoch 46, batch 2000, loss[loss=0.2003, ctc_loss=0.1296, cr_loss=0.3533, over 16898.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1184, cr_loss=0.3351, over 3379378.53 frames. ], batch size: 58, lr: 2.58e-03, grad_scale: 16.0 2024-09-25 20:44:46,864 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=827498.0, ans=0.025 2024-09-25 20:44:58,117 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=827498.0, ans=0.125 2024-09-25 20:45:00,062 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.06 vs. limit=15.0 2024-09-25 20:45:27,582 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=827591.3333333334, ans=0.0 2024-09-25 20:45:33,011 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.75 vs. limit=6.0 2024-09-25 20:45:40,033 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=827638.0, ans=0.2 2024-09-25 20:45:44,119 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.77 vs. limit=6.0 2024-09-25 20:46:05,207 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=827684.6666666666, ans=0.04949747468305833 2024-09-25 20:46:11,124 INFO [train.py:1198] (2/4) Epoch 46, batch 2050, loss[loss=0.1793, ctc_loss=0.1106, cr_loss=0.3434, over 17171.00 frames. ], tot_loss[loss=0.1853, ctc_loss=0.1183, cr_loss=0.335, over 3375220.43 frames. ], batch size: 45, lr: 2.58e-03, grad_scale: 16.0 2024-09-25 20:46:13,094 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=827731.3333333334, ans=0.1 2024-09-25 20:46:18,715 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.306e+02 1.402e+02 1.494e+02 2.074e+02, threshold=2.805e+02, percent-clipped=0.0 2024-09-25 20:46:25,347 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=827731.3333333334, ans=0.1 2024-09-25 20:46:52,387 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=827824.6666666666, ans=0.125 2024-09-25 20:47:00,886 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=827871.3333333334, ans=0.1 2024-09-25 20:47:20,512 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=13.41 vs. limit=22.5 2024-09-25 20:47:28,302 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=827918.0, ans=0.125 2024-09-25 20:47:29,837 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=827918.0, ans=0.125 2024-09-25 20:47:32,984 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=827964.6666666666, ans=0.125 2024-09-25 20:47:34,396 INFO [train.py:1198] (2/4) Epoch 46, batch 2100, loss[loss=0.1771, ctc_loss=0.1123, cr_loss=0.324, over 17190.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1186, cr_loss=0.3343, over 3355372.29 frames. ], batch size: 47, lr: 2.58e-03, grad_scale: 16.0 2024-09-25 20:48:11,360 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=8.98 vs. limit=10.0 2024-09-25 20:48:50,797 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=828151.3333333334, ans=0.0 2024-09-25 20:48:56,845 INFO [train.py:1198] (2/4) Epoch 46, batch 2150, loss[loss=0.2179, ctc_loss=0.1382, cr_loss=0.3987, over 17016.00 frames. ], tot_loss[loss=0.1866, ctc_loss=0.1194, cr_loss=0.3363, over 3353413.92 frames. ], batch size: 56, lr: 2.57e-03, grad_scale: 16.0 2024-09-25 20:49:01,529 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.174e+02 1.324e+02 1.399e+02 1.517e+02 1.900e+02, threshold=2.799e+02, percent-clipped=0.0 2024-09-25 20:49:02,424 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.51 vs. limit=15.0 2024-09-25 20:49:29,237 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=828291.3333333334, ans=0.1 2024-09-25 20:49:30,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=828291.3333333334, ans=0.0 2024-09-25 20:49:32,360 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=828291.3333333334, ans=0.0 2024-09-25 20:49:37,442 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=7.12 vs. limit=15.0 2024-09-25 20:49:40,447 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=828291.3333333334, ans=0.125 2024-09-25 20:49:54,687 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=828338.0, ans=0.2 2024-09-25 20:50:03,485 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=20.53 vs. limit=22.5 2024-09-25 20:50:17,902 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=828431.3333333334, ans=0.0 2024-09-25 20:50:19,237 INFO [train.py:1198] (2/4) Epoch 46, batch 2200, loss[loss=0.1638, ctc_loss=0.1025, cr_loss=0.3061, over 17053.00 frames. ], tot_loss[loss=0.1866, ctc_loss=0.1194, cr_loss=0.3361, over 3341408.52 frames. ], batch size: 39, lr: 2.57e-03, grad_scale: 16.0 2024-09-25 20:50:19,964 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.17 vs. limit=15.0 2024-09-25 20:50:24,428 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=828431.3333333334, ans=0.05 2024-09-25 20:50:31,015 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=828431.3333333334, ans=0.07 2024-09-25 20:50:40,783 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.max_abs, batch_count=828478.0, ans=10.0 2024-09-25 20:51:07,942 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=9.54 vs. limit=15.0 2024-09-25 20:51:09,099 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=828571.3333333334, ans=0.0 2024-09-25 20:51:13,711 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=16.92 vs. limit=22.5 2024-09-25 20:51:44,711 INFO [train.py:1198] (2/4) Epoch 46, batch 2250, loss[loss=0.1885, ctc_loss=0.1232, cr_loss=0.3265, over 17284.00 frames. ], tot_loss[loss=0.1864, ctc_loss=0.1192, cr_loss=0.336, over 3357083.71 frames. ], batch size: 49, lr: 2.57e-03, grad_scale: 16.0 2024-09-25 20:51:49,433 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.075e+02 1.294e+02 1.364e+02 1.489e+02 2.080e+02, threshold=2.728e+02, percent-clipped=0.0 2024-09-25 20:51:53,525 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=8.18 vs. limit=15.0 2024-09-25 20:52:01,446 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.55 vs. limit=15.0 2024-09-25 20:52:22,451 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=13.07 vs. limit=15.0 2024-09-25 20:53:04,509 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=828851.3333333334, ans=0.05 2024-09-25 20:53:07,428 INFO [train.py:1198] (2/4) Epoch 46, batch 2300, loss[loss=0.1721, ctc_loss=0.1108, cr_loss=0.3062, over 17058.00 frames. ], tot_loss[loss=0.1867, ctc_loss=0.1195, cr_loss=0.3363, over 3352911.87 frames. ], batch size: 46, lr: 2.57e-03, grad_scale: 16.0 2024-09-25 20:53:16,153 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.06 vs. limit=15.0 2024-09-25 20:53:29,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff3_skip_rate, batch_count=828944.6666666666, ans=0.0 2024-09-25 20:53:35,644 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=10.72 vs. limit=15.0 2024-09-25 20:54:04,075 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=8.98 vs. limit=15.0 2024-09-25 20:54:18,481 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=7.73 vs. limit=15.0 2024-09-25 20:54:27,182 INFO [train.py:1198] (2/4) Epoch 46, batch 2350, loss[loss=0.2068, ctc_loss=0.1305, cr_loss=0.3811, over 17032.00 frames. ], tot_loss[loss=0.186, ctc_loss=0.1189, cr_loss=0.3353, over 3351320.69 frames. ], batch size: 56, lr: 2.57e-03, grad_scale: 16.0 2024-09-25 20:54:31,908 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.139e+02 1.291e+02 1.368e+02 1.435e+02 1.902e+02, threshold=2.737e+02, percent-clipped=0.0 2024-09-25 20:54:40,150 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=829131.3333333334, ans=0.2 2024-09-25 20:54:56,395 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=829178.0, ans=0.125 2024-09-25 20:55:07,041 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=829224.6666666666, ans=0.125 2024-09-25 20:55:20,624 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=11.67 vs. limit=15.0 2024-09-25 20:55:31,365 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=829271.3333333334, ans=0.125 2024-09-25 20:55:36,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=829318.0, ans=0.1 2024-09-25 20:55:46,498 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.65 vs. limit=6.0 2024-09-25 20:55:50,418 INFO [train.py:1198] (2/4) Epoch 46, batch 2400, loss[loss=0.2125, ctc_loss=0.1371, cr_loss=0.3772, over 15083.00 frames. ], tot_loss[loss=0.187, ctc_loss=0.1197, cr_loss=0.3368, over 3346895.52 frames. ], batch size: 89, lr: 2.57e-03, grad_scale: 32.0 2024-09-25 20:56:11,975 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=829411.3333333334, ans=0.0 2024-09-25 20:56:12,389 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=14.24 vs. limit=15.0 2024-09-25 20:56:37,377 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=829458.0, ans=0.015 2024-09-25 20:56:39,261 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=829458.0, ans=0.125 2024-09-25 20:57:15,351 INFO [train.py:1198] (2/4) Epoch 46, batch 2450, loss[loss=0.1833, ctc_loss=0.1172, cr_loss=0.3309, over 17130.00 frames. ], tot_loss[loss=0.1882, ctc_loss=0.1205, cr_loss=0.3386, over 3348922.53 frames. ], batch size: 48, lr: 2.57e-03, grad_scale: 32.0 2024-09-25 20:57:18,939 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=829598.0, ans=0.0 2024-09-25 20:57:20,236 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.115e+02 1.311e+02 1.415e+02 1.511e+02 3.334e+02, threshold=2.830e+02, percent-clipped=1.0 2024-09-25 20:58:01,426 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=829691.3333333334, ans=0.125 2024-09-25 20:58:14,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=829738.0, ans=0.125 2024-09-25 20:58:33,656 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=829784.6666666666, ans=0.0 2024-09-25 20:58:38,052 INFO [train.py:1198] (2/4) Epoch 46, batch 2500, loss[loss=0.1528, ctc_loss=0.09543, cr_loss=0.2866, over 17099.00 frames. ], tot_loss[loss=0.1867, ctc_loss=0.1194, cr_loss=0.3367, over 3357979.82 frames. ], batch size: 40, lr: 2.57e-03, grad_scale: 32.0 2024-09-25 20:58:43,469 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=8.39 vs. limit=15.0 2024-09-25 20:59:34,826 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=829971.3333333334, ans=0.1 2024-09-25 20:59:36,536 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=829971.3333333334, ans=0.0 2024-09-25 20:59:38,069 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=829971.3333333334, ans=0.2 2024-09-25 20:59:38,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=829971.3333333334, ans=0.0 2024-09-25 20:59:44,829 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=830018.0, ans=0.1 2024-09-25 20:59:54,151 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=830018.0, ans=0.125 2024-09-25 20:59:55,793 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=830018.0, ans=0.1 2024-09-25 20:59:58,772 INFO [train.py:1198] (2/4) Epoch 46, batch 2550, loss[loss=0.2119, ctc_loss=0.1364, cr_loss=0.3772, over 16072.00 frames. ], tot_loss[loss=0.1873, ctc_loss=0.1198, cr_loss=0.3374, over 3348117.38 frames. ], batch size: 74, lr: 2.57e-03, grad_scale: 32.0 2024-09-25 21:00:06,164 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.070e+02 1.313e+02 1.390e+02 1.516e+02 1.832e+02, threshold=2.779e+02, percent-clipped=0.0 2024-09-25 21:00:32,212 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=830158.0, ans=0.2 2024-09-25 21:01:13,041 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=9.23 vs. limit=15.0 2024-09-25 21:01:15,852 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=830251.3333333334, ans=0.0 2024-09-25 21:01:26,887 INFO [train.py:1198] (2/4) Epoch 46, batch 2600, loss[loss=0.1686, ctc_loss=0.1077, cr_loss=0.3045, over 17080.00 frames. ], tot_loss[loss=0.1882, ctc_loss=0.1204, cr_loss=0.3387, over 3348877.27 frames. ], batch size: 43, lr: 2.57e-03, grad_scale: 32.0 2024-09-25 21:01:53,418 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=830344.6666666666, ans=0.125 2024-09-25 21:02:42,688 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=830484.6666666666, ans=0.125 2024-09-25 21:02:49,477 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=830531.3333333334, ans=0.1 2024-09-25 21:02:50,768 INFO [train.py:1198] (2/4) Epoch 46, batch 2650, loss[loss=0.209, ctc_loss=0.1436, cr_loss=0.327, over 11666.00 frames. ], tot_loss[loss=0.188, ctc_loss=0.1203, cr_loss=0.3384, over 3346946.59 frames. ], batch size: 123, lr: 2.57e-03, grad_scale: 32.0 2024-09-25 21:02:55,451 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.099e+02 1.287e+02 1.345e+02 1.480e+02 2.035e+02, threshold=2.690e+02, percent-clipped=0.0 2024-09-25 21:03:00,638 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=830531.3333333334, ans=0.1 2024-09-25 21:03:03,693 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=830531.3333333334, ans=0.125 2024-09-25 21:03:11,726 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=830578.0, ans=0.025 2024-09-25 21:03:29,315 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=830624.6666666666, ans=0.125 2024-09-25 21:03:54,931 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=830718.0, ans=0.04949747468305833 2024-09-25 21:04:10,490 INFO [train.py:1198] (2/4) Epoch 46, batch 2700, loss[loss=0.1947, ctc_loss=0.125, cr_loss=0.3484, over 17040.00 frames. ], tot_loss[loss=0.1869, ctc_loss=0.1195, cr_loss=0.3372, over 3362933.09 frames. ], batch size: 46, lr: 2.57e-03, grad_scale: 32.0 2024-09-25 21:04:12,276 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=830764.6666666666, ans=0.0 2024-09-25 21:04:47,111 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=830858.0, ans=0.125 2024-09-25 21:04:50,236 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=830858.0, ans=0.125 2024-09-25 21:04:50,559 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=11.94 vs. limit=15.0 2024-09-25 21:05:02,330 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=830904.6666666666, ans=0.1 2024-09-25 21:05:05,327 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=830904.6666666666, ans=0.1 2024-09-25 21:05:15,406 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.11 vs. limit=15.0 2024-09-25 21:05:32,278 INFO [train.py:1198] (2/4) Epoch 46, batch 2750, loss[loss=0.2157, ctc_loss=0.1411, cr_loss=0.3731, over 16439.00 frames. ], tot_loss[loss=0.1864, ctc_loss=0.1192, cr_loss=0.336, over 3363157.01 frames. ], batch size: 66, lr: 2.57e-03, grad_scale: 32.0 2024-09-25 21:05:37,114 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.184e+02 1.313e+02 1.406e+02 1.516e+02 1.975e+02, threshold=2.813e+02, percent-clipped=0.0 2024-09-25 21:06:09,821 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=831091.3333333334, ans=0.1 2024-09-25 21:06:16,156 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=831091.3333333334, ans=0.1 2024-09-25 21:06:18,385 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.90 vs. limit=22.5 2024-09-25 21:06:30,520 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=831138.0, ans=0.0 2024-09-25 21:06:32,153 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=831138.0, ans=0.0 2024-09-25 21:06:35,291 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=831138.0, ans=0.125 2024-09-25 21:06:38,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=831138.0, ans=0.125 2024-09-25 21:06:41,695 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=831184.6666666666, ans=0.125 2024-09-25 21:06:46,018 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.81 vs. limit=6.0 2024-09-25 21:06:57,840 INFO [train.py:1198] (2/4) Epoch 46, batch 2800, loss[loss=0.196, ctc_loss=0.127, cr_loss=0.3451, over 17003.00 frames. ], tot_loss[loss=0.1867, ctc_loss=0.1195, cr_loss=0.3359, over 3362019.08 frames. ], batch size: 51, lr: 2.57e-03, grad_scale: 32.0 2024-09-25 21:07:05,404 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=4.35 vs. limit=15.0 2024-09-25 21:07:14,540 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=831278.0, ans=0.125 2024-09-25 21:07:24,029 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=831278.0, ans=0.125 2024-09-25 21:07:31,341 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=831324.6666666666, ans=0.1 2024-09-25 21:07:39,253 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.min_positive, batch_count=831324.6666666666, ans=0.05 2024-09-25 21:07:47,325 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=831371.3333333334, ans=0.0 2024-09-25 21:08:11,555 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=831418.0, ans=0.04949747468305833 2024-09-25 21:08:13,463 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.76 vs. limit=10.0 2024-09-25 21:08:16,212 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=831418.0, ans=0.125 2024-09-25 21:08:20,587 INFO [train.py:1198] (2/4) Epoch 46, batch 2850, loss[loss=0.1878, ctc_loss=0.118, cr_loss=0.3488, over 17255.00 frames. ], tot_loss[loss=0.1851, ctc_loss=0.1184, cr_loss=0.3332, over 3364148.49 frames. ], batch size: 44, lr: 2.57e-03, grad_scale: 32.0 2024-09-25 21:08:26,958 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.083e+02 1.275e+02 1.372e+02 1.492e+02 2.089e+02, threshold=2.744e+02, percent-clipped=0.0 2024-09-25 21:08:40,503 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.98 vs. limit=15.0 2024-09-25 21:09:22,711 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.34 vs. limit=15.0 2024-09-25 21:09:33,281 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=831651.3333333334, ans=0.0 2024-09-25 21:09:39,525 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=831698.0, ans=0.125 2024-09-25 21:09:40,971 INFO [train.py:1198] (2/4) Epoch 46, batch 2900, loss[loss=0.1675, ctc_loss=0.1053, cr_loss=0.3106, over 16744.00 frames. ], tot_loss[loss=0.1844, ctc_loss=0.1179, cr_loss=0.3321, over 3371571.84 frames. ], batch size: 37, lr: 2.57e-03, grad_scale: 16.0 2024-09-25 21:09:49,526 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=831698.0, ans=0.125 2024-09-25 21:10:04,760 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=831744.6666666666, ans=0.07 2024-09-25 21:10:24,802 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module1.whiten, num_groups=1, num_channels=192, metric=8.75 vs. limit=15.0 2024-09-25 21:10:59,285 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=831884.6666666666, ans=0.0 2024-09-25 21:11:00,896 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=831884.6666666666, ans=0.2 2024-09-25 21:11:08,571 INFO [train.py:1198] (2/4) Epoch 46, batch 2950, loss[loss=0.1537, ctc_loss=0.09384, cr_loss=0.2991, over 17097.00 frames. ], tot_loss[loss=0.1838, ctc_loss=0.1175, cr_loss=0.3318, over 3381272.88 frames. ], batch size: 43, lr: 2.57e-03, grad_scale: 16.0 2024-09-25 21:11:08,815 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=831931.3333333334, ans=0.1 2024-09-25 21:11:08,817 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=831931.3333333334, ans=10.0 2024-09-25 21:11:16,441 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.188e+02 1.273e+02 1.341e+02 1.442e+02 4.466e+02, threshold=2.682e+02, percent-clipped=1.0 2024-09-25 21:11:18,427 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=831931.3333333334, ans=0.1 2024-09-25 21:11:38,003 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=8.59 vs. limit=15.0 2024-09-25 21:11:52,669 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn1.whiten, num_groups=1, num_channels=192, metric=9.18 vs. limit=22.5 2024-09-25 21:12:15,683 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=832118.0, ans=0.025 2024-09-25 21:12:27,906 INFO [train.py:1198] (2/4) Epoch 46, batch 3000, loss[loss=0.1767, ctc_loss=0.113, cr_loss=0.3182, over 16872.00 frames. ], tot_loss[loss=0.1848, ctc_loss=0.1181, cr_loss=0.3334, over 3381519.50 frames. ], batch size: 58, lr: 2.57e-03, grad_scale: 16.0 2024-09-25 21:12:27,906 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 21:12:43,451 INFO [train.py:1230] (2/4) Epoch 46, validation: loss=0.03583, ctc_loss=0.03583, cr_loss=1.006e-14, over 944034.00 frames. 2024-09-25 21:12:43,451 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 21:12:46,886 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=832164.6666666666, ans=0.2 2024-09-25 21:12:53,310 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=832164.6666666666, ans=0.125 2024-09-25 21:13:18,657 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=14.06 vs. limit=15.0 2024-09-25 21:13:21,913 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=7.15 vs. limit=15.0 2024-09-25 21:13:51,287 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=10.84 vs. limit=15.0 2024-09-25 21:13:57,226 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=832351.3333333334, ans=0.0 2024-09-25 21:14:01,693 INFO [train.py:1198] (2/4) Epoch 46, batch 3050, loss[loss=0.2069, ctc_loss=0.1353, cr_loss=0.3577, over 16870.00 frames. ], tot_loss[loss=0.1836, ctc_loss=0.1173, cr_loss=0.3317, over 3386836.46 frames. ], batch size: 58, lr: 2.57e-03, grad_scale: 16.0 2024-09-25 21:14:09,459 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.315e+02 1.415e+02 1.524e+02 2.421e+02, threshold=2.829e+02, percent-clipped=0.0 2024-09-25 21:14:10,036 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=6.48 vs. limit=15.0 2024-09-25 21:14:23,986 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=832444.6666666666, ans=0.0 2024-09-25 21:14:31,688 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=832491.3333333334, ans=0.04949747468305833 2024-09-25 21:14:42,661 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=832491.3333333334, ans=0.035 2024-09-25 21:14:55,977 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=8.94 vs. limit=15.0 2024-09-25 21:14:58,418 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=832538.0, ans=0.0 2024-09-25 21:15:03,153 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=832584.6666666666, ans=0.1 2024-09-25 21:15:20,316 INFO [train.py:1198] (2/4) Epoch 46, batch 3100, loss[loss=0.1809, ctc_loss=0.1171, cr_loss=0.319, over 16999.00 frames. ], tot_loss[loss=0.1846, ctc_loss=0.118, cr_loss=0.3328, over 3390231.32 frames. ], batch size: 51, lr: 2.57e-03, grad_scale: 16.0 2024-09-25 21:15:33,045 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=832631.3333333334, ans=0.125 2024-09-25 21:15:45,570 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=832678.0, ans=0.09899494936611666 2024-09-25 21:15:47,172 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=832678.0, ans=0.0 2024-09-25 21:15:51,789 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=832724.6666666666, ans=0.125 2024-09-25 21:15:52,628 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=17.35 vs. limit=22.5 2024-09-25 21:16:04,593 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 21:16:38,987 INFO [train.py:1198] (2/4) Epoch 46, batch 3150, loss[loss=0.1926, ctc_loss=0.1226, cr_loss=0.3501, over 17072.00 frames. ], tot_loss[loss=0.1841, ctc_loss=0.1176, cr_loss=0.3324, over 3386089.97 frames. ], batch size: 46, lr: 2.57e-03, grad_scale: 16.0 2024-09-25 21:16:46,803 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.123e+02 1.285e+02 1.343e+02 1.430e+02 2.452e+02, threshold=2.687e+02, percent-clipped=0.0 2024-09-25 21:16:50,625 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.27 vs. limit=10.0 2024-09-25 21:16:58,344 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=832911.3333333334, ans=0.125 2024-09-25 21:16:58,799 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.55 vs. limit=15.0 2024-09-25 21:17:50,376 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=833051.3333333334, ans=0.0 2024-09-25 21:17:59,372 INFO [train.py:1198] (2/4) Epoch 46, batch 3200, loss[loss=0.1627, ctc_loss=0.1029, cr_loss=0.2993, over 17030.00 frames. ], tot_loss[loss=0.184, ctc_loss=0.1175, cr_loss=0.3326, over 3390316.70 frames. ], batch size: 39, lr: 2.57e-03, grad_scale: 32.0 2024-09-25 21:18:04,728 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.84 vs. limit=15.0 2024-09-25 21:18:06,354 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=512, metric=7.22 vs. limit=15.0 2024-09-25 21:18:18,519 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=833144.6666666666, ans=0.125 2024-09-25 21:18:21,537 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.min_positive, batch_count=833144.6666666666, ans=0.025 2024-09-25 21:18:22,089 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.14 vs. limit=15.0 2024-09-25 21:18:24,604 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.max_abs, batch_count=833144.6666666666, ans=10.0 2024-09-25 21:18:33,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=833191.3333333334, ans=0.125 2024-09-25 21:19:00,608 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=833284.6666666666, ans=0.125 2024-09-25 21:19:17,328 INFO [train.py:1198] (2/4) Epoch 46, batch 3250, loss[loss=0.1779, ctc_loss=0.1147, cr_loss=0.3161, over 17224.00 frames. ], tot_loss[loss=0.185, ctc_loss=0.1182, cr_loss=0.3342, over 3385549.37 frames. ], batch size: 50, lr: 2.57e-03, grad_scale: 32.0 2024-09-25 21:19:22,877 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=15.22 vs. limit=22.5 2024-09-25 21:19:25,151 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.170e+02 1.305e+02 1.369e+02 1.483e+02 2.240e+02, threshold=2.739e+02, percent-clipped=0.0 2024-09-25 21:19:48,057 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=833378.0, ans=0.125 2024-09-25 21:19:52,754 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=833424.6666666666, ans=0.0 2024-09-25 21:20:14,512 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=833471.3333333334, ans=0.0 2024-09-25 21:20:34,659 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer_na.min_abs, batch_count=833518.0, ans=0.02 2024-09-25 21:20:40,789 INFO [train.py:1198] (2/4) Epoch 46, batch 3300, loss[loss=0.2155, ctc_loss=0.1401, cr_loss=0.3773, over 17080.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1184, cr_loss=0.3349, over 3387859.75 frames. ], batch size: 49, lr: 2.57e-03, grad_scale: 32.0 2024-09-25 21:21:04,687 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=833611.3333333334, ans=0.5 2024-09-25 21:21:04,835 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=833611.3333333334, ans=0.125 2024-09-25 21:21:15,760 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=833658.0, ans=0.125 2024-09-25 21:21:25,254 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=833658.0, ans=0.0 2024-09-25 21:21:58,829 INFO [train.py:1198] (2/4) Epoch 46, batch 3350, loss[loss=0.1423, ctc_loss=0.08881, cr_loss=0.2676, over 16238.00 frames. ], tot_loss[loss=0.1852, ctc_loss=0.1184, cr_loss=0.3343, over 3383753.05 frames. ], batch size: 36, lr: 2.57e-03, grad_scale: 32.0 2024-09-25 21:22:06,614 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.162e+02 1.301e+02 1.404e+02 1.465e+02 2.030e+02, threshold=2.807e+02, percent-clipped=0.0 2024-09-25 21:22:20,919 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 21:22:27,029 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=833844.6666666666, ans=0.1 2024-09-25 21:22:50,753 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.93 vs. limit=22.5 2024-09-25 21:23:16,305 INFO [train.py:1198] (2/4) Epoch 46, batch 3400, loss[loss=0.1783, ctc_loss=0.1127, cr_loss=0.3283, over 16970.00 frames. ], tot_loss[loss=0.1849, ctc_loss=0.1181, cr_loss=0.3341, over 3376730.59 frames. ], batch size: 58, lr: 2.57e-03, grad_scale: 32.0 2024-09-25 21:23:27,640 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=834031.3333333334, ans=0.2 2024-09-25 21:24:21,647 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=8.30 vs. limit=15.0 2024-09-25 21:24:30,451 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 21:24:35,927 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=5.07 vs. limit=15.0 2024-09-25 21:24:36,348 INFO [train.py:1198] (2/4) Epoch 46, batch 3450, loss[loss=0.1545, ctc_loss=0.09763, cr_loss=0.2845, over 17285.00 frames. ], tot_loss[loss=0.1855, ctc_loss=0.1185, cr_loss=0.3347, over 3370606.16 frames. ], batch size: 42, lr: 2.57e-03, grad_scale: 16.0 2024-09-25 21:24:45,570 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.164e+02 1.317e+02 1.418e+02 1.501e+02 3.351e+02, threshold=2.836e+02, percent-clipped=1.0 2024-09-25 21:25:04,454 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=834311.3333333334, ans=0.125 2024-09-25 21:25:11,465 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=10.63 vs. limit=15.0 2024-09-25 21:25:23,427 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=834404.6666666666, ans=0.0 2024-09-25 21:25:23,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=834404.6666666666, ans=0.1 2024-09-25 21:25:28,252 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=834404.6666666666, ans=0.125 2024-09-25 21:25:31,427 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=12.55 vs. limit=12.0 2024-09-25 21:25:34,680 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=9.96 vs. limit=22.5 2024-09-25 21:25:35,946 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=834404.6666666666, ans=0.125 2024-09-25 21:25:54,707 INFO [train.py:1198] (2/4) Epoch 46, batch 3500, loss[loss=0.1651, ctc_loss=0.1039, cr_loss=0.3063, over 17209.00 frames. ], tot_loss[loss=0.1851, ctc_loss=0.1184, cr_loss=0.3335, over 3357805.42 frames. ], batch size: 47, lr: 2.57e-03, grad_scale: 16.0 2024-09-25 21:26:11,896 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=834544.6666666666, ans=0.0 2024-09-25 21:26:12,500 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.55 vs. limit=12.0 2024-09-25 21:26:35,376 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=834591.3333333334, ans=0.125 2024-09-25 21:27:03,626 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=834684.6666666666, ans=0.125 2024-09-25 21:27:12,618 INFO [train.py:1198] (2/4) Epoch 46, batch 3550, loss[loss=0.2368, ctc_loss=0.1582, cr_loss=0.3932, over 16448.00 frames. ], tot_loss[loss=0.1858, ctc_loss=0.1189, cr_loss=0.3349, over 3354857.84 frames. ], batch size: 66, lr: 2.56e-03, grad_scale: 16.0 2024-09-25 21:27:12,973 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.min_abs, batch_count=834731.3333333334, ans=0.5 2024-09-25 21:27:21,879 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.171e+02 1.283e+02 1.360e+02 1.445e+02 2.258e+02, threshold=2.720e+02, percent-clipped=0.0 2024-09-25 21:27:45,909 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=834824.6666666666, ans=0.0 2024-09-25 21:27:49,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=834824.6666666666, ans=0.2 2024-09-25 21:27:59,791 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=834871.3333333334, ans=0.125 2024-09-25 21:28:10,960 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=834871.3333333334, ans=0.0 2024-09-25 21:28:32,452 INFO [train.py:1198] (2/4) Epoch 46, batch 3600, loss[loss=0.1952, ctc_loss=0.1238, cr_loss=0.3571, over 17313.00 frames. ], tot_loss[loss=0.1862, ctc_loss=0.1192, cr_loss=0.3352, over 3341764.11 frames. ], batch size: 51, lr: 2.56e-03, grad_scale: 32.0 2024-09-25 21:28:45,230 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=834964.6666666666, ans=0.07 2024-09-25 21:29:14,145 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=7.29 vs. limit=15.0 2024-09-25 21:29:55,383 INFO [train.py:1198] (2/4) Epoch 46, batch 3650, loss[loss=0.1675, ctc_loss=0.1061, cr_loss=0.3068, over 17104.00 frames. ], tot_loss[loss=0.1865, ctc_loss=0.1194, cr_loss=0.3354, over 3344312.19 frames. ], batch size: 43, lr: 2.56e-03, grad_scale: 32.0 2024-09-25 21:30:04,585 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.106e+02 1.303e+02 1.373e+02 1.458e+02 2.085e+02, threshold=2.745e+02, percent-clipped=0.0 2024-09-25 21:30:33,153 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 21:30:50,708 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=835338.0, ans=0.125 2024-09-25 21:30:53,896 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=835338.0, ans=0.125 2024-09-25 21:31:14,426 INFO [train.py:1198] (2/4) Epoch 46, batch 3700, loss[loss=0.1896, ctc_loss=0.1206, cr_loss=0.3448, over 17368.00 frames. ], tot_loss[loss=0.187, ctc_loss=0.1197, cr_loss=0.3363, over 3355260.35 frames. ], batch size: 48, lr: 2.56e-03, grad_scale: 16.0 2024-09-25 21:31:30,460 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=835478.0, ans=0.125 2024-09-25 21:32:01,794 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=835571.3333333334, ans=0.125 2024-09-25 21:32:32,251 INFO [train.py:1198] (2/4) Epoch 46, batch 3750, loss[loss=0.18, ctc_loss=0.1164, cr_loss=0.3182, over 17362.00 frames. ], tot_loss[loss=0.187, ctc_loss=0.1197, cr_loss=0.3364, over 3356609.23 frames. ], batch size: 48, lr: 2.56e-03, grad_scale: 16.0 2024-09-25 21:32:43,197 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.138e+02 1.294e+02 1.384e+02 1.512e+02 2.088e+02, threshold=2.767e+02, percent-clipped=0.0 2024-09-25 21:32:59,177 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.34 vs. limit=15.0 2024-09-25 21:33:00,443 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=835711.3333333334, ans=0.07 2024-09-25 21:33:09,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=835758.0, ans=0.125 2024-09-25 21:33:10,657 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 21:33:15,790 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=11.30 vs. limit=22.5 2024-09-25 21:33:18,530 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.const_attention_rate, batch_count=835804.6666666666, ans=0.025 2024-09-25 21:33:49,436 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=835898.0, ans=0.2 2024-09-25 21:33:50,842 INFO [train.py:1198] (2/4) Epoch 46, batch 3800, loss[loss=0.2002, ctc_loss=0.1292, cr_loss=0.3549, over 16528.00 frames. ], tot_loss[loss=0.1874, ctc_loss=0.12, cr_loss=0.3372, over 3340521.85 frames. ], batch size: 66, lr: 2.56e-03, grad_scale: 16.0 2024-09-25 21:33:52,624 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=835898.0, ans=0.125 2024-09-25 21:34:16,134 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=835944.6666666666, ans=0.125 2024-09-25 21:34:16,434 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.37 vs. limit=15.0 2024-09-25 21:34:27,238 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 21:34:36,457 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=836038.0, ans=0.125 2024-09-25 21:34:55,334 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=836084.6666666666, ans=0.2 2024-09-25 21:35:08,992 INFO [train.py:1198] (2/4) Epoch 46, batch 3850, loss[loss=0.2386, ctc_loss=0.1583, cr_loss=0.4016, over 15124.00 frames. ], tot_loss[loss=0.1887, ctc_loss=0.1211, cr_loss=0.3381, over 3301674.74 frames. ], batch size: 89, lr: 2.56e-03, grad_scale: 16.0 2024-09-25 21:35:19,791 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.194e+02 1.341e+02 1.442e+02 1.559e+02 2.635e+02, threshold=2.885e+02, percent-clipped=0.0 2024-09-25 21:35:44,557 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=836224.6666666666, ans=0.125 2024-09-25 21:35:56,743 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=836271.3333333334, ans=0.125 2024-09-25 21:36:04,906 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=12.40 vs. limit=15.0 2024-09-25 21:37:05,581 INFO [train.py:1198] (2/4) Epoch 47, batch 0, loss[loss=0.1906, ctc_loss=0.1219, cr_loss=0.3439, over 17236.00 frames. ], tot_loss[loss=0.1906, ctc_loss=0.1219, cr_loss=0.3439, over 17236.00 frames. ], batch size: 50, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 21:37:05,581 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 21:37:16,718 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.1.encoder.layers.0.self_attn_weights, attn_weights_entropy = tensor([4.3032, 4.1961, 3.7787, 4.2960], device='cuda:2') 2024-09-25 21:37:22,170 INFO [train.py:1230] (2/4) Epoch 47, validation: loss=0.03509, ctc_loss=0.03509, cr_loss=1.062e-14, over 944034.00 frames. 2024-09-25 21:37:22,170 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 21:37:25,689 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=836346.0, ans=0.1 2024-09-25 21:37:25,710 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=836346.0, ans=0.04949747468305833 2024-09-25 21:37:35,167 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=836346.0, ans=0.0 2024-09-25 21:38:00,049 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.07 vs. limit=15.0 2024-09-25 21:38:01,657 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=5.31 vs. limit=12.0 2024-09-25 21:38:07,544 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=836439.3333333334, ans=0.125 2024-09-25 21:38:33,888 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff2_skip_rate, batch_count=836532.6666666666, ans=0.0 2024-09-25 21:38:44,942 INFO [train.py:1198] (2/4) Epoch 47, batch 50, loss[loss=0.1761, ctc_loss=0.111, cr_loss=0.3254, over 17089.00 frames. ], tot_loss[loss=0.1878, ctc_loss=0.1202, cr_loss=0.3379, over 755921.35 frames. ], batch size: 43, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 21:38:55,066 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=836579.3333333334, ans=0.0 2024-09-25 21:38:56,631 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=836579.3333333334, ans=0.125 2024-09-25 21:39:02,614 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.187e+02 1.320e+02 1.514e+02 1.646e+02 2.881e+02, threshold=3.028e+02, percent-clipped=0.0 2024-09-25 21:39:03,034 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=836626.0, ans=0.1 2024-09-25 21:39:18,867 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 21:39:33,472 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=13.07 vs. limit=15.0 2024-09-25 21:40:05,068 INFO [train.py:1198] (2/4) Epoch 47, batch 100, loss[loss=0.1961, ctc_loss=0.123, cr_loss=0.3651, over 17163.00 frames. ], tot_loss[loss=0.1856, ctc_loss=0.1186, cr_loss=0.335, over 1336566.01 frames. ], batch size: 45, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 21:40:11,807 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.max_abs, batch_count=836812.6666666666, ans=10.0 2024-09-25 21:40:18,170 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=836812.6666666666, ans=0.035 2024-09-25 21:40:28,435 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=6.16 vs. limit=15.0 2024-09-25 21:40:39,463 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=836906.0, ans=0.04949747468305833 2024-09-25 21:40:48,937 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=836906.0, ans=0.04949747468305833 2024-09-25 21:41:08,424 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=4.81 vs. limit=12.0 2024-09-25 21:41:16,318 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.41 vs. limit=15.0 2024-09-25 21:41:19,167 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=836999.3333333334, ans=0.1 2024-09-25 21:41:26,868 INFO [train.py:1198] (2/4) Epoch 47, batch 150, loss[loss=0.1697, ctc_loss=0.1072, cr_loss=0.3124, over 17257.00 frames. ], tot_loss[loss=0.184, ctc_loss=0.1176, cr_loss=0.332, over 1790979.96 frames. ], batch size: 42, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 21:41:28,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=837046.0, ans=0.125 2024-09-25 21:41:44,260 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.065e+02 1.291e+02 1.353e+02 1.432e+02 2.053e+02, threshold=2.706e+02, percent-clipped=0.0 2024-09-25 21:42:30,930 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=837186.0, ans=0.0 2024-09-25 21:42:48,765 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=9.02 vs. limit=15.0 2024-09-25 21:42:51,753 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=837279.3333333334, ans=0.1 2024-09-25 21:42:53,059 INFO [train.py:1198] (2/4) Epoch 47, batch 200, loss[loss=0.1638, ctc_loss=0.1029, cr_loss=0.3045, over 17240.00 frames. ], tot_loss[loss=0.1836, ctc_loss=0.1171, cr_loss=0.3325, over 2149992.79 frames. ], batch size: 42, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 21:43:18,779 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=837326.0, ans=0.125 2024-09-25 21:43:21,049 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=3.48 vs. limit=6.0 2024-09-25 21:43:21,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward3.hidden_balancer.prob, batch_count=837326.0, ans=0.125 2024-09-25 21:43:39,073 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=837372.6666666666, ans=0.0 2024-09-25 21:43:58,558 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=8.40 vs. limit=15.0 2024-09-25 21:43:59,747 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=837466.0, ans=0.125 2024-09-25 21:44:02,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=837466.0, ans=0.0 2024-09-25 21:44:03,755 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=3.83 vs. limit=10.0 2024-09-25 21:44:10,730 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=837466.0, ans=0.125 2024-09-25 21:44:15,402 INFO [train.py:1198] (2/4) Epoch 47, batch 250, loss[loss=0.1648, ctc_loss=0.1027, cr_loss=0.3105, over 17109.00 frames. ], tot_loss[loss=0.1853, ctc_loss=0.1184, cr_loss=0.3349, over 2425968.69 frames. ], batch size: 40, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 21:44:28,331 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=837512.6666666666, ans=0.1 2024-09-25 21:44:29,989 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=837559.3333333334, ans=0.2 2024-09-25 21:44:32,923 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.146e+02 1.332e+02 1.402e+02 1.492e+02 3.444e+02, threshold=2.804e+02, percent-clipped=1.0 2024-09-25 21:44:36,479 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=3.114e-02 2024-09-25 21:44:42,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=837559.3333333334, ans=0.125 2024-09-25 21:44:55,931 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=837606.0, ans=0.04949747468305833 2024-09-25 21:44:58,093 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=12.99 vs. limit=15.0 2024-09-25 21:45:13,516 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=837652.6666666666, ans=0.2 2024-09-25 21:45:23,078 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=837699.3333333334, ans=10.0 2024-09-25 21:45:35,450 INFO [train.py:1198] (2/4) Epoch 47, batch 300, loss[loss=0.2083, ctc_loss=0.1416, cr_loss=0.3337, over 12095.00 frames. ], tot_loss[loss=0.1862, ctc_loss=0.119, cr_loss=0.3358, over 2630215.79 frames. ], batch size: 124, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 21:45:48,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=837746.0, ans=0.025 2024-09-25 21:46:05,711 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=837792.6666666666, ans=0.1 2024-09-25 21:46:28,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=837886.0, ans=0.125 2024-09-25 21:46:36,178 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=837886.0, ans=0.0 2024-09-25 21:46:49,127 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=837932.6666666666, ans=0.125 2024-09-25 21:47:01,258 INFO [train.py:1198] (2/4) Epoch 47, batch 350, loss[loss=0.2144, ctc_loss=0.1411, cr_loss=0.3665, over 16749.00 frames. ], tot_loss[loss=0.1868, ctc_loss=0.1194, cr_loss=0.3369, over 2789574.81 frames. ], batch size: 61, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 21:47:16,049 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=838026.0, ans=0.1 2024-09-25 21:47:18,803 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.156e+02 1.331e+02 1.404e+02 1.493e+02 2.320e+02, threshold=2.808e+02, percent-clipped=0.0 2024-09-25 21:47:25,069 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=838026.0, ans=0.07 2024-09-25 21:47:26,551 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=838026.0, ans=0.125 2024-09-25 21:47:26,650 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=838026.0, ans=0.0 2024-09-25 21:47:33,090 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=838026.0, ans=0.0 2024-09-25 21:47:33,388 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=9.39 vs. limit=15.0 2024-09-25 21:47:39,818 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=838072.6666666666, ans=0.125 2024-09-25 21:48:16,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=838166.0, ans=0.1 2024-09-25 21:48:24,403 INFO [train.py:1198] (2/4) Epoch 47, batch 400, loss[loss=0.2052, ctc_loss=0.1315, cr_loss=0.3685, over 17214.00 frames. ], tot_loss[loss=0.1888, ctc_loss=0.1208, cr_loss=0.34, over 2921287.63 frames. ], batch size: 55, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 21:48:40,103 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=838212.6666666666, ans=0.125 2024-09-25 21:48:44,970 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.min_positive, batch_count=838259.3333333334, ans=0.05 2024-09-25 21:48:46,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass_mid.scale_min, batch_count=838259.3333333334, ans=0.2 2024-09-25 21:49:09,790 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=3.81 vs. limit=10.0 2024-09-25 21:49:23,568 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=838352.6666666666, ans=0.125 2024-09-25 21:49:47,315 INFO [train.py:1198] (2/4) Epoch 47, batch 450, loss[loss=0.1878, ctc_loss=0.1208, cr_loss=0.3352, over 17049.00 frames. ], tot_loss[loss=0.1882, ctc_loss=0.1204, cr_loss=0.339, over 3028041.78 frames. ], batch size: 52, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 21:49:50,844 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=838446.0, ans=0.025 2024-09-25 21:49:54,198 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=838446.0, ans=0.2 2024-09-25 21:50:04,976 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.186e+02 1.297e+02 1.382e+02 1.499e+02 1.706e+02, threshold=2.764e+02, percent-clipped=0.0 2024-09-25 21:50:41,995 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=838586.0, ans=0.0 2024-09-25 21:51:09,781 INFO [train.py:1198] (2/4) Epoch 47, batch 500, loss[loss=0.1891, ctc_loss=0.1206, cr_loss=0.3423, over 17362.00 frames. ], tot_loss[loss=0.1889, ctc_loss=0.1209, cr_loss=0.3399, over 3102477.96 frames. ], batch size: 48, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 21:51:14,840 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=838679.3333333334, ans=0.025 2024-09-25 21:51:19,504 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=838679.3333333334, ans=0.125 2024-09-25 21:51:22,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=838679.3333333334, ans=0.125 2024-09-25 21:51:26,200 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=14.19 vs. limit=22.5 2024-09-25 21:51:36,433 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.85 vs. limit=6.0 2024-09-25 21:52:22,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=838866.0, ans=0.025 2024-09-25 21:52:32,588 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=838866.0, ans=0.125 2024-09-25 21:52:35,634 INFO [train.py:1198] (2/4) Epoch 47, batch 550, loss[loss=0.1841, ctc_loss=0.1152, cr_loss=0.3442, over 17214.00 frames. ], tot_loss[loss=0.1888, ctc_loss=0.1209, cr_loss=0.3397, over 3156462.23 frames. ], batch size: 47, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 21:52:53,166 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 1.330e+02 1.398e+02 1.505e+02 2.211e+02, threshold=2.797e+02, percent-clipped=0.0 2024-09-25 21:53:09,412 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=839006.0, ans=0.025 2024-09-25 21:53:39,506 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=839052.6666666666, ans=0.0 2024-09-25 21:53:58,552 INFO [train.py:1198] (2/4) Epoch 47, batch 600, loss[loss=0.1871, ctc_loss=0.1169, cr_loss=0.3509, over 17017.00 frames. ], tot_loss[loss=0.1873, ctc_loss=0.1197, cr_loss=0.3376, over 3208930.99 frames. ], batch size: 44, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 21:54:21,308 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=839192.6666666666, ans=0.125 2024-09-25 21:54:38,885 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=839239.3333333334, ans=0.125 2024-09-25 21:55:06,025 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=839332.6666666666, ans=0.125 2024-09-25 21:55:13,911 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=839332.6666666666, ans=0.125 2024-09-25 21:55:18,521 INFO [train.py:1198] (2/4) Epoch 47, batch 650, loss[loss=0.2238, ctc_loss=0.1463, cr_loss=0.3877, over 12021.00 frames. ], tot_loss[loss=0.1883, ctc_loss=0.1205, cr_loss=0.3386, over 3224835.77 frames. ], batch size: 123, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 21:55:32,367 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=4.17 vs. limit=6.0 2024-09-25 21:55:36,234 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.288e+02 1.367e+02 1.485e+02 1.946e+02, threshold=2.733e+02, percent-clipped=0.0 2024-09-25 21:55:43,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=839426.0, ans=0.125 2024-09-25 21:56:41,549 INFO [train.py:1198] (2/4) Epoch 47, batch 700, loss[loss=0.1976, ctc_loss=0.1276, cr_loss=0.3498, over 17026.00 frames. ], tot_loss[loss=0.1878, ctc_loss=0.1202, cr_loss=0.338, over 3255838.39 frames. ], batch size: 53, lr: 2.53e-03, grad_scale: 16.0 2024-09-25 21:56:45,032 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=839612.6666666666, ans=0.125 2024-09-25 21:57:02,089 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=839659.3333333334, ans=0.2 2024-09-25 21:57:07,157 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=839659.3333333334, ans=0.125 2024-09-25 21:58:06,323 INFO [train.py:1198] (2/4) Epoch 47, batch 750, loss[loss=0.1831, ctc_loss=0.1193, cr_loss=0.3188, over 17002.00 frames. ], tot_loss[loss=0.1869, ctc_loss=0.1195, cr_loss=0.3367, over 3269801.40 frames. ], batch size: 51, lr: 2.53e-03, grad_scale: 16.0 2024-09-25 21:58:20,150 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=839846.0, ans=0.025 2024-09-25 21:58:27,884 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.164e+02 1.312e+02 1.373e+02 1.486e+02 2.253e+02, threshold=2.746e+02, percent-clipped=0.0 2024-09-25 21:58:33,111 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=839892.6666666666, ans=0.0 2024-09-25 21:58:53,985 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=839939.3333333334, ans=0.125 2024-09-25 21:58:57,344 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=839986.0, ans=0.1 2024-09-25 21:59:18,082 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.72 vs. limit=6.0 2024-09-25 21:59:31,441 INFO [train.py:1198] (2/4) Epoch 47, batch 800, loss[loss=0.1915, ctc_loss=0.1245, cr_loss=0.3349, over 15933.00 frames. ], tot_loss[loss=0.186, ctc_loss=0.1189, cr_loss=0.3355, over 3294960.66 frames. ], batch size: 74, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 21:59:31,688 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=840079.3333333334, ans=0.0 2024-09-25 21:59:31,792 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=840079.3333333334, ans=0.2 2024-09-25 21:59:41,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=840079.3333333334, ans=0.1 2024-09-25 22:00:54,191 INFO [train.py:1198] (2/4) Epoch 47, batch 850, loss[loss=0.1843, ctc_loss=0.1162, cr_loss=0.3402, over 17274.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1184, cr_loss=0.335, over 3316048.10 frames. ], batch size: 42, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 22:01:13,303 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.171e+02 1.297e+02 1.378e+02 1.477e+02 2.622e+02, threshold=2.756e+02, percent-clipped=0.0 2024-09-25 22:01:13,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=840359.3333333334, ans=0.2 2024-09-25 22:01:15,192 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.skip_rate, batch_count=840359.3333333334, ans=0.035 2024-09-25 22:01:29,426 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=840406.0, ans=0.0 2024-09-25 22:01:33,329 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.72 vs. limit=6.0 2024-09-25 22:01:45,274 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=840452.6666666666, ans=0.0 2024-09-25 22:02:15,139 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=8.70 vs. limit=15.0 2024-09-25 22:02:19,133 INFO [train.py:1198] (2/4) Epoch 47, batch 900, loss[loss=0.2023, ctc_loss=0.1329, cr_loss=0.3468, over 16595.00 frames. ], tot_loss[loss=0.1864, ctc_loss=0.1191, cr_loss=0.3366, over 3319006.16 frames. ], batch size: 66, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 22:02:32,171 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 22:02:48,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=840592.6666666666, ans=0.125 2024-09-25 22:02:48,080 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=840592.6666666666, ans=0.125 2024-09-25 22:03:03,990 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=840639.3333333334, ans=0.125 2024-09-25 22:03:11,905 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=840686.0, ans=0.2 2024-09-25 22:03:20,753 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=840686.0, ans=0.0 2024-09-25 22:03:25,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=840732.6666666666, ans=0.125 2024-09-25 22:03:41,091 INFO [train.py:1198] (2/4) Epoch 47, batch 950, loss[loss=0.1957, ctc_loss=0.1241, cr_loss=0.358, over 17339.00 frames. ], tot_loss[loss=0.1863, ctc_loss=0.119, cr_loss=0.3362, over 3330290.90 frames. ], batch size: 48, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 22:03:46,320 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=840779.3333333334, ans=0.0 2024-09-25 22:03:51,130 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=840779.3333333334, ans=0.04949747468305833 2024-09-25 22:04:00,131 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.146e+02 1.323e+02 1.399e+02 1.542e+02 2.460e+02, threshold=2.799e+02, percent-clipped=0.0 2024-09-25 22:04:11,634 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=840872.6666666666, ans=0.0 2024-09-25 22:04:11,680 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=840872.6666666666, ans=0.07 2024-09-25 22:04:30,441 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=840919.3333333334, ans=0.125 2024-09-25 22:04:33,914 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 22:05:00,651 INFO [train.py:1198] (2/4) Epoch 47, batch 1000, loss[loss=0.1715, ctc_loss=0.1084, cr_loss=0.3156, over 16955.00 frames. ], tot_loss[loss=0.1864, ctc_loss=0.1192, cr_loss=0.336, over 3326103.08 frames. ], batch size: 42, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 22:05:08,937 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=841012.6666666666, ans=0.1 2024-09-25 22:05:24,738 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=841059.3333333334, ans=0.1 2024-09-25 22:06:09,091 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=841199.3333333334, ans=0.125 2024-09-25 22:06:15,901 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=7.72 vs. limit=15.0 2024-09-25 22:06:23,151 INFO [train.py:1198] (2/4) Epoch 47, batch 1050, loss[loss=0.1537, ctc_loss=0.09575, cr_loss=0.2898, over 16338.00 frames. ], tot_loss[loss=0.186, ctc_loss=0.1189, cr_loss=0.3354, over 3334007.74 frames. ], batch size: 36, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 22:06:41,080 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=841292.6666666666, ans=0.125 2024-09-25 22:06:42,387 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.129e+02 1.322e+02 1.409e+02 1.520e+02 2.848e+02, threshold=2.818e+02, percent-clipped=1.0 2024-09-25 22:06:58,636 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.35 vs. limit=15.0 2024-09-25 22:07:04,530 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=841339.3333333334, ans=0.0 2024-09-25 22:07:08,731 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=841339.3333333334, ans=0.125 2024-09-25 22:07:47,935 INFO [train.py:1198] (2/4) Epoch 47, batch 1100, loss[loss=0.1558, ctc_loss=0.09934, cr_loss=0.2822, over 17100.00 frames. ], tot_loss[loss=0.1858, ctc_loss=0.1186, cr_loss=0.3356, over 3344865.26 frames. ], batch size: 40, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 22:07:49,943 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=841479.3333333334, ans=0.1 2024-09-25 22:07:55,251 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.19 vs. limit=12.0 2024-09-25 22:08:01,267 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=841479.3333333334, ans=0.125 2024-09-25 22:09:09,639 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=8.97 vs. limit=15.0 2024-09-25 22:09:10,494 INFO [train.py:1198] (2/4) Epoch 47, batch 1150, loss[loss=0.1845, ctc_loss=0.1215, cr_loss=0.3147, over 17202.00 frames. ], tot_loss[loss=0.1864, ctc_loss=0.1191, cr_loss=0.3362, over 3352764.09 frames. ], batch size: 47, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 22:09:21,842 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=841712.6666666666, ans=0.04949747468305833 2024-09-25 22:09:29,554 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.082e+02 1.291e+02 1.365e+02 1.481e+02 2.112e+02, threshold=2.730e+02, percent-clipped=0.0 2024-09-25 22:09:34,618 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=841759.3333333334, ans=0.125 2024-09-25 22:09:58,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=841852.6666666666, ans=0.0 2024-09-25 22:10:09,972 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=841852.6666666666, ans=0.1 2024-09-25 22:10:17,841 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=841899.3333333334, ans=0.125 2024-09-25 22:10:26,006 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.attention_skip_rate, batch_count=841899.3333333334, ans=0.0 2024-09-25 22:10:29,181 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=841946.0, ans=0.1 2024-09-25 22:10:29,455 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=13.85 vs. limit=15.0 2024-09-25 22:10:32,842 INFO [train.py:1198] (2/4) Epoch 47, batch 1200, loss[loss=0.1627, ctc_loss=0.1006, cr_loss=0.3105, over 17170.00 frames. ], tot_loss[loss=0.1864, ctc_loss=0.1191, cr_loss=0.3364, over 3363035.12 frames. ], batch size: 41, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 22:10:47,607 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 22:11:07,090 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=842039.3333333334, ans=0.0 2024-09-25 22:11:12,617 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=6.77 vs. limit=8.0 2024-09-25 22:11:33,351 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.74 vs. limit=22.5 2024-09-25 22:11:40,715 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=842132.6666666666, ans=0.0 2024-09-25 22:11:55,780 INFO [train.py:1198] (2/4) Epoch 47, batch 1250, loss[loss=0.1717, ctc_loss=0.109, cr_loss=0.3134, over 17060.00 frames. ], tot_loss[loss=0.1857, ctc_loss=0.1186, cr_loss=0.3351, over 3364356.30 frames. ], batch size: 46, lr: 2.53e-03, grad_scale: 32.0 2024-09-25 22:12:04,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=842179.3333333334, ans=0.0 2024-09-25 22:12:13,576 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.86 vs. limit=10.0 2024-09-25 22:12:17,777 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.161e+02 1.299e+02 1.369e+02 1.454e+02 1.923e+02, threshold=2.738e+02, percent-clipped=0.0 2024-09-25 22:12:51,755 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=842319.3333333334, ans=0.0 2024-09-25 22:12:53,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=842319.3333333334, ans=0.09899494936611666 2024-09-25 22:13:02,725 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=842366.0, ans=0.2 2024-09-25 22:13:05,921 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.min_positive, batch_count=842366.0, ans=0.05 2024-09-25 22:13:20,996 INFO [train.py:1198] (2/4) Epoch 47, batch 1300, loss[loss=0.1811, ctc_loss=0.1129, cr_loss=0.341, over 16944.00 frames. ], tot_loss[loss=0.1856, ctc_loss=0.1186, cr_loss=0.3353, over 3367991.62 frames. ], batch size: 42, lr: 2.53e-03, grad_scale: 16.0 2024-09-25 22:13:37,513 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=842459.3333333334, ans=0.0 2024-09-25 22:13:48,707 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=842459.3333333334, ans=0.125 2024-09-25 22:14:02,958 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=842506.0, ans=0.0 2024-09-25 22:14:40,803 INFO [train.py:1198] (2/4) Epoch 47, batch 1350, loss[loss=0.2007, ctc_loss=0.1301, cr_loss=0.3535, over 16999.00 frames. ], tot_loss[loss=0.1857, ctc_loss=0.1186, cr_loss=0.3351, over 3363461.30 frames. ], batch size: 51, lr: 2.53e-03, grad_scale: 16.0 2024-09-25 22:15:01,641 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.184e+02 1.286e+02 1.352e+02 1.458e+02 2.037e+02, threshold=2.703e+02, percent-clipped=0.0 2024-09-25 22:15:10,622 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=4.94 vs. limit=15.0 2024-09-25 22:15:19,505 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=842739.3333333334, ans=0.0 2024-09-25 22:15:19,599 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=842739.3333333334, ans=0.125 2024-09-25 22:15:34,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.max_abs, batch_count=842786.0, ans=10.0 2024-09-25 22:16:03,149 INFO [train.py:1198] (2/4) Epoch 47, batch 1400, loss[loss=0.2014, ctc_loss=0.1262, cr_loss=0.3759, over 17130.00 frames. ], tot_loss[loss=0.1861, ctc_loss=0.1189, cr_loss=0.3359, over 3368738.14 frames. ], batch size: 48, lr: 2.52e-03, grad_scale: 16.0 2024-09-25 22:16:03,496 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=842879.3333333334, ans=0.125 2024-09-25 22:16:22,749 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=842926.0, ans=0.1 2024-09-25 22:16:37,035 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=842972.6666666666, ans=0.025 2024-09-25 22:16:37,037 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=842972.6666666666, ans=0.125 2024-09-25 22:17:18,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=843066.0, ans=0.2 2024-09-25 22:17:23,293 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=843066.0, ans=0.125 2024-09-25 22:17:27,856 INFO [train.py:1198] (2/4) Epoch 47, batch 1450, loss[loss=0.1692, ctc_loss=0.1056, cr_loss=0.318, over 17091.00 frames. ], tot_loss[loss=0.1858, ctc_loss=0.1188, cr_loss=0.335, over 3363970.91 frames. ], batch size: 40, lr: 2.52e-03, grad_scale: 16.0 2024-09-25 22:17:32,989 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=843112.6666666666, ans=0.125 2024-09-25 22:17:39,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=843112.6666666666, ans=0.07 2024-09-25 22:17:48,458 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.105e+02 1.274e+02 1.355e+02 1.464e+02 2.562e+02, threshold=2.711e+02, percent-clipped=0.0 2024-09-25 22:18:04,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=843206.0, ans=0.07 2024-09-25 22:18:47,145 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.max_positive, batch_count=843299.3333333334, ans=0.95 2024-09-25 22:18:50,192 INFO [train.py:1198] (2/4) Epoch 47, batch 1500, loss[loss=0.1971, ctc_loss=0.1264, cr_loss=0.3534, over 16026.00 frames. ], tot_loss[loss=0.1855, ctc_loss=0.1186, cr_loss=0.3347, over 3365902.58 frames. ], batch size: 74, lr: 2.52e-03, grad_scale: 16.0 2024-09-25 22:19:45,027 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=843486.0, ans=0.05 2024-09-25 22:19:49,836 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=843486.0, ans=0.2 2024-09-25 22:19:59,502 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=843532.6666666666, ans=0.0 2024-09-25 22:20:10,578 INFO [train.py:1198] (2/4) Epoch 47, batch 1550, loss[loss=0.1742, ctc_loss=0.1095, cr_loss=0.3237, over 17243.00 frames. ], tot_loss[loss=0.1862, ctc_loss=0.119, cr_loss=0.3357, over 3347093.38 frames. ], batch size: 44, lr: 2.52e-03, grad_scale: 16.0 2024-09-25 22:20:34,259 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.106e+02 1.301e+02 1.382e+02 1.468e+02 1.851e+02, threshold=2.764e+02, percent-clipped=0.0 2024-09-25 22:21:10,290 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer_na.min_abs, batch_count=843719.3333333334, ans=0.02 2024-09-25 22:21:33,263 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.99 vs. limit=22.5 2024-09-25 22:21:34,158 INFO [train.py:1198] (2/4) Epoch 47, batch 1600, loss[loss=0.1916, ctc_loss=0.1218, cr_loss=0.3486, over 17217.00 frames. ], tot_loss[loss=0.1864, ctc_loss=0.1192, cr_loss=0.3362, over 3357956.81 frames. ], batch size: 50, lr: 2.52e-03, grad_scale: 32.0 2024-09-25 22:21:35,966 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.max_abs, batch_count=843812.6666666666, ans=10.0 2024-09-25 22:22:00,306 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=843859.3333333334, ans=0.0 2024-09-25 22:22:05,138 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=843859.3333333334, ans=0.1 2024-09-25 22:22:07,098 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=8.20 vs. limit=15.0 2024-09-25 22:22:08,275 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=843859.3333333334, ans=0.0 2024-09-25 22:22:59,305 INFO [train.py:1198] (2/4) Epoch 47, batch 1650, loss[loss=0.2209, ctc_loss=0.1428, cr_loss=0.3905, over 16681.00 frames. ], tot_loss[loss=0.187, ctc_loss=0.1196, cr_loss=0.3369, over 3359388.57 frames. ], batch size: 61, lr: 2.52e-03, grad_scale: 32.0 2024-09-25 22:23:22,602 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.136e+02 1.287e+02 1.363e+02 1.436e+02 1.851e+02, threshold=2.725e+02, percent-clipped=0.0 2024-09-25 22:23:26,009 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.min_positive, batch_count=844092.6666666666, ans=0.025 2024-09-25 22:23:34,561 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.17 vs. limit=15.0 2024-09-25 22:23:42,143 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=844139.3333333334, ans=0.125 2024-09-25 22:23:51,901 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=844186.0, ans=0.0 2024-09-25 22:24:21,989 INFO [train.py:1198] (2/4) Epoch 47, batch 1700, loss[loss=0.194, ctc_loss=0.1254, cr_loss=0.3429, over 17015.00 frames. ], tot_loss[loss=0.1862, ctc_loss=0.119, cr_loss=0.3358, over 3367155.40 frames. ], batch size: 51, lr: 2.52e-03, grad_scale: 16.0 2024-09-25 22:24:23,853 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=844279.3333333334, ans=0.0 2024-09-25 22:25:33,159 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=844466.0, ans=0.0 2024-09-25 22:25:44,105 INFO [train.py:1198] (2/4) Epoch 47, batch 1750, loss[loss=0.186, ctc_loss=0.1202, cr_loss=0.3288, over 17136.00 frames. ], tot_loss[loss=0.1863, ctc_loss=0.1191, cr_loss=0.3362, over 3360816.27 frames. ], batch size: 48, lr: 2.52e-03, grad_scale: 16.0 2024-09-25 22:26:06,156 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.120e+02 1.334e+02 1.394e+02 1.508e+02 1.965e+02, threshold=2.789e+02, percent-clipped=0.0 2024-09-25 22:26:15,184 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.85 vs. limit=15.0 2024-09-25 22:26:24,587 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.23 vs. limit=10.0 2024-09-25 22:26:40,190 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=844652.6666666666, ans=0.125 2024-09-25 22:26:41,709 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=844652.6666666666, ans=0.0 2024-09-25 22:27:08,725 INFO [train.py:1198] (2/4) Epoch 47, batch 1800, loss[loss=0.1591, ctc_loss=0.09933, cr_loss=0.2988, over 17099.00 frames. ], tot_loss[loss=0.1865, ctc_loss=0.1193, cr_loss=0.3363, over 3362374.81 frames. ], batch size: 43, lr: 2.52e-03, grad_scale: 16.0 2024-09-25 22:27:44,137 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=844839.3333333334, ans=0.1 2024-09-25 22:28:00,353 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=844886.0, ans=0.0 2024-09-25 22:28:31,339 INFO [train.py:1198] (2/4) Epoch 47, batch 1850, loss[loss=0.1873, ctc_loss=0.1204, cr_loss=0.3346, over 17103.00 frames. ], tot_loss[loss=0.1859, ctc_loss=0.1189, cr_loss=0.3353, over 3360567.93 frames. ], batch size: 49, lr: 2.52e-03, grad_scale: 16.0 2024-09-25 22:28:41,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=844979.3333333334, ans=0.2 2024-09-25 22:28:53,365 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.135e+02 1.318e+02 1.395e+02 1.509e+02 1.851e+02, threshold=2.789e+02, percent-clipped=0.0 2024-09-25 22:29:04,995 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=845072.6666666666, ans=0.09899494936611666 2024-09-25 22:29:30,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=845119.3333333334, ans=0.125 2024-09-25 22:29:36,610 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer_ff3.min_abs, batch_count=845166.0, ans=0.2 2024-09-25 22:29:40,412 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.80 vs. limit=15.0 2024-09-25 22:29:44,688 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff3_skip_rate, batch_count=845166.0, ans=0.0 2024-09-25 22:29:49,198 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=845212.6666666666, ans=0.2 2024-09-25 22:29:50,498 INFO [train.py:1198] (2/4) Epoch 47, batch 1900, loss[loss=0.1734, ctc_loss=0.1098, cr_loss=0.3182, over 17002.00 frames. ], tot_loss[loss=0.1864, ctc_loss=0.1193, cr_loss=0.3359, over 3358423.93 frames. ], batch size: 56, lr: 2.52e-03, grad_scale: 16.0 2024-09-25 22:30:34,662 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=845306.0, ans=0.1 2024-09-25 22:31:00,783 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.28 vs. limit=15.0 2024-09-25 22:31:01,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=845399.3333333334, ans=0.0 2024-09-25 22:31:10,010 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=845399.3333333334, ans=0.1 2024-09-25 22:31:12,815 INFO [train.py:1198] (2/4) Epoch 47, batch 1950, loss[loss=0.1796, ctc_loss=0.1149, cr_loss=0.3232, over 17211.00 frames. ], tot_loss[loss=0.1881, ctc_loss=0.1204, cr_loss=0.3385, over 3358190.96 frames. ], batch size: 50, lr: 2.52e-03, grad_scale: 16.0 2024-09-25 22:31:30,714 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=845492.6666666666, ans=0.125 2024-09-25 22:31:35,175 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.125e+02 1.279e+02 1.368e+02 1.443e+02 1.975e+02, threshold=2.736e+02, percent-clipped=0.0 2024-09-25 22:31:52,171 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=845539.3333333334, ans=0.0 2024-09-25 22:31:52,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=845539.3333333334, ans=0.0 2024-09-25 22:31:58,222 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=845539.3333333334, ans=0.0 2024-09-25 22:32:03,336 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.38 vs. limit=15.0 2024-09-25 22:32:09,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.min_positive, batch_count=845586.0, ans=0.05 2024-09-25 22:32:36,586 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module1.whiten, num_groups=1, num_channels=192, metric=8.05 vs. limit=15.0 2024-09-25 22:32:38,557 INFO [train.py:1198] (2/4) Epoch 47, batch 2000, loss[loss=0.2377, ctc_loss=0.1533, cr_loss=0.4219, over 17002.00 frames. ], tot_loss[loss=0.1884, ctc_loss=0.1206, cr_loss=0.339, over 3364930.72 frames. ], batch size: 53, lr: 2.52e-03, grad_scale: 32.0 2024-09-25 22:32:48,787 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=845679.3333333334, ans=0.1 2024-09-25 22:33:08,245 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=16.06 vs. limit=22.5 2024-09-25 22:33:13,904 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=845772.6666666666, ans=0.125 2024-09-25 22:33:20,306 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=845772.6666666666, ans=0.1 2024-09-25 22:33:32,181 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn1.whiten, num_groups=1, num_channels=192, metric=9.15 vs. limit=22.5 2024-09-25 22:33:51,212 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.67 vs. limit=15.0 2024-09-25 22:33:52,626 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=4.86 vs. limit=15.0 2024-09-25 22:33:55,511 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=845866.0, ans=0.2 2024-09-25 22:34:01,510 INFO [train.py:1198] (2/4) Epoch 47, batch 2050, loss[loss=0.1973, ctc_loss=0.1282, cr_loss=0.3458, over 16715.00 frames. ], tot_loss[loss=0.1872, ctc_loss=0.1198, cr_loss=0.337, over 3368584.72 frames. ], batch size: 61, lr: 2.52e-03, grad_scale: 16.0 2024-09-25 22:34:09,756 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=845912.6666666666, ans=0.1 2024-09-25 22:34:25,350 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.136e+02 1.271e+02 1.350e+02 1.448e+02 1.978e+02, threshold=2.700e+02, percent-clipped=0.0 2024-09-25 22:34:43,896 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=13.65 vs. limit=22.5 2024-09-25 22:34:58,092 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.75 vs. limit=22.5 2024-09-25 22:35:00,940 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=846052.6666666666, ans=0.025 2024-09-25 22:35:23,976 INFO [train.py:1198] (2/4) Epoch 47, batch 2100, loss[loss=0.1835, ctc_loss=0.118, cr_loss=0.3276, over 17233.00 frames. ], tot_loss[loss=0.1864, ctc_loss=0.1191, cr_loss=0.3365, over 3368950.43 frames. ], batch size: 50, lr: 2.52e-03, grad_scale: 8.0 2024-09-25 22:35:24,358 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=846146.0, ans=0.1 2024-09-25 22:35:33,996 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=846146.0, ans=0.125 2024-09-25 22:35:42,680 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.50 vs. limit=10.0 2024-09-25 22:35:48,198 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=846192.6666666666, ans=0.2 2024-09-25 22:36:25,177 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=846286.0, ans=0.0 2024-09-25 22:36:27,073 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.86 vs. limit=22.5 2024-09-25 22:36:28,977 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=17.48 vs. limit=22.5 2024-09-25 22:36:31,499 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=846332.6666666666, ans=0.125 2024-09-25 22:36:34,578 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=846332.6666666666, ans=0.1 2024-09-25 22:36:46,573 INFO [train.py:1198] (2/4) Epoch 47, batch 2150, loss[loss=0.1806, ctc_loss=0.1151, cr_loss=0.3276, over 17290.00 frames. ], tot_loss[loss=0.1868, ctc_loss=0.1194, cr_loss=0.337, over 3376196.60 frames. ], batch size: 49, lr: 2.52e-03, grad_scale: 8.0 2024-09-25 22:36:51,589 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=846379.3333333334, ans=0.0 2024-09-25 22:36:58,902 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=846379.3333333334, ans=0.125 2024-09-25 22:36:58,965 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=846379.3333333334, ans=0.0 2024-09-25 22:37:03,818 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=846426.0, ans=0.0 2024-09-25 22:37:14,615 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.166e+02 1.323e+02 1.402e+02 1.477e+02 1.932e+02, threshold=2.805e+02, percent-clipped=0.0 2024-09-25 22:37:21,617 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 22:37:37,466 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=846519.3333333334, ans=0.0 2024-09-25 22:37:44,726 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.03 vs. limit=6.0 2024-09-25 22:38:04,443 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.min_positive, batch_count=846566.0, ans=0.025 2024-09-25 22:38:05,110 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=5.44 vs. limit=15.0 2024-09-25 22:38:12,283 INFO [train.py:1198] (2/4) Epoch 47, batch 2200, loss[loss=0.2063, ctc_loss=0.1343, cr_loss=0.3603, over 17031.00 frames. ], tot_loss[loss=0.1885, ctc_loss=0.1207, cr_loss=0.3389, over 3366646.91 frames. ], batch size: 52, lr: 2.52e-03, grad_scale: 8.0 2024-09-25 22:38:32,035 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=846659.3333333334, ans=0.125 2024-09-25 22:38:57,551 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=846706.0, ans=0.125 2024-09-25 22:39:32,143 INFO [train.py:1198] (2/4) Epoch 47, batch 2250, loss[loss=0.1778, ctc_loss=0.1125, cr_loss=0.3263, over 17083.00 frames. ], tot_loss[loss=0.1878, ctc_loss=0.1202, cr_loss=0.3382, over 3368528.18 frames. ], batch size: 43, lr: 2.52e-03, grad_scale: 8.0 2024-09-25 22:39:50,492 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=846892.6666666666, ans=0.125 2024-09-25 22:39:58,070 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.141e+02 1.312e+02 1.382e+02 1.535e+02 2.267e+02, threshold=2.763e+02, percent-clipped=0.0 2024-09-25 22:40:01,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=846892.6666666666, ans=0.125 2024-09-25 22:40:12,770 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=5.49 vs. limit=15.0 2024-09-25 22:40:20,517 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.whiten, num_groups=1, num_channels=512, metric=4.32 vs. limit=12.0 2024-09-25 22:40:41,141 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=847032.6666666666, ans=0.0 2024-09-25 22:40:55,215 INFO [train.py:1198] (2/4) Epoch 47, batch 2300, loss[loss=0.1955, ctc_loss=0.1231, cr_loss=0.362, over 17223.00 frames. ], tot_loss[loss=0.1873, ctc_loss=0.1199, cr_loss=0.3368, over 3362060.91 frames. ], batch size: 50, lr: 2.52e-03, grad_scale: 8.0 2024-09-25 22:41:06,867 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=847079.3333333334, ans=0.125 2024-09-25 22:41:36,701 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=847172.6666666666, ans=0.125 2024-09-25 22:41:55,506 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=847219.3333333334, ans=0.125 2024-09-25 22:42:20,986 INFO [train.py:1198] (2/4) Epoch 47, batch 2350, loss[loss=0.1856, ctc_loss=0.1171, cr_loss=0.3427, over 17225.00 frames. ], tot_loss[loss=0.1867, ctc_loss=0.1195, cr_loss=0.3361, over 3361317.70 frames. ], batch size: 47, lr: 2.52e-03, grad_scale: 8.0 2024-09-25 22:42:28,258 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=512, metric=14.49 vs. limit=22.5 2024-09-25 22:42:34,457 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=847312.6666666666, ans=0.0 2024-09-25 22:42:40,799 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=847359.3333333334, ans=0.07 2024-09-25 22:42:46,752 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.178e+02 1.289e+02 1.363e+02 1.471e+02 2.156e+02, threshold=2.725e+02, percent-clipped=0.0 2024-09-25 22:43:30,113 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=847499.3333333334, ans=0.125 2024-09-25 22:43:31,004 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.82 vs. limit=6.0 2024-09-25 22:43:44,375 INFO [train.py:1198] (2/4) Epoch 47, batch 2400, loss[loss=0.16, ctc_loss=0.09962, cr_loss=0.3018, over 16706.00 frames. ], tot_loss[loss=0.1861, ctc_loss=0.119, cr_loss=0.3358, over 3366378.09 frames. ], batch size: 37, lr: 2.52e-03, grad_scale: 16.0 2024-09-25 22:43:46,770 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=3.12 vs. limit=6.0 2024-09-25 22:43:52,672 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=847546.0, ans=0.09899494936611666 2024-09-25 22:44:04,379 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=7.52 vs. limit=15.0 2024-09-25 22:44:20,136 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=847639.3333333334, ans=0.2 2024-09-25 22:44:21,647 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=847639.3333333334, ans=0.125 2024-09-25 22:44:29,896 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=9.12 vs. limit=12.0 2024-09-25 22:44:39,368 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=847686.0, ans=0.0 2024-09-25 22:44:52,205 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=847732.6666666666, ans=0.2 2024-09-25 22:45:07,168 INFO [train.py:1198] (2/4) Epoch 47, batch 2450, loss[loss=0.2217, ctc_loss=0.1428, cr_loss=0.3946, over 16902.00 frames. ], tot_loss[loss=0.1857, ctc_loss=0.1187, cr_loss=0.3347, over 3369915.72 frames. ], batch size: 58, lr: 2.52e-03, grad_scale: 16.0 2024-09-25 22:45:15,553 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=847779.3333333334, ans=0.0 2024-09-25 22:45:28,471 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=14.69 vs. limit=15.0 2024-09-25 22:45:32,659 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.112e+02 1.309e+02 1.414e+02 1.505e+02 2.738e+02, threshold=2.828e+02, percent-clipped=1.0 2024-09-25 22:45:32,919 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=847826.0, ans=0.125 2024-09-25 22:45:45,999 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=6.96 vs. limit=15.0 2024-09-25 22:45:46,426 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=11.82 vs. limit=22.5 2024-09-25 22:45:52,251 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=847872.6666666666, ans=0.125 2024-09-25 22:46:08,262 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=847919.3333333334, ans=0.2 2024-09-25 22:46:27,292 INFO [train.py:1198] (2/4) Epoch 47, batch 2500, loss[loss=0.2048, ctc_loss=0.1291, cr_loss=0.3783, over 17363.00 frames. ], tot_loss[loss=0.1857, ctc_loss=0.1187, cr_loss=0.3349, over 3363472.18 frames. ], batch size: 52, lr: 2.52e-03, grad_scale: 16.0 2024-09-25 22:46:29,279 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=848012.6666666666, ans=0.125 2024-09-25 22:47:00,392 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=848059.3333333334, ans=0.025 2024-09-25 22:47:16,434 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=10.10 vs. limit=12.0 2024-09-25 22:47:17,706 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.out_combiner.scale_min, batch_count=848106.0, ans=0.2 2024-09-25 22:47:25,846 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=848152.6666666666, ans=0.0 2024-09-25 22:47:55,379 INFO [train.py:1198] (2/4) Epoch 47, batch 2550, loss[loss=0.1675, ctc_loss=0.1054, cr_loss=0.3104, over 17026.00 frames. ], tot_loss[loss=0.1859, ctc_loss=0.1189, cr_loss=0.335, over 3358572.39 frames. ], batch size: 44, lr: 2.52e-03, grad_scale: 16.0 2024-09-25 22:48:20,933 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.155e+02 1.321e+02 1.406e+02 1.538e+02 2.240e+02, threshold=2.812e+02, percent-clipped=0.0 2024-09-25 22:48:40,609 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=848339.3333333334, ans=0.1 2024-09-25 22:48:48,957 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=848386.0, ans=0.0 2024-09-25 22:49:00,029 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=848432.6666666666, ans=0.0 2024-09-25 22:49:08,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=848432.6666666666, ans=0.125 2024-09-25 22:49:15,634 INFO [train.py:1198] (2/4) Epoch 47, batch 2600, loss[loss=0.1598, ctc_loss=0.1018, cr_loss=0.2902, over 17201.00 frames. ], tot_loss[loss=0.186, ctc_loss=0.119, cr_loss=0.3353, over 3368668.72 frames. ], batch size: 41, lr: 2.52e-03, grad_scale: 16.0 2024-09-25 22:49:23,973 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=848479.3333333334, ans=0.0 2024-09-25 22:49:30,573 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=848526.0, ans=0.2 2024-09-25 22:49:58,699 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=848572.6666666666, ans=0.125 2024-09-25 22:50:08,131 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=848619.3333333334, ans=0.125 2024-09-25 22:50:38,208 INFO [train.py:1198] (2/4) Epoch 47, batch 2650, loss[loss=0.1752, ctc_loss=0.1117, cr_loss=0.3175, over 17210.00 frames. ], tot_loss[loss=0.1864, ctc_loss=0.1192, cr_loss=0.3358, over 3371868.56 frames. ], batch size: 47, lr: 2.52e-03, grad_scale: 16.0 2024-09-25 22:50:45,098 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=848712.6666666666, ans=0.0 2024-09-25 22:51:03,872 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.172e+02 1.306e+02 1.404e+02 1.475e+02 2.254e+02, threshold=2.808e+02, percent-clipped=0.0 2024-09-25 22:51:13,668 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=848806.0, ans=0.125 2024-09-25 22:51:29,868 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=848852.6666666666, ans=0.125 2024-09-25 22:51:42,071 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=848852.6666666666, ans=0.2 2024-09-25 22:51:46,973 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=848899.3333333334, ans=0.125 2024-09-25 22:51:52,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=848899.3333333334, ans=0.0 2024-09-25 22:52:03,698 INFO [train.py:1198] (2/4) Epoch 47, batch 2700, loss[loss=0.2022, ctc_loss=0.1275, cr_loss=0.3733, over 17140.00 frames. ], tot_loss[loss=0.1865, ctc_loss=0.1192, cr_loss=0.3362, over 3378989.89 frames. ], batch size: 48, lr: 2.52e-03, grad_scale: 16.0 2024-09-25 22:52:13,549 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 22:52:16,722 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=848946.0, ans=0.05 2024-09-25 22:52:19,825 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=848992.6666666666, ans=0.1 2024-09-25 22:53:14,914 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=849132.6666666666, ans=10.0 2024-09-25 22:53:25,683 INFO [train.py:1198] (2/4) Epoch 47, batch 2750, loss[loss=0.2197, ctc_loss=0.1437, cr_loss=0.38, over 16549.00 frames. ], tot_loss[loss=0.1881, ctc_loss=0.1205, cr_loss=0.3384, over 3367355.94 frames. ], batch size: 66, lr: 2.52e-03, grad_scale: 16.0 2024-09-25 22:53:27,817 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=12.70 vs. limit=12.0 2024-09-25 22:53:51,198 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.120e+02 1.348e+02 1.413e+02 1.514e+02 2.410e+02, threshold=2.827e+02, percent-clipped=0.0 2024-09-25 22:53:54,853 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff2_skip_rate, batch_count=849226.0, ans=0.0 2024-09-25 22:53:59,699 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=849272.6666666666, ans=0.125 2024-09-25 22:54:26,837 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=849319.3333333334, ans=0.2 2024-09-25 22:54:33,347 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=849366.0, ans=0.125 2024-09-25 22:54:39,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=849366.0, ans=0.0 2024-09-25 22:54:45,976 INFO [train.py:1198] (2/4) Epoch 47, batch 2800, loss[loss=0.1936, ctc_loss=0.1268, cr_loss=0.3343, over 16987.00 frames. ], tot_loss[loss=0.1882, ctc_loss=0.1204, cr_loss=0.3388, over 3357073.54 frames. ], batch size: 56, lr: 2.51e-03, grad_scale: 32.0 2024-09-25 22:54:59,353 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=11.75 vs. limit=22.5 2024-09-25 22:55:28,468 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=849506.0, ans=0.0 2024-09-25 22:55:44,695 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=849552.6666666666, ans=0.035 2024-09-25 22:56:08,515 INFO [train.py:1198] (2/4) Epoch 47, batch 2850, loss[loss=0.1754, ctc_loss=0.114, cr_loss=0.3072, over 17022.00 frames. ], tot_loss[loss=0.188, ctc_loss=0.1202, cr_loss=0.3387, over 3354824.78 frames. ], batch size: 44, lr: 2.51e-03, grad_scale: 16.0 2024-09-25 22:56:33,598 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.min_positive, batch_count=849692.6666666666, ans=0.05 2024-09-25 22:56:38,076 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.322e+02 1.395e+02 1.477e+02 2.122e+02, threshold=2.790e+02, percent-clipped=0.0 2024-09-25 22:57:12,660 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=849786.0, ans=0.1 2024-09-25 22:57:33,014 INFO [train.py:1198] (2/4) Epoch 47, batch 2900, loss[loss=0.158, ctc_loss=0.0979, cr_loss=0.3005, over 17252.00 frames. ], tot_loss[loss=0.1877, ctc_loss=0.12, cr_loss=0.3388, over 3356762.26 frames. ], batch size: 42, lr: 2.51e-03, grad_scale: 8.0 2024-09-25 22:57:35,201 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.13 vs. limit=15.0 2024-09-25 22:57:59,100 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=12.27 vs. limit=22.5 2024-09-25 22:58:06,243 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=849972.6666666666, ans=0.125 2024-09-25 22:58:07,747 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=849972.6666666666, ans=0.125 2024-09-25 22:58:23,114 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.83 vs. limit=6.0 2024-09-25 22:58:39,349 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=11.73 vs. limit=22.5 2024-09-25 22:58:53,194 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.const_attention_rate, batch_count=850066.0, ans=0.025 2024-09-25 22:58:56,180 INFO [train.py:1198] (2/4) Epoch 47, batch 2950, loss[loss=0.1805, ctc_loss=0.1141, cr_loss=0.332, over 17343.00 frames. ], tot_loss[loss=0.1877, ctc_loss=0.1201, cr_loss=0.3384, over 3357178.54 frames. ], batch size: 48, lr: 2.51e-03, grad_scale: 8.0 2024-09-25 22:58:56,599 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=850112.6666666666, ans=0.125 2024-09-25 22:58:59,721 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=850112.6666666666, ans=0.125 2024-09-25 22:59:01,844 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=3.94 vs. limit=6.0 2024-09-25 22:59:03,743 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.33 vs. limit=15.0 2024-09-25 22:59:08,095 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.67 vs. limit=15.0 2024-09-25 22:59:24,951 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.161e+02 1.302e+02 1.410e+02 1.499e+02 2.139e+02, threshold=2.820e+02, percent-clipped=0.0 2024-09-25 22:59:25,366 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=850159.3333333334, ans=0.0 2024-09-25 22:59:36,608 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=850206.0, ans=0.125 2024-09-25 22:59:37,085 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=4.62 vs. limit=15.0 2024-09-25 22:59:55,568 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=850252.6666666666, ans=0.125 2024-09-25 23:00:17,881 INFO [train.py:1198] (2/4) Epoch 47, batch 3000, loss[loss=0.1722, ctc_loss=0.1075, cr_loss=0.3236, over 17024.00 frames. ], tot_loss[loss=0.1869, ctc_loss=0.1195, cr_loss=0.337, over 3364939.69 frames. ], batch size: 44, lr: 2.51e-03, grad_scale: 8.0 2024-09-25 23:00:17,881 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 23:00:33,699 INFO [train.py:1230] (2/4) Epoch 47, validation: loss=0.0348, ctc_loss=0.0348, cr_loss=1.036e-14, over 944034.00 frames. 2024-09-25 23:00:33,700 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 23:00:54,677 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=850392.6666666666, ans=0.95 2024-09-25 23:01:05,574 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=850439.3333333334, ans=0.2 2024-09-25 23:01:22,844 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=850486.0, ans=0.125 2024-09-25 23:01:36,130 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.95 vs. limit=15.0 2024-09-25 23:01:52,183 INFO [train.py:1198] (2/4) Epoch 47, batch 3050, loss[loss=0.2429, ctc_loss=0.1599, cr_loss=0.4149, over 15047.00 frames. ], tot_loss[loss=0.1867, ctc_loss=0.1194, cr_loss=0.3365, over 3363159.82 frames. ], batch size: 89, lr: 2.51e-03, grad_scale: 8.0 2024-09-25 23:02:14,228 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=850626.0, ans=0.0 2024-09-25 23:02:20,304 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.194e+02 1.305e+02 1.413e+02 1.506e+02 4.087e+02, threshold=2.825e+02, percent-clipped=1.0 2024-09-25 23:02:33,437 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=850672.6666666666, ans=0.125 2024-09-25 23:02:33,819 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.conv_module2.whiten, num_groups=1, num_channels=512, metric=8.63 vs. limit=15.0 2024-09-25 23:02:56,135 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=850766.0, ans=0.125 2024-09-25 23:03:13,092 INFO [train.py:1198] (2/4) Epoch 47, batch 3100, loss[loss=0.1618, ctc_loss=0.1024, cr_loss=0.2968, over 17069.00 frames. ], tot_loss[loss=0.1871, ctc_loss=0.1198, cr_loss=0.3364, over 3345514.28 frames. ], batch size: 43, lr: 2.51e-03, grad_scale: 8.0 2024-09-25 23:03:13,297 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=850812.6666666666, ans=0.125 2024-09-25 23:03:48,293 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=850906.0, ans=0.125 2024-09-25 23:04:03,905 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=850952.6666666666, ans=0.125 2024-09-25 23:04:05,631 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=850952.6666666666, ans=0.05 2024-09-25 23:04:11,993 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 23:04:18,226 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=850999.3333333334, ans=0.0 2024-09-25 23:04:22,606 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=850999.3333333334, ans=0.0 2024-09-25 23:04:27,581 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=850999.3333333334, ans=0.125 2024-09-25 23:04:33,493 INFO [train.py:1198] (2/4) Epoch 47, batch 3150, loss[loss=0.1696, ctc_loss=0.1075, cr_loss=0.3107, over 17187.00 frames. ], tot_loss[loss=0.1876, ctc_loss=0.1202, cr_loss=0.3373, over 3340137.21 frames. ], batch size: 41, lr: 2.51e-03, grad_scale: 8.0 2024-09-25 23:04:49,371 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=851092.6666666666, ans=0.0 2024-09-25 23:05:01,409 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.168e+02 1.375e+02 1.455e+02 1.638e+02 2.123e+02, threshold=2.910e+02, percent-clipped=0.0 2024-09-25 23:05:23,568 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=851186.0, ans=0.125 2024-09-25 23:05:33,045 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=851186.0, ans=0.125 2024-09-25 23:05:48,099 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=851232.6666666666, ans=0.125 2024-09-25 23:05:54,051 INFO [train.py:1198] (2/4) Epoch 47, batch 3200, loss[loss=0.1951, ctc_loss=0.1227, cr_loss=0.3625, over 17090.00 frames. ], tot_loss[loss=0.1878, ctc_loss=0.1203, cr_loss=0.3374, over 3342481.61 frames. ], batch size: 43, lr: 2.51e-03, grad_scale: 16.0 2024-09-25 23:05:57,861 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys.whitening_limit, batch_count=851279.3333333334, ans=6.0 2024-09-25 23:06:02,103 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=851279.3333333334, ans=0.2 2024-09-25 23:06:15,075 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=5.96 vs. limit=12.0 2024-09-25 23:06:15,078 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=11.18 vs. limit=15.0 2024-09-25 23:06:49,010 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=851419.3333333334, ans=0.0 2024-09-25 23:07:01,377 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 23:07:03,458 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=8.49 vs. limit=15.0 2024-09-25 23:07:04,564 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=851466.0, ans=0.2 2024-09-25 23:07:12,127 INFO [train.py:1198] (2/4) Epoch 47, batch 3250, loss[loss=0.1715, ctc_loss=0.1083, cr_loss=0.316, over 17082.00 frames. ], tot_loss[loss=0.1869, ctc_loss=0.1196, cr_loss=0.3365, over 3345537.25 frames. ], batch size: 43, lr: 2.51e-03, grad_scale: 16.0 2024-09-25 23:07:16,370 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.59 vs. limit=6.0 2024-09-25 23:07:21,145 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.49 vs. limit=15.0 2024-09-25 23:07:25,672 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=7.51 vs. limit=15.0 2024-09-25 23:07:34,293 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.attention_skip_rate, batch_count=851559.3333333334, ans=0.0 2024-09-25 23:07:40,357 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.291e+02 1.374e+02 1.460e+02 1.837e+02, threshold=2.749e+02, percent-clipped=0.0 2024-09-25 23:08:03,808 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=851652.6666666666, ans=0.125 2024-09-25 23:08:07,146 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=12.78 vs. limit=15.0 2024-09-25 23:08:29,432 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=13.36 vs. limit=15.0 2024-09-25 23:08:30,208 INFO [train.py:1198] (2/4) Epoch 47, batch 3300, loss[loss=0.1968, ctc_loss=0.1259, cr_loss=0.3545, over 17262.00 frames. ], tot_loss[loss=0.1867, ctc_loss=0.1194, cr_loss=0.3365, over 3349502.11 frames. ], batch size: 44, lr: 2.51e-03, grad_scale: 16.0 2024-09-25 23:08:33,756 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=12.14 vs. limit=12.0 2024-09-25 23:09:00,259 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=851839.3333333334, ans=0.125 2024-09-25 23:09:44,281 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=15.48 vs. limit=15.0 2024-09-25 23:09:48,391 INFO [train.py:1198] (2/4) Epoch 47, batch 3350, loss[loss=0.1842, ctc_loss=0.1179, cr_loss=0.3317, over 17353.00 frames. ], tot_loss[loss=0.1863, ctc_loss=0.1192, cr_loss=0.3356, over 3336278.16 frames. ], batch size: 48, lr: 2.51e-03, grad_scale: 8.0 2024-09-25 23:09:48,786 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=851979.3333333334, ans=0.5 2024-09-25 23:10:09,395 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=852026.0, ans=0.07 2024-09-25 23:10:20,045 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.314e+02 1.397e+02 1.497e+02 2.232e+02, threshold=2.793e+02, percent-clipped=0.0 2024-09-25 23:10:35,861 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.min_positive, batch_count=852119.3333333334, ans=0.025 2024-09-25 23:10:42,773 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=8.94 vs. limit=15.0 2024-09-25 23:11:08,897 INFO [train.py:1198] (2/4) Epoch 47, batch 3400, loss[loss=0.177, ctc_loss=0.1173, cr_loss=0.2989, over 12245.00 frames. ], tot_loss[loss=0.1872, ctc_loss=0.1198, cr_loss=0.337, over 3330232.25 frames. ], batch size: 123, lr: 2.51e-03, grad_scale: 8.0 2024-09-25 23:11:13,838 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=852212.6666666666, ans=0.125 2024-09-25 23:11:17,179 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=852212.6666666666, ans=0.125 2024-09-25 23:11:21,959 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=852212.6666666666, ans=0.125 2024-09-25 23:11:23,347 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=852259.3333333334, ans=10.0 2024-09-25 23:11:30,309 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.32 vs. limit=15.0 2024-09-25 23:12:04,002 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=852352.6666666666, ans=0.0 2024-09-25 23:12:08,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=852352.6666666666, ans=0.125 2024-09-25 23:12:21,695 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.68 vs. limit=15.0 2024-09-25 23:12:27,290 INFO [train.py:1198] (2/4) Epoch 47, batch 3450, loss[loss=0.18, ctc_loss=0.114, cr_loss=0.3302, over 17355.00 frames. ], tot_loss[loss=0.1873, ctc_loss=0.1199, cr_loss=0.3373, over 3333016.34 frames. ], batch size: 48, lr: 2.51e-03, grad_scale: 8.0 2024-09-25 23:12:33,149 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=5.94 vs. limit=15.0 2024-09-25 23:12:34,317 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.92 vs. limit=22.5 2024-09-25 23:12:55,690 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=852492.6666666666, ans=0.0 2024-09-25 23:12:56,966 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.186e+02 1.319e+02 1.407e+02 1.479e+02 2.069e+02, threshold=2.813e+02, percent-clipped=0.0 2024-09-25 23:12:57,266 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=852539.3333333334, ans=0.125 2024-09-25 23:13:16,718 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.11 vs. limit=15.0 2024-09-25 23:13:29,106 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=8.93 vs. limit=15.0 2024-09-25 23:13:36,357 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=852632.6666666666, ans=0.125 2024-09-25 23:13:37,988 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=852632.6666666666, ans=0.1 2024-09-25 23:13:49,279 INFO [train.py:1198] (2/4) Epoch 47, batch 3500, loss[loss=0.1647, ctc_loss=0.1019, cr_loss=0.314, over 17035.00 frames. ], tot_loss[loss=0.1871, ctc_loss=0.1197, cr_loss=0.337, over 3335448.05 frames. ], batch size: 39, lr: 2.51e-03, grad_scale: 8.0 2024-09-25 23:13:49,535 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=852679.3333333334, ans=0.1 2024-09-25 23:14:25,473 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=852772.6666666666, ans=0.09899494936611666 2024-09-25 23:14:34,836 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=852819.3333333334, ans=0.0 2024-09-25 23:14:35,239 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.11 vs. limit=22.5 2024-09-25 23:15:07,125 INFO [train.py:1198] (2/4) Epoch 47, batch 3550, loss[loss=0.1752, ctc_loss=0.1107, cr_loss=0.3227, over 17285.00 frames. ], tot_loss[loss=0.1867, ctc_loss=0.1194, cr_loss=0.3363, over 3347684.28 frames. ], batch size: 51, lr: 2.51e-03, grad_scale: 8.0 2024-09-25 23:15:23,329 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.36 vs. limit=10.0 2024-09-25 23:15:38,866 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.176e+02 1.284e+02 1.338e+02 1.473e+02 3.313e+02, threshold=2.677e+02, percent-clipped=1.0 2024-09-25 23:15:40,788 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=853006.0, ans=0.125 2024-09-25 23:15:48,767 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=853006.0, ans=0.1 2024-09-25 23:15:50,153 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=853006.0, ans=0.125 2024-09-25 23:16:01,358 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=853052.6666666666, ans=0.07 2024-09-25 23:16:02,997 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=853052.6666666666, ans=0.1 2024-09-25 23:16:06,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=853052.6666666666, ans=0.1 2024-09-25 23:16:20,615 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.11 vs. limit=12.0 2024-09-25 23:16:21,638 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=853099.3333333334, ans=0.125 2024-09-25 23:16:21,656 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.min_positive, batch_count=853099.3333333334, ans=0.05 2024-09-25 23:16:25,454 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.whiten, num_groups=1, num_channels=192, metric=5.12 vs. limit=12.0 2024-09-25 23:16:27,424 INFO [train.py:1198] (2/4) Epoch 47, batch 3600, loss[loss=0.1968, ctc_loss=0.1253, cr_loss=0.3579, over 16941.00 frames. ], tot_loss[loss=0.1874, ctc_loss=0.1199, cr_loss=0.3374, over 3341634.72 frames. ], batch size: 42, lr: 2.51e-03, grad_scale: 16.0 2024-09-25 23:16:29,177 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.scale_min, batch_count=853146.0, ans=0.2 2024-09-25 23:16:38,549 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=853146.0, ans=0.125 2024-09-25 23:16:40,092 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=853146.0, ans=0.125 2024-09-25 23:16:43,188 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=853192.6666666666, ans=0.1 2024-09-25 23:16:43,277 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=853192.6666666666, ans=0.2 2024-09-25 23:16:47,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=853192.6666666666, ans=0.0 2024-09-25 23:16:56,965 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=853239.3333333334, ans=0.125 2024-09-25 23:17:20,424 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=853286.0, ans=0.125 2024-09-25 23:17:33,630 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=14.39 vs. limit=22.5 2024-09-25 23:17:34,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=853332.6666666666, ans=0.125 2024-09-25 23:17:40,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=853332.6666666666, ans=0.0 2024-09-25 23:17:45,126 INFO [train.py:1198] (2/4) Epoch 47, batch 3650, loss[loss=0.2032, ctc_loss=0.133, cr_loss=0.3509, over 17027.00 frames. ], tot_loss[loss=0.1878, ctc_loss=0.1202, cr_loss=0.338, over 3346920.04 frames. ], batch size: 52, lr: 2.51e-03, grad_scale: 16.0 2024-09-25 23:17:48,581 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 23:17:50,152 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=853379.3333333334, ans=0.125 2024-09-25 23:18:12,655 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=7.27 vs. limit=15.0 2024-09-25 23:18:13,791 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=853426.0, ans=0.025 2024-09-25 23:18:14,971 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.311e+02 1.426e+02 1.502e+02 2.686e+02, threshold=2.852e+02, percent-clipped=1.0 2024-09-25 23:18:18,392 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=853472.6666666666, ans=0.0 2024-09-25 23:18:28,559 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=853472.6666666666, ans=0.1 2024-09-25 23:18:53,646 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 23:19:04,231 INFO [train.py:1198] (2/4) Epoch 47, batch 3700, loss[loss=0.2113, ctc_loss=0.1366, cr_loss=0.3733, over 17257.00 frames. ], tot_loss[loss=0.1872, ctc_loss=0.1198, cr_loss=0.3372, over 3352331.61 frames. ], batch size: 44, lr: 2.51e-03, grad_scale: 16.0 2024-09-25 23:19:27,526 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=853659.3333333334, ans=0.0 2024-09-25 23:19:41,260 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=853706.0, ans=0.125 2024-09-25 23:19:47,496 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=853706.0, ans=0.2 2024-09-25 23:19:53,940 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=853752.6666666666, ans=0.0 2024-09-25 23:19:55,531 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=853752.6666666666, ans=0.1 2024-09-25 23:19:56,068 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=11.86 vs. limit=15.0 2024-09-25 23:20:04,627 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=853752.6666666666, ans=0.015 2024-09-25 23:20:09,713 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=853799.3333333334, ans=0.125 2024-09-25 23:20:23,763 INFO [train.py:1198] (2/4) Epoch 47, batch 3750, loss[loss=0.2006, ctc_loss=0.1269, cr_loss=0.3682, over 17306.00 frames. ], tot_loss[loss=0.1861, ctc_loss=0.119, cr_loss=0.3359, over 3351017.89 frames. ], batch size: 49, lr: 2.51e-03, grad_scale: 16.0 2024-09-25 23:20:44,270 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=853892.6666666666, ans=0.1 2024-09-25 23:20:53,580 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.156e+02 1.332e+02 1.393e+02 1.534e+02 1.821e+02, threshold=2.786e+02, percent-clipped=0.0 2024-09-25 23:21:43,155 INFO [train.py:1198] (2/4) Epoch 47, batch 3800, loss[loss=0.2217, ctc_loss=0.1487, cr_loss=0.3647, over 11635.00 frames. ], tot_loss[loss=0.1853, ctc_loss=0.1183, cr_loss=0.335, over 3349741.46 frames. ], batch size: 123, lr: 2.51e-03, grad_scale: 16.0 2024-09-25 23:21:54,958 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.78 vs. limit=22.5 2024-09-25 23:22:01,001 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=854126.0, ans=0.2 2024-09-25 23:22:05,736 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=854126.0, ans=0.125 2024-09-25 23:22:07,383 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=854126.0, ans=0.125 2024-09-25 23:22:19,905 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=854172.6666666666, ans=0.125 2024-09-25 23:22:21,485 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=854172.6666666666, ans=0.07 2024-09-25 23:22:27,871 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=854172.6666666666, ans=0.125 2024-09-25 23:22:31,561 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=854219.3333333334, ans=0.0 2024-09-25 23:23:03,554 INFO [train.py:1198] (2/4) Epoch 47, batch 3850, loss[loss=0.1956, ctc_loss=0.1234, cr_loss=0.3613, over 16977.00 frames. ], tot_loss[loss=0.1833, ctc_loss=0.1169, cr_loss=0.3321, over 3334253.80 frames. ], batch size: 42, lr: 2.51e-03, grad_scale: 16.0 2024-09-25 23:23:32,701 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.186e+02 1.338e+02 1.449e+02 1.615e+02 3.869e+02, threshold=2.899e+02, percent-clipped=2.0 2024-09-25 23:23:43,575 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=854406.0, ans=0.09899494936611666 2024-09-25 23:23:51,412 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-25 23:24:00,261 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=854452.6666666666, ans=0.2 2024-09-25 23:25:01,208 INFO [train.py:1198] (2/4) Epoch 48, batch 0, loss[loss=0.1935, ctc_loss=0.1232, cr_loss=0.3514, over 17090.00 frames. ], tot_loss[loss=0.1935, ctc_loss=0.1232, cr_loss=0.3514, over 17090.00 frames. ], batch size: 49, lr: 2.48e-03, grad_scale: 32.0 2024-09-25 23:25:01,208 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-25 23:25:16,464 INFO [train.py:1230] (2/4) Epoch 48, validation: loss=0.0347, ctc_loss=0.0347, cr_loss=1.045e-14, over 944034.00 frames. 2024-09-25 23:25:16,465 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-25 23:25:52,233 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=854620.6666666666, ans=0.125 2024-09-25 23:26:16,698 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=854667.3333333334, ans=0.0 2024-09-25 23:26:21,694 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=854714.0, ans=0.125 2024-09-25 23:26:38,950 INFO [train.py:1198] (2/4) Epoch 48, batch 50, loss[loss=0.1895, ctc_loss=0.123, cr_loss=0.3325, over 17133.00 frames. ], tot_loss[loss=0.1891, ctc_loss=0.1211, cr_loss=0.3402, over 751331.98 frames. ], batch size: 48, lr: 2.48e-03, grad_scale: 32.0 2024-09-25 23:27:11,080 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=854854.0, ans=0.125 2024-09-25 23:27:12,388 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=854854.0, ans=0.0 2024-09-25 23:27:16,936 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.326e+02 1.422e+02 1.657e+02 2.406e+02, threshold=2.844e+02, percent-clipped=0.0 2024-09-25 23:27:19,762 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=9.06 vs. limit=15.0 2024-09-25 23:27:46,668 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer1.prob, batch_count=854947.3333333334, ans=0.125 2024-09-25 23:27:50,103 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=854947.3333333334, ans=0.0 2024-09-25 23:27:51,649 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=854947.3333333334, ans=0.1 2024-09-25 23:28:02,159 INFO [train.py:1198] (2/4) Epoch 48, batch 100, loss[loss=0.1689, ctc_loss=0.1061, cr_loss=0.3138, over 17299.00 frames. ], tot_loss[loss=0.188, ctc_loss=0.1201, cr_loss=0.3398, over 1333767.87 frames. ], batch size: 46, lr: 2.48e-03, grad_scale: 16.0 2024-09-25 23:28:08,817 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=854994.0, ans=0.125 2024-09-25 23:28:15,206 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=854994.0, ans=0.1 2024-09-25 23:28:23,213 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=855040.6666666666, ans=0.1 2024-09-25 23:28:36,165 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=855087.3333333334, ans=0.125 2024-09-25 23:28:47,443 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=855087.3333333334, ans=0.0 2024-09-25 23:29:14,294 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=855180.6666666666, ans=0.125 2024-09-25 23:29:22,470 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=855180.6666666666, ans=0.0 2024-09-25 23:29:25,170 INFO [train.py:1198] (2/4) Epoch 48, batch 150, loss[loss=0.1876, ctc_loss=0.1224, cr_loss=0.3262, over 17055.00 frames. ], tot_loss[loss=0.1885, ctc_loss=0.1204, cr_loss=0.3403, over 1783520.27 frames. ], batch size: 56, lr: 2.48e-03, grad_scale: 16.0 2024-09-25 23:29:25,510 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=855227.3333333334, ans=0.125 2024-09-25 23:29:28,670 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=855227.3333333334, ans=0.125 2024-09-25 23:29:31,704 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=855227.3333333334, ans=0.125 2024-09-25 23:29:56,788 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=855274.0, ans=0.2 2024-09-25 23:30:05,999 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.318e+02 1.406e+02 1.496e+02 3.141e+02, threshold=2.812e+02, percent-clipped=1.0 2024-09-25 23:30:09,641 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-25 23:30:09,646 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.max_abs, batch_count=855320.6666666666, ans=10.0 2024-09-25 23:30:27,511 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=855367.3333333334, ans=0.0 2024-09-25 23:30:34,104 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.05 vs. limit=15.0 2024-09-25 23:30:36,925 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=855414.0, ans=10.0 2024-09-25 23:30:47,886 INFO [train.py:1198] (2/4) Epoch 48, batch 200, loss[loss=0.2332, ctc_loss=0.1539, cr_loss=0.3965, over 11816.00 frames. ], tot_loss[loss=0.1867, ctc_loss=0.1192, cr_loss=0.3378, over 2124369.48 frames. ], batch size: 123, lr: 2.48e-03, grad_scale: 16.0 2024-09-25 23:31:13,448 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=855507.3333333334, ans=0.1 2024-09-25 23:31:18,518 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=7.77 vs. limit=15.0 2024-09-25 23:31:25,987 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=855554.0, ans=0.125 2024-09-25 23:31:37,357 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=855600.6666666666, ans=0.125 2024-09-25 23:32:10,689 INFO [train.py:1198] (2/4) Epoch 48, batch 250, loss[loss=0.1804, ctc_loss=0.1127, cr_loss=0.3387, over 17016.00 frames. ], tot_loss[loss=0.1865, ctc_loss=0.119, cr_loss=0.3372, over 2401330.35 frames. ], batch size: 44, lr: 2.48e-03, grad_scale: 16.0 2024-09-25 23:32:12,711 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=855694.0, ans=0.0 2024-09-25 23:32:29,253 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.06 vs. limit=15.0 2024-09-25 23:32:33,660 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.69 vs. limit=22.5 2024-09-25 23:32:49,017 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.111e+02 1.307e+02 1.374e+02 1.472e+02 2.103e+02, threshold=2.747e+02, percent-clipped=0.0 2024-09-25 23:33:00,048 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=855834.0, ans=0.1 2024-09-25 23:33:29,748 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.05 vs. limit=15.0 2024-09-25 23:33:33,796 INFO [train.py:1198] (2/4) Epoch 48, batch 300, loss[loss=0.1758, ctc_loss=0.1156, cr_loss=0.3009, over 16053.00 frames. ], tot_loss[loss=0.185, ctc_loss=0.118, cr_loss=0.3349, over 2622053.13 frames. ], batch size: 74, lr: 2.48e-03, grad_scale: 16.0 2024-09-25 23:33:38,847 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=855927.3333333334, ans=0.0 2024-09-25 23:33:55,934 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=855974.0, ans=0.09899494936611666 2024-09-25 23:34:39,029 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=856114.0, ans=0.125 2024-09-25 23:34:59,435 INFO [train.py:1198] (2/4) Epoch 48, batch 350, loss[loss=0.2032, ctc_loss=0.1314, cr_loss=0.359, over 17019.00 frames. ], tot_loss[loss=0.1845, ctc_loss=0.1177, cr_loss=0.334, over 2794722.25 frames. ], batch size: 51, lr: 2.48e-03, grad_scale: 16.0 2024-09-25 23:35:10,844 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=856160.6666666666, ans=0.0 2024-09-25 23:35:28,348 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=856207.3333333334, ans=0.125 2024-09-25 23:35:35,504 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=13.12 vs. limit=22.5 2024-09-25 23:35:37,610 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.143e+02 1.298e+02 1.363e+02 1.438e+02 2.006e+02, threshold=2.725e+02, percent-clipped=0.0 2024-09-25 23:35:43,505 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=3.90 vs. limit=15.0 2024-09-25 23:35:44,323 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=856254.0, ans=0.125 2024-09-25 23:36:06,621 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=856347.3333333334, ans=0.0 2024-09-25 23:36:12,983 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.prob, batch_count=856347.3333333334, ans=0.125 2024-09-25 23:36:18,621 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.conv_module2.whiten, num_groups=1, num_channels=192, metric=7.59 vs. limit=15.0 2024-09-25 23:36:22,273 INFO [train.py:1198] (2/4) Epoch 48, batch 400, loss[loss=0.1518, ctc_loss=0.09667, cr_loss=0.2756, over 16667.00 frames. ], tot_loss[loss=0.1851, ctc_loss=0.1182, cr_loss=0.3344, over 2922334.76 frames. ], batch size: 37, lr: 2.48e-03, grad_scale: 32.0 2024-09-25 23:36:37,037 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=856440.6666666666, ans=0.125 2024-09-25 23:36:47,591 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=6.66 vs. limit=15.0 2024-09-25 23:37:20,037 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=856534.0, ans=0.125 2024-09-25 23:37:20,085 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=856534.0, ans=0.125 2024-09-25 23:37:26,549 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=856580.6666666666, ans=0.125 2024-09-25 23:37:39,424 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=856580.6666666666, ans=0.09899494936611666 2024-09-25 23:37:42,215 INFO [train.py:1198] (2/4) Epoch 48, batch 450, loss[loss=0.1786, ctc_loss=0.1148, cr_loss=0.3188, over 17208.00 frames. ], tot_loss[loss=0.1847, ctc_loss=0.1179, cr_loss=0.334, over 3016961.18 frames. ], batch size: 50, lr: 2.48e-03, grad_scale: 32.0 2024-09-25 23:37:47,686 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=6.44 vs. limit=15.0 2024-09-25 23:37:50,564 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=856627.3333333334, ans=0.1 2024-09-25 23:38:01,187 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=856674.0, ans=0.0 2024-09-25 23:38:23,492 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.150e+02 1.314e+02 1.437e+02 1.529e+02 1.990e+02, threshold=2.873e+02, percent-clipped=0.0 2024-09-25 23:38:25,539 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=856720.6666666666, ans=0.1 2024-09-25 23:39:07,875 INFO [train.py:1198] (2/4) Epoch 48, batch 500, loss[loss=0.1546, ctc_loss=0.09628, cr_loss=0.2914, over 17268.00 frames. ], tot_loss[loss=0.1849, ctc_loss=0.118, cr_loss=0.3344, over 3098338.66 frames. ], batch size: 42, lr: 2.48e-03, grad_scale: 32.0 2024-09-25 23:39:17,863 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=856860.6666666666, ans=0.1 2024-09-25 23:39:40,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=856954.0, ans=0.125 2024-09-25 23:40:08,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=857000.6666666666, ans=0.125 2024-09-25 23:40:13,974 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=4.87 vs. limit=15.0 2024-09-25 23:40:30,796 INFO [train.py:1198] (2/4) Epoch 48, batch 550, loss[loss=0.1966, ctc_loss=0.1281, cr_loss=0.3424, over 17015.00 frames. ], tot_loss[loss=0.1846, ctc_loss=0.1178, cr_loss=0.3341, over 3156483.36 frames. ], batch size: 56, lr: 2.48e-03, grad_scale: 32.0 2024-09-25 23:40:42,541 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=8.40 vs. limit=15.0 2024-09-25 23:41:04,199 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=857187.3333333334, ans=0.125 2024-09-25 23:41:11,883 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.169e+02 1.319e+02 1.428e+02 1.533e+02 1.901e+02, threshold=2.856e+02, percent-clipped=0.0 2024-09-25 23:41:25,200 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=857234.0, ans=0.125 2024-09-25 23:41:39,594 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=857280.6666666666, ans=0.1 2024-09-25 23:41:53,734 INFO [train.py:1198] (2/4) Epoch 48, batch 600, loss[loss=0.2053, ctc_loss=0.132, cr_loss=0.3665, over 16669.00 frames. ], tot_loss[loss=0.185, ctc_loss=0.1181, cr_loss=0.3343, over 3205515.78 frames. ], batch size: 61, lr: 2.48e-03, grad_scale: 32.0 2024-09-25 23:42:03,744 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.min_positive, batch_count=857327.3333333334, ans=0.025 2024-09-25 23:42:08,452 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.prob, batch_count=857374.0, ans=0.125 2024-09-25 23:42:16,279 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=857374.0, ans=0.0 2024-09-25 23:42:26,787 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.61 vs. limit=5.0 2024-09-25 23:42:27,775 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=857420.6666666666, ans=0.125 2024-09-25 23:42:32,501 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=857420.6666666666, ans=0.125 2024-09-25 23:42:42,084 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.skip_rate, batch_count=857467.3333333334, ans=0.04949747468305833 2024-09-25 23:42:57,404 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=857467.3333333334, ans=0.2 2024-09-25 23:43:16,491 INFO [train.py:1198] (2/4) Epoch 48, batch 650, loss[loss=0.2008, ctc_loss=0.13, cr_loss=0.3539, over 17230.00 frames. ], tot_loss[loss=0.1852, ctc_loss=0.1183, cr_loss=0.3345, over 3232363.15 frames. ], batch size: 50, lr: 2.48e-03, grad_scale: 16.0 2024-09-25 23:43:34,522 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=857607.3333333334, ans=0.125 2024-09-25 23:43:59,395 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 1.322e+02 1.414e+02 1.494e+02 2.236e+02, threshold=2.829e+02, percent-clipped=0.0 2024-09-25 23:44:01,254 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=857654.0, ans=0.1 2024-09-25 23:44:06,147 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=857700.6666666666, ans=0.125 2024-09-25 23:44:06,254 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=857700.6666666666, ans=0.2 2024-09-25 23:44:12,764 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=857700.6666666666, ans=0.125 2024-09-25 23:44:40,139 INFO [train.py:1198] (2/4) Epoch 48, batch 700, loss[loss=0.2003, ctc_loss=0.129, cr_loss=0.3565, over 16118.00 frames. ], tot_loss[loss=0.1838, ctc_loss=0.1172, cr_loss=0.3329, over 3271527.24 frames. ], batch size: 74, lr: 2.48e-03, grad_scale: 16.0 2024-09-25 23:45:08,419 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=857840.6666666666, ans=0.0 2024-09-25 23:45:56,343 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=857980.6666666666, ans=0.125 2024-09-25 23:46:05,007 INFO [train.py:1198] (2/4) Epoch 48, batch 750, loss[loss=0.1645, ctc_loss=0.1029, cr_loss=0.3081, over 17018.00 frames. ], tot_loss[loss=0.1835, ctc_loss=0.117, cr_loss=0.3323, over 3287950.23 frames. ], batch size: 44, lr: 2.48e-03, grad_scale: 16.0 2024-09-25 23:46:10,369 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.76 vs. limit=15.0 2024-09-25 23:46:11,851 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=858027.3333333334, ans=0.0 2024-09-25 23:46:11,867 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=858027.3333333334, ans=0.07 2024-09-25 23:46:16,563 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=858027.3333333334, ans=0.125 2024-09-25 23:46:21,485 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=858074.0, ans=0.0 2024-09-25 23:46:30,940 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=858074.0, ans=0.125 2024-09-25 23:46:43,619 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=858120.6666666666, ans=0.1 2024-09-25 23:46:44,963 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.152e+02 1.340e+02 1.448e+02 1.538e+02 3.559e+02, threshold=2.895e+02, percent-clipped=1.0 2024-09-25 23:47:19,078 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=858214.0, ans=0.05 2024-09-25 23:47:25,136 INFO [train.py:1198] (2/4) Epoch 48, batch 800, loss[loss=0.166, ctc_loss=0.1068, cr_loss=0.2961, over 17213.00 frames. ], tot_loss[loss=0.1828, ctc_loss=0.1165, cr_loss=0.3313, over 3312745.29 frames. ], batch size: 47, lr: 2.48e-03, grad_scale: 32.0 2024-09-25 23:47:33,744 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=858260.6666666666, ans=0.0 2024-09-25 23:47:38,772 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=13.50 vs. limit=15.0 2024-09-25 23:47:56,020 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=858354.0, ans=0.1 2024-09-25 23:48:00,175 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.min_positive, batch_count=858354.0, ans=0.05 2024-09-25 23:48:08,293 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=858354.0, ans=0.2 2024-09-25 23:48:16,299 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=858400.6666666666, ans=0.1 2024-09-25 23:48:19,568 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=858400.6666666666, ans=0.0 2024-09-25 23:48:32,777 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.35 vs. limit=15.0 2024-09-25 23:48:47,864 INFO [train.py:1198] (2/4) Epoch 48, batch 850, loss[loss=0.1604, ctc_loss=0.1014, cr_loss=0.2954, over 16939.00 frames. ], tot_loss[loss=0.1833, ctc_loss=0.117, cr_loss=0.3317, over 3320380.02 frames. ], batch size: 42, lr: 2.48e-03, grad_scale: 32.0 2024-09-25 23:49:05,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.prob, batch_count=858540.6666666666, ans=0.125 2024-09-25 23:49:13,591 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=858540.6666666666, ans=0.125 2024-09-25 23:49:14,003 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_whiten.whitening_limit, batch_count=858540.6666666666, ans=15.0 2024-09-25 23:49:14,065 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=9.54 vs. limit=15.0 2024-09-25 23:49:21,669 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=858587.3333333334, ans=0.125 2024-09-25 23:49:30,829 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.140e+02 1.305e+02 1.378e+02 1.479e+02 2.667e+02, threshold=2.756e+02, percent-clipped=0.0 2024-09-25 23:50:15,698 INFO [train.py:1198] (2/4) Epoch 48, batch 900, loss[loss=0.2024, ctc_loss=0.1313, cr_loss=0.3554, over 15037.00 frames. ], tot_loss[loss=0.1838, ctc_loss=0.1173, cr_loss=0.3325, over 3323540.94 frames. ], batch size: 89, lr: 2.47e-03, grad_scale: 32.0 2024-09-25 23:50:20,737 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=858727.3333333334, ans=0.0 2024-09-25 23:50:22,273 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=858727.3333333334, ans=0.125 2024-09-25 23:50:26,878 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=858727.3333333334, ans=0.1 2024-09-25 23:50:28,411 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=858727.3333333334, ans=0.025 2024-09-25 23:50:34,913 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=858774.0, ans=0.09899494936611666 2024-09-25 23:50:47,508 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=858820.6666666666, ans=0.125 2024-09-25 23:50:55,698 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=858820.6666666666, ans=0.025 2024-09-25 23:50:57,734 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=4.58 vs. limit=10.0 2024-09-25 23:50:58,801 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=858820.6666666666, ans=0.0 2024-09-25 23:51:02,266 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=7.26 vs. limit=15.0 2024-09-25 23:51:18,950 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=858867.3333333334, ans=0.2 2024-09-25 23:51:30,505 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=8.17 vs. limit=15.0 2024-09-25 23:51:37,631 INFO [train.py:1198] (2/4) Epoch 48, batch 950, loss[loss=0.1913, ctc_loss=0.1233, cr_loss=0.3399, over 17316.00 frames. ], tot_loss[loss=0.1838, ctc_loss=0.1173, cr_loss=0.3327, over 3334695.90 frames. ], batch size: 49, lr: 2.47e-03, grad_scale: 16.0 2024-09-25 23:51:39,742 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=858960.6666666666, ans=0.04949747468305833 2024-09-25 23:51:46,102 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer1.max_abs, batch_count=858960.6666666666, ans=10.0 2024-09-25 23:51:52,360 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=859007.3333333334, ans=0.125 2024-09-25 23:51:54,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=859007.3333333334, ans=0.0 2024-09-25 23:51:55,691 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=859007.3333333334, ans=0.125 2024-09-25 23:51:57,721 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.12 vs. limit=6.0 2024-09-25 23:52:19,388 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.144e+02 1.296e+02 1.387e+02 1.508e+02 2.071e+02, threshold=2.775e+02, percent-clipped=0.0 2024-09-25 23:52:32,412 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=859100.6666666666, ans=0.1 2024-09-25 23:53:00,135 INFO [train.py:1198] (2/4) Epoch 48, batch 1000, loss[loss=0.1946, ctc_loss=0.1233, cr_loss=0.3567, over 17161.00 frames. ], tot_loss[loss=0.1852, ctc_loss=0.1182, cr_loss=0.3348, over 3350449.94 frames. ], batch size: 45, lr: 2.47e-03, grad_scale: 16.0 2024-09-25 23:53:56,202 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=859334.0, ans=0.125 2024-09-25 23:54:01,194 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.79 vs. limit=22.5 2024-09-25 23:54:22,550 INFO [train.py:1198] (2/4) Epoch 48, batch 1050, loss[loss=0.1934, ctc_loss=0.1245, cr_loss=0.3442, over 17333.00 frames. ], tot_loss[loss=0.1859, ctc_loss=0.1189, cr_loss=0.3353, over 3350920.94 frames. ], batch size: 48, lr: 2.47e-03, grad_scale: 16.0 2024-09-25 23:54:24,584 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=859427.3333333334, ans=0.0 2024-09-25 23:54:30,884 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=859427.3333333334, ans=0.015 2024-09-25 23:54:32,590 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=859427.3333333334, ans=0.0 2024-09-25 23:55:06,816 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.127e+02 1.301e+02 1.364e+02 1.474e+02 5.159e+02, threshold=2.728e+02, percent-clipped=1.0 2024-09-25 23:55:18,407 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff2_skip_rate, batch_count=859567.3333333334, ans=0.0 2024-09-25 23:55:19,898 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=859567.3333333334, ans=0.125 2024-09-25 23:55:23,082 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=859567.3333333334, ans=0.125 2024-09-25 23:55:45,300 INFO [train.py:1198] (2/4) Epoch 48, batch 1100, loss[loss=0.1884, ctc_loss=0.1216, cr_loss=0.3342, over 17307.00 frames. ], tot_loss[loss=0.1853, ctc_loss=0.1185, cr_loss=0.3342, over 3351219.59 frames. ], batch size: 49, lr: 2.47e-03, grad_scale: 16.0 2024-09-25 23:55:47,296 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=859660.6666666666, ans=0.05 2024-09-25 23:55:55,345 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=859660.6666666666, ans=0.0 2024-09-25 23:56:14,317 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.43 vs. limit=22.5 2024-09-25 23:56:18,568 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=859754.0, ans=0.2 2024-09-25 23:56:21,806 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=859754.0, ans=0.125 2024-09-25 23:56:23,340 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=859754.0, ans=0.0 2024-09-25 23:56:26,654 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-25 23:56:45,837 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=859800.6666666666, ans=10.0 2024-09-25 23:56:46,077 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=8.74 vs. limit=15.0 2024-09-25 23:57:07,815 INFO [train.py:1198] (2/4) Epoch 48, batch 1150, loss[loss=0.1621, ctc_loss=0.1041, cr_loss=0.29, over 17293.00 frames. ], tot_loss[loss=0.1859, ctc_loss=0.1189, cr_loss=0.3349, over 3348820.54 frames. ], batch size: 51, lr: 2.47e-03, grad_scale: 16.0 2024-09-25 23:57:17,705 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=859894.0, ans=0.125 2024-09-25 23:57:25,570 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=859940.6666666666, ans=0.07 2024-09-25 23:57:35,198 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=859940.6666666666, ans=0.125 2024-09-25 23:57:52,001 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.124e+02 1.298e+02 1.394e+02 1.477e+02 1.999e+02, threshold=2.788e+02, percent-clipped=0.0 2024-09-25 23:58:08,302 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=860034.0, ans=0.09899494936611666 2024-09-25 23:58:08,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=860034.0, ans=0.1 2024-09-25 23:58:25,777 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=860080.6666666666, ans=0.125 2024-09-25 23:58:27,454 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=860080.6666666666, ans=0.125 2024-09-25 23:58:30,253 INFO [train.py:1198] (2/4) Epoch 48, batch 1200, loss[loss=0.1935, ctc_loss=0.1256, cr_loss=0.3397, over 16968.00 frames. ], tot_loss[loss=0.1864, ctc_loss=0.1193, cr_loss=0.3354, over 3345130.59 frames. ], batch size: 56, lr: 2.47e-03, grad_scale: 32.0 2024-09-25 23:58:54,141 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=860174.0, ans=0.025 2024-09-25 23:59:19,794 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=860267.3333333334, ans=0.1 2024-09-25 23:59:24,754 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer1.prob, batch_count=860267.3333333334, ans=0.125 2024-09-25 23:59:45,130 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer2.prob, batch_count=860314.0, ans=0.125 2024-09-25 23:59:56,069 INFO [train.py:1198] (2/4) Epoch 48, batch 1250, loss[loss=0.1987, ctc_loss=0.1286, cr_loss=0.3505, over 17133.00 frames. ], tot_loss[loss=0.1856, ctc_loss=0.1187, cr_loss=0.3345, over 3342049.40 frames. ], batch size: 48, lr: 2.47e-03, grad_scale: 16.0 2024-09-26 00:00:01,894 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.93 vs. limit=10.0 2024-09-26 00:00:11,096 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.31 vs. limit=15.0 2024-09-26 00:00:12,150 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=860407.3333333334, ans=0.125 2024-09-26 00:00:39,299 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.147e+02 1.302e+02 1.385e+02 1.504e+02 2.588e+02, threshold=2.770e+02, percent-clipped=0.0 2024-09-26 00:00:49,120 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=860500.6666666666, ans=0.2 2024-09-26 00:01:03,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=860547.3333333334, ans=0.2 2024-09-26 00:01:18,608 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=860594.0, ans=0.07 2024-09-26 00:01:19,890 INFO [train.py:1198] (2/4) Epoch 48, batch 1300, loss[loss=0.1588, ctc_loss=0.1025, cr_loss=0.2815, over 17172.00 frames. ], tot_loss[loss=0.1858, ctc_loss=0.1188, cr_loss=0.3349, over 3347248.98 frames. ], batch size: 41, lr: 2.47e-03, grad_scale: 16.0 2024-09-26 00:01:23,287 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=860594.0, ans=0.1 2024-09-26 00:01:29,859 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=860594.0, ans=0.2 2024-09-26 00:01:32,536 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=9.65 vs. limit=22.5 2024-09-26 00:01:33,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=860594.0, ans=0.125 2024-09-26 00:01:41,543 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=860640.6666666666, ans=0.0 2024-09-26 00:01:49,470 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=860640.6666666666, ans=0.0 2024-09-26 00:01:49,545 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=860640.6666666666, ans=0.125 2024-09-26 00:02:05,793 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=860687.3333333334, ans=0.1 2024-09-26 00:02:40,987 INFO [train.py:1198] (2/4) Epoch 48, batch 1350, loss[loss=0.1857, ctc_loss=0.1153, cr_loss=0.3519, over 17359.00 frames. ], tot_loss[loss=0.1858, ctc_loss=0.1187, cr_loss=0.3353, over 3355796.54 frames. ], batch size: 48, lr: 2.47e-03, grad_scale: 16.0 2024-09-26 00:02:56,573 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=860827.3333333334, ans=0.2 2024-09-26 00:02:56,634 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=860827.3333333334, ans=0.125 2024-09-26 00:02:56,686 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=860827.3333333334, ans=0.0 2024-09-26 00:03:11,051 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=860874.0, ans=0.0 2024-09-26 00:03:14,929 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.64 vs. limit=15.0 2024-09-26 00:03:19,268 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=860920.6666666666, ans=0.1 2024-09-26 00:03:26,899 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.141e+02 1.284e+02 1.347e+02 1.442e+02 2.015e+02, threshold=2.695e+02, percent-clipped=0.0 2024-09-26 00:04:07,154 INFO [train.py:1198] (2/4) Epoch 48, batch 1400, loss[loss=0.1912, ctc_loss=0.1221, cr_loss=0.3452, over 17270.00 frames. ], tot_loss[loss=0.1859, ctc_loss=0.1189, cr_loss=0.3351, over 3362187.42 frames. ], batch size: 44, lr: 2.47e-03, grad_scale: 16.0 2024-09-26 00:04:09,092 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=861060.6666666666, ans=0.125 2024-09-26 00:04:23,549 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=861107.3333333334, ans=0.125 2024-09-26 00:04:39,720 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.ff2_skip_rate, batch_count=861154.0, ans=0.0 2024-09-26 00:05:12,946 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.max_abs, batch_count=861247.3333333334, ans=10.0 2024-09-26 00:05:26,357 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.24 vs. limit=15.0 2024-09-26 00:05:30,162 INFO [train.py:1198] (2/4) Epoch 48, batch 1450, loss[loss=0.1886, ctc_loss=0.1204, cr_loss=0.3409, over 17346.00 frames. ], tot_loss[loss=0.1846, ctc_loss=0.1179, cr_loss=0.3332, over 3364125.45 frames. ], batch size: 48, lr: 2.47e-03, grad_scale: 16.0 2024-09-26 00:05:43,329 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=861294.0, ans=0.2 2024-09-26 00:05:48,249 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=861340.6666666666, ans=0.1 2024-09-26 00:05:56,396 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=9.45 vs. limit=15.0 2024-09-26 00:05:59,385 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-26 00:06:08,448 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=861387.3333333334, ans=0.09899494936611666 2024-09-26 00:06:15,972 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.169e+02 1.346e+02 1.409e+02 1.541e+02 2.437e+02, threshold=2.817e+02, percent-clipped=0.0 2024-09-26 00:06:30,964 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=861434.0, ans=0.04949747468305833 2024-09-26 00:06:34,204 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=861434.0, ans=0.125 2024-09-26 00:06:34,599 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=4.31 vs. limit=15.0 2024-09-26 00:06:40,689 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=861480.6666666666, ans=0.0 2024-09-26 00:06:49,399 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=4.14 vs. limit=15.0 2024-09-26 00:06:52,970 INFO [train.py:1198] (2/4) Epoch 48, batch 1500, loss[loss=0.2049, ctc_loss=0.1298, cr_loss=0.3755, over 17345.00 frames. ], tot_loss[loss=0.1845, ctc_loss=0.1177, cr_loss=0.3337, over 3375328.01 frames. ], batch size: 48, lr: 2.47e-03, grad_scale: 16.0 2024-09-26 00:07:01,300 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=861527.3333333334, ans=0.1 2024-09-26 00:07:18,543 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=861574.0, ans=0.125 2024-09-26 00:07:28,792 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.75 vs. limit=15.0 2024-09-26 00:07:33,112 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=861620.6666666666, ans=0.2 2024-09-26 00:08:03,207 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=10.47 vs. limit=22.5 2024-09-26 00:08:15,403 INFO [train.py:1198] (2/4) Epoch 48, batch 1550, loss[loss=0.173, ctc_loss=0.1104, cr_loss=0.3134, over 17356.00 frames. ], tot_loss[loss=0.1843, ctc_loss=0.1176, cr_loss=0.3336, over 3364460.54 frames. ], batch size: 48, lr: 2.47e-03, grad_scale: 16.0 2024-09-26 00:08:21,065 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=4.05 vs. limit=15.0 2024-09-26 00:08:30,665 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=5.35 vs. limit=15.0 2024-09-26 00:08:36,245 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=861807.3333333334, ans=0.125 2024-09-26 00:09:00,538 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.173e+02 1.301e+02 1.350e+02 1.480e+02 2.232e+02, threshold=2.701e+02, percent-clipped=0.0 2024-09-26 00:09:05,875 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=10.47 vs. limit=12.0 2024-09-26 00:09:14,205 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.96 vs. limit=15.0 2024-09-26 00:09:18,362 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=861900.6666666666, ans=0.07 2024-09-26 00:09:37,073 INFO [train.py:1198] (2/4) Epoch 48, batch 1600, loss[loss=0.1913, ctc_loss=0.123, cr_loss=0.3414, over 17019.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1184, cr_loss=0.3347, over 3358381.12 frames. ], batch size: 44, lr: 2.47e-03, grad_scale: 32.0 2024-09-26 00:09:38,864 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=861994.0, ans=0.125 2024-09-26 00:10:30,053 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.32 vs. limit=15.0 2024-09-26 00:10:39,916 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=3.82 vs. limit=10.0 2024-09-26 00:11:01,993 INFO [train.py:1198] (2/4) Epoch 48, batch 1650, loss[loss=0.1638, ctc_loss=0.1046, cr_loss=0.2958, over 17100.00 frames. ], tot_loss[loss=0.1855, ctc_loss=0.1186, cr_loss=0.3347, over 3352660.16 frames. ], batch size: 40, lr: 2.47e-03, grad_scale: 32.0 2024-09-26 00:11:08,978 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=862227.3333333334, ans=0.125 2024-09-26 00:11:09,248 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten.whitening_limit, batch_count=862227.3333333334, ans=15.0 2024-09-26 00:11:12,010 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=862227.3333333334, ans=0.2 2024-09-26 00:11:16,812 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=862274.0, ans=0.125 2024-09-26 00:11:34,681 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=862320.6666666666, ans=0.0 2024-09-26 00:11:47,276 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.142e+02 1.306e+02 1.368e+02 1.453e+02 1.769e+02, threshold=2.736e+02, percent-clipped=0.0 2024-09-26 00:12:22,676 INFO [train.py:1198] (2/4) Epoch 48, batch 1700, loss[loss=0.1948, ctc_loss=0.1262, cr_loss=0.3429, over 16455.00 frames. ], tot_loss[loss=0.1853, ctc_loss=0.1184, cr_loss=0.3345, over 3345855.12 frames. ], batch size: 66, lr: 2.47e-03, grad_scale: 16.0 2024-09-26 00:12:29,256 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=862460.6666666666, ans=0.015 2024-09-26 00:12:42,211 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=862507.3333333334, ans=0.1 2024-09-26 00:13:02,333 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=862554.0, ans=0.1 2024-09-26 00:13:45,463 INFO [train.py:1198] (2/4) Epoch 48, batch 1750, loss[loss=0.1824, ctc_loss=0.1179, cr_loss=0.3224, over 17273.00 frames. ], tot_loss[loss=0.1851, ctc_loss=0.1182, cr_loss=0.3343, over 3349730.66 frames. ], batch size: 44, lr: 2.47e-03, grad_scale: 16.0 2024-09-26 00:13:54,860 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.min_abs, batch_count=862694.0, ans=0.5 2024-09-26 00:14:31,540 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=10.55 vs. limit=12.0 2024-09-26 00:14:32,512 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.170e+02 1.295e+02 1.382e+02 1.510e+02 2.646e+02, threshold=2.763e+02, percent-clipped=0.0 2024-09-26 00:14:32,882 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=862787.3333333334, ans=0.125 2024-09-26 00:14:37,631 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=862834.0, ans=0.125 2024-09-26 00:14:38,336 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=192, metric=6.29 vs. limit=15.0 2024-09-26 00:15:05,643 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=862880.6666666666, ans=0.0 2024-09-26 00:15:07,325 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.22 vs. limit=15.0 2024-09-26 00:15:09,851 INFO [train.py:1198] (2/4) Epoch 48, batch 1800, loss[loss=0.1899, ctc_loss=0.122, cr_loss=0.3393, over 17214.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1184, cr_loss=0.3349, over 3352356.97 frames. ], batch size: 55, lr: 2.47e-03, grad_scale: 16.0 2024-09-26 00:15:16,895 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.04 vs. limit=15.0 2024-09-26 00:15:18,168 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=862927.3333333334, ans=0.2 2024-09-26 00:15:29,377 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=862974.0, ans=0.125 2024-09-26 00:15:32,530 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-26 00:16:07,018 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=863067.3333333334, ans=0.0 2024-09-26 00:16:10,189 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=863067.3333333334, ans=0.1 2024-09-26 00:16:31,980 INFO [train.py:1198] (2/4) Epoch 48, batch 1850, loss[loss=0.1462, ctc_loss=0.09185, cr_loss=0.2717, over 16215.00 frames. ], tot_loss[loss=0.1864, ctc_loss=0.1191, cr_loss=0.3361, over 3353363.35 frames. ], batch size: 36, lr: 2.47e-03, grad_scale: 16.0 2024-09-26 00:16:37,172 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=863160.6666666666, ans=0.125 2024-09-26 00:16:46,549 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=863207.3333333334, ans=0.025 2024-09-26 00:16:51,582 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=863207.3333333334, ans=0.0 2024-09-26 00:16:52,998 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=863207.3333333334, ans=0.1 2024-09-26 00:17:16,786 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.173e+02 1.322e+02 1.421e+02 1.543e+02 2.119e+02, threshold=2.843e+02, percent-clipped=0.0 2024-09-26 00:17:17,028 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=863254.0, ans=0.0 2024-09-26 00:17:35,975 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.86 vs. limit=12.0 2024-09-26 00:17:41,111 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=863347.3333333334, ans=0.125 2024-09-26 00:17:41,678 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=12.90 vs. limit=15.0 2024-09-26 00:17:54,932 INFO [train.py:1198] (2/4) Epoch 48, batch 1900, loss[loss=0.2084, ctc_loss=0.1376, cr_loss=0.354, over 16748.00 frames. ], tot_loss[loss=0.187, ctc_loss=0.1195, cr_loss=0.3374, over 3359069.44 frames. ], batch size: 61, lr: 2.47e-03, grad_scale: 16.0 2024-09-26 00:18:47,674 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-26 00:18:58,623 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=863534.0, ans=0.125 2024-09-26 00:19:00,196 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=863580.6666666666, ans=0.1 2024-09-26 00:19:17,621 INFO [train.py:1198] (2/4) Epoch 48, batch 1950, loss[loss=0.1917, ctc_loss=0.1224, cr_loss=0.3467, over 17303.00 frames. ], tot_loss[loss=0.1863, ctc_loss=0.119, cr_loss=0.3365, over 3359864.38 frames. ], batch size: 51, lr: 2.47e-03, grad_scale: 16.0 2024-09-26 00:19:35,280 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=9.95 vs. limit=22.5 2024-09-26 00:19:41,240 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=863674.0, ans=0.125 2024-09-26 00:19:42,992 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer1.prob, batch_count=863674.0, ans=0.125 2024-09-26 00:19:52,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=863720.6666666666, ans=0.2 2024-09-26 00:20:03,936 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=863720.6666666666, ans=0.0 2024-09-26 00:20:05,198 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.164e+02 1.335e+02 1.400e+02 1.522e+02 2.815e+02, threshold=2.800e+02, percent-clipped=0.0 2024-09-26 00:20:37,396 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=863814.0, ans=0.1 2024-09-26 00:20:40,168 INFO [train.py:1198] (2/4) Epoch 48, batch 2000, loss[loss=0.2068, ctc_loss=0.1344, cr_loss=0.362, over 15189.00 frames. ], tot_loss[loss=0.1869, ctc_loss=0.1194, cr_loss=0.3371, over 3348182.05 frames. ], batch size: 89, lr: 2.47e-03, grad_scale: 32.0 2024-09-26 00:20:46,281 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=863860.6666666666, ans=0.07 2024-09-26 00:22:02,510 INFO [train.py:1198] (2/4) Epoch 48, batch 2050, loss[loss=0.1672, ctc_loss=0.1061, cr_loss=0.3054, over 17263.00 frames. ], tot_loss[loss=0.1866, ctc_loss=0.1193, cr_loss=0.3365, over 3347046.97 frames. ], batch size: 42, lr: 2.47e-03, grad_scale: 32.0 2024-09-26 00:22:51,183 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.098e+02 1.300e+02 1.355e+02 1.459e+02 2.186e+02, threshold=2.709e+02, percent-clipped=0.0 2024-09-26 00:23:09,422 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=864280.6666666666, ans=0.125 2024-09-26 00:23:25,052 INFO [train.py:1198] (2/4) Epoch 48, batch 2100, loss[loss=0.2302, ctc_loss=0.1522, cr_loss=0.39, over 15004.00 frames. ], tot_loss[loss=0.1868, ctc_loss=0.1194, cr_loss=0.3368, over 3354839.35 frames. ], batch size: 89, lr: 2.47e-03, grad_scale: 16.0 2024-09-26 00:24:15,058 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=864467.3333333334, ans=0.125 2024-09-26 00:24:24,958 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=3.40 vs. limit=15.0 2024-09-26 00:24:50,876 INFO [train.py:1198] (2/4) Epoch 48, batch 2150, loss[loss=0.2273, ctc_loss=0.1524, cr_loss=0.3743, over 14950.00 frames. ], tot_loss[loss=0.1862, ctc_loss=0.119, cr_loss=0.3357, over 3361233.55 frames. ], batch size: 89, lr: 2.47e-03, grad_scale: 16.0 2024-09-26 00:25:00,848 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=864560.6666666666, ans=0.2 2024-09-26 00:25:37,464 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.112e+02 1.295e+02 1.362e+02 1.467e+02 1.863e+02, threshold=2.724e+02, percent-clipped=0.0 2024-09-26 00:25:53,124 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.out_combiner.scale_min, batch_count=864700.6666666666, ans=0.2 2024-09-26 00:25:54,895 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=864700.6666666666, ans=0.1 2024-09-26 00:26:13,769 INFO [train.py:1198] (2/4) Epoch 48, batch 2200, loss[loss=0.1875, ctc_loss=0.1169, cr_loss=0.3527, over 17294.00 frames. ], tot_loss[loss=0.1863, ctc_loss=0.1192, cr_loss=0.3355, over 3365865.78 frames. ], batch size: 46, lr: 2.47e-03, grad_scale: 16.0 2024-09-26 00:27:03,822 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=864934.0, ans=0.0 2024-09-26 00:27:14,929 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=864934.0, ans=0.125 2024-09-26 00:27:18,262 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=864980.6666666666, ans=0.025 2024-09-26 00:27:36,712 INFO [train.py:1198] (2/4) Epoch 48, batch 2250, loss[loss=0.2039, ctc_loss=0.1333, cr_loss=0.3532, over 17012.00 frames. ], tot_loss[loss=0.185, ctc_loss=0.1182, cr_loss=0.334, over 3369379.42 frames. ], batch size: 53, lr: 2.47e-03, grad_scale: 16.0 2024-09-26 00:27:41,766 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=865027.3333333334, ans=0.1 2024-09-26 00:27:59,319 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=865074.0, ans=0.125 2024-09-26 00:28:12,416 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=865120.6666666666, ans=0.0 2024-09-26 00:28:23,489 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.181e+02 1.300e+02 1.376e+02 1.452e+02 2.028e+02, threshold=2.752e+02, percent-clipped=0.0 2024-09-26 00:28:57,254 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=865214.0, ans=0.0 2024-09-26 00:28:57,290 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=865214.0, ans=0.125 2024-09-26 00:29:00,162 INFO [train.py:1198] (2/4) Epoch 48, batch 2300, loss[loss=0.1663, ctc_loss=0.1053, cr_loss=0.3049, over 17304.00 frames. ], tot_loss[loss=0.1849, ctc_loss=0.1181, cr_loss=0.3339, over 3372558.06 frames. ], batch size: 46, lr: 2.47e-03, grad_scale: 16.0 2024-09-26 00:29:23,211 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.whiten, num_groups=1, num_channels=512, metric=4.54 vs. limit=12.0 2024-09-26 00:29:40,340 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.38 vs. limit=22.5 2024-09-26 00:29:49,541 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=865400.6666666666, ans=0.0 2024-09-26 00:29:58,251 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=4.91 vs. limit=10.0 2024-09-26 00:30:18,654 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=865447.3333333334, ans=0.1 2024-09-26 00:30:20,152 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=865447.3333333334, ans=0.025 2024-09-26 00:30:22,865 INFO [train.py:1198] (2/4) Epoch 48, batch 2350, loss[loss=0.204, ctc_loss=0.1326, cr_loss=0.3569, over 17087.00 frames. ], tot_loss[loss=0.186, ctc_loss=0.1189, cr_loss=0.3355, over 3366887.08 frames. ], batch size: 46, lr: 2.47e-03, grad_scale: 16.0 2024-09-26 00:30:57,467 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=865587.3333333334, ans=0.125 2024-09-26 00:31:11,656 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.174e+02 1.301e+02 1.386e+02 1.481e+02 2.528e+02, threshold=2.772e+02, percent-clipped=0.0 2024-09-26 00:31:18,347 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=865634.0, ans=0.1 2024-09-26 00:31:30,943 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=865680.6666666666, ans=0.125 2024-09-26 00:31:35,667 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=865680.6666666666, ans=0.1 2024-09-26 00:31:44,986 INFO [train.py:1198] (2/4) Epoch 48, batch 2400, loss[loss=0.1962, ctc_loss=0.1246, cr_loss=0.3579, over 17014.00 frames. ], tot_loss[loss=0.1858, ctc_loss=0.1188, cr_loss=0.3349, over 3368769.51 frames. ], batch size: 56, lr: 2.46e-03, grad_scale: 32.0 2024-09-26 00:31:48,868 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=9.04 vs. limit=15.0 2024-09-26 00:31:53,588 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=865727.3333333334, ans=0.125 2024-09-26 00:32:06,932 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=6.95 vs. limit=15.0 2024-09-26 00:33:07,653 INFO [train.py:1198] (2/4) Epoch 48, batch 2450, loss[loss=0.1759, ctc_loss=0.1133, cr_loss=0.3129, over 17163.00 frames. ], tot_loss[loss=0.1857, ctc_loss=0.1188, cr_loss=0.3345, over 3360062.07 frames. ], batch size: 45, lr: 2.46e-03, grad_scale: 32.0 2024-09-26 00:33:14,385 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=865960.6666666666, ans=0.025 2024-09-26 00:33:14,392 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=865960.6666666666, ans=0.2 2024-09-26 00:33:54,506 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.19 vs. limit=10.0 2024-09-26 00:33:58,228 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.118e+02 1.328e+02 1.396e+02 1.507e+02 2.681e+02, threshold=2.792e+02, percent-clipped=0.0 2024-09-26 00:33:58,618 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-26 00:34:03,233 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=866100.6666666666, ans=0.2 2024-09-26 00:34:16,401 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=866147.3333333334, ans=0.0 2024-09-26 00:34:32,988 INFO [train.py:1198] (2/4) Epoch 48, batch 2500, loss[loss=0.212, ctc_loss=0.1392, cr_loss=0.364, over 17037.00 frames. ], tot_loss[loss=0.1859, ctc_loss=0.119, cr_loss=0.3346, over 3354072.26 frames. ], batch size: 56, lr: 2.46e-03, grad_scale: 16.0 2024-09-26 00:34:44,478 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=866194.0, ans=0.1 2024-09-26 00:34:55,778 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.balancer2.prob, batch_count=866240.6666666666, ans=0.125 2024-09-26 00:35:29,732 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=866334.0, ans=0.1 2024-09-26 00:35:30,565 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module1.whiten, num_groups=1, num_channels=192, metric=10.65 vs. limit=15.0 2024-09-26 00:35:31,368 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=866334.0, ans=0.2 2024-09-26 00:35:56,123 INFO [train.py:1198] (2/4) Epoch 48, batch 2550, loss[loss=0.2066, ctc_loss=0.1326, cr_loss=0.3698, over 16997.00 frames. ], tot_loss[loss=0.187, ctc_loss=0.1197, cr_loss=0.3366, over 3358591.87 frames. ], batch size: 53, lr: 2.46e-03, grad_scale: 16.0 2024-09-26 00:36:13,976 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=866474.0, ans=0.0 2024-09-26 00:36:43,904 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.101e+02 1.316e+02 1.412e+02 1.521e+02 2.388e+02, threshold=2.824e+02, percent-clipped=0.0 2024-09-26 00:36:52,552 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=866567.3333333334, ans=0.125 2024-09-26 00:36:54,299 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=866567.3333333334, ans=0.0 2024-09-26 00:36:55,770 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=866567.3333333334, ans=0.0 2024-09-26 00:37:15,194 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=866660.6666666666, ans=0.125 2024-09-26 00:37:16,412 INFO [train.py:1198] (2/4) Epoch 48, batch 2600, loss[loss=0.2126, ctc_loss=0.1435, cr_loss=0.3455, over 11564.00 frames. ], tot_loss[loss=0.1866, ctc_loss=0.1194, cr_loss=0.3362, over 3353005.20 frames. ], batch size: 123, lr: 2.46e-03, grad_scale: 16.0 2024-09-26 00:37:18,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=866660.6666666666, ans=0.0 2024-09-26 00:37:25,030 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.36 vs. limit=6.0 2024-09-26 00:37:43,784 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=2.41 vs. limit=6.0 2024-09-26 00:37:48,190 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-26 00:37:51,840 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=6.41 vs. limit=15.0 2024-09-26 00:38:00,963 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=866754.0, ans=0.1 2024-09-26 00:38:03,246 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.86 vs. limit=15.0 2024-09-26 00:38:21,563 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=866847.3333333334, ans=0.1 2024-09-26 00:38:23,564 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.24 vs. limit=15.0 2024-09-26 00:38:35,147 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.58 vs. limit=15.0 2024-09-26 00:38:41,784 INFO [train.py:1198] (2/4) Epoch 48, batch 2650, loss[loss=0.1608, ctc_loss=0.1015, cr_loss=0.2965, over 16950.00 frames. ], tot_loss[loss=0.1862, ctc_loss=0.1191, cr_loss=0.3353, over 3349512.28 frames. ], batch size: 42, lr: 2.46e-03, grad_scale: 16.0 2024-09-26 00:38:43,597 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=866894.0, ans=0.125 2024-09-26 00:38:45,280 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=866894.0, ans=0.025 2024-09-26 00:38:54,807 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=866894.0, ans=0.125 2024-09-26 00:39:06,029 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.scale_min, batch_count=866940.6666666666, ans=0.2 2024-09-26 00:39:21,268 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=7.21 vs. limit=15.0 2024-09-26 00:39:30,971 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=867034.0, ans=0.1 2024-09-26 00:39:32,299 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.147e+02 1.306e+02 1.381e+02 1.471e+02 2.187e+02, threshold=2.762e+02, percent-clipped=0.0 2024-09-26 00:40:04,502 INFO [train.py:1198] (2/4) Epoch 48, batch 2700, loss[loss=0.1683, ctc_loss=0.1073, cr_loss=0.3048, over 17047.00 frames. ], tot_loss[loss=0.1861, ctc_loss=0.1191, cr_loss=0.3351, over 3343729.16 frames. ], batch size: 39, lr: 2.46e-03, grad_scale: 16.0 2024-09-26 00:40:04,901 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=867127.3333333334, ans=0.125 2024-09-26 00:40:12,963 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=867127.3333333334, ans=0.0 2024-09-26 00:40:27,163 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=867174.0, ans=0.1 2024-09-26 00:40:50,777 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=867220.6666666666, ans=0.0 2024-09-26 00:41:03,411 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer_ff2.min_abs, batch_count=867267.3333333334, ans=0.1 2024-09-26 00:41:06,881 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-26 00:41:11,766 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=867314.0, ans=0.1 2024-09-26 00:41:15,918 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=7.11 vs. limit=8.0 2024-09-26 00:41:27,333 INFO [train.py:1198] (2/4) Epoch 48, batch 2750, loss[loss=0.2089, ctc_loss=0.1351, cr_loss=0.3694, over 16761.00 frames. ], tot_loss[loss=0.1859, ctc_loss=0.119, cr_loss=0.3348, over 3353072.81 frames. ], batch size: 61, lr: 2.46e-03, grad_scale: 16.0 2024-09-26 00:41:32,462 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=867360.6666666666, ans=0.5 2024-09-26 00:41:33,970 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=867360.6666666666, ans=0.2 2024-09-26 00:42:04,347 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer1.prob, batch_count=867454.0, ans=0.125 2024-09-26 00:42:10,724 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=867454.0, ans=0.025 2024-09-26 00:42:15,064 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.121e+02 1.302e+02 1.400e+02 1.492e+02 2.005e+02, threshold=2.800e+02, percent-clipped=0.0 2024-09-26 00:42:40,658 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=867547.3333333334, ans=0.125 2024-09-26 00:42:42,348 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=867547.3333333334, ans=0.025 2024-09-26 00:42:49,903 INFO [train.py:1198] (2/4) Epoch 48, batch 2800, loss[loss=0.1514, ctc_loss=0.09556, cr_loss=0.2792, over 17125.00 frames. ], tot_loss[loss=0.1856, ctc_loss=0.1187, cr_loss=0.3345, over 3356111.69 frames. ], batch size: 40, lr: 2.46e-03, grad_scale: 32.0 2024-09-26 00:42:56,729 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=867594.0, ans=0.0 2024-09-26 00:42:58,470 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=11.98 vs. limit=15.0 2024-09-26 00:43:04,627 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=867640.6666666666, ans=0.125 2024-09-26 00:43:14,068 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=867640.6666666666, ans=0.2 2024-09-26 00:43:22,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=867687.3333333334, ans=0.1 2024-09-26 00:43:58,077 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=867780.6666666666, ans=0.2 2024-09-26 00:43:58,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=867780.6666666666, ans=0.1 2024-09-26 00:44:01,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass_mid.scale_min, batch_count=867780.6666666666, ans=0.2 2024-09-26 00:44:02,949 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=867780.6666666666, ans=0.1 2024-09-26 00:44:12,427 INFO [train.py:1198] (2/4) Epoch 48, batch 2850, loss[loss=0.1364, ctc_loss=0.08351, cr_loss=0.2643, over 16710.00 frames. ], tot_loss[loss=0.1858, ctc_loss=0.1189, cr_loss=0.3348, over 3338719.00 frames. ], batch size: 37, lr: 2.46e-03, grad_scale: 32.0 2024-09-26 00:44:20,765 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=867827.3333333334, ans=0.0 2024-09-26 00:44:32,946 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-26 00:44:36,095 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer2.prob, batch_count=867874.0, ans=0.125 2024-09-26 00:44:42,641 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=867874.0, ans=0.07 2024-09-26 00:45:02,970 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.151e+02 1.286e+02 1.342e+02 1.424e+02 1.734e+02, threshold=2.685e+02, percent-clipped=0.0 2024-09-26 00:45:13,401 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=512, metric=5.42 vs. limit=15.0 2024-09-26 00:45:36,620 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=868060.6666666666, ans=0.2 2024-09-26 00:45:37,911 INFO [train.py:1198] (2/4) Epoch 48, batch 2900, loss[loss=0.2002, ctc_loss=0.1299, cr_loss=0.3515, over 17217.00 frames. ], tot_loss[loss=0.1864, ctc_loss=0.1193, cr_loss=0.3356, over 3336932.29 frames. ], batch size: 50, lr: 2.46e-03, grad_scale: 32.0 2024-09-26 00:46:05,164 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=868107.3333333334, ans=0.025 2024-09-26 00:46:19,758 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=868154.0, ans=0.0 2024-09-26 00:46:40,038 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=868247.3333333334, ans=0.125 2024-09-26 00:46:57,807 INFO [train.py:1198] (2/4) Epoch 48, batch 2950, loss[loss=0.2262, ctc_loss=0.1481, cr_loss=0.3903, over 17058.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1186, cr_loss=0.3344, over 3346425.22 frames. ], batch size: 52, lr: 2.46e-03, grad_scale: 32.0 2024-09-26 00:47:12,819 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=868340.6666666666, ans=0.125 2024-09-26 00:47:28,109 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.const_attention_rate, batch_count=868340.6666666666, ans=0.025 2024-09-26 00:47:28,203 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff2_skip_rate, batch_count=868340.6666666666, ans=0.0 2024-09-26 00:47:48,716 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=868434.0, ans=0.09899494936611666 2024-09-26 00:47:50,008 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.142e+02 1.284e+02 1.365e+02 1.462e+02 2.440e+02, threshold=2.731e+02, percent-clipped=0.0 2024-09-26 00:48:12,680 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=868480.6666666666, ans=0.025 2024-09-26 00:48:20,378 INFO [train.py:1198] (2/4) Epoch 48, batch 3000, loss[loss=0.182, ctc_loss=0.1154, cr_loss=0.333, over 17299.00 frames. ], tot_loss[loss=0.1851, ctc_loss=0.1182, cr_loss=0.3342, over 3352439.95 frames. ], batch size: 51, lr: 2.46e-03, grad_scale: 16.0 2024-09-26 00:48:20,378 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-26 00:48:38,776 INFO [train.py:1230] (2/4) Epoch 48, validation: loss=0.03527, ctc_loss=0.03527, cr_loss=1.067e-14, over 944034.00 frames. 2024-09-26 00:48:38,777 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-26 00:49:03,912 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=868574.0, ans=0.025 2024-09-26 00:49:26,129 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=11.97 vs. limit=15.0 2024-09-26 00:49:29,198 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=868667.3333333334, ans=0.125 2024-09-26 00:49:57,146 INFO [train.py:1198] (2/4) Epoch 48, batch 3050, loss[loss=0.2028, ctc_loss=0.1307, cr_loss=0.3607, over 16988.00 frames. ], tot_loss[loss=0.1863, ctc_loss=0.1191, cr_loss=0.336, over 3344475.65 frames. ], batch size: 53, lr: 2.46e-03, grad_scale: 16.0 2024-09-26 00:50:06,687 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=868760.6666666666, ans=0.0 2024-09-26 00:50:32,830 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=868854.0, ans=0.125 2024-09-26 00:50:36,137 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=868854.0, ans=0.0 2024-09-26 00:50:40,619 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=868854.0, ans=0.035 2024-09-26 00:50:46,112 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=6.46 vs. limit=10.0 2024-09-26 00:50:48,353 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.116e+02 1.292e+02 1.394e+02 1.474e+02 1.856e+02, threshold=2.789e+02, percent-clipped=0.0 2024-09-26 00:50:54,693 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=868900.6666666666, ans=0.025 2024-09-26 00:51:13,749 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=868947.3333333334, ans=0.1 2024-09-26 00:51:18,421 INFO [train.py:1198] (2/4) Epoch 48, batch 3100, loss[loss=0.1766, ctc_loss=0.1107, cr_loss=0.3296, over 17286.00 frames. ], tot_loss[loss=0.1861, ctc_loss=0.1189, cr_loss=0.336, over 3358671.53 frames. ], batch size: 51, lr: 2.46e-03, grad_scale: 16.0 2024-09-26 00:51:26,427 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=868994.0, ans=0.09899494936611666 2024-09-26 00:51:34,123 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=869040.6666666666, ans=0.0 2024-09-26 00:51:35,489 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=869040.6666666666, ans=10.0 2024-09-26 00:51:48,050 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=869087.3333333334, ans=0.2 2024-09-26 00:52:06,569 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=869134.0, ans=0.125 2024-09-26 00:52:06,687 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.min_positive, batch_count=869134.0, ans=0.05 2024-09-26 00:52:17,505 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=869134.0, ans=0.125 2024-09-26 00:52:35,870 INFO [train.py:1198] (2/4) Epoch 48, batch 3150, loss[loss=0.1998, ctc_loss=0.1286, cr_loss=0.3564, over 17029.00 frames. ], tot_loss[loss=0.1858, ctc_loss=0.1188, cr_loss=0.3353, over 3363396.67 frames. ], batch size: 56, lr: 2.46e-03, grad_scale: 16.0 2024-09-26 00:53:09,410 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=869320.6666666666, ans=0.0 2024-09-26 00:53:25,441 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=512, metric=12.07 vs. limit=22.5 2024-09-26 00:53:26,182 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.144e+02 1.286e+02 1.377e+02 1.489e+02 2.573e+02, threshold=2.753e+02, percent-clipped=0.0 2024-09-26 00:53:51,604 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.scale_min, batch_count=869414.0, ans=0.2 2024-09-26 00:53:53,134 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.min_positive, batch_count=869414.0, ans=0.025 2024-09-26 00:53:55,915 INFO [train.py:1198] (2/4) Epoch 48, batch 3200, loss[loss=0.2434, ctc_loss=0.1652, cr_loss=0.3913, over 11667.00 frames. ], tot_loss[loss=0.1862, ctc_loss=0.119, cr_loss=0.3358, over 3359420.03 frames. ], batch size: 123, lr: 2.46e-03, grad_scale: 32.0 2024-09-26 00:54:26,448 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=4.76 vs. limit=15.0 2024-09-26 00:54:30,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=869554.0, ans=0.2 2024-09-26 00:54:31,137 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=256, metric=7.89 vs. limit=15.0 2024-09-26 00:54:56,214 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=869600.6666666666, ans=0.0 2024-09-26 00:55:05,433 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=869647.3333333334, ans=0.0 2024-09-26 00:55:10,186 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=869647.3333333334, ans=0.125 2024-09-26 00:55:14,463 INFO [train.py:1198] (2/4) Epoch 48, batch 3250, loss[loss=0.1805, ctc_loss=0.1167, cr_loss=0.3191, over 16935.00 frames. ], tot_loss[loss=0.1869, ctc_loss=0.1196, cr_loss=0.3366, over 3365195.95 frames. ], batch size: 58, lr: 2.46e-03, grad_scale: 32.0 2024-09-26 00:55:15,165 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.76 vs. limit=22.5 2024-09-26 00:55:25,687 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.nonlin_attention.balancer.max_positive, batch_count=869694.0, ans=0.95 2024-09-26 00:55:44,618 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-26 00:56:03,196 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.191e+02 1.301e+02 1.439e+02 1.548e+02 2.033e+02, threshold=2.879e+02, percent-clipped=0.0 2024-09-26 00:56:33,152 INFO [train.py:1198] (2/4) Epoch 48, batch 3300, loss[loss=0.2022, ctc_loss=0.1294, cr_loss=0.3635, over 16869.00 frames. ], tot_loss[loss=0.1871, ctc_loss=0.1197, cr_loss=0.3371, over 3362385.81 frames. ], batch size: 58, lr: 2.46e-03, grad_scale: 32.0 2024-09-26 00:57:54,014 INFO [train.py:1198] (2/4) Epoch 48, batch 3350, loss[loss=0.2286, ctc_loss=0.1468, cr_loss=0.409, over 17051.00 frames. ], tot_loss[loss=0.1873, ctc_loss=0.1198, cr_loss=0.3375, over 3358609.06 frames. ], batch size: 52, lr: 2.46e-03, grad_scale: 32.0 2024-09-26 00:57:59,011 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=870160.6666666666, ans=0.0 2024-09-26 00:58:10,380 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=12.19 vs. limit=22.5 2024-09-26 00:58:42,523 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.149e+02 1.313e+02 1.389e+02 1.485e+02 2.642e+02, threshold=2.779e+02, percent-clipped=0.0 2024-09-26 00:58:56,794 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=870347.3333333334, ans=0.125 2024-09-26 00:58:58,563 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.24 vs. limit=15.0 2024-09-26 00:59:12,293 INFO [train.py:1198] (2/4) Epoch 48, batch 3400, loss[loss=0.1504, ctc_loss=0.09326, cr_loss=0.2858, over 17258.00 frames. ], tot_loss[loss=0.1871, ctc_loss=0.1198, cr_loss=0.3368, over 3358775.09 frames. ], batch size: 42, lr: 2.46e-03, grad_scale: 32.0 2024-09-26 00:59:23,975 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=512, metric=15.39 vs. limit=22.5 2024-09-26 00:59:37,543 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.hidden_balancer.prob, batch_count=870440.6666666666, ans=0.125 2024-09-26 00:59:42,203 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=870487.3333333334, ans=0.125 2024-09-26 00:59:47,669 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=870487.3333333334, ans=0.0 2024-09-26 00:59:58,384 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=870487.3333333334, ans=0.1 2024-09-26 01:00:02,175 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=3.90 vs. limit=6.0 2024-09-26 01:00:03,127 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.bypass.scale_min, batch_count=870534.0, ans=0.2 2024-09-26 01:00:32,483 INFO [train.py:1198] (2/4) Epoch 48, batch 3450, loss[loss=0.1701, ctc_loss=0.1066, cr_loss=0.3176, over 17007.00 frames. ], tot_loss[loss=0.1862, ctc_loss=0.119, cr_loss=0.3358, over 3356881.23 frames. ], batch size: 44, lr: 2.46e-03, grad_scale: 32.0 2024-09-26 01:01:06,138 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=870720.6666666666, ans=0.0 2024-09-26 01:01:13,837 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=870720.6666666666, ans=0.125 2024-09-26 01:01:24,344 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.161e+02 1.301e+02 1.395e+02 1.467e+02 2.192e+02, threshold=2.789e+02, percent-clipped=0.0 2024-09-26 01:01:32,849 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer_ff3.min_abs, batch_count=870767.3333333334, ans=0.2 2024-09-26 01:01:52,534 INFO [train.py:1198] (2/4) Epoch 48, batch 3500, loss[loss=0.1983, ctc_loss=0.1279, cr_loss=0.3521, over 17330.00 frames. ], tot_loss[loss=0.1858, ctc_loss=0.1187, cr_loss=0.3356, over 3358089.58 frames. ], batch size: 51, lr: 2.46e-03, grad_scale: 16.0 2024-09-26 01:02:14,934 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=870907.3333333334, ans=0.125 2024-09-26 01:02:19,781 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=870907.3333333334, ans=0.1 2024-09-26 01:02:42,184 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=871000.6666666666, ans=0.1 2024-09-26 01:02:53,920 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=192, metric=8.59 vs. limit=15.0 2024-09-26 01:02:54,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=871000.6666666666, ans=0.2 2024-09-26 01:03:06,885 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=871047.3333333334, ans=0.0 2024-09-26 01:03:06,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=871047.3333333334, ans=0.125 2024-09-26 01:03:08,900 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.44 vs. limit=12.0 2024-09-26 01:03:12,761 INFO [train.py:1198] (2/4) Epoch 48, batch 3550, loss[loss=0.1932, ctc_loss=0.1232, cr_loss=0.3502, over 17072.00 frames. ], tot_loss[loss=0.1851, ctc_loss=0.1182, cr_loss=0.3347, over 3360340.25 frames. ], batch size: 46, lr: 2.46e-03, grad_scale: 16.0 2024-09-26 01:03:17,866 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=871094.0, ans=0.0 2024-09-26 01:03:17,881 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=871094.0, ans=0.1 2024-09-26 01:03:25,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer2.prob, batch_count=871094.0, ans=0.125 2024-09-26 01:03:41,271 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=871140.6666666666, ans=0.125 2024-09-26 01:03:51,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=871187.3333333334, ans=0.0 2024-09-26 01:04:02,352 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.323e+02 1.392e+02 1.491e+02 3.535e+02, threshold=2.785e+02, percent-clipped=1.0 2024-09-26 01:04:30,422 INFO [train.py:1198] (2/4) Epoch 48, batch 3600, loss[loss=0.17, ctc_loss=0.1063, cr_loss=0.3184, over 17305.00 frames. ], tot_loss[loss=0.1851, ctc_loss=0.1181, cr_loss=0.3347, over 3358803.27 frames. ], batch size: 49, lr: 2.46e-03, grad_scale: 32.0 2024-09-26 01:04:48,023 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.min_positive, batch_count=871374.0, ans=0.05 2024-09-26 01:05:06,807 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=871420.6666666666, ans=0.2 2024-09-26 01:05:12,850 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=871420.6666666666, ans=0.0 2024-09-26 01:05:48,461 INFO [train.py:1198] (2/4) Epoch 48, batch 3650, loss[loss=0.2065, ctc_loss=0.1336, cr_loss=0.3641, over 15937.00 frames. ], tot_loss[loss=0.1853, ctc_loss=0.1183, cr_loss=0.335, over 3358000.36 frames. ], batch size: 74, lr: 2.46e-03, grad_scale: 32.0 2024-09-26 01:06:06,492 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-26 01:06:32,720 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=871654.0, ans=0.125 2024-09-26 01:06:34,339 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=871654.0, ans=0.0 2024-09-26 01:06:39,033 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=871700.6666666666, ans=0.125 2024-09-26 01:06:40,145 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.328e+02 1.397e+02 1.530e+02 1.851e+02, threshold=2.795e+02, percent-clipped=0.0 2024-09-26 01:06:41,964 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.attention_skip_rate, batch_count=871700.6666666666, ans=0.0 2024-09-26 01:06:43,851 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=11.11 vs. limit=12.0 2024-09-26 01:06:53,985 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=871747.3333333334, ans=0.125 2024-09-26 01:07:04,933 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.min_positive, batch_count=871747.3333333334, ans=0.05 2024-09-26 01:07:07,937 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=871794.0, ans=0.025 2024-09-26 01:07:09,262 INFO [train.py:1198] (2/4) Epoch 48, batch 3700, loss[loss=0.1787, ctc_loss=0.1102, cr_loss=0.3424, over 17181.00 frames. ], tot_loss[loss=0.1858, ctc_loss=0.1188, cr_loss=0.3352, over 3354797.92 frames. ], batch size: 45, lr: 2.46e-03, grad_scale: 32.0 2024-09-26 01:07:15,771 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=871794.0, ans=0.0 2024-09-26 01:07:38,077 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=6.11 vs. limit=12.0 2024-09-26 01:07:48,828 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=871887.3333333334, ans=0.04949747468305833 2024-09-26 01:08:15,539 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.out_whiten.whitening_limit, batch_count=871980.6666666666, ans=8.0 2024-09-26 01:08:28,606 INFO [train.py:1198] (2/4) Epoch 48, batch 3750, loss[loss=0.1841, ctc_loss=0.1168, cr_loss=0.3367, over 17019.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1184, cr_loss=0.3348, over 3349780.83 frames. ], batch size: 51, lr: 2.46e-03, grad_scale: 16.0 2024-09-26 01:08:32,476 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.72 vs. limit=15.0 2024-09-26 01:08:46,620 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.42 vs. limit=6.0 2024-09-26 01:09:19,539 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=872167.3333333334, ans=0.2 2024-09-26 01:09:20,720 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.132e+02 1.319e+02 1.400e+02 1.543e+02 5.735e+02, threshold=2.801e+02, percent-clipped=1.0 2024-09-26 01:09:29,477 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=872167.3333333334, ans=0.125 2024-09-26 01:09:47,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=872260.6666666666, ans=0.0 2024-09-26 01:09:48,870 INFO [train.py:1198] (2/4) Epoch 48, batch 3800, loss[loss=0.2037, ctc_loss=0.1344, cr_loss=0.3465, over 11928.00 frames. ], tot_loss[loss=0.1862, ctc_loss=0.119, cr_loss=0.3356, over 3328486.36 frames. ], batch size: 124, lr: 2.46e-03, grad_scale: 16.0 2024-09-26 01:09:49,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=872260.6666666666, ans=0.0 2024-09-26 01:10:12,819 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=872307.3333333334, ans=0.125 2024-09-26 01:10:22,603 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.76 vs. limit=15.0 2024-09-26 01:10:57,801 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.const_attention_rate, batch_count=872447.3333333334, ans=0.025 2024-09-26 01:11:06,798 INFO [train.py:1198] (2/4) Epoch 48, batch 3850, loss[loss=0.2341, ctc_loss=0.1596, cr_loss=0.3725, over 12052.00 frames. ], tot_loss[loss=0.1895, ctc_loss=0.1216, cr_loss=0.3395, over 3275437.57 frames. ], batch size: 123, lr: 2.46e-03, grad_scale: 8.0 2024-09-26 01:11:08,616 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=872494.0, ans=0.2 2024-09-26 01:11:08,689 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=872494.0, ans=0.025 2024-09-26 01:11:17,311 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn1.whiten, num_groups=1, num_channels=192, metric=11.76 vs. limit=22.5 2024-09-26 01:11:47,363 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.attention_skip_rate, batch_count=872587.3333333334, ans=0.0 2024-09-26 01:11:48,788 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=872587.3333333334, ans=0.1 2024-09-26 01:11:59,142 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.219e+02 1.389e+02 1.529e+02 1.705e+02 2.274e+02, threshold=3.058e+02, percent-clipped=0.0 2024-09-26 01:13:04,118 INFO [train.py:1198] (2/4) Epoch 49, batch 0, loss[loss=0.2134, ctc_loss=0.1377, cr_loss=0.3788, over 16683.00 frames. ], tot_loss[loss=0.2134, ctc_loss=0.1377, cr_loss=0.3788, over 16683.00 frames. ], batch size: 61, lr: 2.43e-03, grad_scale: 16.0 2024-09-26 01:13:04,118 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-26 01:13:19,509 INFO [train.py:1230] (2/4) Epoch 49, validation: loss=0.03487, ctc_loss=0.03487, cr_loss=1.087e-14, over 944034.00 frames. 2024-09-26 01:13:19,510 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-26 01:13:33,492 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.prob, batch_count=872708.6666666666, ans=0.125 2024-09-26 01:13:36,585 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass_mid.scale_min, batch_count=872755.3333333334, ans=0.2 2024-09-26 01:13:52,406 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=872802.0, ans=0.1 2024-09-26 01:14:18,164 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.77 vs. limit=10.0 2024-09-26 01:14:21,328 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.98 vs. limit=6.0 2024-09-26 01:14:22,615 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=872848.6666666666, ans=0.09899494936611666 2024-09-26 01:14:44,179 INFO [train.py:1198] (2/4) Epoch 49, batch 50, loss[loss=0.1869, ctc_loss=0.1188, cr_loss=0.3405, over 17094.00 frames. ], tot_loss[loss=0.1862, ctc_loss=0.1187, cr_loss=0.3372, over 761168.54 frames. ], batch size: 49, lr: 2.43e-03, grad_scale: 16.0 2024-09-26 01:15:39,911 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=873082.0, ans=0.125 2024-09-26 01:15:46,355 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=873082.0, ans=0.1 2024-09-26 01:15:47,637 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.099e+02 1.321e+02 1.394e+02 1.552e+02 2.223e+02, threshold=2.788e+02, percent-clipped=0.0 2024-09-26 01:15:51,255 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=873128.6666666666, ans=0.2 2024-09-26 01:15:59,243 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=873128.6666666666, ans=0.125 2024-09-26 01:16:02,520 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-26 01:16:07,026 INFO [train.py:1198] (2/4) Epoch 49, batch 100, loss[loss=0.1857, ctc_loss=0.1166, cr_loss=0.3457, over 17103.00 frames. ], tot_loss[loss=0.1875, ctc_loss=0.1197, cr_loss=0.3389, over 1334584.90 frames. ], batch size: 49, lr: 2.43e-03, grad_scale: 16.0 2024-09-26 01:16:47,208 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=873268.6666666666, ans=0.125 2024-09-26 01:17:11,162 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=873362.0, ans=0.0 2024-09-26 01:17:29,482 INFO [train.py:1198] (2/4) Epoch 49, batch 150, loss[loss=0.1899, ctc_loss=0.1208, cr_loss=0.3459, over 17300.00 frames. ], tot_loss[loss=0.1845, ctc_loss=0.1177, cr_loss=0.334, over 1781914.81 frames. ], batch size: 46, lr: 2.43e-03, grad_scale: 16.0 2024-09-26 01:17:53,626 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=873455.3333333334, ans=0.2 2024-09-26 01:18:29,798 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=873548.6666666666, ans=0.1 2024-09-26 01:18:32,853 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.108e+02 1.298e+02 1.390e+02 1.504e+02 2.457e+02, threshold=2.779e+02, percent-clipped=0.0 2024-09-26 01:18:42,128 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.17 vs. limit=15.0 2024-09-26 01:18:52,325 INFO [train.py:1198] (2/4) Epoch 49, batch 200, loss[loss=0.1522, ctc_loss=0.09556, cr_loss=0.2833, over 17126.00 frames. ], tot_loss[loss=0.1828, ctc_loss=0.1165, cr_loss=0.3317, over 2133125.02 frames. ], batch size: 40, lr: 2.43e-03, grad_scale: 16.0 2024-09-26 01:18:57,474 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=873642.0, ans=0.2 2024-09-26 01:19:15,005 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=873688.6666666666, ans=0.025 2024-09-26 01:19:23,065 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=873735.3333333334, ans=0.1 2024-09-26 01:19:31,073 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=873735.3333333334, ans=0.1 2024-09-26 01:19:47,543 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.63 vs. limit=5.0 2024-09-26 01:20:17,863 INFO [train.py:1198] (2/4) Epoch 49, batch 250, loss[loss=0.1861, ctc_loss=0.1176, cr_loss=0.3427, over 16994.00 frames. ], tot_loss[loss=0.1852, ctc_loss=0.1182, cr_loss=0.335, over 2404783.81 frames. ], batch size: 53, lr: 2.43e-03, grad_scale: 16.0 2024-09-26 01:20:35,418 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.72 vs. limit=6.0 2024-09-26 01:21:10,617 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=874015.3333333334, ans=0.025 2024-09-26 01:21:18,234 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.091e+02 1.257e+02 1.339e+02 1.417e+02 1.603e+02, threshold=2.678e+02, percent-clipped=0.0 2024-09-26 01:21:25,074 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=874062.0, ans=0.2 2024-09-26 01:21:30,746 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=17.71 vs. limit=22.5 2024-09-26 01:21:31,707 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=874062.0, ans=0.125 2024-09-26 01:21:33,239 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=874062.0, ans=0.05 2024-09-26 01:21:37,797 INFO [train.py:1198] (2/4) Epoch 49, batch 300, loss[loss=0.198, ctc_loss=0.1294, cr_loss=0.3429, over 17003.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1183, cr_loss=0.3355, over 2622162.90 frames. ], batch size: 51, lr: 2.43e-03, grad_scale: 16.0 2024-09-26 01:22:02,680 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=11.47 vs. limit=15.0 2024-09-26 01:22:11,728 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.balancer2.prob, batch_count=874202.0, ans=0.125 2024-09-26 01:22:22,650 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=874202.0, ans=0.125 2024-09-26 01:22:29,712 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=9.19 vs. limit=15.0 2024-09-26 01:23:00,851 INFO [train.py:1198] (2/4) Epoch 49, batch 350, loss[loss=0.1728, ctc_loss=0.1087, cr_loss=0.3204, over 17094.00 frames. ], tot_loss[loss=0.1863, ctc_loss=0.1188, cr_loss=0.337, over 2778620.50 frames. ], batch size: 49, lr: 2.43e-03, grad_scale: 16.0 2024-09-26 01:23:43,538 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=874435.3333333334, ans=0.125 2024-09-26 01:23:56,410 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=874482.0, ans=0.0 2024-09-26 01:24:04,153 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.165e+02 1.315e+02 1.403e+02 1.517e+02 7.992e+02, threshold=2.807e+02, percent-clipped=1.0 2024-09-26 01:24:15,825 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.max_abs, batch_count=874528.6666666666, ans=10.0 2024-09-26 01:24:23,274 INFO [train.py:1198] (2/4) Epoch 49, batch 400, loss[loss=0.1616, ctc_loss=0.1005, cr_loss=0.3058, over 17197.00 frames. ], tot_loss[loss=0.1874, ctc_loss=0.1196, cr_loss=0.339, over 2912151.63 frames. ], batch size: 41, lr: 2.43e-03, grad_scale: 32.0 2024-09-26 01:24:23,902 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.93 vs. limit=15.0 2024-09-26 01:24:26,752 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=874575.3333333334, ans=0.025 2024-09-26 01:24:35,202 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.16 vs. limit=15.0 2024-09-26 01:24:51,810 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-26 01:25:03,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=874668.6666666666, ans=0.125 2024-09-26 01:25:23,319 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=874715.3333333334, ans=0.125 2024-09-26 01:25:31,298 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=874762.0, ans=0.05 2024-09-26 01:25:38,993 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=874762.0, ans=0.0 2024-09-26 01:25:48,233 INFO [train.py:1198] (2/4) Epoch 49, batch 450, loss[loss=0.2088, ctc_loss=0.1367, cr_loss=0.3607, over 15997.00 frames. ], tot_loss[loss=0.1876, ctc_loss=0.1198, cr_loss=0.3393, over 3017177.21 frames. ], batch size: 74, lr: 2.43e-03, grad_scale: 32.0 2024-09-26 01:25:48,629 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.max_abs, batch_count=874808.6666666666, ans=10.0 2024-09-26 01:26:01,316 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=874808.6666666666, ans=0.2 2024-09-26 01:26:10,865 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=874855.3333333334, ans=0.125 2024-09-26 01:26:15,781 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=874855.3333333334, ans=0.1 2024-09-26 01:26:30,236 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=874902.0, ans=0.125 2024-09-26 01:26:49,065 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.141e+02 1.319e+02 1.417e+02 1.516e+02 2.635e+02, threshold=2.833e+02, percent-clipped=0.0 2024-09-26 01:27:08,163 INFO [train.py:1198] (2/4) Epoch 49, batch 500, loss[loss=0.1896, ctc_loss=0.1213, cr_loss=0.3412, over 17138.00 frames. ], tot_loss[loss=0.1874, ctc_loss=0.1197, cr_loss=0.3385, over 3093266.59 frames. ], batch size: 48, lr: 2.43e-03, grad_scale: 32.0 2024-09-26 01:27:20,796 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-26 01:27:23,798 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=875042.0, ans=0.125 2024-09-26 01:27:33,479 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=875088.6666666666, ans=0.2 2024-09-26 01:27:39,918 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=875088.6666666666, ans=0.125 2024-09-26 01:27:42,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=875135.3333333334, ans=0.1 2024-09-26 01:27:42,948 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=875135.3333333334, ans=0.2 2024-09-26 01:27:46,207 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=875135.3333333334, ans=0.125 2024-09-26 01:27:50,859 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=875135.3333333334, ans=0.125 2024-09-26 01:27:55,722 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=875135.3333333334, ans=0.0 2024-09-26 01:28:29,776 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=875228.6666666666, ans=0.125 2024-09-26 01:28:32,675 INFO [train.py:1198] (2/4) Epoch 49, batch 550, loss[loss=0.1757, ctc_loss=0.1117, cr_loss=0.3203, over 17061.00 frames. ], tot_loss[loss=0.186, ctc_loss=0.1188, cr_loss=0.3361, over 3146167.22 frames. ], batch size: 46, lr: 2.43e-03, grad_scale: 32.0 2024-09-26 01:28:40,931 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=875275.3333333334, ans=0.125 2024-09-26 01:28:56,907 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=875322.0, ans=0.0 2024-09-26 01:28:58,409 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=875322.0, ans=0.1 2024-09-26 01:29:16,167 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.prob, batch_count=875368.6666666666, ans=0.125 2024-09-26 01:29:33,607 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.295e+02 1.361e+02 1.476e+02 2.484e+02, threshold=2.722e+02, percent-clipped=0.0 2024-09-26 01:29:51,593 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.95 vs. limit=15.0 2024-09-26 01:29:58,449 INFO [train.py:1198] (2/4) Epoch 49, batch 600, loss[loss=0.1772, ctc_loss=0.1112, cr_loss=0.3301, over 17203.00 frames. ], tot_loss[loss=0.1853, ctc_loss=0.1182, cr_loss=0.3355, over 3198429.16 frames. ], batch size: 47, lr: 2.43e-03, grad_scale: 32.0 2024-09-26 01:30:10,052 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-26 01:30:54,375 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.scale_min, batch_count=875648.6666666666, ans=0.2 2024-09-26 01:31:18,110 INFO [train.py:1198] (2/4) Epoch 49, batch 650, loss[loss=0.1621, ctc_loss=0.09941, cr_loss=0.3135, over 17119.00 frames. ], tot_loss[loss=0.186, ctc_loss=0.1188, cr_loss=0.3361, over 3213927.49 frames. ], batch size: 40, lr: 2.43e-03, grad_scale: 32.0 2024-09-26 01:31:37,763 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.balancer1.prob, batch_count=875788.6666666666, ans=0.125 2024-09-26 01:31:46,037 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-26 01:31:54,237 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=512, metric=4.74 vs. limit=15.0 2024-09-26 01:32:01,759 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=875835.3333333334, ans=0.025 2024-09-26 01:32:17,396 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.01 vs. limit=15.0 2024-09-26 01:32:21,519 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.113e+02 1.285e+02 1.382e+02 1.503e+02 2.355e+02, threshold=2.765e+02, percent-clipped=0.0 2024-09-26 01:32:26,689 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=875928.6666666666, ans=0.0 2024-09-26 01:32:26,963 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.27 vs. limit=10.0 2024-09-26 01:32:40,448 INFO [train.py:1198] (2/4) Epoch 49, batch 700, loss[loss=0.183, ctc_loss=0.1189, cr_loss=0.3202, over 17269.00 frames. ], tot_loss[loss=0.1853, ctc_loss=0.1183, cr_loss=0.3347, over 3250355.36 frames. ], batch size: 44, lr: 2.42e-03, grad_scale: 32.0 2024-09-26 01:32:40,781 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=875975.3333333334, ans=0.125 2024-09-26 01:33:03,782 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.whiten, num_groups=1, num_channels=192, metric=4.86 vs. limit=12.0 2024-09-26 01:33:08,272 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=11.36 vs. limit=22.5 2024-09-26 01:33:29,799 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=876115.3333333334, ans=0.0 2024-09-26 01:33:31,780 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.59 vs. limit=15.0 2024-09-26 01:33:49,355 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=876162.0, ans=0.1 2024-09-26 01:34:03,488 INFO [train.py:1198] (2/4) Epoch 49, batch 750, loss[loss=0.1783, ctc_loss=0.1131, cr_loss=0.3262, over 17089.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1184, cr_loss=0.3349, over 3268670.63 frames. ], batch size: 43, lr: 2.42e-03, grad_scale: 32.0 2024-09-26 01:34:29,323 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=876255.3333333334, ans=0.0 2024-09-26 01:34:44,913 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=876302.0, ans=0.125 2024-09-26 01:35:09,764 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.132e+02 1.321e+02 1.400e+02 1.492e+02 1.805e+02, threshold=2.801e+02, percent-clipped=0.0 2024-09-26 01:35:28,995 INFO [train.py:1198] (2/4) Epoch 49, batch 800, loss[loss=0.1689, ctc_loss=0.1028, cr_loss=0.3306, over 17018.00 frames. ], tot_loss[loss=0.1846, ctc_loss=0.1178, cr_loss=0.3338, over 3286930.79 frames. ], batch size: 39, lr: 2.42e-03, grad_scale: 32.0 2024-09-26 01:35:35,635 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_skip_rate, batch_count=876442.0, ans=0.0 2024-09-26 01:35:42,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=876442.0, ans=0.1 2024-09-26 01:35:54,763 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=876488.6666666666, ans=0.125 2024-09-26 01:35:56,324 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=876488.6666666666, ans=0.125 2024-09-26 01:35:59,748 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=876535.3333333334, ans=0.0 2024-09-26 01:36:02,966 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=876535.3333333334, ans=0.2 2024-09-26 01:36:46,207 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=876628.6666666666, ans=0.0 2024-09-26 01:36:49,021 INFO [train.py:1198] (2/4) Epoch 49, batch 850, loss[loss=0.2096, ctc_loss=0.1357, cr_loss=0.3695, over 17043.00 frames. ], tot_loss[loss=0.185, ctc_loss=0.1181, cr_loss=0.3342, over 3309756.24 frames. ], batch size: 56, lr: 2.42e-03, grad_scale: 32.0 2024-09-26 01:36:54,096 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=876675.3333333334, ans=0.1 2024-09-26 01:37:41,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=876815.3333333334, ans=0.125 2024-09-26 01:37:51,966 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.174e+02 1.324e+02 1.412e+02 1.511e+02 2.737e+02, threshold=2.823e+02, percent-clipped=0.0 2024-09-26 01:37:57,212 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=876862.0, ans=0.0 2024-09-26 01:37:57,709 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.whiten, num_groups=1, num_channels=512, metric=6.10 vs. limit=12.0 2024-09-26 01:38:06,777 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=876862.0, ans=0.125 2024-09-26 01:38:11,408 INFO [train.py:1198] (2/4) Epoch 49, batch 900, loss[loss=0.1686, ctc_loss=0.1082, cr_loss=0.3019, over 16935.00 frames. ], tot_loss[loss=0.1841, ctc_loss=0.1176, cr_loss=0.3327, over 3318035.84 frames. ], batch size: 42, lr: 2.42e-03, grad_scale: 32.0 2024-09-26 01:38:15,897 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=6.34 vs. limit=15.0 2024-09-26 01:38:16,605 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.ff3_skip_rate, batch_count=876908.6666666666, ans=0.0 2024-09-26 01:38:36,911 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.whiten, num_groups=1, num_channels=512, metric=4.43 vs. limit=12.0 2024-09-26 01:38:46,092 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.attention_skip_rate, batch_count=877002.0, ans=0.0 2024-09-26 01:38:50,368 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.convnext.out_whiten, num_groups=1, num_channels=128, metric=4.52 vs. limit=5.0 2024-09-26 01:38:51,450 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=4.52 vs. limit=15.0 2024-09-26 01:39:13,489 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=877048.6666666666, ans=0.2 2024-09-26 01:39:18,239 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=877095.3333333334, ans=0.07 2024-09-26 01:39:33,924 INFO [train.py:1198] (2/4) Epoch 49, batch 950, loss[loss=0.2058, ctc_loss=0.1352, cr_loss=0.3532, over 16880.00 frames. ], tot_loss[loss=0.1844, ctc_loss=0.1178, cr_loss=0.3331, over 3329713.71 frames. ], batch size: 58, lr: 2.42e-03, grad_scale: 16.0 2024-09-26 01:40:08,231 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=877188.6666666666, ans=0.0 2024-09-26 01:40:41,474 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.100e+02 1.300e+02 1.395e+02 1.483e+02 3.362e+02, threshold=2.790e+02, percent-clipped=1.0 2024-09-26 01:40:53,841 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=877328.6666666666, ans=0.125 2024-09-26 01:41:01,590 INFO [train.py:1198] (2/4) Epoch 49, batch 1000, loss[loss=0.1643, ctc_loss=0.1035, cr_loss=0.3038, over 17099.00 frames. ], tot_loss[loss=0.1848, ctc_loss=0.1181, cr_loss=0.3338, over 3333501.44 frames. ], batch size: 49, lr: 2.42e-03, grad_scale: 16.0 2024-09-26 01:41:03,433 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=877375.3333333334, ans=0.1 2024-09-26 01:41:08,233 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=877375.3333333334, ans=0.125 2024-09-26 01:41:24,009 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff3_skip_rate, batch_count=877422.0, ans=0.0 2024-09-26 01:42:12,148 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-26 01:42:23,941 INFO [train.py:1198] (2/4) Epoch 49, batch 1050, loss[loss=0.1742, ctc_loss=0.1104, cr_loss=0.3188, over 17211.00 frames. ], tot_loss[loss=0.1846, ctc_loss=0.1178, cr_loss=0.3338, over 3344414.34 frames. ], batch size: 50, lr: 2.42e-03, grad_scale: 16.0 2024-09-26 01:42:24,250 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=877608.6666666666, ans=0.125 2024-09-26 01:42:30,748 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=877608.6666666666, ans=0.2 2024-09-26 01:42:35,526 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer1.prob, batch_count=877608.6666666666, ans=0.125 2024-09-26 01:42:57,649 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=877702.0, ans=0.0 2024-09-26 01:43:04,622 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=8.41 vs. limit=15.0 2024-09-26 01:43:16,998 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.55 vs. limit=22.5 2024-09-26 01:43:19,918 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=877748.6666666666, ans=0.07 2024-09-26 01:43:27,730 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=7.77 vs. limit=15.0 2024-09-26 01:43:28,469 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.167e+02 1.276e+02 1.380e+02 1.490e+02 2.320e+02, threshold=2.759e+02, percent-clipped=0.0 2024-09-26 01:43:46,129 INFO [train.py:1198] (2/4) Epoch 49, batch 1100, loss[loss=0.1918, ctc_loss=0.1227, cr_loss=0.3452, over 17027.00 frames. ], tot_loss[loss=0.1852, ctc_loss=0.1182, cr_loss=0.335, over 3349782.19 frames. ], batch size: 56, lr: 2.42e-03, grad_scale: 16.0 2024-09-26 01:43:46,436 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.const_attention_rate, batch_count=877842.0, ans=0.025 2024-09-26 01:43:55,863 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=877842.0, ans=0.0 2024-09-26 01:44:07,005 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer2.prob, batch_count=877888.6666666666, ans=0.125 2024-09-26 01:44:07,026 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=877888.6666666666, ans=0.125 2024-09-26 01:44:32,475 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.max_positive, batch_count=877982.0, ans=0.95 2024-09-26 01:44:44,897 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=877982.0, ans=0.0 2024-09-26 01:45:01,467 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.max_positive, batch_count=878028.6666666666, ans=0.95 2024-09-26 01:45:10,651 INFO [train.py:1198] (2/4) Epoch 49, batch 1150, loss[loss=0.1599, ctc_loss=0.09978, cr_loss=0.3007, over 16950.00 frames. ], tot_loss[loss=0.1844, ctc_loss=0.1177, cr_loss=0.3337, over 3353017.31 frames. ], batch size: 42, lr: 2.42e-03, grad_scale: 16.0 2024-09-26 01:45:18,948 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-26 01:45:25,305 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff2_skip_rate, batch_count=878122.0, ans=0.0 2024-09-26 01:45:46,594 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=878168.6666666666, ans=0.0 2024-09-26 01:45:49,750 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=878168.6666666666, ans=0.125 2024-09-26 01:45:57,627 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=878215.3333333334, ans=0.125 2024-09-26 01:46:04,024 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=878215.3333333334, ans=0.07 2024-09-26 01:46:11,827 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=878215.3333333334, ans=0.0 2024-09-26 01:46:13,008 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.291e+02 1.352e+02 1.450e+02 2.079e+02, threshold=2.704e+02, percent-clipped=0.0 2024-09-26 01:46:18,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass.scale_min, batch_count=878262.0, ans=0.2 2024-09-26 01:46:24,876 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=878262.0, ans=0.1 2024-09-26 01:46:30,812 INFO [train.py:1198] (2/4) Epoch 49, batch 1200, loss[loss=0.1924, ctc_loss=0.1226, cr_loss=0.3492, over 16785.00 frames. ], tot_loss[loss=0.1846, ctc_loss=0.1178, cr_loss=0.3342, over 3360621.39 frames. ], batch size: 61, lr: 2.42e-03, grad_scale: 32.0 2024-09-26 01:46:32,624 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=878308.6666666666, ans=0.1 2024-09-26 01:46:51,767 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.90 vs. limit=6.0 2024-09-26 01:46:54,617 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.attention_skip_rate, batch_count=878355.3333333334, ans=0.0 2024-09-26 01:47:09,064 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-26 01:47:28,870 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=878448.6666666666, ans=0.1 2024-09-26 01:47:33,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward3.hidden_balancer.prob, batch_count=878448.6666666666, ans=0.125 2024-09-26 01:47:52,403 INFO [train.py:1198] (2/4) Epoch 49, batch 1250, loss[loss=0.1645, ctc_loss=0.1046, cr_loss=0.2993, over 16734.00 frames. ], tot_loss[loss=0.184, ctc_loss=0.1173, cr_loss=0.3335, over 3361493.66 frames. ], batch size: 37, lr: 2.42e-03, grad_scale: 32.0 2024-09-26 01:47:59,696 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=384, metric=19.61 vs. limit=22.5 2024-09-26 01:48:13,355 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=878588.6666666666, ans=0.0 2024-09-26 01:48:17,230 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=256, metric=11.58 vs. limit=22.5 2024-09-26 01:48:22,187 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=878588.6666666666, ans=10.0 2024-09-26 01:48:31,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=878635.3333333334, ans=0.0 2024-09-26 01:48:40,919 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=878682.0, ans=0.125 2024-09-26 01:48:53,676 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer2.prob, batch_count=878682.0, ans=0.125 2024-09-26 01:48:56,555 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.071e+02 1.287e+02 1.357e+02 1.444e+02 1.832e+02, threshold=2.714e+02, percent-clipped=0.0 2024-09-26 01:49:14,199 INFO [train.py:1198] (2/4) Epoch 49, batch 1300, loss[loss=0.1527, ctc_loss=0.09584, cr_loss=0.2842, over 16948.00 frames. ], tot_loss[loss=0.1835, ctc_loss=0.1171, cr_loss=0.3324, over 3365914.65 frames. ], batch size: 42, lr: 2.42e-03, grad_scale: 32.0 2024-09-26 01:49:27,181 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.scale_min, batch_count=878775.3333333334, ans=0.2 2024-09-26 01:49:47,432 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-26 01:49:56,235 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=878868.6666666666, ans=0.0 2024-09-26 01:49:57,907 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer1.prob, batch_count=878868.6666666666, ans=0.125 2024-09-26 01:50:34,996 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=12.50 vs. limit=15.0 2024-09-26 01:50:39,184 INFO [train.py:1198] (2/4) Epoch 49, batch 1350, loss[loss=0.222, ctc_loss=0.1427, cr_loss=0.3964, over 17211.00 frames. ], tot_loss[loss=0.1851, ctc_loss=0.1183, cr_loss=0.3344, over 3356111.14 frames. ], batch size: 55, lr: 2.42e-03, grad_scale: 32.0 2024-09-26 01:50:49,095 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=879008.6666666666, ans=0.1 2024-09-26 01:50:58,567 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=879055.3333333334, ans=0.1 2024-09-26 01:51:08,170 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.out_combiner.scale_min, batch_count=879055.3333333334, ans=0.2 2024-09-26 01:51:22,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=879102.0, ans=0.1 2024-09-26 01:51:22,738 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=879102.0, ans=0.0 2024-09-26 01:51:27,468 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-26 01:51:33,934 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer1.prob, batch_count=879148.6666666666, ans=0.125 2024-09-26 01:51:41,923 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.182e+02 1.323e+02 1.398e+02 1.529e+02 2.878e+02, threshold=2.797e+02, percent-clipped=1.0 2024-09-26 01:51:59,737 INFO [train.py:1198] (2/4) Epoch 49, batch 1400, loss[loss=0.1505, ctc_loss=0.0943, cr_loss=0.2808, over 16704.00 frames. ], tot_loss[loss=0.1858, ctc_loss=0.1187, cr_loss=0.3355, over 3352020.30 frames. ], batch size: 37, lr: 2.42e-03, grad_scale: 16.0 2024-09-26 01:52:02,299 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=6.89 vs. limit=15.0 2024-09-26 01:52:15,174 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer2.prob, batch_count=879242.0, ans=0.125 2024-09-26 01:52:21,831 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-26 01:52:31,073 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=879288.6666666666, ans=0.125 2024-09-26 01:52:31,226 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=12.13 vs. limit=12.0 2024-09-26 01:52:46,813 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.const_attention_rate, batch_count=879335.3333333334, ans=0.025 2024-09-26 01:52:47,443 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=5.88 vs. limit=15.0 2024-09-26 01:53:08,940 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=879428.6666666666, ans=0.0 2024-09-26 01:53:12,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer1.max_abs, batch_count=879428.6666666666, ans=10.0 2024-09-26 01:53:19,391 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=879428.6666666666, ans=0.1 2024-09-26 01:53:24,001 INFO [train.py:1198] (2/4) Epoch 49, batch 1450, loss[loss=0.1443, ctc_loss=0.08902, cr_loss=0.2763, over 16960.00 frames. ], tot_loss[loss=0.1858, ctc_loss=0.1187, cr_loss=0.3356, over 3355112.97 frames. ], batch size: 42, lr: 2.42e-03, grad_scale: 16.0 2024-09-26 01:53:34,057 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-26 01:54:20,606 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=879615.3333333334, ans=0.0 2024-09-26 01:54:28,634 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.139e+02 1.317e+02 1.389e+02 1.499e+02 2.448e+02, threshold=2.778e+02, percent-clipped=0.0 2024-09-26 01:54:29,029 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=879662.0, ans=0.2 2024-09-26 01:54:30,665 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=879662.0, ans=0.0 2024-09-26 01:54:37,321 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.72 vs. limit=10.0 2024-09-26 01:54:47,225 INFO [train.py:1198] (2/4) Epoch 49, batch 1500, loss[loss=0.1606, ctc_loss=0.1003, cr_loss=0.3014, over 17194.00 frames. ], tot_loss[loss=0.186, ctc_loss=0.1187, cr_loss=0.3361, over 3360729.58 frames. ], batch size: 41, lr: 2.42e-03, grad_scale: 16.0 2024-09-26 01:54:50,727 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=879708.6666666666, ans=0.125 2024-09-26 01:55:03,518 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=879755.3333333334, ans=0.125 2024-09-26 01:55:40,500 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff2_skip_rate, batch_count=879848.6666666666, ans=0.0 2024-09-26 01:56:07,498 INFO [train.py:1198] (2/4) Epoch 49, batch 1550, loss[loss=0.1808, ctc_loss=0.116, cr_loss=0.324, over 17019.00 frames. ], tot_loss[loss=0.1864, ctc_loss=0.1191, cr_loss=0.3365, over 3359810.45 frames. ], batch size: 44, lr: 2.42e-03, grad_scale: 16.0 2024-09-26 01:56:12,618 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=879942.0, ans=0.125 2024-09-26 01:56:30,790 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=384, metric=6.33 vs. limit=10.0 2024-09-26 01:56:33,504 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=879988.6666666666, ans=0.1 2024-09-26 01:57:13,894 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.158e+02 1.314e+02 1.386e+02 1.460e+02 1.935e+02, threshold=2.771e+02, percent-clipped=0.0 2024-09-26 01:57:28,770 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=880175.3333333334, ans=0.0 2024-09-26 01:57:30,178 INFO [train.py:1198] (2/4) Epoch 49, batch 1600, loss[loss=0.1698, ctc_loss=0.1044, cr_loss=0.3272, over 17110.00 frames. ], tot_loss[loss=0.1863, ctc_loss=0.119, cr_loss=0.3364, over 3358756.99 frames. ], batch size: 40, lr: 2.42e-03, grad_scale: 32.0 2024-09-26 01:57:32,110 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=880175.3333333334, ans=0.125 2024-09-26 01:57:36,960 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=880175.3333333334, ans=0.0 2024-09-26 01:57:38,476 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=880175.3333333334, ans=0.1 2024-09-26 01:57:38,492 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=880175.3333333334, ans=0.125 2024-09-26 01:58:20,996 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=880315.3333333334, ans=0.025 2024-09-26 01:58:49,810 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=880362.0, ans=0.0 2024-09-26 01:58:51,299 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=880408.6666666666, ans=0.1 2024-09-26 01:58:52,672 INFO [train.py:1198] (2/4) Epoch 49, batch 1650, loss[loss=0.1815, ctc_loss=0.1165, cr_loss=0.3251, over 17243.00 frames. ], tot_loss[loss=0.185, ctc_loss=0.118, cr_loss=0.3349, over 3361463.56 frames. ], batch size: 50, lr: 2.42e-03, grad_scale: 32.0 2024-09-26 01:58:59,588 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-26 01:59:13,709 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass.skip_rate, batch_count=880455.3333333334, ans=0.04949747468305833 2024-09-26 01:59:13,739 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=880455.3333333334, ans=0.125 2024-09-26 01:59:25,327 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.06 vs. limit=15.0 2024-09-26 01:59:53,286 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=4.76 vs. limit=15.0 2024-09-26 02:00:01,887 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.163e+02 1.295e+02 1.382e+02 1.560e+02 2.405e+02, threshold=2.764e+02, percent-clipped=0.0 2024-09-26 02:00:17,669 INFO [train.py:1198] (2/4) Epoch 49, batch 1700, loss[loss=0.224, ctc_loss=0.1447, cr_loss=0.3964, over 17030.00 frames. ], tot_loss[loss=0.1858, ctc_loss=0.1186, cr_loss=0.336, over 3365527.91 frames. ], batch size: 56, lr: 2.42e-03, grad_scale: 32.0 2024-09-26 02:00:25,904 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=880642.0, ans=0.1 2024-09-26 02:00:35,439 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=880688.6666666666, ans=0.125 2024-09-26 02:00:36,924 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.5.prob, batch_count=880688.6666666666, ans=0.125 2024-09-26 02:01:10,577 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=880782.0, ans=0.1 2024-09-26 02:01:12,561 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.99 vs. limit=15.0 2024-09-26 02:01:17,528 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.98 vs. limit=6.0 2024-09-26 02:01:33,413 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=14.30 vs. limit=22.5 2024-09-26 02:01:37,517 INFO [train.py:1198] (2/4) Epoch 49, batch 1750, loss[loss=0.2064, ctc_loss=0.1392, cr_loss=0.3359, over 15267.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1182, cr_loss=0.3357, over 3361521.63 frames. ], batch size: 89, lr: 2.42e-03, grad_scale: 16.0 2024-09-26 02:01:48,940 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff3_skip_rate, batch_count=880875.3333333334, ans=0.0 2024-09-26 02:02:13,748 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=880968.6666666666, ans=0.125 2024-09-26 02:02:45,571 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.131e+02 1.296e+02 1.405e+02 1.536e+02 1.900e+02, threshold=2.809e+02, percent-clipped=0.0 2024-09-26 02:02:59,569 INFO [train.py:1198] (2/4) Epoch 49, batch 1800, loss[loss=0.1637, ctc_loss=0.1045, cr_loss=0.2957, over 17248.00 frames. ], tot_loss[loss=0.1852, ctc_loss=0.1182, cr_loss=0.3352, over 3365658.79 frames. ], batch size: 44, lr: 2.42e-03, grad_scale: 16.0 2024-09-26 02:03:23,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward2.hidden_balancer.prob, batch_count=881155.3333333334, ans=0.125 2024-09-26 02:03:28,157 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=881155.3333333334, ans=0.125 2024-09-26 02:03:50,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.nonlin_attention.balancer.prob, batch_count=881248.6666666666, ans=0.125 2024-09-26 02:03:55,064 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=881248.6666666666, ans=0.1 2024-09-26 02:03:56,755 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=881248.6666666666, ans=0.125 2024-09-26 02:04:15,838 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=881295.3333333334, ans=0.125 2024-09-26 02:04:21,861 INFO [train.py:1198] (2/4) Epoch 49, batch 1850, loss[loss=0.2084, ctc_loss=0.1407, cr_loss=0.3383, over 11654.00 frames. ], tot_loss[loss=0.1858, ctc_loss=0.1186, cr_loss=0.3358, over 3354143.31 frames. ], batch size: 123, lr: 2.42e-03, grad_scale: 16.0 2024-09-26 02:04:35,632 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=8.88 vs. limit=15.0 2024-09-26 02:04:46,653 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.const_attention_rate, batch_count=881388.6666666666, ans=0.025 2024-09-26 02:04:53,079 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=881388.6666666666, ans=0.1 2024-09-26 02:05:07,486 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=881435.3333333334, ans=0.1 2024-09-26 02:05:08,162 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=6.02 vs. limit=15.0 2024-09-26 02:05:12,206 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.layerdrop_rate, batch_count=881435.3333333334, ans=0.015 2024-09-26 02:05:13,861 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.scale_min, batch_count=881482.0, ans=0.2 2024-09-26 02:05:18,856 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass.skip_rate, batch_count=881482.0, ans=0.09899494936611666 2024-09-26 02:05:31,719 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=881528.6666666666, ans=0.035 2024-09-26 02:05:33,165 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.195e+02 1.348e+02 1.416e+02 1.489e+02 2.449e+02, threshold=2.831e+02, percent-clipped=0.0 2024-09-26 02:05:38,379 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=881528.6666666666, ans=0.07 2024-09-26 02:05:47,592 INFO [train.py:1198] (2/4) Epoch 49, batch 1900, loss[loss=0.2009, ctc_loss=0.1262, cr_loss=0.3735, over 15850.00 frames. ], tot_loss[loss=0.1858, ctc_loss=0.1185, cr_loss=0.3361, over 3358917.49 frames. ], batch size: 74, lr: 2.42e-03, grad_scale: 16.0 2024-09-26 02:05:47,957 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.scale_min, batch_count=881575.3333333334, ans=0.2 2024-09-26 02:05:59,082 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=881575.3333333334, ans=0.125 2024-09-26 02:05:59,799 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.63 vs. limit=22.5 2024-09-26 02:06:34,511 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=881715.3333333334, ans=0.09899494936611666 2024-09-26 02:06:51,060 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=16.01 vs. limit=22.5 2024-09-26 02:06:56,970 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=881762.0, ans=0.0 2024-09-26 02:07:10,385 INFO [train.py:1198] (2/4) Epoch 49, batch 1950, loss[loss=0.164, ctc_loss=0.1022, cr_loss=0.3093, over 17123.00 frames. ], tot_loss[loss=0.1853, ctc_loss=0.1182, cr_loss=0.3352, over 3348577.70 frames. ], batch size: 40, lr: 2.42e-03, grad_scale: 16.0 2024-09-26 02:07:37,811 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.ff2_skip_rate, batch_count=881855.3333333334, ans=0.0 2024-09-26 02:07:40,272 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=9.53 vs. limit=15.0 2024-09-26 02:07:41,063 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=881902.0, ans=0.125 2024-09-26 02:08:14,420 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=881995.3333333334, ans=0.0 2024-09-26 02:08:18,104 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.175e+02 1.309e+02 1.411e+02 1.494e+02 2.442e+02, threshold=2.822e+02, percent-clipped=0.0 2024-09-26 02:08:32,518 INFO [train.py:1198] (2/4) Epoch 49, batch 2000, loss[loss=0.2087, ctc_loss=0.1348, cr_loss=0.3694, over 17217.00 frames. ], tot_loss[loss=0.1852, ctc_loss=0.1181, cr_loss=0.3351, over 3338710.60 frames. ], batch size: 55, lr: 2.42e-03, grad_scale: 32.0 2024-09-26 02:08:57,291 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module1.whiten, num_groups=1, num_channels=192, metric=8.83 vs. limit=15.0 2024-09-26 02:09:03,185 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=882135.3333333334, ans=0.5 2024-09-26 02:09:57,472 INFO [train.py:1198] (2/4) Epoch 49, batch 2050, loss[loss=0.1913, ctc_loss=0.1264, cr_loss=0.3247, over 15123.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1184, cr_loss=0.3351, over 3345583.82 frames. ], batch size: 88, lr: 2.42e-03, grad_scale: 32.0 2024-09-26 02:10:01,426 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.06 vs. limit=12.0 2024-09-26 02:10:40,778 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=882368.6666666666, ans=0.125 2024-09-26 02:11:03,374 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=882462.0, ans=0.1 2024-09-26 02:11:04,655 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.134e+02 1.334e+02 1.409e+02 1.527e+02 2.352e+02, threshold=2.819e+02, percent-clipped=0.0 2024-09-26 02:11:09,963 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=882462.0, ans=0.1 2024-09-26 02:11:17,568 INFO [train.py:1198] (2/4) Epoch 49, batch 2100, loss[loss=0.163, ctc_loss=0.1029, cr_loss=0.3003, over 16926.00 frames. ], tot_loss[loss=0.185, ctc_loss=0.1181, cr_loss=0.3342, over 3345736.57 frames. ], batch size: 42, lr: 2.42e-03, grad_scale: 16.0 2024-09-26 02:11:49,918 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=882602.0, ans=0.1 2024-09-26 02:11:59,836 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=5.19 vs. limit=15.0 2024-09-26 02:12:08,570 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=882648.6666666666, ans=0.07 2024-09-26 02:12:20,330 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=256, metric=5.31 vs. limit=15.0 2024-09-26 02:12:39,852 INFO [train.py:1198] (2/4) Epoch 49, batch 2150, loss[loss=0.1922, ctc_loss=0.1251, cr_loss=0.3353, over 16491.00 frames. ], tot_loss[loss=0.1856, ctc_loss=0.1186, cr_loss=0.3353, over 3354340.45 frames. ], batch size: 66, lr: 2.42e-03, grad_scale: 16.0 2024-09-26 02:12:48,320 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=882742.0, ans=0.1 2024-09-26 02:12:52,839 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=882742.0, ans=0.125 2024-09-26 02:13:02,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer1.prob, batch_count=882788.6666666666, ans=0.125 2024-09-26 02:13:21,496 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.22 vs. limit=15.0 2024-09-26 02:13:49,689 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.094e+02 1.355e+02 1.416e+02 1.495e+02 3.199e+02, threshold=2.832e+02, percent-clipped=1.0 2024-09-26 02:14:02,463 INFO [train.py:1198] (2/4) Epoch 49, batch 2200, loss[loss=0.1791, ctc_loss=0.1122, cr_loss=0.3348, over 17023.00 frames. ], tot_loss[loss=0.185, ctc_loss=0.1181, cr_loss=0.3346, over 3362796.83 frames. ], batch size: 44, lr: 2.42e-03, grad_scale: 16.0 2024-09-26 02:14:10,975 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=5.09 vs. limit=15.0 2024-09-26 02:14:26,690 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=883022.0, ans=0.125 2024-09-26 02:14:28,505 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=883022.0, ans=0.2 2024-09-26 02:14:37,355 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=883068.6666666666, ans=0.2 2024-09-26 02:14:49,438 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-26 02:15:02,251 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.ff2_skip_rate, batch_count=883115.3333333334, ans=0.0 2024-09-26 02:15:03,962 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.ff3_skip_rate, batch_count=883115.3333333334, ans=0.0 2024-09-26 02:15:18,358 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=883162.0, ans=0.0 2024-09-26 02:15:27,837 INFO [train.py:1198] (2/4) Epoch 49, batch 2250, loss[loss=0.2091, ctc_loss=0.1357, cr_loss=0.3671, over 17214.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1183, cr_loss=0.3353, over 3367298.25 frames. ], batch size: 55, lr: 2.41e-03, grad_scale: 16.0 2024-09-26 02:15:29,755 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=883208.6666666666, ans=0.0 2024-09-26 02:15:32,114 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=4.68 vs. limit=15.0 2024-09-26 02:15:44,124 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=883255.3333333334, ans=0.0 2024-09-26 02:15:44,185 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module1.balancer2.prob, batch_count=883255.3333333334, ans=0.125 2024-09-26 02:16:34,923 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.163e+02 1.298e+02 1.376e+02 1.462e+02 1.948e+02, threshold=2.752e+02, percent-clipped=0.0 2024-09-26 02:16:47,762 INFO [train.py:1198] (2/4) Epoch 49, batch 2300, loss[loss=0.1518, ctc_loss=0.0953, cr_loss=0.2823, over 17188.00 frames. ], tot_loss[loss=0.1845, ctc_loss=0.1177, cr_loss=0.334, over 3366792.92 frames. ], batch size: 41, lr: 2.41e-03, grad_scale: 16.0 2024-09-26 02:16:51,332 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-26 02:16:51,914 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=4.06 vs. limit=15.0 2024-09-26 02:16:54,419 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=883442.0, ans=0.2 2024-09-26 02:17:02,160 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=883442.0, ans=0.125 2024-09-26 02:17:13,444 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=883488.6666666666, ans=0.0 2024-09-26 02:17:21,649 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer2.prob, batch_count=883535.3333333334, ans=0.125 2024-09-26 02:17:56,346 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=883628.6666666666, ans=0.0 2024-09-26 02:17:58,065 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.min_positive, batch_count=883628.6666666666, ans=0.05 2024-09-26 02:18:13,002 INFO [train.py:1198] (2/4) Epoch 49, batch 2350, loss[loss=0.1784, ctc_loss=0.1139, cr_loss=0.3227, over 16784.00 frames. ], tot_loss[loss=0.1842, ctc_loss=0.1175, cr_loss=0.3336, over 3368456.68 frames. ], batch size: 61, lr: 2.41e-03, grad_scale: 16.0 2024-09-26 02:18:21,420 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=883675.3333333334, ans=0.125 2024-09-26 02:18:37,734 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.min_positive, batch_count=883722.0, ans=0.05 2024-09-26 02:18:38,182 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn1.whiten, num_groups=1, num_channels=512, metric=12.63 vs. limit=22.5 2024-09-26 02:18:44,190 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer1.prob, batch_count=883768.6666666666, ans=0.125 2024-09-26 02:18:50,427 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=883768.6666666666, ans=0.0 2024-09-26 02:19:04,630 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=883815.3333333334, ans=0.125 2024-09-26 02:19:11,016 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=883815.3333333334, ans=0.125 2024-09-26 02:19:20,302 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.282e+02 1.377e+02 1.471e+02 1.818e+02, threshold=2.754e+02, percent-clipped=0.0 2024-09-26 02:19:35,627 INFO [train.py:1198] (2/4) Epoch 49, batch 2400, loss[loss=0.198, ctc_loss=0.1288, cr_loss=0.3462, over 16639.00 frames. ], tot_loss[loss=0.1836, ctc_loss=0.1171, cr_loss=0.3329, over 3363313.48 frames. ], batch size: 61, lr: 2.41e-03, grad_scale: 32.0 2024-09-26 02:19:54,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer1.prob, batch_count=883955.3333333334, ans=0.125 2024-09-26 02:19:57,765 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_skip_rate, batch_count=883955.3333333334, ans=0.0 2024-09-26 02:20:28,092 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=884048.6666666666, ans=0.0 2024-09-26 02:20:58,123 INFO [train.py:1198] (2/4) Epoch 49, batch 2450, loss[loss=0.2062, ctc_loss=0.1289, cr_loss=0.3864, over 17005.00 frames. ], tot_loss[loss=0.1846, ctc_loss=0.1178, cr_loss=0.3344, over 3357058.53 frames. ], batch size: 53, lr: 2.41e-03, grad_scale: 16.0 2024-09-26 02:21:08,081 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=884142.0, ans=0.125 2024-09-26 02:21:46,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=884282.0, ans=0.1 2024-09-26 02:22:09,462 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.145e+02 1.309e+02 1.374e+02 1.463e+02 2.759e+02, threshold=2.748e+02, percent-clipped=1.0 2024-09-26 02:22:20,946 INFO [train.py:1198] (2/4) Epoch 49, batch 2500, loss[loss=0.1787, ctc_loss=0.115, cr_loss=0.3185, over 17015.00 frames. ], tot_loss[loss=0.1843, ctc_loss=0.1175, cr_loss=0.3337, over 3360009.13 frames. ], batch size: 44, lr: 2.41e-03, grad_scale: 16.0 2024-09-26 02:22:24,604 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_skip_rate, batch_count=884375.3333333334, ans=0.0 2024-09-26 02:22:34,429 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=884375.3333333334, ans=0.1 2024-09-26 02:23:01,765 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=884468.6666666666, ans=0.0 2024-09-26 02:23:17,221 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer2.prob, batch_count=884515.3333333334, ans=0.125 2024-09-26 02:23:23,651 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=884515.3333333334, ans=0.125 2024-09-26 02:23:44,123 INFO [train.py:1198] (2/4) Epoch 49, batch 2550, loss[loss=0.1796, ctc_loss=0.1148, cr_loss=0.3238, over 15980.00 frames. ], tot_loss[loss=0.1855, ctc_loss=0.1184, cr_loss=0.3356, over 3357712.05 frames. ], batch size: 74, lr: 2.41e-03, grad_scale: 16.0 2024-09-26 02:23:50,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=884608.6666666666, ans=0.0 2024-09-26 02:24:16,424 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer2.prob, batch_count=884702.0, ans=0.125 2024-09-26 02:24:26,791 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=3.81 vs. limit=15.0 2024-09-26 02:24:33,487 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.2.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-26 02:24:33,927 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=512, metric=3.77 vs. limit=15.0 2024-09-26 02:24:58,343 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.136e+02 1.313e+02 1.367e+02 1.472e+02 2.257e+02, threshold=2.734e+02, percent-clipped=0.0 2024-09-26 02:25:01,958 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.min_positive, batch_count=884795.3333333334, ans=0.025 2024-09-26 02:25:09,543 INFO [train.py:1198] (2/4) Epoch 49, batch 2600, loss[loss=0.2065, ctc_loss=0.1339, cr_loss=0.363, over 17170.00 frames. ], tot_loss[loss=0.1853, ctc_loss=0.1183, cr_loss=0.3351, over 3359598.73 frames. ], batch size: 55, lr: 2.41e-03, grad_scale: 16.0 2024-09-26 02:25:19,416 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward2.hidden_balancer.prob, batch_count=884842.0, ans=0.125 2024-09-26 02:25:32,269 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=884888.6666666666, ans=0.125 2024-09-26 02:25:54,724 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.scale_min, batch_count=884935.3333333334, ans=0.2 2024-09-26 02:26:29,481 INFO [train.py:1198] (2/4) Epoch 49, batch 2650, loss[loss=0.1616, ctc_loss=0.1021, cr_loss=0.2973, over 17127.00 frames. ], tot_loss[loss=0.1849, ctc_loss=0.118, cr_loss=0.3341, over 3355484.53 frames. ], batch size: 40, lr: 2.41e-03, grad_scale: 16.0 2024-09-26 02:26:31,398 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=885075.3333333334, ans=0.125 2024-09-26 02:26:42,672 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=885075.3333333334, ans=0.125 2024-09-26 02:27:25,278 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=885215.3333333334, ans=0.125 2024-09-26 02:27:34,748 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=885262.0, ans=0.125 2024-09-26 02:27:40,642 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.183e+02 1.322e+02 1.427e+02 1.515e+02 2.814e+02, threshold=2.853e+02, percent-clipped=1.0 2024-09-26 02:27:49,131 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=885262.0, ans=0.05 2024-09-26 02:27:51,958 INFO [train.py:1198] (2/4) Epoch 49, batch 2700, loss[loss=0.1677, ctc_loss=0.1064, cr_loss=0.3062, over 17282.00 frames. ], tot_loss[loss=0.1852, ctc_loss=0.1184, cr_loss=0.3342, over 3347657.42 frames. ], batch size: 49, lr: 2.41e-03, grad_scale: 16.0 2024-09-26 02:28:21,771 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=885355.3333333334, ans=0.125 2024-09-26 02:28:24,949 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=885402.0, ans=0.0 2024-09-26 02:28:46,775 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=192, metric=7.21 vs. limit=15.0 2024-09-26 02:28:50,700 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=885448.6666666666, ans=0.1 2024-09-26 02:29:14,280 INFO [train.py:1198] (2/4) Epoch 49, batch 2750, loss[loss=0.1725, ctc_loss=0.1094, cr_loss=0.3155, over 17244.00 frames. ], tot_loss[loss=0.1861, ctc_loss=0.119, cr_loss=0.3358, over 3350259.09 frames. ], batch size: 44, lr: 2.41e-03, grad_scale: 16.0 2024-09-26 02:29:15,409 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.92 vs. limit=6.0 2024-09-26 02:29:24,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=885542.0, ans=0.125 2024-09-26 02:29:43,711 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=885588.6666666666, ans=0.0 2024-09-26 02:29:45,143 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.attention_skip_rate, batch_count=885588.6666666666, ans=0.0 2024-09-26 02:29:56,457 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.3.self_attn_weights, loss-sum=0.000e+00 2024-09-26 02:30:02,648 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.max_abs, batch_count=885635.3333333334, ans=10.0 2024-09-26 02:30:11,593 INFO [scaling.py:1024] (2/4) Whitening: name=encoder_embed.out_whiten, num_groups=1, num_channels=192, metric=7.27 vs. limit=8.0 2024-09-26 02:30:28,233 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.217e+02 1.341e+02 1.441e+02 1.583e+02 2.930e+02, threshold=2.882e+02, percent-clipped=1.0 2024-09-26 02:30:39,197 INFO [train.py:1198] (2/4) Epoch 49, batch 2800, loss[loss=0.19, ctc_loss=0.1242, cr_loss=0.3291, over 17237.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1186, cr_loss=0.3342, over 3338814.66 frames. ], batch size: 55, lr: 2.41e-03, grad_scale: 32.0 2024-09-26 02:30:43,277 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.55 vs. limit=15.0 2024-09-26 02:31:16,624 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=256, metric=7.21 vs. limit=15.0 2024-09-26 02:31:20,270 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=5.43 vs. limit=12.0 2024-09-26 02:32:01,786 INFO [train.py:1198] (2/4) Epoch 49, batch 2850, loss[loss=0.205, ctc_loss=0.1301, cr_loss=0.3747, over 17152.00 frames. ], tot_loss[loss=0.1859, ctc_loss=0.1188, cr_loss=0.3352, over 3347414.04 frames. ], batch size: 48, lr: 2.41e-03, grad_scale: 32.0 2024-09-26 02:32:40,035 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=11.24 vs. limit=22.5 2024-09-26 02:32:40,904 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward2.hidden_balancer.prob, batch_count=886102.0, ans=0.125 2024-09-26 02:32:44,163 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=886102.0, ans=0.125 2024-09-26 02:32:50,746 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=886148.6666666666, ans=0.1 2024-09-26 02:33:13,930 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.143e+02 1.301e+02 1.414e+02 1.534e+02 2.528e+02, threshold=2.827e+02, percent-clipped=0.0 2024-09-26 02:33:20,973 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=12.46 vs. limit=15.0 2024-09-26 02:33:25,145 INFO [train.py:1198] (2/4) Epoch 49, batch 2900, loss[loss=0.2079, ctc_loss=0.1361, cr_loss=0.3591, over 17209.00 frames. ], tot_loss[loss=0.1863, ctc_loss=0.1191, cr_loss=0.3363, over 3352477.21 frames. ], batch size: 55, lr: 2.41e-03, grad_scale: 32.0 2024-09-26 02:33:27,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=886242.0, ans=0.1 2024-09-26 02:33:47,533 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=886288.6666666666, ans=0.0 2024-09-26 02:33:54,461 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=13.79 vs. limit=15.0 2024-09-26 02:34:00,249 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=886335.3333333334, ans=0.2 2024-09-26 02:34:37,446 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module2.balancer2.prob, batch_count=886428.6666666666, ans=0.125 2024-09-26 02:34:45,381 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.min_abs, batch_count=886428.6666666666, ans=0.5 2024-09-26 02:34:49,794 INFO [train.py:1198] (2/4) Epoch 49, batch 2950, loss[loss=0.2156, ctc_loss=0.142, cr_loss=0.3681, over 16457.00 frames. ], tot_loss[loss=0.1852, ctc_loss=0.1182, cr_loss=0.3349, over 3360804.37 frames. ], batch size: 66, lr: 2.41e-03, grad_scale: 32.0 2024-09-26 02:34:54,918 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=886475.3333333334, ans=0.07 2024-09-26 02:35:03,008 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.const_attention_rate, batch_count=886475.3333333334, ans=0.025 2024-09-26 02:35:15,747 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=886522.0, ans=0.0 2024-09-26 02:35:25,165 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=886568.6666666666, ans=0.125 2024-09-26 02:35:58,783 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.170e+02 1.303e+02 1.409e+02 1.487e+02 2.405e+02, threshold=2.817e+02, percent-clipped=0.0 2024-09-26 02:35:59,150 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=886662.0, ans=0.1 2024-09-26 02:36:00,505 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=886662.0, ans=0.125 2024-09-26 02:36:03,861 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=886662.0, ans=0.125 2024-09-26 02:36:09,095 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=10.62 vs. limit=15.0 2024-09-26 02:36:09,977 INFO [train.py:1198] (2/4) Epoch 49, batch 3000, loss[loss=0.181, ctc_loss=0.1148, cr_loss=0.3308, over 17074.00 frames. ], tot_loss[loss=0.1852, ctc_loss=0.1182, cr_loss=0.335, over 3370045.11 frames. ], batch size: 43, lr: 2.41e-03, grad_scale: 32.0 2024-09-26 02:36:09,977 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-26 02:36:25,647 INFO [train.py:1230] (2/4) Epoch 49, validation: loss=0.03501, ctc_loss=0.03501, cr_loss=1.043e-14, over 944034.00 frames. 2024-09-26 02:36:25,648 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-26 02:36:27,258 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=886708.6666666666, ans=0.125 2024-09-26 02:36:28,983 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.const_attention_rate, batch_count=886708.6666666666, ans=0.025 2024-09-26 02:36:29,089 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer2.prob, batch_count=886708.6666666666, ans=0.125 2024-09-26 02:36:49,055 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-26 02:37:17,782 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.whiten, num_groups=1, num_channels=256, metric=10.62 vs. limit=12.0 2024-09-26 02:37:39,922 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=886895.3333333334, ans=0.1 2024-09-26 02:37:43,055 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass.skip_rate, batch_count=886895.3333333334, ans=0.07 2024-09-26 02:37:47,486 INFO [train.py:1198] (2/4) Epoch 49, batch 3050, loss[loss=0.1635, ctc_loss=0.1041, cr_loss=0.2971, over 16666.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1183, cr_loss=0.3353, over 3369462.73 frames. ], batch size: 37, lr: 2.41e-03, grad_scale: 32.0 2024-09-26 02:37:47,712 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=886942.0, ans=0.0 2024-09-26 02:38:48,963 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-26 02:38:54,743 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.193e+02 1.323e+02 1.385e+02 1.464e+02 2.781e+02, threshold=2.770e+02, percent-clipped=0.0 2024-09-26 02:39:04,520 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=887175.3333333334, ans=0.0 2024-09-26 02:39:05,809 INFO [train.py:1198] (2/4) Epoch 49, batch 3100, loss[loss=0.198, ctc_loss=0.124, cr_loss=0.3701, over 17352.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1183, cr_loss=0.3356, over 3374786.37 frames. ], batch size: 48, lr: 2.41e-03, grad_scale: 32.0 2024-09-26 02:39:07,610 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.nonlin_attention.balancer.prob, batch_count=887175.3333333334, ans=0.125 2024-09-26 02:39:11,217 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=6.45 vs. limit=15.0 2024-09-26 02:39:12,413 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer2.prob, batch_count=887175.3333333334, ans=0.125 2024-09-26 02:39:14,013 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=887175.3333333334, ans=0.125 2024-09-26 02:39:40,245 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=887268.6666666666, ans=0.0 2024-09-26 02:39:43,440 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=887268.6666666666, ans=0.125 2024-09-26 02:39:45,068 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-26 02:39:52,168 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=3.96 vs. limit=15.0 2024-09-26 02:40:02,530 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=887315.3333333334, ans=0.125 2024-09-26 02:40:05,871 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=384, metric=4.14 vs. limit=15.0 2024-09-26 02:40:26,354 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.25 vs. limit=15.0 2024-09-26 02:40:26,987 INFO [train.py:1198] (2/4) Epoch 49, batch 3150, loss[loss=0.199, ctc_loss=0.1243, cr_loss=0.3739, over 17316.00 frames. ], tot_loss[loss=0.1856, ctc_loss=0.1184, cr_loss=0.3356, over 3362451.16 frames. ], batch size: 51, lr: 2.41e-03, grad_scale: 32.0 2024-09-26 02:40:47,671 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=887455.3333333334, ans=0.04949747468305833 2024-09-26 02:40:56,939 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=887502.0, ans=0.1 2024-09-26 02:41:06,786 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=512, metric=10.36 vs. limit=15.0 2024-09-26 02:41:35,930 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.172e+02 1.299e+02 1.369e+02 1.479e+02 2.318e+02, threshold=2.737e+02, percent-clipped=0.0 2024-09-26 02:41:45,297 INFO [train.py:1198] (2/4) Epoch 49, batch 3200, loss[loss=0.2034, ctc_loss=0.1293, cr_loss=0.3705, over 16544.00 frames. ], tot_loss[loss=0.1857, ctc_loss=0.1185, cr_loss=0.3359, over 3354448.75 frames. ], batch size: 66, lr: 2.41e-03, grad_scale: 32.0 2024-09-26 02:41:50,351 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=887642.0, ans=0.125 2024-09-26 02:42:05,719 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=887688.6666666666, ans=0.025 2024-09-26 02:42:15,212 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=887735.3333333334, ans=0.0 2024-09-26 02:42:18,720 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.92 vs. limit=15.0 2024-09-26 02:42:43,438 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer1.prob, batch_count=887782.0, ans=0.125 2024-09-26 02:43:05,707 INFO [train.py:1198] (2/4) Epoch 49, batch 3250, loss[loss=0.2073, ctc_loss=0.1345, cr_loss=0.3639, over 16861.00 frames. ], tot_loss[loss=0.1846, ctc_loss=0.1178, cr_loss=0.334, over 3355807.77 frames. ], batch size: 58, lr: 2.41e-03, grad_scale: 32.0 2024-09-26 02:43:10,726 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass_mid.scale_min, batch_count=887875.3333333334, ans=0.2 2024-09-26 02:43:44,393 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.84 vs. limit=10.0 2024-09-26 02:44:09,260 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=888062.0, ans=0.125 2024-09-26 02:44:18,280 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.165e+02 1.305e+02 1.378e+02 1.462e+02 1.990e+02, threshold=2.757e+02, percent-clipped=0.0 2024-09-26 02:44:26,140 INFO [train.py:1198] (2/4) Epoch 49, batch 3300, loss[loss=0.1884, ctc_loss=0.1184, cr_loss=0.3501, over 17027.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1184, cr_loss=0.3351, over 3351242.67 frames. ], batch size: 44, lr: 2.41e-03, grad_scale: 16.0 2024-09-26 02:44:59,886 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.80 vs. limit=22.5 2024-09-26 02:45:13,375 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=888248.6666666666, ans=0.0 2024-09-26 02:45:26,016 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=888248.6666666666, ans=0.2 2024-09-26 02:45:30,521 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=888295.3333333334, ans=0.125 2024-09-26 02:45:33,885 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=888295.3333333334, ans=0.2 2024-09-26 02:45:37,556 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=5.54 vs. limit=12.0 2024-09-26 02:45:44,479 INFO [train.py:1198] (2/4) Epoch 49, batch 3350, loss[loss=0.182, ctc_loss=0.1141, cr_loss=0.3396, over 17071.00 frames. ], tot_loss[loss=0.1843, ctc_loss=0.1176, cr_loss=0.3335, over 3358313.20 frames. ], batch size: 46, lr: 2.41e-03, grad_scale: 16.0 2024-09-26 02:45:50,935 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-26 02:45:57,123 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=888342.0, ans=0.1 2024-09-26 02:46:00,452 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=888388.6666666666, ans=0.125 2024-09-26 02:46:12,916 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer2.prob, batch_count=888388.6666666666, ans=0.125 2024-09-26 02:46:26,928 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=888435.3333333334, ans=0.0 2024-09-26 02:46:28,755 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=888435.3333333334, ans=0.125 2024-09-26 02:46:33,274 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=888482.0, ans=0.0 2024-09-26 02:46:37,958 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=888482.0, ans=0.2 2024-09-26 02:46:49,056 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff2_skip_rate, batch_count=888528.6666666666, ans=0.0 2024-09-26 02:46:54,936 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.181e+02 1.320e+02 1.389e+02 1.476e+02 1.760e+02, threshold=2.779e+02, percent-clipped=0.0 2024-09-26 02:46:57,421 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.30 vs. limit=15.0 2024-09-26 02:47:02,746 INFO [train.py:1198] (2/4) Epoch 49, batch 3400, loss[loss=0.1752, ctc_loss=0.1101, cr_loss=0.3258, over 17246.00 frames. ], tot_loss[loss=0.185, ctc_loss=0.1181, cr_loss=0.3344, over 3357670.73 frames. ], batch size: 44, lr: 2.41e-03, grad_scale: 16.0 2024-09-26 02:47:03,046 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=888575.3333333334, ans=0.1 2024-09-26 02:47:36,483 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=888668.6666666666, ans=0.125 2024-09-26 02:48:23,485 INFO [train.py:1198] (2/4) Epoch 49, batch 3450, loss[loss=0.1801, ctc_loss=0.1135, cr_loss=0.3332, over 16721.00 frames. ], tot_loss[loss=0.1841, ctc_loss=0.1175, cr_loss=0.3329, over 3355999.75 frames. ], batch size: 61, lr: 2.41e-03, grad_scale: 16.0 2024-09-26 02:48:26,921 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=888808.6666666666, ans=0.125 2024-09-26 02:48:37,090 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.72 vs. limit=6.0 2024-09-26 02:48:53,183 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=888902.0, ans=0.1 2024-09-26 02:48:57,717 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=888902.0, ans=0.0 2024-09-26 02:49:03,835 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=888902.0, ans=0.125 2024-09-26 02:49:05,353 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=888902.0, ans=0.04949747468305833 2024-09-26 02:49:25,503 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=888995.3333333334, ans=0.0 2024-09-26 02:49:34,581 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.147e+02 1.304e+02 1.377e+02 1.518e+02 2.125e+02, threshold=2.754e+02, percent-clipped=0.0 2024-09-26 02:49:34,859 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=888995.3333333334, ans=0.0 2024-09-26 02:49:40,815 INFO [train.py:1198] (2/4) Epoch 49, batch 3500, loss[loss=0.1864, ctc_loss=0.1228, cr_loss=0.3182, over 17092.00 frames. ], tot_loss[loss=0.1856, ctc_loss=0.1187, cr_loss=0.335, over 3348843.33 frames. ], batch size: 49, lr: 2.41e-03, grad_scale: 8.0 2024-09-26 02:50:16,244 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=889135.3333333334, ans=0.125 2024-09-26 02:50:34,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.8.prob, batch_count=889182.0, ans=0.125 2024-09-26 02:50:39,683 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=889182.0, ans=0.125 2024-09-26 02:50:48,346 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module2.whiten, num_groups=1, num_channels=192, metric=5.12 vs. limit=15.0 2024-09-26 02:50:49,439 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=9.36 vs. limit=15.0 2024-09-26 02:50:52,142 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=889228.6666666666, ans=0.0 2024-09-26 02:50:52,259 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.out_combiner.scale_min, batch_count=889228.6666666666, ans=0.2 2024-09-26 02:51:01,304 INFO [train.py:1198] (2/4) Epoch 49, batch 3550, loss[loss=0.2347, ctc_loss=0.1555, cr_loss=0.3958, over 17198.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1185, cr_loss=0.3344, over 3348704.58 frames. ], batch size: 55, lr: 2.41e-03, grad_scale: 8.0 2024-09-26 02:51:12,642 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=889275.3333333334, ans=0.05 2024-09-26 02:52:01,837 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=11.56 vs. limit=15.0 2024-09-26 02:52:13,579 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.119e+02 1.292e+02 1.382e+02 1.482e+02 3.513e+02, threshold=2.765e+02, percent-clipped=1.0 2024-09-26 02:52:19,985 INFO [train.py:1198] (2/4) Epoch 49, batch 3600, loss[loss=0.2062, ctc_loss=0.1315, cr_loss=0.3734, over 17002.00 frames. ], tot_loss[loss=0.1853, ctc_loss=0.1183, cr_loss=0.3346, over 3358779.18 frames. ], batch size: 53, lr: 2.41e-03, grad_scale: 16.0 2024-09-26 02:52:58,257 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=889602.0, ans=0.04949747468305833 2024-09-26 02:52:59,772 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=889602.0, ans=0.0 2024-09-26 02:53:14,128 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.prob, batch_count=889648.6666666666, ans=0.125 2024-09-26 02:53:27,183 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_skip_rate, batch_count=889695.3333333334, ans=0.0 2024-09-26 02:53:30,493 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=889695.3333333334, ans=0.125 2024-09-26 02:53:42,723 INFO [train.py:1198] (2/4) Epoch 49, batch 3650, loss[loss=0.1905, ctc_loss=0.1219, cr_loss=0.3431, over 17220.00 frames. ], tot_loss[loss=0.1848, ctc_loss=0.1179, cr_loss=0.3342, over 3353316.46 frames. ], batch size: 47, lr: 2.41e-03, grad_scale: 8.0 2024-09-26 02:53:43,650 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=256, metric=5.05 vs. limit=15.0 2024-09-26 02:53:47,726 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=889742.0, ans=0.1 2024-09-26 02:53:57,212 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=889788.6666666666, ans=0.1 2024-09-26 02:54:16,235 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module1.whiten, num_groups=1, num_channels=512, metric=7.74 vs. limit=15.0 2024-09-26 02:54:46,208 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=5.06 vs. limit=15.0 2024-09-26 02:54:47,200 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=889928.6666666666, ans=0.1 2024-09-26 02:54:56,317 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.170e+02 1.314e+02 1.382e+02 1.463e+02 2.217e+02, threshold=2.764e+02, percent-clipped=0.0 2024-09-26 02:55:01,083 INFO [train.py:1198] (2/4) Epoch 49, batch 3700, loss[loss=0.1618, ctc_loss=0.1044, cr_loss=0.2872, over 17299.00 frames. ], tot_loss[loss=0.1852, ctc_loss=0.1182, cr_loss=0.3354, over 3361553.10 frames. ], batch size: 46, lr: 2.41e-03, grad_scale: 8.0 2024-09-26 02:55:04,019 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=13.19 vs. limit=15.0 2024-09-26 02:55:05,315 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=889975.3333333334, ans=0.0 2024-09-26 02:55:23,238 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.71 vs. limit=6.0 2024-09-26 02:55:48,923 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=890115.3333333334, ans=0.025 2024-09-26 02:56:01,953 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.const_attention_rate, batch_count=890115.3333333334, ans=0.025 2024-09-26 02:56:10,120 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=512, metric=6.25 vs. limit=15.0 2024-09-26 02:56:15,928 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=890162.0, ans=0.1 2024-09-26 02:56:20,583 INFO [train.py:1198] (2/4) Epoch 49, batch 3750, loss[loss=0.2014, ctc_loss=0.1293, cr_loss=0.3607, over 17009.00 frames. ], tot_loss[loss=0.1858, ctc_loss=0.1186, cr_loss=0.3359, over 3357444.25 frames. ], batch size: 51, lr: 2.41e-03, grad_scale: 8.0 2024-09-26 02:56:22,385 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=890208.6666666666, ans=0.1 2024-09-26 02:57:10,433 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=890348.6666666666, ans=0.125 2024-09-26 02:57:35,557 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.130e+02 1.349e+02 1.442e+02 1.558e+02 6.975e+02, threshold=2.884e+02, percent-clipped=1.0 2024-09-26 02:57:40,293 INFO [train.py:1198] (2/4) Epoch 49, batch 3800, loss[loss=0.1889, ctc_loss=0.1209, cr_loss=0.3398, over 16569.00 frames. ], tot_loss[loss=0.1878, ctc_loss=0.1202, cr_loss=0.3379, over 3320741.14 frames. ], batch size: 66, lr: 2.40e-03, grad_scale: 8.0 2024-09-26 02:57:49,765 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=890442.0, ans=0.0 2024-09-26 02:58:21,358 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=890535.3333333334, ans=0.125 2024-09-26 02:58:23,025 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer2.prob, batch_count=890535.3333333334, ans=0.125 2024-09-26 02:58:57,230 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.ff3_skip_rate, batch_count=890675.3333333334, ans=0.0 2024-09-26 02:58:58,407 INFO [train.py:1198] (2/4) Epoch 49, batch 3850, loss[loss=0.1895, ctc_loss=0.1225, cr_loss=0.3349, over 12090.00 frames. ], tot_loss[loss=0.1908, ctc_loss=0.1225, cr_loss=0.3412, over 3261026.15 frames. ], batch size: 123, lr: 2.40e-03, grad_scale: 8.0 2024-09-26 02:59:05,406 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=890675.3333333334, ans=0.0 2024-09-26 02:59:25,598 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten1.whitening_limit, batch_count=890722.0, ans=10.0 2024-09-26 03:00:03,487 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=7.50 vs. limit=15.0 2024-09-26 03:00:56,113 INFO [train.py:1198] (2/4) Epoch 50, batch 0, loss[loss=0.179, ctc_loss=0.1114, cr_loss=0.3378, over 17108.00 frames. ], tot_loss[loss=0.179, ctc_loss=0.1114, cr_loss=0.3378, over 17108.00 frames. ], batch size: 49, lr: 2.38e-03, grad_scale: 16.0 2024-09-26 03:00:56,114 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-26 03:01:04,864 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.5.encoder.layers.1.self_attn_weights, attn_weights_entropy = tensor([3.5214, 5.0785, 4.8752, 5.0479], device='cuda:2') 2024-09-26 03:01:07,251 INFO [zipformer.py:1858] (2/4) name=encoder.encoders.5.encoder.layers.0.self_attn_weights, attn_weights_entropy = tensor([2.6162, 4.3467, 4.0024, 4.7988], device='cuda:2') 2024-09-26 03:01:12,090 INFO [train.py:1230] (2/4) Epoch 50, validation: loss=0.03452, ctc_loss=0.03452, cr_loss=1.145e-14, over 944034.00 frames. 2024-09-26 03:01:12,090 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-26 03:01:13,603 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.288e+02 1.424e+02 1.584e+02 1.731e+02 2.410e+02, threshold=3.169e+02, percent-clipped=0.0 2024-09-26 03:01:20,233 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=890890.0, ans=0.1 2024-09-26 03:01:20,932 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.whiten, num_groups=1, num_channels=256, metric=3.82 vs. limit=12.0 2024-09-26 03:02:22,577 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=18.56 vs. limit=22.5 2024-09-26 03:02:34,238 INFO [train.py:1198] (2/4) Epoch 50, batch 50, loss[loss=0.1953, ctc_loss=0.1257, cr_loss=0.3483, over 17205.00 frames. ], tot_loss[loss=0.1869, ctc_loss=0.1195, cr_loss=0.3368, over 758990.80 frames. ], batch size: 55, lr: 2.38e-03, grad_scale: 16.0 2024-09-26 03:02:52,706 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.95 vs. limit=6.0 2024-09-26 03:03:03,485 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.conv.2.prob, batch_count=891170.0, ans=0.125 2024-09-26 03:03:28,794 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=891263.3333333334, ans=0.025 2024-09-26 03:03:32,509 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.43 vs. limit=12.0 2024-09-26 03:03:35,165 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.ff3_skip_rate, batch_count=891263.3333333334, ans=0.0 2024-09-26 03:03:57,326 INFO [train.py:1198] (2/4) Epoch 50, batch 100, loss[loss=0.1852, ctc_loss=0.1172, cr_loss=0.3402, over 17137.00 frames. ], tot_loss[loss=0.1872, ctc_loss=0.1199, cr_loss=0.3367, over 1319100.67 frames. ], batch size: 48, lr: 2.38e-03, grad_scale: 16.0 2024-09-26 03:03:58,965 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.109e+02 1.291e+02 1.362e+02 1.431e+02 2.417e+02, threshold=2.724e+02, percent-clipped=0.0 2024-09-26 03:03:59,360 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward3.hidden_balancer.prob, batch_count=891356.6666666666, ans=0.125 2024-09-26 03:04:06,289 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=13.00 vs. limit=22.5 2024-09-26 03:04:19,099 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=4.29 vs. limit=15.0 2024-09-26 03:04:20,060 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer1.prob, batch_count=891403.3333333334, ans=0.125 2024-09-26 03:04:21,717 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=891403.3333333334, ans=0.2 2024-09-26 03:04:24,829 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=891403.3333333334, ans=0.025 2024-09-26 03:04:45,021 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=891450.0, ans=0.125 2024-09-26 03:04:48,012 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer2.prob, batch_count=891496.6666666666, ans=0.125 2024-09-26 03:05:02,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_skip_rate, batch_count=891543.3333333334, ans=0.0 2024-09-26 03:05:22,552 INFO [train.py:1198] (2/4) Epoch 50, batch 150, loss[loss=0.1799, ctc_loss=0.1136, cr_loss=0.3318, over 17022.00 frames. ], tot_loss[loss=0.1885, ctc_loss=0.1207, cr_loss=0.3391, over 1775165.10 frames. ], batch size: 44, lr: 2.38e-03, grad_scale: 16.0 2024-09-26 03:05:42,402 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=891636.6666666666, ans=0.125 2024-09-26 03:06:15,957 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=891730.0, ans=0.0 2024-09-26 03:06:19,146 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.prob, batch_count=891730.0, ans=0.125 2024-09-26 03:06:19,230 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=891730.0, ans=0.2 2024-09-26 03:06:45,776 INFO [train.py:1198] (2/4) Epoch 50, batch 200, loss[loss=0.2249, ctc_loss=0.1486, cr_loss=0.3816, over 15919.00 frames. ], tot_loss[loss=0.1865, ctc_loss=0.1192, cr_loss=0.3366, over 2123505.14 frames. ], batch size: 74, lr: 2.38e-03, grad_scale: 16.0 2024-09-26 03:06:47,296 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.198e+02 1.325e+02 1.401e+02 1.516e+02 2.050e+02, threshold=2.802e+02, percent-clipped=0.0 2024-09-26 03:06:49,186 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=891823.3333333334, ans=0.1 2024-09-26 03:06:52,288 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=891823.3333333334, ans=0.125 2024-09-26 03:07:09,791 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=891870.0, ans=0.1 2024-09-26 03:07:24,958 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=18.30 vs. limit=22.5 2024-09-26 03:07:26,165 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.01 vs. limit=6.0 2024-09-26 03:07:33,945 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=891963.3333333334, ans=0.0 2024-09-26 03:07:38,582 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass.skip_rate, batch_count=891963.3333333334, ans=0.035 2024-09-26 03:07:46,611 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=891963.3333333334, ans=0.125 2024-09-26 03:08:02,607 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=892010.0, ans=0.1 2024-09-26 03:08:05,418 INFO [train.py:1198] (2/4) Epoch 50, batch 250, loss[loss=0.1916, ctc_loss=0.1241, cr_loss=0.3378, over 15989.00 frames. ], tot_loss[loss=0.1855, ctc_loss=0.1183, cr_loss=0.3358, over 2402680.91 frames. ], batch size: 74, lr: 2.38e-03, grad_scale: 16.0 2024-09-26 03:08:25,980 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=892103.3333333334, ans=0.125 2024-09-26 03:08:34,070 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=892103.3333333334, ans=0.0 2024-09-26 03:08:35,713 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=892103.3333333334, ans=0.0 2024-09-26 03:08:51,530 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=892150.0, ans=0.0 2024-09-26 03:08:53,136 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=892150.0, ans=0.125 2024-09-26 03:09:14,803 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=192, metric=13.57 vs. limit=15.0 2024-09-26 03:09:28,275 INFO [train.py:1198] (2/4) Epoch 50, batch 300, loss[loss=0.1742, ctc_loss=0.1128, cr_loss=0.3066, over 17029.00 frames. ], tot_loss[loss=0.185, ctc_loss=0.118, cr_loss=0.335, over 2616479.90 frames. ], batch size: 51, lr: 2.38e-03, grad_scale: 16.0 2024-09-26 03:09:29,767 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.113e+02 1.289e+02 1.359e+02 1.478e+02 2.731e+02, threshold=2.717e+02, percent-clipped=0.0 2024-09-26 03:09:44,412 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=892290.0, ans=0.2 2024-09-26 03:09:50,552 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=892336.6666666666, ans=0.125 2024-09-26 03:09:51,245 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=512, metric=9.10 vs. limit=15.0 2024-09-26 03:09:58,863 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.ff3_skip_rate, batch_count=892336.6666666666, ans=0.0 2024-09-26 03:10:16,627 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=892383.3333333334, ans=0.1 2024-09-26 03:10:32,877 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=892430.0, ans=0.09899494936611666 2024-09-26 03:10:40,859 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=892476.6666666666, ans=0.125 2024-09-26 03:10:55,338 INFO [train.py:1198] (2/4) Epoch 50, batch 350, loss[loss=0.1873, ctc_loss=0.1187, cr_loss=0.3431, over 17243.00 frames. ], tot_loss[loss=0.1849, ctc_loss=0.1179, cr_loss=0.3349, over 2783840.27 frames. ], batch size: 44, lr: 2.38e-03, grad_scale: 16.0 2024-09-26 03:11:34,418 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=10.03 vs. limit=15.0 2024-09-26 03:11:44,757 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass_mid.scale_min, batch_count=892663.3333333334, ans=0.2 2024-09-26 03:11:49,705 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module2.balancer2.min_abs, batch_count=892663.3333333334, ans=0.5 2024-09-26 03:11:56,146 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=11.58 vs. limit=15.0 2024-09-26 03:12:17,699 INFO [train.py:1198] (2/4) Epoch 50, batch 400, loss[loss=0.183, ctc_loss=0.1156, cr_loss=0.3368, over 17371.00 frames. ], tot_loss[loss=0.1846, ctc_loss=0.1177, cr_loss=0.3343, over 2918722.14 frames. ], batch size: 48, lr: 2.38e-03, grad_scale: 32.0 2024-09-26 03:12:19,212 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.189e+02 1.302e+02 1.364e+02 1.437e+02 1.796e+02, threshold=2.729e+02, percent-clipped=0.0 2024-09-26 03:12:42,169 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=4.14 vs. limit=6.0 2024-09-26 03:13:39,998 INFO [train.py:1198] (2/4) Epoch 50, batch 450, loss[loss=0.1867, ctc_loss=0.1177, cr_loss=0.3454, over 16947.00 frames. ], tot_loss[loss=0.184, ctc_loss=0.1173, cr_loss=0.3335, over 3007789.20 frames. ], batch size: 42, lr: 2.38e-03, grad_scale: 16.0 2024-09-26 03:13:59,301 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.out_combiner.scale_min, batch_count=893036.6666666666, ans=0.2 2024-09-26 03:14:31,120 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.54 vs. limit=15.0 2024-09-26 03:15:02,680 INFO [train.py:1198] (2/4) Epoch 50, batch 500, loss[loss=0.2158, ctc_loss=0.1406, cr_loss=0.3764, over 16530.00 frames. ], tot_loss[loss=0.1845, ctc_loss=0.1176, cr_loss=0.3342, over 3093016.79 frames. ], batch size: 66, lr: 2.38e-03, grad_scale: 16.0 2024-09-26 03:15:03,007 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-26 03:15:05,799 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.155e+02 1.293e+02 1.377e+02 1.474e+02 1.981e+02, threshold=2.754e+02, percent-clipped=0.0 2024-09-26 03:15:16,379 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=512, metric=5.79 vs. limit=15.0 2024-09-26 03:15:17,435 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=893270.0, ans=0.1 2024-09-26 03:15:32,425 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.05 vs. limit=6.0 2024-09-26 03:16:15,738 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.prob, batch_count=893410.0, ans=0.125 2024-09-26 03:16:25,401 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=893456.6666666666, ans=0.125 2024-09-26 03:16:26,694 INFO [train.py:1198] (2/4) Epoch 50, batch 550, loss[loss=0.2368, ctc_loss=0.154, cr_loss=0.4143, over 17030.00 frames. ], tot_loss[loss=0.1851, ctc_loss=0.1181, cr_loss=0.3351, over 3157773.57 frames. ], batch size: 52, lr: 2.38e-03, grad_scale: 16.0 2024-09-26 03:16:27,120 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass.skip_rate, batch_count=893456.6666666666, ans=0.07 2024-09-26 03:16:37,569 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.attention_skip_rate, batch_count=893456.6666666666, ans=0.0 2024-09-26 03:17:03,316 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=893550.0, ans=0.1 2024-09-26 03:17:11,143 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=893550.0, ans=0.125 2024-09-26 03:17:11,379 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=893550.0, ans=0.1 2024-09-26 03:17:49,349 INFO [train.py:1198] (2/4) Epoch 50, batch 600, loss[loss=0.2027, ctc_loss=0.1289, cr_loss=0.3691, over 17059.00 frames. ], tot_loss[loss=0.1832, ctc_loss=0.1168, cr_loss=0.3322, over 3203445.24 frames. ], batch size: 46, lr: 2.38e-03, grad_scale: 16.0 2024-09-26 03:17:52,578 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.148e+02 1.327e+02 1.372e+02 1.501e+02 3.825e+02, threshold=2.745e+02, percent-clipped=1.0 2024-09-26 03:17:52,867 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=893690.0, ans=0.0 2024-09-26 03:18:15,519 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=11.04 vs. limit=22.5 2024-09-26 03:18:16,697 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.10 vs. limit=15.0 2024-09-26 03:18:18,396 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=4.92 vs. limit=15.0 2024-09-26 03:18:21,445 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=256, metric=9.94 vs. limit=22.5 2024-09-26 03:18:23,066 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.38 vs. limit=10.0 2024-09-26 03:18:24,336 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module1.balancer2.prob, batch_count=893783.3333333334, ans=0.125 2024-09-26 03:18:25,944 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=893783.3333333334, ans=0.125 2024-09-26 03:18:43,759 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=893830.0, ans=0.125 2024-09-26 03:19:12,059 INFO [train.py:1198] (2/4) Epoch 50, batch 650, loss[loss=0.2179, ctc_loss=0.1476, cr_loss=0.3513, over 11891.00 frames. ], tot_loss[loss=0.1847, ctc_loss=0.1178, cr_loss=0.3342, over 3229604.81 frames. ], batch size: 123, lr: 2.38e-03, grad_scale: 16.0 2024-09-26 03:19:23,690 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer2.prob, batch_count=893923.3333333334, ans=0.125 2024-09-26 03:20:03,300 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=894063.3333333334, ans=0.125 2024-09-26 03:20:07,952 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=894063.3333333334, ans=0.1 2024-09-26 03:20:37,744 INFO [train.py:1198] (2/4) Epoch 50, batch 700, loss[loss=0.1944, ctc_loss=0.1266, cr_loss=0.3388, over 17349.00 frames. ], tot_loss[loss=0.186, ctc_loss=0.1188, cr_loss=0.336, over 3263478.15 frames. ], batch size: 52, lr: 2.38e-03, grad_scale: 16.0 2024-09-26 03:20:40,921 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.133e+02 1.326e+02 1.434e+02 1.550e+02 1.872e+02, threshold=2.869e+02, percent-clipped=0.0 2024-09-26 03:20:46,058 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer1.prob, batch_count=894156.6666666666, ans=0.125 2024-09-26 03:20:55,715 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass_mid.scale_min, batch_count=894203.3333333334, ans=0.2 2024-09-26 03:21:14,810 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.pos_emb_skip_rate, batch_count=894250.0, ans=0.0 2024-09-26 03:21:25,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=894296.6666666666, ans=0.125 2024-09-26 03:21:47,640 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer2.prob, batch_count=894343.3333333334, ans=0.125 2024-09-26 03:22:00,146 INFO [train.py:1198] (2/4) Epoch 50, batch 750, loss[loss=0.2227, ctc_loss=0.1446, cr_loss=0.3905, over 17004.00 frames. ], tot_loss[loss=0.1862, ctc_loss=0.1189, cr_loss=0.3364, over 3290735.23 frames. ], batch size: 53, lr: 2.38e-03, grad_scale: 16.0 2024-09-26 03:22:11,678 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=894390.0, ans=0.07 2024-09-26 03:22:46,604 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.balancer2.prob, batch_count=894530.0, ans=0.125 2024-09-26 03:23:05,689 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.2.self_attn_weights, loss-sum=0.000e+00 2024-09-26 03:23:05,920 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.whiten.whitening_limit, batch_count=894576.6666666666, ans=15.0 2024-09-26 03:23:22,023 INFO [train.py:1198] (2/4) Epoch 50, batch 800, loss[loss=0.1584, ctc_loss=0.1011, cr_loss=0.2867, over 16377.00 frames. ], tot_loss[loss=0.1862, ctc_loss=0.1189, cr_loss=0.3365, over 3315914.62 frames. ], batch size: 36, lr: 2.37e-03, grad_scale: 32.0 2024-09-26 03:23:25,275 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.157e+02 1.307e+02 1.380e+02 1.476e+02 1.772e+02, threshold=2.760e+02, percent-clipped=0.0 2024-09-26 03:23:25,607 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=894623.3333333334, ans=0.125 2024-09-26 03:23:30,541 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=894623.3333333334, ans=0.2 2024-09-26 03:23:33,606 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=894623.3333333334, ans=0.1 2024-09-26 03:23:51,480 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.12 vs. limit=22.5 2024-09-26 03:23:54,167 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=894716.6666666666, ans=0.05 2024-09-26 03:24:10,478 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=894763.3333333334, ans=0.1 2024-09-26 03:24:23,363 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=894763.3333333334, ans=0.0 2024-09-26 03:24:45,377 INFO [train.py:1198] (2/4) Epoch 50, batch 850, loss[loss=0.2062, ctc_loss=0.1323, cr_loss=0.3695, over 17012.00 frames. ], tot_loss[loss=0.1859, ctc_loss=0.1187, cr_loss=0.3362, over 3331901.72 frames. ], batch size: 56, lr: 2.37e-03, grad_scale: 32.0 2024-09-26 03:25:41,608 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=894996.6666666666, ans=0.125 2024-09-26 03:25:43,255 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=894996.6666666666, ans=0.125 2024-09-26 03:26:08,943 INFO [train.py:1198] (2/4) Epoch 50, batch 900, loss[loss=0.2156, ctc_loss=0.1414, cr_loss=0.3707, over 16607.00 frames. ], tot_loss[loss=0.1853, ctc_loss=0.1181, cr_loss=0.3358, over 3347765.48 frames. ], batch size: 66, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 03:26:09,363 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer2.prob, batch_count=895090.0, ans=0.125 2024-09-26 03:26:13,763 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.191e+02 1.310e+02 1.400e+02 1.511e+02 3.836e+02, threshold=2.800e+02, percent-clipped=1.0 2024-09-26 03:26:15,534 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=895090.0, ans=0.1 2024-09-26 03:26:54,903 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=895183.3333333334, ans=0.125 2024-09-26 03:26:56,605 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=895183.3333333334, ans=0.0 2024-09-26 03:26:58,206 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward3.hidden_balancer.prob, batch_count=895230.0, ans=0.125 2024-09-26 03:27:09,946 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=9.32 vs. limit=12.0 2024-09-26 03:27:20,897 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=895276.6666666666, ans=0.125 2024-09-26 03:27:22,820 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=12.42 vs. limit=15.0 2024-09-26 03:27:28,821 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff2_skip_rate, batch_count=895276.6666666666, ans=0.0 2024-09-26 03:27:31,734 INFO [train.py:1198] (2/4) Epoch 50, batch 950, loss[loss=0.1652, ctc_loss=0.1034, cr_loss=0.3092, over 17311.00 frames. ], tot_loss[loss=0.1852, ctc_loss=0.1181, cr_loss=0.3358, over 3352098.64 frames. ], batch size: 51, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 03:27:33,654 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=895323.3333333334, ans=0.125 2024-09-26 03:27:33,776 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=895323.3333333334, ans=0.1 2024-09-26 03:27:46,655 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.3.self_attn1.whiten, num_groups=1, num_channels=512, metric=10.35 vs. limit=22.5 2024-09-26 03:28:03,722 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=384, metric=17.03 vs. limit=22.5 2024-09-26 03:28:20,582 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_skip_rate, batch_count=895463.3333333334, ans=0.0 2024-09-26 03:28:41,612 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer1.min_positive, batch_count=895510.0, ans=0.025 2024-09-26 03:28:54,018 INFO [train.py:1198] (2/4) Epoch 50, batch 1000, loss[loss=0.1893, ctc_loss=0.1219, cr_loss=0.3372, over 17215.00 frames. ], tot_loss[loss=0.1861, ctc_loss=0.1187, cr_loss=0.3369, over 3350092.80 frames. ], batch size: 50, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 03:28:58,809 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.214e+02 1.309e+02 1.391e+02 1.502e+02 1.865e+02, threshold=2.782e+02, percent-clipped=0.0 2024-09-26 03:29:00,710 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=895556.6666666666, ans=0.1 2024-09-26 03:29:07,313 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=895556.6666666666, ans=0.125 2024-09-26 03:30:00,970 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=895743.3333333334, ans=10.0 2024-09-26 03:30:05,663 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=895743.3333333334, ans=0.125 2024-09-26 03:30:09,062 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=895743.3333333334, ans=0.125 2024-09-26 03:30:19,365 INFO [train.py:1198] (2/4) Epoch 50, batch 1050, loss[loss=0.1758, ctc_loss=0.1147, cr_loss=0.3056, over 17364.00 frames. ], tot_loss[loss=0.1865, ctc_loss=0.119, cr_loss=0.3374, over 3347306.57 frames. ], batch size: 48, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 03:30:36,093 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=13.33 vs. limit=22.5 2024-09-26 03:30:43,820 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=895836.6666666666, ans=0.125 2024-09-26 03:30:55,588 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.73 vs. limit=10.0 2024-09-26 03:31:31,119 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=895976.6666666666, ans=0.07 2024-09-26 03:31:44,417 INFO [train.py:1198] (2/4) Epoch 50, batch 1100, loss[loss=0.1693, ctc_loss=0.109, cr_loss=0.3016, over 16908.00 frames. ], tot_loss[loss=0.1859, ctc_loss=0.1186, cr_loss=0.3364, over 3348121.41 frames. ], batch size: 58, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 03:31:44,804 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module2.balancer1.prob, batch_count=896023.3333333334, ans=0.125 2024-09-26 03:31:49,230 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.142e+02 1.312e+02 1.377e+02 1.461e+02 2.179e+02, threshold=2.754e+02, percent-clipped=0.0 2024-09-26 03:31:51,761 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=18.96 vs. limit=22.5 2024-09-26 03:31:54,423 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=896023.3333333334, ans=0.1 2024-09-26 03:32:08,717 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=896070.0, ans=0.1 2024-09-26 03:32:26,337 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.bypass_mid.scale_min, batch_count=896116.6666666666, ans=0.2 2024-09-26 03:32:37,963 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.72 vs. limit=6.0 2024-09-26 03:32:48,657 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.attention_skip_rate, batch_count=896210.0, ans=0.0 2024-09-26 03:32:59,988 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass_mid.scale_min, batch_count=896210.0, ans=0.2 2024-09-26 03:33:06,911 INFO [train.py:1198] (2/4) Epoch 50, batch 1150, loss[loss=0.1959, ctc_loss=0.1285, cr_loss=0.337, over 16492.00 frames. ], tot_loss[loss=0.1865, ctc_loss=0.1191, cr_loss=0.3369, over 3334538.02 frames. ], batch size: 66, lr: 2.37e-03, grad_scale: 8.0 2024-09-26 03:33:07,290 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=896256.6666666666, ans=0.0 2024-09-26 03:33:12,121 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=896256.6666666666, ans=0.125 2024-09-26 03:33:26,441 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.scale_min, batch_count=896303.3333333334, ans=0.2 2024-09-26 03:33:44,109 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-26 03:33:58,671 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module2.balancer1.prob, batch_count=896396.6666666666, ans=0.125 2024-09-26 03:34:08,909 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.21 vs. limit=15.0 2024-09-26 03:34:11,611 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=896443.3333333334, ans=0.0 2024-09-26 03:34:29,984 INFO [train.py:1198] (2/4) Epoch 50, batch 1200, loss[loss=0.1967, ctc_loss=0.1265, cr_loss=0.3515, over 17168.00 frames. ], tot_loss[loss=0.1858, ctc_loss=0.1186, cr_loss=0.3357, over 3333680.30 frames. ], batch size: 45, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 03:34:30,389 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=896490.0, ans=0.0 2024-09-26 03:34:36,176 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.154e+02 1.309e+02 1.371e+02 1.493e+02 2.008e+02, threshold=2.741e+02, percent-clipped=0.0 2024-09-26 03:34:47,708 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-26 03:34:52,289 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=896536.6666666666, ans=0.125 2024-09-26 03:34:56,927 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass_mid.scale_min, batch_count=896536.6666666666, ans=0.2 2024-09-26 03:35:14,834 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=896583.3333333334, ans=0.025 2024-09-26 03:35:25,779 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=896630.0, ans=0.1 2024-09-26 03:35:32,218 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer1.max_abs, batch_count=896630.0, ans=10.0 2024-09-26 03:35:47,019 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.27 vs. limit=15.0 2024-09-26 03:35:52,907 INFO [train.py:1198] (2/4) Epoch 50, batch 1250, loss[loss=0.2049, ctc_loss=0.132, cr_loss=0.3647, over 16751.00 frames. ], tot_loss[loss=0.1858, ctc_loss=0.1186, cr_loss=0.3359, over 3340878.88 frames. ], batch size: 61, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 03:35:54,852 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=896723.3333333334, ans=0.1 2024-09-26 03:35:58,268 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=896723.3333333334, ans=0.125 2024-09-26 03:36:15,816 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=896770.0, ans=0.0 2024-09-26 03:36:15,857 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.bypass_mid.scale_min, batch_count=896770.0, ans=0.2 2024-09-26 03:36:18,965 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=896770.0, ans=0.125 2024-09-26 03:36:26,261 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward3.out_whiten, num_groups=1, num_channels=256, metric=14.09 vs. limit=15.0 2024-09-26 03:37:12,600 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.attention_skip_rate, batch_count=896910.0, ans=0.0 2024-09-26 03:37:15,373 INFO [train.py:1198] (2/4) Epoch 50, batch 1300, loss[loss=0.2195, ctc_loss=0.1376, cr_loss=0.4092, over 17009.00 frames. ], tot_loss[loss=0.186, ctc_loss=0.1188, cr_loss=0.336, over 3345566.67 frames. ], batch size: 56, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 03:37:21,094 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.70 vs. limit=15.0 2024-09-26 03:37:21,703 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.165e+02 1.310e+02 1.392e+02 1.499e+02 2.433e+02, threshold=2.784e+02, percent-clipped=0.0 2024-09-26 03:37:26,819 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=896956.6666666666, ans=0.125 2024-09-26 03:37:36,712 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=14.02 vs. limit=22.5 2024-09-26 03:38:03,023 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.min_abs, batch_count=897050.0, ans=0.5 2024-09-26 03:38:04,785 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=897096.6666666666, ans=0.125 2024-09-26 03:38:06,361 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=897096.6666666666, ans=0.1 2024-09-26 03:38:23,241 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=3.83 vs. limit=15.0 2024-09-26 03:38:31,009 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.24 vs. limit=15.0 2024-09-26 03:38:38,299 INFO [train.py:1198] (2/4) Epoch 50, batch 1350, loss[loss=0.1409, ctc_loss=0.0853, cr_loss=0.2781, over 17206.00 frames. ], tot_loss[loss=0.1856, ctc_loss=0.1185, cr_loss=0.3356, over 3356096.98 frames. ], batch size: 41, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 03:38:38,648 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.attention_skip_rate, batch_count=897190.0, ans=0.0 2024-09-26 03:38:43,392 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=897190.0, ans=0.1 2024-09-26 03:39:04,230 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=897236.6666666666, ans=0.0 2024-09-26 03:39:18,814 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=897283.3333333334, ans=0.09899494936611666 2024-09-26 03:39:48,439 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=897376.6666666666, ans=0.0 2024-09-26 03:40:01,284 INFO [train.py:1198] (2/4) Epoch 50, batch 1400, loss[loss=0.1559, ctc_loss=0.09783, cr_loss=0.2902, over 16927.00 frames. ], tot_loss[loss=0.1842, ctc_loss=0.1175, cr_loss=0.3337, over 3364960.32 frames. ], batch size: 42, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 03:40:04,802 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=897423.3333333334, ans=0.025 2024-09-26 03:40:07,585 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.145e+02 1.312e+02 1.402e+02 1.516e+02 2.766e+02, threshold=2.805e+02, percent-clipped=0.0 2024-09-26 03:40:30,455 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=512, metric=9.34 vs. limit=15.0 2024-09-26 03:40:31,488 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.const_attention_rate, batch_count=897470.0, ans=0.025 2024-09-26 03:40:31,592 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.out_proj.dropout_p, batch_count=897470.0, ans=0.1 2024-09-26 03:40:33,118 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=897470.0, ans=0.0 2024-09-26 03:40:39,492 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=897516.6666666666, ans=0.025 2024-09-26 03:41:24,326 INFO [train.py:1198] (2/4) Epoch 50, batch 1450, loss[loss=0.2174, ctc_loss=0.1411, cr_loss=0.3811, over 16517.00 frames. ], tot_loss[loss=0.1843, ctc_loss=0.1176, cr_loss=0.3339, over 3370181.56 frames. ], batch size: 66, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 03:41:31,047 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=17.35 vs. limit=22.5 2024-09-26 03:41:33,873 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=8.77 vs. limit=15.0 2024-09-26 03:41:52,671 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=897703.3333333334, ans=0.0 2024-09-26 03:42:10,209 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=897750.0, ans=0.125 2024-09-26 03:42:11,837 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=897750.0, ans=0.125 2024-09-26 03:42:13,416 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.prob, batch_count=897796.6666666666, ans=0.125 2024-09-26 03:42:31,670 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.34 vs. limit=15.0 2024-09-26 03:42:46,928 INFO [train.py:1198] (2/4) Epoch 50, batch 1500, loss[loss=0.1415, ctc_loss=0.08658, cr_loss=0.2744, over 16247.00 frames. ], tot_loss[loss=0.1837, ctc_loss=0.117, cr_loss=0.3332, over 3372278.35 frames. ], batch size: 36, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 03:42:53,350 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.088e+02 1.299e+02 1.380e+02 1.479e+02 2.541e+02, threshold=2.761e+02, percent-clipped=0.0 2024-09-26 03:43:41,227 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=898030.0, ans=0.0 2024-09-26 03:44:09,795 INFO [train.py:1198] (2/4) Epoch 50, batch 1550, loss[loss=0.2222, ctc_loss=0.1504, cr_loss=0.3592, over 11588.00 frames. ], tot_loss[loss=0.1845, ctc_loss=0.1176, cr_loss=0.3341, over 3362399.81 frames. ], batch size: 123, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 03:45:12,049 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=256, metric=19.11 vs. limit=22.5 2024-09-26 03:45:23,746 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.51 vs. limit=6.0 2024-09-26 03:45:35,380 INFO [train.py:1198] (2/4) Epoch 50, batch 1600, loss[loss=0.1722, ctc_loss=0.1101, cr_loss=0.3106, over 17217.00 frames. ], tot_loss[loss=0.1849, ctc_loss=0.118, cr_loss=0.3345, over 3356371.79 frames. ], batch size: 50, lr: 2.37e-03, grad_scale: 32.0 2024-09-26 03:45:41,791 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.345e+02 1.418e+02 1.518e+02 2.144e+02, threshold=2.837e+02, percent-clipped=0.0 2024-09-26 03:45:47,323 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=512, metric=5.61 vs. limit=15.0 2024-09-26 03:45:52,143 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.07 vs. limit=15.0 2024-09-26 03:46:17,034 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=898450.0, ans=0.125 2024-09-26 03:46:20,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=898450.0, ans=0.1 2024-09-26 03:46:57,036 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=512, metric=3.86 vs. limit=15.0 2024-09-26 03:46:57,839 INFO [train.py:1198] (2/4) Epoch 50, batch 1650, loss[loss=0.1725, ctc_loss=0.112, cr_loss=0.3027, over 17019.00 frames. ], tot_loss[loss=0.1837, ctc_loss=0.1171, cr_loss=0.3332, over 3363610.27 frames. ], batch size: 44, lr: 2.37e-03, grad_scale: 32.0 2024-09-26 03:47:46,934 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module2.whiten, num_groups=1, num_channels=384, metric=4.17 vs. limit=15.0 2024-09-26 03:47:49,525 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-26 03:48:09,629 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.min_positive, batch_count=898776.6666666666, ans=0.025 2024-09-26 03:48:20,712 INFO [train.py:1198] (2/4) Epoch 50, batch 1700, loss[loss=0.1858, ctc_loss=0.1203, cr_loss=0.3273, over 17098.00 frames. ], tot_loss[loss=0.1845, ctc_loss=0.1177, cr_loss=0.3341, over 3362508.66 frames. ], batch size: 49, lr: 2.37e-03, grad_scale: 32.0 2024-09-26 03:48:27,051 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.137e+02 1.297e+02 1.382e+02 1.469e+02 2.239e+02, threshold=2.764e+02, percent-clipped=0.0 2024-09-26 03:48:30,471 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=898823.3333333334, ans=0.1 2024-09-26 03:48:55,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=898916.6666666666, ans=0.125 2024-09-26 03:49:07,247 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_skip_rate, batch_count=898963.3333333334, ans=0.0 2024-09-26 03:49:09,690 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.0.self_attn2.whiten, num_groups=1, num_channels=192, metric=11.98 vs. limit=22.5 2024-09-26 03:49:11,044 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=8.36 vs. limit=15.0 2024-09-26 03:49:27,523 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.skip_rate, batch_count=899010.0, ans=0.09899494936611666 2024-09-26 03:49:42,838 INFO [train.py:1198] (2/4) Epoch 50, batch 1750, loss[loss=0.1431, ctc_loss=0.08804, cr_loss=0.2751, over 17076.00 frames. ], tot_loss[loss=0.1838, ctc_loss=0.1172, cr_loss=0.3332, over 3366388.54 frames. ], batch size: 43, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 03:49:51,626 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.45 vs. limit=12.0 2024-09-26 03:49:54,407 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=899056.6666666666, ans=0.0 2024-09-26 03:50:38,555 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.conv_module1.balancer1.prob, batch_count=899196.6666666666, ans=0.125 2024-09-26 03:50:51,420 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.min_positive, batch_count=899243.3333333334, ans=0.05 2024-09-26 03:51:02,648 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.bypass.skip_rate, batch_count=899243.3333333334, ans=0.04949747468305833 2024-09-26 03:51:05,453 INFO [train.py:1198] (2/4) Epoch 50, batch 1800, loss[loss=0.1598, ctc_loss=0.0989, cr_loss=0.3047, over 17309.00 frames. ], tot_loss[loss=0.1844, ctc_loss=0.1176, cr_loss=0.334, over 3361440.87 frames. ], batch size: 42, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 03:51:13,301 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.136e+02 1.324e+02 1.397e+02 1.479e+02 2.541e+02, threshold=2.793e+02, percent-clipped=0.0 2024-09-26 03:51:15,639 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.25 vs. limit=10.0 2024-09-26 03:51:20,172 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=899336.6666666666, ans=0.125 2024-09-26 03:51:32,496 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=899336.6666666666, ans=0.1 2024-09-26 03:52:15,195 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.conv_module1.whiten, num_groups=1, num_channels=192, metric=7.72 vs. limit=15.0 2024-09-26 03:52:23,776 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=899476.6666666666, ans=0.0 2024-09-26 03:52:28,317 INFO [train.py:1198] (2/4) Epoch 50, batch 1850, loss[loss=0.1877, ctc_loss=0.1173, cr_loss=0.3521, over 17313.00 frames. ], tot_loss[loss=0.184, ctc_loss=0.1172, cr_loss=0.3339, over 3366511.34 frames. ], batch size: 51, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 03:52:31,879 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.min_positive, batch_count=899523.3333333334, ans=0.05 2024-09-26 03:52:38,631 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=15.16 vs. limit=15.0 2024-09-26 03:52:44,973 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=899570.0, ans=0.0 2024-09-26 03:52:45,251 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=1.70 vs. limit=6.0 2024-09-26 03:53:16,132 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=899616.6666666666, ans=0.125 2024-09-26 03:53:21,638 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.whiten.whitening_limit, batch_count=899663.3333333334, ans=15.0 2024-09-26 03:53:41,861 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.skip_rate, batch_count=899710.0, ans=0.04949747468305833 2024-09-26 03:53:51,421 INFO [train.py:1198] (2/4) Epoch 50, batch 1900, loss[loss=0.182, ctc_loss=0.1134, cr_loss=0.3432, over 17080.00 frames. ], tot_loss[loss=0.1848, ctc_loss=0.1179, cr_loss=0.3346, over 3372896.72 frames. ], batch size: 43, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 03:53:59,441 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.136e+02 1.288e+02 1.384e+02 1.484e+02 2.274e+02, threshold=2.768e+02, percent-clipped=0.0 2024-09-26 03:54:06,196 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=899803.3333333334, ans=0.1 2024-09-26 03:54:07,855 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward2.hidden_balancer.prob, batch_count=899803.3333333334, ans=0.125 2024-09-26 03:54:11,104 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.attention_skip_rate, batch_count=899803.3333333334, ans=0.0 2024-09-26 03:54:38,132 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=3.24 vs. limit=15.0 2024-09-26 03:54:48,822 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer1.prob, batch_count=899896.6666666666, ans=0.125 2024-09-26 03:55:03,912 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=4.63 vs. limit=15.0 2024-09-26 03:55:14,170 INFO [train.py:1198] (2/4) Epoch 50, batch 1950, loss[loss=0.1937, ctc_loss=0.1225, cr_loss=0.3558, over 16165.00 frames. ], tot_loss[loss=0.1845, ctc_loss=0.1176, cr_loss=0.3341, over 3367972.93 frames. ], batch size: 74, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 03:55:26,904 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-26 03:56:00,321 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=900083.3333333334, ans=0.0 2024-09-26 03:56:16,355 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.1.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-26 03:56:22,158 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=900176.6666666666, ans=0.1 2024-09-26 03:56:28,612 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=900176.6666666666, ans=0.1 2024-09-26 03:56:39,349 INFO [train.py:1198] (2/4) Epoch 50, batch 2000, loss[loss=0.1761, ctc_loss=0.1116, cr_loss=0.3222, over 17353.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1184, cr_loss=0.3351, over 3346961.73 frames. ], batch size: 48, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 03:56:48,773 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.198e+02 1.333e+02 1.399e+02 1.506e+02 2.393e+02, threshold=2.799e+02, percent-clipped=0.0 2024-09-26 03:56:49,154 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=900223.3333333334, ans=0.125 2024-09-26 03:57:09,985 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=900316.6666666666, ans=0.125 2024-09-26 03:57:38,271 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.feed_forward1.out_proj.dropout_p, batch_count=900363.3333333334, ans=0.1 2024-09-26 03:57:50,142 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=11.66 vs. limit=22.5 2024-09-26 03:58:01,483 INFO [train.py:1198] (2/4) Epoch 50, batch 2050, loss[loss=0.1766, ctc_loss=0.1125, cr_loss=0.3206, over 17003.00 frames. ], tot_loss[loss=0.1852, ctc_loss=0.1182, cr_loss=0.3351, over 3353892.28 frames. ], batch size: 39, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 03:58:01,755 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module1.balancer2.prob, batch_count=900456.6666666666, ans=0.125 2024-09-26 03:58:03,299 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer2.prob, batch_count=900456.6666666666, ans=0.125 2024-09-26 03:58:06,478 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=900456.6666666666, ans=0.1 2024-09-26 03:58:54,054 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=900596.6666666666, ans=0.0 2024-09-26 03:58:55,777 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=900596.6666666666, ans=0.125 2024-09-26 03:59:22,788 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.ff2_skip_rate, batch_count=900690.0, ans=0.0 2024-09-26 03:59:24,081 INFO [train.py:1198] (2/4) Epoch 50, batch 2100, loss[loss=0.2049, ctc_loss=0.1339, cr_loss=0.3553, over 14897.00 frames. ], tot_loss[loss=0.1843, ctc_loss=0.1175, cr_loss=0.3338, over 3353973.17 frames. ], batch size: 89, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 03:59:24,409 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass.skip_rate, batch_count=900690.0, ans=0.09899494936611666 2024-09-26 03:59:33,716 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.174e+02 1.304e+02 1.389e+02 1.531e+02 2.569e+02, threshold=2.777e+02, percent-clipped=0.0 2024-09-26 03:59:35,728 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.scale_min, batch_count=900690.0, ans=0.2 2024-09-26 03:59:51,679 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.prob, batch_count=900736.6666666666, ans=0.125 2024-09-26 03:59:56,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=900783.3333333334, ans=0.0 2024-09-26 04:00:14,322 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=900830.0, ans=0.0 2024-09-26 04:00:39,826 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=4.08 vs. limit=12.0 2024-09-26 04:00:40,864 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.prob, batch_count=900876.6666666666, ans=0.125 2024-09-26 04:00:46,959 INFO [train.py:1198] (2/4) Epoch 50, batch 2150, loss[loss=0.1762, ctc_loss=0.1142, cr_loss=0.31, over 17218.00 frames. ], tot_loss[loss=0.1841, ctc_loss=0.1175, cr_loss=0.333, over 3350234.96 frames. ], batch size: 50, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 04:01:39,598 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=901063.3333333334, ans=0.0 2024-09-26 04:02:10,064 INFO [train.py:1198] (2/4) Epoch 50, batch 2200, loss[loss=0.2393, ctc_loss=0.1585, cr_loss=0.404, over 17221.00 frames. ], tot_loss[loss=0.184, ctc_loss=0.1175, cr_loss=0.3327, over 3352266.34 frames. ], batch size: 55, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 04:02:19,447 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.179e+02 1.317e+02 1.377e+02 1.488e+02 2.594e+02, threshold=2.754e+02, percent-clipped=0.0 2024-09-26 04:02:19,734 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=901156.6666666666, ans=0.0 2024-09-26 04:03:01,909 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-26 04:03:04,996 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=901296.6666666666, ans=0.125 2024-09-26 04:03:05,151 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.hidden_balancer.prob, batch_count=901296.6666666666, ans=0.125 2024-09-26 04:03:32,031 INFO [train.py:1198] (2/4) Epoch 50, batch 2250, loss[loss=0.2155, ctc_loss=0.1386, cr_loss=0.3843, over 17004.00 frames. ], tot_loss[loss=0.185, ctc_loss=0.1181, cr_loss=0.3346, over 3351786.61 frames. ], batch size: 56, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 04:03:46,780 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff2_skip_rate, batch_count=901436.6666666666, ans=0.0 2024-09-26 04:03:48,564 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=11.50 vs. limit=15.0 2024-09-26 04:03:54,690 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.balancer1.prob, batch_count=901436.6666666666, ans=0.125 2024-09-26 04:04:18,145 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=901483.3333333334, ans=0.1 2024-09-26 04:04:27,603 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=901530.0, ans=0.1 2024-09-26 04:04:50,993 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.59 vs. limit=15.0 2024-09-26 04:04:54,691 INFO [train.py:1198] (2/4) Epoch 50, batch 2300, loss[loss=0.1809, ctc_loss=0.1153, cr_loss=0.3281, over 16754.00 frames. ], tot_loss[loss=0.1836, ctc_loss=0.1172, cr_loss=0.332, over 3348919.90 frames. ], batch size: 61, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 04:04:59,887 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.balancer2.prob, batch_count=901623.3333333334, ans=0.125 2024-09-26 04:05:04,460 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.141e+02 1.310e+02 1.388e+02 1.508e+02 2.945e+02, threshold=2.775e+02, percent-clipped=1.0 2024-09-26 04:05:16,427 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.self_attn2.whiten, num_groups=1, num_channels=512, metric=11.59 vs. limit=22.5 2024-09-26 04:05:20,042 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer2.prob, batch_count=901670.0, ans=0.125 2024-09-26 04:05:24,722 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_skip_rate, batch_count=901670.0, ans=0.0 2024-09-26 04:05:42,227 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.balancer.min_positive, batch_count=901716.6666666666, ans=0.05 2024-09-26 04:05:50,239 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=901763.3333333334, ans=0.025 2024-09-26 04:05:59,893 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=901810.0, ans=10.0 2024-09-26 04:06:19,856 INFO [train.py:1198] (2/4) Epoch 50, batch 2350, loss[loss=0.1901, ctc_loss=0.1237, cr_loss=0.3318, over 17040.00 frames. ], tot_loss[loss=0.183, ctc_loss=0.1167, cr_loss=0.3312, over 3353859.11 frames. ], batch size: 51, lr: 2.37e-03, grad_scale: 16.0 2024-09-26 04:06:26,527 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.balancer1.prob, batch_count=901856.6666666666, ans=0.125 2024-09-26 04:06:31,437 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.balancer.prob, batch_count=901856.6666666666, ans=0.125 2024-09-26 04:07:11,703 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.whiten, num_groups=1, num_channels=384, metric=3.48 vs. limit=12.0 2024-09-26 04:07:28,583 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.balancer.min_positive, batch_count=902043.3333333334, ans=0.05 2024-09-26 04:07:39,158 INFO [train.py:1198] (2/4) Epoch 50, batch 2400, loss[loss=0.1816, ctc_loss=0.1139, cr_loss=0.3385, over 17274.00 frames. ], tot_loss[loss=0.1823, ctc_loss=0.1162, cr_loss=0.3307, over 3365899.22 frames. ], batch size: 44, lr: 2.37e-03, grad_scale: 32.0 2024-09-26 04:07:48,369 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=902090.0, ans=0.125 2024-09-26 04:07:51,372 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.173e+02 1.310e+02 1.369e+02 1.448e+02 2.051e+02, threshold=2.738e+02, percent-clipped=0.0 2024-09-26 04:08:05,021 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.15 vs. limit=15.0 2024-09-26 04:08:23,981 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.balancer2.prob, batch_count=902183.3333333334, ans=0.125 2024-09-26 04:08:30,557 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=902230.0, ans=0.125 2024-09-26 04:09:02,438 INFO [train.py:1198] (2/4) Epoch 50, batch 2450, loss[loss=0.1558, ctc_loss=0.09763, cr_loss=0.2906, over 17069.00 frames. ], tot_loss[loss=0.1824, ctc_loss=0.1162, cr_loss=0.3311, over 3357627.81 frames. ], batch size: 40, lr: 2.36e-03, grad_scale: 16.0 2024-09-26 04:09:18,574 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.self_attn_weights.whiten_keys.whitening_limit, batch_count=902323.3333333334, ans=6.0 2024-09-26 04:10:04,698 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.whiten_keys, num_groups=4, num_channels=128, metric=2.99 vs. limit=6.0 2024-09-26 04:10:09,394 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.30 vs. limit=10.0 2024-09-26 04:10:27,375 INFO [train.py:1198] (2/4) Epoch 50, batch 2500, loss[loss=0.2076, ctc_loss=0.1324, cr_loss=0.3764, over 16705.00 frames. ], tot_loss[loss=0.1837, ctc_loss=0.1171, cr_loss=0.333, over 3352577.35 frames. ], batch size: 61, lr: 2.36e-03, grad_scale: 16.0 2024-09-26 04:10:27,729 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.ff3_skip_rate, batch_count=902556.6666666666, ans=0.0 2024-09-26 04:10:35,445 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=902556.6666666666, ans=0.0 2024-09-26 04:10:38,429 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.191e+02 1.330e+02 1.408e+02 1.529e+02 2.285e+02, threshold=2.816e+02, percent-clipped=0.0 2024-09-26 04:10:51,623 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=902603.3333333334, ans=0.125 2024-09-26 04:11:04,251 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=902650.0, ans=0.125 2024-09-26 04:11:05,836 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=902650.0, ans=0.0 2024-09-26 04:11:19,168 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.hidden_balancer.prob, batch_count=902696.6666666666, ans=0.125 2024-09-26 04:11:34,133 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=384, metric=8.67 vs. limit=15.0 2024-09-26 04:11:37,170 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.28 vs. limit=10.0 2024-09-26 04:11:49,378 INFO [train.py:1198] (2/4) Epoch 50, batch 2550, loss[loss=0.1953, ctc_loss=0.1273, cr_loss=0.3401, over 11977.00 frames. ], tot_loss[loss=0.1843, ctc_loss=0.1175, cr_loss=0.3338, over 3354624.71 frames. ], batch size: 123, lr: 2.36e-03, grad_scale: 8.0 2024-09-26 04:12:05,941 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.4.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-26 04:12:07,414 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.self_attn_weights.pos_emb_skip_rate, batch_count=902836.6666666666, ans=0.0 2024-09-26 04:12:28,457 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.const_attention_rate, batch_count=902883.3333333334, ans=0.025 2024-09-26 04:13:03,060 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-26 04:13:12,204 INFO [train.py:1198] (2/4) Epoch 50, batch 2600, loss[loss=0.188, ctc_loss=0.1195, cr_loss=0.3428, over 17306.00 frames. ], tot_loss[loss=0.1846, ctc_loss=0.1177, cr_loss=0.3345, over 3354471.63 frames. ], batch size: 51, lr: 2.36e-03, grad_scale: 8.0 2024-09-26 04:13:15,842 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.balancer1.prob, batch_count=903023.3333333334, ans=0.125 2024-09-26 04:13:18,890 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.min_positive, batch_count=903023.3333333334, ans=0.025 2024-09-26 04:13:25,160 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.157e+02 1.312e+02 1.402e+02 1.467e+02 1.641e+02, threshold=2.804e+02, percent-clipped=0.0 2024-09-26 04:13:33,438 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.const_attention_rate, batch_count=903070.0, ans=0.025 2024-09-26 04:13:35,494 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=288, metric=3.81 vs. limit=10.0 2024-09-26 04:13:41,939 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=11.17 vs. limit=22.5 2024-09-26 04:13:43,241 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer2.prob, batch_count=903116.6666666666, ans=0.125 2024-09-26 04:13:56,089 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.const_attention_rate, batch_count=903116.6666666666, ans=0.025 2024-09-26 04:14:17,794 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-26 04:14:22,407 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.dropout.p, batch_count=903210.0, ans=0.1 2024-09-26 04:14:33,560 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.bypass.skip_rate, batch_count=903256.6666666666, ans=0.07 2024-09-26 04:14:34,856 INFO [train.py:1198] (2/4) Epoch 50, batch 2650, loss[loss=0.1953, ctc_loss=0.1247, cr_loss=0.3528, over 16889.00 frames. ], tot_loss[loss=0.1842, ctc_loss=0.1175, cr_loss=0.3336, over 3350443.40 frames. ], batch size: 58, lr: 2.36e-03, grad_scale: 8.0 2024-09-26 04:14:38,723 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.71 vs. limit=22.5 2024-09-26 04:14:57,439 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module1.balancer2.prob, batch_count=903303.3333333334, ans=0.125 2024-09-26 04:15:22,122 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module1.balancer2.prob, batch_count=903350.0, ans=0.125 2024-09-26 04:15:28,616 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.balancer2.prob, batch_count=903396.6666666666, ans=0.125 2024-09-26 04:15:42,989 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_module2.balancer1.prob, batch_count=903443.3333333334, ans=0.125 2024-09-26 04:15:47,755 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer1.prob, batch_count=903443.3333333334, ans=0.125 2024-09-26 04:15:49,794 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_proj.dropout_p, batch_count=903443.3333333334, ans=0.1 2024-09-26 04:15:57,314 INFO [train.py:1198] (2/4) Epoch 50, batch 2700, loss[loss=0.1597, ctc_loss=0.09955, cr_loss=0.301, over 16402.00 frames. ], tot_loss[loss=0.1836, ctc_loss=0.1171, cr_loss=0.3326, over 3349151.11 frames. ], batch size: 36, lr: 2.36e-03, grad_scale: 8.0 2024-09-26 04:16:12,403 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.141e+02 1.330e+02 1.437e+02 1.530e+02 2.387e+02, threshold=2.875e+02, percent-clipped=0.0 2024-09-26 04:16:17,536 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=903536.6666666666, ans=0.125 2024-09-26 04:16:17,582 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=903536.6666666666, ans=0.125 2024-09-26 04:16:22,212 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=903536.6666666666, ans=0.1 2024-09-26 04:16:35,286 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward1.out_proj.dropout_p, batch_count=903583.3333333334, ans=0.1 2024-09-26 04:17:15,637 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=903676.6666666666, ans=0.0 2024-09-26 04:17:15,669 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff2_skip_rate, batch_count=903676.6666666666, ans=0.0 2024-09-26 04:17:17,425 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=12.98 vs. limit=15.0 2024-09-26 04:17:20,227 INFO [train.py:1198] (2/4) Epoch 50, batch 2750, loss[loss=0.1823, ctc_loss=0.1159, cr_loss=0.3323, over 17317.00 frames. ], tot_loss[loss=0.1845, ctc_loss=0.1177, cr_loss=0.3341, over 3338326.76 frames. ], batch size: 51, lr: 2.36e-03, grad_scale: 8.0 2024-09-26 04:17:23,365 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.hidden_balancer.prob, batch_count=903723.3333333334, ans=0.125 2024-09-26 04:17:42,324 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=903770.0, ans=0.1 2024-09-26 04:17:43,141 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=5.17 vs. limit=15.0 2024-09-26 04:17:51,118 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.2.self_attn2.whiten, num_groups=1, num_channels=384, metric=16.36 vs. limit=22.5 2024-09-26 04:17:55,382 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.attention_skip_rate, batch_count=903816.6666666666, ans=0.0 2024-09-26 04:18:06,309 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=903816.6666666666, ans=0.1 2024-09-26 04:18:25,621 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_skip_rate, batch_count=903910.0, ans=0.0 2024-09-26 04:18:28,656 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.conv_skip_rate, batch_count=903910.0, ans=0.0 2024-09-26 04:18:42,701 INFO [train.py:1198] (2/4) Epoch 50, batch 2800, loss[loss=0.2096, ctc_loss=0.1351, cr_loss=0.3727, over 17196.00 frames. ], tot_loss[loss=0.1857, ctc_loss=0.1185, cr_loss=0.3359, over 3342301.08 frames. ], batch size: 55, lr: 2.36e-03, grad_scale: 16.0 2024-09-26 04:18:57,718 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.167e+02 1.318e+02 1.422e+02 1.529e+02 1.768e+02, threshold=2.845e+02, percent-clipped=0.0 2024-09-26 04:19:02,893 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.bypass_mid.scale_min, batch_count=904003.3333333334, ans=0.2 2024-09-26 04:19:52,357 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.self_attn_weights.pos_emb_skip_rate, batch_count=904143.3333333334, ans=0.0 2024-09-26 04:20:04,785 INFO [train.py:1198] (2/4) Epoch 50, batch 2850, loss[loss=0.1646, ctc_loss=0.102, cr_loss=0.3131, over 16949.00 frames. ], tot_loss[loss=0.1848, ctc_loss=0.1179, cr_loss=0.3346, over 3347615.92 frames. ], batch size: 42, lr: 2.36e-03, grad_scale: 16.0 2024-09-26 04:20:47,296 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.balancer1.prob, batch_count=904283.3333333334, ans=0.125 2024-09-26 04:21:07,148 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=904330.0, ans=0.125 2024-09-26 04:21:20,315 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.nonlin_attention.whiten2, num_groups=1, num_channels=384, metric=3.44 vs. limit=15.0 2024-09-26 04:21:29,402 INFO [train.py:1198] (2/4) Epoch 50, batch 2900, loss[loss=0.2052, ctc_loss=0.1322, cr_loss=0.3648, over 15976.00 frames. ], tot_loss[loss=0.1847, ctc_loss=0.1178, cr_loss=0.3342, over 3344217.58 frames. ], batch size: 74, lr: 2.36e-03, grad_scale: 16.0 2024-09-26 04:21:42,285 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.130e+02 1.318e+02 1.386e+02 1.483e+02 2.505e+02, threshold=2.771e+02, percent-clipped=0.0 2024-09-26 04:21:42,571 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.feed_forward2.hidden_balancer.prob, batch_count=904423.3333333334, ans=0.125 2024-09-26 04:21:45,713 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.balancer1.prob, batch_count=904470.0, ans=0.125 2024-09-26 04:21:58,528 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=904470.0, ans=10.0 2024-09-26 04:22:05,119 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=904516.6666666666, ans=0.1 2024-09-26 04:22:17,539 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.bypass_mid.scale_min, batch_count=904563.3333333334, ans=0.2 2024-09-26 04:22:52,047 INFO [train.py:1198] (2/4) Epoch 50, batch 2950, loss[loss=0.1604, ctc_loss=0.09922, cr_loss=0.306, over 16343.00 frames. ], tot_loss[loss=0.1848, ctc_loss=0.1179, cr_loss=0.3347, over 3351205.98 frames. ], batch size: 36, lr: 2.36e-03, grad_scale: 16.0 2024-09-26 04:22:55,543 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=904656.6666666666, ans=0.0 2024-09-26 04:23:15,565 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.55 vs. limit=15.0 2024-09-26 04:23:27,763 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.bypass_mid.scale_min, batch_count=904750.0, ans=0.2 2024-09-26 04:23:31,093 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.ff3_skip_rate, batch_count=904750.0, ans=0.0 2024-09-26 04:23:34,427 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.const_attention_rate, batch_count=904750.0, ans=0.025 2024-09-26 04:23:49,072 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.2.self_attn_weights.whiten_keys, num_groups=8, num_channels=256, metric=3.27 vs. limit=6.0 2024-09-26 04:24:14,375 INFO [train.py:1198] (2/4) Epoch 50, batch 3000, loss[loss=0.1678, ctc_loss=0.1039, cr_loss=0.3196, over 17015.00 frames. ], tot_loss[loss=0.1841, ctc_loss=0.1174, cr_loss=0.3338, over 3357163.93 frames. ], batch size: 39, lr: 2.36e-03, grad_scale: 16.0 2024-09-26 04:24:14,375 INFO [train.py:1221] (2/4) Computing validation loss 2024-09-26 04:24:30,363 INFO [train.py:1230] (2/4) Epoch 50, validation: loss=0.03495, ctc_loss=0.03495, cr_loss=1.037e-14, over 944034.00 frames. 2024-09-26 04:24:30,364 INFO [train.py:1231] (2/4) Maximum memory allocated so far is 21244MB 2024-09-26 04:24:38,665 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.ff3_skip_rate, batch_count=904890.0, ans=0.0 2024-09-26 04:24:42,934 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.200e+02 1.332e+02 1.410e+02 1.508e+02 3.404e+02, threshold=2.821e+02, percent-clipped=1.0 2024-09-26 04:24:43,157 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.balancer.prob, batch_count=904890.0, ans=0.125 2024-09-26 04:24:44,886 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=904936.6666666666, ans=0.1 2024-09-26 04:24:46,281 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.min_positive, batch_count=904936.6666666666, ans=0.025 2024-09-26 04:25:08,467 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_module2.balancer2.prob, batch_count=904983.3333333334, ans=0.125 2024-09-26 04:25:33,492 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module1.balancer1.max_abs, batch_count=905076.6666666666, ans=10.0 2024-09-26 04:25:48,679 INFO [train.py:1198] (2/4) Epoch 50, batch 3050, loss[loss=0.1997, ctc_loss=0.1283, cr_loss=0.3572, over 16898.00 frames. ], tot_loss[loss=0.1849, ctc_loss=0.1179, cr_loss=0.3349, over 3363829.32 frames. ], batch size: 58, lr: 2.36e-03, grad_scale: 16.0 2024-09-26 04:25:52,928 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.0.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=144, metric=4.81 vs. limit=10.0 2024-09-26 04:26:16,107 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.self_attn_weights.pos_emb_skip_rate, batch_count=905170.0, ans=0.0 2024-09-26 04:26:20,994 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.const_attention_rate, batch_count=905216.6666666666, ans=0.025 2024-09-26 04:26:22,617 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.ff3_skip_rate, batch_count=905216.6666666666, ans=0.0 2024-09-26 04:26:24,699 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=384, metric=9.71 vs. limit=15.0 2024-09-26 04:26:48,900 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer2.prob, batch_count=905263.3333333334, ans=0.125 2024-09-26 04:26:52,187 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.balancer2.prob, batch_count=905310.0, ans=0.125 2024-09-26 04:27:06,933 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=12.77 vs. limit=22.5 2024-09-26 04:27:09,366 INFO [train.py:1198] (2/4) Epoch 50, batch 3100, loss[loss=0.1616, ctc_loss=0.1016, cr_loss=0.3001, over 17251.00 frames. ], tot_loss[loss=0.1852, ctc_loss=0.1181, cr_loss=0.3356, over 3363516.03 frames. ], batch size: 42, lr: 2.36e-03, grad_scale: 16.0 2024-09-26 04:27:12,793 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.bypass_mid.scale_min, batch_count=905356.6666666666, ans=0.2 2024-09-26 04:27:12,873 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=905356.6666666666, ans=0.125 2024-09-26 04:27:17,366 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.balancer1.prob, batch_count=905356.6666666666, ans=0.125 2024-09-26 04:27:21,791 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.114e+02 1.333e+02 1.400e+02 1.486e+02 1.966e+02, threshold=2.799e+02, percent-clipped=0.0 2024-09-26 04:27:25,150 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.nonlin_attention.balancer.prob, batch_count=905403.3333333334, ans=0.125 2024-09-26 04:27:43,851 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.conv_module2.balancer2.prob, batch_count=905450.0, ans=0.125 2024-09-26 04:27:57,102 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module1.balancer1.prob, batch_count=905496.6666666666, ans=0.125 2024-09-26 04:28:08,309 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=905496.6666666666, ans=0.125 2024-09-26 04:28:23,924 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.3.encoder.layers.1.self_attn_weights, loss-sum=0.000e+00 2024-09-26 04:28:25,659 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module2.balancer2.prob, batch_count=905543.3333333334, ans=0.125 2024-09-26 04:28:29,960 INFO [train.py:1198] (2/4) Epoch 50, batch 3150, loss[loss=0.2005, ctc_loss=0.1308, cr_loss=0.3483, over 15936.00 frames. ], tot_loss[loss=0.1862, ctc_loss=0.1189, cr_loss=0.3365, over 3361880.56 frames. ], batch size: 74, lr: 2.36e-03, grad_scale: 16.0 2024-09-26 04:28:31,933 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.feed_forward3.hidden_balancer.prob, batch_count=905590.0, ans=0.125 2024-09-26 04:29:09,416 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.attention_skip_rate, batch_count=905683.3333333334, ans=0.0 2024-09-26 04:29:31,478 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.prob, batch_count=905776.6666666666, ans=0.125 2024-09-26 04:29:36,773 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=8.07 vs. limit=15.0 2024-09-26 04:29:47,546 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=905823.3333333334, ans=0.0 2024-09-26 04:29:48,813 INFO [train.py:1198] (2/4) Epoch 50, batch 3200, loss[loss=0.1691, ctc_loss=0.1077, cr_loss=0.3066, over 17004.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1183, cr_loss=0.3355, over 3361697.33 frames. ], batch size: 51, lr: 2.36e-03, grad_scale: 32.0 2024-09-26 04:30:01,215 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.176e+02 1.309e+02 1.397e+02 1.487e+02 2.037e+02, threshold=2.793e+02, percent-clipped=0.0 2024-09-26 04:30:04,718 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.1.conv_skip_rate, batch_count=905870.0, ans=0.0 2024-09-26 04:30:25,473 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.bypass.scale_min, batch_count=905916.6666666666, ans=0.2 2024-09-26 04:31:07,457 INFO [train.py:1198] (2/4) Epoch 50, batch 3250, loss[loss=0.1465, ctc_loss=0.09375, cr_loss=0.2638, over 16688.00 frames. ], tot_loss[loss=0.1851, ctc_loss=0.118, cr_loss=0.3351, over 3366917.01 frames. ], batch size: 37, lr: 2.36e-03, grad_scale: 32.0 2024-09-26 04:31:07,637 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=906056.6666666666, ans=0.0 2024-09-26 04:32:27,560 INFO [train.py:1198] (2/4) Epoch 50, batch 3300, loss[loss=0.1832, ctc_loss=0.1138, cr_loss=0.3469, over 17158.00 frames. ], tot_loss[loss=0.185, ctc_loss=0.118, cr_loss=0.3348, over 3359294.84 frames. ], batch size: 45, lr: 2.36e-03, grad_scale: 16.0 2024-09-26 04:32:41,661 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.177e+02 1.281e+02 1.376e+02 1.500e+02 2.072e+02, threshold=2.752e+02, percent-clipped=0.0 2024-09-26 04:32:44,236 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.self_attn1.whiten, num_groups=1, num_channels=384, metric=19.56 vs. limit=22.5 2024-09-26 04:33:08,546 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.bypass_mid.scale_min, batch_count=906383.3333333334, ans=0.2 2024-09-26 04:33:35,553 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.feed_forward1.out_proj.dropout_p, batch_count=906476.6666666666, ans=0.1 2024-09-26 04:33:37,201 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.ff3_skip_rate, batch_count=906476.6666666666, ans=0.0 2024-09-26 04:33:43,558 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.conv_module1.balancer2.min_positive, batch_count=906476.6666666666, ans=0.05 2024-09-26 04:33:46,249 INFO [train.py:1198] (2/4) Epoch 50, batch 3350, loss[loss=0.2177, ctc_loss=0.1405, cr_loss=0.3857, over 17327.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1183, cr_loss=0.3354, over 3354442.23 frames. ], batch size: 51, lr: 2.36e-03, grad_scale: 16.0 2024-09-26 04:34:20,390 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.1.self_attn1.whiten, num_groups=1, num_channels=256, metric=12.70 vs. limit=22.5 2024-09-26 04:34:26,407 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.feed_forward1.hidden_balancer.prob, batch_count=906616.6666666666, ans=0.125 2024-09-26 04:34:27,961 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward2.hidden_balancer.prob, batch_count=906616.6666666666, ans=0.125 2024-09-26 04:34:41,909 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder_embed.convnext.out_balancer.prob, batch_count=906663.3333333334, ans=0.125 2024-09-26 04:34:42,003 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff2_skip_rate, batch_count=906663.3333333334, ans=0.0 2024-09-26 04:34:50,139 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=4.09 vs. limit=10.0 2024-09-26 04:35:06,872 INFO [train.py:1198] (2/4) Epoch 50, batch 3400, loss[loss=0.1915, ctc_loss=0.1224, cr_loss=0.3456, over 17217.00 frames. ], tot_loss[loss=0.1861, ctc_loss=0.1187, cr_loss=0.3369, over 3361659.97 frames. ], batch size: 55, lr: 2.36e-03, grad_scale: 16.0 2024-09-26 04:35:20,635 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.200e+02 1.315e+02 1.394e+02 1.507e+02 2.409e+02, threshold=2.789e+02, percent-clipped=0.0 2024-09-26 04:35:30,808 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.feed_forward2.out_whiten, num_groups=1, num_channels=512, metric=7.67 vs. limit=15.0 2024-09-26 04:35:33,498 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=906803.3333333334, ans=0.0 2024-09-26 04:35:39,752 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=906850.0, ans=0.0 2024-09-26 04:35:41,323 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_module2.balancer1.prob, batch_count=906850.0, ans=0.125 2024-09-26 04:35:47,876 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=14.06 vs. limit=15.0 2024-09-26 04:36:06,329 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer2.prob, batch_count=906896.6666666666, ans=0.125 2024-09-26 04:36:23,439 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.prob, batch_count=906990.0, ans=0.125 2024-09-26 04:36:24,724 INFO [train.py:1198] (2/4) Epoch 50, batch 3450, loss[loss=0.1772, ctc_loss=0.1146, cr_loss=0.3129, over 17297.00 frames. ], tot_loss[loss=0.1859, ctc_loss=0.1185, cr_loss=0.3366, over 3363288.95 frames. ], batch size: 46, lr: 2.36e-03, grad_scale: 16.0 2024-09-26 04:36:24,934 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer2.prob, batch_count=906990.0, ans=0.125 2024-09-26 04:36:32,565 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.conv_skip_rate, batch_count=906990.0, ans=0.0 2024-09-26 04:36:44,885 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.nonlin_attention.whiten1, num_groups=1, num_channels=192, metric=3.78 vs. limit=10.0 2024-09-26 04:37:01,193 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.scale_min, batch_count=907083.3333333334, ans=0.2 2024-09-26 04:37:02,627 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.ff2_skip_rate, batch_count=907083.3333333334, ans=0.0 2024-09-26 04:37:21,217 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=907130.0, ans=0.1 2024-09-26 04:37:24,514 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer2.prob, batch_count=907130.0, ans=0.125 2024-09-26 04:37:30,835 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.nonlin_attention.balancer.prob, batch_count=907176.6666666666, ans=0.125 2024-09-26 04:37:32,247 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.1.encoder.layers.0.feed_forward1.hidden_balancer.prob, batch_count=907176.6666666666, ans=0.125 2024-09-26 04:37:42,082 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_skip_rate, batch_count=907176.6666666666, ans=0.0 2024-09-26 04:37:44,967 INFO [train.py:1198] (2/4) Epoch 50, batch 3500, loss[loss=0.1852, ctc_loss=0.1197, cr_loss=0.3274, over 17289.00 frames. ], tot_loss[loss=0.1847, ctc_loss=0.1176, cr_loss=0.3351, over 3372305.82 frames. ], batch size: 51, lr: 2.36e-03, grad_scale: 16.0 2024-09-26 04:37:45,232 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.bypass.skip_rate, batch_count=907223.3333333334, ans=0.09899494936611666 2024-09-26 04:37:58,718 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.132e+02 1.340e+02 1.405e+02 1.556e+02 2.203e+02, threshold=2.811e+02, percent-clipped=0.0 2024-09-26 04:38:03,848 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.5.encoder.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-26 04:38:21,509 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=907316.6666666666, ans=0.1 2024-09-26 04:38:23,063 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.2.bypass.skip_rate, batch_count=907316.6666666666, ans=0.07 2024-09-26 04:38:32,394 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module1.balancer2.min_positive, batch_count=907363.3333333334, ans=0.05 2024-09-26 04:38:35,431 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.balancer1.prob, batch_count=907363.3333333334, ans=0.125 2024-09-26 04:38:47,941 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.1.conv_module2.balancer1.min_positive, batch_count=907410.0, ans=0.025 2024-09-26 04:38:54,410 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.3.bypass.scale_min, batch_count=907410.0, ans=0.2 2024-09-26 04:39:05,112 INFO [train.py:1198] (2/4) Epoch 50, batch 3550, loss[loss=0.2131, ctc_loss=0.1389, cr_loss=0.3707, over 16991.00 frames. ], tot_loss[loss=0.1852, ctc_loss=0.118, cr_loss=0.3357, over 3370594.53 frames. ], batch size: 53, lr: 2.36e-03, grad_scale: 16.0 2024-09-26 04:39:19,364 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_skip_rate, batch_count=907503.3333333334, ans=0.0 2024-09-26 04:39:20,993 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.attention_skip_rate, batch_count=907503.3333333334, ans=0.0 2024-09-26 04:39:24,025 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer1.prob, batch_count=907503.3333333334, ans=0.125 2024-09-26 04:39:30,160 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=907503.3333333334, ans=0.125 2024-09-26 04:39:32,274 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.conv_module1.whiten, num_groups=1, num_channels=384, metric=5.68 vs. limit=15.0 2024-09-26 04:39:43,114 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.3.encoder.layers.1.whiten, num_groups=1, num_channels=512, metric=5.17 vs. limit=12.0 2024-09-26 04:39:47,221 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.bypass.skip_rate, batch_count=907550.0, ans=0.07 2024-09-26 04:39:50,959 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.feed_forward3.out_whiten, num_groups=1, num_channels=384, metric=11.62 vs. limit=15.0 2024-09-26 04:40:04,496 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.conv_module2.balancer1.prob, batch_count=907596.6666666666, ans=0.125 2024-09-26 04:40:04,502 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.balancer2.prob, batch_count=907596.6666666666, ans=0.125 2024-09-26 04:40:22,837 INFO [train.py:1198] (2/4) Epoch 50, batch 3600, loss[loss=0.1904, ctc_loss=0.1219, cr_loss=0.3428, over 16868.00 frames. ], tot_loss[loss=0.184, ctc_loss=0.1172, cr_loss=0.3339, over 3371563.62 frames. ], batch size: 58, lr: 2.36e-03, grad_scale: 32.0 2024-09-26 04:40:36,919 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.188e+02 1.290e+02 1.354e+02 1.454e+02 2.078e+02, threshold=2.707e+02, percent-clipped=0.0 2024-09-26 04:40:37,276 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.1.conv_module2.balancer1.max_abs, batch_count=907736.6666666666, ans=10.0 2024-09-26 04:41:26,327 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.5.encoder.layers.0.feed_forward2.out_whiten, num_groups=1, num_channels=256, metric=13.21 vs. limit=15.0 2024-09-26 04:41:42,941 INFO [train.py:1198] (2/4) Epoch 50, batch 3650, loss[loss=0.1993, ctc_loss=0.1275, cr_loss=0.3593, over 17309.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1182, cr_loss=0.336, over 3368165.01 frames. ], batch size: 49, lr: 2.36e-03, grad_scale: 16.0 2024-09-26 04:42:07,963 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_skip_rate, batch_count=907970.0, ans=0.0 2024-09-26 04:42:13,401 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=4.95 vs. limit=15.0 2024-09-26 04:42:14,161 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.feed_forward3.hidden_balancer.prob, batch_count=908016.6666666666, ans=0.125 2024-09-26 04:42:37,491 INFO [scaling.py:1120] (2/4) WithLoss: name=encoder.encoders.0.layers.0.self_attn_weights, loss-sum=0.000e+00 2024-09-26 04:43:01,636 INFO [train.py:1198] (2/4) Epoch 50, batch 3700, loss[loss=0.1828, ctc_loss=0.1151, cr_loss=0.3384, over 17300.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1182, cr_loss=0.336, over 3367781.59 frames. ], batch size: 51, lr: 2.36e-03, grad_scale: 16.0 2024-09-26 04:43:03,449 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module2.balancer1.prob, batch_count=908156.6666666666, ans=0.125 2024-09-26 04:43:17,257 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.236e+02 1.309e+02 1.373e+02 1.463e+02 1.965e+02, threshold=2.746e+02, percent-clipped=0.0 2024-09-26 04:43:40,223 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.2.encoder.layers.0.whiten, num_groups=1, num_channels=384, metric=5.57 vs. limit=12.0 2024-09-26 04:43:41,168 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=908250.0, ans=0.1 2024-09-26 04:43:59,438 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.feed_forward1.hidden_balancer.prob, batch_count=908296.6666666666, ans=0.125 2024-09-26 04:44:16,545 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.0.conv_module1.balancer2.min_positive, batch_count=908343.3333333334, ans=0.05 2024-09-26 04:44:21,191 INFO [train.py:1198] (2/4) Epoch 50, batch 3750, loss[loss=0.1757, ctc_loss=0.1125, cr_loss=0.3161, over 17301.00 frames. ], tot_loss[loss=0.1855, ctc_loss=0.1183, cr_loss=0.3362, over 3362088.16 frames. ], batch size: 51, lr: 2.36e-03, grad_scale: 16.0 2024-09-26 04:44:32,403 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.feed_forward1.out_proj.dropout_p, batch_count=908390.0, ans=0.1 2024-09-26 04:44:52,602 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.0.layers.0.ff3_skip_rate, batch_count=908483.3333333334, ans=0.0 2024-09-26 04:44:52,651 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.skip_rate, batch_count=908483.3333333334, ans=0.07 2024-09-26 04:45:12,256 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.1.conv_module1.whiten, num_groups=1, num_channels=256, metric=4.80 vs. limit=15.0 2024-09-26 04:45:21,164 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.1.conv_skip_rate, batch_count=908530.0, ans=0.0 2024-09-26 04:45:24,312 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module2.balancer1.prob, batch_count=908576.6666666666, ans=0.125 2024-09-26 04:45:32,232 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.2.conv_module2.balancer1.prob, batch_count=908576.6666666666, ans=0.125 2024-09-26 04:45:39,937 INFO [train.py:1198] (2/4) Epoch 50, batch 3800, loss[loss=0.2187, ctc_loss=0.1457, cr_loss=0.3649, over 11934.00 frames. ], tot_loss[loss=0.1854, ctc_loss=0.1184, cr_loss=0.3353, over 3343996.01 frames. ], batch size: 123, lr: 2.36e-03, grad_scale: 16.0 2024-09-26 04:45:48,160 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.1.conv_module1.balancer2.prob, batch_count=908623.3333333334, ans=0.125 2024-09-26 04:45:55,606 WARNING [optim.py:487] (2/4) Clipping_scale=2.0, grad-norm quartiles 1.149e+02 1.341e+02 1.416e+02 1.505e+02 2.139e+02, threshold=2.833e+02, percent-clipped=0.0 2024-09-26 04:46:02,837 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.1.encoder.layers.0.feed_forward1.out_whiten, num_groups=1, num_channels=256, metric=4.82 vs. limit=15.0 2024-09-26 04:46:44,333 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.5.encoder.layers.0.conv_module1.balancer1.prob, batch_count=908810.0, ans=0.125 2024-09-26 04:46:58,241 INFO [train.py:1198] (2/4) Epoch 50, batch 3850, loss[loss=0.2287, ctc_loss=0.1454, cr_loss=0.4164, over 15069.00 frames. ], tot_loss[loss=0.1882, ctc_loss=0.1204, cr_loss=0.3388, over 3298936.89 frames. ], batch size: 89, lr: 2.36e-03, grad_scale: 16.0 2024-09-26 04:47:17,659 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.3.encoder.layers.0.conv_module2.balancer2.min_positive, batch_count=908903.3333333334, ans=0.05 2024-09-26 04:47:29,075 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.4.encoder.layers.0.conv_module1.balancer1.prob, batch_count=908950.0, ans=0.125 2024-09-26 04:48:01,336 INFO [scaling.py:214] (2/4) ScheduledFloat: name=encoder.encoders.2.encoder.layers.2.bypass.scale_min, batch_count=909043.3333333334, ans=0.2 2024-09-26 04:48:04,653 INFO [scaling.py:1024] (2/4) Whitening: name=encoder.encoders.4.encoder.layers.1.whiten, num_groups=1, num_channels=384, metric=9.82 vs. limit=12.0 2024-09-26 04:48:10,121 INFO [train.py:1496] (2/4) Done!