ESM1b_AAV2_classification

To load tokenizer from ESM, you need to install transformers with this version as follow:

!git clone -b add_esm-proper --single-branch https://github.com/liujas000/transformers.git !pip -q install ./transformers

This model is a fine-tuned version of facebook/esm-1b on AAV2 dataset with ~230k sequences (Bryant et al 2020).

The WT sequence (aa561-588): D E E E I R T T N P V A T E Q Y G S V S T N L Q R G N R Maximum length: 50

It achieves the following results on the evaluation set. Note:this is result of the last epoch, I think the pushed model is loaded with best checkpoint - best val_loss, I'm not so sure though :/

• Loss: 0.2250
• Accuracy: 0.9620
• F1: 0.9632
• Precision: 0.9642
• Recall: 0.9622
• Auroc: 0.9620

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 16
• eval_batch_size: 32
• seed: 42
• gradient_accumulation_steps: 64
• total_train_batch_size: 1024
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 200
• num_epochs: 8
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	F1	Precision	Recall	Auroc
No log	1.0	232	0.1311	0.9495	0.9501	0.9711	0.9299	0.9502
No log	2.0	464	0.1032	0.9606	0.9620	0.9583	0.9657	0.9604
0.1924	3.0	696	0.0995	0.9627	0.9641	0.9584	0.9700	0.9625
0.1924	4.0	928	0.1218	0.9611	0.9624	0.9607	0.9641	0.9610
0.067	5.0	1160	0.1187	0.9622	0.9633	0.9678	0.9588	0.9623
0.067	6.0	1392	0.1514	0.9612	0.9621	0.9710	0.9534	0.9615
0.0271	7.0	1624	0.1890	0.9612	0.9626	0.9580	0.9673	0.9610
0.0271	8.0	1856	0.2250	0.9620	0.9632	0.9642	0.9622	0.9620

Framework versions

• Transformers 4.13.0.dev0
• Pytorch 1.11.0+cu113
• Datasets 2.1.0
• Tokenizers 0.10.3