sino
commited on
Commit
·
68c85ec
1
Parent(s):
a62a2b1
Upload configuration_maelm.py
Browse files- configuration_maelm.py +131 -0
configuration_maelm.py
ADDED
@@ -0,0 +1,131 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2023 the Falcon authors and HuggingFace Inc. team. All rights reserved.
|
3 |
+
#
|
4 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
5 |
+
# you may not use this file except in compliance with the License.
|
6 |
+
# You may obtain a copy of the License at
|
7 |
+
#
|
8 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
9 |
+
#
|
10 |
+
# Unless required by applicable law or agreed to in writing, software
|
11 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
12 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
+
# See the License for the specific language governing permissions and
|
14 |
+
# limitations under the License.
|
15 |
+
""" Falcon configuration"""
|
16 |
+
from transformers.configuration_utils import PretrainedConfig
|
17 |
+
from transformers.utils import logging
|
18 |
+
from transformers import AutoConfig
|
19 |
+
|
20 |
+
logger = logging.get_logger(__name__)
|
21 |
+
|
22 |
+
|
23 |
+
class MAELMConfig(PretrainedConfig):
|
24 |
+
"""
|
25 |
+
This is the configuration class to store the configuration of a [`FalconModel`]. It is used to instantiate a Falcon
|
26 |
+
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
|
27 |
+
defaults will yield a similar configuration to that of the
|
28 |
+
[tiiuae/falcon-7b](https://huggingface.co/tiiuae/falcon-7b) architecture.
|
29 |
+
|
30 |
+
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
31 |
+
documentation from [`PretrainedConfig`] for more information.
|
32 |
+
|
33 |
+
|
34 |
+
Args:
|
35 |
+
vocab_size (`int`, *optional*, defaults to 65024):
|
36 |
+
Vocabulary size of the Falcon model. Defines the number of different tokens that can be represented by the
|
37 |
+
`inputs_ids` passed when calling [`FalconModel`]
|
38 |
+
hidden_size (`int`, *optional*, defaults to 4544):
|
39 |
+
Dimension of the hidden representations.
|
40 |
+
num_hidden_layers (`int`, *optional*, defaults to 32):
|
41 |
+
Number of hidden layers in the Transformer decoder.
|
42 |
+
num_attention_heads (`int`, *optional*, defaults to 71):
|
43 |
+
Number of attention heads for each attention layer in the Transformer encoder.
|
44 |
+
layer_norm_epsilon (`float`, *optional*, defaults to 1e-05):
|
45 |
+
The epsilon used by the layer normalization layers.
|
46 |
+
initializer_range (`float`, *optional*, defaults to 0.02):
|
47 |
+
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
|
48 |
+
use_cache (`bool`, *optional*, defaults to `True`):
|
49 |
+
Whether the model should return the last key/values attentions (not used by all models). Only relevant if
|
50 |
+
`config.is_decoder=True`.
|
51 |
+
hidden_dropout (`float`, *optional*, defaults to 0.0):
|
52 |
+
The dropout probability for MLP layers.
|
53 |
+
attention_dropout (`float`, *optional*, defaults to 0.0):
|
54 |
+
The dropout probability for attention layers.
|
55 |
+
num_kv_heads (`int`, *optional*):
|
56 |
+
Number of key-value heads to use per attention layer. If unset, defaults to the same value as
|
57 |
+
`num_attention_heads`.
|
58 |
+
alibi (`bool`, *optional*, defaults to `False`):
|
59 |
+
Whether to use ALiBi positional biases during self-attention.
|
60 |
+
new_decoder_architecture (`bool`, *optional*, defaults to `False`):
|
61 |
+
Whether to use the new (Falcon-40B) decoder architecture. If `True`, the `multi_query` and `parallel_attn`
|
62 |
+
arguments are ignored, as the new decoder always uses parallel attention.
|
63 |
+
multi_query (`bool`, *optional*, defaults to `True`):
|
64 |
+
Whether to use multi-query attention in the decoder. Ignored when `new_decoder_architecture` is `True`.
|
65 |
+
parallel_attn (`bool`, *optional*, defaults to `True`):
|
66 |
+
Whether to compute attention in parallel with the feedforward layer. If False, they are consecutive
|
67 |
+
instead, as in the original Transformer architecture. Ignored when `new_decoder_architecture` is `True`.
|
68 |
+
bias (`bool`, *optional*, defaults to `False`):
|
69 |
+
Whether to use bias on Linear layers.
|
70 |
+
max_position_embeddings (`int`, *optional*, defaults to 2048):
|
71 |
+
The maximum sequence length that this model might ever be used with, when `alibi` is `False`. Pretrained
|
72 |
+
Falcon models with RoPE support up to 2048 tokens.
|
73 |
+
rope_theta (`float`, *optional*, defaults to 10000.0):
|
74 |
+
The base period of the RoPE embeddings.
|
75 |
+
rope_scaling (`Dict`, *optional*):
|
76 |
+
Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling
|
77 |
+
strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is
|
78 |
+
`{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update
|
79 |
+
`max_position_embeddings` to the expected new maximum. See the following thread for more information on how
|
80 |
+
these scaling strategies behave:
|
81 |
+
https://www.reddit.com/r/LocalLLaMA/comments/14mrgpr/dynamically_scaled_rope_further_increases/. This is an
|
82 |
+
experimental feature, subject to breaking API changes in future versions.
|
83 |
+
bos_token_id (`int`, *optional*, defaults to 11):
|
84 |
+
The id of the "beginning-of-sequence" token.
|
85 |
+
eos_token_id (`int`, *optional*, defaults to 11):
|
86 |
+
The id of the "end-of-sequence" token.
|
87 |
+
"""
|
88 |
+
|
89 |
+
model_type = "MAELM"
|
90 |
+
|
91 |
+
|
92 |
+
def __init__(
|
93 |
+
self,
|
94 |
+
seed=42,
|
95 |
+
cache_dir=None,
|
96 |
+
do_train=True,
|
97 |
+
do_eval=False,
|
98 |
+
do_test=False,
|
99 |
+
dataset_name=None,
|
100 |
+
spect_len=2992,
|
101 |
+
train_dataset_list=[{'train_file': '/mnt/bn/music-nas-dxj1/datasets/MCC_AIGC/mccaigc_train_1w.csv', \
|
102 |
+
'train_tokenized_data': None, 'train_data_root': '/mnt/bn/music-nas-dxj1/datasets/MCC_AIGC/logmel',}],
|
103 |
+
per_device_eval_batch_size=32,
|
104 |
+
preprocessing_num_workers=64,
|
105 |
+
overwrite_cache=True,
|
106 |
+
output_dir='/mnt/bn/music-nas-dxj1/VWork/ckpts_vault/cap_lynx-apm_umg_PT-mccaigc1w_FT',
|
107 |
+
save_interval_steps=1000,
|
108 |
+
overwrite_output_dir=True,
|
109 |
+
gradient_accumulation_steps=1,
|
110 |
+
num_train_epochs=50,
|
111 |
+
per_device_train_batch_size=12,
|
112 |
+
learning_rate=0.00005,
|
113 |
+
lm_lr_ratio=0.1,
|
114 |
+
tokenizer_name='Llama-2-7b-hf',
|
115 |
+
resume_from_checkpoint=None,
|
116 |
+
resume_from_pth='epoch_4-step_8639-allstep_60000.pth',
|
117 |
+
backbone={'name': 'MAEViT', 'arch': 'b', 'patch_size': 16, 'mask_ratio': 0.0, 'img_size': [80, 2992], \
|
118 |
+
'ckpt': 'epoch_20.pth'},
|
119 |
+
neck={'name': 'LMDecoder', 'patch_size': 16, 'img_size': [80, 2992], 'in_chans': 3, 'embed_dim': 768, \
|
120 |
+
'decoder_embed_dim': 4544, 'freeze_decoder': True, 'decoder_type': 'Llama-2-7b-hf'},
|
121 |
+
wandb={'proj': 'ATRena_cap', 'expname': 'cap_lynx_apmPT_mccaigc1wFT'},
|
122 |
+
**kwargs,
|
123 |
+
):
|
124 |
+
self.backbone = backbone
|
125 |
+
self.neck = neck
|
126 |
+
self.tokenizer_name = tokenizer_name
|
127 |
+
self._name_or_path = None
|
128 |
+
self.resume_from_checkpoint = resume_from_checkpoint
|
129 |
+
self.resume_from_pth = resume_from_pth
|
130 |
+
|
131 |
+
AutoConfig.register("MAELM", MAELMConfig)
|