File size: 14,876 Bytes
63d0aed |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 |
# coding=utf-8
# Copyright 2022 Meta and The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" ESM model configuration"""
from dataclasses import asdict, dataclass
from typing import Optional
from transformers import PretrainedConfig, logging
logger = logging.get_logger(__name__)
# TODO Update this
ESM_PRETRAINED_CONFIG_ARCHIVE_MAP = {
"facebook/esm-1b": "https://huggingface.co/facebook/esm-1b/resolve/main/config.json",
# See all ESM models at https://huggingface.co/models?filter=esm
}
class EsmConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`ESMModel`]. It is used to instantiate a ESM model
according to the specified arguments, defining the model architecture. Instantiating a configuration with the
defaults will yield a similar configuration to that of the ESM
[facebook/esm-1b](https://huggingface.co/facebook/esm-1b) architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
vocab_size (`int`, *optional*):
Vocabulary size of the ESM model. Defines the number of different tokens that can be represented by the
`inputs_ids` passed when calling [`ESMModel`].
mask_token_id (`int`, *optional*):
The index of the mask token in the vocabulary. This must be included in the config because of the
"mask-dropout" scaling trick, which will scale the inputs depending on the number of masked tokens.
pad_token_id (`int`, *optional*):
The index of the padding token in the vocabulary. This must be included in the config because certain parts
of the ESM code use this instead of the attention mask.
hidden_size (`int`, *optional*, defaults to 768):
Dimensionality of the encoder layers and the pooler layer.
num_hidden_layers (`int`, *optional*, defaults to 12):
Number of hidden layers in the Transformer encoder.
num_attention_heads (`int`, *optional*, defaults to 12):
Number of attention heads for each attention layer in the Transformer encoder.
intermediate_size (`int`, *optional*, defaults to 3072):
Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
The dropout ratio for the attention probabilities.
max_position_embeddings (`int`, *optional*, defaults to 1026):
The maximum sequence length that this model might ever be used with. Typically set this to something large
just in case (e.g., 512 or 1024 or 2048).
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
layer_norm_eps (`float`, *optional*, defaults to 1e-12):
The epsilon used by the layer normalization layers.
position_embedding_type (`str`, *optional*, defaults to `"absolute"`):
Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query", "rotary"`.
For positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to
[Self-Attention with Relative Position Representations (Shaw et al.)](https://arxiv.org/abs/1803.02155).
For more information on `"relative_key_query"`, please refer to *Method 4* in [Improve Transformer Models
with Better Relative Position Embeddings (Huang et al.)](https://arxiv.org/abs/2009.13658).
is_decoder (`bool`, *optional*, defaults to `False`):
Whether the model is used as a decoder or not. If `False`, the model is used as an encoder.
use_cache (`bool`, *optional*, defaults to `True`):
Whether or not the model should return the last key/values attentions (not used by all models). Only
relevant if `config.is_decoder=True`.
emb_layer_norm_before (`bool`, *optional*):
Whether to apply layer normalization after embeddings but before the main stem of the network.
token_dropout (`bool`, defaults to `False`):
When this is enabled, masked tokens are treated as if they had been dropped out by input dropout.
Examples:
```python
>>> from transformers import EsmModel, EsmConfig
>>> # Initializing a ESM facebook/esm-1b style configuration >>> configuration = EsmConfig()
>>> # Initializing a model from the configuration >>> model = ESMModel(configuration)
>>> # Accessing the model configuration >>> configuration = model.config
```"""
model_type = "esm"
def __init__(
self,
vocab_size=None,
mask_token_id=None,
pad_token_id=None,
hidden_size=768,
num_hidden_layers=12,
num_attention_heads=12,
intermediate_size=3072,
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=1026,
initializer_range=0.02,
layer_norm_eps=1e-12,
position_embedding_type="absolute",
use_cache=True,
emb_layer_norm_before=None,
token_dropout=False,
is_folding_model=False,
esmfold_config=None,
vocab_list=None,
add_bias_fnn=True,
**kwargs,
):
super().__init__(
pad_token_id=pad_token_id, mask_token_id=mask_token_id, **kwargs
)
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.intermediate_size = intermediate_size
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.initializer_range = initializer_range
self.layer_norm_eps = layer_norm_eps
self.position_embedding_type = position_embedding_type
self.use_cache = use_cache
self.emb_layer_norm_before = emb_layer_norm_before
self.token_dropout = token_dropout
self.is_folding_model = is_folding_model
# Arguments needed for Dalmatian
self.add_bias_fnn = add_bias_fnn
if is_folding_model:
if esmfold_config is None:
logger.info(
"No esmfold_config supplied for folding model, using default values."
)
esmfold_config = EsmFoldConfig()
elif isinstance(esmfold_config, dict):
esmfold_config = EsmFoldConfig(**esmfold_config)
self.esmfold_config = esmfold_config
if vocab_list is None:
logger.warning(
"No vocab_list supplied for folding model, assuming the ESM-2 vocabulary!"
)
self.vocab_list = get_default_vocab_list()
else:
self.vocab_list = vocab_list
else:
self.esmfold_config = None
self.vocab_list = None
if self.esmfold_config is not None and getattr(
self.esmfold_config, "use_esm_attn_map", False
):
raise ValueError(
"The HuggingFace port of ESMFold does not support use_esm_attn_map at this time!"
)
def to_dict(self):
"""
Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
Returns:
`Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
"""
output = super().to_dict()
if isinstance(self.esmfold_config, EsmFoldConfig):
output["esmfold_config"] = self.esmfold_config.to_dict()
return output
@dataclass
class EsmFoldConfig:
esm_type: str = None
fp16_esm: bool = True
use_esm_attn_map: bool = False
esm_ablate_pairwise: bool = False
esm_ablate_sequence: bool = False
esm_input_dropout: float = 0
embed_aa: bool = True
bypass_lm: bool = False
lddt_head_hid_dim: int = 128
trunk: "TrunkConfig" = None
def __post_init__(self):
if self.trunk is None:
self.trunk = TrunkConfig()
elif isinstance(self.trunk, dict):
self.trunk = TrunkConfig(**self.trunk)
def to_dict(self):
"""
Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
Returns:
`Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
"""
output = asdict(self)
output["trunk"] = self.trunk.to_dict()
return output
@dataclass
class TrunkConfig:
num_blocks: int = 48
sequence_state_dim: int = 1024
pairwise_state_dim: int = 128
sequence_head_width: int = 32
pairwise_head_width: int = 32
position_bins: int = 32
dropout: float = 0
layer_drop: float = 0
cpu_grad_checkpoint: bool = False
max_recycles: int = 4
chunk_size: Optional[int] = 128
structure_module: "StructureModuleConfig" = None
def __post_init__(self):
if self.structure_module is None:
self.structure_module = StructureModuleConfig()
elif isinstance(self.structure_module, dict):
self.structure_module = StructureModuleConfig(**self.structure_module)
if self.max_recycles <= 0:
raise ValueError(
f"`max_recycles` should be positive, got {self.max_recycles}."
)
if self.sequence_state_dim % self.sequence_state_dim != 0:
raise ValueError(
"`sequence_state_dim` should be a round multiple of `sequence_state_dim`, got"
f" {self.sequence_state_dim} and {self.sequence_state_dim}."
)
if self.pairwise_state_dim % self.pairwise_state_dim != 0:
raise ValueError(
"`pairwise_state_dim` should be a round multiple of `pairwise_state_dim`, got"
f" {self.pairwise_state_dim} and {self.pairwise_state_dim}."
)
sequence_num_heads = self.sequence_state_dim // self.sequence_head_width
pairwise_num_heads = self.pairwise_state_dim // self.pairwise_head_width
if self.sequence_state_dim != sequence_num_heads * self.sequence_head_width:
raise ValueError(
"`sequence_state_dim` should be equal to `sequence_num_heads * sequence_head_width, got"
f" {self.sequence_state_dim} != {sequence_num_heads} * {self.sequence_head_width}."
)
if self.pairwise_state_dim != pairwise_num_heads * self.pairwise_head_width:
raise ValueError(
"`pairwise_state_dim` should be equal to `pairwise_num_heads * pairwise_head_width, got"
f" {self.pairwise_state_dim} != {pairwise_num_heads} * {self.pairwise_head_width}."
)
if self.pairwise_state_dim % 2 != 0:
raise ValueError(
f"`pairwise_state_dim` should be even, got {self.pairwise_state_dim}."
)
if self.dropout >= 0.4:
raise ValueError(
f"`dropout` should not be greater than 0.4, got {self.dropout}."
)
def to_dict(self):
"""
Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
Returns:
`Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
"""
output = asdict(self)
output["structure_module"] = self.structure_module.to_dict()
return output
@dataclass
class StructureModuleConfig:
"""
Args:
sequence_dim:
Single representation channel dimension
pairwise_dim:
Pair representation channel dimension
ipa_dim:
IPA hidden channel dimension
resnet_dim:
Angle resnet (Alg. 23 lines 11-14) hidden channel dimension
num_heads_ipa:
Number of IPA heads
num_qk_points:
Number of query/key points to generate during IPA
num_v_points:
Number of value points to generate during IPA
dropout_rate:
Dropout rate used throughout the layer
num_blocks:
Number of structure module blocks
num_transition_layers:
Number of layers in the single representation transition (Alg. 23 lines 8-9)
num_resnet_blocks:
Number of blocks in the angle resnet
num_angles:
Number of angles to generate in the angle resnet
trans_scale_factor:
Scale of single representation transition hidden dimension
epsilon:
Small number used in angle resnet normalization
inf:
Large number used for attention masking
"""
sequence_dim: int = 384
pairwise_dim: int = 128
ipa_dim: int = 16
resnet_dim: int = 128
num_heads_ipa: int = 12
num_qk_points: int = 4
num_v_points: int = 8
dropout_rate: float = 0.1
num_blocks: int = 8
num_transition_layers: int = 1
num_resnet_blocks: int = 2
num_angles: int = 7
trans_scale_factor: int = 10
epsilon: float = 1e-8
inf: float = 1e5
def to_dict(self):
return asdict(self)
def get_default_vocab_list():
return (
"<cls>",
"<pad>",
"<eos>",
"<unk>",
"L",
"A",
"G",
"V",
"S",
"E",
"R",
"T",
"I",
"D",
"P",
"K",
"Q",
"N",
"F",
"Y",
"M",
"H",
"W",
"C",
"X",
"B",
"U",
"Z",
"O",
".",
"-",
"<null_1>",
"<mask>",
)
|