Upload modeling_fastesm.py with huggingface_hub
Browse files- modeling_fastesm.py +332 -152
modeling_fastesm.py
CHANGED
|
@@ -1,11 +1,13 @@
|
|
| 1 |
import torch
|
| 2 |
import torch.nn as nn
|
|
|
|
| 3 |
from torch.nn import functional as F
|
| 4 |
-
from torch.utils.data import Dataset
|
| 5 |
-
from
|
|
|
|
| 6 |
from einops import rearrange
|
| 7 |
from dataclasses import dataclass
|
| 8 |
-
from transformers import PreTrainedModel, PretrainedConfig, EsmTokenizer
|
| 9 |
from transformers.modeling_outputs import (
|
| 10 |
ModelOutput,
|
| 11 |
BaseModelOutputWithPastAndCrossAttentions,
|
|
@@ -26,31 +28,31 @@ from tqdm.auto import tqdm
|
|
| 26 |
|
| 27 |
@dataclass
|
| 28 |
class EsmMaskedLMOutput(ModelOutput):
|
| 29 |
-
loss: Optional[torch.
|
| 30 |
-
logits: Optional[torch.
|
| 31 |
-
last_hidden_state: Optional[torch.
|
| 32 |
-
hidden_states: Optional[Tuple[torch.
|
| 33 |
-
attentions: Optional[Tuple[torch.
|
| 34 |
|
| 35 |
|
| 36 |
class FastEsmConfig(PretrainedConfig):
|
| 37 |
model_type = "fast_esm"
|
| 38 |
def __init__(
|
| 39 |
self,
|
| 40 |
-
vocab_size=None,
|
| 41 |
-
mask_token_id=None,
|
| 42 |
-
pad_token_id=None,
|
| 43 |
-
hidden_size=768,
|
| 44 |
-
num_hidden_layers=12,
|
| 45 |
-
num_attention_heads=12,
|
| 46 |
-
intermediate_size=3072,
|
| 47 |
-
hidden_dropout_prob=0.1,
|
| 48 |
-
attention_probs_dropout_prob=0.1,
|
| 49 |
-
max_position_embeddings=1026,
|
| 50 |
-
initializer_range=0.02,
|
| 51 |
-
layer_norm_eps=1e-12,
|
| 52 |
-
position_embedding_type="absolute",
|
| 53 |
-
emb_layer_norm_before=None,
|
| 54 |
**kwargs,
|
| 55 |
):
|
| 56 |
super().__init__(pad_token_id=pad_token_id, mask_token_id=mask_token_id, **kwargs)
|
|
@@ -68,35 +70,35 @@ class FastEsmConfig(PretrainedConfig):
|
|
| 68 |
self.position_embedding_type = position_embedding_type
|
| 69 |
self.emb_layer_norm_before = emb_layer_norm_before
|
| 70 |
|
| 71 |
-
def to_dict(self):
|
| 72 |
"""
|
| 73 |
Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
|
| 74 |
|
| 75 |
Returns:
|
| 76 |
-
`Dict[str, any]`:
|
| 77 |
"""
|
| 78 |
output = super().to_dict()
|
| 79 |
return output
|
| 80 |
|
| 81 |
|
| 82 |
-
def rotate_half(x):
|
| 83 |
x1, x2 = x.chunk(2, dim=-1)
|
| 84 |
return torch.cat((-x2, x1), dim=-1)
|
| 85 |
|
| 86 |
|
| 87 |
-
def apply_rotary_pos_emb(x, cos, sin):
|
| 88 |
cos = cos[:, :, : x.shape[-2], :]
|
| 89 |
sin = sin[:, :, : x.shape[-2], :]
|
| 90 |
|
| 91 |
return (x * cos) + (rotate_half(x) * sin)
|
| 92 |
|
| 93 |
|
| 94 |
-
def symmetrize(x):
|
| 95 |
"Make layer symmetric in final two dimensions, used for contact prediction."
|
| 96 |
return x + x.transpose(-1, -2)
|
| 97 |
|
| 98 |
|
| 99 |
-
def average_product_correct(x):
|
| 100 |
"Perform average product correct, used for contact prediction."
|
| 101 |
a1 = x.sum(-1, keepdims=True)
|
| 102 |
a2 = x.sum(-2, keepdims=True)
|
|
@@ -114,18 +116,18 @@ class EsmContactPredictionHead(nn.Module):
|
|
| 114 |
def __init__(
|
| 115 |
self,
|
| 116 |
in_features: int,
|
| 117 |
-
bias=True,
|
| 118 |
eos_idx: int = 2,
|
| 119 |
):
|
| 120 |
super().__init__()
|
| 121 |
self.in_features = in_features
|
| 122 |
self.eos_idx = eos_idx
|
| 123 |
-
self.regression = nn.Linear(in_features, 1, bias)
|
| 124 |
self.activation = nn.Sigmoid()
|
| 125 |
|
| 126 |
-
def forward(self,
|
| 127 |
# remove eos token attentions
|
| 128 |
-
eos_mask =
|
| 129 |
eos_mask = eos_mask.unsqueeze(1) * eos_mask.unsqueeze(2)
|
| 130 |
attentions = attentions * eos_mask[:, None, None, :, :]
|
| 131 |
attentions = attentions[..., :-1, :-1]
|
|
@@ -161,7 +163,7 @@ class RotaryEmbedding(torch.nn.Module):
|
|
| 161 |
self._cos_cached = None
|
| 162 |
self._sin_cached = None
|
| 163 |
|
| 164 |
-
def _update_cos_sin_tables(self, x, seq_dimension=2):
|
| 165 |
seq_len = x.shape[seq_dimension]
|
| 166 |
|
| 167 |
# Reset the tables if the sequence length has changed,
|
|
@@ -204,7 +206,12 @@ class EsmEmbeddings(nn.Module):
|
|
| 204 |
)
|
| 205 |
|
| 206 |
def forward(
|
| 207 |
-
self,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 208 |
):
|
| 209 |
if inputs_embeds is None:
|
| 210 |
inputs_embeds = self.word_embeddings(input_ids)
|
|
@@ -236,7 +243,7 @@ class EsmEmbeddings(nn.Module):
|
|
| 236 |
|
| 237 |
|
| 238 |
class EsmSelfAttention(nn.Module):
|
| 239 |
-
def __init__(self, config, position_embedding_type=None):
|
| 240 |
super().__init__()
|
| 241 |
if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
|
| 242 |
raise ValueError(
|
|
@@ -267,8 +274,8 @@ class EsmSelfAttention(nn.Module):
|
|
| 267 |
def forward(
|
| 268 |
self,
|
| 269 |
hidden_states: torch.Tensor,
|
| 270 |
-
attention_mask: Optional[torch.
|
| 271 |
-
output_attentions: bool = False,
|
| 272 |
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
| 273 |
"""Forward pass for self attention.
|
| 274 |
|
|
@@ -321,8 +328,8 @@ class EsmAttention(nn.Module):
|
|
| 321 |
def forward(
|
| 322 |
self,
|
| 323 |
hidden_states: torch.Tensor,
|
| 324 |
-
attention_mask: Optional[torch.
|
| 325 |
-
output_attentions: bool = False,
|
| 326 |
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
| 327 |
"""Forward pass for attention layer.
|
| 328 |
|
|
@@ -362,8 +369,8 @@ class EsmLayer(nn.Module):
|
|
| 362 |
def forward(
|
| 363 |
self,
|
| 364 |
hidden_states: torch.Tensor,
|
| 365 |
-
attention_mask: Optional[torch.
|
| 366 |
-
output_attentions: bool = False,
|
| 367 |
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
| 368 |
"""Forward pass for transformer layer.
|
| 369 |
|
|
@@ -410,9 +417,9 @@ class EsmEncoder(nn.Module):
|
|
| 410 |
def forward(
|
| 411 |
self,
|
| 412 |
hidden_states: torch.Tensor,
|
| 413 |
-
attention_mask: Optional[torch.
|
| 414 |
-
output_hidden_states: bool = False,
|
| 415 |
-
output_attentions: bool = False,
|
| 416 |
) -> BaseModelOutputWithPastAndCrossAttentions:
|
| 417 |
"""Forward pass for transformer encoder.
|
| 418 |
|
|
@@ -465,8 +472,90 @@ class EsmEncoder(nn.Module):
|
|
| 465 |
)
|
| 466 |
|
| 467 |
|
| 468 |
-
###
|
| 469 |
-
class
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 470 |
"""Simple dataset for protein sequences."""
|
| 471 |
def __init__(self, sequences: list[str]):
|
| 472 |
self.sequences = sequences
|
|
@@ -478,52 +567,22 @@ class ProteinDataset(Dataset):
|
|
| 478 |
return self.sequences[idx]
|
| 479 |
|
| 480 |
|
| 481 |
-
|
| 482 |
-
|
| 483 |
-
|
| 484 |
-
|
| 485 |
-
|
| 486 |
-
config_class = FastEsmConfig
|
| 487 |
-
base_model_prefix = "fastesm"
|
| 488 |
-
supports_gradient_checkpointing = True
|
| 489 |
-
tokenizer = EsmTokenizer.from_pretrained("facebook/esm2_t33_650M_UR50D")
|
| 490 |
-
def _init_weights(self, module):
|
| 491 |
-
"""Initialize the weights"""
|
| 492 |
-
if isinstance(module, nn.Linear):
|
| 493 |
-
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
| 494 |
-
if module.bias is not None:
|
| 495 |
-
module.bias.data.zero_()
|
| 496 |
-
elif isinstance(module, nn.Embedding):
|
| 497 |
-
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
| 498 |
-
if module.padding_idx is not None:
|
| 499 |
-
module.weight.data[module.padding_idx].zero_()
|
| 500 |
-
elif isinstance(module, nn.LayerNorm):
|
| 501 |
-
module.bias.data.zero_()
|
| 502 |
-
module.weight.data.fill_(1.0)
|
| 503 |
|
| 504 |
-
|
| 505 |
-
|
| 506 |
-
|
| 507 |
-
|
| 508 |
-
return self.esm.embeddings.word_embeddings
|
| 509 |
|
| 510 |
@property
|
| 511 |
def device(self) -> torch.device:
|
| 512 |
"""Get the device of the model."""
|
| 513 |
return next(self.parameters()).device
|
| 514 |
|
| 515 |
-
def mean_pooling(self, x: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
|
| 516 |
-
"""Apply mean pooling to sequence outputs."""
|
| 517 |
-
if attention_mask is None:
|
| 518 |
-
return x.mean(dim=1)
|
| 519 |
-
else:
|
| 520 |
-
attention_mask = attention_mask.unsqueeze(-1)
|
| 521 |
-
return (x * attention_mask).sum(dim=1) / attention_mask.sum(dim=1)
|
| 522 |
-
|
| 523 |
-
def _collate_fn(self, sequences: list[str]) -> tuple[torch.Tensor, torch.Tensor]:
|
| 524 |
-
"""Collate function for batching sequences."""
|
| 525 |
-
return self.tokenizer(sequences, return_tensors="pt", padding='longest', pad_to_multiple_of=8)
|
| 526 |
-
|
| 527 |
def _read_sequences_from_db(self, db_path: str) -> set[str]:
|
| 528 |
"""Read sequences from SQLite database."""
|
| 529 |
import sqlite3
|
|
@@ -540,15 +599,18 @@ class FastEsmPreTrainedModel(PreTrainedModel):
|
|
| 540 |
|
| 541 |
def embed_dataset(
|
| 542 |
self,
|
| 543 |
-
sequences:
|
|
|
|
| 544 |
batch_size: int = 2,
|
| 545 |
max_len: int = 512,
|
| 546 |
full_embeddings: bool = False,
|
| 547 |
-
|
| 548 |
-
|
| 549 |
num_workers: int = 0,
|
| 550 |
sql: bool = False,
|
|
|
|
| 551 |
sql_db_path: str = 'embeddings.db',
|
|
|
|
| 552 |
) -> Optional[dict[str, torch.Tensor]]:
|
| 553 |
"""Embed a dataset of protein sequences.
|
| 554 |
|
|
@@ -557,7 +619,6 @@ class FastEsmPreTrainedModel(PreTrainedModel):
|
|
| 557 |
batch_size: Batch size for processing
|
| 558 |
max_len: Maximum sequence length
|
| 559 |
full_embeddings: Whether to return full residue-wise (True) embeddings or pooled (False)
|
| 560 |
-
full_precision: Whether to cast to full precision (float32) before storage - relevant for dict storage
|
| 561 |
pooling_type: Type of pooling ('mean' or 'cls')
|
| 562 |
num_workers: Number of workers for data loading, 0 for the main process
|
| 563 |
sql: Whether to store embeddings in SQLite database - will be stored in float32
|
|
@@ -565,18 +626,46 @@ class FastEsmPreTrainedModel(PreTrainedModel):
|
|
| 565 |
|
| 566 |
Returns:
|
| 567 |
Dictionary mapping sequences to embeddings, or None if sql=True
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 568 |
"""
|
|
|
|
|
|
|
|
|
|
| 569 |
device = self.device
|
|
|
|
| 570 |
|
| 571 |
def get_embeddings(residue_embeddings: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
|
| 572 |
-
if full_embeddings:
|
| 573 |
return residue_embeddings
|
| 574 |
-
elif pooling_type == 'mean':
|
| 575 |
-
return self.mean_pooling(residue_embeddings, attention_mask)
|
| 576 |
else:
|
| 577 |
-
return residue_embeddings
|
| 578 |
|
| 579 |
-
sequences = list(set([seq[:max_len] for seq in sequences]))
|
| 580 |
if sql:
|
| 581 |
import sqlite3
|
| 582 |
conn = sqlite3.connect(sql_db_path)
|
|
@@ -587,52 +676,94 @@ class FastEsmPreTrainedModel(PreTrainedModel):
|
|
| 587 |
print(f"Found {len(already_embedded)} already embedded sequences in {sql_db_path}")
|
| 588 |
print(f"Embedding {len(to_embed)} new sequences")
|
| 589 |
if len(to_embed) > 0:
|
| 590 |
-
to_embed = sorted(to_embed, key=len, reverse=True)
|
| 591 |
dataset = ProteinDataset(to_embed)
|
| 592 |
-
dataloader = DataLoader(dataset, batch_size=batch_size, num_workers=num_workers, collate_fn=
|
| 593 |
with torch.no_grad():
|
| 594 |
for i, batch in tqdm(enumerate(dataloader), total=len(dataloader), desc='Embedding batches'):
|
| 595 |
-
seqs =
|
| 596 |
input_ids, attention_mask = batch['input_ids'].to(device), batch['attention_mask'].to(device)
|
| 597 |
-
residue_embeddings = self.
|
| 598 |
embeddings = get_embeddings(residue_embeddings, attention_mask).cpu()
|
| 599 |
-
|
| 600 |
-
|
|
|
|
| 601 |
c.execute("INSERT OR REPLACE INTO embeddings VALUES (?, ?)",
|
| 602 |
(seq, emb.cpu().numpy().tobytes()))
|
| 603 |
|
| 604 |
if (i + 1) % 100 == 0:
|
| 605 |
conn.commit()
|
| 606 |
-
|
| 607 |
conn.commit()
|
| 608 |
conn.close()
|
| 609 |
return None
|
| 610 |
-
|
| 611 |
-
sequences = list(set([seq[:max_len] for seq in sequences]))
|
| 612 |
-
sequences = sorted(sequences, key=len, reverse=True)
|
| 613 |
-
dataset = ProteinDataset(sequences)
|
| 614 |
-
dataloader = DataLoader(dataset, batch_size=batch_size, num_workers=num_workers, collate_fn=self._collate_fn, shuffle=False)
|
| 615 |
embeddings_dict = {}
|
| 616 |
-
|
| 617 |
-
|
| 618 |
-
|
| 619 |
-
|
| 620 |
-
|
| 621 |
-
|
| 622 |
-
|
| 623 |
-
|
| 624 |
-
|
| 625 |
-
|
| 626 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 627 |
return embeddings_dict
|
| 628 |
|
| 629 |
|
| 630 |
-
class
|
| 631 |
-
|
| 632 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 633 |
self.config = config
|
| 634 |
self.embeddings = EsmEmbeddings(config)
|
| 635 |
self.encoder = EsmEncoder(config)
|
|
|
|
|
|
|
|
|
|
| 636 |
# Initialize weights and apply final processing
|
| 637 |
self.post_init()
|
| 638 |
|
|
@@ -642,12 +773,36 @@ class FAST_ESM_ENCODER(FastEsmPreTrainedModel):
|
|
| 642 |
def set_input_embeddings(self, value):
|
| 643 |
self.embeddings.word_embeddings = value
|
| 644 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 645 |
def forward(
|
| 646 |
self,
|
| 647 |
-
input_ids: Optional[torch.
|
| 648 |
attention_mask: Optional[torch.Tensor] = None,
|
| 649 |
-
position_ids: Optional[torch.
|
| 650 |
-
inputs_embeds: Optional[torch.
|
| 651 |
output_attentions: Optional[bool] = None,
|
| 652 |
output_hidden_states: Optional[bool] = None,
|
| 653 |
return_dict: Optional[bool] = None, # to play nice with HF adjacent packages
|
|
@@ -679,7 +834,7 @@ class FAST_ESM_ENCODER(FastEsmPreTrainedModel):
|
|
| 679 |
raise ValueError("You have to specify either input_ids or inputs_embeds")
|
| 680 |
|
| 681 |
batch_size, seq_length = input_shape
|
| 682 |
-
|
| 683 |
input_ids=input_ids,
|
| 684 |
position_ids=position_ids,
|
| 685 |
attention_mask=attention_mask,
|
|
@@ -694,7 +849,7 @@ class FAST_ESM_ENCODER(FastEsmPreTrainedModel):
|
|
| 694 |
extended_attention_mask = None
|
| 695 |
|
| 696 |
encoder_outputs = self.encoder(
|
| 697 |
-
|
| 698 |
attention_mask=extended_attention_mask,
|
| 699 |
output_hidden_states=output_hidden_states,
|
| 700 |
output_attentions=output_attentions,
|
|
@@ -708,9 +863,9 @@ class FAST_ESM_ENCODER(FastEsmPreTrainedModel):
|
|
| 708 |
)
|
| 709 |
|
| 710 |
|
| 711 |
-
class FastEsmModel(FastEsmPreTrainedModel):
|
| 712 |
-
def __init__(self, config, add_pooling_layer=True):
|
| 713 |
-
super().__init__(config)
|
| 714 |
self.config = config
|
| 715 |
self.esm = FAST_ESM_ENCODER(config)
|
| 716 |
self.pooler = EsmPooler(config) if add_pooling_layer else None
|
|
@@ -723,12 +878,18 @@ class FastEsmModel(FastEsmPreTrainedModel):
|
|
| 723 |
def set_input_embeddings(self, value):
|
| 724 |
self.embeddings.word_embeddings = value
|
| 725 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 726 |
def forward(
|
| 727 |
self,
|
| 728 |
-
input_ids: Optional[torch.
|
| 729 |
attention_mask: Optional[torch.Tensor] = None,
|
| 730 |
-
position_ids: Optional[torch.
|
| 731 |
-
inputs_embeds: Optional[torch.
|
| 732 |
output_attentions: Optional[bool] = None,
|
| 733 |
output_hidden_states: Optional[bool] = None,
|
| 734 |
return_dict: Optional[bool] = None, # to play nice with HF adjacent packages
|
|
@@ -778,11 +939,11 @@ class FastEsmModel(FastEsmPreTrainedModel):
|
|
| 778 |
)
|
| 779 |
|
| 780 |
|
| 781 |
-
class FastEsmForMaskedLM(FastEsmPreTrainedModel):
|
| 782 |
_tied_weights_keys = ["lm_head.decoder.weight"]
|
| 783 |
|
| 784 |
def __init__(self, config):
|
| 785 |
-
super().__init__(config)
|
| 786 |
self.esm = FAST_ESM_ENCODER(config, add_pooling_layer=False)
|
| 787 |
self.lm_head = EsmLMHead(config)
|
| 788 |
self.loss_fct = nn.CrossEntropyLoss()
|
|
@@ -794,13 +955,19 @@ class FastEsmForMaskedLM(FastEsmPreTrainedModel):
|
|
| 794 |
def set_output_embeddings(self, new_embeddings):
|
| 795 |
self.lm_head.decoder = new_embeddings
|
| 796 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 797 |
def forward(
|
| 798 |
self,
|
| 799 |
-
input_ids: Optional[torch.
|
| 800 |
attention_mask: Optional[torch.Tensor] = None,
|
| 801 |
-
position_ids: Optional[torch.
|
| 802 |
-
inputs_embeds: Optional[torch.
|
| 803 |
-
labels: Optional[torch.
|
| 804 |
output_attentions: Optional[bool] = None,
|
| 805 |
output_hidden_states: Optional[bool] = None,
|
| 806 |
return_dict: Optional[bool] = None, # to play nice with HF adjacent packages
|
|
@@ -829,13 +996,10 @@ class FastEsmForMaskedLM(FastEsmPreTrainedModel):
|
|
| 829 |
attentions=outputs.attentions,
|
| 830 |
)
|
| 831 |
|
| 832 |
-
def predict_contacts(self, tokens, attention_mask):
|
| 833 |
-
raise NotImplementedError("predict_contacts is not supported by F.scaled_dot_product_attention")
|
| 834 |
-
|
| 835 |
|
| 836 |
-
class FastEsmForSequenceClassification(FastEsmPreTrainedModel):
|
| 837 |
def __init__(self, config):
|
| 838 |
-
super().__init__(config)
|
| 839 |
self.num_labels = config.num_labels
|
| 840 |
self.config = config
|
| 841 |
self.esm = FAST_ESM_ENCODER(config, add_pooling_layer=False)
|
|
@@ -845,13 +1009,19 @@ class FastEsmForSequenceClassification(FastEsmPreTrainedModel):
|
|
| 845 |
self.bce = nn.BCEWithLogitsLoss()
|
| 846 |
self.init_weights()
|
| 847 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 848 |
def forward(
|
| 849 |
self,
|
| 850 |
-
input_ids: Optional[torch.
|
| 851 |
attention_mask: Optional[torch.Tensor] = None,
|
| 852 |
-
position_ids: Optional[torch.
|
| 853 |
-
inputs_embeds: Optional[torch.
|
| 854 |
-
labels: Optional[torch.
|
| 855 |
output_attentions: Optional[bool] = None,
|
| 856 |
output_hidden_states: Optional[bool] = None,
|
| 857 |
return_dict: Optional[bool] = None
|
|
@@ -896,9 +1066,9 @@ class FastEsmForSequenceClassification(FastEsmPreTrainedModel):
|
|
| 896 |
)
|
| 897 |
|
| 898 |
|
| 899 |
-
class FastEsmForTokenClassification(FastEsmPreTrainedModel):
|
| 900 |
def __init__(self, config):
|
| 901 |
-
super().__init__(config)
|
| 902 |
self.num_labels = config.num_labels
|
| 903 |
self.esm = FAST_ESM_ENCODER(config, add_pooling_layer=False)
|
| 904 |
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
|
@@ -906,13 +1076,19 @@ class FastEsmForTokenClassification(FastEsmPreTrainedModel):
|
|
| 906 |
self.loss_fct = nn.CrossEntropyLoss()
|
| 907 |
self.init_weights()
|
| 908 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 909 |
def forward(
|
| 910 |
self,
|
| 911 |
-
input_ids: Optional[torch.
|
| 912 |
attention_mask: Optional[torch.Tensor] = None,
|
| 913 |
-
position_ids: Optional[torch.
|
| 914 |
-
inputs_embeds: Optional[torch.
|
| 915 |
-
labels: Optional[torch.
|
| 916 |
output_attentions: Optional[bool] = None,
|
| 917 |
output_hidden_states: Optional[bool] = None,
|
| 918 |
return_dict: Optional[bool] = None
|
|
@@ -972,7 +1148,11 @@ if __name__ == "__main__":
|
|
| 972 |
tokenizer = EsmTokenizer.from_pretrained(model_path)
|
| 973 |
config = FastEsmConfig.from_pretrained(model_path)
|
| 974 |
fast_model = FastEsmForMaskedLM(config).from_pretrained(model_path).to(device)
|
|
|
|
|
|
|
| 975 |
model = TransformersEsmModel.from_pretrained(model_path, token_dropout=False).to(device)
|
|
|
|
|
|
|
| 976 |
|
| 977 |
counts = [0] * len(tolerances)
|
| 978 |
for _ in range(seq_count):
|
|
|
|
| 1 |
import torch
|
| 2 |
import torch.nn as nn
|
| 3 |
+
import os
|
| 4 |
from torch.nn import functional as F
|
| 5 |
+
from torch.utils.data import Dataset as TorchDataset
|
| 6 |
+
from torch.utils.data import DataLoader as DataLoader
|
| 7 |
+
from typing import Optional, Tuple, Union, Callable, List, Dict, Any
|
| 8 |
from einops import rearrange
|
| 9 |
from dataclasses import dataclass
|
| 10 |
+
from transformers import PreTrainedModel, PretrainedConfig, EsmTokenizer, PreTrainedTokenizerBase
|
| 11 |
from transformers.modeling_outputs import (
|
| 12 |
ModelOutput,
|
| 13 |
BaseModelOutputWithPastAndCrossAttentions,
|
|
|
|
| 28 |
|
| 29 |
@dataclass
|
| 30 |
class EsmMaskedLMOutput(ModelOutput):
|
| 31 |
+
loss: Optional[torch.Tensor] = None
|
| 32 |
+
logits: Optional[torch.Tensor] = None
|
| 33 |
+
last_hidden_state: Optional[torch.Tensor] = None
|
| 34 |
+
hidden_states: Optional[Tuple[torch.Tensor, ...]] = None
|
| 35 |
+
attentions: Optional[Tuple[torch.Tensor, ...]] = None
|
| 36 |
|
| 37 |
|
| 38 |
class FastEsmConfig(PretrainedConfig):
|
| 39 |
model_type = "fast_esm"
|
| 40 |
def __init__(
|
| 41 |
self,
|
| 42 |
+
vocab_size: int = None,
|
| 43 |
+
mask_token_id: int = None,
|
| 44 |
+
pad_token_id: int = None,
|
| 45 |
+
hidden_size: int = 768,
|
| 46 |
+
num_hidden_layers: int = 12,
|
| 47 |
+
num_attention_heads: int = 12,
|
| 48 |
+
intermediate_size: int = 3072,
|
| 49 |
+
hidden_dropout_prob: float = 0.1,
|
| 50 |
+
attention_probs_dropout_prob: float = 0.1,
|
| 51 |
+
max_position_embeddings: int = 1026,
|
| 52 |
+
initializer_range: float = 0.02,
|
| 53 |
+
layer_norm_eps: float = 1e-12,
|
| 54 |
+
position_embedding_type: str = "absolute",
|
| 55 |
+
emb_layer_norm_before: bool = None,
|
| 56 |
**kwargs,
|
| 57 |
):
|
| 58 |
super().__init__(pad_token_id=pad_token_id, mask_token_id=mask_token_id, **kwargs)
|
|
|
|
| 70 |
self.position_embedding_type = position_embedding_type
|
| 71 |
self.emb_layer_norm_before = emb_layer_norm_before
|
| 72 |
|
| 73 |
+
def to_dict(self) -> Dict[str, Any]:
|
| 74 |
"""
|
| 75 |
Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
|
| 76 |
|
| 77 |
Returns:
|
| 78 |
+
`Dict[str, any]`: Dictionar y of all the attributes that make up this configuration instance,
|
| 79 |
"""
|
| 80 |
output = super().to_dict()
|
| 81 |
return output
|
| 82 |
|
| 83 |
|
| 84 |
+
def rotate_half(x: torch.Tensor) -> torch.Tensor:
|
| 85 |
x1, x2 = x.chunk(2, dim=-1)
|
| 86 |
return torch.cat((-x2, x1), dim=-1)
|
| 87 |
|
| 88 |
|
| 89 |
+
def apply_rotary_pos_emb(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor) -> torch.Tensor:
|
| 90 |
cos = cos[:, :, : x.shape[-2], :]
|
| 91 |
sin = sin[:, :, : x.shape[-2], :]
|
| 92 |
|
| 93 |
return (x * cos) + (rotate_half(x) * sin)
|
| 94 |
|
| 95 |
|
| 96 |
+
def symmetrize(x: torch.Tensor) -> torch.Tensor:
|
| 97 |
"Make layer symmetric in final two dimensions, used for contact prediction."
|
| 98 |
return x + x.transpose(-1, -2)
|
| 99 |
|
| 100 |
|
| 101 |
+
def average_product_correct(x: torch.Tensor) -> torch.Tensor:
|
| 102 |
"Perform average product correct, used for contact prediction."
|
| 103 |
a1 = x.sum(-1, keepdims=True)
|
| 104 |
a2 = x.sum(-2, keepdims=True)
|
|
|
|
| 116 |
def __init__(
|
| 117 |
self,
|
| 118 |
in_features: int,
|
| 119 |
+
bias: bool = True,
|
| 120 |
eos_idx: int = 2,
|
| 121 |
):
|
| 122 |
super().__init__()
|
| 123 |
self.in_features = in_features
|
| 124 |
self.eos_idx = eos_idx
|
| 125 |
+
self.regression = nn.Linear(in_features, 1, bias=bias)
|
| 126 |
self.activation = nn.Sigmoid()
|
| 127 |
|
| 128 |
+
def forward(self, input_ids: torch.Tensor, attentions: torch.Tensor) -> torch.Tensor:
|
| 129 |
# remove eos token attentions
|
| 130 |
+
eos_mask = input_ids.ne(self.eos_idx).to(attentions)
|
| 131 |
eos_mask = eos_mask.unsqueeze(1) * eos_mask.unsqueeze(2)
|
| 132 |
attentions = attentions * eos_mask[:, None, None, :, :]
|
| 133 |
attentions = attentions[..., :-1, :-1]
|
|
|
|
| 163 |
self._cos_cached = None
|
| 164 |
self._sin_cached = None
|
| 165 |
|
| 166 |
+
def _update_cos_sin_tables(self, x: torch.Tensor, seq_dimension: int = 2) -> Tuple[torch.Tensor, torch.Tensor]:
|
| 167 |
seq_len = x.shape[seq_dimension]
|
| 168 |
|
| 169 |
# Reset the tables if the sequence length has changed,
|
|
|
|
| 206 |
)
|
| 207 |
|
| 208 |
def forward(
|
| 209 |
+
self,
|
| 210 |
+
input_ids: Optional[torch.Tensor] = None,
|
| 211 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 212 |
+
position_ids: Optional[torch.Tensor] = None,
|
| 213 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
| 214 |
+
past_key_values_length: Optional[int] = 0,
|
| 215 |
):
|
| 216 |
if inputs_embeds is None:
|
| 217 |
inputs_embeds = self.word_embeddings(input_ids)
|
|
|
|
| 243 |
|
| 244 |
|
| 245 |
class EsmSelfAttention(nn.Module):
|
| 246 |
+
def __init__(self, config, position_embedding_type: Optional[str] = None):
|
| 247 |
super().__init__()
|
| 248 |
if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
|
| 249 |
raise ValueError(
|
|
|
|
| 274 |
def forward(
|
| 275 |
self,
|
| 276 |
hidden_states: torch.Tensor,
|
| 277 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 278 |
+
output_attentions: Optional[bool] = False,
|
| 279 |
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
| 280 |
"""Forward pass for self attention.
|
| 281 |
|
|
|
|
| 328 |
def forward(
|
| 329 |
self,
|
| 330 |
hidden_states: torch.Tensor,
|
| 331 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 332 |
+
output_attentions: Optional[bool] = False,
|
| 333 |
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
| 334 |
"""Forward pass for attention layer.
|
| 335 |
|
|
|
|
| 369 |
def forward(
|
| 370 |
self,
|
| 371 |
hidden_states: torch.Tensor,
|
| 372 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 373 |
+
output_attentions: Optional[bool] = False,
|
| 374 |
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
| 375 |
"""Forward pass for transformer layer.
|
| 376 |
|
|
|
|
| 417 |
def forward(
|
| 418 |
self,
|
| 419 |
hidden_states: torch.Tensor,
|
| 420 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 421 |
+
output_hidden_states: Optional[bool] = False,
|
| 422 |
+
output_attentions: Optional[bool] = False,
|
| 423 |
) -> BaseModelOutputWithPastAndCrossAttentions:
|
| 424 |
"""Forward pass for transformer encoder.
|
| 425 |
|
|
|
|
| 472 |
)
|
| 473 |
|
| 474 |
|
| 475 |
+
### Support for embedding datasets with low code
|
| 476 |
+
class Pooler:
|
| 477 |
+
def __init__(self, pooling_types: List[str]):
|
| 478 |
+
self.pooling_types = pooling_types
|
| 479 |
+
self.pooling_options = {
|
| 480 |
+
'mean': self.mean_pooling,
|
| 481 |
+
'max': self.max_pooling,
|
| 482 |
+
'min': self.min_pooling,
|
| 483 |
+
'norm': self.norm_pooling,
|
| 484 |
+
'prod': self.prod_pooling,
|
| 485 |
+
'median': self.median_pooling,
|
| 486 |
+
'std': self.std_pooling,
|
| 487 |
+
'var': self.var_pooling,
|
| 488 |
+
'cls': self.cls_pooling,
|
| 489 |
+
}
|
| 490 |
+
|
| 491 |
+
def mean_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None): # (b, L, d) -> (b, d)
|
| 492 |
+
if attention_mask is None:
|
| 493 |
+
return emb.mean(dim=1)
|
| 494 |
+
else:
|
| 495 |
+
attention_mask = attention_mask.unsqueeze(-1)
|
| 496 |
+
return (emb * attention_mask).sum(dim=1) / attention_mask.sum(dim=1)
|
| 497 |
+
|
| 498 |
+
def max_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None): # (b, L, d) -> (b, d)
|
| 499 |
+
if attention_mask is None:
|
| 500 |
+
return emb.max(dim=1).values
|
| 501 |
+
else:
|
| 502 |
+
attention_mask = attention_mask.unsqueeze(-1)
|
| 503 |
+
return (emb * attention_mask).max(dim=1).values
|
| 504 |
+
|
| 505 |
+
def min_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None): # (b, L, d) -> (b, d)
|
| 506 |
+
if attention_mask is None:
|
| 507 |
+
return emb.min(dim=1).values
|
| 508 |
+
else:
|
| 509 |
+
attention_mask = attention_mask.unsqueeze(-1)
|
| 510 |
+
return (emb * attention_mask).min(dim=1).values
|
| 511 |
+
|
| 512 |
+
def norm_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None): # (b, L, d) -> (b, d)
|
| 513 |
+
if attention_mask is None:
|
| 514 |
+
return emb.norm(dim=1, p=2)
|
| 515 |
+
else:
|
| 516 |
+
attention_mask = attention_mask.unsqueeze(-1)
|
| 517 |
+
return (emb * attention_mask).norm(dim=1, p=2)
|
| 518 |
+
|
| 519 |
+
def prod_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None): # (b, L, d) -> (b, d)
|
| 520 |
+
length = emb.shape[1]
|
| 521 |
+
if attention_mask is None:
|
| 522 |
+
return emb.prod(dim=1) / length
|
| 523 |
+
else:
|
| 524 |
+
attention_mask = attention_mask.unsqueeze(-1)
|
| 525 |
+
return ((emb * attention_mask).prod(dim=1) / attention_mask.sum(dim=1)) / length
|
| 526 |
+
|
| 527 |
+
def median_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None): # (b, L, d) -> (b, d)
|
| 528 |
+
if attention_mask is None:
|
| 529 |
+
return emb.median(dim=1).values
|
| 530 |
+
else:
|
| 531 |
+
attention_mask = attention_mask.unsqueeze(-1)
|
| 532 |
+
return (emb * attention_mask).median(dim=1).values
|
| 533 |
+
|
| 534 |
+
def std_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None): # (b, L, d) -> (b, d)
|
| 535 |
+
if attention_mask is None:
|
| 536 |
+
return emb.std(dim=1)
|
| 537 |
+
else:
|
| 538 |
+
attention_mask = attention_mask.unsqueeze(-1)
|
| 539 |
+
return (emb * attention_mask).std(dim=1)
|
| 540 |
+
|
| 541 |
+
def var_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None): # (b, L, d) -> (b, d)
|
| 542 |
+
if attention_mask is None:
|
| 543 |
+
return emb.var(dim=1)
|
| 544 |
+
else:
|
| 545 |
+
attention_mask = attention_mask.unsqueeze(-1)
|
| 546 |
+
return (emb * attention_mask).var(dim=1)
|
| 547 |
+
|
| 548 |
+
def cls_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None): # (b, L, d) -> (b, d)
|
| 549 |
+
return emb[:, 0, :]
|
| 550 |
+
|
| 551 |
+
def __call__(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None): # [mean, max]
|
| 552 |
+
final_emb = []
|
| 553 |
+
for pooling_type in self.pooling_types:
|
| 554 |
+
final_emb.append(self.pooling_options[pooling_type](emb, attention_mask)) # (b, d)
|
| 555 |
+
return torch.cat(final_emb, dim=-1) # (b, n_pooling_types * d)
|
| 556 |
+
|
| 557 |
+
|
| 558 |
+
class ProteinDataset(TorchDataset):
|
| 559 |
"""Simple dataset for protein sequences."""
|
| 560 |
def __init__(self, sequences: list[str]):
|
| 561 |
self.sequences = sequences
|
|
|
|
| 567 |
return self.sequences[idx]
|
| 568 |
|
| 569 |
|
| 570 |
+
def build_collator(tokenizer) -> Callable[[list[str]], tuple[torch.Tensor, torch.Tensor]]:
|
| 571 |
+
def _collate_fn(sequences: list[str]) -> tuple[torch.Tensor, torch.Tensor]:
|
| 572 |
+
"""Collate function for batching sequences."""
|
| 573 |
+
return tokenizer(sequences, return_tensors="pt", padding='longest', pad_to_multiple_of=8)
|
| 574 |
+
return _collate_fn
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 575 |
|
| 576 |
+
|
| 577 |
+
class EmbeddingMixin:
|
| 578 |
+
def _embed(self, input_ids: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
|
| 579 |
+
raise NotImplementedError
|
|
|
|
| 580 |
|
| 581 |
@property
|
| 582 |
def device(self) -> torch.device:
|
| 583 |
"""Get the device of the model."""
|
| 584 |
return next(self.parameters()).device
|
| 585 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 586 |
def _read_sequences_from_db(self, db_path: str) -> set[str]:
|
| 587 |
"""Read sequences from SQLite database."""
|
| 588 |
import sqlite3
|
|
|
|
| 599 |
|
| 600 |
def embed_dataset(
|
| 601 |
self,
|
| 602 |
+
sequences: List[str],
|
| 603 |
+
tokenizer: PreTrainedTokenizerBase,
|
| 604 |
batch_size: int = 2,
|
| 605 |
max_len: int = 512,
|
| 606 |
full_embeddings: bool = False,
|
| 607 |
+
embed_dtype: torch.dtype = torch.float32,
|
| 608 |
+
pooling_types: List[str] = ['mean'],
|
| 609 |
num_workers: int = 0,
|
| 610 |
sql: bool = False,
|
| 611 |
+
save: bool = True,
|
| 612 |
sql_db_path: str = 'embeddings.db',
|
| 613 |
+
save_path: str = 'embeddings.pth',
|
| 614 |
) -> Optional[dict[str, torch.Tensor]]:
|
| 615 |
"""Embed a dataset of protein sequences.
|
| 616 |
|
|
|
|
| 619 |
batch_size: Batch size for processing
|
| 620 |
max_len: Maximum sequence length
|
| 621 |
full_embeddings: Whether to return full residue-wise (True) embeddings or pooled (False)
|
|
|
|
| 622 |
pooling_type: Type of pooling ('mean' or 'cls')
|
| 623 |
num_workers: Number of workers for data loading, 0 for the main process
|
| 624 |
sql: Whether to store embeddings in SQLite database - will be stored in float32
|
|
|
|
| 626 |
|
| 627 |
Returns:
|
| 628 |
Dictionary mapping sequences to embeddings, or None if sql=True
|
| 629 |
+
|
| 630 |
+
Note:
|
| 631 |
+
- If sql=True, embeddings can only be stored in float32
|
| 632 |
+
- sql is ideal if you need to stream a very large dataset for training in real-time
|
| 633 |
+
- save=True is ideal if you can store the entire embedding dictionary in RAM
|
| 634 |
+
- sql will be used if it is True and save is True or False
|
| 635 |
+
- If your sql database or .pth file is already present, they will be scanned first for already embedded sequences
|
| 636 |
+
- Sequences will be truncated to max_len and sorted by length in descending order for faster processing
|
| 637 |
+
|
| 638 |
+
Example:
|
| 639 |
+
>>> embedder = EmbeddingMixin()
|
| 640 |
+
>>> embedding_dict = embedder.embed_dataset(
|
| 641 |
+
sequences=[
|
| 642 |
+
'MALWMRLLPLLALLALWGPDPAAA', ... # list of protein sequences
|
| 643 |
+
],
|
| 644 |
+
batch_size=2, # adjust for your GPU memory
|
| 645 |
+
max_len=512, # adjust for your needs
|
| 646 |
+
full_embeddings=False, # if True, no pooling is performed
|
| 647 |
+
embed_dtype=torch.float32, # cast to what dtype you want
|
| 648 |
+
pooling_type=['mean', 'cls'], # more than one pooling type will be concatenated together
|
| 649 |
+
num_workers=0, # if you have many cpu cores, we find that num_workers = 4 is fast for large datasets
|
| 650 |
+
sql=False, # if True, embeddings will be stored in SQLite database
|
| 651 |
+
sql_db_path='embeddings.db',
|
| 652 |
+
save=True, # if True, embeddings will be saved as a .pth file
|
| 653 |
+
save_path='embeddings.pth',
|
| 654 |
+
)
|
| 655 |
+
>>> # embedding_dict is a dictionary mapping sequences to their embeddings as tensors for .pth or numpy arrays for sql
|
| 656 |
"""
|
| 657 |
+
sequences = list(set([seq[:max_len] for seq in sequences]))
|
| 658 |
+
sequences = sorted(sequences, key=len, reverse=True)
|
| 659 |
+
collate_fn = build_collator(tokenizer)
|
| 660 |
device = self.device
|
| 661 |
+
pooler = Pooler(pooling_types) if not full_embeddings else None
|
| 662 |
|
| 663 |
def get_embeddings(residue_embeddings: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
|
| 664 |
+
if full_embeddings or residue_embeddings.ndim == 2: # if already pooled or want residue-wise embeddings
|
| 665 |
return residue_embeddings
|
|
|
|
|
|
|
| 666 |
else:
|
| 667 |
+
return pooler(residue_embeddings, attention_mask)
|
| 668 |
|
|
|
|
| 669 |
if sql:
|
| 670 |
import sqlite3
|
| 671 |
conn = sqlite3.connect(sql_db_path)
|
|
|
|
| 676 |
print(f"Found {len(already_embedded)} already embedded sequences in {sql_db_path}")
|
| 677 |
print(f"Embedding {len(to_embed)} new sequences")
|
| 678 |
if len(to_embed) > 0:
|
|
|
|
| 679 |
dataset = ProteinDataset(to_embed)
|
| 680 |
+
dataloader = DataLoader(dataset, batch_size=batch_size, num_workers=num_workers, collate_fn=collate_fn, shuffle=False)
|
| 681 |
with torch.no_grad():
|
| 682 |
for i, batch in tqdm(enumerate(dataloader), total=len(dataloader), desc='Embedding batches'):
|
| 683 |
+
seqs = to_embed[i * batch_size:(i + 1) * batch_size]
|
| 684 |
input_ids, attention_mask = batch['input_ids'].to(device), batch['attention_mask'].to(device)
|
| 685 |
+
residue_embeddings = self._embed(input_ids, attention_mask).float() # sql requires float32
|
| 686 |
embeddings = get_embeddings(residue_embeddings, attention_mask).cpu()
|
| 687 |
+
for seq, emb, mask in zip(seqs, embeddings, attention_mask):
|
| 688 |
+
if full_embeddings:
|
| 689 |
+
emb = emb[mask.bool()]
|
| 690 |
c.execute("INSERT OR REPLACE INTO embeddings VALUES (?, ?)",
|
| 691 |
(seq, emb.cpu().numpy().tobytes()))
|
| 692 |
|
| 693 |
if (i + 1) % 100 == 0:
|
| 694 |
conn.commit()
|
| 695 |
+
|
| 696 |
conn.commit()
|
| 697 |
conn.close()
|
| 698 |
return None
|
| 699 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
| 700 |
embeddings_dict = {}
|
| 701 |
+
if os.path.exists(save_path):
|
| 702 |
+
embeddings_dict = torch.load(save_path, map_location='cpu', weights_only=True)
|
| 703 |
+
to_embed = [seq for seq in sequences if seq not in embeddings_dict]
|
| 704 |
+
print(f"Found {len(embeddings_dict)} already embedded sequences in {save_path}")
|
| 705 |
+
print(f"Embedding {len(to_embed)} new sequences")
|
| 706 |
+
else:
|
| 707 |
+
to_embed = sequences
|
| 708 |
+
print(f"Embedding {len(to_embed)} new sequences")
|
| 709 |
+
|
| 710 |
+
if len(to_embed) > 0:
|
| 711 |
+
dataset = ProteinDataset(to_embed)
|
| 712 |
+
dataloader = DataLoader(dataset, batch_size=batch_size, num_workers=num_workers, collate_fn=collate_fn, shuffle=False)
|
| 713 |
+
with torch.no_grad():
|
| 714 |
+
for i, batch in tqdm(enumerate(dataloader), total=len(dataloader), desc='Embedding batches'):
|
| 715 |
+
seqs = to_embed[i * batch_size:(i + 1) * batch_size]
|
| 716 |
+
input_ids, attention_mask = batch['input_ids'].to(device), batch['attention_mask'].to(device)
|
| 717 |
+
residue_embeddings = self._embed(input_ids, attention_mask)
|
| 718 |
+
embeddings = get_embeddings(residue_embeddings, attention_mask).to(embed_dtype).cpu()
|
| 719 |
+
for seq, emb in zip(seqs, embeddings):
|
| 720 |
+
embeddings_dict[seq] = emb
|
| 721 |
+
|
| 722 |
+
if save:
|
| 723 |
+
torch.save(embeddings_dict, save_path)
|
| 724 |
+
|
| 725 |
return embeddings_dict
|
| 726 |
|
| 727 |
|
| 728 |
+
class FastEsmPreTrainedModel(PreTrainedModel):
|
| 729 |
+
"""
|
| 730 |
+
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
|
| 731 |
+
models.
|
| 732 |
+
"""
|
| 733 |
+
config_class = FastEsmConfig
|
| 734 |
+
base_model_prefix = "fastesm"
|
| 735 |
+
supports_gradient_checkpointing = True
|
| 736 |
+
tokenizer = EsmTokenizer.from_pretrained("facebook/esm2_t6_8M_UR50D")
|
| 737 |
+
def _init_weights(self, module):
|
| 738 |
+
"""Initialize the weights"""
|
| 739 |
+
if isinstance(module, nn.Linear):
|
| 740 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
| 741 |
+
if module.bias is not None:
|
| 742 |
+
module.bias.data.zero_()
|
| 743 |
+
elif isinstance(module, nn.Embedding):
|
| 744 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
| 745 |
+
if module.padding_idx is not None:
|
| 746 |
+
module.weight.data[module.padding_idx].zero_()
|
| 747 |
+
elif isinstance(module, nn.LayerNorm):
|
| 748 |
+
module.bias.data.zero_()
|
| 749 |
+
module.weight.data.fill_(1.0)
|
| 750 |
+
|
| 751 |
+
def get_input_embeddings(self) -> nn.Module:
|
| 752 |
+
try:
|
| 753 |
+
return self.embeddings.word_embeddings
|
| 754 |
+
except AttributeError:
|
| 755 |
+
return self.esm.embeddings.word_embeddings
|
| 756 |
+
|
| 757 |
+
|
| 758 |
+
class FAST_ESM_ENCODER(FastEsmPreTrainedModel, EmbeddingMixin):
|
| 759 |
+
def __init__(self, config, add_pooling_layer: Optional[bool] = True):
|
| 760 |
+
super(FastEsmPreTrainedModel, self).__init__(config)
|
| 761 |
self.config = config
|
| 762 |
self.embeddings = EsmEmbeddings(config)
|
| 763 |
self.encoder = EsmEncoder(config)
|
| 764 |
+
self.contact_head = EsmContactPredictionHead(
|
| 765 |
+
in_features=config.num_hidden_layers * config.num_attention_heads, bias=True
|
| 766 |
+
)
|
| 767 |
# Initialize weights and apply final processing
|
| 768 |
self.post_init()
|
| 769 |
|
|
|
|
| 773 |
def set_input_embeddings(self, value):
|
| 774 |
self.embeddings.word_embeddings = value
|
| 775 |
|
| 776 |
+
def _embed(self, input_ids: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
|
| 777 |
+
token_embedding_output = self.embeddings(input_ids, attention_mask=attention_mask)
|
| 778 |
+
batch_size, seq_length = input_ids.shape
|
| 779 |
+
if attention_mask is not None:
|
| 780 |
+
extended_attention_mask = attention_mask[:, None, None, :].expand(
|
| 781 |
+
batch_size, 1, seq_length, seq_length
|
| 782 |
+
).bool()
|
| 783 |
+
else:
|
| 784 |
+
extended_attention_mask = None
|
| 785 |
+
encoder_outputs = self.encoder(
|
| 786 |
+
token_embedding_output,
|
| 787 |
+
attention_mask=extended_attention_mask,
|
| 788 |
+
output_hidden_states=False,
|
| 789 |
+
output_attentions=False,
|
| 790 |
+
)
|
| 791 |
+
return encoder_outputs.last_hidden_state
|
| 792 |
+
|
| 793 |
+
def predict_contacts(self, input_ids: torch.Tensor, attention_mask: torch.Tensor) -> torch.Tensor:
|
| 794 |
+
attns = self(input_ids, attention_mask=attention_mask, output_attentions=True).attentions
|
| 795 |
+
attns = torch.stack(attns, dim=1)
|
| 796 |
+
attns *= attention_mask.unsqueeze(1).unsqueeze(2).unsqueeze(3)
|
| 797 |
+
attns *= attention_mask.unsqueeze(1).unsqueeze(2).unsqueeze(4)
|
| 798 |
+
return self.contact_head(input_ids, attns)
|
| 799 |
+
|
| 800 |
def forward(
|
| 801 |
self,
|
| 802 |
+
input_ids: Optional[torch.Tensor] = None,
|
| 803 |
attention_mask: Optional[torch.Tensor] = None,
|
| 804 |
+
position_ids: Optional[torch.Tensor] = None,
|
| 805 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
| 806 |
output_attentions: Optional[bool] = None,
|
| 807 |
output_hidden_states: Optional[bool] = None,
|
| 808 |
return_dict: Optional[bool] = None, # to play nice with HF adjacent packages
|
|
|
|
| 834 |
raise ValueError("You have to specify either input_ids or inputs_embeds")
|
| 835 |
|
| 836 |
batch_size, seq_length = input_shape
|
| 837 |
+
token_embedding_output = self.embeddings(
|
| 838 |
input_ids=input_ids,
|
| 839 |
position_ids=position_ids,
|
| 840 |
attention_mask=attention_mask,
|
|
|
|
| 849 |
extended_attention_mask = None
|
| 850 |
|
| 851 |
encoder_outputs = self.encoder(
|
| 852 |
+
token_embedding_output,
|
| 853 |
attention_mask=extended_attention_mask,
|
| 854 |
output_hidden_states=output_hidden_states,
|
| 855 |
output_attentions=output_attentions,
|
|
|
|
| 863 |
)
|
| 864 |
|
| 865 |
|
| 866 |
+
class FastEsmModel(FastEsmPreTrainedModel, EmbeddingMixin):
|
| 867 |
+
def __init__(self, config, add_pooling_layer: Optional[bool] = True):
|
| 868 |
+
super(FastEsmPreTrainedModel, self).__init__(config)
|
| 869 |
self.config = config
|
| 870 |
self.esm = FAST_ESM_ENCODER(config)
|
| 871 |
self.pooler = EsmPooler(config) if add_pooling_layer else None
|
|
|
|
| 878 |
def set_input_embeddings(self, value):
|
| 879 |
self.embeddings.word_embeddings = value
|
| 880 |
|
| 881 |
+
def _embed(self, input_ids: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
|
| 882 |
+
return self.esm._embed(input_ids, attention_mask)
|
| 883 |
+
|
| 884 |
+
def predict_contacts(self, input_ids: torch.Tensor, attention_mask: torch.Tensor) -> torch.Tensor:
|
| 885 |
+
return self.esm.predict_contacts(input_ids, attention_mask=attention_mask)
|
| 886 |
+
|
| 887 |
def forward(
|
| 888 |
self,
|
| 889 |
+
input_ids: Optional[torch.Tensor] = None,
|
| 890 |
attention_mask: Optional[torch.Tensor] = None,
|
| 891 |
+
position_ids: Optional[torch.Tensor] = None,
|
| 892 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
| 893 |
output_attentions: Optional[bool] = None,
|
| 894 |
output_hidden_states: Optional[bool] = None,
|
| 895 |
return_dict: Optional[bool] = None, # to play nice with HF adjacent packages
|
|
|
|
| 939 |
)
|
| 940 |
|
| 941 |
|
| 942 |
+
class FastEsmForMaskedLM(FastEsmPreTrainedModel, EmbeddingMixin):
|
| 943 |
_tied_weights_keys = ["lm_head.decoder.weight"]
|
| 944 |
|
| 945 |
def __init__(self, config):
|
| 946 |
+
super(FastEsmPreTrainedModel, self).__init__(config)
|
| 947 |
self.esm = FAST_ESM_ENCODER(config, add_pooling_layer=False)
|
| 948 |
self.lm_head = EsmLMHead(config)
|
| 949 |
self.loss_fct = nn.CrossEntropyLoss()
|
|
|
|
| 955 |
def set_output_embeddings(self, new_embeddings):
|
| 956 |
self.lm_head.decoder = new_embeddings
|
| 957 |
|
| 958 |
+
def _embed(self, input_ids: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
|
| 959 |
+
return self.esm._embed(input_ids, attention_mask)
|
| 960 |
+
|
| 961 |
+
def predict_contacts(self, input_ids: torch.Tensor, attention_mask: torch.Tensor) -> torch.Tensor:
|
| 962 |
+
return self.esm.predict_contacts(input_ids, attention_mask=attention_mask)
|
| 963 |
+
|
| 964 |
def forward(
|
| 965 |
self,
|
| 966 |
+
input_ids: Optional[torch.Tensor] = None,
|
| 967 |
attention_mask: Optional[torch.Tensor] = None,
|
| 968 |
+
position_ids: Optional[torch.Tensor] = None,
|
| 969 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
| 970 |
+
labels: Optional[torch.Tensor] = None,
|
| 971 |
output_attentions: Optional[bool] = None,
|
| 972 |
output_hidden_states: Optional[bool] = None,
|
| 973 |
return_dict: Optional[bool] = None, # to play nice with HF adjacent packages
|
|
|
|
| 996 |
attentions=outputs.attentions,
|
| 997 |
)
|
| 998 |
|
|
|
|
|
|
|
|
|
|
| 999 |
|
| 1000 |
+
class FastEsmForSequenceClassification(FastEsmPreTrainedModel, EmbeddingMixin):
|
| 1001 |
def __init__(self, config):
|
| 1002 |
+
super(FastEsmPreTrainedModel, self).__init__(config)
|
| 1003 |
self.num_labels = config.num_labels
|
| 1004 |
self.config = config
|
| 1005 |
self.esm = FAST_ESM_ENCODER(config, add_pooling_layer=False)
|
|
|
|
| 1009 |
self.bce = nn.BCEWithLogitsLoss()
|
| 1010 |
self.init_weights()
|
| 1011 |
|
| 1012 |
+
def _embed(self, input_ids: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
|
| 1013 |
+
return self.esm._embed(input_ids, attention_mask)
|
| 1014 |
+
|
| 1015 |
+
def predict_contacts(self, input_ids: torch.Tensor, attention_mask: torch.Tensor) -> torch.Tensor:
|
| 1016 |
+
return self.esm.predict_contacts(input_ids, attention_mask=attention_mask)
|
| 1017 |
+
|
| 1018 |
def forward(
|
| 1019 |
self,
|
| 1020 |
+
input_ids: Optional[torch.Tensor] = None,
|
| 1021 |
attention_mask: Optional[torch.Tensor] = None,
|
| 1022 |
+
position_ids: Optional[torch.Tensor] = None,
|
| 1023 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
| 1024 |
+
labels: Optional[torch.Tensor] = None,
|
| 1025 |
output_attentions: Optional[bool] = None,
|
| 1026 |
output_hidden_states: Optional[bool] = None,
|
| 1027 |
return_dict: Optional[bool] = None
|
|
|
|
| 1066 |
)
|
| 1067 |
|
| 1068 |
|
| 1069 |
+
class FastEsmForTokenClassification(FastEsmPreTrainedModel, EmbeddingMixin):
|
| 1070 |
def __init__(self, config):
|
| 1071 |
+
super(FastEsmPreTrainedModel, self).__init__(config)
|
| 1072 |
self.num_labels = config.num_labels
|
| 1073 |
self.esm = FAST_ESM_ENCODER(config, add_pooling_layer=False)
|
| 1074 |
self.dropout = nn.Dropout(config.hidden_dropout_prob)
|
|
|
|
| 1076 |
self.loss_fct = nn.CrossEntropyLoss()
|
| 1077 |
self.init_weights()
|
| 1078 |
|
| 1079 |
+
def _embed(self, input_ids: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
|
| 1080 |
+
return self.esm._embed(input_ids, attention_mask)
|
| 1081 |
+
|
| 1082 |
+
def predict_contacts(self, input_ids: torch.Tensor, attention_mask: torch.Tensor) -> torch.Tensor:
|
| 1083 |
+
return self.esm.predict_contacts(input_ids, attention_mask=attention_mask)
|
| 1084 |
+
|
| 1085 |
def forward(
|
| 1086 |
self,
|
| 1087 |
+
input_ids: Optional[torch.Tensor] = None,
|
| 1088 |
attention_mask: Optional[torch.Tensor] = None,
|
| 1089 |
+
position_ids: Optional[torch.Tensor] = None,
|
| 1090 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
| 1091 |
+
labels: Optional[torch.Tensor] = None,
|
| 1092 |
output_attentions: Optional[bool] = None,
|
| 1093 |
output_hidden_states: Optional[bool] = None,
|
| 1094 |
return_dict: Optional[bool] = None
|
|
|
|
| 1148 |
tokenizer = EsmTokenizer.from_pretrained(model_path)
|
| 1149 |
config = FastEsmConfig.from_pretrained(model_path)
|
| 1150 |
fast_model = FastEsmForMaskedLM(config).from_pretrained(model_path).to(device)
|
| 1151 |
+
print('fast model')
|
| 1152 |
+
print(fast_model)
|
| 1153 |
model = TransformersEsmModel.from_pretrained(model_path, token_dropout=False).to(device)
|
| 1154 |
+
print('transformers model')
|
| 1155 |
+
print(model)
|
| 1156 |
|
| 1157 |
counts = [0] * len(tolerances)
|
| 1158 |
for _ in range(seq_count):
|