File size: 63,060 Bytes
b233db2 cdc6f1b fd7a6b7 cdc6f1b b233db2 7bb2dfa b233db2 7bb2dfa b233db2 7bb2dfa b233db2 7bb2dfa b233db2 f39adad b233db2 f39adad b233db2 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 | ### Embedding Mixin + Pooler
import os
import sqlite3
import networkx as nx
import numpy as np
import torch
from tqdm.auto import tqdm
from typing import Callable, List, Optional
from torch.utils.data import DataLoader
from torch.utils.data import Dataset as TorchDataset
from transformers import PreTrainedTokenizerBase
class Pooler:
def __init__(self, pooling_types: List[str]):
self.pooling_types = pooling_types
self.pooling_options = {
'mean': self.mean_pooling,
'max': self.max_pooling,
'norm': self.norm_pooling,
'median': self.median_pooling,
'std': self.std_pooling,
'var': self.var_pooling,
'cls': self.cls_pooling,
'parti': self._pool_parti,
}
def _create_pooled_matrices_across_layers(self, attentions: torch.Tensor) -> torch.Tensor:
maxed_attentions = torch.max(attentions, dim=1)[0]
return maxed_attentions
def _page_rank(self, attention_matrix, personalization=None, nstart=None, prune_type="top_k_outdegree"):
# Run PageRank on the attention matrix converted to a graph.
# Raises exceptions if the graph doesn't match the token sequence or has no edges.
# Returns the PageRank scores for each token node.
G = self._convert_to_graph(attention_matrix)
if G.number_of_nodes() != attention_matrix.shape[0]:
raise Exception(
f"The number of nodes in the graph should be equal to the number of tokens in sequence! You have {G.number_of_nodes()} nodes for {attention_matrix.shape[0]} tokens.")
if G.number_of_edges() == 0:
raise Exception(f"You don't seem to have any attention edges left in the graph.")
return nx.pagerank(G, alpha=0.85, tol=1e-06, weight='weight', personalization=personalization, nstart=nstart, max_iter=100)
def _convert_to_graph(self, matrix):
# Convert a matrix (e.g., attention scores) to a directed graph using networkx.
# Each element in the matrix represents a directed edge with a weight.
G = nx.from_numpy_array(matrix, create_using=nx.DiGraph)
return G
def _calculate_importance_weights(self, dict_importance, attention_mask: Optional[torch.Tensor] = None):
# Remove keys where attention_mask is 0
if attention_mask is not None:
for k in list(dict_importance.keys()):
if attention_mask[k] == 0:
del dict_importance[k]
#dict_importance[0] # remove cls
#dict_importance[-1] # remove eos
total = sum(dict_importance.values())
return np.array([v / total for _, v in dict_importance.items()])
def _pool_parti(self, emb: torch.Tensor, attentions: torch.Tensor, attention_mask: Optional[torch.Tensor] = None): # (b, L, d) -> (b, d)
maxed_attentions = self._create_pooled_matrices_across_layers(attentions).numpy()
# emb is (b, L, d), maxed_attentions is (b, L, L)
emb_pooled = []
for e, a, mask in zip(emb, maxed_attentions, attention_mask):
dict_importance = self._page_rank(a)
importance_weights = self._calculate_importance_weights(dict_importance, mask)
num_tokens = int(mask.sum().item())
emb_pooled.append(np.average(e[:num_tokens], weights=importance_weights, axis=0))
pooled = torch.tensor(np.array(emb_pooled))
return pooled
def mean_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, **kwargs): # (b, L, d) -> (b, d)
if attention_mask is None:
return emb.mean(dim=1)
else:
attention_mask = attention_mask.unsqueeze(-1)
return (emb * attention_mask).sum(dim=1) / attention_mask.sum(dim=1)
def max_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, **kwargs): # (b, L, d) -> (b, d)
if attention_mask is None:
return emb.max(dim=1).values
else:
attention_mask = attention_mask.unsqueeze(-1)
return (emb * attention_mask).max(dim=1).values
def norm_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, **kwargs): # (b, L, d) -> (b, d)
if attention_mask is None:
return emb.norm(dim=1, p=2)
else:
attention_mask = attention_mask.unsqueeze(-1)
return (emb * attention_mask).norm(dim=1, p=2)
def median_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, **kwargs): # (b, L, d) -> (b, d)
if attention_mask is None:
return emb.median(dim=1).values
else:
attention_mask = attention_mask.unsqueeze(-1)
return (emb * attention_mask).median(dim=1).values
def std_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, **kwargs): # (b, L, d) -> (b, d)
if attention_mask is None:
return emb.std(dim=1)
else:
# Compute variance correctly over non-masked positions, then take sqrt
var = self.var_pooling(emb, attention_mask, **kwargs)
return torch.sqrt(var)
def var_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, **kwargs): # (b, L, d) -> (b, d)
if attention_mask is None:
return emb.var(dim=1)
else:
# Correctly compute variance over only non-masked positions
attention_mask = attention_mask.unsqueeze(-1) # (b, L, 1)
# Compute mean over non-masked positions
mean = (emb * attention_mask).sum(dim=1) / attention_mask.sum(dim=1) # (b, d)
mean = mean.unsqueeze(1) # (b, 1, d)
# Compute squared differences from mean, only over non-masked positions
squared_diff = (emb - mean) ** 2 # (b, L, d)
# Sum squared differences over non-masked positions and divide by count
var = (squared_diff * attention_mask).sum(dim=1) / attention_mask.sum(dim=1) # (b, d)
return var
def cls_pooling(self, emb: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, **kwargs): # (b, L, d) -> (b, d)
return emb[:, 0, :]
def __call__(
self,
emb: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
attentions: Optional[torch.Tensor] = None
): # [mean, max]
final_emb = []
for pooling_type in self.pooling_types:
final_emb.append(self.pooling_options[pooling_type](emb=emb, attention_mask=attention_mask, attentions=attentions)) # (b, d)
return torch.cat(final_emb, dim=-1) # (b, n_pooling_types * d)
class ProteinDataset(TorchDataset):
"""Simple dataset for protein sequences."""
def __init__(self, sequences: list[str]):
self.sequences = sequences
def __len__(self) -> int:
return len(self.sequences)
def __getitem__(self, idx: int) -> str:
return self.sequences[idx]
def build_collator(tokenizer: PreTrainedTokenizerBase) -> Callable[[list[str]], dict[str, torch.Tensor]]:
def _collate_fn(sequences: list[str]) -> dict[str, torch.Tensor]:
return tokenizer(sequences, return_tensors="pt", padding='longest')
return _collate_fn
class EmbeddingMixin:
def _embed(self, input_ids: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
raise NotImplementedError
@property
def device(self) -> torch.device:
"""Get the device of the model."""
return next(self.parameters()).device
def _read_sequences_from_db(self, db_path: str) -> set[str]:
"""Read sequences from SQLite database."""
sequences = []
with sqlite3.connect(db_path) as conn:
c = conn.cursor()
c.execute("SELECT sequence FROM embeddings")
while True:
row = c.fetchone()
if row is None:
break
sequences.append(row[0])
return set(sequences)
def _ensure_embeddings_table(self, conn: sqlite3.Connection) -> None:
cursor = conn.cursor()
cursor.execute(
"CREATE TABLE IF NOT EXISTS embeddings ("
"sequence TEXT PRIMARY KEY, "
"embedding BLOB NOT NULL, "
"shape TEXT, "
"dtype TEXT"
")"
)
cursor.execute("PRAGMA table_info(embeddings)")
rows = cursor.fetchall()
column_names = [row[1] for row in rows]
if "shape" not in column_names:
cursor.execute("ALTER TABLE embeddings ADD COLUMN shape TEXT")
if "dtype" not in column_names:
cursor.execute("ALTER TABLE embeddings ADD COLUMN dtype TEXT")
conn.commit()
def load_embeddings_from_pth(self, save_path: str) -> dict[str, torch.Tensor]:
assert os.path.exists(save_path), f"Embedding file does not exist: {save_path}"
payload = torch.load(save_path, map_location="cpu", weights_only=True)
assert isinstance(payload, dict), "Expected .pth embeddings file to contain a dictionary."
for sequence, tensor in payload.items():
assert isinstance(sequence, str), "Expected embedding dictionary keys to be sequences (str)."
assert isinstance(tensor, torch.Tensor), "Expected embedding dictionary values to be tensors."
return payload
def load_embeddings_from_db(self, db_path: str, sequences: Optional[List[str]] = None) -> dict[str, torch.Tensor]:
assert os.path.exists(db_path), f"Embedding database does not exist: {db_path}"
loaded: dict[str, torch.Tensor] = {}
with sqlite3.connect(db_path) as conn:
self._ensure_embeddings_table(conn)
cursor = conn.cursor()
if sequences is None:
cursor.execute("SELECT sequence, embedding, shape, dtype FROM embeddings")
else:
if len(sequences) == 0:
return loaded
placeholders = ",".join(["?"] * len(sequences))
cursor.execute(
f"SELECT sequence, embedding, shape, dtype FROM embeddings WHERE sequence IN ({placeholders})",
tuple(sequences),
)
rows = cursor.fetchall()
for row in rows:
sequence = row[0]
embedding_bytes = row[1]
shape_text = row[2]
dtype_text = row[3]
assert shape_text is not None, "Missing shape metadata in embeddings table."
assert dtype_text is not None, "Missing dtype metadata in embeddings table."
shape_values = [int(value) for value in shape_text.split(",") if len(value) > 0]
assert len(shape_values) > 0, f"Invalid shape metadata for sequence: {sequence}"
expected_size = int(np.prod(shape_values))
np_dtype = np.dtype(dtype_text)
array = np.frombuffer(embedding_bytes, dtype=np_dtype)
assert array.size == expected_size, f"Shape mismatch while reading sequence: {sequence}"
reshaped = array.copy().reshape(tuple(shape_values))
loaded[sequence] = torch.from_numpy(reshaped)
return loaded
def embed_dataset(
self,
sequences: List[str],
tokenizer: Optional[PreTrainedTokenizerBase] = None,
batch_size: int = 2,
max_len: int = 512,
truncate: bool = True,
full_embeddings: bool = False,
embed_dtype: torch.dtype = torch.float32,
pooling_types: List[str] = ['mean'],
num_workers: int = 0,
sql: bool = False,
save: bool = True,
sql_db_path: str = 'embeddings.db',
save_path: str = 'embeddings.pth',
**kwargs,
) -> Optional[dict[str, torch.Tensor]]:
"""
Embed a dataset of protein sequences.
Supports two modes:
- Tokenizer mode (ESM2/ESM++): provide `tokenizer`, `_embed(input_ids, attention_mask)` is used.
- Sequence mode (E1): pass `tokenizer=None`, `_embed(sequences, return_attention_mask=True, **kwargs)` is used.
"""
sequences = list(set([seq[:max_len] if truncate else seq for seq in sequences]))
sequences = sorted(sequences, key=len, reverse=True)
hidden_size = self.config.hidden_size
pooler = Pooler(pooling_types) if not full_embeddings else None
tokenizer_mode = tokenizer is not None
if tokenizer_mode:
collate_fn = build_collator(tokenizer)
device = self.device
else:
collate_fn = None
device = None
def get_embeddings(residue_embeddings: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
if full_embeddings or residue_embeddings.ndim == 2:
return residue_embeddings
return pooler(residue_embeddings, attention_mask)
def iter_batches(to_embed: List[str]):
if tokenizer_mode:
assert collate_fn is not None
assert device is not None
dataset = ProteinDataset(to_embed)
dataloader = DataLoader(dataset, batch_size=batch_size, num_workers=num_workers, collate_fn=collate_fn, shuffle=False)
for i, batch in tqdm(enumerate(dataloader), total=len(dataloader), desc='Embedding batches'):
seqs = to_embed[i * batch_size:(i + 1) * batch_size]
input_ids = batch['input_ids'].to(device)
attention_mask = batch['attention_mask'].to(device)
residue_embeddings = self._embed(input_ids, attention_mask)
yield seqs, residue_embeddings, attention_mask
else:
for batch_start in tqdm(range(0, len(to_embed), batch_size), desc='Embedding batches'):
seqs = to_embed[batch_start:batch_start + batch_size]
batch_output = self._embed(seqs, return_attention_mask=True, **kwargs)
assert isinstance(batch_output, tuple), "Sequence mode _embed must return (last_hidden_state, attention_mask)."
assert len(batch_output) == 2, "Sequence mode _embed must return exactly two values."
residue_embeddings, attention_mask = batch_output
assert isinstance(attention_mask, torch.Tensor), "Sequence mode _embed must return attention_mask as a torch.Tensor."
yield seqs, residue_embeddings, attention_mask
if sql:
conn = sqlite3.connect(sql_db_path)
self._ensure_embeddings_table(conn)
c = conn.cursor()
already_embedded = self._read_sequences_from_db(sql_db_path)
to_embed = [seq for seq in sequences if seq not in already_embedded]
print(f"Found {len(already_embedded)} already embedded sequences in {sql_db_path}")
print(f"Embedding {len(to_embed)} new sequences")
if len(to_embed) > 0:
with torch.no_grad():
for i, (seqs, residue_embeddings, attention_mask) in enumerate(iter_batches(to_embed)):
embeddings = get_embeddings(residue_embeddings, attention_mask).to(embed_dtype)
for seq, emb, mask in zip(seqs, embeddings, attention_mask):
if full_embeddings:
emb = emb[mask.bool()].reshape(-1, hidden_size)
emb_np = emb.cpu().numpy()
emb_shape = ",".join([str(dim) for dim in emb_np.shape])
emb_dtype = str(emb_np.dtype)
c.execute(
"INSERT OR REPLACE INTO embeddings (sequence, embedding, shape, dtype) VALUES (?, ?, ?, ?)",
(seq, emb_np.tobytes(), emb_shape, emb_dtype),
)
if tokenizer_mode and (i + 1) % 100 == 0:
conn.commit()
conn.commit()
conn.close()
return None
embeddings_dict = {}
if os.path.exists(save_path):
embeddings_dict = self.load_embeddings_from_pth(save_path)
to_embed = [seq for seq in sequences if seq not in embeddings_dict]
print(f"Found {len(embeddings_dict)} already embedded sequences in {save_path}")
print(f"Embedding {len(to_embed)} new sequences")
else:
to_embed = sequences
print(f"Embedding {len(to_embed)} new sequences")
if len(to_embed) > 0:
with torch.no_grad():
for seqs, residue_embeddings, attention_mask in iter_batches(to_embed):
embeddings = get_embeddings(residue_embeddings, attention_mask).to(embed_dtype)
for seq, emb, mask in zip(seqs, embeddings, attention_mask):
if full_embeddings:
emb = emb[mask.bool()].reshape(-1, hidden_size)
embeddings_dict[seq] = emb.cpu()
if save:
torch.save(embeddings_dict, save_path)
return embeddings_dict
"""
ESM++ model implementation.
ESM++ is a faithful implementation of ESMC that allows for batching and standard Huggingface compatibility
The ESM Python package is not required
Modified from https://github.com/evolutionaryscale/esm
License: https://www.evolutionaryscale.ai/policies/cambrian-non-commercial-license-agreement
"""
import math
import os
import torch
import torch.nn as nn
import torch.nn.functional as F
from dataclasses import dataclass
from functools import cache, partial
from pathlib import Path
from typing import Optional, Tuple, Union, List
from einops import rearrange, repeat
from huggingface_hub import snapshot_download
from tokenizers import Tokenizer
from tokenizers.models import BPE
from tokenizers.processors import TemplateProcessing
from transformers import PreTrainedModel, PreTrainedTokenizerFast, PretrainedConfig
from transformers.modeling_outputs import ModelOutput
try:
from torch.nn.attention.flex_attention import create_block_mask
from torch.nn.attention.flex_attention import flex_attention
except ImportError:
create_block_mask = None
flex_attention = None
def get_attention_mask(
attn_backend: str,
batch_size: int,
seq_len: int,
device: torch.device,
attention_mask: Optional[torch.Tensor] = None
) -> torch.Tensor:
if attention_mask is None:
token_attention_mask = torch.ones((batch_size, seq_len), device=device).bool()
else:
token_attention_mask = attention_mask.bool()
if attn_backend == "flex":
assert create_block_mask is not None, "Flex attention backend requested but torch.create_block_mask is unavailable."
if attention_mask is None:
flex_block_mask = None
else:
def mask_mod(batch_idx, head_idx, q_idx, kv_idx):
return (token_attention_mask[batch_idx, q_idx] == token_attention_mask[batch_idx, kv_idx]) & (token_attention_mask[batch_idx, q_idx] != 0)
flex_block_mask = create_block_mask(
mask_mod,
batch_size,
1,
seq_len,
seq_len,
device=device,
)
extended_attention_mask = None
else:
flex_block_mask = None
extended_attention_mask = token_attention_mask[:, None, :, None] & token_attention_mask[:, None, None, :]
return extended_attention_mask, flex_block_mask
class ESMplusplusConfig(PretrainedConfig):
"""Configuration class for ESM++ model.
Args:
vocab_size: Size of the vocabulary
hidden_size: Dimension of hidden layers
num_attention_heads: Number of attention heads
num_hidden_layers: Number of transformer layers
num_labels: Number of output labels for classification
problem_type: Type of problem - regression, single/multi label classification
"""
model_type = "ESMplusplus"
def __init__(
self,
vocab_size: int = 64,
hidden_size: int = 960,
num_attention_heads: int = 15,
num_hidden_layers: int = 30,
num_labels: int = 2,
problem_type: str | None = None,
dropout: float = 0.0,
initializer_range: float = 0.02,
attn_backend: str = "sdpa",
**kwargs,
):
super().__init__(**kwargs)
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_attention_heads = num_attention_heads
self.num_hidden_layers = num_hidden_layers
self.num_labels = num_labels
self.problem_type = problem_type
self.dropout = dropout
self.initializer_range = initializer_range
self.tie_word_embeddings = False
self.attn_backend = attn_backend
### Rotary Embeddings
def rotate_half(x: torch.Tensor, interleaved: bool = False) -> torch.Tensor:
"""Rotates half the hidden dims of the input."""
if not interleaved:
x1, x2 = x.chunk(2, dim=-1)
return torch.cat((-x2, x1), dim=-1)
else:
x1, x2 = x[..., ::2], x[..., 1::2]
return rearrange(
torch.stack((-x2, x1), dim=-1), "... d two -> ... (d two)", two=2
)
def apply_rotary_emb_torch(
x: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
interleaved: bool = False,
_inplace: bool = False,
) -> torch.Tensor:
"""Apply rotary embeddings to input based on cos and sin."""
ro_dim = cos.shape[-1] * 2
assert ro_dim <= x.shape[-1]
seqlen = x.size(1)
cos = cos[:seqlen]
sin = sin[:seqlen]
cos = repeat(cos, "s d -> s 1 (2 d)")
sin = repeat(sin, "s d -> s 1 (2 d)")
return torch.cat(
[
x[..., :ro_dim] * cos + rotate_half(x[..., :ro_dim], interleaved) * sin,
x[..., ro_dim:],
],
dim=-1,
)
class RotaryEmbedding(torch.nn.Module):
"""Rotary position embeddings.
Based on the paper "RoFormer: Enhanced Transformer with Rotary Position Embedding"
Args:
dim: Dimension of the embedding
base: Base for computing angular frequencies
interleaved: Whether to use interleaved rotations
scale_base: Base for scaling
scaling_factor: Factor for scaling positions
pos_idx_in_fp32: Whether to compute position indices in fp32
device: Computation device
"""
def __init__(
self,
dim: int,
base: float = 10000.0,
interleaved: bool = False,
scale_base: Optional[float] = None,
scaling_factor: float = 1.0,
pos_idx_in_fp32: bool = True,
device: Optional[torch.device] = None,
):
super().__init__()
self.dim = dim
self.base = float(base)
self.pos_idx_in_fp32 = pos_idx_in_fp32
self.interleaved = interleaved
self.scale_base = scale_base
self.scaling_factor = scaling_factor
self.device = device
self._seq_len_cached = 0
self._cos_cached = None
self._sin_cached = None
self._cos_k_cached = None
self._sin_k_cached = None
self._inv_freq_compute_device: Optional[torch.device] = None
self.reset_parameters()
def reset_parameters(self):
"""Reset the parameters of the embedding."""
if "inv_freq" in self._buffers and isinstance(self._buffers["inv_freq"], torch.Tensor):
buffer_device = self._buffers["inv_freq"].device
else:
buffer_device = self.device
inv_freq = self._compute_inv_freq(buffer_device)
self._inv_freq_compute_device = inv_freq.device
self._seq_len_cached = 0
self._cos_cached = None
self._sin_cached = None
self._cos_k_cached = None
self._sin_k_cached = None
self.register_buffer("inv_freq", inv_freq, persistent=False)
arange = torch.arange(0, self.dim, 2, device=buffer_device, dtype=torch.float32)
scale = (
(arange + 0.4 * self.dim) / (1.4 * self.dim)
if self.scale_base is not None
else None
)
self.register_buffer("scale", scale)
def _compute_inv_freq(self, device: Optional[torch.device] = None) -> torch.Tensor:
"""Compute inverse frequency bands."""
return 1 / (
self.base
** (
torch.arange(0, self.dim, 2, device=device, dtype=torch.float32)
/ self.dim
)
)
def _update_cos_sin_cache(self, seqlen: int, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None):
"""Update the cached cosine and sine values."""
if (
seqlen > self._seq_len_cached
or self._cos_cached is None
or self._cos_cached.device != device
or self._cos_cached.dtype != dtype
or (self.training and self._cos_cached.is_inference())
):
self._seq_len_cached = seqlen
if self.pos_idx_in_fp32:
t = torch.arange(seqlen, device=device, dtype=torch.float32)
t /= self.scaling_factor
if self.inv_freq.dtype != torch.float32:
inv_freq = self.inv_freq.to(torch.float32)
else:
inv_freq = self.inv_freq
else:
t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
t /= self.scaling_factor
inv_freq = self.inv_freq
freqs = torch.outer(t, inv_freq)
if self.scale is None:
self._cos_cached = torch.cos(freqs).to(dtype)
self._sin_cached = torch.sin(freqs).to(dtype)
else:
power = (
torch.arange(
seqlen, dtype=self.scale.dtype, device=self.scale.device
)
- seqlen // 2
) / self.scale_base
scale = self.scale.to(device=power.device) ** power.unsqueeze(-1)
self._cos_cached = (torch.cos(freqs) * scale).to(dtype)
self._sin_cached = (torch.sin(freqs) * scale).to(dtype)
self._cos_k_cached = (torch.cos(freqs) / scale).to(dtype)
self._sin_k_cached = (torch.sin(freqs) / scale).to(dtype)
def forward(self, q: torch.Tensor, k: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
"""Apply rotary embeddings to queries and keys.
Args:
q: Query tensor of shape (batch, seqlen, nheads, headdim)
k: Key tensor of shape (batch, seqlen, nheads, headdim)
Returns:
Tuple of rotated query and key tensors
"""
assert self._inv_freq_compute_device is not None, "Rotary inv_freq compute device should be set after initialization."
if self._inv_freq_compute_device != q.device:
self.reset_parameters()
self._update_cos_sin_cache(q.shape[1], device=q.device, dtype=q.dtype)
assert self._cos_cached is not None
assert self._sin_cached is not None
if self.scale is None:
return (
apply_rotary_emb_torch(
q,
self._cos_cached,
self._sin_cached,
self.interleaved,
True, # inplace=True
),
apply_rotary_emb_torch(
k,
self._cos_cached,
self._sin_cached,
self.interleaved,
True, # inplace=True
),
) # type: ignore
else:
assert False
### Feedforward Network Components
def swiglu_correction_fn(expansion_ratio: float, d_model: int) -> int:
"""Compute corrected dimension for SwiGLU."""
return int(((expansion_ratio * d_model) + 255) // 256 * 256)
class SwiGLU(nn.Module):
"""SwiGLU activation function."""
def __init__(self):
super(SwiGLU, self).__init__()
def forward(self, x: torch.Tensor) -> torch.Tensor:
x1, x2 = x.chunk(2, dim=-1)
return F.silu(x1) * x2
def swiglu_ln_ffn(d_model: int, expansion_ratio: float) -> nn.Sequential:
"""Create SwiGLU feedforward network with layer normalization."""
return nn.Sequential(
nn.LayerNorm(d_model),
nn.Linear(
d_model, swiglu_correction_fn(expansion_ratio, d_model) * 2, bias=False
),
SwiGLU(),
nn.Linear(swiglu_correction_fn(expansion_ratio, d_model), d_model, bias=False),
)
### Attention
class MultiHeadAttention(nn.Module):
"""Multi-head attention with rotary embeddings.
Args:
d_model: Model dimension
n_heads: Number of attention heads
"""
def __init__(
self,
d_model: int,
n_heads: int,
attn_backend: str = "sdpa",
):
super().__init__()
self.d_model = d_model
self.n_heads = n_heads
self.d_head = self.d_model // self.n_heads
self.attn_backend = attn_backend
self.layernorm_qkv = nn.Sequential(
nn.LayerNorm(d_model), nn.Linear(d_model, d_model * 3, bias=False)
)
self.out_proj = nn.Linear(d_model, d_model, bias=False)
self.q_ln = nn.LayerNorm(d_model, bias=False)
self.k_ln = nn.LayerNorm(d_model, bias=False)
self.reshaper = partial(rearrange, pattern="b s (h d) -> b h s d", h=n_heads)
self.rotary = RotaryEmbedding(d_model // n_heads)
def _apply_rotary(self, q: torch.Tensor, k: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
"""Apply rotary embeddings to query and key."""
q = q.unflatten(-1, (self.n_heads, self.d_head))
k = k.unflatten(-1, (self.n_heads, self.d_head))
q, k = self.rotary(q, k)
q = q.flatten(-2, -1)
k = k.flatten(-2, -1)
return q, k
def forward(
self,
x: torch.Tensor,
attention_mask: torch.Tensor,
flex_block_mask: object,
output_attentions: bool = False,
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
"""
Args:
x: Input tensor
attention_mask: 4D attention mask
flex_block_mask: Flex attention block mask
output_attentions: Whether to return attention weights
Returns:
Output tensor after self attention, and optionally attention weights
"""
attn_weights = None
qkv_BLD3 = self.layernorm_qkv(x)
query_BLD, key_BLD, value_BLD = torch.chunk(qkv_BLD3, 3, dim=-1)
query_BLD, key_BLD = (
self.q_ln(query_BLD).to(query_BLD.dtype),
self.k_ln(key_BLD).to(query_BLD.dtype),
)
query_BLD, key_BLD = self._apply_rotary(query_BLD, key_BLD)
query_BHLD, key_BHLD, value_BHLD = map(self.reshaper, (query_BLD, key_BLD, value_BLD))
scale = 1 / math.sqrt(self.d_head)
if output_attentions: # Manual attention computation
attn_weights = torch.matmul(query_BHLD, key_BHLD.transpose(-2, -1)) * scale
attn_weights = attn_weights.masked_fill(attention_mask.logical_not(), float('-inf'))
attn_weights = F.softmax(attn_weights, dim=-1)
context_BHLD = torch.matmul(attn_weights, value_BHLD)
else:
if self.attn_backend == "flex":
assert flex_attention is not None, "Flex attention backend requested but torch.flex_attention is unavailable."
assert query_BHLD.dtype in (torch.float16, torch.bfloat16), f"Flex attention backend requires float16 or bfloat16, got {query_BHLD.dtype}."
assert flex_block_mask is not None, "Flex attention backend requires a block mask when attention_mask is provided."
context_BHLD = flex_attention(
query_BHLD,
key_BHLD,
value_BHLD,
block_mask=flex_block_mask,
scale=scale,
)
else:
context_BHLD = F.scaled_dot_product_attention(
query_BHLD,
key_BHLD,
value_BHLD,
attn_mask=attention_mask,
scale=scale,
)
context_BLD = rearrange(context_BHLD, "b h s d -> b s (h d)")
output = self.out_proj(context_BLD)
return output, attn_weights
### Regression Head
def RegressionHead(d_model: int, output_dim: int, hidden_dim: Optional[int] = None) -> nn.Module:
"""Create a regression head with optional hidden dimension.
Args:
d_model: Input dimension
output_dim: Output dimension
hidden_dim: Optional hidden dimension (defaults to d_model)
"""
hidden_dim = hidden_dim if hidden_dim is not None else d_model
return nn.Sequential(
nn.Linear(d_model, hidden_dim),
nn.GELU(),
nn.LayerNorm(hidden_dim),
nn.Linear(hidden_dim, output_dim),
)
### Transformer Block
class UnifiedTransformerBlock(nn.Module):
"""Transformer block with attention and feedforward layers.
Args:
d_model: Model dimension
n_heads: Number of attention heads
residue_scaling_factor: Factor for scaling residual connections
expansion_ratio: Expansion ratio for feedforward network
"""
def __init__(
self,
d_model: int,
n_heads: int,
residue_scaling_factor: float = 1,
expansion_ratio: float = 8 / 3,
dropout: float = 0.0,
attn_backend: str = "sdpa",
):
super().__init__()
self.attn = MultiHeadAttention(
d_model=d_model,
n_heads=n_heads,
attn_backend=attn_backend,
)
self.ffn = swiglu_ln_ffn(d_model, expansion_ratio)
self.scaling_factor = residue_scaling_factor
self.dropout = nn.Dropout(dropout)
def forward(
self,
x: torch.Tensor,
attention_mask: torch.Tensor,
flex_block_mask: object,
output_attentions: bool = False,
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
"""
Args:
x: Input tensor
attention_mask: 4D attention mask
flex_block_mask: Flex attention block mask
output_attentions: Whether to return attention weights
Returns:
Output tensor after transformer block, and optionally attention weights
"""
attn_output, attn_weights = self.attn(
x,
attention_mask,
flex_block_mask,
output_attentions,
)
x = x + self.dropout(attn_output) / self.scaling_factor
x = x + self.dropout(self.ffn(x)) / self.scaling_factor
return x, attn_weights
### Model Outputs
@dataclass
class TransformerOutput(ModelOutput):
"""Output type for transformer encoder."""
last_hidden_state: Optional[torch.Tensor] = None
hidden_states: Optional[Tuple[torch.Tensor]] = None
attentions: Optional[Tuple[torch.Tensor]] = None
@dataclass
class ESMplusplusOutput(ModelOutput):
"""Output type for ESM++ models."""
loss: Optional[torch.Tensor] = None
logits: Optional[torch.Tensor] = None
last_hidden_state: Optional[torch.Tensor] = None
hidden_states: Optional[Tuple[torch.Tensor]] = None
attentions: Optional[Tuple[torch.Tensor]] = None
### Transformer Stack
class TransformerStack(nn.Module):
"""Stack of transformer blocks.
Args:
d_model: Model dimension
n_heads: Number of attention heads
n_layers: Number of transformer layers
dropout: Dropout rate
"""
def __init__(
self,
d_model: int,
n_heads: int,
n_layers: int,
dropout: float = 0.0,
attn_backend: str = "sdpa",
):
super().__init__()
self._attn_backend = attn_backend
self.blocks = nn.ModuleList(
[
UnifiedTransformerBlock(
d_model,
n_heads,
residue_scaling_factor=math.sqrt(n_layers / 36),
dropout=dropout,
attn_backend=attn_backend,
)
for i in range(n_layers)
]
)
self.norm = nn.LayerNorm(d_model, bias=False)
self.gradient_checkpointing = False
self.attn_backend = attn_backend
@property
def attn_backend(self) -> str:
return self._attn_backend
@attn_backend.setter
def attn_backend(self, backend: str) -> None:
assert backend in ("sdpa", "flex"), f"Unsupported attn_backend: {backend}"
self._attn_backend = backend
for block in self.blocks:
block.attn.attn_backend = backend
def forward(
self,
x: torch.Tensor,
attention_mask: Optional[torch.Tensor] = None,
output_hidden_states: Optional[bool] = False,
output_attentions: Optional[bool] = False,
) -> TransformerOutput:
"""
Args:
x: Input tensor
attention_mask: Optional 2D attention mask
output_hidden_states: Whether to return all hidden states
output_attentions: Whether to return attention weights
Returns:
TransformerOutput containing last hidden state and optionally all hidden states and attention weights
"""
hidden_states = () if output_hidden_states else None
attentions = () if output_attentions else None
# move to 4D attention mask or flex block mask
attention_mask, flex_block_mask = get_attention_mask(
attn_backend=self._attn_backend,
batch_size=x.shape[0],
seq_len=x.shape[1],
device=x.device,
attention_mask=attention_mask,
)
for block in self.blocks:
if self.gradient_checkpointing and self.training:
x, attn_weights = self._gradient_checkpointing_func(
block.__call__,
x=x,
attention_mask=attention_mask,
flex_block_mask=flex_block_mask,
output_attentions=output_attentions,
)
else:
x, attn_weights = block(
x=x,
attention_mask=attention_mask,
flex_block_mask=flex_block_mask,
output_attentions=output_attentions,
)
if attentions is not None:
attentions += (attn_weights,)
if output_hidden_states:
assert hidden_states is not None
hidden_states += (x,)
last_hidden_state = self.norm(x)
if output_hidden_states:
hidden_states += (last_hidden_state,)
return TransformerOutput(
last_hidden_state=last_hidden_state,
hidden_states=hidden_states,
attentions=attentions
)
class PreTrainedESMplusplusModel(PreTrainedModel):
"""
init weights for ESM++ models
"""
config_class = ESMplusplusConfig
base_model_prefix = "esm++"
supports_gradient_checkpointing = True
all_tied_weights_keys = {}
def _init_weights(self, module):
"""Initialize the weights"""
# HF from_pretrained marks loaded parameters with `_is_hf_initialized`.
# Skip this module if any local parameter is already marked as loaded.
for parameter in module.parameters(recurse=False):
if "_is_hf_initialized" in parameter.__dict__ and parameter.__dict__["_is_hf_initialized"]:
return
if isinstance(module, nn.Linear):
nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
if module.bias is not None:
nn.init.zeros_(module.bias)
elif isinstance(module, nn.Embedding):
nn.init.normal_(module.weight, mean=0.0, std=self.config.initializer_range)
if module.padding_idx is not None:
with torch.no_grad():
module.weight[module.padding_idx].zero_()
elif isinstance(module, nn.LayerNorm):
if module.bias is not None:
nn.init.zeros_(module.bias)
nn.init.ones_(module.weight)
@property
def attn_backend(self) -> str:
return self.config.attn_backend
@attn_backend.setter
def attn_backend(self, backend: str) -> None:
assert backend in ("sdpa", "flex"), f"Unsupported attn_backend: {backend}"
self.config.attn_backend = backend
for module in self.modules():
if isinstance(module, TransformerStack):
module.attn_backend = backend
def _reset_rotary_embeddings(self):
"""Refresh non-persistent rotary buffers after checkpoint loading."""
for module in self.modules():
if isinstance(module, RotaryEmbedding):
module.reset_parameters()
@classmethod
def from_pretrained(cls, pretrained_model_name_or_path, *model_args, **kwargs):
output_loading_info = bool(kwargs["output_loading_info"]) if "output_loading_info" in kwargs else False
loaded = super().from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
if output_loading_info:
model, loading_info = loaded
model._reset_rotary_embeddings()
return model, loading_info
loaded._reset_rotary_embeddings()
return loaded
@classmethod
def from_pretrained_esm(cls, model_name: str):
"""Load a pretrained ESM++ model."""
if '300' in model_name:
return ESMplusplus_300M()
elif '600' in model_name:
return ESMplusplus_600M()
else:
raise ValueError(f"Invalid model name: {model_name}")
### ESM++ Models
class ESMplusplusModel(PreTrainedESMplusplusModel, EmbeddingMixin):
"""
ESM++ model. transformer model with no heads
"""
config_class = ESMplusplusConfig
def __init__(self, config: ESMplusplusConfig, **kwargs):
PreTrainedESMplusplusModel.__init__(self, config, **kwargs)
self.config = config
self.vocab_size = config.vocab_size
self.embed = nn.Embedding(self.vocab_size, config.hidden_size)
self.transformer = TransformerStack(
d_model=config.hidden_size,
n_heads=config.num_attention_heads,
n_layers=config.num_hidden_layers,
dropout=config.dropout,
attn_backend=config.attn_backend,
)
self.tokenizer = EsmSequenceTokenizer()
self.init_weights()
def get_input_embeddings(self):
return self.embed
def set_input_embeddings(self, value):
self.embed = value
def _embed(self, input_ids: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
x = self.embed(input_ids)
return self.transformer(
x=x,
attention_mask=attention_mask,
output_hidden_states=False,
output_attentions=False,
).last_hidden_state
def forward(
self,
input_ids: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
inputs_embeds: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None, # to play nice with HF adjacent packages
**kwargs,
) -> TransformerOutput:
"""Forward pass for masked language modeling.
Args:
input_ids: Input token IDs
attention_mask: Attention mask
inputs_embeds: Optional precomputed embeddings
output_hidden_states: Whether to return all hidden states
output_attentions: Whether to return attention weights
Returns:
TransformerOutput containing last hidden state and optionally all hidden states and attention weights
"""
assert input_ids is not None or inputs_embeds is not None, "You have to specify either input_ids or inputs_embeds"
assert not (input_ids is not None and inputs_embeds is not None), "You cannot specify both input_ids and inputs_embeds at the same time"
if inputs_embeds is None:
x = self.embed(input_ids)
else:
x = inputs_embeds
transformer_output = self.transformer(
x=x,
attention_mask=attention_mask,
output_hidden_states=output_hidden_states,
output_attentions=output_attentions,
)
return ESMplusplusOutput(
last_hidden_state=transformer_output.last_hidden_state,
hidden_states=transformer_output.hidden_states,
attentions=transformer_output.attentions,
)
class ESMplusplusForMaskedLM(PreTrainedESMplusplusModel, EmbeddingMixin):
"""
ESM++ model for masked language modeling.
Implements the base ESM++ architecture with a masked language modeling head.
"""
config_class = ESMplusplusConfig
def __init__(self, config: ESMplusplusConfig, **kwargs):
PreTrainedESMplusplusModel.__init__(self, config, **kwargs)
self.config = config
self.vocab_size = config.vocab_size
self.embed = nn.Embedding(self.vocab_size, config.hidden_size)
self.transformer = TransformerStack(
d_model=config.hidden_size,
n_heads=config.num_attention_heads,
n_layers=config.num_hidden_layers,
dropout=config.dropout,
attn_backend=config.attn_backend,
)
self.sequence_head = RegressionHead(config.hidden_size, self.vocab_size)
self.ce_loss = nn.CrossEntropyLoss()
self.tokenizer = EsmSequenceTokenizer()
self.init_weights()
def get_input_embeddings(self):
return self.embed
def set_input_embeddings(self, value):
self.embed = value
def get_output_embeddings(self):
return self.sequence_head[-1]
def set_output_embeddings(self, new_embeddings):
self.sequence_head[-1] = new_embeddings
def _embed(self, input_ids: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
x = self.embed(input_ids)
return self.transformer(
x=x,
attention_mask=attention_mask,
output_hidden_states=False,
output_attentions=False,
).last_hidden_state
def forward(
self,
input_ids: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
inputs_embeds: Optional[torch.Tensor] = None,
labels: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None, # to play nice with HF adjacent packages
**kwargs,
) -> ESMplusplusOutput:
"""Forward pass for masked language modeling.
Args:
input_ids: Input token IDs
attention_mask: Attention mask
inputs_embeds: Optional precomputed embeddings
labels: Optional labels for masked tokens
output_hidden_states: Whether to return all hidden states
output_attentions: Whether to return attention weights
Returns:
ESMplusplusOutput containing loss, logits, hidden states and attention weights
"""
if inputs_embeds is None:
x = self.embed(input_ids)
else:
x = inputs_embeds
output = self.transformer(
x=x,
attention_mask=attention_mask,
output_hidden_states=output_hidden_states,
output_attentions=output_attentions,
)
last_hidden_state = output.last_hidden_state
logits = self.sequence_head(last_hidden_state)
loss = None
if labels is not None:
loss = self.ce_loss(logits.view(-1, self.vocab_size), labels.view(-1))
return ESMplusplusOutput(
loss=loss,
logits=logits,
last_hidden_state=last_hidden_state,
hidden_states=output.hidden_states,
attentions=output.attentions,
)
class ESMplusplusForSequenceClassification(ESMplusplusForMaskedLM, EmbeddingMixin):
"""
ESM++ model for sequence classification.
Extends the base ESM++ model with a classification head.
"""
def __init__(self, config: ESMplusplusConfig, **kwargs):
ESMplusplusForMaskedLM.__init__(self, config, **kwargs)
self.config = config
self.num_labels = config.num_labels
self.classifier = RegressionHead(config.hidden_size * 2, config.num_labels, config.hidden_size * 4)
# Large intermediate projections help with sequence classification tasks (*4)
self.mse = nn.MSELoss()
self.ce = nn.CrossEntropyLoss()
self.bce = nn.BCEWithLogitsLoss()
# if kwargs has pooling_types, use them, otherwise use ['cls', 'mean']
if 'pooling_types' in kwargs and isinstance(kwargs['pooling_types'], List[str]) and len(kwargs['pooling_types']) > 0:
pooling_types = kwargs['pooling_types']
else:
pooling_types = ['mean', 'var']
self.pooler = Pooler(pooling_types)
self.init_weights()
def _embed(self, input_ids: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
x = self.embed(input_ids)
return self.transformer(
x=x,
attention_mask=attention_mask,
output_hidden_states=False,
output_attentions=False,
).last_hidden_state
def forward(
self,
input_ids: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
inputs_embeds: Optional[torch.Tensor] = None,
labels: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None, # to play nice with HF adjacent packages
**kwargs,
) -> ESMplusplusOutput:
"""Forward pass for sequence classification.
Args:
input_ids: Input token IDs
attention_mask: Attention mask
inputs_embeds: Optional precomputed embeddings
labels: Optional labels for classification
output_hidden_states: Whether to return all hidden states
output_attentions: Whether to return attention weights
Returns:
ESMplusplusOutput containing loss, logits, and hidden states
"""
output = super().forward(
input_ids=input_ids,
attention_mask=attention_mask,
inputs_embeds=inputs_embeds,
labels=None,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states
)
last_hidden_state = output.last_hidden_state
features = self.pooler(last_hidden_state, attention_mask) # pooler expects 2d attention mask
logits = self.classifier(features)
loss = None
if labels is not None:
labels = labels.to(logits.device)
if self.config.problem_type is None:
if self.num_labels == 1:
self.config.problem_type = "regression"
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
self.config.problem_type = "single_label_classification"
else:
self.config.problem_type = "multi_label_classification"
if self.config.problem_type == "regression":
if self.num_labels == 1:
loss = self.mse(logits.flatten(), labels.flatten())
else:
loss = self.mse(logits, labels)
elif self.config.problem_type == "single_label_classification":
loss = self.ce(logits.view(-1, self.num_labels), labels.view(-1))
elif self.config.problem_type == "multi_label_classification":
loss = self.bce(logits, labels)
return ESMplusplusOutput(
loss=loss,
logits=logits,
last_hidden_state=last_hidden_state,
hidden_states=output.hidden_states,
attentions=output.attentions,
)
class ESMplusplusForTokenClassification(ESMplusplusForMaskedLM, EmbeddingMixin):
"""
ESM++ model for token classification.
Extends the base ESM++ model with a token classification head.
"""
def __init__(self, config: ESMplusplusConfig, **kwargs):
ESMplusplusForMaskedLM.__init__(self, config, **kwargs)
self.config = config
self.num_labels = config.num_labels
self.classifier = RegressionHead(config.hidden_size, config.num_labels, config.hidden_size * 4)
# Large intermediate projections help with sequence classification tasks (*4)
self.loss_fct = nn.CrossEntropyLoss()
self.init_weights()
def _embed(self, input_ids: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
x = self.embed(input_ids)
return self.transformer(x, attention_mask, output_hidden_states=False, output_attentions=False).last_hidden_state
def forward(
self,
input_ids: Optional[torch.Tensor] = None,
attention_mask: Optional[torch.Tensor] = None,
inputs_embeds: Optional[torch.Tensor] = None,
labels: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None, # to play nice with HF adjacent packages
**kwargs,
) -> ESMplusplusOutput:
"""Forward pass for token classification.
Args:
input_ids: Input token IDs
attention_mask: Attention mask
inputs_embeds: Optional precomputed embeddings
labels: Optional labels for token classification
output_hidden_states: Whether to return all hidden states
output_attentions: Whether to return attention weights
Returns:
ESMplusplusOutput containing loss, logits, and hidden states
"""
output = super().forward(
input_ids=input_ids,
attention_mask=attention_mask,
inputs_embeds=inputs_embeds,
labels=None,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states
)
last_hidden_state = output.last_hidden_state
logits = self.classifier(last_hidden_state)
loss = None
if labels is not None:
loss = self.loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
return ESMplusplusOutput(
loss=loss,
logits=logits,
last_hidden_state=last_hidden_state,
hidden_states=output.hidden_states,
attentions=output.attentions,
)
### Loading from EvolutionaryScale
_ESMC_CHECKPOINT_SPECS = {
"esmc-300": {
"repo_id": "EvolutionaryScale/esmc-300m-2024-12",
"weights_relpath": "data/weights/esmc_300m_2024_12_v0.pth",
"hidden_size": 960,
"num_attention_heads": 15,
"num_hidden_layers": 30,
},
"esmc-600": {
"repo_id": "EvolutionaryScale/esmc-600m-2024-12",
"weights_relpath": "data/weights/esmc_600m_2024_12_v0.pth",
"hidden_size": 1152,
"num_attention_heads": 18,
"num_hidden_layers": 36,
},
}
def _resolve_esmc_checkpoint_key(model: str) -> str:
if "esmc-300" in model:
return "esmc-300"
if "esmc-600" in model:
return "esmc-600"
raise ValueError(f"{model=} is an invalid ESMC model name.")
@staticmethod
@cache
def data_root(model: str):
if "INFRA_PROVIDER" in os.environ:
return Path("")
key = _resolve_esmc_checkpoint_key(model)
return Path(snapshot_download(repo_id=_ESMC_CHECKPOINT_SPECS[key]["repo_id"]))
def get_esmc_checkpoint_path(model: str) -> Path:
key = _resolve_esmc_checkpoint_key(model)
return data_root(key) / _ESMC_CHECKPOINT_SPECS[key]["weights_relpath"]
def _load_esmc_checkpoint_model(
config: ESMplusplusConfig,
model: str,
device: torch.device | str = "cpu",
) -> ESMplusplusForMaskedLM:
key = _resolve_esmc_checkpoint_key(model)
spec = _ESMC_CHECKPOINT_SPECS[key]
assert config.hidden_size == spec["hidden_size"], (
f"ESMC loader expected hidden_size={spec['hidden_size']} for {key}, "
f"but got {config.hidden_size}."
)
assert config.num_attention_heads == spec["num_attention_heads"], (
f"ESMC loader expected num_attention_heads={spec['num_attention_heads']} for {key}, "
f"but got {config.num_attention_heads}."
)
assert config.num_hidden_layers == spec["num_hidden_layers"], (
f"ESMC loader expected num_hidden_layers={spec['num_hidden_layers']} for {key}, "
f"but got {config.num_hidden_layers}."
)
with torch.device(device):
model_obj = ESMplusplusForMaskedLM(config)
state_dict = torch.load(get_esmc_checkpoint_path(key), map_location=device)
model_obj.load_state_dict(state_dict)
return model_obj
def ESMplusplus_300M(device: torch.device | str = "cpu"):
config = ESMplusplusConfig(
hidden_size=960,
num_attention_heads=15,
num_hidden_layers=30,
)
return _load_esmc_checkpoint_model(config=config, model="esmc-300", device=device)
def ESMplusplus_600M(device: torch.device | str = "cpu"):
config = ESMplusplusConfig(
hidden_size=1152,
num_attention_heads=18,
num_hidden_layers=36,
)
return _load_esmc_checkpoint_model(config=config, model="esmc-600", device=device)
### Tokenization
SEQUENCE_VOCAB = [
"<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", ".", "-", "|",
"<mask>",
]
class EsmSequenceTokenizer(PreTrainedTokenizerFast):
model_input_names = ["input_ids", "attention_mask"]
def __init__(
self,
unk_token="<unk>",
cls_token="<cls>",
pad_token="<pad>",
mask_token="<mask>",
eos_token="<eos>",
chain_break_token="|",
**kwargs,
):
all_tokens = SEQUENCE_VOCAB
token_to_id = {tok: ind for ind, tok in enumerate(all_tokens)}
# a character-level tokenizer is the same as BPE with no token merges
bpe = BPE(token_to_id, merges=[], unk_token=unk_token)
tokenizer = Tokenizer(bpe)
special_tokens = [
cls_token,
pad_token,
mask_token,
eos_token,
chain_break_token,
]
self.cb_token = chain_break_token
additional_special_tokens = [chain_break_token]
tokenizer.add_special_tokens(special_tokens)
# This is where we configure the automatic addition of special tokens when we call
# tokenizer(text, add_special_tokens=True). Note that you can also configure how two
# sequences are merged if you want.
tokenizer.post_processor = TemplateProcessing( # type: ignore
single="<cls> $A <eos>",
pair="<cls>:0 $A:0 <eos>:0 $B:1 <eos>:1",
special_tokens=[
("<cls>", tokenizer.token_to_id("<cls>")),
("<eos>", tokenizer.token_to_id("<eos>")),
],
)
super().__init__(
tokenizer_object=tokenizer,
unk_token=unk_token,
cls_token=cls_token,
pad_token=pad_token,
mask_token=mask_token,
eos_token=eos_token,
additional_special_tokens=additional_special_tokens,
**kwargs,
)
# These are a footgun, we never use the `bos` token anywhere so we're just overriding it here.
@property
def bos_token(self):
return self.cls_token
@property
def bos_token_id(self):
return self.cls_token_id
@property
def chain_break_token(self):
return self.cb_token
@property
def chain_break_token_id(self):
return self.convert_tokens_to_ids(self.chain_break_token)
@property
def all_token_ids(self):
return list(range(self.vocab_size))
@property
def special_token_ids(self):
return self.all_special_ids
|