File size: 63,353 Bytes
f34d690 |
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 |
"""
DEXTR with LiLT - 3-Step Document Extraction.
Architecture:
- Document encoder: LiLT (XLM-RoBERTa + Layout Transformer), extended to 1024 tokens
- Query encoder: Sentence Transformer (frozen, for semantic similarity)
- Step 1: Token Classification (Q/A/T/H/O) - TRAINED
- Step 2: Query-Question Matching (ZERO-SHOT) - NO TRAINING
- Step 3: Table Head (hierarchical with attention-based column assignment) - TRAINED
Key insight: Query→Question matching is semantic similarity. Sentence transformers
provide proper semantic embeddings for matching queries to question text.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import LiltModel, LiltConfig
from sentence_transformers import SentenceTransformer
from typing import Optional, Dict, List, Tuple
# Label mappings for Step 1
LABEL2ID = {
"O": 0,
"B-Q": 1, "I-Q": 2, # Question
"B-A": 3, "I-A": 4, # Answer
"B-H": 5, "I-H": 6, # Header
"B-TABLE": 7, "I-TABLE": 8, # Table
}
ID2LABEL = {v: k for k, v in LABEL2ID.items()}
NUM_LABELS = len(LABEL2ID)
class TokenClassificationHead(nn.Module):
"""
Step 1: Token Classification Head.
Predicts Q/A/H/T/O labels for each token (FUNSD-style).
"""
def __init__(self, hidden_size: int = 768, num_labels: int = NUM_LABELS, dropout: float = 0.1):
super().__init__()
self.classifier = nn.Sequential(
nn.Dropout(dropout),
nn.Linear(hidden_size, 256),
nn.GELU(),
nn.Dropout(dropout),
nn.Linear(256, num_labels),
)
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
return self.classifier(hidden_states)
class QALinker(nn.Module):
"""
Step 1.5: Q-A Linker.
Predicts which Answer span links to which Question span.
Uses both semantic similarity and spatial features.
"""
def __init__(self, hidden_size: int = 768, dropout: float = 0.1):
super().__init__()
self.hidden_size = hidden_size
# Project Q and A to same space for semantic matching
self.q_proj = nn.Sequential(
nn.Linear(hidden_size, 256),
nn.LayerNorm(256),
nn.GELU(),
nn.Dropout(dropout),
)
self.a_proj = nn.Sequential(
nn.Linear(hidden_size, 256),
nn.LayerNorm(256),
nn.GELU(),
nn.Dropout(dropout),
)
# Spatial feature scorer
# Features: x_dist, y_dist, is_right, is_below, is_same_line, width_ratio, height_ratio
self.spatial_scorer = nn.Sequential(
nn.Linear(7, 32),
nn.GELU(),
nn.Linear(32, 1),
)
# Learnable temperature
self.log_temp = nn.Parameter(torch.tensor(1.0))
def compute_spatial_features(
self,
q_bboxes: torch.Tensor, # (num_q, 4) - [x1, y1, x2, y2]
a_bboxes: torch.Tensor, # (num_a, 4)
) -> torch.Tensor:
"""
Compute spatial features for all Q-A pairs.
Returns: (num_q, num_a, 7) feature tensor
"""
num_q = q_bboxes.shape[0]
num_a = a_bboxes.shape[0]
device = q_bboxes.device
# Compute centers and sizes
q_cx = (q_bboxes[:, 0] + q_bboxes[:, 2]) / 2 # (num_q,)
q_cy = (q_bboxes[:, 1] + q_bboxes[:, 3]) / 2
q_w = q_bboxes[:, 2] - q_bboxes[:, 0]
q_h = q_bboxes[:, 3] - q_bboxes[:, 1]
a_cx = (a_bboxes[:, 0] + a_bboxes[:, 2]) / 2 # (num_a,)
a_cy = (a_bboxes[:, 1] + a_bboxes[:, 3]) / 2
a_w = a_bboxes[:, 2] - a_bboxes[:, 0]
a_h = a_bboxes[:, 3] - a_bboxes[:, 1]
# Expand for pairwise computation
q_cx = q_cx.unsqueeze(1).expand(num_q, num_a) # (num_q, num_a)
q_cy = q_cy.unsqueeze(1).expand(num_q, num_a)
q_x2 = q_bboxes[:, 2].unsqueeze(1).expand(num_q, num_a)
q_y2 = q_bboxes[:, 3].unsqueeze(1).expand(num_q, num_a)
q_w = q_w.unsqueeze(1).expand(num_q, num_a)
q_h = q_h.unsqueeze(1).expand(num_q, num_a)
a_cx = a_cx.unsqueeze(0).expand(num_q, num_a)
a_cy = a_cy.unsqueeze(0).expand(num_q, num_a)
a_x1 = a_bboxes[:, 0].unsqueeze(0).expand(num_q, num_a)
a_y1 = a_bboxes[:, 1].unsqueeze(0).expand(num_q, num_a)
a_w = a_w.unsqueeze(0).expand(num_q, num_a)
a_h = a_h.unsqueeze(0).expand(num_q, num_a)
# Compute features (all normalized to [0, 1] range roughly)
x_dist = (a_x1 - q_x2) / 1000.0 # Horizontal distance (positive = A right of Q)
y_dist = (a_cy - q_cy).abs() / 1000.0 # Vertical distance
is_right = (a_x1 > q_x2).float() # A is to the right of Q
is_below = (a_y1 > q_y2).float() # A is below Q
is_same_line = (y_dist < 0.03).float() # Same line (small y distance)
width_ratio = (a_w / (q_w + 1e-6)).clamp(0, 5) / 5.0 # Relative width
height_ratio = (a_h / (q_h + 1e-6)).clamp(0, 5) / 5.0 # Relative height
# Stack features: (num_q, num_a, 7)
features = torch.stack([
x_dist, y_dist, is_right, is_below, is_same_line, width_ratio, height_ratio
], dim=-1)
return features
def forward(
self,
q_embeds: torch.Tensor, # (num_q, hidden)
a_embeds: torch.Tensor, # (num_a, hidden)
q_bboxes: torch.Tensor, # (num_q, 4)
a_bboxes: torch.Tensor, # (num_a, 4)
) -> torch.Tensor:
"""
Compute Q-A link scores.
Returns: (num_q, num_a) score matrix
"""
# Semantic similarity
q_proj = self.q_proj(q_embeds) # (num_q, 256)
a_proj = self.a_proj(a_embeds) # (num_a, 256)
q_proj = F.normalize(q_proj, dim=-1)
a_proj = F.normalize(a_proj, dim=-1)
semantic_scores = torch.matmul(q_proj, a_proj.t()) # (num_q, num_a)
# Spatial scores
spatial_feats = self.compute_spatial_features(q_bboxes, a_bboxes) # (num_q, num_a, 7)
spatial_scores = self.spatial_scorer(spatial_feats).squeeze(-1) # (num_q, num_a)
# Combine with learnable temperature
temperature = self.log_temp.exp().clamp(min=0.1, max=10.0)
combined_scores = (semantic_scores + spatial_scores) * temperature
return combined_scores
class QuestionPredictor(nn.Module):
"""
Predicts Question embedding from Answer embedding.
Used when explicit Question tokens are missing in the document.
Input: LiLT A embedding (768 dim)
Output: Predicted Q embedding in sentence transformer space (384 dim)
Training:
- For Q-A pairs with Q: randomly mask Q and train to predict it
- Learns what Q "should look like" given the answer
"""
def __init__(self, input_dim: int = 768, output_dim: int = 384, dropout: float = 0.1):
super().__init__()
# Input: LiLT answer embedding (768)
# Output: predicted question embedding in sentence transformer space (384)
self.predictor = nn.Sequential(
nn.Linear(input_dim, input_dim),
nn.LayerNorm(input_dim),
nn.GELU(),
nn.Dropout(dropout),
nn.Linear(input_dim, output_dim),
nn.LayerNorm(output_dim),
)
def forward(self, answer_embed: torch.Tensor) -> torch.Tensor:
"""
Predict question embedding from answer embedding.
Args:
answer_embed: (batch, hidden) or (hidden,) - LiLT answer span embedding
Returns:
predicted_q_embed: (batch, output_dim) or (output_dim,) - in sentence transformer space
"""
return self.predictor(answer_embed)
class QueryAnswerMatcher(nn.Module):
"""
Step 2: Query-Answer Matching.
Given answer span embeddings and query embedding, scores which span matches.
Features:
- Start+End pooling: spans use [h_start; h_end] instead of mean (captures boundaries)
- Bbox features: spatial position helps distinguish similar text
- Learnable temperature for score scaling
- Cosine similarity with projection to lower dim
"""
def __init__(
self,
hidden_size: int = 768,
proj_dim: int = 256,
dropout: float = 0.1,
use_bbox_features: bool = True,
num_bbox_features: int = 6, # x1, y1, x2, y2, width, height
):
super().__init__()
self.proj_dim = proj_dim
self.use_bbox_features = use_bbox_features
self.hidden_size = hidden_size
# Span input: [start; end] = 2*hidden, optionally + bbox
span_input_dim = 2 * hidden_size
if use_bbox_features:
span_input_dim += num_bbox_features
self.span_proj = nn.Sequential(
nn.Linear(span_input_dim, proj_dim),
nn.LayerNorm(proj_dim),
nn.GELU(),
nn.Dropout(dropout),
)
self.query_proj = nn.Sequential(
nn.Linear(hidden_size, proj_dim),
nn.LayerNorm(proj_dim),
nn.GELU(),
nn.Dropout(dropout),
)
# Learnable temperature (initialized to sqrt(proj_dim) like standard attention)
self.log_temp = nn.Parameter(torch.tensor(proj_dim ** 0.5).log())
def forward(
self,
span_embeddings: torch.Tensor, # (batch, num_spans, span_input_dim)
query_embedding: torch.Tensor, # (batch, hidden)
span_mask: Optional[torch.Tensor] = None,
) -> torch.Tensor:
span_proj = self.span_proj(span_embeddings)
query_proj = self.query_proj(query_embedding).unsqueeze(1)
# Normalize for cosine similarity-like behavior
span_proj = F.normalize(span_proj, dim=-1)
query_proj = F.normalize(query_proj, dim=-1)
# Scaled dot product with learnable temperature
temperature = self.log_temp.exp().clamp(min=0.1, max=100.0)
scores = torch.bmm(query_proj, span_proj.transpose(1, 2)).squeeze(1) * temperature
if span_mask is not None:
scores = scores.masked_fill(~span_mask.bool(), float('-inf'))
return scores
class TableHead(nn.Module):
"""
Step 3: Hierarchical Table Extraction Head.
Sub-steps:
1. Table detection (TABLE/O per token)
2. Header detection (HEADER/CELL per token)
3. Row segmentation (B-ROW/I-ROW/O)
4. Column assignment (cell → header via attention)
"""
def __init__(self, hidden_size: int = 768, dropout: float = 0.1):
super().__init__()
self.table_detector = nn.Sequential(
nn.Dropout(dropout),
nn.Linear(hidden_size, 256),
nn.GELU(),
nn.Linear(256, 2),
)
self.header_detector = nn.Sequential(
nn.Dropout(dropout),
nn.Linear(hidden_size, 256),
nn.GELU(),
nn.Linear(256, 2),
)
self.row_tagger = nn.Sequential(
nn.Dropout(dropout),
nn.Linear(hidden_size, 256),
nn.GELU(),
nn.Linear(256, 3),
)
self.col_key_proj = nn.Linear(hidden_size, 128)
self.col_query_proj = nn.Linear(hidden_size, 128)
def forward(self, hidden_states: torch.Tensor) -> Dict[str, torch.Tensor]:
table_logits = self.table_detector(hidden_states)
header_logits = self.header_detector(hidden_states)
row_logits = self.row_tagger(hidden_states)
col_keys = self.col_key_proj(hidden_states)
col_queries = self.col_query_proj(hidden_states)
col_scores = torch.bmm(col_queries, col_keys.transpose(1, 2)) / (128 ** 0.5)
return {
"table_logits": table_logits,
"header_logits": header_logits,
"row_logits": row_logits,
"col_scores": col_scores,
}
class DEXTRLiLT(nn.Module):
"""
DEXTR with LiLT: Multi-Step Document Extraction with Zero-Shot Query Matching.
Architecture:
- Step 1: Token Classification (Q/A/H/T/O) - TRAINED
- Step 1.5: Q-A Linker (links Question spans to Answer spans) - TRAINED
- Step 2: Query → Question Matching (ZERO-SHOT with Sentence Transformer)
- Step 3: Table Head - TRAINED
Uses SEPARATE encoders:
- Document encoder: LiLT (XLM-RoBERTa + Layout Transformer) for layout-aware encoding
- Query encoder: Sentence Transformer (frozen) for semantic similarity
Zero-shot matching:
- Query text encoded with frozen Sentence Transformer
- Question text (from predicted Q regions) encoded with frozen Sentence Transformer
- Direct cosine similarity (no trainable projections)
- Fallback: if no Q regions, use QuestionPredictor on A text
"""
def __init__(
self,
model_name: str = "nielsr/lilt-xlm-roberta-base",
query_model_name: str = "paraphrase-multilingual-MiniLM-L12-v2",
max_seq_len: int = 1024,
hidden_size: int = 768,
dropout: float = 0.1,
q_mask_prob: float = 0.3, # Probability of masking Q during training (for Q-A linker)
):
super().__init__()
self.max_seq_len = max_seq_len
self.hidden_size = hidden_size
self.q_mask_prob = q_mask_prob
# Document encoder: LiLT (layout-aware)
if max_seq_len > 512:
self.encoder = self._load_lilt_extended(model_name, max_seq_len)
else:
self.encoder = LiltModel.from_pretrained(model_name)
# Query encoder: Sentence Transformer (frozen, zero-shot)
# Provides proper semantic similarity for query→question matching
self.query_encoder = SentenceTransformer(query_model_name)
self.query_encoder.requires_grad_(False) # Freeze for zero-shot
self.query_embed_dim = self.query_encoder.get_sentence_embedding_dimension()
print(f"Loaded frozen query encoder: {query_model_name} (dim={self.query_embed_dim})")
# Step 1: Token Classification Head
self.token_classifier = TokenClassificationHead(hidden_size, NUM_LABELS, dropout)
# Step 1.5: Q-A Linker (links Q spans to A spans)
self.qa_linker = QALinker(hidden_size, dropout)
# Step 2: QuestionPredictor for documents without explicit Q labels (e.g., receipts)
# Input: LiLT A embedding (768 dim), Output: Q embedding in sentence transformer space (384 dim)
self.question_predictor = QuestionPredictor(hidden_size, self.query_embed_dim, dropout)
# Step 3: Table Head
self.table_head = TableHead(hidden_size, dropout)
def _load_lilt_extended(self, model_name: str, max_seq_len: int) -> LiltModel:
"""Load LiLT with extended position embeddings."""
model = LiltModel.from_pretrained(model_name)
original_max_pos = model.config.max_position_embeddings
required_positions = max_seq_len + 2
if required_positions <= original_max_pos:
return model
# 1. Extend text position embeddings
old_pos_emb = model.embeddings.position_embeddings.weight.data
hidden_size = old_pos_emb.shape[1]
padding_idx = model.embeddings.position_embeddings.padding_idx
new_pos_emb = nn.Embedding(required_positions, hidden_size, padding_idx=padding_idx)
new_pos_emb.weight.data[:original_max_pos] = old_pos_emb
new_pos_emb.weight.data[original_max_pos:].normal_(mean=0.0, std=0.02)
model.embeddings.position_embeddings = new_pos_emb
# 2. Extend layout box_position_embeddings (same sequence length limit)
old_box_pos = model.layout_embeddings.box_position_embeddings.weight.data
box_hidden = old_box_pos.shape[1] # 192
new_box_pos = nn.Embedding(required_positions, box_hidden)
new_box_pos.weight.data[:original_max_pos] = old_box_pos
new_box_pos.weight.data[original_max_pos:].normal_(mean=0.0, std=0.02)
model.layout_embeddings.box_position_embeddings = new_box_pos
# 3. Update config
model.config.max_position_embeddings = required_positions
# 4. Extend position_ids buffer
new_position_ids = torch.arange(required_positions).unsqueeze(0)
model.embeddings.register_buffer("position_ids", new_position_ids, persistent=False)
print(f"Extended LiLT positions: {original_max_pos} → {required_positions}")
return model
def encode(
self,
input_ids: torch.Tensor,
attention_mask: torch.Tensor,
bbox: torch.Tensor,
) -> torch.Tensor:
"""Encode tokens with layout using shared LiLT encoder."""
outputs = self.encoder(
input_ids=input_ids,
attention_mask=attention_mask,
bbox=bbox,
)
return outputs.last_hidden_state
def encode_texts(
self,
texts: List[str],
device: torch.device,
) -> torch.Tensor:
"""
Encode text strings using frozen Sentence Transformer (zero-shot).
Used for encoding Q text extracted from predicted regions.
Args:
texts: List of text strings to encode
device: Device to put tensors on
Returns:
embeddings: (num_texts, embed_dim) tensor
"""
if not texts:
return torch.zeros(0, self.query_embed_dim, device=device)
with torch.no_grad():
embeddings = self.query_encoder.encode(
texts,
convert_to_tensor=True,
device=device,
)
# Clone to exit inference mode (needed for autograd compatibility)
embeddings = embeddings.clone()
return embeddings # (num_texts, embed_dim)
def pool_spans(
self,
hidden_states: torch.Tensor,
span_indices: List[List[Tuple[int, int]]],
bbox: Optional[torch.Tensor] = None,
use_bbox_features: bool = True,
) -> Tuple[torch.Tensor, torch.Tensor]:
"""
Pool hidden states for answer spans using start+end tokens.
Returns [h_start; h_end; bbox_features] for each span instead of mean pooling.
This preserves boundary information which is crucial for extraction tasks.
Args:
hidden_states: (batch, seq, hidden) - document encodings
span_indices: List of (start, end) tuples per batch item
bbox: (batch, seq, 4) - bounding boxes [x1, y1, x2, y2] normalized to [0, 1000]
use_bbox_features: whether to include bbox in span representation
Returns:
span_embeddings: (batch, max_spans, span_dim) where span_dim = 2*hidden + 6 (if bbox)
span_mask: (batch, max_spans) bool mask
"""
batch_size = hidden_states.shape[0]
hidden_size = hidden_states.shape[2]
device = hidden_states.device
max_spans = max(len(spans) for spans in span_indices) if span_indices else 1
max_spans = max(max_spans, 1)
# Output dimension: [h_start; h_end] + optionally [x1, y1, x2, y2, width, height]
span_dim = 2 * hidden_size
if use_bbox_features and bbox is not None:
span_dim += 6 # normalized bbox features
span_embeddings = torch.zeros(batch_size, max_spans, span_dim, device=device)
span_mask = torch.zeros(batch_size, max_spans, dtype=torch.bool, device=device)
for b, spans in enumerate(span_indices):
for s, (start, end) in enumerate(spans):
if s >= max_spans:
break
if start >= hidden_states.shape[1] or end > hidden_states.shape[1]:
continue
# Start+End pooling: [h_start; h_end]
h_start = hidden_states[b, start, :]
h_end = hidden_states[b, end - 1, :] # end is exclusive
span_repr = torch.cat([h_start, h_end], dim=0)
# Add bbox features if available
if use_bbox_features and bbox is not None:
# Get bbox of span (union of start and end token bboxes)
bbox_start = bbox[b, start, :] # [x1, y1, x2, y2]
bbox_end = bbox[b, end - 1, :]
# Compute span bounding box (min x1/y1, max x2/y2)
span_x1 = torch.min(bbox_start[0], bbox_end[0])
span_y1 = torch.min(bbox_start[1], bbox_end[1])
span_x2 = torch.max(bbox_start[2], bbox_end[2])
span_y2 = torch.max(bbox_start[3], bbox_end[3])
# Normalize to [0, 1] and add width/height
bbox_feat = torch.tensor([
span_x1 / 1000.0,
span_y1 / 1000.0,
span_x2 / 1000.0,
span_y2 / 1000.0,
(span_x2 - span_x1) / 1000.0, # width
(span_y2 - span_y1) / 1000.0, # height
], device=device)
span_repr = torch.cat([span_repr, bbox_feat], dim=0)
span_embeddings[b, s] = span_repr
span_mask[b, s] = True
return span_embeddings, span_mask
def pool_single_span(
self,
hidden_states: torch.Tensor, # (seq, hidden) - single sample
span: Tuple[int, int],
bbox: Optional[torch.Tensor] = None, # (seq, 4)
) -> torch.Tensor:
"""
Pool a single span to get its embedding.
Returns mean of start and end tokens (hidden_size,).
"""
start, end = span
if start >= hidden_states.shape[0] or end > hidden_states.shape[0]:
return torch.zeros(self.hidden_size, device=hidden_states.device)
h_start = hidden_states[start, :]
h_end = hidden_states[end - 1, :]
# Mean of start and end for Q embedding (simpler than concat for matching)
return (h_start + h_end) / 2
def get_span_bbox(
self,
bbox: torch.Tensor, # (seq, 4)
span: Tuple[int, int],
) -> torch.Tensor:
"""Get bounding box for a span (union of start and end token bboxes)."""
start, end = span
bbox_start = bbox[start, :]
bbox_end = bbox[end - 1, :]
span_bbox = torch.stack([
torch.min(bbox_start[0], bbox_end[0]), # x1
torch.min(bbox_start[1], bbox_end[1]), # y1
torch.max(bbox_start[2], bbox_end[2]), # x2
torch.max(bbox_start[3], bbox_end[3]), # y2
])
return span_bbox
def extract_span_text(
self,
span: Tuple[int, int],
tokens: List[str],
subword_to_word: Dict[int, int],
) -> str:
"""
Extract text from a span using tokens and subword-to-word mapping.
Args:
span: (start, end) subword indices (exclusive end)
tokens: List of word tokens
subword_to_word: Dict mapping subword idx -> word idx
Returns:
Text string for the span
"""
start, end = span
# Get word indices for this span
word_indices = set()
for subword_idx in range(start, end):
if subword_idx in subword_to_word:
word_indices.add(subword_to_word[subword_idx])
if not word_indices:
return ""
# Get contiguous word range
min_word = min(word_indices)
max_word = max(word_indices)
# Extract and join tokens
span_tokens = tokens[min_word:max_word + 1]
return " ".join(span_tokens)
def extract_qa_regions(
self,
token_logits: torch.Tensor, # (batch, seq, num_labels)
attention_mask: torch.Tensor, # (batch, seq)
) -> Tuple[List[List[Tuple[int, int]]], List[List[Tuple[int, int]]]]:
"""
Extract Q and A regions from Step 1 token predictions.
This enables the cascading architecture where Step 2 uses
Step 1's predictions instead of ground truth spans.
Returns:
q_regions: List of Q spans per batch item [(start, end), ...]
a_regions: List of A spans per batch item [(start, end), ...]
"""
batch_size = token_logits.shape[0]
preds = token_logits.argmax(dim=-1) # (batch, seq)
q_regions = []
a_regions = []
for b in range(batch_size):
sample_q_regions = []
sample_a_regions = []
seq_len = attention_mask[b].sum().item()
pred_seq = preds[b, :int(seq_len)].cpu().tolist()
# Extract Q spans (B-Q=1, I-Q=2)
current_span = None
for i, label in enumerate(pred_seq):
if label == 1: # B-Q
if current_span is not None:
sample_q_regions.append(current_span)
current_span = (i, i + 1)
elif label == 2: # I-Q
if current_span is not None:
current_span = (current_span[0], i + 1)
else:
if current_span is not None:
sample_q_regions.append(current_span)
current_span = None
if current_span is not None:
sample_q_regions.append(current_span)
# Extract A spans (B-A=3, I-A=4)
current_span = None
for i, label in enumerate(pred_seq):
if label == 3: # B-A
if current_span is not None:
sample_a_regions.append(current_span)
current_span = (i, i + 1)
elif label == 4: # I-A
if current_span is not None:
current_span = (current_span[0], i + 1)
else:
if current_span is not None:
sample_a_regions.append(current_span)
current_span = None
if current_span is not None:
sample_a_regions.append(current_span)
q_regions.append(sample_q_regions)
a_regions.append(sample_a_regions)
return q_regions, a_regions
def match_regions_to_gt(
self,
pred_regions: List[Tuple[int, int]],
gt_regions: List[Tuple[int, int]],
) -> Tuple[List[int], List[int]]:
"""
Match predicted regions to GT regions by overlap.
Returns:
gt_to_pred: For each GT region, index of best matching pred region (-1 if none)
pred_to_gt: For each pred region, index of best matching GT region (-1 if none)
"""
gt_to_pred = []
for gt_start, gt_end in gt_regions:
best_pred_idx = -1
best_overlap = 0
for pred_idx, (pred_start, pred_end) in enumerate(pred_regions):
overlap_start = max(gt_start, pred_start)
overlap_end = min(gt_end, pred_end)
overlap = max(0, overlap_end - overlap_start)
if overlap > best_overlap:
best_overlap = overlap
best_pred_idx = pred_idx
gt_to_pred.append(best_pred_idx)
pred_to_gt = []
for pred_start, pred_end in pred_regions:
best_gt_idx = -1
best_overlap = 0
for gt_idx, (gt_start, gt_end) in enumerate(gt_regions):
overlap_start = max(gt_start, pred_start)
overlap_end = min(gt_end, pred_end)
overlap = max(0, overlap_end - overlap_start)
if overlap > best_overlap:
best_overlap = overlap
best_gt_idx = gt_idx
pred_to_gt.append(best_gt_idx)
return gt_to_pred, pred_to_gt
def forward(
self,
input_ids: torch.Tensor,
attention_mask: torch.Tensor,
bbox: torch.Tensor,
tokens: Optional[List[List[str]]] = None,
subword_to_word: Optional[List[Dict[int, int]]] = None,
query_texts: Optional[List[str]] = None,
gt_answer_spans: Optional[List[List[Tuple[int, int]]]] = None,
gt_question_spans: Optional[List[List[Optional[Tuple[int, int]]]]] = None,
target_field_idx: Optional[List[int]] = None,
training: bool = True,
) -> Dict[str, torch.Tensor]:
"""
Forward pass with CASCADING architecture and ZERO-SHOT query matching.
Key features:
- Step 1: Token classification to predict Q/A regions
- Step 1.5: Q-A linking using LiLT embeddings + spatial features
- Step 2: ZERO-SHOT query→question matching using Sentence Transformer
- GT spans are only used to compute labels (which predicted region is correct)
Args:
input_ids: (batch, seq) token IDs
attention_mask: (batch, seq) attention mask
bbox: (batch, seq, 4) bounding boxes
tokens: List of word tokens per batch item (for zero-shot Q text encoding)
subword_to_word: List of dicts mapping subword idx to word idx
query_texts: List of raw query strings (for sentence transformer)
gt_answer_spans: GT answer spans per batch item (for loss labels only)
gt_question_spans: GT question spans per batch item (for loss labels only)
target_field_idx: Index of GT field that query should match (for loss labels)
training: whether in training mode (affects Q masking for Q-A linker)
"""
batch_size = input_ids.shape[0]
device = input_ids.device
# Encode document
hidden_states = self.encode(input_ids, attention_mask, bbox)
# Step 1: Token Classification
token_logits = self.token_classifier(hidden_states)
# Step 3: Table Head
table_outputs = self.table_head(hidden_states)
outputs = {
"hidden_states": hidden_states,
"token_logits": token_logits,
**table_outputs,
}
# Step 1.5 + Step 2: Q-A Linking and Query-Question Matching
# CASCADING: Extract Q/A regions from Step 1 PREDICTIONS
if query_texts is not None:
# Extract predicted Q and A regions from Step 1 output
pred_q_regions, pred_a_regions = self.extract_qa_regions(token_logits, attention_mask)
outputs["pred_q_regions"] = pred_q_regions
outputs["pred_a_regions"] = pred_a_regions
# Encode query using sentence transformer
query_emb = self.encode_texts(query_texts, device)
outputs["query_embedding"] = query_emb
# Process each sample in batch using PREDICTED regions
all_q_embeds_lilt = [] # Q embeddings from LiLT (768 dim) for QA Linker
all_q_embeds_st = [] # Q embeddings from sentence transformer (384 dim) for query matching
all_a_embeds = [] # A embeddings from predicted regions (768 dim)
all_q_bboxes = [] # Q bboxes for QA linker
all_a_bboxes = [] # A bboxes for QA linker
all_pred_q_from_a = [] # Predicted Q from A (for QuestionPredictor loss)
all_real_q_from_pred = [] # Real Q from predicted regions (for Q masking loss)
all_real_q_embeds = [] # Real Q embeds from GT (for aux loss)
all_gt_to_pred_a_idx = [] # Maps GT field idx to predicted A region idx
for b in range(batch_size):
sample_pred_q = pred_q_regions[b] # Predicted Q spans from Step 1
sample_pred_a = pred_a_regions[b] # Predicted A spans from Step 1
# Skip if no predicted regions
if len(sample_pred_a) == 0:
all_q_embeds_lilt.append(None)
all_q_embeds_st.append(None)
all_a_embeds.append(None)
all_q_bboxes.append(None)
all_a_bboxes.append(None)
all_pred_q_from_a.append(None)
all_real_q_from_pred.append(None)
all_real_q_embeds.append(None)
all_gt_to_pred_a_idx.append(None)
continue
# Pool A embeddings from predicted regions
sample_a_embeds = []
sample_a_bboxes = []
for a_span in sample_pred_a:
a_emb = self.pool_single_span(hidden_states[b], a_span, bbox[b])
sample_a_embeds.append(a_emb)
sample_a_bboxes.append(self.get_span_bbox(bbox[b], a_span))
# Two Q representations:
# 1. q_embeds_lilt (768) - LiLT pooled, for QA Linker
# 2. q_embeds_st (384) - Sentence transformer, for query matching
sample_q_embeds_lilt = [] # For QA Linker (768 dim)
sample_q_embeds_st = [] # For query matching (384 dim)
sample_q_bboxes = []
sample_real_q_from_pred = [] # Real Q embeds (ST) for QuestionPredictor loss
sample_pred_q_from_a = [] # Predicted Q embeds from A (for loss)
if len(sample_pred_q) > 0:
# We have predicted Q regions
# Get Q texts for sentence transformer encoding
q_texts = []
if tokens is not None and subword_to_word is not None:
for q_span in sample_pred_q:
q_text = self.extract_span_text(q_span, tokens[b], subword_to_word[b])
q_texts.append(q_text if q_text else "unknown")
# Batch encode all Q texts with sentence transformer
real_q_embeds_st = self.encode_texts(q_texts, device) # (num_q, 384)
else:
real_q_embeds_st = None
for i, q_span in enumerate(sample_pred_q):
# LiLT embedding for QA Linker
q_emb_lilt = self.pool_single_span(hidden_states[b], q_span, bbox[b])
sample_q_embeds_lilt.append(q_emb_lilt)
q_bbox = self.get_span_bbox(bbox[b], q_span)
sample_q_bboxes.append(q_bbox)
# Sentence transformer embedding for query matching
if real_q_embeds_st is not None:
real_q_emb_st = real_q_embeds_st[i] # 384 dim
else:
real_q_emb_st = None
# Q MASKING: during training, randomly mask Q and use predictor
should_mask = training and (torch.rand(1).item() < self.q_mask_prob)
if should_mask and i < len(sample_a_embeds):
# Mask: use QuestionPredictor instead of real Q for query matching
pred_q_emb = self.question_predictor(sample_a_embeds[i]) # 384 dim
sample_q_embeds_st.append(pred_q_emb)
sample_real_q_from_pred.append(real_q_emb_st) # Real Q for loss
sample_pred_q_from_a.append(pred_q_emb) # Predicted Q for loss
else:
# No mask: use real Q (sentence transformer encoded)
sample_q_embeds_st.append(real_q_emb_st if real_q_emb_st is not None else torch.zeros(self.query_embed_dim, device=device))
sample_real_q_from_pred.append(real_q_emb_st)
sample_pred_q_from_a.append(None) # No prediction, no loss
else:
# No Q regions predicted - use QuestionPredictor on all A
for a_emb in sample_a_embeds:
pred_q = self.question_predictor(a_emb) # 384 dim
sample_q_embeds_st.append(pred_q)
sample_pred_q_from_a.append(pred_q)
sample_real_q_from_pred.append(None) # No real Q from prediction
# No LiLT Q embeds when no Q predicted
sample_q_embeds_lilt = []
# Use A bboxes as proxy for Q bboxes
sample_q_bboxes = sample_a_bboxes.copy()
# Store real Q embeds from GT (for aux loss when Q was masked)
# Encode GT Q text with sentence transformer (384 dim)
sample_real_q = []
if gt_question_spans is not None and b < len(gt_question_spans) and tokens is not None and subword_to_word is not None:
gt_q_texts = []
gt_q_valid_indices = []
for idx, gt_q_span in enumerate(gt_question_spans[b]):
if gt_q_span is not None:
q_text = self.extract_span_text(gt_q_span, tokens[b], subword_to_word[b])
gt_q_texts.append(q_text if q_text else "unknown")
gt_q_valid_indices.append(idx)
# Batch encode GT Q texts
if gt_q_texts:
gt_q_embeds = self.encode_texts(gt_q_texts, device) # (num_valid, 384)
embed_idx = 0
for idx, gt_q_span in enumerate(gt_question_spans[b]):
if gt_q_span is not None:
sample_real_q.append(gt_q_embeds[embed_idx])
embed_idx += 1
else:
sample_real_q.append(None)
else:
sample_real_q = [None] * len(gt_question_spans[b])
# Match GT A spans to predicted A regions (for label computation)
gt_to_pred_a = None
if gt_answer_spans is not None and b < len(gt_answer_spans):
gt_a_spans = gt_answer_spans[b]
gt_to_pred_a, _ = self.match_regions_to_gt(sample_pred_a, gt_a_spans)
all_q_embeds_lilt.append(torch.stack(sample_q_embeds_lilt) if sample_q_embeds_lilt else None)
all_q_embeds_st.append(torch.stack(sample_q_embeds_st) if sample_q_embeds_st else None)
all_a_embeds.append(torch.stack(sample_a_embeds) if sample_a_embeds else None)
all_q_bboxes.append(torch.stack(sample_q_bboxes) if sample_q_bboxes else None)
all_a_bboxes.append(torch.stack(sample_a_bboxes) if sample_a_bboxes else None)
# For Q predictor loss: parallel lists with None for non-masked positions
all_pred_q_from_a.append(sample_pred_q_from_a if sample_pred_q_from_a else None)
all_real_q_from_pred.append(sample_real_q_from_pred if sample_real_q_from_pred else None)
all_real_q_embeds.append(sample_real_q if sample_real_q else None)
all_gt_to_pred_a_idx.append(gt_to_pred_a)
# Filter out None entries and compute outputs
# Use LiLT Q embeds for valid check (QA Linker needs them)
valid_indices_lilt = [i for i, q in enumerate(all_q_embeds_lilt) if q is not None]
# Also track valid ST Q embeds for query matching
valid_indices_st = [i for i, q in enumerate(all_q_embeds_st) if q is not None]
if valid_indices_lilt:
# Get max spans for padding (use LiLT Q for QA Linker)
max_q_spans_lilt = max(all_q_embeds_lilt[i].shape[0] for i in valid_indices_lilt)
max_a_spans = max(all_a_embeds[i].shape[0] for i in valid_indices_lilt)
# Create padded tensors for Q (LiLT, 768 dim) - for QA Linker
q_embeds_lilt_padded = torch.zeros(batch_size, max_q_spans_lilt, self.hidden_size, device=device)
q_bboxes_padded = torch.zeros(batch_size, max_q_spans_lilt, 4, device=device)
q_span_mask = torch.zeros(batch_size, max_q_spans_lilt, dtype=torch.bool, device=device)
# Create padded tensors for A (LiLT, 768 dim)
a_embeds_padded = torch.zeros(batch_size, max_a_spans, self.hidden_size, device=device)
a_bboxes_padded = torch.zeros(batch_size, max_a_spans, 4, device=device)
a_span_mask = torch.zeros(batch_size, max_a_spans, dtype=torch.bool, device=device)
for b in valid_indices_lilt:
nq = all_q_embeds_lilt[b].shape[0]
q_embeds_lilt_padded[b, :nq] = all_q_embeds_lilt[b]
q_bboxes_padded[b, :nq] = all_q_bboxes[b]
q_span_mask[b, :nq] = True
na = all_a_embeds[b].shape[0]
a_embeds_padded[b, :na] = all_a_embeds[b]
a_bboxes_padded[b, :na] = all_a_bboxes[b]
a_span_mask[b, :na] = True
# Step 1.5: Q-A Linker - predict which Q links to which A (uses LiLT embeds)
qa_link_scores_list = []
for b in valid_indices_lilt:
nq = all_q_embeds_lilt[b].shape[0]
na = all_a_embeds[b].shape[0]
if nq > 0 and na > 0:
link_scores = self.qa_linker(
all_q_embeds_lilt[b], # (num_q, 768) LiLT
all_a_embeds[b], # (num_a, 768) LiLT
all_q_bboxes[b], # (num_q, 4)
all_a_bboxes[b], # (num_a, 4)
)
qa_link_scores_list.append(link_scores)
else:
qa_link_scores_list.append(None)
outputs["qa_link_scores"] = qa_link_scores_list
# Create padded tensors for Q (sentence transformer, 384 dim) - for query matching
if valid_indices_st:
max_q_spans_st = max(all_q_embeds_st[i].shape[0] for i in valid_indices_st)
q_embeds_st_padded = torch.zeros(batch_size, max_q_spans_st, self.query_embed_dim, device=device)
q_span_mask_st = torch.zeros(batch_size, max_q_spans_st, dtype=torch.bool, device=device)
for b in valid_indices_st:
nq = all_q_embeds_st[b].shape[0]
q_embeds_st_padded[b, :nq] = all_q_embeds_st[b]
q_span_mask_st[b, :nq] = True
outputs["q_embeds_st"] = q_embeds_st_padded # For query matching
outputs["q_span_mask_st"] = q_span_mask_st
# GT-based QA link scores (for training - uses GT spans directly, LiLT embeddings)
if training and gt_question_spans is not None and gt_answer_spans is not None:
gt_qa_link_scores_list = []
gt_valid_q_indices_list = [] # Track which Q indices are valid (not None)
for b in range(batch_size):
gt_q_spans = gt_question_spans[b] if b < len(gt_question_spans) else []
gt_a_spans = gt_answer_spans[b] if b < len(gt_answer_spans) else []
# Filter valid Q spans (not None) and track indices
valid_q_indices = [i for i, q in enumerate(gt_q_spans) if q is not None]
valid_q_spans = [gt_q_spans[i] for i in valid_q_indices]
if len(valid_q_spans) > 0 and len(gt_a_spans) > 0:
# Pool embeddings from GT spans (LiLT, 768 dim)
gt_q_embeds = torch.stack([
self.pool_single_span(hidden_states[b], q_span, bbox[b])
for q_span in valid_q_spans
])
gt_a_embeds = torch.stack([
self.pool_single_span(hidden_states[b], a_span, bbox[b])
for a_span in gt_a_spans
])
gt_q_bboxes = torch.stack([
self.get_span_bbox(bbox[b], q_span)
for q_span in valid_q_spans
])
gt_a_bboxes = torch.stack([
self.get_span_bbox(bbox[b], a_span)
for a_span in gt_a_spans
])
# Compute QA link scores on GT
gt_link_scores = self.qa_linker(
gt_q_embeds, gt_a_embeds, gt_q_bboxes, gt_a_bboxes
)
gt_qa_link_scores_list.append(gt_link_scores)
gt_valid_q_indices_list.append(valid_q_indices)
else:
gt_qa_link_scores_list.append(None)
gt_valid_q_indices_list.append(None)
outputs["gt_qa_link_scores"] = gt_qa_link_scores_list
outputs["gt_valid_q_indices"] = gt_valid_q_indices_list
# Step 2: Query-Question Matching (ZERO-SHOT with Sentence Transformer)
# Use pre-computed q_embeds_st for query matching
if valid_indices_st and "q_embeds_st" in outputs:
q_embeds_st_padded = outputs["q_embeds_st"]
q_span_mask_st = outputs["q_span_mask_st"]
# Compute cosine similarity between query and Q embeddings
query_norm = F.normalize(query_emb, dim=-1)
q_norm = F.normalize(q_embeds_st_padded, dim=-1)
match_scores = torch.bmm(q_norm, query_norm.unsqueeze(-1)).squeeze(-1)
match_scores = match_scores.masked_fill(~q_span_mask_st, float('-inf'))
outputs["match_scores"] = match_scores
outputs["match_span_mask"] = q_span_mask_st
# Store LiLT embeddings for QA Linker
if valid_indices_lilt:
outputs["q_span_mask"] = q_span_mask
outputs["a_span_mask"] = a_span_mask
outputs["q_embeds"] = q_embeds_lilt_padded
outputs["a_embeds"] = a_embeds_padded
outputs["q_bboxes"] = q_bboxes_padded
outputs["a_bboxes"] = a_bboxes_padded
# Store for loss computation (Q masking / QuestionPredictor)
outputs["pred_q_from_a"] = all_pred_q_from_a
outputs["real_q_from_pred"] = all_real_q_from_pred # Real Q (ST) when masked
outputs["real_q_embeds"] = all_real_q_embeds # GT Q embeds (ST)
outputs["gt_to_pred_a_idx"] = all_gt_to_pred_a_idx
outputs["valid_batch_indices_lilt"] = valid_indices_lilt
outputs["valid_batch_indices_st"] = valid_indices_st
return outputs
def compute_loss(
outputs: Dict[str, torch.Tensor],
token_labels: torch.Tensor,
attention_mask: torch.Tensor,
table_labels: Optional[torch.Tensor] = None,
row_labels: Optional[torch.Tensor] = None,
header_labels: Optional[torch.Tensor] = None,
match_labels: Optional[torch.Tensor] = None,
col_labels: Optional[torch.Tensor] = None,
class_weights: Optional[torch.Tensor] = None,
qa_link_labels: Optional[List[torch.Tensor]] = None, # NEW: Q-A link labels
) -> Dict[str, torch.Tensor]:
"""Compute joint loss for all steps including Q-A linking."""
device = token_labels.device
losses = {}
# Step 1: Token Classification
token_logits = outputs["token_logits"]
token_logits_flat = token_logits.view(-1, token_logits.shape[-1])
token_labels_flat = token_labels.view(-1)
attn_flat = attention_mask.view(-1).bool()
if class_weights is not None:
step1_loss = F.cross_entropy(
token_logits_flat[attn_flat],
token_labels_flat[attn_flat],
weight=class_weights,
)
else:
step1_loss = F.cross_entropy(
token_logits_flat[attn_flat],
token_labels_flat[attn_flat],
)
losses["step1_loss"] = step1_loss
# Step 1.5: Q-A Linker Loss (uses GT spans directly - no dependency on Step 1 predictions)
if "gt_qa_link_scores" in outputs and qa_link_labels is not None:
qa_link_losses = []
gt_scores_list = outputs["gt_qa_link_scores"]
gt_valid_q_indices = outputs.get("gt_valid_q_indices", [None] * len(gt_scores_list))
for scores, labels, valid_indices in zip(gt_scores_list, qa_link_labels, gt_valid_q_indices):
if scores is None or labels is None or valid_indices is None:
continue
# Filter labels to only include valid Q indices (matching scores shape)
filtered_labels = labels[valid_indices] if len(valid_indices) > 0 else labels
if scores.numel() > 0 and filtered_labels.numel() > 0:
# scores: (num_valid_q, num_gt_a), filtered_labels: (num_valid_q,)
num_a = scores.shape[1]
valid_mask = (filtered_labels >= 0) & (filtered_labels < num_a)
if valid_mask.any():
link_loss = F.cross_entropy(scores[valid_mask], filtered_labels[valid_mask])
qa_link_losses.append(link_loss)
if qa_link_losses:
losses["qa_link_loss"] = torch.stack(qa_link_losses).mean()
else:
losses["qa_link_loss"] = torch.tensor(0.0, device=device)
else:
losses["qa_link_loss"] = torch.tensor(0.0, device=device)
# Question Predictor Loss (MSE between predicted Q and real Q in Sentence Transformer space)
# QuestionPredictor learns: A_lilt_embed (768) → Q_st_embed (384)
# Training: when Q is masked, compare predicted Q with real Q (sentence transformer)
if "pred_q_from_a" in outputs and "real_q_from_pred" in outputs:
pred_q_from_a = outputs["pred_q_from_a"] # List of lists
real_q_from_pred = outputs["real_q_from_pred"] # List of lists
q_predict_losses = []
for batch_pred, batch_real in zip(pred_q_from_a, real_q_from_pred):
if batch_pred is None or batch_real is None:
continue
for pred_q, real_q in zip(batch_pred, batch_real):
if pred_q is not None and real_q is not None:
# MSE loss to make predicted Q similar to real Q
mse = F.mse_loss(pred_q, real_q.detach())
q_predict_losses.append(mse)
if q_predict_losses:
losses["q_predict_loss"] = torch.stack(q_predict_losses).mean()
else:
losses["q_predict_loss"] = torch.tensor(0.0, device=device)
else:
losses["q_predict_loss"] = torch.tensor(0.0, device=device)
# Step 2: Query-Question Matching - ZERO-SHOT (no loss)
# Zero-shot Q text matching - no training needed
# QuestionPredictor IS trained (for A→Q prediction when no Q regions)
losses["step2_loss"] = torch.tensor(0.0, device=device)
# Step 3: Table Losses
if table_labels is not None:
table_logits = outputs["table_logits"]
table_logits_flat = table_logits.view(-1, 2)
table_labels_flat = table_labels.view(-1)
table_det_loss = F.cross_entropy(
table_logits_flat[attn_flat],
table_labels_flat[attn_flat],
)
losses["table_det_loss"] = table_det_loss
else:
losses["table_det_loss"] = torch.tensor(0.0, device=device)
if header_labels is not None and table_labels is not None:
header_logits = outputs["header_logits"]
table_mask = (table_labels == 1) & attention_mask.bool()
if table_mask.any():
header_logits_flat = header_logits.view(-1, 2)
header_labels_flat = header_labels.view(-1)
table_mask_flat = table_mask.view(-1)
header_loss = F.cross_entropy(
header_logits_flat[table_mask_flat],
header_labels_flat[table_mask_flat],
)
losses["header_loss"] = header_loss
else:
losses["header_loss"] = torch.tensor(0.0, device=device)
else:
losses["header_loss"] = torch.tensor(0.0, device=device)
if row_labels is not None and table_labels is not None:
row_logits = outputs["row_logits"]
table_mask = (table_labels == 1) & attention_mask.bool()
if table_mask.any():
row_logits_flat = row_logits.view(-1, 3)
row_labels_flat = row_labels.view(-1)
table_mask_flat = table_mask.view(-1)
row_loss = F.cross_entropy(
row_logits_flat[table_mask_flat],
row_labels_flat[table_mask_flat],
)
losses["row_loss"] = row_loss
else:
losses["row_loss"] = torch.tensor(0.0, device=device)
else:
losses["row_loss"] = torch.tensor(0.0, device=device)
if col_labels is not None and table_labels is not None and header_labels is not None:
col_scores = outputs["col_scores"]
cell_mask = (table_labels == 1) & (header_labels == 0) & attention_mask.bool()
if cell_mask.any():
col_scores_flat = col_scores.view(-1, col_scores.shape[-1])
col_labels_flat = col_labels.view(-1)
cell_mask_flat = cell_mask.view(-1)
col_loss = F.cross_entropy(
col_scores_flat[cell_mask_flat],
col_labels_flat[cell_mask_flat],
)
losses["col_loss"] = col_loss
else:
losses["col_loss"] = torch.tensor(0.0, device=device)
else:
losses["col_loss"] = torch.tensor(0.0, device=device)
# Total loss with weighted components
total_loss = (
losses["step1_loss"] + # Token classification (main task)
losses["qa_link_loss"] * 0.3 + # Q-A linker (reduced to not compete with step1)
losses["q_predict_loss"] * 0.2 + # Q prediction aux (reduced)
losses["step2_loss"] * 0.5 + # Query-Q matching (reduced)
losses["table_det_loss"] * 0.5 + # Table detection
losses["header_loss"] * 0.3 + # Header detection
losses["row_loss"] * 0.3 + # Row segmentation
losses["col_loss"] * 0.3 # Column assignment
)
losses["loss"] = total_loss
return losses
def compute_contrastive_loss(
query_embedding: torch.Tensor, # (batch, hidden)
span_embeddings: torch.Tensor, # (batch, num_spans, span_dim)
span_mask: torch.Tensor, # (batch, num_spans)
match_labels: torch.Tensor, # (batch,) - index of correct span
temperature: float = 0.07,
hard_negative_weight: float = 2.0,
) -> torch.Tensor:
"""
Contrastive loss for query-answer matching.
InfoNCE-style loss that:
- Pulls query towards correct answer span
- Pushes query away from all other spans (in-batch negatives)
- Applies extra weight to hard negatives (high-scoring wrong answers)
Args:
query_embedding: Query [CLS] embeddings
span_embeddings: Pooled span embeddings [h_start; h_end; bbox]
span_mask: Valid span mask
match_labels: Index of correct span for each query
temperature: Softmax temperature (lower = harder)
hard_negative_weight: Extra weight for hard negatives
Returns:
Contrastive loss scalar
"""
batch_size = query_embedding.shape[0]
device = query_embedding.device
if batch_size == 0:
return torch.tensor(0.0, device=device)
# Project spans to same dimension as query if needed
# (This is already done by QueryAnswerMatcher, so we use the projected scores)
# Here we compute a simpler version using the match_scores from forward
# For now, use a margin-based contrastive approach
# We want: sim(q, correct_span) > sim(q, wrong_span) + margin
losses = []
for b in range(batch_size):
valid_spans = span_mask[b].sum().item()
if valid_spans <= 1:
continue
correct_idx = match_labels[b].item()
if correct_idx >= valid_spans:
continue
# Get span embeddings for this sample
spans = span_embeddings[b, :int(valid_spans), :] # (num_valid, dim)
query = query_embedding[b] # (hidden,)
# Simple approach: normalize and compute similarities
# The projection happens in QueryAnswerMatcher, so we compute raw similarities here
# This loss is auxiliary to the CE loss
# Normalize for cosine similarity
spans_norm = F.normalize(spans[:, :query.shape[0]], dim=-1) # Use first hidden_size dims
query_norm = F.normalize(query, dim=-1)
# Compute similarities
sims = torch.matmul(spans_norm, query_norm) / temperature # (num_valid,)
# Create target (one-hot)
target = torch.zeros(int(valid_spans), device=device)
target[correct_idx] = 1.0
# InfoNCE: -log(exp(sim_pos) / sum(exp(sim_all)))
# With hard negative weighting
weights = torch.ones(int(valid_spans), device=device)
weights[correct_idx] = 1.0
# Hard negatives: spans with high similarity but wrong
with torch.no_grad():
hard_neg_mask = (sims > sims[correct_idx] - 0.5) & (torch.arange(int(valid_spans), device=device) != correct_idx)
weights[hard_neg_mask] = hard_negative_weight
# Weighted softmax cross-entropy
log_probs = F.log_softmax(sims, dim=0)
loss = -log_probs[correct_idx]
# Add margin loss for hard negatives
for i in range(int(valid_spans)):
if i != correct_idx and hard_neg_mask[i] if hard_neg_mask.any() else False:
margin_loss = F.relu(sims[i] - sims[correct_idx] + 0.3)
loss = loss + 0.1 * margin_loss
losses.append(loss)
if not losses:
return torch.tensor(0.0, device=device)
return torch.stack(losses).mean()
def compute_loss_with_contrastive(
outputs: Dict[str, torch.Tensor],
token_labels: torch.Tensor,
attention_mask: torch.Tensor,
table_labels: Optional[torch.Tensor] = None,
row_labels: Optional[torch.Tensor] = None,
header_labels: Optional[torch.Tensor] = None,
match_labels: Optional[torch.Tensor] = None,
col_labels: Optional[torch.Tensor] = None,
class_weights: Optional[torch.Tensor] = None,
qa_link_labels: Optional[List[torch.Tensor]] = None,
contrastive_weight: float = 0.5,
) -> Dict[str, torch.Tensor]:
"""
Compute joint loss with contrastive learning for Step 2.
Adds InfoNCE-style contrastive loss to help learn better span representations.
"""
# Get base losses
losses = compute_loss(
outputs=outputs,
token_labels=token_labels,
attention_mask=attention_mask,
table_labels=table_labels,
row_labels=row_labels,
header_labels=header_labels,
match_labels=match_labels,
col_labels=col_labels,
class_weights=class_weights,
qa_link_labels=qa_link_labels,
)
# Contrastive loss for Step 2 is DISABLED for zero-shot matching
# Zero-shot matching - no contrastive loss needed
losses["contrastive_loss"] = torch.tensor(0.0, device=token_labels.device)
return losses
if __name__ == "__main__":
print("Testing DEXTR LiLT model with CASCADING architecture...")
print("=" * 60)
model = DEXTRLiLT(
model_name="nielsr/lilt-xlm-roberta-base",
max_seq_len=1024,
)
total_params = sum(p.numel() for p in model.parameters())
trainable_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
print(f"Total parameters: {total_params:,}")
print(f"Trainable parameters: {trainable_params:,}")
# Test forward
batch_size = 2
seq_len = 128
query_len = 16
input_ids = torch.randint(0, 1000, (batch_size, seq_len))
attention_mask = torch.ones(batch_size, seq_len, dtype=torch.long)
# Create valid bboxes: [x1, y1, x2, y2] where x1<x2 and y1<y2, values in [0, 1000]
x1 = torch.randint(0, 500, (batch_size, seq_len))
y1 = torch.randint(0, 500, (batch_size, seq_len))
x2 = x1 + torch.randint(10, 500, (batch_size, seq_len))
y2 = y1 + torch.randint(10, 500, (batch_size, seq_len))
# Clamp to valid range [0, 1000]
bbox = torch.stack([x1, y1, x2.clamp(max=1000), y2.clamp(max=1000)], dim=-1)
query_input_ids = torch.randint(0, 1000, (batch_size, query_len))
query_attention_mask = torch.ones(batch_size, query_len, dtype=torch.long)
# GT spans (only used for loss labels in cascading architecture)
gt_answer_spans = [[(5, 10), (20, 25)], [(8, 12)]]
gt_question_spans = [[(2, 5), (17, 20)], [(5, 8)]]
print("\nRunning forward pass (CASCADING architecture)...")
print(" - Step 1: Token classification → predict Q/A/H/TABLE/O")
print(" - Step 1.5: Extract predicted regions, QALinker pairs them")
print(" - Step 2: Query matches to predicted Q embeddings")
model.eval()
with torch.no_grad():
outputs = model(
input_ids=input_ids,
attention_mask=attention_mask,
bbox=bbox,
query_input_ids=query_input_ids,
query_attention_mask=query_attention_mask,
gt_answer_spans=gt_answer_spans,
gt_question_spans=gt_question_spans,
)
print(f"\nOutput shapes:")
print(f" token_logits: {outputs['token_logits'].shape}")
print(f" table_logits: {outputs['table_logits'].shape}")
print(f" header_logits: {outputs['header_logits'].shape}")
print(f" row_logits: {outputs['row_logits'].shape}")
print(f" col_scores: {outputs['col_scores'].shape}")
print(f" query_embedding: {outputs['query_embedding'].shape}")
# Predicted regions from Step 1
print(f"\nPredicted regions from Step 1:")
print(f" pred_q_regions: {outputs['pred_q_regions']}")
print(f" pred_a_regions: {outputs['pred_a_regions']}")
# Match scores may not exist if no regions predicted
if "match_scores" in outputs:
print(f" match_scores: {outputs['match_scores'].shape}")
else:
print(" match_scores: None (no regions predicted)")
# Test loss
print("\nTesting loss computation...")
token_labels = torch.randint(0, NUM_LABELS, (batch_size, seq_len))
table_labels = torch.randint(0, 2, (batch_size, seq_len))
row_labels = torch.randint(0, 3, (batch_size, seq_len))
header_labels = torch.randint(0, 2, (batch_size, seq_len))
# For cascading: match_labels should point to predicted Q region index
# This would normally be computed by matching GT fields to predicted regions
match_labels = None
if "match_scores" in outputs and outputs["match_scores"].shape[1] > 0:
match_labels = torch.zeros(batch_size, dtype=torch.long)
losses = compute_loss(
outputs,
token_labels=token_labels,
attention_mask=attention_mask,
table_labels=table_labels,
row_labels=row_labels,
header_labels=header_labels,
match_labels=match_labels,
)
print(f"\nLosses:")
for name, value in losses.items():
print(f" {name}: {value.item():.4f}")
print("\n" + "=" * 60)
print("DEXTR LiLT CASCADING architecture test passed!")
|