File size: 94,054 Bytes
6247296 |
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 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 |
# coding=utf-8
# Copyright 2024 The HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import copy
import math
import warnings
import zlib
from typing import Callable, Iterator, List, Optional, Tuple, Union
import numpy as np
import torch
import torch.nn.functional as F
from torch import nn
from transformers.cache_utils import EncoderDecoderCache
from transformers.generation.configuration_utils import GenerationConfig
from transformers.generation.logits_process import (
LogitsProcessorList,
SuppressTokensAtBeginLogitsProcessor,
SuppressTokensLogitsProcessor,
WhisperNoSpeechDetection,
WhisperTimeStampLogitsProcessor,
)
from transformers.generation.stopping_criteria import StoppingCriteriaList
from transformers.modeling_outputs import BaseModelOutput
from transformers.utils import logging
from transformers.models.whisper.tokenization_whisper import TASK_IDS, TO_LANGUAGE_CODE
logger = logging.get_logger(__name__)
def _median_filter(inputs: torch.Tensor, filter_width: int) -> torch.Tensor:
"""
Applies a median filter of width `filter_width` along the last dimension of the input.
The `inputs` tensor is assumed to be 3- or 4-dimensional.
"""
if filter_width <= 0 or filter_width % 2 != 1:
raise ValueError("`filter_width` should be an odd number")
pad_width = filter_width // 2
if inputs.shape[-1] <= pad_width:
return inputs
# Pad the left and right edges.
inputs = nn.functional.pad(inputs, (pad_width, pad_width, 0, 0), mode="reflect")
# sort() is faster than torch.median (https://github.com/pytorch/pytorch/issues/51450)
result = inputs.unfold(-1, filter_width, 1).sort()[0][..., pad_width]
return result
def _dynamic_time_warping(matrix: np.ndarray):
"""
Measures similarity between two temporal sequences: the input audio and the output tokens. Used to generate
token-level timestamps.
"""
output_length, input_length = matrix.shape
cost = np.ones((output_length + 1, input_length + 1), dtype=np.float32) * np.inf
trace = -np.ones((output_length + 1, input_length + 1), dtype=np.float32)
cost[0, 0] = 0
for j in range(1, input_length + 1):
for i in range(1, output_length + 1):
c0 = cost[i - 1, j - 1]
c1 = cost[i - 1, j]
c2 = cost[i, j - 1]
if c0 < c1 and c0 < c2:
c, t = c0, 0
elif c1 < c0 and c1 < c2:
c, t = c1, 1
else:
c, t = c2, 2
cost[i, j] = matrix[i - 1, j - 1] + c
trace[i, j] = t
# backtrace
i = trace.shape[0] - 1
j = trace.shape[1] - 1
trace[0, :] = 2
trace[:, 0] = 1
text_indices = []
time_indices = []
while i > 0 or j > 0:
text_indices.append(i - 1)
time_indices.append(j - 1)
if trace[i, j] == 0:
i -= 1
j -= 1
elif trace[i, j] == 1:
i -= 1
elif trace[i, j] == 2:
j -= 1
else:
raise RuntimeError(
f"Internal error in dynamic time warping. Unexpected trace[{i}, {j}]. Please file a bug report."
)
text_indices = np.array(text_indices)[::-1]
time_indices = np.array(time_indices)[::-1]
return text_indices, time_indices
def _get_attr_from_logit_processors(logits_processor, logit_processor_class, attribute_name):
if logits_processor is not None:
logit_processor = next((cls for cls in logits_processor if isinstance(cls, logit_processor_class)), None)
if logit_processor:
return getattr(logit_processor, attribute_name, None)
return None
def _pad_to_max_length(
current_segments,
pad_token_id,
device,
padding_side="right",
padding="longest",
bos_token_tensor=None,
cut_off_length=None,
):
max_total_length = 0
sequences = []
if padding_side not in ["right", "left"]:
raise ValueError(f"`padding_side` must be either 'right' or 'left', not {padding_side}")
if padding not in ["longest", "max_length"]:
raise ValueError(f"`padding` must be either 'longest' or 'max_length', not {padding}")
elif padding == "max_length" and cut_off_length is None:
raise ValueError("`cut_off_length` must be specified when `padding='max_length'`")
for current_segment_list in current_segments:
if current_segment_list is not None and len([d["tokens"] for d in current_segment_list]) > 0:
sequence = torch.cat([d["tokens"] for d in current_segment_list], dim=-1)
if cut_off_length is not None:
sequence = sequence[-cut_off_length:]
if bos_token_tensor is not None:
sequence = torch.cat([bos_token_tensor, sequence])
sequences.append(sequence)
max_total_length = max(max_total_length, len(sequences[-1]))
elif bos_token_tensor is not None:
sequences.append(bos_token_tensor)
else:
sequences.append(torch.tensor([], device=device))
max_total_length = cut_off_length + 1 if padding == "max_length" else max_total_length
for i in range(len(current_segments)):
pad_length = max_total_length - len(sequences[i])
pad = (0, pad_length) if padding_side == "right" else (pad_length, 0)
sequences[i] = F.pad(sequences[i], pad=pad, value=pad_token_id)
sequences = torch.stack(sequences, dim=0)
return sequences
class WhisperGenerationMixin:
def _extract_token_timestamps(self, generate_outputs, alignment_heads, time_precision=0.02, num_frames=None):
"""
Calculates token-level timestamps using the encoder-decoder cross-attentions and dynamic time-warping (DTW) to
map each output token to a position in the input audio. If `num_frames` is specified, the encoder-decoder
cross-attentions will be cropped before applying DTW.
Returns:
tensor containing the timestamps in seconds for each predicted token
"""
# Create a list with `decoder_layers` elements, each a tensor of shape
# (batch size, attention_heads, output length, input length).
cross_attentions = []
for i in range(self.config.decoder_layers):
cross_attentions.append(torch.cat([x[i] for x in generate_outputs.cross_attentions], dim=2))
# Select specific cross-attention layers and heads. This is a tensor
# of shape (batch size, num selected, output length, input length).
weights = torch.stack([cross_attentions[l][:, h] for l, h in alignment_heads])
weights = weights.permute([1, 0, 2, 3])
weight_length = None
if "beam_indices" in generate_outputs:
# If beam search has been used, the output sequences may have been generated for more timesteps than their sequence_lengths
# since the beam search strategy chooses the most probable sequences at the end of the search.
# In that case, the cross_attentions weights are too long and we have to make sure that they have the right output_length
weight_length = (generate_outputs.beam_indices != -1).sum(-1).max()
weights = weights[:, :, :weight_length]
# If beam index is still -1, it means that the associated token id is EOS
# We need to replace the index with 0 since index_select gives an error if any of the indexes is -1.
beam_indices = generate_outputs.beam_indices[:, :weight_length]
beam_indices = beam_indices.masked_fill(beam_indices == -1, 0)
# Select the cross attention from the right beam for each output sequences
weights = torch.stack(
[
torch.index_select(weights[:, :, i, :], dim=0, index=beam_indices[:, i])
for i in range(beam_indices.shape[1])
],
dim=2,
)
# make sure timestamps are as long as weights
input_length = weight_length or cross_attentions[0].shape[2]
timestamps = torch.zeros_like(generate_outputs.sequences, dtype=torch.float32)[:, : input_length + 1]
batch_size = timestamps.shape[0]
if num_frames is not None:
# two cases:
# 1. num_frames is the same for each sample -> compute the DTW matrix for each sample in parallel
# 2. num_frames is different, compute the DTW matrix for each sample sequentially
# we're using np.unique because num_frames can be int/list/tuple
if isinstance(num_frames, int):
weights = weights[..., : num_frames // 2]
elif isinstance(num_frames, (list, tuple, np.ndarray)) and len(np.unique(num_frames)) == 1:
weights = weights[..., : num_frames[0] // 2]
elif isinstance(num_frames, (torch.Tensor)) and len(torch.unique(num_frames)) == 1:
weights = weights[..., : num_frames[0] // 2]
else:
# num_frames is of shape (batch_size,) whereas batch_size is truely batch_size*num_return_sequences
repeat_time = batch_size if isinstance(num_frames, int) else batch_size // len(num_frames)
num_frames = np.repeat(num_frames, repeat_time)
if num_frames is None or isinstance(num_frames, int):
# Normalize and smoothen the weights.
std = torch.std(weights, dim=-2, keepdim=True, unbiased=False)
mean = torch.mean(weights, dim=-2, keepdim=True)
weights = (weights - mean) / std
weights = _median_filter(weights, self.config.median_filter_width)
# Average the different cross-attention heads.
weights = weights.mean(dim=1)
# Perform dynamic time warping on each element of the batch.
for batch_idx in range(batch_size):
if num_frames is not None and isinstance(num_frames, (tuple, list, np.ndarray, torch.Tensor)):
matrix = weights[batch_idx, ..., : num_frames[batch_idx] // 2]
# Normalize and smoothen the weights.
std = torch.std(matrix, dim=-2, keepdim=True, unbiased=False)
mean = torch.mean(matrix, dim=-2, keepdim=True)
matrix = (matrix - mean) / std
matrix = _median_filter(matrix, self.config.median_filter_width)
# Average the different cross-attention heads.
matrix = matrix.mean(dim=0)
else:
matrix = weights[batch_idx]
text_indices, time_indices = _dynamic_time_warping(-matrix.cpu().double().numpy())
jumps = np.pad(np.diff(text_indices), (1, 0), constant_values=1).astype(bool)
jump_times = time_indices[jumps] * time_precision
timestamps[batch_idx, 1:] = torch.tensor(jump_times)
return timestamps
def generate(
self,
input_features: Optional[torch.Tensor] = None,
generation_config: Optional[GenerationConfig] = None,
logits_processor: Optional[LogitsProcessorList] = None,
stopping_criteria: Optional[StoppingCriteriaList] = None,
prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None,
synced_gpus: bool = False,
return_timestamps: Optional[bool] = None,
task: Optional[str] = None,
language: Optional[Union[str, List[str]]] = None,
is_multilingual: Optional[bool] = None,
prompt_ids: Optional[torch.Tensor] = None,
prompt_condition_type: Optional[str] = None, # first-segment, all-segments
condition_on_prev_tokens: Optional[bool] = None,
temperature: Optional[Union[float, Tuple[float, ...]]] = None,
compression_ratio_threshold: Optional[float] = None,
logprob_threshold: Optional[float] = None,
no_speech_threshold: Optional[float] = None,
num_segment_frames: Optional[int] = None,
attention_mask: Optional[torch.Tensor] = None,
time_precision: float = 0.02,
return_token_timestamps: Optional[bool] = None,
return_segments: bool = False,
return_dict_in_generate: Optional[bool] = None,
**kwargs,
):
"""
Transcribes or translates log-mel input features to a sequence of auto-regressively generated token ids.
<Tip warning={true}>
Most generation-controlling parameters are set in `generation_config` which, if not passed, will be set to the
model's default generation configuration. You can override any `generation_config` by passing the corresponding
parameters to generate(), e.g. `.generate(inputs, num_beams=4, do_sample=True)`.
For an overview of generation strategies and code examples, check out the [following
guide](./generation_strategies).
</Tip>
Parameters:
input_features (`torch.Tensor` of shape `(batch_size, feature_size, sequence_length)`, *optional*):
Float values of log-mel features extracted from the raw speech waveform. The raw speech waveform can be obtained by
loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via
the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
[`AutoFeatureExtractor`] should be used for extracting the mel features, padding and conversion into a
tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`] for details.
generation_config (`~generation.GenerationConfig`, *optional*):
The generation configuration to be used as base parametrization for the generation call. `**kwargs`
passed to generate matching the attributes of `generation_config` will override them. If
`generation_config` is not provided, the default will be used, which had the following loading
priority: 1) from the `generation_config.json` model file, if it exists; 2) from the model
configuration. Please note that unspecified parameters will inherit [`~generation.GenerationConfig`]'s
default values, whose documentation should be checked to parameterize generation.
logits_processor (`LogitsProcessorList`, *optional*):
Custom logits processors that complement the default logits processors built from arguments and
generation config. If a logit processor is passed that is already created with the arguments or a
generation config an error is thrown. This feature is intended for advanced users.
stopping_criteria (`StoppingCriteriaList`, *optional*):
Custom stopping criteria that complement the default stopping criteria built from arguments and a
generation config. If a stopping criteria is passed that is already created with the arguments or a
generation config an error is thrown. This feature is intended for advanced users.
prefix_allowed_tokens_fn (`Callable[[int, torch.Tensor], List[int]]`, *optional*):
If provided, this function constraints the beam search to allowed tokens only at each step. If not
provided no constraint is applied. This function takes 2 arguments: the batch ID `batch_id` and
`input_ids`. It has to return a list with the allowed tokens for the next generation step conditioned
on the batch ID `batch_id` and the previously generated tokens `inputs_ids`. This argument is useful
for constrained generation conditioned on the prefix, as described in [Autoregressive Entity
Retrieval](https://arxiv.org/abs/2010.00904).
synced_gpus (`bool`, *optional*, defaults to `False`):
Whether to continue running the while loop until max_length (needed for ZeRO stage 3)
return_timestamps (`bool`, *optional*):
Whether to return the timestamps with the text. This enables the `WhisperTimestampsLogitsProcessor`.
task (`str`, *optional*):
Task to use for generation, either "translate" or "transcribe". The `model.config.forced_decoder_ids`
will be updated accordingly.
language (`str` or list of `str`, *optional*):
Language token to use for generation, can be either in the form of `<|en|>`, `en` or `english`. For
batched generation, a list of language tokens can be passed. You can find all the possible language
tokens in the `model.generation_config.lang_to_id` dictionary.
is_multilingual (`bool`, *optional*):
Whether or not the model is multilingual.
prompt_ids (`torch.Tensor`, *optional*):
Rank-1 tensor of token IDs created by passing text to [`~WhisperProcessor.get_prompt_ids`] that is
provided as a prompt to each chunk. This can be used to provide or "prompt-engineer" a context for
transcription, e.g. custom vocabularies or proper nouns to make it more likely to predict those words
correctly. It cannot be used in conjunction with `decoder_start_token_id` as it overwrites this value.
prompt_condition_type (`str`, *optional*):
Only relevant for long-form transcription. Condition type of `prompt_ids`. 'first-segment' means only the first segment is conditioned on `prompt_ids`. 'all-segments' means each segment is conditioned on `prompt_ids`. Make sure to enable `condition_on_prev_tokens` for 'all-segments'.
Defaults to 'first-segment'. For short-term transcription only 'first-segment' is possible.
condition_on_prev_tokens (`bool`, *optional*):
Only relevant for long-form transcription. Whether to condition each segment on the previous segment.
As shown in the [the Whisper paper](https://cdn.openai.com/papers/whisper.pdf), this can help to improve
performance.
temperature (`float` or list of `float`, *optional*):
The temperature to be used for generation. Passing a single `float` value and `do_sample=True` activates
generation using sampling. For long-form transcription, temperature fallback can be activated by passing
a list of float values such as (0.0, 0.2, 0.4, 0.6, 0.8, 1.0). As shown in the [the Whisper paper](https://cdn.openai.com/papers/whisper.pdf), this can help to improve
performance.
compression_ratio_threshold (`float`, *optional*):
Only relevant for long-form transcription. If defined, the zlib compression rate of each segment will be computed. If the compression rate of
a segment is higher than `compression_ratio_threshold`, temperature fallback is activated: the generated segment is discarded and the generation is
repeated using a higher temperature. The intuition behind this feature is that segments with very high compression rates
suffer from a lot of repetition. The unwanted repetition can be reduced by injecting more randomness by increasing the temperature. If `compression_ratio_threshold` is defined
make sure that `temperature` is a list of values. A common value for `compression_ratio_threshold` is 1.35.
As shown in the [the Whisper paper](https://cdn.openai.com/papers/whisper.pdf), this can help to improve
performance.
logprob_threshold (`float`, *optional*):
Only relevant for long-form transcription. If defined, the average log-probability of each segment will be computed. If the log-probability of
a given segment is lower than `logprob_threshold`, temperature fallback is activated: the generated segment is discarded and the generation is
repeated using a higher temperature. The intuition behind this feature is that segments of low log-probability
can be improved by injecting more randomness by increasing the temperature. If `logprob_threshold` is defined
make sure that `temperature` is a list of values. A common value for `logprob_threshold` is -1.0.
As shown in the [the Whisper paper](https://cdn.openai.com/papers/whisper.pdf), this can help to improve
performance.
no_speech_threshold (`float`, *optional*):
Only relevant for long-form transcription. If defined, the "no-speech" token combined with the `logprob_threshold`
is used to determine whether a segment contains only silence. In this case, the transcription for this segment
is skipped.
As shown in the [the Whisper paper](https://cdn.openai.com/papers/whisper.pdf), this can help to improve
performance.
num_segment_frames (`int`, *optional*):
The number of frames a single segment is made of. If not defined, `num_segment_frames` defaults to the model's stride
times the maximum input length.
attention_mask (`torch.Tensor`, *optional*):
`attention_mask` needs to be passed when doing long-form transcription using a batch size > 1.
time_precision (`int`, *optional*, defaults to 0.02):
The duration of output token in seconds. *E.g.* 0.02 means that a generated token on average accounts
for 20 ms.
return_token_timestamps (`bool`, *optional*):
Whether to return token-level timestamps with the text. This can be used with or without the
`return_timestamps` option. To get word-level timestamps, use the tokenizer to group the tokens into
words.
return_segments (`bool`, *optional*, defaults to `False`):
Whether to additionally return a list of all segments. Note that this option can only be enabled
when doing long-form transcription.
return_dict_in_generate (`bool`, *optional*, defaults to `False`):
Whether or not to return a [`~utils.ModelOutput`] instead of just returning the generated tokens.
Note that when doing long-form transcription, `return_dict_in_generate` can only be enabled when
`return_segments` is set True. In this case the generation outputs of each segment is added to each
segment.
kwargs (`Dict[str, Any]`, *optional*):
Ad hoc parametrization of `generate_config` and/or additional model-specific kwargs that will be
forwarded to the `forward` function of the model. If the model is an encoder-decoder model, encoder
specific kwargs should not be prefixed and decoder specific kwargs should be prefixed with *decoder_*.
Return:
[`~utils.ModelOutput`] or `torch.LongTensor` or `Dict[str, Any]`: A [`~utils.ModelOutput`] (if `return_dict_in_generate=True`
or when `config.return_dict_in_generate=True`) or a `torch.FloatTensor` or a dict of segments when `return_segments=True`.
If the passed input is > 30 seconds / > 3000 mel input features and `return_segments=True` then a dictionary of generated sequence ids, called `sequences` and a list of each generated segment is returned.
else if the passed input is <= 30 seconds / >= 3000 mel input features, the possible [`~utils.ModelOutput`] types are:
- [`~generation.GenerateEncoderDecoderOutput`],
- [`~generation.GenerateBeamEncoderDecoderOutput`]
else only the generated output sequence ids are returned.
Example:
- *Longform transcription*: To transcribe or translate audios longer than 30 seconds, process the audio files without truncation and pass all mel features at once to generate.
```python
>>> import torch
>>> from transformers import AutoProcessor, WhisperForConditionalGeneration
>>> from datasets import load_dataset, Audio
>>> processor = AutoProcessor.from_pretrained("openai/whisper-tiny.en")
>>> model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")
>>> model.cuda() # doctest: +IGNORE_RESULT
>>> # load audios > 30 seconds
>>> ds = load_dataset("distil-whisper/meanwhile", "default")["test"]
>>> # resample to 16kHz
>>> ds = ds.cast_column("audio", Audio(sampling_rate=16000))
>>> # take first 8 audios and retrieve array
>>> audio = ds[:8]["audio"]
>>> audio = [x["array"] for x in audio]
>>> # make sure to NOT truncate the input audio, to return the `attention_mask` and to pad to the longest audio
>>> inputs = processor(audio, return_tensors="pt", truncation=False, padding="longest", return_attention_mask=True, sampling_rate=16_000)
>>> inputs = inputs.to("cuda", torch.float32)
>>> # transcribe audio to ids
>>> generated_ids = model.generate(**inputs)
>>> transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)
>>> transcription[0]
" Folks, if you watch the show, you know, I spent a lot of time right over there. Patiently and astutely scrutinizing the boxwood and mahogany chest set of the day's biggest stories developing the central headline pawns, definitely maneuvering an oso topical night to F6, fainting a classic Sicilian, nade door variation on the news, all the while seeing eight moves deep and patiently marshalling the latest press releases into a fisher's shows in Lip Nitsky attack that culminates in the elegant lethal slow-played, all-passant checkmate that is my nightly monologue. But sometimes, sometimes, folks, I. CHEERING AND APPLAUSE Sometimes I startle away, cubside down in the monkey bars of a condemned playground on a super fun site. Get all hept up on goofballs. Rummage that were discarded tag bag of defective toys. Yank out a fist bowl of disembodied doll limbs, toss them on a stained kid's place mat from a defunct dennies. set up a table inside a rusty cargo container down by the Wharf and challenged toothless drifters to the godless bughouse blitz of tournament that is my segment. Meanwhile."
```
- *Shortform transcription*: If passed mel input features are < 30 seconds, the whole audio will be transcribed with a single call to generate.
```python
>>> import torch
>>> from transformers import AutoProcessor, WhisperForConditionalGeneration
>>> from datasets import load_dataset
>>> processor = AutoProcessor.from_pretrained("openai/whisper-tiny.en")
>>> model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")
>>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
>>> inputs = processor(ds[0]["audio"]["array"], return_tensors="pt")
>>> input_features = inputs.input_features
>>> generated_ids = model.generate(inputs=input_features)
>>> transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
>>> transcription
' Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.'
```
"""
# 0. deprecate old inputs
if "inputs" in kwargs:
input_features = kwargs.pop("inputs")
warnings.warn(
"The input name `inputs` is deprecated. Please make sure to use `input_features` instead.",
FutureWarning,
)
# 1. prepare generation config
generation_config, kwargs = self._prepare_generation_config(generation_config, **kwargs)
# 2. set global generate variables
input_stride = self.model.encoder.conv1.stride[0] * self.model.encoder.conv2.stride[0]
num_segment_frames = input_stride * self.config.max_source_positions
batch_size, total_input_frames = self._retrieve_total_input_frames(
input_features=input_features, input_stride=input_stride, kwargs=kwargs
)
is_shortform = total_input_frames <= num_segment_frames
# 3. Make sure generation config is correctly set
# Make sure the generation config is correctly set depending on whether timestamps are to be returned or not
return_dict_in_generate = self._set_return_outputs(
return_dict_in_generate=return_dict_in_generate,
return_token_timestamps=return_token_timestamps,
logprob_threshold=logprob_threshold,
generation_config=generation_config,
)
timestamp_begin = self._set_return_timestamps(
return_timestamps=return_timestamps, is_shortform=is_shortform, generation_config=generation_config
)
self._set_language_and_task(
language=language, task=task, is_multilingual=is_multilingual, generation_config=generation_config
)
self._set_num_frames(
return_token_timestamps=return_token_timestamps, generation_config=generation_config, kwargs=kwargs
)
self._set_thresholds_and_condition(
generation_config=generation_config,
logprob_threshold=logprob_threshold,
compression_ratio_threshold=compression_ratio_threshold,
no_speech_threshold=no_speech_threshold,
condition_on_prev_tokens=condition_on_prev_tokens,
)
self._set_prompt_condition_type(
generation_config=generation_config,
prompt_condition_type=prompt_condition_type,
)
kwargs["attention_mask"] = attention_mask
# pass self.config for backward compatibility
init_tokens = self._retrieve_init_tokens(
input_features,
batch_size=batch_size,
generation_config=generation_config,
config=self.config,
num_segment_frames=num_segment_frames,
kwargs=kwargs,
)
# passing `decoder_input_ids` is deprecated - the only exception is for assisted generation
# where the input ids are handled explicitly by the generate method
self._check_decoder_input_ids(kwargs=kwargs)
# 3. Retrieve logits processors
device = kwargs["encoder_outputs"][0].device if "encoder_outputs" in kwargs else input_features.device
begin_index = init_tokens.shape[1]
logits_processor = self._retrieve_logit_processors(
generation_config=generation_config,
logits_processor=logits_processor,
begin_index=begin_index, # begin index is index of first generated decoder token
num_beams=kwargs.get("num_beams", 1),
device=device,
)
# 4 Set and retrieve global generation variables
self._set_condition_on_prev_tokens(
condition_on_prev_tokens=condition_on_prev_tokens, generation_config=generation_config
)
temperatures = [temperature] if not isinstance(temperature, (list, tuple)) else temperature
temperature = temperatures[0]
max_frames, seek = self._retrieve_max_frames_and_seek(
batch_size=batch_size,
attention_mask=attention_mask,
total_input_frames=total_input_frames,
is_shortform=is_shortform,
)
# 5 Prepare running variables, list for generation
num_return_sequences = generation_config.num_return_sequences
(
batch_idx_map,
cur_bsz,
input_features,
seek,
max_frames,
init_tokens,
do_condition_on_prev_tokens,
) = self._expand_variables_for_generation(
input_features=input_features,
seek=seek,
max_frames=max_frames,
init_tokens=init_tokens,
batch_size=batch_size,
condition_on_prev_tokens=condition_on_prev_tokens,
generation_config=generation_config,
)
current_segments = self._prepare_segments(
prompt_ids=prompt_ids,
batch_size=cur_bsz,
generation_config=generation_config,
)
# 6 Transcribe audio until we reach the end of all input audios
while (seek < max_frames).any():
# 6.1 NOTE: When in longform transcription mode and batch size > 1 we need to dynamically reduce the batch size during the loop
# in case one audio finished earlier than another one. Thus, we need to keep a table of "previous-index-2-current-index" in order
# to know which original audio is being decoded
# Set updated index map, duration of previously decoded chunks and number of max frames of current decoding chunk
input_features, cur_bsz, batch_idx_map = self._maybe_reduce_batch(
input_features=input_features,
seek=seek,
max_frames=max_frames,
cur_bsz=cur_bsz,
batch_idx_map=batch_idx_map,
)
time_offset = seek * time_precision / input_stride
seek_num_frames = (max_frames - seek).clamp(max=num_segment_frames)
# 6.2 cut out next 30s segment from input features
segment_input = self._get_input_segment(
input_features=input_features,
seek=seek,
seek_num_frames=seek_num_frames,
num_segment_frames=num_segment_frames,
cur_bsz=cur_bsz,
batch_idx_map=batch_idx_map,
)
# 6.3 prepare decoder input ids
suppress_tokens = _get_attr_from_logit_processors(
logits_processor, SuppressTokensLogitsProcessor, "suppress_tokens"
)
decoder_input_ids, kwargs = self._prepare_decoder_input_ids(
cur_bsz=cur_bsz,
init_tokens=init_tokens,
current_segments=current_segments,
batch_idx_map=batch_idx_map,
do_condition_on_prev_tokens=do_condition_on_prev_tokens,
prompt_ids=prompt_ids,
generation_config=generation_config,
config=self.config,
device=init_tokens.device,
suppress_tokens=suppress_tokens,
kwargs=kwargs,
)
# 6.4 set max new tokens or max length
self._set_max_new_tokens_and_length(
config=self.config,
decoder_input_ids=decoder_input_ids,
generation_config=generation_config,
)
# 6.5 Set current `begin_index` for all logit processors
if logits_processor is not None:
for proc in logits_processor:
if hasattr(proc, "set_begin_index"):
proc.set_begin_index(decoder_input_ids.shape[-1])
# 6.6 Run generate with fallback
(
seek_sequences,
seek_outputs,
should_skip,
do_condition_on_prev_tokens,
model_output_type,
) = self.generate_with_fallback(
segment_input=segment_input,
decoder_input_ids=decoder_input_ids,
cur_bsz=cur_bsz,
batch_idx_map=batch_idx_map,
seek=seek,
num_segment_frames=num_segment_frames,
max_frames=max_frames,
temperatures=temperatures,
generation_config=generation_config,
logits_processor=logits_processor,
stopping_criteria=stopping_criteria,
prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
synced_gpus=synced_gpus,
return_token_timestamps=return_token_timestamps,
do_condition_on_prev_tokens=do_condition_on_prev_tokens,
is_shortform=is_shortform,
batch_size=batch_size,
kwargs=kwargs,
)
# 6.7 In every generated sequence, split by timestamp tokens and extract segments
for i, seek_sequence in enumerate(seek_sequences):
prev_i = batch_idx_map[i]
if should_skip[i]:
seek[prev_i] += seek_num_frames[prev_i]
continue
segments, segment_offset = self._retrieve_segment(
seek_sequence=seek_sequence,
seek_outputs=seek_outputs,
time_offset=time_offset,
timestamp_begin=timestamp_begin,
seek_num_frames=seek_num_frames,
time_precision=time_precision,
input_stride=input_stride,
prev_idx=prev_i,
idx=i,
return_token_timestamps=return_token_timestamps,
)
current_segments[prev_i] += segments
if is_shortform:
seek[prev_i] += max_frames[i]
else:
seek[prev_i] += segment_offset
# 7. Once all segments are added to the list of all segments, called `current_segments`, we extract the predicted
# output tokens from the list of dicts. If we use batch size > 1, we make sure to pad the output
final_segments = (
[x[1:] for x in current_segments]
if (prompt_ids is not None and generation_config.prompt_condition_type == "first-segment")
else current_segments
)
sequences = _pad_to_max_length(
final_segments, generation_config.pad_token_id, device=self.device, padding_side="right"
)
# 8. If we return all segments, the predicted output sequences are put under `"sequences"`.
if return_segments:
return {"sequences": sequences, "segments": final_segments}
if is_shortform:
# add eos token:
if generation_config.max_new_tokens is None and generation_config.max_length is None:
eos_tokens = torch.full((sequences.shape[0], 1), generation_config.eos_token_id)
sequences = torch.cat([sequences, eos_tokens], dim=-1)
if return_token_timestamps:
outputs = {}
outputs["sequences"] = sequences
outputs["token_timestamps"] = torch.stack([d["token_timestamps"] for d in seek_outputs], dim=0)
else:
outputs = sequences
if return_dict_in_generate and generation_config.return_dict_in_generate:
dict_outputs = self._stack_split_outputs(seek_outputs, model_output_type, sequences.device, kwargs)
if num_return_sequences > 1:
if hasattr(dict_outputs, "encoder_attentions") and dict_outputs.encoder_attentions is not None:
dict_outputs.encoder_attentions = tuple(
dict_outputs.encoder_attentions[i][::num_return_sequences]
for i in range(len(dict_outputs.encoder_attentions))
)
if (
hasattr(dict_outputs, "encoder_hidden_states")
and dict_outputs.encoder_hidden_states is not None
):
dict_outputs.encoder_hidden_states = tuple(
dict_outputs.encoder_hidden_states[i][::num_return_sequences]
for i in range(len(dict_outputs.encoder_hidden_states))
)
if return_token_timestamps:
dict_outputs["token_timestamps"] = outputs["token_timestamps"]
return dict_outputs
return outputs
return sequences
def generate_with_fallback(
self,
segment_input,
decoder_input_ids,
cur_bsz,
batch_idx_map,
seek,
num_segment_frames,
max_frames,
temperatures,
generation_config,
logits_processor,
stopping_criteria,
prefix_allowed_tokens_fn,
synced_gpus,
return_token_timestamps,
do_condition_on_prev_tokens,
is_shortform,
batch_size,
kwargs,
):
kwargs = copy.copy(kwargs)
# 6.6 Batch generate current chunk
seek_sequence_list = [None for _ in range(cur_bsz)]
seek_outputs_list = [None for _ in range(cur_bsz)]
needs_fallback = [False for _ in range(cur_bsz)]
should_skip = [False for _ in range(cur_bsz)]
fallback_index_map = list(range(cur_bsz))
if generation_config.no_speech_threshold is not None:
self._setup_no_speech_detection(logits_processor, segment_input, decoder_input_ids, kwargs)
for fallback_idx, temperature in enumerate(temperatures):
generation_config.do_sample = temperature is not None and temperature > 0.0
generation_config.temperature = temperature if generation_config.do_sample else 1.0
if generation_config.do_sample:
generation_config.num_beams = 1
generate_kwargs = copy.copy(kwargs)
for key in ["do_sample", "temperature", "num_beams"]:
if key in generate_kwargs:
del generate_kwargs[key]
cur_bsz = decoder_input_ids.shape[0]
if generation_config.cache_implementation == "static" and cur_bsz < batch_size:
segment_input = F.pad(segment_input, (0, 0, 0, 0, 0, batch_size - cur_bsz), value=0)
decoder_input_ids = F.pad(
decoder_input_ids, (0, 0, 0, batch_size - cur_bsz), value=generation_config.pad_token_id
)
if generate_kwargs.get("decoder_attention_mask") is not None:
generate_kwargs["decoder_attention_mask"] = F.pad(
generate_kwargs["decoder_attention_mask"], (0, 0, 0, batch_size - cur_bsz), value=True
)
if generate_kwargs.get("encoder_outputs") is not None:
generate_kwargs["encoder_outputs"] = F.pad(
generate_kwargs["encoder_outputs"], (0, 0, 0, 0, 0, batch_size - cur_bsz), value=0
)
seek_outputs = super().generate(
segment_input,
generation_config=generation_config,
logits_processor=logits_processor,
stopping_criteria=stopping_criteria,
prefix_allowed_tokens_fn=prefix_allowed_tokens_fn,
synced_gpus=synced_gpus,
decoder_input_ids=decoder_input_ids,
**generate_kwargs,
)
model_output_type = type(seek_outputs)
# post-process sequence tokens and outputs to be in list form
seek_sequences, seek_outputs = self._postprocess_outputs(
seek_outputs=seek_outputs,
decoder_input_ids=decoder_input_ids,
return_token_timestamps=return_token_timestamps,
generation_config=generation_config,
is_shortform=is_shortform,
)
if cur_bsz < batch_size:
seek_sequences = seek_sequences[:cur_bsz]
seek_outputs = seek_outputs[:cur_bsz]
# 6.7 Extract cut sequences from every sequence and check if fallback should be applied
# Loop over each decoded audio individually as each decoding can be of a different length
new_fallback_index_map = []
new_segment_input = []
new_decoder_input_ids = []
new_decoder_attention_mask = []
for i, seek_sequence in enumerate(seek_sequences):
# make sure we cut a predicted EOS token if we are not finished with the generation yet
prev_i = batch_idx_map[fallback_index_map[i]]
is_not_final = (seek[prev_i] + num_segment_frames) < max_frames[prev_i]
# remove eos token id
if is_not_final and seek_sequence[-1] == generation_config.eos_token_id:
seek_sequence = seek_sequence[:-1]
if return_token_timestamps and not is_shortform:
seek_outputs[i]["token_timestamps"] = seek_outputs[i]["token_timestamps"][:-1]
# remove all padding tokens
if seek_sequence[-1] == generation_config.pad_token_id:
num_paddings = (seek_sequence == generation_config.pad_token_id).sum()
seek_sequence = seek_sequence[:-num_paddings]
if return_token_timestamps and not is_shortform:
seek_outputs[i]["token_timestamps"] = seek_outputs[i]["token_timestamps"][:-num_paddings]
# check which sequences in batch need fallback & which should be skipped
needs_fallback[i], should_skip[i] = self._need_fallback(
seek_sequence,
seek_outputs,
i,
logits_processor,
generation_config,
self.config.vocab_size,
temperature,
)
seek_sequence_list[fallback_index_map[i]] = seek_sequence
seek_outputs_list[fallback_index_map[i]] = seek_outputs[i]
is_low_temperature = temperature is None or temperature < 0.5
do_condition_on_prev_tokens[fallback_index_map[i]] = (
generation_config.condition_on_prev_tokens and is_low_temperature
)
if needs_fallback[i]:
new_fallback_index_map.append(fallback_index_map[i])
new_segment_input.append(segment_input[i])
new_decoder_input_ids.append(decoder_input_ids[i])
if "decoder_attention_mask" in kwargs:
new_decoder_attention_mask.append(kwargs["decoder_attention_mask"][i])
fallback_index_map = new_fallback_index_map
# if no sequence needs to be run with temperature fallback, we're finished
if len(fallback_index_map) == 0 or fallback_idx == len(temperatures) - 1:
seek_sequences = seek_sequence_list
seek_outputs = seek_outputs_list
break
# if we're still in the loop, make sure that decoder_input_ids and segment inputs are tensors
decoder_input_ids = torch.stack(new_decoder_input_ids)
segment_input = torch.stack(new_segment_input)
if "decoder_attention_mask" in kwargs:
kwargs["decoder_attention_mask"] = torch.stack(new_decoder_attention_mask)
return seek_sequences, seek_outputs, should_skip, do_condition_on_prev_tokens, model_output_type
@staticmethod
def _prepare_segments(prompt_ids, batch_size, generation_config):
if prompt_ids is not None and generation_config.prompt_condition_type == "first-segment":
prev_sot_token_id = getattr(generation_config, "prev_sot_token_id", None)
prompt_ids = prompt_ids[1:] if prompt_ids[0] == prev_sot_token_id else prompt_ids
current_segments = [[{"tokens": prompt_ids}] for _ in range(batch_size)]
else:
current_segments = [[] for _ in range(batch_size)]
return current_segments
def _postprocess_outputs(
self, seek_outputs, decoder_input_ids, return_token_timestamps, generation_config, is_shortform
):
# remove all previously passed decoder input ids
start_idx = decoder_input_ids.shape[-1] if not is_shortform else torch.tensor(0)
if isinstance(seek_outputs, torch.Tensor):
seek_outputs = seek_outputs[:, start_idx:]
return seek_outputs, seek_outputs
if return_token_timestamps and hasattr(generation_config, "alignment_heads"):
num_frames = getattr(generation_config, "num_frames", None)
seek_outputs["token_timestamps"] = self._extract_token_timestamps(
seek_outputs, generation_config.alignment_heads, num_frames=num_frames
)
seek_outputs["token_timestamps"] = seek_outputs["token_timestamps"][:, start_idx:]
seek_outputs["sequences"] = seek_outputs["sequences"][:, start_idx:]
def split_by_batch_index(values, key, batch_idx, is_shortform):
if key in ["scores", "encoder_attentions", "encoder_hidden_states", "logits"]:
return [v[batch_idx].cpu() for v in values]
if key in ["decoder_attentions", "decoder_hidden_states", "cross_attentions"]:
return tuple(tuple(w[batch_idx][None].cpu() for w in v) for v in values)
elif key == "past_key_values":
if not is_shortform:
# we don't save `past_key_values` as this is too costly for longform
return None
elif isinstance(values, EncoderDecoderCache):
all_past_key_values = []
for layer_idx in range(self.config.decoder_layers):
layer_past_key_values = []
for cache_cls in [values.self_attention_cache, values.cross_attention_cache]:
for v in [cache_cls.key_cache, cache_cls.value_cache]:
layer_past_key_values.append(v[layer_idx][batch_idx][None].cpu())
all_past_key_values.append(tuple(layer_past_key_values))
return tuple(all_past_key_values)
else:
all_past_key_values = []
for v in range(len(values)):
layer_past_key_values = []
for w in values[v]:
layer_past_key_values.append(w[batch_idx][None].cpu())
all_past_key_values.append(tuple(layer_past_key_values))
return tuple(all_past_key_values)
return values[batch_idx].cpu()
sequence_tokens = seek_outputs["sequences"]
seek_outputs = [
{k: split_by_batch_index(v, k, i, is_shortform) for k, v in seek_outputs.items()}
for i in range(sequence_tokens.shape[0])
]
return sequence_tokens, seek_outputs
def _stack_split_outputs(self, seek_outputs, model_output_type, device, kwargs):
# Stack back seek_outputs tensors after splitting them with the split_by_batch_index method
outputs = {}
for key in seek_outputs[0].keys():
if key == "sequences":
outputs[key] = torch.stack([v[key] for v in seek_outputs], dim=0).to(device)
if key in ["scores", "encoder_attentions", "encoder_hidden_states", "logits"]:
outputs[key] = tuple(
torch.stack([v[key][i] for v in seek_outputs]).to(device) for i in range(len(seek_outputs[0][key]))
)
if key in ["decoder_attentions", "decoder_hidden_states", "cross_attentions"]:
outputs[key] = tuple(
tuple(
torch.stack([v[key][i][j] for v in seek_outputs]).squeeze(1).to(device)
for j in range(len(seek_outputs[0][key][0]))
)
for i in range(len(seek_outputs[0][key]))
)
if key == "past_key_values":
past_key_value_type = kwargs.get("past_key_values")
if seek_outputs[0][key] is not None:
outputs[key] = tuple(
tuple(
torch.stack([v[key][i][j] for v in seek_outputs]).squeeze(1).to(device)
for j in range(len(seek_outputs[0][key][0]))
)
for i in range(len(seek_outputs[0][key]))
)
if past_key_value_type is not None and isinstance(past_key_value_type, EncoderDecoderCache):
outputs[key] = past_key_value_type.from_legacy_cache(outputs[key])
else:
outputs[key] = None
return model_output_type(**outputs)
def _need_fallback(
self,
seek_sequence,
seek_outputs,
index,
logits_processor,
generation_config,
vocab_size,
temperature,
):
needs_fallback = False
should_skip = False
if generation_config.compression_ratio_threshold is not None:
compression_ratio = self._retrieve_compression_ratio(seek_sequence, vocab_size)
if compression_ratio > generation_config.compression_ratio_threshold:
needs_fallback = True
if generation_config.logprob_threshold is not None:
if hasattr(seek_outputs[0], "sequences_scores"):
logprobs = [s["sequences_scores"] for s in seek_outputs][index]
else:
scores = seek_outputs[index]["scores"]
logprobs = self._retrieve_avg_logprobs(
scores, seek_sequence, generation_config.eos_token_id, temperature
)
if logprobs < generation_config.logprob_threshold:
needs_fallback = True
if generation_config.no_speech_threshold is not None:
no_speech_prob = _get_attr_from_logit_processors(
logits_processor, WhisperNoSpeechDetection, "no_speech_prob"
)
if (
logprobs < generation_config.logprob_threshold
and no_speech_prob[index] > generation_config.no_speech_threshold
):
needs_fallback = False
should_skip = True
return needs_fallback, should_skip
def _expand_variables_for_generation(
self, input_features, seek, max_frames, init_tokens, batch_size, condition_on_prev_tokens, generation_config
):
if generation_config.num_return_sequences is not None and generation_config.num_return_sequences > 1:
batch_idx_map = list(range(batch_size * generation_config.num_return_sequences))
cur_bsz = len(batch_idx_map)
do_condition_on_prev_tokens = [condition_on_prev_tokens for _ in range(len(batch_idx_map))]
input_features = input_features.repeat_interleave(generation_config.num_return_sequences, dim=0)
seek = seek.repeat_interleave(generation_config.num_return_sequences, dim=0)
max_frames = max_frames.repeat_interleave(generation_config.num_return_sequences, dim=0)
init_tokens = init_tokens.repeat_interleave(generation_config.num_return_sequences, dim=0)
generation_config.num_return_sequences = 1
else:
cur_bsz = batch_size
batch_idx_map = list(range(cur_bsz))
do_condition_on_prev_tokens = [condition_on_prev_tokens for _ in range(cur_bsz)]
return (
batch_idx_map,
cur_bsz,
input_features,
seek,
max_frames,
init_tokens,
do_condition_on_prev_tokens,
)
@staticmethod
def _setup_no_speech_detection(logits_processor, segment_input, decoder_input_ids, kwargs):
set_inputs = _get_attr_from_logit_processors(logits_processor, WhisperNoSpeechDetection, "set_inputs")
extra_kwargs = {k: v for k, v in kwargs.items() if torch.is_tensor(v)}
set_inputs({"inputs": segment_input, "decoder_input_ids": decoder_input_ids, **extra_kwargs})
@staticmethod
def _retrieve_total_input_frames(input_features, input_stride, kwargs):
if input_features is not None:
return input_features.shape[0], input_features.shape[-1]
if "encoder_outputs" in kwargs:
encoder_outputs_shape = (
kwargs["encoder_outputs"][0].shape
if isinstance(kwargs["encoder_outputs"], BaseModelOutput)
else kwargs["encoder_outputs"].shape
)
return encoder_outputs_shape[0], encoder_outputs_shape[1] * input_stride
raise ValueError("Make sure to provide either `input_features` or `encoder_outputs` to `generate`.")
@staticmethod
def _maybe_warn_unused_inputs(
condition_on_prev_tokens,
temperature,
compression_ratio_threshold,
logprob_threshold,
no_speech_threshold,
total_input_frames,
):
warning_prefix = (
f"Audio input consists of only {total_input_frames}. "
"Short-form transcription is activated."
"{}, but will be ignored."
)
if condition_on_prev_tokens is not None:
logger.warning(warning_prefix.format(f"condition_on_prev_tokens is set to {condition_on_prev_tokens}"))
if compression_ratio_threshold is not None:
logger.warning(
warning_prefix.format(f"compression_ratio_threshold is set to {compression_ratio_threshold}")
)
if logprob_threshold is not None:
logger.warning(warning_prefix.format(f"logprob_threshold is set to {logprob_threshold}"))
if no_speech_threshold is not None:
logger.warning(warning_prefix.format(f"no_speech_threshold is set to {no_speech_threshold}"))
# when passing temperature as a list it cannot just be ignored => throw error in this case
if isinstance(temperature, (list, tuple)):
raise ValueError(
f"Audio input consists of only {total_input_frames}. Short-form transcription is activated."
f"temperature cannot be set to {temperature} which can only be used for temperature fallback for long-form generation. Make sure to set `temperature` to a float value or `None` for short-form generation."
)
@staticmethod
def _set_return_outputs(return_dict_in_generate, return_token_timestamps, logprob_threshold, generation_config):
if return_dict_in_generate is None:
return_dict_in_generate = generation_config.return_dict_in_generate
else:
generation_config.return_dict_in_generate = return_dict_in_generate
generation_config.return_token_timestamps = return_token_timestamps
if return_token_timestamps:
generation_config.return_dict_in_generate = True
generation_config.output_attentions = True
generation_config.output_scores = True
if logprob_threshold is not None:
generation_config.return_dict_in_generate = True
generation_config.output_scores = True
return return_dict_in_generate
def _set_return_timestamps(self, return_timestamps, is_shortform, generation_config):
if return_timestamps is None and hasattr(generation_config, "return_timestamps"):
return_timestamps = generation_config.return_timestamps
if not is_shortform:
if return_timestamps is False:
raise ValueError(
"You have passed more than 3000 mel input features (> 30 seconds) which automatically enables long-form generation which "
"requires the model to predict timestamp tokens. Please either pass `return_timestamps=True` or make sure to pass no more than 3000 mel input features."
)
logger.info("Setting `return_timestamps=True` for long-form generation.")
return_timestamps = True
if return_timestamps and not hasattr(generation_config, "no_timestamps_token_id"):
raise ValueError(
"You are trying to return timestamps, but the generation config is not properly set. "
"Make sure to initialize the generation config with the correct attributes that are needed such as `no_timestamps_token_id`. "
"For more details on how to generate the approtiate config, refer to https://github.com/huggingface/transformers/issues/21878#issuecomment-1451902363"
)
generation_config.return_timestamps = return_timestamps
if hasattr(generation_config, "no_timestamps_token_id"):
timestamp_begin = generation_config.no_timestamps_token_id + 1
else:
# BC for models missing the `no_timestamps_token_id` in the generation config when generating short-form with no timestamps
# We set the timestamp begin token larger than the vocab size, such that the timestamp condition is never met in the decoding loop
timestamp_begin = self.config.vocab_size + 1
return timestamp_begin
@staticmethod
def _set_language_and_task(language, task, is_multilingual, generation_config):
if is_multilingual is not None:
if not hasattr(generation_config, "is_multilingual"):
raise ValueError(
"The generation config is outdated and is thus not compatible with the `is_multilingual` argument "
"to `generate`. Please update the generation config as per the instructions "
"https://github.com/huggingface/transformers/issues/25084#issuecomment-1664398224"
)
generation_config.is_multilingual = is_multilingual
if hasattr(generation_config, "is_multilingual") and not generation_config.is_multilingual:
if task is not None or language is not None:
raise ValueError(
"Cannot specify `task` or `language` for an English-only model. If the model is intended to be "
"multilingual, pass `is_multilingual=True` to generate, or update the generation config."
)
if language is not None:
if not hasattr(generation_config, "lang_to_id"):
raise ValueError(
"The generation config is outdated and is thus not compatible with the `language` argument "
"to `generate`. Either set the language using the `forced_decoder_ids` in the model config, "
"or update the generation config as per the instructions https://github.com/huggingface/transformers/issues/25084#issuecomment-1664398224"
)
generation_config.language = language
if task is not None:
if not hasattr(generation_config, "task_to_id"):
raise ValueError(
"The generation config is outdated and is thus not compatible with the `task` argument "
"to `generate`. Either set the task using the `forced_decoder_ids` in the model config, "
"or update the generation config as per the instructions https://github.com/huggingface/transformers/issues/25084#issuecomment-1664398224"
)
generation_config.task = task
def _retrieve_init_tokens(self, input_features, batch_size, generation_config, config, num_segment_frames, kwargs):
def replace_or_add(lst: List[int], num: int, itr: Iterator[int]):
"""short function to replace num with a itr in lst"""
found = any(i in lst for i in itr)
if found:
lst = [num if i in itr else i for i in lst]
else:
lst.append(num)
return lst
def language_to_id(language: str) -> int:
language = language.lower()
if language in generation_config.lang_to_id.keys():
language_token = language
elif language in TO_LANGUAGE_CODE.keys():
language_token = f"<|{TO_LANGUAGE_CODE[language]}|>"
elif language in TO_LANGUAGE_CODE.values():
language_token = f"<|{language}|>"
else:
is_language_code = len(language) == 2
raise ValueError(
f"Unsupported language: {language}. Language should be one of:"
f" {list(TO_LANGUAGE_CODE.values()) if is_language_code else list(TO_LANGUAGE_CODE.keys())}."
)
if language_token not in generation_config.lang_to_id:
raise ValueError(
f"{language_token} is not supported by this specific model as it is not in the `generation_config.lang_to_id`."
"(You should just add it to the generation config)"
)
return generation_config.lang_to_id[language_token]
task = getattr(generation_config, "task", None)
language = getattr(generation_config, "language", None)
forced_decoder_ids = generation_config.forced_decoder_ids
if forced_decoder_ids is not None:
if language is None and task is None and forced_decoder_ids[0][1] is None:
logger.warning_once(
"Due to a bug fix in https://github.com/huggingface/transformers/pull/28687 transcription using a multilingual Whisper will default to language detection followed by transcription instead of translation to English."
"This might be a breaking change for your use case. If you want to instead always translate your audio to English, make sure to pass `language='en'`."
)
elif hasattr(config, "forced_decoder_ids") and config.forced_decoder_ids is not None:
forced_decoder_ids = config.forced_decoder_ids
if forced_decoder_ids is not None and task is not None:
logger.warning_once(
f"You have passed task={task}, but also have set `forced_decoder_ids` to {forced_decoder_ids} which creates a conflict. `forced_decoder_ids` will be ignored in favor of task={task}."
)
forced_decoder_ids = None
elif forced_decoder_ids is not None and language is not None:
logger.warning_once(
f"You have passed language={language}, but also have set `forced_decoder_ids` to {forced_decoder_ids} which creates a conflict. `forced_decoder_ids` will be ignored in favor of language={language}."
)
forced_decoder_ids = None
init_tokens = [generation_config.decoder_start_token_id]
if forced_decoder_ids is not None and forced_decoder_ids[0][0] == 1:
i = 1
while len(forced_decoder_ids) > 0 and forced_decoder_ids[0][0] == i:
init_tokens += [forced_decoder_ids[0][1]]
forced_decoder_ids = forced_decoder_ids[1:]
i += 1
if len(forced_decoder_ids) > 0:
raise ValueError(
f"You are using token ids in `forced_decoder_ids` that do not seem to correctly follow the prompt pattern of Whisper. Make sure that {forced_decoder_ids} has an entry for all indices >= 1 and < {forced_decoder_ids[0][0]}.",
)
# from v4.39 the forced decoder ids are always None in favour of decoder input ids
generation_config.forced_decoder_ids = None
is_lang_id_undefined = len(init_tokens) <= 1 or (len(init_tokens) > 1 and init_tokens[1] is None)
# Make sure language is a list of strings of the correct length
if isinstance(language, (list, tuple)):
if any(l is None for l in language):
raise TypeError(
"Expected `language` to be `None`, a single string (e.g. `'en'`), or a list of strings with length equal to the batch size (e.g. `('en', 'fr')` for a batch size of 2). Got a list containing `None`."
)
if len(language) != batch_size:
raise ValueError(
"When passing a list of languages, the length of the list must match the batch size. "
f"Expected length of {batch_size}, but got {len(language)} languages."
)
languages = language
elif language is None:
# Language will be detected for each item in batch
languages = [None] * batch_size
else:
languages = [language] # Use a length-1 list now, broadcast later
# Separate init_tokens for each language
init_tokens = [copy.copy(init_tokens) for _ in languages]
# Update init_tokens with languages
lang_ids = None
if language is not None:
lang_ids = [language_to_id(l) for l in languages]
elif hasattr(generation_config, "lang_to_id") and is_lang_id_undefined:
# language is not defined or intentially set to `None` to trigger language detection
lang_ids = self.detect_language(
input_features=input_features,
encoder_outputs=kwargs.get("encoder_outputs", None),
attention_mask=kwargs.get("attention_mask", None),
generation_config=generation_config,
num_segment_frames=num_segment_frames,
).tolist()
if lang_ids is not None:
# append or replace lang_ids to init_tokens
for i in range(len(init_tokens)):
if len(init_tokens[i]) > 1:
init_tokens[i][1] = lang_ids[i]
else:
init_tokens[i].append(lang_ids[i])
del languages
# Update init_tokens with task
for i in range(len(init_tokens)):
if task is not None:
if task in TASK_IDS:
init_tokens[i].append(generation_config.task_to_id[generation_config.task])
task_id = generation_config.task_to_id[generation_config.task]
# if task is defined it'll overwrite task ids that might have already been defined via the generation_config
replace_or_add(init_tokens[i], task_id, generation_config.task_to_id.values())
else:
raise ValueError(f"The `{task}`task is not supported. The task should be one of `{TASK_IDS}`")
elif language is not None and hasattr(generation_config, "task_to_id"):
# if language is defined, but no task id is in `init_tokens`, default to transcribe
if not any(ti in init_tokens[i] for ti in generation_config.task_to_id.values()):
init_tokens[i].append(generation_config.task_to_id["transcribe"])
if (
not generation_config.return_timestamps
and hasattr(generation_config, "no_timestamps_token_id")
and init_tokens[i][-1] != generation_config.no_timestamps_token_id
):
init_tokens[i].append(generation_config.no_timestamps_token_id)
elif (
generation_config.return_timestamps and init_tokens[i][-1] == generation_config.no_timestamps_token_id
):
logger.info(
"<|notimestamps|> prompt token is removed from generation_config since `return_timestamps` is set to `'True'`."
)
init_tokens[i] = init_tokens[i][:-1]
# let's make sure we don't pass `None` tokens as prompt tokens
init_tokens[i] = [t for t in init_tokens[i] if t is not None]
return torch.as_tensor(init_tokens, dtype=torch.long, device=self.device).expand(batch_size, -1)
def detect_language(
self,
input_features: Optional[torch.FloatTensor] = None,
attention_mask: Optional[torch.LongTensor] = None,
encoder_outputs: Optional[Union[torch.FloatTensor, BaseModelOutput]] = None,
generation_config: Optional[GenerationConfig] = None,
num_segment_frames: int = 3000,
) -> torch.Tensor:
"""
Detects language from log-mel input features or encoder_outputs
Parameters:
input_features (`torch.Tensor` of shape `(batch_size, feature_size, sequence_length)`, *optional*):
Float values of log-mel features extracted from the raw speech waveform. The raw speech waveform can be obtained by
loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.* via
the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
[`AutoFeatureExtractor`] should be used for extracting the mel features, padding and conversion into a
tensor of type `torch.FloatTensor`. See [`~WhisperFeatureExtractor.__call__`] for details.
encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*):
Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`)
`last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) is a sequence of
hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
generation_config (`~generation.GenerationConfig`, *optional*):
The generation configuration to be used as base parametrization for the generation call. `**kwargs`
passed to generate matching the attributes of `generation_config` will override them. If
`generation_config` is not provided, the default will be used, which had the following loading
priority: 1) from the `generation_config.json` model file, if it exists; 2) from the model
configuration. Please note that unspecified parameters will inherit [`~generation.GenerationConfig`]'s
default values, whose documentation should be checked to parameterize generation.
num_segment_frames (`int`, *optional*, defaults to 3000):
The number of log-mel frames the model expects
Return:
A `torch.LongTensor` representing the detected language ids.
"""
if input_features is None and encoder_outputs is None:
raise ValueError("You have to specify either `input_features` or `encoder_outputs`")
elif input_features is not None and encoder_outputs is not None:
raise ValueError("Make sure to specificy only one of `input_features` or `encoder_outputs` - not both!")
elif input_features is not None:
inputs = {"input_features": input_features[:, :, :num_segment_frames]}
batch_size = input_features.shape[0]
elif encoder_outputs is not None:
inputs = {"encoder_outputs": encoder_outputs}
batch_size = (
encoder_outputs[0].shape[0] if isinstance(encoder_outputs, BaseModelOutput) else encoder_outputs[0]
)
if attention_mask is not None:
inputs["attention_mask"] = attention_mask
generation_config = generation_config or self.generation_config
decoder_input_ids = (
torch.ones((batch_size, 1), device=self.device, dtype=torch.long)
* generation_config.decoder_start_token_id
)
with torch.no_grad():
logits = self(**inputs, decoder_input_ids=decoder_input_ids).logits[:, -1]
non_lang_mask = torch.ones_like(logits[0], dtype=torch.bool)
non_lang_mask[list(generation_config.lang_to_id.values())] = False
logits[:, non_lang_mask] = -np.inf
lang_ids = logits.argmax(-1)
return lang_ids
@staticmethod
def _check_decoder_input_ids(kwargs):
decoder_input_ids = kwargs.get("decoder_input_ids", None)
assistant_model = kwargs.get("assistant_model", None)
if decoder_input_ids is not None and assistant_model is not None:
raise ValueError(
"Passing `decoder_input_ids` is deprecated. Consider passing `prompt_ids` instead.",
)
@staticmethod
def _set_num_frames(return_token_timestamps, generation_config, kwargs):
if return_token_timestamps:
if getattr(generation_config, "task", None) == "translate":
logger.warning("Token-level timestamps may not be reliable for task 'translate'.")
if not hasattr(generation_config, "alignment_heads"):
raise ValueError(
"Model generation config has no `alignment_heads`, token-level timestamps not available. "
"See https://gist.github.com/hollance/42e32852f24243b748ae6bc1f985b13a on how to add this property to the generation config."
)
generation_config.num_frames = kwargs.pop("num_frames", None)
@staticmethod
def _set_thresholds_and_condition(
generation_config,
logprob_threshold,
compression_ratio_threshold,
no_speech_threshold,
condition_on_prev_tokens,
):
generation_config.logprob_threshold = (
logprob_threshold
if logprob_threshold is not None
else getattr(generation_config, "logprob_threshold", None)
)
generation_config.compression_ratio_threshold = (
compression_ratio_threshold
if compression_ratio_threshold is not None
else getattr(generation_config, "compression_ratio_threshold", None)
)
generation_config.no_speech_threshold = (
no_speech_threshold
if no_speech_threshold is not None
else getattr(generation_config, "no_speech_threshold", None)
)
generation_config.condition_on_prev_tokens = (
condition_on_prev_tokens
if condition_on_prev_tokens is not None
else getattr(generation_config, "condition_on_prev_tokens", None)
)
@staticmethod
def _set_prompt_condition_type(generation_config, prompt_condition_type):
allowed_cond_types = ["first-segment", "all-segments"]
# default to "first-segment"
prompt_condition_type = prompt_condition_type or allowed_cond_types[0]
if prompt_condition_type not in allowed_cond_types:
raise ValueError(
f"`prompt_condition_type={prompt_condition_type} does not exist. Make sure to set `prompt_condition_type` to one of {', '.join(allowed_cond_types)}"
)
if generation_config.condition_on_prev_tokens is not True and prompt_condition_type == "all-segments":
raise ValueError(
"Make sure to set `condition_on_prev_tokens=True` when setting `prompt_condition_type='all-segments'`."
)
generation_config.prompt_condition_type = prompt_condition_type
@staticmethod
def _set_condition_on_prev_tokens(condition_on_prev_tokens, generation_config):
condition_on_prev_tokens = (
condition_on_prev_tokens
if condition_on_prev_tokens is not None
else getattr(generation_config, "condition_on_prev_tokens", False)
)
generation_config.condition_on_prev_tokens = condition_on_prev_tokens
@staticmethod
def _retrieve_max_frames_and_seek(batch_size, attention_mask, total_input_frames, is_shortform):
if batch_size > 1 and not is_shortform and attention_mask is None:
raise ValueError(
"When doing batched long-form audio transcription, make sure to pass an `attention_mask`. You can retrieve the `attention_mask` by doing `processor(audio, ..., return_attention_mask=True)` "
)
elif batch_size > 1 and not is_shortform:
max_frames = attention_mask.sum(-1).cpu().to(torch.long)
seek = torch.zeros((batch_size,), dtype=torch.long)
else:
max_frames = torch.ones((batch_size,), dtype=torch.long) * total_input_frames
seek = torch.zeros((batch_size,), dtype=torch.long)
return max_frames, seek
def _retrieve_logit_processors(self, generation_config, logits_processor, begin_index, num_beams, device):
if generation_config.return_timestamps is True:
timestamp_processor = WhisperTimeStampLogitsProcessor(generation_config, begin_index=begin_index)
logits_processor = (
[timestamp_processor] if logits_processor is None else [timestamp_processor] + logits_processor
)
if generation_config.suppress_tokens is not None:
suppress_tokens_processor = SuppressTokensLogitsProcessor(generation_config.suppress_tokens, device=device)
logits_processor = (
[suppress_tokens_processor]
if logits_processor is None
else [suppress_tokens_processor] + logits_processor
)
generation_config.suppress_tokens = None
if generation_config.begin_suppress_tokens is not None:
begin_suppress_processor = SuppressTokensAtBeginLogitsProcessor(
generation_config.begin_suppress_tokens, begin_index=begin_index, device=device
)
logits_processor = (
[begin_suppress_processor]
if logits_processor is None
else [begin_suppress_processor] + logits_processor
)
generation_config.begin_suppress_tokens = None
if generation_config.no_speech_threshold is not None:
no_speech_detector = WhisperNoSpeechDetection(
no_speech_token=generation_config.no_timestamps_token_id - 1,
begin_index=begin_index,
scores_is_logprobs=num_beams > 1,
)
logits_processor = (
[no_speech_detector] if logits_processor is None else [no_speech_detector] + logits_processor
)
no_speech_detector.set_model(self)
return logits_processor
@staticmethod
def _maybe_reduce_batch(input_features, seek, max_frames, cur_bsz, batch_idx_map):
prev_bsz = cur_bsz
new_batch_idx_map = []
for i in range(prev_bsz):
prev_i = batch_idx_map[i]
if seek[prev_i] >= max_frames[prev_i]:
cut_index = i + (cur_bsz - prev_bsz)
cur_bsz -= 1
input_features = torch.cat([input_features[:cut_index], input_features[cut_index + 1 :]], dim=0)
else:
# cut out index that goes away
new_batch_idx_map.append(prev_i)
return input_features, cur_bsz, new_batch_idx_map
@staticmethod
def _get_input_segment(input_features, seek, seek_num_frames, num_segment_frames, cur_bsz, batch_idx_map):
if input_features is None:
return None
segment_input = []
for i in range(cur_bsz):
prev_i = batch_idx_map[i]
segment_input_slice = input_features[i : i + 1, :, seek[prev_i] : seek[prev_i] + seek_num_frames[prev_i]]
if segment_input_slice.shape[-1] < num_segment_frames:
# pad to 3000 if necessary
segment_input_slice = F.pad(
segment_input_slice, pad=(0, num_segment_frames - segment_input_slice.shape[-1])
)
segment_input.append(segment_input_slice)
segment_input = torch.cat(segment_input, dim=0)
return segment_input
@staticmethod
def _prepare_decoder_input_ids(
cur_bsz,
init_tokens,
current_segments,
batch_idx_map,
do_condition_on_prev_tokens,
prompt_ids,
generation_config,
config,
device,
suppress_tokens,
kwargs,
):
if "decoder_input_ids" in kwargs:
decoder_input_ids = kwargs.pop("decoder_input_ids")
return decoder_input_ids, kwargs
cut_off_length = config.max_target_positions // 2 - 1
decoder_input_ids = init_tokens[batch_idx_map]
prev_start_of_text = getattr(generation_config, "prev_sot_token_id", None)
if prev_start_of_text is None:
prev_start_of_text = suppress_tokens[-2] if suppress_tokens is not None else None
if any(do_condition_on_prev_tokens) and len(current_segments[0]) > 0:
# according to https://github.com/openai/whisper/blob/e58f28804528831904c3b6f2c0e473f346223433/whisper/decoding.py#L609
active_segments = [current_segments[i] if do_condition_on_prev_tokens[i] else None for i in batch_idx_map]
if prompt_ids is not None and generation_config.prompt_condition_type == "all-segments":
prev_ids = prompt_ids
else:
one_tensor = torch.ones((cur_bsz, 1), device=device, dtype=torch.long)
prev_ids = prev_start_of_text * one_tensor[0] if prev_start_of_text is not None else None
padding = "max_length" if generation_config.cache_implementation == "static" else "longest"
prev_tokens = _pad_to_max_length(
active_segments,
generation_config.pad_token_id,
device=device,
padding_side="left",
padding=padding,
bos_token_tensor=prev_ids,
cut_off_length=cut_off_length,
)
decoder_input_ids = torch.cat([prev_tokens, decoder_input_ids], dim=-1)
kwargs["decoder_attention_mask"] = decoder_input_ids != generation_config.pad_token_id
elif prompt_ids is not None:
prev_tokens = prompt_ids[None].repeat(decoder_input_ids.shape[0], 1)
decoder_input_ids = torch.cat([prev_tokens, decoder_input_ids], dim=-1)
# make sure `"decoder_attention_mask"` is not passed to forward
kwargs.pop("decoder_attention_mask", None)
else:
# make sure `"decoder_attention_mask"` is not passed to forward
kwargs.pop("decoder_attention_mask", None)
return decoder_input_ids, kwargs
def _set_max_new_tokens_and_length(self, config, decoder_input_ids, generation_config):
max_new_tokens = generation_config.max_new_tokens if generation_config.max_new_tokens is not None else 0
if max_new_tokens + decoder_input_ids.shape[-1] > self.config.max_target_positions:
raise ValueError(
f"The length of `decoder_input_ids` equal `prompt_ids` plus special start tokens is {decoder_input_ids.shape[-1]}, and the `max_new_tokens` "
f"is {max_new_tokens}. Thus, the combined length of "
f"`decoder_input_ids` and `max_new_tokens` is: {max_new_tokens + decoder_input_ids.shape[-1]}. This exceeds the "
f"`max_target_positions` of the Whisper model: {self.config.max_target_positions}. "
"You should either reduce the length of your prompt, or reduce the value of `max_new_tokens`, "
f"so that their combined length is less than {self.config.max_target_positions}."
)
num_initial_tokens = min(config.max_target_positions // 2 - 1, decoder_input_ids.shape[-1] - 1)
# Make sure we don't get larger than `max_length`
if generation_config.max_length is not None and generation_config.max_new_tokens is None:
max_length = min(generation_config.max_length + num_initial_tokens, config.max_target_positions)
logger.info(
f"Increase max_length from {generation_config.max_length} to {max_length} since input is conditioned on previous segment."
)
elif (
generation_config.max_new_tokens is not None
and generation_config.max_new_tokens + decoder_input_ids.shape[-1] > config.max_target_positions
):
max_new_tokens = config.max_target_positions - decoder_input_ids.shape[-1]
generation_config.max_new_tokens = max_new_tokens
@staticmethod
def _retrieve_compression_ratio(tokens, vocab_size):
"""Compute byte length of zlib compressed token bytes vs. byte length of raw token bytes"""
length = int(math.log2(vocab_size) / 8) + 1
token_bytes = b"".join([t.to_bytes(length, "little") for t in tokens.tolist()])
compression_ratio = len(token_bytes) / len(zlib.compress(token_bytes))
return compression_ratio
@staticmethod
def _retrieve_avg_logprobs(scores, tokens, eos_token_id, temperature):
rescale_temperature = temperature if temperature > 0.0 else 1
scores = torch.stack(scores).to(tokens.device)
if scores.shape[0] > tokens.shape[0]:
scores = scores[: tokens.shape[0]]
else:
tokens = tokens[-scores.shape[0] :]
logprobs = F.log_softmax((scores * rescale_temperature).float(), dim=-1).to(scores.dtype)
# retrieve logprob of selected tokens and sum
sum_logprobs = sum((logprobs[i][tokens[i]] * (tokens[i] != eos_token_id)) for i in range(logprobs.shape[0]))
length = (tokens != eos_token_id).sum(-1) if eos_token_id is not None else tokens.shape[0]
avg_logprobs = sum_logprobs / (length + 1)
return avg_logprobs
@staticmethod
def _retrieve_segment(
seek_sequence,
seek_outputs,
time_offset,
timestamp_begin,
seek_num_frames,
time_precision,
input_stride,
prev_idx,
idx,
return_token_timestamps,
):
# find the predicted "end of segment" predictions of Whisper
# "end of segment" predictions occur whenever Whisper predicts a timestamp token
timestamp_tokens: torch.Tensor = seek_sequence.ge(timestamp_begin)
single_timestamp_ending = timestamp_tokens[-2:].tolist() == [False, True]
timestamp_segment_indices = torch.where(timestamp_tokens[:-1] & timestamp_tokens[1:])[0]
timestamp_segment_indices.add_(1)
token_timestamps = seek_outputs[idx]["token_timestamps"] if return_token_timestamps else []
# If whisper predicted a "end of segment" via a timestep token, let's go ever each
# "end of segment" prediction and slice the decoding into segments accordingly
if len(timestamp_segment_indices) > 0:
# if the output contains two consecutive timestamp tokens
slices = timestamp_segment_indices.tolist()
segments = []
if single_timestamp_ending:
slices.append(len(seek_sequence))
last_slice = 0
# Add each segment to list of all segments
for current_slice in slices:
sliced_tokens = seek_sequence[last_slice:current_slice]
start_timestamp_pos = sliced_tokens[0].item() - timestamp_begin
end_timestamp_pos = sliced_tokens[-1].item() - timestamp_begin
segments.append(
{
"start": time_offset[prev_idx] + start_timestamp_pos * time_precision,
"end": time_offset[prev_idx] + end_timestamp_pos * time_precision,
"tokens": sliced_tokens,
"result": seek_outputs[idx],
}
)
if return_token_timestamps:
segments[-1]["token_timestamps"] = (
token_timestamps[last_slice:current_slice] + time_offset[prev_idx]
)
last_slice = current_slice
if single_timestamp_ending:
# single timestamp at the end means no speech after the last timestamp.
segment_offset = seek_num_frames[prev_idx]
else:
# otherwise, ignore the unfinished segment and seek to the last timestamp
# here we throw away all predictions after the last predicted "end of segment"
# since we are cutting right in the middle of an audio
last_timestamp_pos = seek_sequence[last_slice - 1].item() - timestamp_begin
segment_offset = last_timestamp_pos * input_stride
else:
# If whisper does not predict any "end of segment" token, then
# the whole decoding is considered a segment and we add it to the list of segments
timestamps = seek_sequence[timestamp_tokens.nonzero().flatten()]
last_timestamp_pos = seek_num_frames[prev_idx]
if timestamps.numel() > 0 and timestamps[-1].item() != timestamp_begin:
# no consecutive timestamps but it has a timestamp; use the last one.
last_timestamp_pos = timestamps[-1].item() - timestamp_begin
segments = [
{
"start": time_offset[prev_idx],
"end": time_offset[prev_idx] + last_timestamp_pos * time_precision,
"tokens": seek_sequence,
"result": seek_outputs[idx],
}
]
if return_token_timestamps:
segments[-1]["token_timestamps"] = token_timestamps + time_offset[prev_idx]
segment_offset = seek_num_frames[prev_idx]
return segments, segment_offset
|