Spaces:
Runtime error
Runtime error
File size: 82,872 Bytes
38818c3 1923610 38818c3 1923610 38818c3 1923610 38818c3 1923610 38818c3 1923610 38818c3 1923610 38818c3 1923610 38818c3 1923610 38818c3 1923610 38818c3 1923610 38818c3 1923610 38818c3 1923610 38818c3 1923610 38818c3 1923610 38818c3 1923610 38818c3 1923610 38818c3 1923610 38818c3 1923610 38818c3 |
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 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 |
# @title Model code (no change needed)
"""Model code"""
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
from __future__ import annotations
import io
import logging
import zlib
from dataclasses import dataclass, field
from typing import Dict, final, Final, List, Literal, Tuple
import librosa
import numpy as np
import soundfile as sf
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchaudio
import torchaudio.functional as audio_F
import uroman
from fairseq2.context import RuntimeContext
from fairseq2.data import VocabularyInfo
from fairseq2.models.asr import AsrModel, AsrModelOutput
from fairseq2.models.llama import LLaMAConfig, LLaMAFactory
from fairseq2.models.seq2seq import Seq2SeqBatch
from fairseq2.models.wav2vec2 import (
StandardWav2Vec2Masker,
Wav2Vec2EncoderConfig,
Wav2Vec2EncoderFactory,
Wav2Vec2Frontend,
Wav2Vec2Masker,
)
from fairseq2.models.wav2vec2.asr import Wav2Vec2AsrConfig
from fairseq2.nn import IncrementalStateBag, Linear, StandardEmbedding
from fairseq2.nn.padding import PaddingMask
from fairseq2.nn.transformer import TransformerDecoder, TransformerEncoder
from torch import Tensor
from torch.nn import Dropout
@final
class Wav2Vec2LlamaModel(AsrModel):
"""Represents a wav2vec 2.0 encoder feeding to a Llama decoder for ASR."""
model_dim: int
encoder_frontend: Wav2Vec2Frontend
encoder: TransformerEncoder
encoder_proj: nn.Module
text_frontend: StandardEmbedding
llama_decoder: TransformerDecoder
final_proj: nn.Module
masker: Wav2Vec2Masker | None
final_dropout: Dropout | None
target_vocab_info: VocabularyInfo
def __init__(
self,
encoder_frontend: Wav2Vec2Frontend,
encoder: TransformerEncoder,
encoder_proj: nn.Module,
text_frontend: StandardEmbedding,
llama_decoder: TransformerDecoder,
final_proj: nn.Module,
target_vocab_info: VocabularyInfo,
*,
masker: Wav2Vec2Masker | None = None,
final_dropout_p: float = 0.0,
max_generation_length: int = 8192,
encoder_stacking: int = 1,
frozen_encoder: bool = False,
random_context_length: bool = True,
) -> None:
"""
:param encoder_frontend:
The encoder frontend.
:param encoder:
The encoder (i.e. context network).
:param encoder_proj:
Normally a linear layer projecting the encoder outputs to the decoder's model dim.
:text_frontend:
The embedding module for text tokens.
:param llama_decoder:
The decoder-only model.
:param final_proj:
The last linear layers projecting from the decoder to logits.
:param target_vocab_info:
The vocabulary information of sequences produced by the model.
:param masker:
The feature masker.
:param final_dropout_p:
The dropout probability on context network outputs.
:param max_generation_length:
The maximum length of generated sequences.
:param encoder_stacking:
The number audio embeddings frames to stack before the decoder calls.
:param frozen_encoder:
If ``True``, the encoder is frozen during training.
"""
super().__init__()
self.model_dim = encoder.model_dim
self.encoder_frontend = encoder_frontend
self.encoder = encoder
self.encoder_proj = encoder_proj
self.text_frontend = text_frontend
self.llama_decoder = llama_decoder
self.final_proj = final_proj
self.target_vocab_info = target_vocab_info
self.max_generation_length = max_generation_length
self.encoder_stacking = encoder_stacking
self.frozen_encoder = frozen_encoder
self.random_context_length = random_context_length
self.context_len_rng = np.random.RandomState(42)
self.register_module("masker", masker)
if final_dropout_p > 0.0:
self.final_dropout = Dropout(final_dropout_p)
else:
self.register_module("final_dropout", None)
def forward(self, batch: Seq2SeqBatch) -> Wav2Vec2LlamaOutput: # type: ignore[override]
"""
:param batch:
The batch of sequences to process.
"""
device = batch.source_seqs.device
dtype = batch.source_seqs.dtype
batch = self.prepare_batch(batch)
inputs = self.create_default_syntax_inference(batch, device)
# Embed all modalities
embedded = self.embed_inputs(inputs, dtype)
# Concat all decoder inputs
(
decoder_inputs,
decoder_inputs_padding_mask,
decoder_context_inputs,
decoder_context_padding_mask,
) = self.concat_inputs(embedded)
# Run the decoder
dec_out, _ = self.llama_decoder(decoder_inputs, decoder_inputs_padding_mask)
logits = self.final_proj(dec_out)
assert self.target_vocab_info.pad_idx is not None
assert self.target_vocab_info.eos_idx is not None
return Wav2Vec2LlamaOutput(
logits=logits,
logits_padding_mask=decoder_inputs_padding_mask,
decoder_context_inputs=decoder_context_inputs,
decoder_context_padding_mask=decoder_context_padding_mask,
model=self,
pad_idx=self.target_vocab_info.pad_idx,
eos_idx=self.target_vocab_info.eos_idx,
padding_mask=None,
)
def prepare_batch(self, batch: Seq2SeqBatch) -> Seq2SeqBatch:
# Create padding masks if there aren't any
if batch.source_padding_mask is None:
lengths = torch.full_like(
batch.source_seqs[:, 0],
fill_value=batch.source_seqs.size(1),
dtype=torch.int64,
)
batch.source_padding_mask = PaddingMask(lengths, int(lengths.max()))
if batch.target_padding_mask is None:
lengths = torch.full_like(
batch.target_seqs[:, 0],
fill_value=batch.target_seqs.size(1),
dtype=torch.int64,
)
batch.target_padding_mask = PaddingMask(lengths, int(lengths.max()))
# Padding masks for context audio and text
if "context_audio" in batch.example:
for i in range(len(batch.example["context_audio"])):
# For audio
seq_lens = batch.example["context_audio"][i]["data"]["waveform"][
"seq_lens"
]
batch.example["context_audio"][i]["data"]["waveform"][
"padding_mask"
] = PaddingMask(seq_lens, int(seq_lens.max()))
# For text
seq_lens = batch.example["context_text"][i]["seq_lens"]
batch.example["context_text"][i]["padding_mask"] = PaddingMask(
seq_lens, int(seq_lens.max())
)
return batch
def create_default_syntax_inference(
self, batch: Seq2SeqBatch, device
) -> List[Dict[str, object]]:
# Create a dict of inputs for the base case. Ths syntax is:
# target audio <bos> target text <eos>
inputs = [
{
"value": {
"seqs": batch.source_seqs,
"padding_mask": batch.source_padding_mask,
},
"type": "audio",
"loss": False,
},
{
"value": {
"seqs": self.create_single_char(
batch, self.target_vocab_info.bos_idx, device
)
},
"type": "text",
"loss": False,
},
]
return inputs
@staticmethod
def create_single_char(batch: Seq2SeqBatch, char: int, device) -> Tensor:
return torch.full_like(
batch.target_seqs[:, :1], fill_value=char, device=device # type: ignore
)
def embed_inputs(
self, inputs: List[Dict[str, object]], dtype: Literal
) -> List[Dict[str, object]]:
# Embed the different modalities
for inp in inputs:
if inp["type"] == "audio":
inp["value"]["seqs"], inp["value"]["padding_mask"] = self.embed_audio(
inp["value"]["seqs"], inp["value"]["padding_mask"]
)
elif inp["type"] == "text":
inp["value"]["seqs"] = self.embed_text(inp["value"]["seqs"], dtype)
else:
raise ValueError(f"Unknown input type: {inp['type']}")
return inputs
def embed_audio(
self, seqs: Tensor, padding_mask: PaddingMask
) -> tuple[Tensor, PaddingMask | None]:
# This is somewhat more memory efficient than setting param.requires_grad to False
# Since the encoder activations will not be saved in the graph too.
with torch.set_grad_enabled(not self.frozen_encoder):
# Run the encoder
enc_out, enc_padding_mask, _ = self.encoder_frontend.extract_features(
seqs, padding_mask
)
enc_out, enc_padding_mask, _ = self.encoder_frontend.process_features(
enc_out, enc_padding_mask, self.masker if self.training else None
)
enc_out, enc_padding_mask = self.encoder(enc_out, enc_padding_mask)
if self.final_dropout is not None:
enc_out = self.final_dropout(enc_out)
# Stack the encoder outputs
if enc_out.size(1) % self.encoder_stacking != 0:
n_padding = self.encoder_stacking - (
enc_out.size(1) % self.encoder_stacking
)
enc_out = F.pad(enc_out, (0, 0, 0, n_padding))
assert enc_out.size(1) % self.encoder_stacking == 0
enc_out = enc_out.view(
enc_out.size(0),
enc_out.size(1) // self.encoder_stacking,
enc_out.size(-1) * self.encoder_stacking,
)
new_lengths = torch.where(
(enc_padding_mask.seq_lens % self.encoder_stacking) == 0,
enc_padding_mask.seq_lens // self.encoder_stacking,
enc_padding_mask.seq_lens // self.encoder_stacking + 1,
)
enc_padding_mask = PaddingMask(new_lengths, int(new_lengths.max()))
# Project encoder outputs to decoder input dimension
enc_out = self.encoder_proj(enc_out)
return enc_out, enc_padding_mask
def embed_text(self, seqs: Tensor, dtype: Literal) -> Tensor:
return self.text_frontend(seqs).to(dtype)
def concat_inputs(
self, inputs: List[Dict[str, object]]
) -> Tuple[Tensor, PaddingMask]:
t = inputs[0]["value"]["seqs"]
device = t.device
dtype = t.dtype
B = t.size(0)
input_dim = t.size(2)
ones = torch.ones(dtype=torch.int64, device=device, size=[B])
# Compute total lengths
lengths = [
(
inp["value"]["padding_mask"].seq_lens
if "padding_mask" in inp["value"]
else ones
)
for inp in inputs
]
total_lengths = sum(lengths)
padding_mask = PaddingMask(total_lengths, int(total_lengths.max()))
# Init the matrix with zeros
decoder_inputs = torch.zeros(
[B, int(total_lengths.max()), input_dim],
device=device,
dtype=dtype,
)
# Put everything in the right place
for b in range(B):
b_inputs = [
inp["value"]["seqs"][b : b + 1, : length[b]]
for (inp, length) in zip(inputs, lengths)
]
b_inputs = torch.cat(b_inputs, dim=1)
assert b_inputs.size(1) == padding_mask.seq_lens[b]
decoder_inputs[b, : b_inputs.size(1)] = b_inputs
# Compute total context length (everything that we don't train the loss for)
context_lengths = [
(
inp["value"]["padding_mask"].seq_lens
if "padding_mask" in inp["value"]
else ones
)
for inp in inputs
if inp["loss"] == False
]
total_context_lengths = sum(context_lengths)
context_padding_mask = PaddingMask(
total_context_lengths, int(total_context_lengths.max())
)
decoder_context_inputs = decoder_inputs[:, : total_context_lengths.max()]
return (
decoder_inputs,
padding_mask,
decoder_context_inputs,
context_padding_mask,
)
@final
@dataclass
class Wav2Vec2LlamaOutput(AsrModelOutput):
logits: Tensor
"""The logits for next-step prediction. *Shape:* :math:`(N,S_{out}, V)`,
where :math:`N` is the batch size, :math:`S_{out}` is the decoder sequence
length, :math:`V` is the size
of the vocabulary."""
logits_padding_mask: PaddingMask
"""The padding mask for the above tensor. *Shape:* :math:`(N,S_{out})`."""
decoder_context_inputs: Tensor
"""
Inputs to the llama decoder for everything except the final text. *Shape:* :math:`(N,S_{out},D)`.
"""
decoder_context_padding_mask: PaddingMask
"""The padding mask for the above tensor. *Shape:* :math:`(N,S_{out})`, where
:math:`N` is the batch size and :math:`S_{out}` a sequence
length."""
model: nn.Module
"""A reference to the model."""
pad_idx: int
"""The index of the padding symbol in the target vocabulary."""
eos_idx: int
"""The index of the end-of-sequence symbol in the target vocabulary."""
def add_eos(
self, targets: Tensor, target_padding_mask: PaddingMask
) -> tuple[Tensor, PaddingMask]:
targets = torch.cat(
[
targets,
torch.full_like(targets[:, :1], fill_value=self.pad_idx),
],
dim=-1,
)
targets[torch.arange(targets.size(0)), target_padding_mask.seq_lens] = (
self.eos_idx
)
target_padding_mask = PaddingMask(
target_padding_mask.seq_lens + 1,
int(target_padding_mask.seq_lens.max()) + 1,
)
return targets, target_padding_mask
def remove_context_logits(
self,
targets: Tensor,
target_padding_mask: PaddingMask,
) -> Tensor:
assert self.decoder_context_padding_mask is not None
logits_no_context = torch.zeros_like(
self.logits[:, : targets.size(1), :],
)
for i in range(self.logits.size(0)):
context_len_i = self.decoder_context_padding_mask.seq_lens[i]
tgt_len_i = target_padding_mask.seq_lens[i]
total_len_i = self.logits_padding_mask.seq_lens[i]
assert context_len_i + tgt_len_i == total_len_i
logits_no_context[i, :tgt_len_i] = self.logits[
i, context_len_i - 1 : context_len_i - 1 + tgt_len_i
]
return logits_no_context
@staticmethod
def combine_masks(mask1: Tensor, mask2: Tensor) -> Tensor:
combined_mask = torch.zeros_like(mask1)
combined_mask[mask1] = mask2
return combined_mask
@staticmethod
def idx_1d_to_2d(idx: Tensor, dim2: int) -> tuple[Tensor, Tensor]:
return idx // dim2, idx % dim2
@staticmethod
def compression_ratio(text: str) -> float:
text_bytes = text.encode("utf-8")
return len(text_bytes) / len(zlib.compress(text_bytes))
@torch.no_grad()
def generate_hypotheses(
self, pad_idx: int, blank_label: int = 0
) -> tuple[Tensor, PaddingMask | None]:
# Some init
nbest = 5
length_norm = False
B = self.decoder_context_inputs.size(0)
device = self.decoder_context_inputs.device
dtype = self.decoder_context_inputs.dtype
ex_separator = torch.arange(B, device=device).unsqueeze(1) * nbest
eos_idx = self.model.target_vocab_info.eos_idx
# Prepare a decoder input matrix, prefill with context
decoder_inputs = torch.zeros(
[
B * nbest,
self.model.max_generation_length,
self.model.llama_decoder.model_dim,
],
device=device,
dtype=dtype,
)
decoder_inputs[:, : self.decoder_context_inputs.size(1)] = (
self.decoder_context_inputs.repeat_interleave(nbest, dim=0)
)
context_lengths = self.decoder_context_padding_mask.seq_lens.repeat_interleave(
nbest
)
# Prepare a token output matrix and a scores matrix
out_tokens = torch.full_like(
decoder_inputs[:, :, 0],
fill_value=pad_idx,
dtype=torch.int,
)
scores = torch.zeros_like(decoder_inputs[:, 0, 0], dtype=torch.float)
# Prefill with shortest context, keep state
state_bag = IncrementalStateBag(max_num_steps=self.model.max_generation_length)
min_context_len = int(context_lengths.min())
_, _ = self.model.llama_decoder(
seqs=decoder_inputs[:, :min_context_len],
padding_mask=None,
state_bag=state_bag,
)
state_bag.increment_step_nr(min_context_len)
# Iterative decoding
# For each sample, choose either context, or emitted text embedding
# If EOS is emitted, the sample is non-active
# Stop when there are no active samples
eos_mask = torch.zeros_like(context_lengths, dtype=torch.bool)
done = False
t = context_lengths.min() - 1
while not done:
# Run the decoder on mixed context and emitted text embeddings
dec_out, _ = self.model.llama_decoder(
seqs=decoder_inputs[:, t : t + 1],
padding_mask=None,
state_bag=state_bag,
)
state_bag.increment_step_nr(1)
logits = self.model.final_proj(dec_out).squeeze(1) # [B * nbest, V]
log_probs = F.log_softmax(logits, dim=-1)
# Choose nbest
if length_norm:
n_tokens = torch.logical_and(
out_tokens[:, :t] != pad_idx, out_tokens[:, :t] != eos_idx
).sum(dim=1, keepdim=True)
candidate_scores = (scores.unsqueeze(1) * n_tokens + log_probs) / (
n_tokens + 1
)
else:
candidate_scores = scores.unsqueeze(1) + log_probs # [B * nbest, V]
candidate_scores[eos_mask] = -torch.inf
candidate_scores[eos_mask, eos_idx] = scores[
eos_mask
] # Don't change scores for ended hypos
top_scores, top_idx = candidate_scores.view(B, -1).topk(
k=nbest, dim=-1, sorted=True
)
top_idx_nbest, top_idx_v = self.idx_1d_to_2d(
top_idx, candidate_scores.size(-1)
)
top_idx_b = (top_idx_nbest + ex_separator).view(-1) # Parent hypos indices
# Reorder some tensors based on parent hypos
out_tokens = out_tokens[top_idx_b]
eos_mask = eos_mask[top_idx_b]
state_bag.reorder(top_idx_b)
scores = torch.where(eos_mask, scores, top_scores.view(-1))
out_tokens[:, t] = top_idx_v.view(-1)
# For hypos that still don't emit tokens, set new tokens to pad_idx, score to 0.
no_token_mask = t < context_lengths - 1
out_tokens[no_token_mask, t] = pad_idx
scores[no_token_mask] = 0.0
# For hypos that had EOS previously, set new tokens to EOS. Scores don't change.
# Set new EOS mask.
out_tokens[eos_mask, t] = eos_idx
new_tokens = out_tokens[:, t : t + 1]
eos_mask = (new_tokens == eos_idx).squeeze(1)
# Run new tokens through frontend, set in decoder input
new_tokens_embedded = self.model.embed_text(new_tokens, dtype=dtype)
decoder_inputs[~no_token_mask, t + 1] = (
new_tokens_embedded[~no_token_mask].to(decoder_inputs.dtype).squeeze(1)
) # Don't override audio encoder outputs
# Early stopping if emitting repeating characters, use compression ratio
# only every t, only when started emitting tokens more than T tokens ago
compression_window = 100
compression_threshold = 4.0
if t % 250 == 0:
cpu_tokens = out_tokens[:, t - compression_window : t].cpu().numpy()
ratios_floats = [
self.compression_ratio(
np.array_str(cpu_tokens[i]).replace("\n", "")
)
for i in range(B * nbest)
]
ratios = torch.tensor(ratios_floats, device=device)
early_stopping_mask = torch.logical_and(
ratios > compression_threshold,
t > context_lengths + compression_window,
)
eos_mask = torch.logical_or(eos_mask, early_stopping_mask)
# Decide if we are done
done = bool(
torch.logical_or(
torch.all(eos_mask),
t == self.model.max_generation_length - 4,
)
)
t += 1
# Get final tokens, only use top hypo
out_tokens = out_tokens[::nbest]
valid_tokens_mask = torch.logical_and(
torch.logical_and(
out_tokens != pad_idx,
out_tokens != self.model.target_vocab_info.bos_idx,
),
out_tokens != eos_idx,
)
valid_tokens_count = valid_tokens_mask.sum(dim=1)
final_tokens = torch.full(
[B, int(valid_tokens_count.max())],
fill_value=pad_idx,
dtype=torch.int64,
device=device,
)
for i in range(B):
final_tokens[i, : valid_tokens_count[i]] = out_tokens[i][
valid_tokens_mask[i]
]
padding_mask = PaddingMask(valid_tokens_count, int(valid_tokens_count.max()))
return final_tokens, padding_mask
class Wav2Vec2LlamaFactory:
_config: Wav2Vec2LlamaConfig
def __init__(
self,
config: Wav2Vec2LlamaConfig,
) -> None:
self._config = config
def create_encoder(self) -> tuple[Wav2Vec2Frontend, TransformerEncoder]:
factory = Wav2Vec2EncoderFactory(self._config.wav2vec_ctc_config.encoder_config)
return factory.create_encoder_frontend(), factory.create_encoder()
def create_masker(self) -> Wav2Vec2Masker:
config = self._config.wav2vec_ctc_config
return StandardWav2Vec2Masker(
config.mask_codebase,
config.encoder_config.model_dim,
config.temporal_mask_span_len,
config.max_temporal_mask_prob,
config.min_num_temporal_mask_spans,
config.spatial_mask_span_len,
config.max_spatial_mask_prob,
config.min_num_spatial_mask_spans,
)
def create_model(self) -> Wav2Vec2LlamaModel:
encoder_frontend, encoder = self.create_encoder()
masker = (
self.create_masker()
if self._config.wav2vec_ctc_config.use_masking
else None
)
encoder_proj = Linear(
self._config.wav2vec_ctc_config.encoder_config.model_dim
* self._config.encoder_stacking,
self._config.llama_config.model_dim,
bias=True,
)
text_frontend = StandardEmbedding(
num_embeddings=self._config.llama_config.vocab_info.size,
embedding_dim=self._config.llama_config.model_dim,
)
llama_decoder = LLaMAFactory(self._config.llama_config).create_decoder()
final_proj = Linear(
self._config.llama_config.model_dim,
self._config.llama_config.vocab_info.size,
bias=False,
)
return Wav2Vec2LlamaModel(
encoder_frontend=encoder_frontend,
encoder=encoder,
encoder_proj=encoder_proj,
text_frontend=text_frontend,
llama_decoder=llama_decoder,
final_proj=final_proj,
target_vocab_info=self._config.wav2vec_ctc_config.vocab_info,
masker=masker,
final_dropout_p=self._config.wav2vec_ctc_config.final_dropout_p,
max_generation_length=self._config.llama_config.max_seq_len,
encoder_stacking=self._config.encoder_stacking,
frozen_encoder=self._config.frozen_encoder,
)
"""Configs"""
from dataclasses import dataclass, field
from typing import Final
from fairseq2.context import RuntimeContext
from fairseq2.data import VocabularyInfo
from fairseq2.models.wav2vec2 import Wav2Vec2EncoderConfig
WAV2VEC2_ASR_MODEL_FAMILY: Final = "wav2vec2_asr"
@dataclass(kw_only=True)
class Wav2Vec2AsrConfig:
"""Holds the configuration of a wav2vec 2.0 ASR model.
The default values correspond to the base 10h architecture as described in
:cite:t:`https://doi.org/10.48550/arxiv.2006.11477`.
"""
encoder_config: Wav2Vec2EncoderConfig = field(
default_factory=lambda: Wav2Vec2EncoderConfig(
feature_gradient_scale=1.0,
dropout_p=0.0,
attn_dropout_p=0.0,
ffn_inner_dropout_p=0.1,
)
)
"""The configuration of the encoder."""
vocab_info: VocabularyInfo = field(
default_factory=lambda: VocabularyInfo(
size=32, unk_idx=3, bos_idx=0, eos_idx=2, pad_idx=1
)
)
"""The vocabulary information."""
final_dropout_p: float = 0.0
"""The dropout probability on the output of the encoder."""
# Mask
mask_codebase: str = "fairseq2"
use_masking: bool = True
"""If ``True``, masks features as regularization."""
temporal_mask_span_len: int = 10
"""The length of each temporal mask span that is applied over time steps."""
max_temporal_mask_prob: float = 0.69
"""The maximum probability of masking a time step. Note that, due to mask
span overlap, the effective probability will be lower."""
min_num_temporal_mask_spans: int = 2
"""The minimum number of temporal masks sampled per sequence."""
spatial_mask_span_len: int = 64
"""The length of each spatial mask span that is applied over features."""
max_spatial_mask_prob: float = 0.55
"""The maximum probability of masking a feature. Note that, due to mask span
overlap, the effective probability will be lower."""
min_num_spatial_mask_spans: int = 2
"""The minimum number of spatial masks sampled per sequence."""
def register_wav2vec2_asr_configs(context: RuntimeContext) -> None:
registry = context.get_config_registry(Wav2Vec2AsrConfig)
wav2vec2_asr_arch = registry.decorator
w2v2_encoder_registry = context.get_config_registry(Wav2Vec2EncoderConfig)
@wav2vec2_asr_arch("base_10h")
def base_10h() -> Wav2Vec2AsrConfig:
return Wav2Vec2AsrConfig()
@wav2vec2_asr_arch("base_100h")
def base_100h() -> Wav2Vec2AsrConfig:
config = base_10h()
config.encoder_config.layer_drop_p = 0.1
return config
@wav2vec2_asr_arch("large_10h")
def large_10h() -> Wav2Vec2AsrConfig:
config = base_10h()
config.encoder_config = w2v2_encoder_registry.get("large")
config.encoder_config.feature_gradient_scale = 1.0
config.encoder_config.dropout_p = 0.0
config.encoder_config.attn_dropout_p = 0.0
config.encoder_config.ffn_inner_dropout_p = 0.1
config.encoder_config.layer_drop_p = 0.1
config.max_temporal_mask_prob = 0.80
config.max_spatial_mask_prob = 0.30
return config
@wav2vec2_asr_arch("large_100h")
def large_100h() -> Wav2Vec2AsrConfig:
config = large_10h()
config.max_temporal_mask_prob = 0.53
config.max_spatial_mask_prob = 0.55
return config
@wav2vec2_asr_arch("large_lv60k_10h")
def large_lv60k_10h() -> Wav2Vec2AsrConfig:
config = base_10h()
config.encoder_config = w2v2_encoder_registry.get("large_lv60k")
config.encoder_config.feature_gradient_scale = 1.0
config.encoder_config.dropout_p = 0.0
config.encoder_config.attn_dropout_p = 0.0
config.encoder_config.ffn_inner_dropout_p = 0.1
config.encoder_config.layer_drop_p = 0.1
config.max_temporal_mask_prob = 0.80
config.max_spatial_mask_prob = 0.30
return config
@wav2vec2_asr_arch("large_lv60k_100h")
def large_lv60k_100h() -> Wav2Vec2AsrConfig:
config = large_lv60k_10h()
config.max_temporal_mask_prob = 0.53
config.max_spatial_mask_prob = 0.55
return config
@wav2vec2_asr_arch("300m_bib61")
def bib61_300m() -> Wav2Vec2AsrConfig:
config = base_10h()
config.encoder_config = w2v2_encoder_registry.get("large_lv60k")
config.encoder_config.feature_gradient_scale = 1.0
config.encoder_config.dropout_p = 0.0
config.encoder_config.attn_dropout_p = 0.0
config.encoder_config.ffn_inner_dropout_p = 0.1
config.encoder_config.layer_drop_p = 0.1
config.use_masking = False
config.max_temporal_mask_prob = 0.0
config.max_spatial_mask_prob = 0.0
config.vocab_info.size = 2475
return config
@wav2vec2_asr_arch("300m_bib1143")
def bib1143_300m() -> Wav2Vec2AsrConfig:
config = bib61_300m()
config.vocab_info.size = 3335
return config
@wav2vec2_asr_arch("1b_bib61")
def bib61_1b() -> Wav2Vec2AsrConfig:
config = base_10h()
config.encoder_config = w2v2_encoder_registry.get("1b")
config.encoder_config.feature_gradient_scale = 1.0
config.encoder_config.dropout_p = 0.0
config.encoder_config.attn_dropout_p = 0.0
config.encoder_config.ffn_inner_dropout_p = 0.1
config.encoder_config.layer_drop_p = 0.1
config.use_masking = False
config.max_temporal_mask_prob = 0.0
config.max_spatial_mask_prob = 0.0
config.vocab_info.size = 2475
return config
@wav2vec2_asr_arch("1b_llama_bib61")
def llama_bib61_1b() -> Wav2Vec2AsrConfig:
config = base_10h()
config.encoder_config = w2v2_encoder_registry.get("1b_llama")
config.encoder_config.feature_gradient_scale = 1.0
config.encoder_config.dropout_p = 0.0
config.encoder_config.attn_dropout_p = 0.0
config.encoder_config.ffn_inner_dropout_p = 0.1
config.encoder_config.layer_drop_p = 0.1
config.use_masking = False
config.max_temporal_mask_prob = 0.0
config.max_spatial_mask_prob = 0.0
config.vocab_info.size = 2475
return config
@wav2vec2_asr_arch("2b_bib61")
def bib61_2b() -> Wav2Vec2AsrConfig:
config = base_10h()
config.encoder_config = w2v2_encoder_registry.get("2b")
config.encoder_config.feature_gradient_scale = 1.0
config.encoder_config.dropout_p = 0.0
config.encoder_config.attn_dropout_p = 0.0
config.encoder_config.ffn_inner_dropout_p = 0.1
config.encoder_config.layer_drop_p = 0.1
config.use_masking = False
config.max_temporal_mask_prob = 0.0
config.max_spatial_mask_prob = 0.0
config.vocab_info.size = 2475
return config
@wav2vec2_asr_arch("3b_bib61")
def bib61_3b() -> Wav2Vec2AsrConfig:
config = base_10h()
config.encoder_config = w2v2_encoder_registry.get("3b")
config.encoder_config.feature_gradient_scale = 1.0
config.encoder_config.dropout_p = 0.0
config.encoder_config.attn_dropout_p = 0.0
config.encoder_config.ffn_inner_dropout_p = 0.1
config.encoder_config.layer_drop_p = 0.1
config.use_masking = False
config.max_temporal_mask_prob = 0.0
config.max_spatial_mask_prob = 0.0
config.vocab_info.size = 2475
return config
@wav2vec2_asr_arch("5b_bib61")
def bib61_5b() -> Wav2Vec2AsrConfig:
config = base_10h()
config.encoder_config = w2v2_encoder_registry.get("5b")
config.encoder_config.feature_gradient_scale = 1.0
config.encoder_config.dropout_p = 0.0
config.encoder_config.attn_dropout_p = 0.0
config.encoder_config.ffn_inner_dropout_p = 0.1
config.encoder_config.layer_drop_p = 0.1
config.use_masking = False
config.max_temporal_mask_prob = 0.0
config.max_spatial_mask_prob = 0.0
config.vocab_info.size = 2475
return config
@wav2vec2_asr_arch("7b_bib61")
def bib61_7b() -> Wav2Vec2AsrConfig:
config = base_10h()
config.encoder_config = w2v2_encoder_registry.get("7b")
config.encoder_config.feature_gradient_scale = 1.0
config.encoder_config.dropout_p = 0.0
config.encoder_config.attn_dropout_p = 0.0
config.encoder_config.ffn_inner_dropout_p = 0.1
config.encoder_config.layer_drop_p = 0.1
config.use_masking = False
config.max_temporal_mask_prob = 0.0
config.max_spatial_mask_prob = 0.0
config.vocab_info.size = 2475
return config
@wav2vec2_asr_arch("3.25b_bib61")
def higher_bib61_3b() -> Wav2Vec2AsrConfig:
config = base_10h()
config.encoder_config = w2v2_encoder_registry.get("3.25b")
config.encoder_config.feature_gradient_scale = 1.0
config.encoder_config.dropout_p = 0.0
config.encoder_config.attn_dropout_p = 0.0
config.encoder_config.ffn_inner_dropout_p = 0.1
config.encoder_config.layer_drop_p = 0.1
config.use_masking = False
config.max_temporal_mask_prob = 0.0
config.max_spatial_mask_prob = 0.0
config.vocab_info.size = 2475
return config
@wav2vec2_asr_arch("5b_front51")
def front51_5b() -> Wav2Vec2AsrConfig:
config = base_10h()
config.encoder_config = w2v2_encoder_registry.get("5b")
config.encoder_config.feature_gradient_scale = 1.0
config.encoder_config.dropout_p = 0.0
config.encoder_config.attn_dropout_p = 0.0
config.encoder_config.ffn_inner_dropout_p = 0.1
config.encoder_config.layer_drop_p = 0.1
config.use_masking = False
config.max_temporal_mask_prob = 0.0
config.max_spatial_mask_prob = 0.0
config.vocab_info.size = 222
return config
@wav2vec2_asr_arch("7b_front51")
def front51_7b() -> Wav2Vec2AsrConfig:
config = base_10h()
config.encoder_config = w2v2_encoder_registry.get("7b")
config.encoder_config.feature_gradient_scale = 1.0
config.encoder_config.dropout_p = 0.0
config.encoder_config.attn_dropout_p = 0.0
config.encoder_config.ffn_inner_dropout_p = 0.1
config.encoder_config.layer_drop_p = 0.1
config.use_masking = False
config.max_temporal_mask_prob = 0.0
config.max_spatial_mask_prob = 0.0
config.vocab_info.size = 222
return config
@wav2vec2_asr_arch("1b_bib1143")
def bib1143_1b() -> Wav2Vec2AsrConfig:
config = base_10h()
config.encoder_config = w2v2_encoder_registry.get("1b")
config.encoder_config.feature_gradient_scale = 1.0
config.encoder_config.dropout_p = 0.0
config.encoder_config.attn_dropout_p = 0.0
config.encoder_config.ffn_inner_dropout_p = 0.1
config.encoder_config.layer_drop_p = 0.1
config.use_masking = False
config.max_temporal_mask_prob = 0.0
config.max_spatial_mask_prob = 0.0
config.vocab_info.size = 3335
return config
@wav2vec2_asr_arch("3b_bib1143")
def bib1143_3b() -> Wav2Vec2AsrConfig:
config = base_10h()
config.encoder_config = w2v2_encoder_registry.get("3b")
config.encoder_config.feature_gradient_scale = 1.0
config.encoder_config.dropout_p = 0.0
config.encoder_config.attn_dropout_p = 0.0
config.encoder_config.ffn_inner_dropout_p = 0.1
config.encoder_config.layer_drop_p = 0.1
config.use_masking = False
config.max_temporal_mask_prob = 0.0
config.max_spatial_mask_prob = 0.0
config.vocab_info.size = 3335
return config
@wav2vec2_asr_arch("5b_bib1143")
def bib1143_5b() -> Wav2Vec2AsrConfig:
config = base_10h()
config.encoder_config = w2v2_encoder_registry.get("5b")
config.encoder_config.feature_gradient_scale = 1.0
config.encoder_config.dropout_p = 0.0
config.encoder_config.attn_dropout_p = 0.0
config.encoder_config.ffn_inner_dropout_p = 0.1
config.encoder_config.layer_drop_p = 0.1
config.use_masking = False
config.max_temporal_mask_prob = 0.0
config.max_spatial_mask_prob = 0.0
config.vocab_info.size = 3335 # following bibfront1194's vocab size
return config
@wav2vec2_asr_arch("7b_bib1143")
def bib1143_7b() -> Wav2Vec2AsrConfig:
config = base_10h()
config.encoder_config = w2v2_encoder_registry.get("7b")
config.encoder_config.feature_gradient_scale = 1.0
config.encoder_config.dropout_p = 0.0
config.encoder_config.attn_dropout_p = 0.0
config.encoder_config.ffn_inner_dropout_p = 0.1
config.encoder_config.layer_drop_p = 0.1
config.use_masking = False
config.max_temporal_mask_prob = 0.0
config.max_spatial_mask_prob = 0.0
config.vocab_info.size = 3335
return config
from dataclasses import dataclass, field
from typing import Final
from fairseq2.context import RuntimeContext
from fairseq2.nn.transformer import TransformerNormOrder
from fairseq2.utils.validation import ValidationError, ValidationResult
WAV2VEC2_MODEL_FAMILY: Final = "wav2vec2"
@dataclass(kw_only=True)
class Wav2Vec2Config:
"""Holds the configuration of a wav2vec 2.0 model.
The default values correspond to the base architecture as described in
:cite:t:`https://doi.org/10.48550/arxiv.2006.11477`.
"""
encoder_config: Wav2Vec2EncoderConfig = field(
default_factory=lambda: Wav2Vec2EncoderConfig()
)
"""The configuration of the wav2vec 2.0 encoder."""
final_dim: int = 256
"""The dimensionality of the final projection that is applied to context
network outputs and quantized targets."""
final_proj_bias: bool = True
"""If ``True``, the final projection learns an additive bias."""
quantizer_encoder_grad: bool = True
"""If ``True``, gradients are propagated from the quantizer through the convolutional
encoder. Otherwise, they are detached and the encoder is only trained with gradients
from the transformer. """
# Mask
mask_codebase: str = "fairseq2"
temporal_mask_span_len: int = 10
"""The length of each temporal mask span that is applied over time steps."""
max_temporal_mask_prob: float = 0.69
"""The maximum probability of masking a time step. Note that, due to mask
span overlap, the effective probability will be lower."""
min_num_temporal_mask_spans: int = 2
"""The minimum number of temporal masks sampled per sequence."""
spatial_mask_span_len: int = 10
"""The length of each spatial mask span that is applied over features."""
max_spatial_mask_prob: float = 0.0
"""The maximum probability of masking a feature. Note that, due to mask span
overlap, the effective probability will be lower."""
min_num_spatial_mask_spans: int = 2
"""The minimum number of spatial masks sampled per sequence."""
# Quantization
quantized_dim: int = 256
"""The output dimensionality of vector quantizer."""
num_codebooks: int = 2
"""The number of codebooks."""
num_codebook_entries: int = 320
"""The number of entries per codebook."""
codebook_sampling_temperature: tuple[float, float, float] = (2.0, 0.5, 0.999995)
"""A tuple of start temperature, end temperature, and decay factor for
codebook entry sampling."""
# Loss
num_distractors: int = 100
"""The number of distractors to use in contrastive prediction."""
logit_temp: float = 0.1
"""The temperature to divide logits by."""
@dataclass(kw_only=True)
class Wav2Vec2EncoderConfig:
"""Holds the configuration of a wav2vec 2.0 encoder.
The default values correspond to the base architecture described in
:cite:t:`https://doi.org/10.48550/arxiv.2006.11477`.
"""
model_dim: int = 768
"""The dimensionality of the model."""
max_seq_len: int = 4096
"""The maximum sequence length after feature extraction."""
# Features
feature_dim: int = 512
"""The dimensionality of extracted features."""
use_fbank: bool = False
"""If ``True``, uses log-mel filterbanks instead of waveforms as input."""
first_pass_dropout_p: float = 0.0
"""The dropout probability on extracted features before masking and
positional encoding."""
layer_norm_features: bool = True
"""If ``True``, applies Layer Normalization to extracted features."""
# Waveform Feature Extractor
feature_extractor_layer_descs: list[tuple[int, int, int]] = field(
default_factory=lambda: [(512, 10, 5)] + [(512, 3, 2)] * 4 + [(512, 2, 2)] * 2
)
"""A tuple of output dimension, kernel size, and stride for each feature
extraction layer."""
feature_extractor_bias: bool = False
"""If ``True``, convolutions in feature extraction layers learn an additive
bias."""
feature_extractor_layer_norm_convs: bool = False
"""If ``True``, applies Layer Normalization to outputs of convolutions in
feature extraction layers."""
feature_gradient_scale: float = 0.1
"""The scale factor for gradients of extracted features. Setting to a value
less than 1.0 allows the feature extractor to learn at a lower rate than the
rest of the model."""
# Filterbank Feature Extractor
num_fbank_channels: int = 0
"""The number of source log-mel filterbank channels."""
fbank_stride: int = 0
sample_fbank_every_k: int = 0
# Position Encoder
pos_encoder_type: str = "conv"
"""The type of position encoder ('conv', 'relative', 'rotary')."""
# Convolutional Position Encoder
pos_encoder_depth: int = 1
"""The number of stacked position encoder layers."""
pos_conv_kernel_size: int = 128
"""The total kernel size of 1D convolutions in position encoder layers."""
num_pos_conv_groups: int = 16
"""The number of convolution groups in position encoder layers."""
# Encoder (i.e. Context Network)
use_conformer: bool = False
"""If ``True``, uses Conformer blocks instead of Transformer encoder layers."""
num_encoder_layers: int = 12
"""The number of encoder layers."""
num_encoder_attn_heads: int = 12
"""The number of attention heads in encoder layers."""
ffn_inner_dim: int = 3072
"""The inner dimensionality of feed-forward networks."""
dropout_p: float = 0.1
"""The dropout probability on outputs of Transformer layers."""
attn_dropout_p: float = 0.1
"""The dropout probability on attention weights."""
ffn_inner_dropout_p: float = 0.0
"""The dropout probability on inner activations of feed-forward networks."""
layer_drop_p: float = 0.05
"""If greater than zero, applies LayerDrop to encoder layers as described in
:cite:t:`https://doi.org/10.48550/arxiv.1909.11556`."""
norm_order: TransformerNormOrder = TransformerNormOrder.POST
"""The Layer Normalization order."""
depthwise_conv_kernel_size: int = 0
"""The kernel size of depthwise convolutions in Conformer blocks."""
def validate(self) -> None:
result = ValidationResult()
if self.use_conformer and self.norm_order != TransformerNormOrder.POST:
result.add_error(
f"`norm_order` must be `POST` when `use_conformer` is `True`, but is `{self.norm_order}` instead."
)
if result.has_error:
raise ValidationError(
"The wav2vec 2.0 encoder configuration has one or more validation errors:", result # fmt: skip
)
def register_wav2vec2_configs(context: RuntimeContext) -> None:
arch = context.get_config_registry(Wav2Vec2Config).decorator
arch_encoder = context.get_config_registry(Wav2Vec2EncoderConfig).decorator
@arch("base")
def base() -> Wav2Vec2Config:
return Wav2Vec2Config()
@arch_encoder("base")
def base_encoder() -> Wav2Vec2EncoderConfig:
return base().encoder_config
@arch("large")
def large() -> Wav2Vec2Config:
config = base()
config.encoder_config.model_dim = 1024
config.encoder_config.num_encoder_layers = 24
config.encoder_config.num_encoder_attn_heads = 16
config.encoder_config.ffn_inner_dim = 4096
config.encoder_config.dropout_p = 0.0
config.encoder_config.layer_drop_p = 0.2
config.quantized_dim = 768
config.final_dim = 768
return config
@arch_encoder("large")
def large_encoder() -> Wav2Vec2EncoderConfig:
return large().encoder_config
@arch("large_lv60k")
def large_lv60k() -> Wav2Vec2Config:
config = large()
config.encoder_config.layer_norm_features = False
config.encoder_config.feature_extractor_bias = True
config.encoder_config.feature_extractor_layer_norm_convs = True
config.encoder_config.layer_drop_p = 0.0
config.encoder_config.norm_order = TransformerNormOrder.PRE
config.codebook_sampling_temperature = (2.0, 0.1, 0.999995)
return config
@arch_encoder("large_lv60k")
def large_lv60k_encoder() -> Wav2Vec2EncoderConfig:
return large_lv60k().encoder_config
@arch("xlsr_base")
def xlsr_base() -> Wav2Vec2Config:
config = large_lv60k()
config.encoder_config.attn_dropout_p = 0.0
config.encoder_config.feature_gradient_scale = 1.0
return config
@arch_encoder("xlsr_base")
def xlsr_base_encoder() -> Wav2Vec2EncoderConfig:
return xlsr_base().encoder_config
@arch("base_conformer")
def base_conformer() -> Wav2Vec2Config:
config = xlsr_base()
config.encoder_config.use_conformer = True
config.encoder_config.norm_order = TransformerNormOrder.POST
config.encoder_config.depthwise_conv_kernel_size = 31
# pos_encoder_type
return config
@arch_encoder("base_conformer")
def base_conformer_encoder() -> Wav2Vec2EncoderConfig:
return base_conformer().encoder_config
@arch("tiny")
def tiny() -> Wav2Vec2Config:
config = xlsr_base()
config.encoder_config.model_dim = 1280
config.encoder_config.num_encoder_layers = 4
config.encoder_config.ffn_inner_dim = 1280
config.encoder_config.dropout_p = 0.0
config.quantized_dim = 512
config.final_dim = 512
config.encoder_config.first_pass_dropout_p = 0.1
return config
@arch_encoder("tiny")
def tiny_encoder() -> Wav2Vec2EncoderConfig:
return tiny().encoder_config
@arch("1b")
def b1() -> Wav2Vec2Config:
config = xlsr_base()
config.encoder_config.model_dim = 1280
config.encoder_config.num_encoder_layers = 48
config.encoder_config.ffn_inner_dim = 5120
config.encoder_config.dropout_p = 0.0
config.quantized_dim = 1024
config.final_dim = 1024
config.encoder_config.first_pass_dropout_p = 0.1
return config
@arch_encoder("1b")
def b1_encoder() -> Wav2Vec2EncoderConfig:
return b1().encoder_config
@arch("2b")
def b2() -> Wav2Vec2Config:
config = b1()
config.encoder_config.model_dim = 1920
config.encoder_config.ffn_inner_dim = 7680
return config
@arch_encoder("2b")
def b2_encoder() -> Wav2Vec2EncoderConfig:
return b2().encoder_config
@arch("3b")
def b3() -> Wav2Vec2Config:
config = b1()
config.encoder_config.num_encoder_layers = 60
config.encoder_config.model_dim = 2048
config.encoder_config.ffn_inner_dim = 8192
return config
@arch_encoder("3b")
def b3_encoder() -> Wav2Vec2EncoderConfig:
return b3().encoder_config
@arch("3b_mel")
def mel_3b() -> Wav2Vec2Config:
config = b3()
config.encoder_config.use_fbank = True
config.encoder_config.num_fbank_channels = 80
config.encoder_config.fbank_stride = 2
config.encoder_config.sample_fbank_every_k = 1
config.encoder_config.feature_dim = 160
return config
@arch_encoder("3b_mel")
def mel_3b_encoder() -> Wav2Vec2EncoderConfig:
return mel_3b().encoder_config
@arch("3.25b")
def higher_3b() -> Wav2Vec2Config:
config = b1()
config.encoder_config.num_encoder_layers = 64
config.encoder_config.model_dim = 2048
config.encoder_config.ffn_inner_dim = 8192
config.encoder_config.num_encoder_attn_heads = 32
config.quantized_dim = 1280
config.final_dim = 1280
return config
@arch_encoder("3.25b")
def higher_3b_encoder() -> Wav2Vec2EncoderConfig:
return higher_3b().encoder_config
@arch("4b")
def b4() -> Wav2Vec2Config:
config = b2()
config.quantized_dim = 1280
config.final_dim = 1280
config.encoder_config.num_encoder_layers = 64
config.encoder_config.model_dim = 2304
config.encoder_config.ffn_inner_dim = 9216
config.encoder_config.num_encoder_attn_heads = 32
return config
@arch_encoder("4b")
def b4_encoder() -> Wav2Vec2EncoderConfig:
return b4().encoder_config
@arch("1b_llama")
def llama_1b() -> Wav2Vec2Config:
config = xlsr_base()
config.encoder_config.model_dim = 2048
config.encoder_config.num_encoder_layers = 16
config.encoder_config.ffn_inner_dim = int(2048 * 4 * 1.5)
config.encoder_config.num_encoder_attn_heads = 32
config.encoder_config.dropout_p = 0.0
config.quantized_dim = 1024
config.final_dim = 1024
config.encoder_config.first_pass_dropout_p = 0.1
return config
@arch_encoder("1b_llama")
def llama_1b_encoder() -> Wav2Vec2EncoderConfig:
return llama_1b().encoder_config
@arch("3b_llama")
def llama_3b() -> Wav2Vec2Config:
config = llama_1b()
config.encoder_config.model_dim = 2560
config.encoder_config.num_encoder_layers = 32
config.encoder_config.ffn_inner_dim = int(2560 * 4 * 1.0)
config.quantized_dim = 2048
config.final_dim = 2048
return config
@arch_encoder("3b_llama")
def llama_3b_encoder() -> Wav2Vec2EncoderConfig:
return llama_3b().encoder_config
@arch("5b")
def b5() -> Wav2Vec2Config:
config = b3()
config.encoder_config.num_encoder_layers = 96
config.encoder_config.model_dim = 2048
config.encoder_config.ffn_inner_dim = 8192
config.encoder_config.num_encoder_attn_heads = 16
config.quantized_dim = 1024
config.final_dim = 1024
return config
@arch_encoder("5b")
def b5_encoder() -> Wav2Vec2EncoderConfig:
return b5().encoder_config
@arch("7b")
def b7() -> Wav2Vec2Config:
config = b5()
config.encoder_config.num_encoder_layers = 128
config.encoder_config.model_dim = 2048
config.encoder_config.ffn_inner_dim = 8192
config.encoder_config.num_encoder_attn_heads = 16
config.quantized_dim = 1024
config.final_dim = 1024
return config
@arch_encoder("7b")
def b7_encoder() -> Wav2Vec2EncoderConfig:
return b7().encoder_config
# @title Create model and load weights
"""Create model and load weights"""
from dataclasses import field
import torch
from fairseq2 import setup_fairseq2
from fairseq2.context import get_runtime_context
from fairseq2.data.text.tokenizers.sentencepiece import RawSentencePieceTokenizer
class Wav2Vec2LlamaConfig:
wav2vec_ctc_config: Wav2Vec2AsrConfig = field()
llama_config: LLaMAConfig = field()
encoder_stacking: int = 1
frozen_encoder: bool = False
def load_mms_model(ckpt_path: str, tokenizer_path: str, device=None):
"""
Load the MMS model and tokenizer from checkpoint files with memory optimization.
Args:
ckpt_path (str): Path to the model checkpoint file
tokenizer_path (str): Path to the tokenizer model file
device: Device to load the model on. If None, auto-detects GPU/CPU
Returns:
tuple: (model, text_decoder, device) where:
- model: The loaded and configured MMS model
- text_decoder: The tokenizer decoder
- device: The device the model is loaded on
"""
import gc
import os
import psutil
logger = logging.getLogger(__name__)
def log_memory_usage(step: str):
"""Log current memory usage."""
process = psutil.Process(os.getpid())
memory_info = process.memory_info()
virtual_memory = psutil.virtual_memory()
logger.info(
f"[{step}] Process RSS: {memory_info.rss / (1024**3):.2f} GB, "
f"System Available: {virtual_memory.available / (1024**3):.2f} GB"
)
logger.info(f"Starting MMS model loading process...")
logger.info(f"Checkpoint path: {ckpt_path}")
logger.info(f"Tokenizer path: {tokenizer_path}")
# Check file size
if os.path.exists(ckpt_path):
ckpt_size_gb = os.path.getsize(ckpt_path) / (1024**3)
logger.info(f"Checkpoint file size: {ckpt_size_gb:.2f} GB")
log_memory_usage("Initial")
# Set device with proper CUDA initialization
if device is None:
try:
# Initialize CUDA context properly
logger.info("Checking CUDA availability...")
if torch.cuda.is_available():
logger.info(
f"CUDA is available. Device count: {torch.cuda.device_count()}"
)
# Initialize CUDA context
torch.cuda.init()
# Set device to first available GPU
device = torch.device("cuda:0")
logger.info(f"CUDA device name: {torch.cuda.get_device_name(0)}")
logger.info(
f"CUDA device memory: {torch.cuda.get_device_properties(0).total_memory / 1e9:.1f} GB"
)
else:
logger.warning("CUDA is not available, falling back to CPU")
device = torch.device("cpu")
except Exception as e:
logger.warning(f"CUDA initialization failed: {e}, falling back to CPU")
device = torch.device("cpu")
else:
device = torch.device("cpu") # Force CPU for memory efficiency
logger.info(f"Using device: {device}")
# Load model parameters from checkpoint with memory optimization
logger.info("Loading model parameters from checkpoint...")
try:
# Try memory-mapped loading first (more memory efficient)
logger.info("Attempting memory-mapped loading...")
model_params = torch.load(ckpt_path, map_location="cpu", mmap=True)
logger.info("β Model parameters loaded successfully (memory-mapped)")
except Exception as e:
logger.warning(f"Memory-mapped loading failed: {e}")
logger.info("Falling back to regular loading...")
# Force garbage collection before loading
gc.collect()
model_params = torch.load(ckpt_path, map_location="cpu")
logger.info("β Model parameters loaded successfully (regular)")
log_memory_usage("After checkpoint load")
# Create context
logger.info("Setting up fairseq2 context and registering configs...")
setup_fairseq2()
context = get_runtime_context()
try:
register_wav2vec2_configs(context)
register_wav2vec2_asr_configs(context)
logger.info("β Configs registered successfully")
except Exception as e:
logger.warning(f"Config registration failed (may already be registered): {e}")
w2v2_ctc_registry = context.get_config_registry(Wav2Vec2AsrConfig)
# Create config
logger.info("Creating model configuration...")
wav2vec_ctc_config = w2v2_ctc_registry.get("7b_bib1143")
logger.info(
f"β wav2vec config loaded: vocab_size={wav2vec_ctc_config.vocab_info.size}"
)
llama_config = LLaMAConfig(
model_dim=4096,
max_seq_len=8192,
vocab_info=wav2vec_ctc_config.vocab_info,
num_layers=12,
num_attn_heads=8,
num_key_value_heads=8,
ffn_inner_dim=4096,
rope_theta=10_000.0,
dropout_p=0.1,
)
logger.info(
f"β LLaMA config created: model_dim={llama_config.model_dim}, layers={llama_config.num_layers}"
)
config = Wav2Vec2LlamaConfig()
config.wav2vec_ctc_config = wav2vec_ctc_config
config.llama_config = llama_config
# Instantiate model
logger.info("Instantiating model from factory...")
factory = Wav2Vec2LlamaFactory(config)
model = factory.create_model()
logger.info("β Model instantiated successfully")
# Load state dict from ckpt
logger.info("Loading model state dictionary...")
model.load_state_dict(model_params["model"])
del model_params
logger.info("β Model weights loaded successfully")
# Move to device and set eval mode
logger.info(f"Moving model to device {device} and setting eval mode...")
model = model.to(device).eval()
logger.info("β Model moved to device and set to eval mode")
# Create tokenizer
logger.info(f"Creating tokenizer from {tokenizer_path}...")
tokenizer = RawSentencePieceTokenizer(tokenizer_path)
text_decoder_1143 = tokenizer.create_decoder()
logger.info("β Tokenizer created successfully")
logger.info("MMS model loading completed successfully!")
return model, text_decoder_1143, device
def prepare_audio_batch(wav_path: str, device, max_duration_seconds=2):
"""
Load a wav file from disk and prepare batch for model inference.
Args:
wav_path (str): Path to the WAV file
device: Device to place the batch on
max_duration_seconds (int): Maximum duration to process (for efficiency)
Returns:
Seq2SeqBatch: Prepared batch for model inference
"""
logger = logging.getLogger(__name__)
logger.info(f"Preparing audio batch from: {wav_path}")
logger.info(f"Max duration: {max_duration_seconds}s, target device: {device}")
# Load the WAV file, resample the data to 16 kHz
logger.info("Loading and resampling audio file...")
data, fs = librosa.load(wav_path)
logger.info(f"Original sample rate: {fs} Hz, duration: {len(data)/fs:.2f}s")
data = librosa.resample(data, orig_sr=fs, target_sr=16000)
logger.info("β Audio resampled to 16kHz")
# If the data is multi-channel, merge all channels
if len(data.shape) > 1:
logger.info("Multi-channel audio detected, merging channels...")
data = np.mean(data, axis=0)
else:
data = data
# Cut to specified duration (for efficiency)
if max_duration_seconds > 0:
original_length = len(data)
data = data[: 16000 * max_duration_seconds]
if len(data) < original_length:
logger.info(
f"Audio truncated from {original_length/16000:.2f}s to {len(data)/16000:.2f}s"
)
# Convert to tensor and normalize
logger.info("Converting to tensor and normalizing...")
# Originally data = torch.Tensor(data).to(torch.bfloat16)
data = torch.Tensor(data).float() # Use float32 to match model expectations
data = F.layer_norm(data, data.shape)
# Create batch
logger.info("Creating batch for inference...")
batch = Seq2SeqBatch(
source_seqs=data.unsqueeze(0).to(device),
source_padding_mask=None,
target_seqs=torch.tensor([1], dtype=torch.long)
.unsqueeze(0)
.to(device), # Not used for inference
target_padding_mask=None,
example=[],
)
logger.info(
f"β Audio batch prepared successfully, shape: {batch.source_seqs.shape}"
)
return batch
def run_inference(model, batch, text_decoder, config, device):
"""
Run model inference on a prepared batch.
Args:
model: The loaded MMS model
batch: Prepared audio batch
text_decoder: Tokenizer decoder
config: Model configuration
device: Device for inference
Returns:
list: Decoded text outputs
"""
logger = logging.getLogger(__name__)
logger.info("Starting model inference...")
logger.info(f"Input batch shape: {batch.source_seqs.shape}, device: {device}")
with torch.no_grad():
ctx = (
torch.cuda.amp.autocast()
if torch.cuda.is_available()
else torch.cpu.amp.autocast(dtype=torch.bfloat16)
)
logger.info(
f"Using autocast context: {'CUDA' if torch.cuda.is_available() else 'CPU'}"
)
with ctx:
logger.info("Running forward pass...")
output = model(batch)
logger.info("β Forward pass completed")
logger.info("Generating hypotheses...")
hyp_seq, hyp_padding_mask = output.generate_hypotheses(
pad_idx=config.llama_config.vocab_info.pad_idx
)
logger.info(f"β Generated {len(hyp_seq)} hypotheses")
logger.info("Decoding text...")
results = [text_decoder(s) for s in hyp_seq]
logger.info(f"β Inference completed, results: {results}")
return results
def transcribe_audio(
wav_path: str,
ckpt_path: str = None,
tokenizer_path: str = None,
max_duration_seconds=2,
):
"""
Complete pipeline to transcribe audio using MMS model.
Uses the singleton model instance from server.py to avoid reloading.
Args:
wav_path (str): Path to the WAV file
ckpt_path (str): Path to the model checkpoint (not used, kept for compatibility)
tokenizer_path (str): Path to the tokenizer (not used, kept for compatibility)
max_duration_seconds (int): Maximum duration to process
Returns:
tuple: (transcription_results, audio_data) where:
- transcription_results: list of transcribed text
- audio_data: processed audio tensor for reuse in alignment
"""
logger = logging.getLogger(__name__)
logger.info("Starting complete audio transcription pipeline...")
try:
# Get model from singleton (don't reload)
logger.info("Getting pre-loaded MMS model from singleton...")
from server import get_device, get_model, get_text_decoder
model = get_model()
text_decoder = get_text_decoder()
device = get_device()
if model is None or text_decoder is None or device is None:
raise RuntimeError("Model not properly loaded in server singleton")
logger.info(f"β Using pre-loaded model on device: {device}")
# Get config (needed for inference)
logger.info("Setting up configuration for inference...")
setup_fairseq2()
context = get_runtime_context()
try:
register_wav2vec2_configs(context)
register_wav2vec2_asr_configs(context)
except Exception as e:
logger.warning(f"Config registration warning: {e}")
w2v2_ctc_registry = context.get_config_registry(Wav2Vec2AsrConfig)
wav2vec_ctc_config = w2v2_ctc_registry.get("7b_bib1143")
llama_config = LLaMAConfig(
model_dim=4096,
max_seq_len=8192,
vocab_info=wav2vec_ctc_config.vocab_info,
num_layers=12,
num_attn_heads=8,
num_key_value_heads=8,
ffn_inner_dim=4096,
rope_theta=10_000.0,
dropout_p=0.1,
)
config = Wav2Vec2LlamaConfig()
config.wav2vec_ctc_config = wav2vec_ctc_config
config.llama_config = llama_config
# Prepare batch
logger.info("Preparing audio batch...")
batch = prepare_audio_batch(wav_path, device, max_duration_seconds)
# Extract the processed audio data for return
audio_data = batch.source_seqs.squeeze(0) # Remove batch dimension
# Run inference
logger.info("Running inference...")
results = run_inference(model, batch, text_decoder, config, device)
logger.info(f"Transcription pipeline completed successfully: {results}")
return results, audio_data
except Exception as e:
logger.error(f"Error in transcription pipeline: {str(e)}", exc_info=True)
raise
def normalize_text_with_uroman(text: str) -> str:
"""
Normalize text using uroman for better forced alignment.
Args:
text (str): Input text to normalize
Returns:
str: Normalized text
"""
logger = logging.getLogger(__name__)
try:
# Use uroman to normalize the text
uroman_instance = uroman.Uroman()
normalized = uroman_instance.romanize_string(text)
logger.info(f"Text normalized: '{text}' -> '{normalized}'")
return normalized
except Exception as e:
logger.warning(f"Failed to normalize text with uroman: {e}")
# Fallback to basic normalization
return text.lower().strip()
def perform_forced_alignment(
audio_data: np.ndarray,
transcription_tokens: List[str],
model,
device,
sample_rate: int = 16000,
) -> List[Dict]:
"""
Perform forced alignment using the AudioAlignment class from audio_sentence_alignment.py.
Uses pre-processed audio data from prepare_audio_batch.
Args:
audio_data (np.ndarray): Pre-processed audio data from prepare_audio_batch
transcription_tokens (List[str]): List of tokens from transcription
model: The loaded MMS model (not used directly, AudioAlignment loads its own)
device: Device for computation
sample_rate (int): Audio sample rate
Returns:
List[Dict]: List of segments with timestamps and text
"""
logger = logging.getLogger(__name__)
try:
logger.info(f"Starting forced alignment with pre-processed audio data")
logger.info(f"Audio shape: {audio_data.shape}, sample_rate: {sample_rate}")
logger.info(f"Tokens to align: {transcription_tokens}")
from audio_reading_tools import wav_to_bytes
# Import AudioAlignment and its config classes
from audio_sentence_alignment import (
AlignmentStruct,
AudioAlignment,
AudioAlignmentConfig,
)
# Use the pre-processed audio data directly
# Convert to the format expected by AudioAlignment.get_one_row_alignments
if hasattr(audio_data, "cpu"):
# If it's a torch tensor, use it directly
audio_tensor = audio_data.float()
else:
# If it's numpy, convert to tensor
audio_tensor = torch.from_numpy(audio_data).float()
# Ensure it's 1D (flatten if needed)
if len(audio_tensor.shape) > 1:
audio_tensor = audio_tensor.flatten()
# Convert audio tensor to bytes format expected by AudioAlignment
# Use wav_to_bytes to create proper audio bytes
audio_arr = wav_to_bytes(audio_tensor, sample_rate=sample_rate, format="wav")
logger.info(
f"Converted audio to bytes: shape={audio_arr.shape}, dtype={audio_arr.dtype}"
)
# Preprocess tokens for MMS alignment model using the same approach as TextRomanizer
# The MMS alignment model expects romanized tokens in the same format as text_sentences_tokens
try:
# Join tokens back to text for uroman processing
transcription_text = " ".join(transcription_tokens)
# Import the required functions from TextRomanizer pipeline
from align_utils import get_uroman_tokens
from text_normalization import text_normalize
# Create uroman instance and process the text the same way as TextRomanizer
uroman_instance = uroman.Uroman()
# Step 1: Normalize the text first using text_normalize function (same as TextRomanizer)
normalized_text = text_normalize(transcription_text.strip(), "en")
# Step 2: Get uroman tokens using the same function as TextRomanizer
# This creates character-level tokens with spaces between characters
uroman_tokens_str = get_uroman_tokens(
[normalized_text], uroman_instance, "en"
)[0]
# Step 3: Split by spaces to get individual character tokens (same as real MMS pipeline)
alignment_tokens = uroman_tokens_str.split()
logger.info(f"Original tokens: {transcription_tokens}")
logger.info(f"Original text: '{transcription_text}'")
logger.info(f"Normalized text: '{normalized_text}'")
logger.info(f"Uroman tokens string: '{uroman_tokens_str}'")
logger.info(
f"Alignment tokens (count={len(alignment_tokens)}): {alignment_tokens[:20]}..."
)
# Additional debugging - check for any unusual characters
for i, token in enumerate(alignment_tokens[:10]): # Check first 10 tokens
logger.info(
f"Token {i}: '{token}' (length={len(token)}, chars={[c for c in token]})"
)
except Exception as e:
logger.warning(
f"Failed to preprocess tokens with TextRomanizer approach: {e}"
)
logger.exception("Full error traceback:")
# Fallback: use simple character-level tokenization
transcription_text = " ".join(transcription_tokens).lower()
# Simple character-level tokenization as fallback
alignment_tokens = []
for char in transcription_text:
if char == " ":
alignment_tokens.append(" ")
else:
alignment_tokens.append(char)
logger.info(f"Using fallback character tokens: {alignment_tokens[:20]}...")
logger.info(
f"Using {len(alignment_tokens)} alignment tokens for forced alignment"
)
# Create alignment configuration
alignment_struct = AlignmentStruct(
segement_tokens="tokens",
audio="audio",
)
config = AudioAlignmentConfig(
alignment_column=alignment_struct,
sample_rate=sample_rate,
device=str(device),
use_star=False, # Set to False for standard alignment
)
# Create AudioAlignment instance
logger.info("Creating AudioAlignment instance...")
alignment = AudioAlignment(config)
# Perform alignment using get_one_row_alignments
logger.info("Performing alignment...")
logger.info(f"About to call get_one_row_alignments with:")
logger.info(f" audio_arr type: {type(audio_arr)}, shape: {audio_arr.shape}")
logger.info(
f" alignment_tokens type: {type(alignment_tokens)}, length: {len(alignment_tokens)}"
)
logger.info(
f" First 10 tokens: {alignment_tokens[:10] if len(alignment_tokens) >= 10 else alignment_tokens}"
)
# Check for any problematic characters in tokens
for i, token in enumerate(alignment_tokens[:5]):
token_chars = [ord(c) for c in str(token)]
logger.info(f" Token {i} '{token}' char codes: {token_chars}")
# Check if tokens contain any RTL characters that might cause the LTR assertion
rtl_chars = []
for i, token in enumerate(alignment_tokens):
for char in str(token):
# Check for Arabic, Hebrew, and other RTL characters
if (
"\u0590" <= char <= "\u08ff"
or "\ufb1d" <= char <= "\ufdff"
or "\ufe70" <= char <= "\ufeff"
):
rtl_chars.append((i, token, char, ord(char)))
if rtl_chars:
logger.warning(f"Found RTL characters in tokens: {rtl_chars[:10]}...")
try:
audio_segments = alignment.get_one_row_alignments(
audio_arr, alignment_tokens
)
except Exception as alignment_error:
logger.error(f"Alignment failed with error: {alignment_error}")
logger.error(f"Error type: {type(alignment_error)}")
# Try to provide more context about the error
if "ltr" in str(alignment_error).lower():
logger.error("LTR assertion error detected. This might be due to:")
logger.error("1. RTL characters in the input tokens")
logger.error(
"2. Incorrect token format - tokens should be individual characters"
)
logger.error("3. Unicode normalization issues")
# Try a simple ASCII-only fallback
logger.info("Attempting ASCII-only fallback...")
ascii_tokens = []
for token in alignment_tokens:
# Keep only ASCII characters
ascii_token = "".join(c for c in str(token) if ord(c) < 128)
if ascii_token:
ascii_tokens.append(ascii_token)
logger.info(
f"ASCII tokens (count={len(ascii_tokens)}): {ascii_tokens[:20]}..."
)
try:
audio_segments = alignment.get_one_row_alignments(
audio_arr, ascii_tokens
)
alignment_tokens = ascii_tokens # Update for later use
logger.info("ASCII fallback successful!")
except Exception as ascii_error:
logger.error(f"ASCII fallback also failed: {ascii_error}")
raise alignment_error
else:
raise
logger.info(
f"Alignment completed, got {len(audio_segments)} character segments"
)
# Convert character-level segments back to word-level segments
# Map character segments to original word tokens
aligned_segments = []
transcription_text = " ".join(transcription_tokens)
word_idx = 0
char_idx = 0
for word in transcription_tokens:
if word_idx >= len(transcription_tokens):
break
# Find the start and end character indices for this word
word_start_char = char_idx
word_end_char = char_idx + len(word)
# Find corresponding segments within this character range
word_segments = []
for seg_idx, segment in enumerate(audio_segments):
if seg_idx >= word_start_char and seg_idx < word_end_char:
word_segments.append(segment)
if word_segments:
# Get timing from first and last character segments of the word
start_time = word_segments[0][alignment_struct.segment_start_sec]
last_segment = word_segments[-1]
end_time = (
last_segment[alignment_struct.segment_start_sec]
+ last_segment[alignment_struct.segment_duration]
)
duration = end_time - start_time
else:
# Fallback timing if no segments found
if word_idx < len(audio_segments):
segment = audio_segments[min(word_idx, len(audio_segments) - 1)]
start_time = segment[alignment_struct.segment_start_sec]
duration = segment[alignment_struct.segment_duration]
end_time = start_time + duration
else:
# Final fallback
duration = 0.5 # Default duration
start_time = word_idx * duration
end_time = start_time + duration
aligned_segments.append(
{
"text": word,
"start": start_time,
"end": end_time,
"duration": duration,
}
)
logger.info(
f"Word '{word}': {start_time:.3f}s - {end_time:.3f}s ({duration:.3f}s)"
)
# Update indices
char_idx += len(word)
if (
char_idx < len(transcription_text)
and transcription_text[char_idx] == " "
):
char_idx += 1 # Skip space
word_idx += 1
logger.info(f"Forced alignment completed: {len(aligned_segments)} segments")
return aligned_segments
except Exception as e:
logger.error(f"Error in forced alignment: {str(e)}", exc_info=True)
# Fallback: create uniform timestamps based on audio data length
logger.info("Using fallback uniform timestamps")
try:
# Calculate duration from the audio data
total_duration = (
len(audio_data) / sample_rate
if len(audio_data) > 0
else len(transcription_tokens) * 0.5
)
except:
total_duration = len(transcription_tokens) * 0.5 # Fallback
segment_duration = (
total_duration / len(transcription_tokens) if transcription_tokens else 1.0
)
fallback_segments = []
for i, token in enumerate(transcription_tokens):
start_time = i * segment_duration
end_time = (i + 1) * segment_duration
fallback_segments.append(
{
"text": token,
"start": start_time,
"end": end_time,
"duration": segment_duration,
}
)
logger.info(
f"Using fallback uniform timestamps: {len(fallback_segments)} segments"
)
return fallback_segments
def transcribe_audio_with_alignment(
wav_path: str,
ckpt_path: str = None,
tokenizer_path: str = None,
max_duration_seconds: int = 2,
) -> Dict:
"""
Complete pipeline to transcribe audio and perform forced alignment.
Uses pre-processed audio data from prepare_audio_batch for both steps.
Args:
wav_path (str): Path to the WAV file
ckpt_path (str): Path to the model checkpoint (not used, kept for compatibility)
tokenizer_path (str): Path to the tokenizer (not used, kept for compatibility)
max_duration_seconds (int): Maximum duration to process
Returns:
Dict: Transcription results with alignment information
"""
logger = logging.getLogger(__name__)
try:
# Get model and device first
from server import get_device, get_model
model = get_model()
device = get_device()
if model is None or device is None:
logger.warning(
"Model not available for alignment, returning transcription only"
)
# Get the transcription and processed audio data
transcription_results, audio_data = transcribe_audio(
wav_path, ckpt_path, tokenizer_path, max_duration_seconds
)
if not transcription_results:
return {
"transcription": "",
"tokens": [],
"aligned_segments": [],
"total_duration": 0.0,
}
transcription_text = (
transcription_results[0]
if isinstance(transcription_results, list)
else str(transcription_results)
)
# Tokenize the transcription for alignment
tokens = transcription_text.split() if transcription_text else []
# Perform forced alignment using the same preprocessed audio data
logger.info("Performing forced alignment with preprocessed audio...")
aligned_segments = perform_forced_alignment(audio_data, tokens, model, device)
# Calculate total duration
total_duration = aligned_segments[-1]["end"] if aligned_segments else 0.0
result = {
"transcription": transcription_text,
"tokens": tokens,
"aligned_segments": aligned_segments,
"total_duration": total_duration,
"num_segments": len(aligned_segments),
}
logger.info(
f"Transcription with alignment completed: {len(aligned_segments)} segments, {total_duration:.2f}s total"
)
return result
except Exception as e:
logger.error(f"Error in transcription with alignment: {str(e)}", exc_info=True)
# Return basic transcription without alignment
try:
transcription_results, _ = transcribe_audio(
wav_path, ckpt_path, tokenizer_path, max_duration_seconds
)
transcription_text = (
transcription_results[0] if transcription_results else ""
)
tokens = transcription_text.split() if transcription_text else []
return {
"transcription": transcription_text,
"tokens": tokens,
"aligned_segments": [],
"total_duration": 0.0,
"alignment_error": str(e),
}
except Exception as e2:
logger.error(f"Error in fallback transcription: {str(e2)}", exc_info=True)
return {
"transcription": "",
"tokens": [],
"aligned_segments": [],
"total_duration": 0.0,
"error": str(e2),
}
|