File size: 101,104 Bytes
bb2b5cc | 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 | # Copyright 2022 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.
"""
Processing saving/loading class for common processors.
"""
import bisect
import copy
import inspect
import json
import os
import sys
import typing
from dataclasses import dataclass
from pathlib import Path
from typing import Annotated, Any, Literal, TypedDict, TypeVar, Union
import numpy as np
import typing_extensions
from huggingface_hub import create_repo, is_offline_mode
from huggingface_hub.dataclasses import validate_typed_dict
from huggingface_hub.errors import EntryNotFoundError
from .audio_utils import AudioInput, load_audio
from .dynamic_module_utils import custom_object_save
from .feature_extraction_utils import BatchFeature
from .image_utils import ChannelDimension, ImageInput, is_vision_available
from .tokenization_utils_base import (
PaddingStrategy,
PreTokenizedInput,
PreTrainedTokenizerBase,
TextInput,
TruncationStrategy,
)
from .utils import (
AUDIO_TOKENIZER_NAME,
CHAT_TEMPLATE_DIR,
CHAT_TEMPLATE_FILE,
LEGACY_PROCESSOR_CHAT_TEMPLATE_FILE,
PROCESSOR_NAME,
PushToHubMixin,
TensorType,
cached_file,
copy_func,
direct_transformers_import,
is_torch_available,
list_repo_templates,
logging,
)
from .utils.chat_template_utils import _get_template_variables, render_jinja_template
from .utils.type_validators import (
device_validator,
image_size_validator,
padding_validator,
positive_any_number,
positive_int,
resampling_validator,
tensor_type_validator,
truncation_validator,
video_metadata_validator,
)
from .video_utils import VideoInput, VideoMetadataType
if is_torch_available():
import torch
from .modeling_utils import PreTrainedAudioTokenizerBase
if is_vision_available():
from .image_utils import PILImageResampling
logger = logging.get_logger(__name__)
# type hinting: specifying the type of processor class that inherits from ProcessorMixin
SpecificProcessorType = TypeVar("SpecificProcessorType", bound="ProcessorMixin")
# Dynamically import the Transformers module to grab the attribute classes of the processor from their names.
transformers_module = direct_transformers_import(Path(__file__).parent)
class _LazyAutoProcessorMapping(dict):
"""
Lazy dictionary to avoid circular imports.
The mapping names are only imported when accessed.
"""
_MAPPING_NAMES = {
"image_processor": ("transformers.models.auto.image_processing_auto", "AutoImageProcessor"),
"video_processor": ("transformers.models.auto.video_processing_auto", "AutoVideoProcessor"),
"feature_extractor": ("transformers.models.auto.feature_extraction_auto", "AutoFeatureExtractor"),
"audio_processor": ("transformers.models.auto.feature_extraction_auto", "AutoFeatureExtractor"),
"tokenizer": ("transformers.models.auto.tokenization_auto", "AutoTokenizer"),
}
def __getitem__(self, key):
if key not in self._MAPPING_NAMES:
raise KeyError(key)
module_name, attr_name = self._MAPPING_NAMES[key]
module = __import__(module_name, fromlist=[attr_name])
return getattr(module, attr_name)
def __contains__(self, key):
return key in self._MAPPING_NAMES
def keys(self):
return self._MAPPING_NAMES.keys()
MODALITY_TO_AUTOPROCESSOR_MAPPING = _LazyAutoProcessorMapping()
MODALITY_TO_BASE_CLASS_MAPPING = {
"audio_tokenizer": (
"HiggsAudioV2TokenizerModel",
"DacModel",
), # TODO: @eustlb, to be replaced with PreTrainedAudioTokenizerBase
"audio_processor": "FeatureExtractionMixin",
"tokenizer": ("PreTrainedTokenizerBase", "MistralCommonBackend"),
"feature_extractor": "FeatureExtractionMixin",
"image_processor": "ImageProcessingMixin",
"video_processor": "BaseVideoProcessor",
}
def _get_modality_for_attribute(attribute_name: str) -> str:
"""
Get the canonical modality type for a given attribute name.
For example:
- "image_processor" -> "image_processor"
- "encoder_image_processor" -> "image_processor"
- "text_tokenizer" -> "tokenizer"
- "my_feature_extractor" -> "feature_extractor"
"""
for modality in MODALITY_TO_AUTOPROCESSOR_MAPPING.keys():
if modality in attribute_name:
return modality
raise ValueError(
f"Cannot determine modality for attribute '{attribute_name}'. "
f"Attribute name must contain one of: {list(MODALITY_TO_AUTOPROCESSOR_MAPPING.keys())}"
)
if sys.version_info >= (3, 11):
Unpack = typing.Unpack
else:
Unpack = typing_extensions.Unpack
class TextKwargs(TypedDict, total=False):
"""
Keyword arguments for text processing. For extended documentation, check out tokenization_utils_base methods and
docstrings associated.
Attributes:
add_special_tokens (`bool`, *optional*)
Whether or not to add special tokens when encoding the sequences.
padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*)
Activates and controls padding.
truncation (`bool`, `str` or [`~tokenization_utils_base.TruncationStrategy`], *optional*):
Activates and controls truncation.
max_length (`int`, *optional*):
Controls the maximum length to use by one of the truncation/padding parameters.
stride (`int`, *optional*):
If set, the overflowing tokens will contain some tokens from the end of the truncated sequence.
is_split_into_words (`bool`, *optional*):
Whether or not the input is already pre-tokenized.
pad_to_multiple_of (`int`, *optional*):
If set, will pad the sequence to a multiple of the provided value.
return_token_type_ids (`bool`, *optional*):
Whether to return token type IDs.
return_attention_mask (`bool`, *optional*):
Whether to return the attention mask.
return_overflowing_tokens (`bool`, *optional*):
Whether or not to return overflowing token sequences.
return_special_tokens_mask (`bool`, *optional*):
Whether or not to return special tokens mask information.
return_offsets_mapping (`bool`, *optional*):
Whether or not to return `(char_start, char_end)` for each token.
return_length (`bool`, *optional*):
Whether or not to return the lengths of the encoded inputs.
verbose (`bool`, *optional*):
Whether or not to print more information and warnings.
padding_side (`str`, *optional*):
The side on which padding will be applied.
return_mm_token_type_ids (`bool`, *optional*):
Whether to return multimodal token type ids indicating mm placeholder token positions.
return_tensors (`str` or [`~utils.TensorType`], *optional*):
If set, will return tensors of a particular framework. Acceptable values are:
- `'pt'`: Return PyTorch `torch.Tensor` objects.
- `'np'`: Return NumPy `np.ndarray` objects.
"""
text_pair: TextInput | PreTokenizedInput | list[TextInput] | list[PreTokenizedInput] | None
text_target: TextInput | PreTokenizedInput | list[TextInput] | list[PreTokenizedInput] | None
text_pair_target: TextInput | PreTokenizedInput | list[TextInput] | list[PreTokenizedInput] | None
add_special_tokens: bool | None
padding: Annotated[bool | str | PaddingStrategy | None, padding_validator()]
truncation: Annotated[bool | str | TruncationStrategy | None, truncation_validator()]
max_length: Annotated[int | None, positive_int()]
stride: Annotated[int | None, positive_int()]
is_split_into_words: bool | None
pad_to_multiple_of: Annotated[int | None, positive_int()]
return_token_type_ids: bool | None
return_attention_mask: bool | None
return_overflowing_tokens: bool | None
return_special_tokens_mask: bool | None
return_offsets_mapping: bool | None
return_length: bool | None
verbose: bool | None
padding_side: Literal["left", "right"] | None
return_mm_token_type_ids: bool | None
return_tensors: Annotated[str | TensorType | None, tensor_type_validator()]
class ImagesKwargs(TypedDict, total=False):
"""
Keyword arguments for image processing. For extended documentation, check the appropriate ImageProcessor
class methods and docstrings.
Attributes:
do_convert_rgb (`bool`):
Whether to convert the image to RGB format.
do_resize (`bool`, *optional*):
Whether to resize the image.
size (`dict[str, int]`, *optional*):
Resize the shorter side of the input to `size["shortest_edge"]`.
default_to_square (`bool`, *optional*, defaults to `self.default_to_square`):
Whether to default to a square when resizing, if size is an int.
crop_size (`dict[str, int]`, *optional*):
Desired output size when applying center-cropping.
resample (`PILImageResampling`, *optional*):
Resampling filter to use if resizing the image.
do_rescale (`bool`, *optional*):
Whether to rescale the image by the specified scale `rescale_factor`.
rescale_factor (`int` or `float`, *optional*):
Scale factor to use if rescaling the image.
do_normalize (`bool`, *optional*):
Whether to normalize the image.
image_mean (`float` or `list[float] or tuple[float, float, float]`, *optional*):
Mean to use if normalizing the image.
image_std (`float` or `list[float] or tuple[float, float, float]`, *optional*):
Standard deviation to use if normalizing the image.
do_pad (`bool`, *optional*):
Whether to pad the images in the batch.
pad_size (`dict[str, int]`, *optional*):
The size `{"height": int, "width" int}` to pad the images to.
do_center_crop (`bool`, *optional*):
Whether to center crop the image.
data_format (`ChannelDimension` or `str`, *optional*):
The channel dimension format for the output image.
input_data_format (`ChannelDimension` or `str`, *optional*):
The channel dimension format for the input image.
device (`Union[str, torch.Tensor]`, *optional*):
The device to use for processing (e.g. "cpu", "cuda"), only relevant for torchvision backend.
return_tensors (`str` or [`~utils.TensorType`], *optional*):
If set, will return tensors of a particular framework. Acceptable values are:
- `'pt'`: Return PyTorch `torch.Tensor` objects.
- `'np'`: Return NumPy `np.ndarray` objects.
disable_grouping (`bool`, *optional*):
Whether to group images by shapes when processing or not, only relevant for torchvision backend.
image_seq_length (`int`, *optional*):
The number of image tokens to be used for each image in the input.
Added for backward compatibility but this should be set as a processor attribute in future models.
"""
do_convert_rgb: bool | None
do_resize: bool | None
size: Annotated[int | list[int] | tuple[int, ...] | dict[str, int] | None, image_size_validator()]
default_to_square: bool | None
crop_size: Annotated[int | list[int] | tuple[int, ...] | dict[str, int] | None, image_size_validator()]
resample: Annotated[Union["PILImageResampling", int] | None, resampling_validator()]
do_rescale: bool | None
rescale_factor: float | None
do_normalize: bool | None
image_mean: float | list[float] | tuple[float, ...] | None
image_std: float | list[float] | tuple[float, ...] | None
do_pad: bool | None
pad_size: Annotated[int | list[int] | tuple[int, ...] | dict[str, int] | None, image_size_validator()]
do_center_crop: bool | None
data_format: str | ChannelDimension | None
input_data_format: str | ChannelDimension | None
device: Annotated[Union[str, "torch.device"] | None, device_validator()]
return_tensors: Annotated[str | TensorType | None, tensor_type_validator()]
disable_grouping: bool | None
image_seq_length: int | None
class VideosKwargs(TypedDict, total=False):
"""
Keyword arguments for video processing.
Attributes:
do_convert_rgb (`bool`):
Whether to convert the video to RGB format.
do_resize (`bool`):
Whether to resize the video.
size (`dict[str, int]`, *optional*):
Resize the shorter side of the input to `size["shortest_edge"]`.
default_to_square (`bool`, *optional*, defaults to `self.default_to_square`):
Whether to default to a square when resizing, if size is an int.
resample (`PILImageResampling`, *optional*):
Resampling filter to use if resizing the video.
do_rescale (`bool`, *optional*):
Whether to rescale the video by the specified scale `rescale_factor`.
rescale_factor (`int` or `float`, *optional*):
Scale factor to use if rescaling the video.
do_normalize (`bool`, *optional*):
Whether to normalize the video.
image_mean (`float` or `list[float] or tuple[float, float, float]`, *optional*):
Mean to use if normalizing the video.
image_std (`float` or `list[float] or tuple[float, float, float]`, *optional*):
Standard deviation to use if normalizing the video.
do_center_crop (`bool`, *optional*):
Whether to center crop the video.
do_pad (`bool`, *optional*):
Whether to pad the images in the batch.
do_sample_frames (`bool`, *optional*):
Whether to sample frames from the video before processing or to process the whole video.
video_metadata (`Union[VideoMetadata, dict]`, *optional*):
Metadata of the video containing information about total duration, fps and total number of frames.
num_frames (`int`, *optional*):
Maximum number of frames to sample when `do_sample_frames=True`.
fps (`int` or `float`, *optional*):
Target frames to sample per second when `do_sample_frames=True`.
crop_size (`dict[str, int]`, *optional*):
Desired output size when applying center-cropping.
data_format (`ChannelDimension` or `str`, *optional*):
The channel dimension format for the output video.
input_data_format (`ChannelDimension` or `str`, *optional*):
The channel dimension format for the input video.
device (`Union[str, torch.Tensor]`, *optional*):
The device to use for processing (e.g. "cpu", "cuda"), only relevant for fast image processing.
return_metadata (`bool`, *optional*):
Whether to return video metadata or not.
return_tensors (`str` or [`~utils.TensorType`], *optional*):
If set, will return tensors of a particular framework. Acceptable values are:
- `'pt'`: Return PyTorch `torch.Tensor` objects.
- `'np'`: Return NumPy `np.ndarray` objects.
"""
do_convert_rgb: bool | None
do_resize: bool | None
size: Annotated[int | list[int] | tuple[int, ...] | dict[str, int] | None, image_size_validator()]
default_to_square: bool | None
resample: Annotated[Union["PILImageResampling", int] | None, resampling_validator()]
do_rescale: bool | None
rescale_factor: float | None
do_normalize: bool | None
image_mean: float | list[float] | tuple[float, ...] | None
image_std: float | list[float] | tuple[float, ...] | None
do_center_crop: bool | None
do_pad: bool | None
crop_size: Annotated[int | list[int] | tuple[int, ...] | dict[str, int] | None, image_size_validator()]
data_format: str | ChannelDimension | None
input_data_format: str | ChannelDimension | None
device: Annotated[Union[str, "torch.device"] | None, device_validator()]
do_sample_frames: bool | None
video_metadata: Annotated[VideoMetadataType | None, video_metadata_validator()]
fps: Annotated[int | float | None, positive_any_number()]
num_frames: Annotated[int | None, positive_int()]
return_metadata: bool | None
return_tensors: Annotated[str | TensorType | None, tensor_type_validator()]
class AudioKwargs(TypedDict, total=False):
"""
Keyword arguments for audio processing.
Attributes:
sampling_rate (`int`, *optional*):
The sampling rate at which the `raw_speech` input was sampled.
raw_speech (`np.ndarray`, `list[float]`, `list[np.ndarray]`, `list[list[float]]`):
The sequence or batch of sequences to be padded. Each sequence can be a numpy array, a list of float
values, a list of numpy arrays or a list of list of float values. Must be mono channel audio, not
stereo, i.e. single float per timestep.
padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*):
Select a strategy to pad the returned sequences (according to the model's padding side and padding
index) among:
- `True` or `'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
sequence if provided).
- `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum
acceptable input length for the model if that argument is not provided.
- `False` or `'do_not_pad'`
max_length (`int`, *optional*):
Maximum length of the returned list and optionally padding length (see above).
truncation (`bool`, *optional*):
Activates truncation to cut input sequences longer than *max_length* to *max_length*.
pad_to_multiple_of (`int`, *optional*):
If set, will pad the sequence to a multiple of the provided value.
return_attention_mask (`bool`, *optional*):
Whether or not [`~ASTFeatureExtractor.__call__`] should return `attention_mask`.
return_tensors (`str` or [`~utils.TensorType`], *optional*):
If set, will return tensors of a particular framework. Acceptable values are:
- `'pt'`: Return PyTorch `torch.Tensor` objects.
- `'np'`: Return NumPy `np.ndarray` objects.
"""
sampling_rate: Annotated[int | None, positive_int()]
raw_speech: Union["np.ndarray", list[float], list["np.ndarray"], list[list[float]]] | None
padding: Annotated[bool | str | PaddingStrategy | None, padding_validator()]
max_length: Annotated[int | None, positive_int()]
truncation: Annotated[bool | str | TruncationStrategy | None, truncation_validator()]
pad_to_multiple_of: Annotated[int | None, positive_int()]
return_attention_mask: bool | None
return_tensors: Annotated[str | TensorType | None, tensor_type_validator()]
class ProcessingKwargs(TypedDict, total=False):
"""
Base class for kwargs passing to processors.
In case a model has specific kwargs that are not present in the base class or default values for existing keys,
it should have its own `ModelProcessorKwargs` class that inherits from `ProcessingKwargs` to provide:
1) Additional typed keys and that this model requires to process inputs.
2) Default values for existing keys under a `_defaults` attribute.
New keys have to be defined as follows to ensure type hinting is done correctly.
```python
# adding a new image kwarg for this model
class ModelImagesKwargs(ImagesKwargs, total=False):
new_image_kwarg: Optional[bool]
class ModelProcessorKwargs(ProcessingKwargs, total=False):
images_kwargs: ModelImagesKwargs
_defaults = {
"images_kwargs: {
"new_image_kwarg": False,
}
"text_kwargs": {
"padding": "max_length",
},
}
```
For Python 3.8 compatibility, when inheriting from this class and overriding one of the kwargs,
you need to manually update the __annotations__ dictionary. This can be done as follows:
```python
class CustomProcessorKwargs(ProcessingKwargs, total=False):
images_kwargs: CustomImagesKwargs
CustomProcessorKwargs.__annotations__["images_kwargs"] = CustomImagesKwargs # python 3.8 compatibility
```
"""
_defaults = {}
text_kwargs: TextKwargs = {
**TextKwargs.__annotations__,
}
images_kwargs: ImagesKwargs = {
**ImagesKwargs.__annotations__,
}
videos_kwargs: VideosKwargs = {
**VideosKwargs.__annotations__,
}
audio_kwargs: AudioKwargs = {
**AudioKwargs.__annotations__,
}
class TokenizerChatTemplateKwargs(TypedDict, total=False):
"""
NOTE: `TokenizerChatTemplateKwargs` is deprecated and will be removed in future versions
Keyword arguments for tokenizer's `apply_chat_template`, when it is called from within a processor.
tools (`list[Dict]`, *optional*):
A list of tools (callable functions) that will be accessible to the model. If the template does not
support function calling, this argument will have no effect. Each tool should be passed as a JSON Schema,
giving the name, description and argument types for the tool. See our
[chat templating guide](https://huggingface.co/docs/transformers/main/en/chat_templating#automated-function-conversion-for-tool-use)
for more information.
documents (`list[dict[str, str]]`, *optional*):
A list of dicts representing documents that will be accessible to the model if it is performing RAG
(retrieval-augmented generation). If the template does not support RAG, this argument will have no
effect. We recommend that each document should be a dict containing "title" and "text" keys. Please
see the RAG section of the [chat templating guide](https://huggingface.co/docs/transformers/main/en/chat_templating#arguments-for-RAG)
for examples of passing documents with chat templates.
add_generation_prompt (bool, *optional*):
If this is set, a prompt with the token(s) that indicate
the start of an assistant message will be appended to the formatted output. This is useful when you want to generate a response from the model.
Note that this argument will be passed to the chat template, and so it must be supported in the
template for this argument to have any effect.
continue_final_message (bool, *optional*):
If this is set, the chat will be formatted so that the final
message in the chat is open-ended, without any EOS tokens. The model will continue this message
rather than starting a new one. This allows you to "prefill" part of
the model's response for it. Cannot be used at the same time as `add_generation_prompt`.
return_assistant_tokens_mask (`bool`, defaults to `False`):
Whether to return a mask of the assistant generated tokens. For tokens generated by the assistant,
the mask will contain 1. For user and system tokens, the mask will contain 0.
This functionality is only available for chat templates that support it via the `{% generation %}` keyword.
reasoning_effort (`str`, *optional*):
The reasoning effort level to use for the model's response. Supported values depend on the model
(e.g. `"none"`, "low"`, `"medium"`, `"high"`). If the template does not support reasoning effort,
this argument will have no effect.
"""
tools: list[dict] | None = None
documents: list[dict[str, str]] | None = None
add_generation_prompt: bool | None = False
continue_final_message: bool | None = False
return_assistant_tokens_mask: bool | None = False
reasoning_effort: str | None = None
class ProcessorChatTemplateKwargs(TokenizerChatTemplateKwargs, total=False):
"""
NOTE: `ProcessorChatTemplateKwargs` is deprecated and will be removed in future versions
Keyword arguments for processor's `apply_chat_template`.
tokenize (`bool`, *optional*, defaults to `False`):
Whether to tokenize the output or not.
return_dict (`bool`, defaults to `False`):
Whether to return a dictionary with named outputs. Has no effect if tokenize is `False`.
load_audio_from_video (`bool`, *optional*, defaults to `False`):
Whether to use the audio track of input video. If `True` the audio track will be loaded and passed to the
processor. This flag has no effect if the model doesn't support audio modality.
"""
tokenize: bool | None = False
return_dict: bool | None = False
load_audio_from_video: bool | None = False
class AllKwargsForChatTemplate(TypedDict, total=False):
"NOTE: `AllKwargsForChatTemplate` is deprecated and will be removed in future versions"
processor_kwargs: ProcessingKwargs
template_kwargs: ProcessorChatTemplateKwargs
@dataclass
class MultiModalData:
"""
Dataclass that holds extra useful data for processing
multimodal data. Processors currently cannot return keys,
unless it is used in model's forward. Thus we have helper
methods that calculate and return useful data from processing
input multimodals (images/videos).
Note that this dataclass is aimed to be used only in vLLM
and we might change its API in the future.
"""
num_image_tokens: list[int] | None = None
num_video_tokens: list[int] | None = None
num_audio_tokens: list[int] | None = None
num_image_patches: list[int] | None = None
def __contains__(self, key):
return hasattr(self, key) and getattr(self, key) is not None
def __getitem__(self, key):
if hasattr(self, key):
return getattr(self, key)
raise AttributeError(f"{self.__class__.__name__} has no attribute {key}")
class ProcessorMixin(PushToHubMixin):
"""
This is a mixin used to provide saving/loading functionality for all processor classes.
"""
# Dynamically set sub-processor attributes. Not every processor has all of these;
# they are populated via setattr in __init__ based on each subclass's `attributes`.
tokenizer: Any
feature_extractor: Any
image_processor: Any
video_processor: Any
chat_template: str | dict[str, str] | None
# Names need to be attr_class for attr in attributes
_auto_class = None
valid_processor_kwargs = ProcessingKwargs
# args have to match the attributes class attribute
def __init__(self, *args, **kwargs):
# First, extract chat template from kwargs. It can never be a positional arg
setattr(self, "chat_template", kwargs.pop("chat_template", None))
self.image_ids = [getattr(self, "image_token_id", None)]
self.video_ids = [getattr(self, "video_token_id", None)]
self.audio_ids = [getattr(self, "audio_token_id", None)]
# Check audio tokenizer for its class but do not treat it as attr to avoid saving weights
if (audio_tokenizer := kwargs.pop("audio_tokenizer", None)) is not None:
proper_class = self.check_argument_for_proper_class("audio_tokenizer", audio_tokenizer)
if not (is_torch_available() and isinstance(audio_tokenizer, PreTrainedAudioTokenizerBase)):
raise ValueError(
f"Tried to use `{proper_class}` for audio tokenization. However, this class is not"
" registered for audio tokenization."
)
setattr(self, "audio_tokenizer", audio_tokenizer)
# Sanitize args and kwargs
for key in kwargs:
if key not in self.get_attributes():
raise TypeError(f"Unexpected keyword argument {key}.")
for arg, attribute_name in zip(args, self.get_attributes()):
if attribute_name in kwargs:
raise TypeError(f"Got multiple values for argument {attribute_name}.")
else:
kwargs[attribute_name] = arg
if len(kwargs) != len(self.get_attributes()):
raise ValueError(
f"This processor requires {len(self.get_attributes())} arguments: {', '.join(self.get_attributes())}. Got "
f"{len(args)} arguments instead."
)
# Check each arg is of the proper class (this will also catch a user initializing in the wrong order)
for attribute_name, arg in kwargs.items():
self.check_argument_for_proper_class(attribute_name, arg)
setattr(self, attribute_name, arg)
def __call__(
self,
images: ImageInput | None = None,
text: TextInput | PreTokenizedInput | list[TextInput] | list[PreTokenizedInput] | None = None,
videos: VideoInput | None = None,
audio: AudioInput | None = None,
**kwargs: Unpack[ProcessingKwargs],
):
"""
Main method to prepare for model inputs. This method forwards the each modality argument to its own processor
along with `kwargs`. Please refer to the docstring of the each processor attributes for more information.
Args:
images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `list[PIL.Image.Image]`, `list[np.ndarray]`, `list[torch.Tensor]`):
The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
tensor. Both channels-first and channels-last formats are supported.
text (`TextInput`, `PreTokenizedInput`, `list[TextInput]`, `list[PreTokenizedInput]`, *optional*):
The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
(pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
`is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
videos (`np.ndarray`, `torch.Tensor`, `List[np.ndarray]`, `List[torch.Tensor]`):
The video or batch of videos to be prepared. Each video can be a 4D NumPy array or PyTorch
tensor, or a nested list of 3D frames. Both channels-first and channels-last formats are supported.
audio (`np.ndarray`, `torch.Tensor`, `list[np.ndarray]`, `list[torch.Tensor]`):
The audio or batch of audio to be prepared. Each audio can be a NumPy array or PyTorch
tensor.
return_tensors (`str` or [`~utils.TensorType`], *optional*):
If set, will return tensors of a particular framework. Acceptable values are:
- `'pt'`: Return PyTorch `torch.Tensor` objects.
- `'np'`: Return NumPy `np.ndarray` objects.
Returns:
[`BatchFeature`]: A [`BatchFeature`] object with processed inputs in a dict format.
"""
if "audios" in kwargs and audio is None:
raise ValueError("You passed keyword argument `audios` which is deprecated. Please use `audio` instead.")
if images is None and text is None and videos is None and audio is None:
raise ValueError(f"You need to provide at least one input to call {self.__class__.__name__}")
kwargs = self._merge_kwargs(
self.valid_processor_kwargs,
tokenizer_init_kwargs=self.tokenizer.init_kwargs if hasattr(self, "tokenizer") else {},
**kwargs,
)
attribute_to_kwargs = {
"tokenizer": (text, "text_kwargs"),
"image_processor": (images, "images_kwargs"),
"video_processor": (videos, "videos_kwargs"),
"feature_extractor": (audio, "audio_kwargs"),
}
outputs = {}
for attribute_name in self.get_attributes():
attribute = getattr(self, attribute_name, None)
input_data, input_kwargs = attribute_to_kwargs[attribute_name]
if input_data is not None and attribute is not None:
attribute_output = attribute(input_data, **kwargs[input_kwargs])
outputs.update(attribute_output)
return BatchFeature(outputs)
def check_argument_for_proper_class(self, argument_name, argument):
"""
Checks the passed argument's class against the expected transformers class. In case of an unexpected
mismatch between expected and actual class, an error is raise. Otherwise, the proper retrieved class
is returned.
"""
# If the exact attribute name is not in the mapping, use its canonical modality
# (e.g., "encoder_tokenizer" -> "tokenizer")
if argument_name not in MODALITY_TO_BASE_CLASS_MAPPING:
argument_name = _get_modality_for_attribute(argument_name)
class_name = MODALITY_TO_BASE_CLASS_MAPPING.get(argument_name)
if isinstance(class_name, tuple):
proper_class = tuple(self.get_possibly_dynamic_module(n) for n in class_name if n is not None)
else:
proper_class = self.get_possibly_dynamic_module(class_name)
if not isinstance(argument, proper_class):
raise TypeError(
f"Received a {type(argument).__name__} for argument {argument_name}, but a {class_name} was expected."
)
return proper_class
def to_dict(self) -> dict[str, Any]:
"""
Serializes this instance to a Python dictionary.
Returns:
`dict[str, Any]`: Dictionary of all the attributes that make up this processor instance.
"""
# Exclude tokenizer attributes before deepcopying to avoid copying large vocab/token structures.
tokenizer_attributes = set()
for attribute in self.__class__.get_attributes():
if attribute in self.__dict__:
modality = _get_modality_for_attribute(attribute)
if modality == "tokenizer":
tokenizer_attributes.add(attribute)
dict_to_copy = {k: v for k, v in self.__dict__.items() if k not in tokenizer_attributes}
output = copy.deepcopy(dict_to_copy)
# Get the kwargs in `__init__`.
sig = inspect.signature(self.__init__)
# Only save the attributes that are presented in the kwargs of `__init__`.
# or in the attributes
attrs_to_save = list(sig.parameters) + self.__class__.get_attributes()
# extra attributes to be kept
attrs_to_save += ["auto_map"]
if "chat_template" in output:
del output["chat_template"]
def cast_array_to_list(dictionary):
"""
Numpy arrays are not serialiazable but can be in pre-processing dicts.
This function casts arrays to list, recusring through the nested configs as well.
"""
for key, value in dictionary.items():
if isinstance(value, np.ndarray):
dictionary[key] = value.tolist()
elif isinstance(value, dict):
dictionary[key] = cast_array_to_list(value)
return dictionary
# Special case, add `audio_tokenizer` dict which points to model weights and path
if "audio_tokenizer" in output:
audio_tokenizer_dict = {
"audio_tokenizer_class": self.audio_tokenizer.__class__.__name__,
"audio_tokenizer_name_or_path": self.audio_tokenizer.name_or_path,
}
output["audio_tokenizer"] = audio_tokenizer_dict
# Serialize attributes as a dict
output = {
k: v.to_dict() if isinstance(v, PushToHubMixin) else v
for k, v in output.items()
if (
k in attrs_to_save # keep all attributes that have to be serialized
and v.__class__.__name__ != "BeamSearchDecoderCTC" # remove attributes with that are objects
)
}
output = cast_array_to_list(output)
output["processor_class"] = self.__class__.__name__
return output
def to_json_string(self) -> str:
"""
Serializes this instance to a JSON string.
Returns:
`str`: String containing all the attributes that make up this feature_extractor instance in JSON format.
"""
dictionary = self.to_dict()
return json.dumps(dictionary, indent=2, sort_keys=True) + "\n"
def to_json_file(self, json_file_path: str | os.PathLike):
"""
Save this instance to a JSON file.
Args:
json_file_path (`str` or `os.PathLike`):
Path to the JSON file in which this processor instance's parameters will be saved.
"""
with open(json_file_path, "w", encoding="utf-8") as writer:
writer.write(self.to_json_string())
def __repr__(self):
attributes_repr = [f"- {name}: {repr(getattr(self, name))}" for name in self.get_attributes()]
attributes_repr = "\n".join(attributes_repr)
return f"{self.__class__.__name__}:\n{attributes_repr}\n\n{self.to_json_string()}"
def save_pretrained(self, save_directory, push_to_hub: bool = False, **kwargs):
"""
Saves the attributes of this processor (feature extractor, tokenizer...) in the specified directory so that it
can be reloaded using the [`~ProcessorMixin.from_pretrained`] method.
<Tip>
This class method is simply calling [`~feature_extraction_utils.FeatureExtractionMixin.save_pretrained`] and
[`~tokenization_utils_base.PreTrainedTokenizerBase.save_pretrained`]. Please refer to the docstrings of the
methods above for more information.
</Tip>
Args:
save_directory (`str` or `os.PathLike`):
Directory where the feature extractor JSON file and the tokenizer files will be saved (directory will
be created if it does not exist).
push_to_hub (`bool`, *optional*, defaults to `False`):
Whether or not to push your model to the Hugging Face model hub after saving it. You can specify the
repository you want to push to with `repo_id` (will default to the name of `save_directory` in your
namespace).
kwargs (`dict[str, Any]`, *optional*):
Additional key word arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
"""
os.makedirs(save_directory, exist_ok=True)
if push_to_hub:
commit_message = kwargs.pop("commit_message", None)
repo_id = kwargs.pop("repo_id", save_directory.split(os.path.sep)[-1])
repo_id = create_repo(repo_id, exist_ok=True, **kwargs).repo_id
files_timestamps = self._get_files_timestamps(save_directory)
# If we have a custom config, we copy the file defining it in the folder and set the attributes so it can be
# loaded from the Hub.
if self._auto_class is not None:
attrs = [getattr(self, attribute_name) for attribute_name in self.get_attributes()]
configs = [(a.init_kwargs if isinstance(a, PreTrainedTokenizerBase) else a) for a in attrs]
configs.append(self)
custom_object_save(self, save_directory, config=configs)
for attribute_name in self.get_attributes():
attribute = getattr(self, attribute_name)
modality = _get_modality_for_attribute(attribute_name)
is_primary = attribute_name == modality
if modality == "tokenizer":
attribute._set_processor_class(self.__class__.__name__)
# Save the tokenizer in its own vocab file. The other attributes are saved as part of `processor_config.json`
if is_primary:
attribute.save_pretrained(save_directory)
else:
# if a model has multiple tokenizers, save the additional tokenizers in their own folders.
attribute.save_pretrained(os.path.join(save_directory, attribute_name))
elif attribute._auto_class is not None:
custom_object_save(attribute, save_directory, config=attribute)
if self._auto_class is not None:
# We added an attribute to the init_kwargs of the tokenizers, which needs to be cleaned up.
for attribute_name in self.get_attributes():
attribute = getattr(self, attribute_name)
if isinstance(attribute, PreTrainedTokenizerBase):
del attribute.init_kwargs["auto_map"]
# If we save using the predefined names, we can load using `from_pretrained`
# plus we save chat_template in its own file
output_processor_file = os.path.join(save_directory, PROCESSOR_NAME)
output_chat_template_file_jinja = os.path.join(save_directory, CHAT_TEMPLATE_FILE)
chat_template_dir = os.path.join(save_directory, CHAT_TEMPLATE_DIR)
# Save `chat_template` in its own file. We can't get it from `processor_dict` as we popped it in `to_dict`
# to avoid serializing chat template in json config file. So let's get it from `self` directly
if isinstance(self.chat_template, str):
# New format for single templates is to save them as chat_template.jinja
with open(output_chat_template_file_jinja, "w", encoding="utf-8") as f:
f.write(self.chat_template)
logger.info(f"chat template saved in {output_chat_template_file_jinja}")
elif isinstance(self.chat_template, dict):
# New format for multiple templates is to save the default as chat_template.jinja
# and the other templates in the chat_templates/ directory
for template_name, template in self.chat_template.items():
if template_name == "default":
with open(output_chat_template_file_jinja, "w", encoding="utf-8") as f:
f.write(self.chat_template["default"])
logger.info(f"chat template saved in {output_chat_template_file_jinja}")
else:
os.makedirs(chat_template_dir, exist_ok=True)
template_filepath = os.path.join(chat_template_dir, f"{template_name}.jinja")
with open(template_filepath, "w", encoding="utf-8") as f:
f.write(template)
logger.info(f"chat template saved in {template_filepath}")
# Create a unified `preprocessor_config.json` and save all attributes as a composite config, except for tokenizers
self.to_json_file(output_processor_file)
logger.info(f"processor saved in {output_processor_file}")
return_files = [output_processor_file]
if push_to_hub:
self._upload_modified_files(
save_directory,
repo_id,
files_timestamps,
commit_message=commit_message,
token=kwargs.get("token"),
)
return return_files
@classmethod
def get_processor_dict(
cls, pretrained_model_name_or_path: str | os.PathLike, **kwargs
) -> tuple[dict[str, Any], dict[str, Any]]:
"""
From a `pretrained_model_name_or_path`, resolve to a dictionary of parameters, to be used for instantiating a
processor of type [`~processing_utils.ProcessingMixin`] using `from_args_and_dict`.
Parameters:
pretrained_model_name_or_path (`str` or `os.PathLike`):
The identifier of the pre-trained checkpoint from which we want the dictionary of parameters.
subfolder (`str`, *optional*, defaults to `""`):
In case the relevant files are located inside a subfolder of the model repo on huggingface.co, you can
specify the folder name here.
Returns:
`tuple[Dict, Dict]`: The dictionary(ies) that will be used to instantiate the processor object.
"""
# holding a copy for optionally loading the audio tokenizer (if available)
audio_tokenizer_kwargs = copy.deepcopy(kwargs)
cache_dir = kwargs.pop("cache_dir", None)
force_download = kwargs.pop("force_download", False)
proxies = kwargs.pop("proxies", None)
token = kwargs.pop("token", None)
local_files_only = kwargs.pop("local_files_only", False)
revision = kwargs.pop("revision", None)
subfolder = kwargs.pop("subfolder", "")
from_pipeline = kwargs.pop("_from_pipeline", None)
from_auto_class = kwargs.pop("_from_auto", False)
user_agent = {"file_type": "processor", "from_auto_class": from_auto_class}
if from_pipeline is not None:
user_agent["using_pipeline"] = from_pipeline
if is_offline_mode() and not local_files_only:
logger.info("Offline mode: forcing local_files_only=True")
local_files_only = True
pretrained_model_name_or_path = str(pretrained_model_name_or_path)
is_local = os.path.isdir(pretrained_model_name_or_path)
if os.path.isdir(pretrained_model_name_or_path):
processor_file = os.path.join(pretrained_model_name_or_path, PROCESSOR_NAME)
additional_chat_template_files = {}
resolved_additional_chat_template_files = {}
if os.path.isfile(pretrained_model_name_or_path):
resolved_processor_file = pretrained_model_name_or_path
# can't load chat-template and audio tokenizer when given a file as pretrained_model_name_or_path
resolved_chat_template_file = None
resolved_raw_chat_template_file = None
resolved_audio_tokenizer_file = None
is_local = True
else:
if is_local:
template_dir = Path(pretrained_model_name_or_path, CHAT_TEMPLATE_DIR)
if template_dir.is_dir():
for template_file in template_dir.glob("*.jinja"):
template_name = template_file.stem
additional_chat_template_files[template_name] = f"{CHAT_TEMPLATE_DIR}/{template_file.name}"
else:
try:
for template in list_repo_templates(
pretrained_model_name_or_path,
local_files_only=local_files_only,
revision=revision,
cache_dir=cache_dir,
token=token,
):
template = template.removesuffix(".jinja")
additional_chat_template_files[template] = f"{CHAT_TEMPLATE_DIR}/{template}.jinja"
except EntryNotFoundError:
pass # No template dir means no template files
processor_file = PROCESSOR_NAME
try:
# Load from local folder or from cache or download from model Hub and cache
resolved_processor_file = cached_file(
pretrained_model_name_or_path,
processor_file,
cache_dir=cache_dir,
force_download=force_download,
proxies=proxies,
local_files_only=local_files_only,
token=token,
user_agent=user_agent,
revision=revision,
subfolder=subfolder,
_raise_exceptions_for_missing_entries=False,
)
# chat_template.json is a legacy file used by the processor class
# a raw chat_template.jinja is preferred in future
resolved_chat_template_file = cached_file(
pretrained_model_name_or_path,
LEGACY_PROCESSOR_CHAT_TEMPLATE_FILE,
cache_dir=cache_dir,
force_download=force_download,
proxies=proxies,
local_files_only=local_files_only,
token=token,
user_agent=user_agent,
revision=revision,
subfolder=subfolder,
_raise_exceptions_for_missing_entries=False,
)
resolved_raw_chat_template_file = cached_file(
pretrained_model_name_or_path,
CHAT_TEMPLATE_FILE,
cache_dir=cache_dir,
force_download=force_download,
proxies=proxies,
local_files_only=local_files_only,
token=token,
user_agent=user_agent,
revision=revision,
subfolder=subfolder,
_raise_exceptions_for_missing_entries=False,
)
resolved_additional_chat_template_files = {
template_name: cached_file(
pretrained_model_name_or_path,
template_file,
cache_dir=cache_dir,
force_download=force_download,
proxies=proxies,
local_files_only=local_files_only,
token=token,
user_agent=user_agent,
revision=revision,
subfolder=subfolder,
_raise_exceptions_for_missing_entries=False,
)
for template_name, template_file in additional_chat_template_files.items()
}
resolved_audio_tokenizer_file = cached_file(
pretrained_model_name_or_path,
AUDIO_TOKENIZER_NAME,
cache_dir=cache_dir,
force_download=force_download,
proxies=proxies,
local_files_only=local_files_only,
token=token,
user_agent=user_agent,
revision=revision,
subfolder=subfolder,
_raise_exceptions_for_missing_entries=False,
)
except OSError:
# Raise any environment error raise by `cached_file`. It will have a helpful error message adapted to
# the original exception.
raise
except Exception:
# For any other exception, we throw a generic error.
raise OSError(
f"Can't load processor for '{pretrained_model_name_or_path}'. If you were trying to load"
" it from 'https://huggingface.co/models', make sure you don't have a local directory with the"
f" same name. Otherwise, make sure '{pretrained_model_name_or_path}' is the correct path to a"
f" directory containing a {PROCESSOR_NAME} file"
)
# Add chat template as kwarg before returning because most models don't have processor config
if resolved_chat_template_file is not None:
# This is the legacy path
with open(resolved_chat_template_file, encoding="utf-8") as reader:
chat_template_json = json.loads(reader.read())
chat_templates = {"default": chat_template_json["chat_template"]}
if resolved_additional_chat_template_files:
raise ValueError(
"Cannot load chat template due to conflicting files - this checkpoint combines "
"a legacy chat_template.json file with separate template files, which is not "
"supported. To resolve this error, replace the legacy chat_template.json file "
"with a modern chat_template.jinja file."
)
else:
chat_templates = {
template_name: open(template_file, "r", encoding="utf-8").read()
for template_name, template_file in resolved_additional_chat_template_files.items()
}
if resolved_raw_chat_template_file is not None:
with open(resolved_raw_chat_template_file, "r", encoding="utf-8") as reader:
chat_templates["default"] = reader.read()
if isinstance(chat_templates, dict) and "default" in chat_templates and len(chat_templates) == 1:
chat_templates = chat_templates["default"] # Flatten when we just have a single template/file
# Existing processors on the Hub created before #27761 being merged don't have `processor_config.json` (if not
# updated afterward), and we need to keep `from_pretrained` work. So here it fallbacks to the empty dict.
# (`cached_file` called using `_raise_exceptions_for_missing_entries=False` to avoid exception)
# However, for models added in the future, we won't get the expected error if this file is missing.
if resolved_processor_file is None:
# In any case we need to pass `chat_template` if it is available
processor_dict = {}
else:
try:
# Load processor dict
with open(resolved_processor_file, encoding="utf-8") as reader:
text = reader.read()
processor_dict = json.loads(text)
except json.JSONDecodeError:
raise OSError(
f"It looks like the config file at '{resolved_processor_file}' is not a valid JSON file."
)
if is_local:
logger.info(f"loading configuration file {resolved_processor_file}")
else:
logger.info(f"loading configuration file {processor_file} from cache at {resolved_processor_file}")
if processor_dict.get("chat_template") is not None:
logger.warning_once(
"Chat templates should be in a 'chat_template.jinja' file but found key='chat_template' "
"in the processor's config. Make sure to move your template to its own file."
)
elif chat_templates:
processor_dict["chat_template"] = chat_templates
# Audio tokenizer needs to load the model checkpoint first, because the saved
# json file contains only references to the model path and repo id
if resolved_audio_tokenizer_file is not None or "audio_tokenizer" in processor_dict:
if resolved_audio_tokenizer_file is not None:
reader = open(resolved_audio_tokenizer_file, "r", encoding="utf-8")
audio_tokenizer_dict = reader.read()
audio_tokenizer_dict = json.loads(audio_tokenizer_dict)
else:
audio_tokenizer_dict = processor_dict["audio_tokenizer"]
audio_tokenizer_class = cls.get_possibly_dynamic_module(audio_tokenizer_dict["audio_tokenizer_class"])
audio_tokenizer_path = audio_tokenizer_dict["audio_tokenizer_name_or_path"]
processor_dict["audio_tokenizer"] = audio_tokenizer_class.from_pretrained(
audio_tokenizer_path, **audio_tokenizer_kwargs
)
return processor_dict, kwargs
@classmethod
def from_args_and_dict(cls, args, processor_dict: dict[str, Any], **kwargs):
"""
Instantiates a type of [`~processing_utils.ProcessingMixin`] from a Python dictionary of parameters.
Args:
processor_dict (`dict[str, Any]`):
Dictionary that will be used to instantiate the processor object. Such a dictionary can be
retrieved from a pretrained checkpoint by leveraging the
[`~processing_utils.ProcessingMixin.to_dict`] method.
kwargs (`dict[str, Any]`):
Additional parameters from which to initialize the processor object.
Returns:
[`~processing_utils.ProcessingMixin`]: The processor object instantiated from those
parameters.
"""
processor_dict = processor_dict.copy()
return_unused_kwargs = kwargs.pop("return_unused_kwargs", False)
# We have to pop up some unused (but specific) kwargs and then validate that it doesn't contain unused kwargs
# If we don't pop, some specific kwargs will raise a warning or error
for unused_kwarg in cls.get_attributes() + ["auto_map", "processor_class"]:
processor_dict.pop(unused_kwarg, None)
# override processor_dict with given kwargs
processor_dict.update(kwargs)
# check if there is an overlap between args and processor_dict
accepted_args_and_kwargs = cls.__init__.__code__.co_varnames[: cls.__init__.__code__.co_argcount][1:]
# validate both processor_dict and given kwargs
unused_kwargs, valid_kwargs = cls.validate_init_kwargs(
processor_config=processor_dict, valid_kwargs=accepted_args_and_kwargs
)
# update args that are already in processor_dict to avoid duplicate arguments
args_to_update = {
i: valid_kwargs.pop(arg)
for i, arg in enumerate(accepted_args_and_kwargs)
if (arg in valid_kwargs and i < len(args))
}
args = [args_to_update.get(i, arg) for i, arg in enumerate(args)]
# instantiate processor with used (and valid) kwargs only
processor = cls(*args, **valid_kwargs)
logger.info(f"Processor {processor}")
if return_unused_kwargs:
return processor, unused_kwargs
else:
return processor
def _merge_kwargs(
self,
ModelProcessorKwargs: ProcessingKwargs,
tokenizer_init_kwargs: dict | None = None,
**kwargs,
) -> dict[str, dict]:
"""
Method to merge dictionaries of kwargs cleanly separated by modality within a Processor instance.
The order of operations is as follows:
1) kwargs passed as before have highest priority to preserve BC.
```python
high_priority_kwargs = {"crop_size" = {"height": 222, "width": 222}, "padding" = "max_length"}
processor(..., **high_priority_kwargs)
```
2) kwargs passed as modality-specific kwargs have second priority. This is the recommended API.
```python
processor(..., text_kwargs={"padding": "max_length"}, images_kwargs={"crop_size": {"height": 222, "width": 222}}})
```
3) kwargs passed during instantiation of a modality processor have fourth priority.
```python
tokenizer = tokenizer_class(..., {"padding": "max_length"})
image_processor = image_processor_class(...)
processor(tokenizer, image_processor) # will pass max_length unless overridden by kwargs at call
```
4) defaults kwargs specified at processor level have lowest priority.
```python
class MyProcessingKwargs(ProcessingKwargs, CommonKwargs, TextKwargs, ImagesKwargs, total=False):
_defaults = {
"text_kwargs": {
"padding": "max_length",
"max_length": 64,
},
}
```
Args:
ModelProcessorKwargs (`ProcessingKwargs`):
Typed dictionary of kwargs specifically required by the model passed.
tokenizer_init_kwargs (`Dict`, *optional*):
Dictionary of kwargs the tokenizer was instantiated with and need to take precedence over defaults.
Returns:
output_kwargs (`Dict`):
Dictionary of per-modality kwargs to be passed to each modality-specific processor.
"""
# holding a copy to avoid mutating user-provided arguments
# Use deepcopy to also copy nested dicts (like videos_kwargs) that will be modified via pop()
kwargs = copy.deepcopy(kwargs)
# Initialize dictionaries
output_kwargs = {
"text_kwargs": {},
"images_kwargs": {},
"audio_kwargs": {},
"videos_kwargs": {},
}
default_kwargs = {
"text_kwargs": {},
"images_kwargs": {},
"audio_kwargs": {},
"videos_kwargs": {},
}
map_preprocessor_kwargs = {
"text_kwargs": "tokenizer",
"images_kwargs": "image_processor",
"audio_kwargs": "feature_extractor",
"videos_kwargs": "video_processor",
}
possible_modality_keywords = {"text", "audio", "videos", "images"}
used_keys = set()
# get defaults from set model processor kwargs if they exist
for modality in default_kwargs:
default_kwargs[modality] = ModelProcessorKwargs._defaults.get(modality, {}).copy()
# Some preprocessors define a set of accepted "valid_kwargs" (currently only vision).
# In those cases, we don’t declare a `ModalityKwargs` attribute in the TypedDict.
# Instead, we dynamically obtain the kwargs from the preprocessor and merge them
# with the general kwargs set. This ensures consistency between preprocessor and
# processor classes, and helps prevent accidental mismatches.
modality_valid_kwargs = set(ModelProcessorKwargs.__annotations__[modality].__annotations__)
if modality in map_preprocessor_kwargs:
preprocessor = getattr(self, map_preprocessor_kwargs[modality], None)
preprocessor_valid_kwargs = (
getattr(preprocessor, "valid_kwargs", None) if preprocessor is not None else None
)
modality_valid_kwargs.update(
set(preprocessor_valid_kwargs.__annotations__ if preprocessor_valid_kwargs is not None else [])
)
# update defaults with arguments from tokenizer init
for modality_key in modality_valid_kwargs:
# init with tokenizer init kwargs if necessary
if tokenizer_init_kwargs is not None and modality_key in tokenizer_init_kwargs:
value = (
getattr(self.tokenizer, modality_key)
if hasattr(self.tokenizer, modality_key)
else tokenizer_init_kwargs[modality_key]
)
default_kwargs[modality][modality_key] = value
# now defaults kwargs are updated with the tokenizers defaults.
# pass defaults to output dictionary
output_kwargs.update(default_kwargs)
# For `common_kwargs` just update all modality-specific kwargs with same key/values
common_kwargs = ModelProcessorKwargs._defaults.get("common_kwargs", {})
common_kwargs.update(kwargs.get("common_kwargs", {}))
if common_kwargs:
for kwarg in output_kwargs.values():
kwarg.update(common_kwargs)
# update modality kwargs with passed kwargs
non_modality_kwargs = set(kwargs) - set(output_kwargs)
for modality, output_kwarg in output_kwargs.items():
modality_valid_kwargs = set(ModelProcessorKwargs.__annotations__[modality].__annotations__)
if modality in map_preprocessor_kwargs:
preprocessor = getattr(self, map_preprocessor_kwargs[modality], None)
preprocessor_valid_kwargs = (
getattr(preprocessor, "valid_kwargs", None) if preprocessor is not None else None
)
modality_valid_kwargs.update(
set(preprocessor_valid_kwargs.__annotations__ if preprocessor_valid_kwargs is not None else [])
)
for modality_key in modality_valid_kwargs:
# check if we received a structured kwarg dict or not to handle it correctly
if modality in kwargs:
kwarg_value = kwargs[modality].pop(modality_key, "__empty__")
# check if this key was passed as a flat kwarg.
if kwarg_value != "__empty__" and modality_key in non_modality_kwargs:
raise ValueError(
f"Keyword argument {modality_key} was passed two times:\n"
f"in a dictionary for {modality} and as a **kwarg."
)
elif modality_key in kwargs:
# we get a modality_key instead of popping it because modality-specific processors
# can have overlapping kwargs
kwarg_value = kwargs.get(modality_key, "__empty__")
else:
kwarg_value = "__empty__"
if not isinstance(kwarg_value, str) or kwarg_value != "__empty__":
output_kwarg[modality_key] = kwarg_value
used_keys.add(modality_key)
# Determine if kwargs is a flat dictionary or contains nested dictionaries
if any(key in default_kwargs for key in kwargs):
# kwargs is dictionary-based, and some keys match modality names
for modality, subdict in kwargs.items():
if modality in default_kwargs:
for subkey, subvalue in subdict.items():
if subkey not in used_keys:
output_kwargs[modality][subkey] = subvalue
used_keys.add(subkey)
else:
# kwargs is a flat dictionary
for key, kwarg in kwargs.items():
if key not in used_keys and key not in possible_modality_keywords:
logger.warning_once(
f"Keyword argument `{key}` is not a valid argument for this processor and will be ignored."
)
for key, typed_dict_obj in ModelProcessorKwargs.__annotations__.items():
if key in map_preprocessor_kwargs:
preprocessor = getattr(self, map_preprocessor_kwargs[key], None)
if preprocessor is None or getattr(preprocessor, "valid_kwargs", None) is None:
continue
preprocessor_typed_dict_obj = getattr(preprocessor, "valid_kwargs")
typed_dict_obj = TypedDict(
"merged_typed_dict",
{**preprocessor_typed_dict_obj.__annotations__, **typed_dict_obj.__annotations__},
total=False,
)
validate_typed_dict(typed_dict_obj, output_kwargs[key])
return output_kwargs
@classmethod
def from_pretrained(
cls: type[SpecificProcessorType],
pretrained_model_name_or_path: str | os.PathLike,
cache_dir: str | os.PathLike | None = None,
force_download: bool = False,
local_files_only: bool = False,
token: str | bool | None = None,
revision: str = "main",
**kwargs,
) -> SpecificProcessorType:
r"""
Instantiate a processor associated with a pretrained model.
<Tip>
This class method is simply calling the feature extractor
[`~feature_extraction_utils.FeatureExtractionMixin.from_pretrained`], image processor
[`~image_processing_utils.ImageProcessingMixin`] and the tokenizer
[`~tokenization_utils_base.PreTrainedTokenizer.from_pretrained`] methods. Please refer to the docstrings of the
methods above for more information.
</Tip>
Args:
pretrained_model_name_or_path (`str` or `os.PathLike`):
This can be either:
- a string, the *model id* of a pretrained feature_extractor hosted inside a model repo on
huggingface.co.
- a path to a *directory* containing a feature extractor file saved using the
[`~SequenceFeatureExtractor.save_pretrained`] method, e.g., `./my_model_directory/`.
- a path to a saved feature extractor JSON *file*, e.g.,
`./my_model_directory/preprocessor_config.json`.
**kwargs
Additional keyword arguments passed along to both
[`~feature_extraction_utils.FeatureExtractionMixin.from_pretrained`] and
[`~tokenization_utils_base.PreTrainedTokenizer.from_pretrained`].
"""
kwargs["cache_dir"] = cache_dir
kwargs["force_download"] = force_download
kwargs["local_files_only"] = local_files_only
kwargs["revision"] = revision
if token is not None:
kwargs["token"] = token
# Get processor_dict first so we can use it to instantiate non-tokenizer sub-processors
processor_dict, instantiation_kwargs = cls.get_processor_dict(pretrained_model_name_or_path, **kwargs)
args = cls._get_arguments_from_pretrained(pretrained_model_name_or_path, processor_dict, **kwargs)
return cls.from_args_and_dict(args, processor_dict, **instantiation_kwargs)
@classmethod
def get_attributes(cls):
args_in_init = inspect.signature(cls.__init__).parameters.keys()
attributes = []
for sub_processor_type in args_in_init:
# don't treat audio_tokenizer as an attribute
if sub_processor_type == "audio_tokenizer":
continue
if any(modality in sub_processor_type for modality in MODALITY_TO_AUTOPROCESSOR_MAPPING.keys()):
attributes.append(sub_processor_type)
# Legacy processors may not override `__init__` and instead expose modality
# attributes via `<attribute>_class`. In that case, `args_in_init` only exposes
# `*args`/`**kwargs`, so we need to infer the attributes from those class-level
# hints to keep backward compatibility (e.g. dynamic processors stored on the Hub).
if not attributes:
for attribute_name, value in cls.__dict__.items():
if value is None or attribute_name == "audio_tokenizer_class" or not attribute_name.endswith("_class"):
continue
inferred_attribute = attribute_name[: -len("_class")]
if inferred_attribute == "audio_tokenizer":
continue
if any(modality in inferred_attribute for modality in MODALITY_TO_AUTOPROCESSOR_MAPPING.keys()):
attributes.append(inferred_attribute)
return attributes
@classmethod
def register_for_auto_class(cls, auto_class="AutoProcessor"):
"""
Register this class with a given auto class. This should only be used for custom feature extractors as the ones
in the library are already mapped with `AutoProcessor`.
Args:
auto_class (`str` or `type`, *optional*, defaults to `"AutoProcessor"`):
The auto class to register this new feature extractor with.
"""
if not isinstance(auto_class, str):
auto_class = auto_class.__name__
import transformers.models.auto as auto_module
if not hasattr(auto_module, auto_class):
raise ValueError(f"{auto_class} is not a valid auto class.")
cls._auto_class = auto_class
@classmethod
def _load_tokenizer_from_pretrained(
cls, sub_processor_type, pretrained_model_name_or_path, subfolder="", **kwargs
):
auto_processor_class = MODALITY_TO_AUTOPROCESSOR_MAPPING["tokenizer"]
is_primary = sub_processor_type == "tokenizer"
if is_primary:
# Primary tokenizer: load from root
tokenizer = auto_processor_class.from_pretrained(
pretrained_model_name_or_path, subfolder=subfolder, **kwargs
)
else:
# Additional tokenizer: load from subfolder (e.g., "decoder_tokenizer")
tokenizer_subfolder = os.path.join(subfolder, sub_processor_type) if subfolder else sub_processor_type
tokenizer = auto_processor_class.from_pretrained(
pretrained_model_name_or_path, subfolder=tokenizer_subfolder, **kwargs
)
return tokenizer
@classmethod
def _get_arguments_from_pretrained(cls, pretrained_model_name_or_path, processor_dict=None, **kwargs):
"""
Identify and instantiate the subcomponents of Processor classes, such as image processors, tokenizers,
and feature extractors. This method inspects the processor's `__init__` signature to identify parameters
that correspond to known modality types (image_processor, tokenizer, feature_extractor, etc.) or contain
modality names in their attribute name.
For tokenizers: Uses the appropriate Auto class (AutoTokenizer) to load via `.from_pretrained()`.
Additional tokenizers (e.g., "decoder_tokenizer") are loaded from subfolders.
For other sub-processors (image_processor, feature_extractor, etc.): Primary ones are loaded via
Auto class. Additional ones are instantiated from the config stored in processor_config.json
(passed as processor_dict).
Args:
pretrained_model_name_or_path: Path or model id to load from.
processor_dict: Optional dict containing processor config (from processor_config.json).
Required when loading additional non-tokenizer sub-processors.
"""
args = []
processor_dict = processor_dict if processor_dict is not None else {}
# Remove subfolder from kwargs to avoid duplicate keyword arguments
subfolder = kwargs.pop("subfolder", "")
# get args from processor init signature
sub_processors = cls.get_attributes()
for sub_processor_type in sub_processors:
modality = _get_modality_for_attribute(sub_processor_type)
is_primary = sub_processor_type == modality
if (
"tokenizer" in sub_processor_type
): # This is only necessary for the checkpoint in test_processing_mistral3.py which has no config.json and
# the tokenizer_config.json references LlamaTokenizerFast. TODO: update the config on the hub.
if "PixtralProcessor" in cls.__name__:
from .tokenization_utils_tokenizers import TokenizersBackend
tokenizer = TokenizersBackend.from_pretrained(
pretrained_model_name_or_path, subfolder=subfolder, **kwargs
)
else:
tokenizer = cls._load_tokenizer_from_pretrained(
sub_processor_type, pretrained_model_name_or_path, subfolder=subfolder, **kwargs
)
args.append(tokenizer)
elif is_primary:
# Primary non-tokenizer sub-processor: load via Auto class
auto_processor_class = MODALITY_TO_AUTOPROCESSOR_MAPPING[sub_processor_type]
# For backward compatibility, check if sub-processor class name is hardcoded as an attribute of the processor class.
if hasattr(cls, sub_processor_type + "_class"):
sub_processor_class_name = getattr(cls, sub_processor_type + "_class")
logger.warning_once(
f"`{cls.__name__}` defines `{sub_processor_type}_class = '{sub_processor_class_name}'`, "
f"which is deprecated. Register the correct mapping in `{auto_processor_class.__name__}` instead.",
)
auto_processor_class = cls.get_possibly_dynamic_module(sub_processor_class_name)
sub_processor = auto_processor_class.from_pretrained(
pretrained_model_name_or_path, subfolder=subfolder, **kwargs
)
args.append(sub_processor)
elif sub_processor_type in processor_dict:
# Additional non-tokenizer sub-processor: instantiate from config in processor_dict
sub_processor_config = processor_dict[sub_processor_type]
if isinstance(sub_processor_config, dict):
# Determine the class to instantiate
# Image processors have 'image_processor_type', feature extractors have 'feature_extractor_type'
type_key = f"{modality}_type"
class_name = sub_processor_config.get(type_key)
if class_name is None:
raise ValueError(
f"Cannot instantiate {sub_processor_type}: missing '{type_key}' in config. "
f"Config keys: {list(sub_processor_config.keys())}"
)
processor_class = cls.get_possibly_dynamic_module(class_name)
sub_processor = processor_class(**sub_processor_config)
args.append(sub_processor)
else:
raise ValueError(
f"Expected dict for {sub_processor_type} in processor_config.json, "
f"got {type(sub_processor_config)}"
)
else:
raise ValueError(
f"Cannot find config for {sub_processor_type} in processor_config.json. "
f"Available keys: {list(processor_dict.keys())}"
)
return args
@staticmethod
def get_possibly_dynamic_module(module_name):
if hasattr(transformers_module, module_name):
return getattr(transformers_module, module_name)
lookup_locations = [
transformers_module.IMAGE_PROCESSOR_MAPPING,
transformers_module.VIDEO_PROCESSOR_MAPPING,
transformers_module.TOKENIZER_MAPPING,
transformers_module.FEATURE_EXTRACTOR_MAPPING,
transformers_module.MODEL_FOR_AUDIO_TOKENIZATION_MAPPING,
]
for lookup_location in lookup_locations:
for custom_class in lookup_location._extra_content.values():
if isinstance(custom_class, tuple):
for custom_subclass in custom_class:
if custom_subclass is not None and custom_subclass.__name__ == module_name:
return custom_subclass
elif custom_class is not None and custom_class.__name__ == module_name:
return custom_class
raise ValueError(
f"Could not find module {module_name} in `transformers`. If this is a custom class, "
f"it should be registered using the relevant `AutoClass.register()` function so that "
f"other functions can find it!"
)
def batch_decode(self, *args, **kwargs):
"""
This method forwards all its arguments to PreTrainedTokenizer's [`~PreTrainedTokenizer.batch_decode`]. Please
refer to the docstring of this method for more information.
"""
if not hasattr(self, "tokenizer"):
raise ValueError(f"Cannot batch decode text: {self.__class__.__name__} has no tokenizer.")
return self.tokenizer.batch_decode(*args, **kwargs)
def decode(self, *args, **kwargs):
"""
This method forwards all its arguments to PreTrainedTokenizer's [`~PreTrainedTokenizer.decode`]. Please refer to
the docstring of this method for more information.
"""
if not hasattr(self, "tokenizer"):
raise ValueError(f"Cannot decode text: {self.__class__.__name__} has no tokenizer.")
return self.tokenizer.decode(*args, **kwargs)
@property
def model_input_names(self):
model_input_names = []
for attribute_name in self.get_attributes():
attribute = getattr(self, attribute_name, None)
attr_input_names = getattr(attribute, "model_input_names")
model_input_names.extend(attr_input_names)
return model_input_names
@staticmethod
def validate_init_kwargs(processor_config, valid_kwargs):
kwargs_from_config = set(processor_config.keys())
valid_kwargs_set = set(valid_kwargs)
unused_keys = kwargs_from_config - valid_kwargs_set
valid_keys = kwargs_from_config & valid_kwargs_set
unused_kwargs = {k: processor_config[k] for k in unused_keys} if unused_keys else {}
valid_kwargs = {k: processor_config[k] for k in valid_keys} if valid_keys else {}
return unused_kwargs, valid_kwargs
def create_mm_token_type_ids(self, input_ids: list) -> list[list[int]]:
# We have to iterate for each list separately because inputs
# might be non-padded lists and we can't cast numpy on that!
# Then cast numpy as each input for faster indexing
mm_token_type_ids = []
for tokenizer_input in input_ids:
tokenizer_input = np.array(tokenizer_input)
mm_token_types = np.zeros_like(tokenizer_input)
mm_token_types[np.isin(tokenizer_input, self.image_ids)] = 1
mm_token_types[np.isin(tokenizer_input, self.video_ids)] = 2
mm_token_types[np.isin(tokenizer_input, self.audio_ids)] = 3
mm_token_type_ids.append(mm_token_types.tolist())
return mm_token_type_ids
def apply_chat_template(
self,
conversation: list[dict[str, str]] | list[list[dict[str, str]]],
chat_template: str | None = None,
tools: list[dict] | None = None,
documents: list[dict[str, str]] | None = None,
add_generation_prompt: bool = False,
continue_final_message: bool = False,
return_assistant_tokens_mask: bool = False,
tokenize: bool = False,
return_tensors: str | TensorType | None = None,
return_dict: bool = False,
load_audio_from_video: bool = False,
processor_kwargs: dict | None = None,
**kwargs,
) -> str:
"""
Similar to the `apply_chat_template` method on tokenizers, this method applies a Jinja template to input
conversations to turn them into a single tokenizable string.
The input is expected to be in the following format, where each message content is a list consisting of text and
optionally image or video inputs. One can also provide an image, video, URL or local path which will be used to form
`pixel_values` when `return_dict=True`. If not provided, one will get only the formatted text, optionally tokenized text.
conversation = [
{
"role": "user",
"content": [
{"type": "image", "url": "https://www.ilankelman.org/stopsigns/australia.jpg"},
{"type": "text", "text": "Please describe this image in detail."},
],
},
]
Args:
conversation (`Union[list[Dict, [str, str]], list[list[dict[str, str]]]]`):
The conversation to format.
chat_template (`Optional[str]`, *optional*):
The Jinja template to use for formatting the conversation. If not provided, the tokenizer's
chat template is used.
"""
processor_kwargs = processor_kwargs or {}
if chat_template is None:
if isinstance(self.chat_template, dict) and "default" in self.chat_template:
chat_template = self.chat_template["default"]
elif isinstance(self.chat_template, dict):
raise ValueError(
'The processor has multiple chat templates but none of them are named "default". You need to specify'
" which one to use by passing the `chat_template` argument. Available templates are: "
f"{', '.join(self.chat_template.keys())}"
)
elif self.chat_template is not None:
chat_template = self.chat_template
else:
raise ValueError(
"Cannot use apply_chat_template because this processor does not have a chat template."
)
else:
if isinstance(self.chat_template, dict) and chat_template in self.chat_template:
# It's the name of a template, not a full template string
chat_template = self.chat_template[chat_template]
else:
# It's a template string, render it directly
pass
# Users might still be passing processing kwargs in `**kwargs` so we need to filter
# out additional kwargs that the template expects via Jinja2 template introspection
template_kwargs = _get_template_variables(chat_template)
processor_kwargs_from_kwargs = {k: v for k, v in kwargs.items() if k not in template_kwargs}
if processor_kwargs_from_kwargs:
logger.warning(
"Kwargs passed to `processor.__call__` have to be in `processor_kwargs` dict, not in `**kwargs`"
)
processor_kwargs = processor_kwargs_from_kwargs
# Check if tokenizer is fast - use backend attribute if available, otherwise fall back to class name
is_tokenizers_fast = False
if hasattr(self, "tokenizer"):
if hasattr(self.tokenizer, "backend"):
is_tokenizers_fast = self.tokenizer.backend == "tokenizers"
else:
# Fallback to class name check
is_tokenizers_fast = self.tokenizer.__class__.__name__.endswith("Fast")
if continue_final_message:
if add_generation_prompt:
raise ValueError(
"continue_final_message and add_generation_prompt are not compatible. Use continue_final_message when you want the model to continue the final message, and add_generation_prompt when you want to add a header that will prompt it to start a new assistant message instead."
)
if return_assistant_tokens_mask:
raise ValueError("continue_final_message is not compatible with return_assistant_tokens_mask.")
if return_assistant_tokens_mask:
if not is_tokenizers_fast:
raise ValueError(
"`return_assistant_tokens_mask` is not possible with slow tokenizers. Make sure you have `tokenizers` installed. "
"If the error persists, open an issue to support a Fast tokenizer for your model."
)
else:
processor_kwargs["return_offsets_mapping"] = (
True # force offset mapping so we can infer token boundaries
)
# Set the sampling rate to load the audio files if user hasn't already passed with `kwargs`
sampling_rate = kwargs.get("sampling_rate", processor_kwargs.get("sampling_rate"))
if sampling_rate is None:
if hasattr(self, "feature_extractor") and hasattr(self.feature_extractor, "sampling_rate"):
sampling_rate = self.feature_extractor.sampling_rate
else:
sampling_rate = 16_000
if isinstance(conversation, (list, tuple)) and (
isinstance(conversation[0], (list, tuple)) or hasattr(conversation[0], "content")
):
is_batched = True
conversations = conversation
else:
is_batched = False
conversations = [conversation]
# Normalize OpenAI-style "image_url" content blocks to HuggingFace-style "image" blocks
# OpenAI format: {"type": "image_url", "image_url": {"url": "..."}}
# HuggingFace format: {"type": "image", "url": "..."}
for conversation_idx, conversation in enumerate(conversations):
for message in conversation:
if not isinstance(message.get("content"), list):
continue
new_content = []
for content in message["content"]:
if isinstance(content, dict) and content.get("type") == "image_url" and "image_url" in content:
image_url_info = content["image_url"]
url = image_url_info.get("url", "") if isinstance(image_url_info, dict) else image_url_info
new_content.append({"type": "image", "url": url})
else:
new_content.append(content)
message["content"] = new_content
if tokenize:
batch_images, batch_videos = [], []
batch_audios = []
for conversation in conversations:
images, videos = [], []
for message in conversation:
content = message.get("content") or []
visuals = [
content_block for content_block in content if content_block["type"] in ["image", "video"]
]
audio_fnames = [
content_block[key]
for content_block in content
for key in ["audio", "url", "path"]
if key in content_block and content_block["type"] == "audio"
]
image_fnames = [
vision_info[key]
for vision_info in visuals
for key in ["image", "url", "path", "base64"]
if key in vision_info and vision_info["type"] == "image"
]
images.extend(image_fnames)
video_fnames = [
vision_info[key]
for vision_info in visuals
for key in ["video", "url", "path"]
if key in vision_info and vision_info["type"] == "video"
]
videos.extend(video_fnames)
# Audio models do not accept nested list of audios (yet!) so we construct a flat input audio list
if not load_audio_from_video:
for fname in audio_fnames:
batch_audios.append(load_audio(fname, sampling_rate=sampling_rate))
else:
for fname in video_fnames:
# This updates the template in-place and adds audio entry
# to ensure `audio` token is added by jinja
message["content"].append({"type": "audio"})
batch_audios.append(load_audio(fname, sampling_rate=sampling_rate))
# Currently all processors can accept nested list of batches, but not flat list of visuals
# So we'll make a batched list of images and let the processor handle it
batch_images.append(images)
batch_videos.append(videos)
# `kwargs` overwrite special tokens if both are present
template_kwargs = {**self.tokenizer.special_tokens_map, **kwargs}
prompt, generation_indices = render_jinja_template(
conversations=conversations,
tools=tools,
documents=documents,
chat_template=chat_template,
return_assistant_tokens_mask=return_assistant_tokens_mask,
continue_final_message=continue_final_message,
add_generation_prompt=add_generation_prompt,
**template_kwargs,
)
if not is_batched:
prompt = prompt[0]
if tokenize:
# Tokenizer's `apply_chat_template` never adds special tokens when tokenizing
# But processor's `apply_chat_template` didn't have an option to tokenize, so users had to format the prompt
# and pass it to the processor. Users thus never worried about special tokens relying on processor handling
# everything internally. The below line is to keep BC for that and be able to work with model that have
# special tokens in the template (consistent with tokenizers). We dont want to raise warning, it will flood command line
# without actionable solution for users
single_prompt = prompt[0] if is_batched else prompt
if self.tokenizer.bos_token is not None and single_prompt.startswith(self.tokenizer.bos_token):
processor_kwargs["add_special_tokens"] = False
# Always sample frames by default unless explicitly set to `False` by users. If users do not pass `num_frames`/`fps`
# sampling should not done for BC.
if "do_sample_frames" not in processor_kwargs and (
processor_kwargs.get("fps") is not None or processor_kwargs.get("num_frames") is not None
):
processor_kwargs["do_sample_frames"] = True
# Set only is user passes a non-None value. Otherwise wa want to use each processor's own defaults
if return_tensors:
processor_kwargs["return_tensors"] = return_tensors
images_exist = any((im is not None) for im_list in batch_images for im in im_list)
videos_exist = any((vid is not None) for vid_list in batch_videos for vid in vid_list)
out = self(
text=prompt,
images=batch_images if images_exist else None,
videos=batch_videos if videos_exist else None,
audio=batch_audios if batch_audios else None,
**processor_kwargs,
)
if return_dict:
if return_assistant_tokens_mask:
assistant_masks = []
offset_mapping = out.pop("offset_mapping")
input_ids = out["input_ids"]
for i in range(len(input_ids)):
current_mask = [0] * len(input_ids[i])
offsets = offset_mapping[i]
offset_starts = [start for start, end in offsets]
for assistant_start_char, assistant_end_char in generation_indices[i]:
start_pos = bisect.bisect_left(offset_starts, assistant_start_char)
end_pos = bisect.bisect_left(offset_starts, assistant_end_char)
if not (
start_pos >= 0
and start_pos < len(offsets)
and offsets[start_pos][0] <= assistant_start_char < offsets[start_pos][1]
):
# start_token is out of bounds maybe due to truncation.
continue
# Ensure end_pos is also within bounds
if end_pos > len(input_ids[i]):
end_pos = len(input_ids[i])
for token_id in range(start_pos, end_pos if end_pos else len(input_ids[i])):
current_mask[token_id] = 1
assistant_masks.append(current_mask)
out["assistant_masks"] = assistant_masks
out.convert_to_tensors(tensor_type=return_tensors)
return out
else:
return out["input_ids"]
return prompt
def parse_response(
self,
response: "str | list[str | int | list[int]] | np.ndarray | torch.Tensor",
schema: list | dict | None = None,
):
"""
Converts an output string created by generating text from a model into a parsed message dictionary.
This method is intended for use with chat models, and will read the tokenizer's `response_schema` attribute to
control parsing, although this can be overridden by passing a `response_schema` argument directly.
Args:
response (`str`):
The output string generated by the model. This can be either a decoded string or list of strings,
or token IDs as a list/array.
schema (`Union[list, dict]`, *optional*):
A response schema that indicates the expected output format and how parsing should be performed.
If not provided, the tokenizer's `response_schema` attribute will be used.
"""
if not hasattr(self, "tokenizer"):
raise ValueError("Can't use parse_response on a processor class without a tokenizer!")
return self.tokenizer.parse_response(response, schema)
def post_process_multimodal_output(
self, generated_outputs, skip_special_tokens=True, generation_mode=None, **kwargs
):
"""
Post-process the output of a multimodal model to return the requested modality output.
If the model cannot generated the requested modality, an error will be raised.
Args:
generated_outputs (`torch.Tensor` or `np.ndarray`):
The output of the model `generate` function. The output is expected to be a tensor of shape `(batch_size, sequence_length)`
or `(sequence_length,)`.
skip_special_tokens (`bool`, *optional*, defaults to `True`):
Whether or not to remove special tokens in the output. Argument passed to the tokenizer's `batch_decode` method.
generation_mode (`str`, *optional*):
Generation mode indicated which modality to output and can be one of `["text", "image", "audio"]`.
**kwargs:
Additional arguments to be passed to the tokenizer's `batch_decode method`.
Returns:
`list[str]`: The decoded text.
"""
if generation_mode is not None and generation_mode != "text":
raise ValueError(
f"{self.__class__.__name__} got an unexpected generation_mode={generation_mode}. Supported options are only [`text`]"
)
return self.post_process_image_text_to_text(
generated_outputs, skip_special_tokens=skip_special_tokens, **kwargs
)
def post_process_image_text_to_text(self, generated_outputs, skip_special_tokens=True, **kwargs):
"""
Post-process the output of a vlm to decode the text.
Args:
generated_outputs (`torch.Tensor` or `np.ndarray`):
The output of the model `generate` function. The output is expected to be a tensor of shape `(batch_size, sequence_length)`
or `(sequence_length,)`.
skip_special_tokens (`bool`, *optional*, defaults to `True`):
Whether or not to remove special tokens in the output. Argument passed to the tokenizer's `decode` method.
**kwargs:
Additional arguments to be passed to the tokenizer's `decode` method.
Returns:
`list[str]`: The decoded text.
"""
return self.tokenizer.decode(generated_outputs, skip_special_tokens=skip_special_tokens, **kwargs)
def _check_special_mm_tokens(self, text: list[str], text_inputs: "BatchFeature", modalities: list[str]):
"""
Checks that number of special tokens in text and processed text is same. The count can be different
if tokenized text was truncated, leading to issues in model code.
"""
for modality in modalities:
token_str = getattr(self, f"{modality}_token")
token_id = getattr(self, f"{modality}_token_id")
ids_count = [list(ids).count(token_id) for ids in text_inputs["input_ids"]]
text_count = [sample.count(token_str) for sample in text]
if ids_count != text_count:
raise ValueError(
f"Mismatch in `{modality}` token count between text and `input_ids`. Got ids={ids_count} and text={text_count}. "
"Likely due to `truncation='max_length'`. Please disable truncation or increase `max_length`."
)
ProcessorMixin.push_to_hub = copy_func(ProcessorMixin.push_to_hub)
if ProcessorMixin.push_to_hub.__doc__ is not None:
ProcessorMixin.push_to_hub.__doc__ = ProcessorMixin.push_to_hub.__doc__.format(
object="processor", object_class="AutoProcessor", object_files="processor files"
)
|