Spaces:
Runtime error
Runtime error
File size: 121,886 Bytes
455a40f |
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 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 |
# coding=utf-8
# Copyright 2019 HuggingFace Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import copy
import inspect
import json
import os
import random
import tempfile
import unittest
import unittest.mock as mock
from dataclasses import fields
from importlib import import_module
from math import isnan
from typing import List, Tuple, get_type_hints
from datasets import Dataset
from huggingface_hub import HfFolder, Repository, delete_repo
from huggingface_hub.file_download import http_get
from requests.exceptions import HTTPError
from transformers import is_tf_available, is_torch_available
from transformers.configuration_utils import PretrainedConfig
from transformers.models.auto import get_values
from transformers.testing_utils import ( # noqa: F401
TOKEN,
USER,
CaptureLogger,
CaptureStdout,
_tf_gpu_memory_limit,
is_pt_tf_cross_test,
is_staging_test,
require_safetensors,
require_tf,
require_tf2onnx,
slow,
tooslow,
torch_device,
)
from transformers.utils import (
CONFIG_NAME,
GENERATION_CONFIG_NAME,
SAFE_WEIGHTS_NAME,
TF2_WEIGHTS_INDEX_NAME,
TF2_WEIGHTS_NAME,
logging,
)
from transformers.utils.generic import ModelOutput
logger = logging.get_logger(__name__)
if is_tf_available():
import h5py
import numpy as np
import tensorflow as tf
from transformers import (
TF_MODEL_FOR_CAUSAL_LM_MAPPING,
TF_MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING,
TF_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING,
TF_MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING,
TF_MODEL_FOR_MASKED_LM_MAPPING,
TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING,
TF_MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING,
TF_MODEL_FOR_PRETRAINING_MAPPING,
TF_MODEL_FOR_QUESTION_ANSWERING_MAPPING,
TF_MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING,
TF_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING,
TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING,
TF_MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING,
TF_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING,
BertConfig,
PreTrainedModel,
PushToHubCallback,
RagRetriever,
TFAutoModel,
TFAutoModelForSequenceClassification,
TFBertForMaskedLM,
TFBertForSequenceClassification,
TFBertModel,
TFPreTrainedModel,
TFRagModel,
TFSharedEmbeddings,
)
from transformers.generation import (
TFBeamSampleDecoderOnlyOutput,
TFBeamSampleEncoderDecoderOutput,
TFBeamSearchDecoderOnlyOutput,
TFBeamSearchEncoderDecoderOutput,
TFGreedySearchDecoderOnlyOutput,
TFGreedySearchEncoderDecoderOutput,
TFSampleDecoderOnlyOutput,
TFSampleEncoderDecoderOutput,
)
from transformers.modeling_tf_utils import tf_shard_checkpoint, unpack_inputs
from transformers.tf_utils import stable_softmax
tf.config.experimental.enable_tensor_float_32_execution(False)
if _tf_gpu_memory_limit is not None:
gpus = tf.config.list_physical_devices("GPU")
for gpu in gpus:
# Restrict TensorFlow to only allocate x GB of memory on the GPUs
try:
tf.config.set_logical_device_configuration(
gpu, [tf.config.LogicalDeviceConfiguration(memory_limit=_tf_gpu_memory_limit)]
)
logical_gpus = tf.config.list_logical_devices("GPU")
print("Logical GPUs", logical_gpus)
except RuntimeError as e:
# Virtual devices must be set before GPUs have been initialized
print(e)
if is_torch_available():
import torch
from transformers import BertModel
def _config_zero_init(config):
configs_no_init = copy.deepcopy(config)
for key in configs_no_init.__dict__.keys():
if "_range" in key or "_std" in key:
setattr(configs_no_init, key, 0.0)
return configs_no_init
def _return_type_has_loss(model):
return_type = get_type_hints(model.call)
if "return" not in return_type:
return False
return_type = return_type["return"]
if hasattr(return_type, "__args__"): # Awkward check for union because UnionType only turns up in 3.10
for type_annotation in return_type.__args__:
if inspect.isclass(type_annotation) and issubclass(type_annotation, ModelOutput):
field_names = [field.name for field in fields(type_annotation)]
if "loss" in field_names:
return True
return False
elif isinstance(return_type, tuple):
return False
elif isinstance(return_type, ModelOutput):
class_fields = fields(return_type)
return "loss" in class_fields
return False
@require_tf
class TFModelTesterMixin:
model_tester = None
all_model_classes = ()
all_generative_model_classes = ()
test_mismatched_shapes = True
test_resize_embeddings = True
test_head_masking = True
is_encoder_decoder = False
has_attentions = True
def _prepare_for_class(self, inputs_dict, model_class, return_labels=False) -> dict:
inputs_dict = copy.deepcopy(inputs_dict)
if model_class in get_values(TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING):
inputs_dict = {
k: tf.tile(tf.expand_dims(v, 1), (1, self.model_tester.num_choices) + (1,) * (v.ndim - 1))
if isinstance(v, tf.Tensor) and v.ndim > 0
else v
for k, v in inputs_dict.items()
}
if return_labels:
if model_class in get_values(TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING):
inputs_dict["labels"] = tf.ones(self.model_tester.batch_size, dtype=tf.int32)
elif model_class in [
*get_values(TF_MODEL_FOR_QUESTION_ANSWERING_MAPPING),
*get_values(TF_MODEL_FOR_DOCUMENT_QUESTION_ANSWERING_MAPPING),
]:
inputs_dict["start_positions"] = tf.zeros(self.model_tester.batch_size, dtype=tf.int32)
inputs_dict["end_positions"] = tf.zeros(self.model_tester.batch_size, dtype=tf.int32)
elif model_class in [
*get_values(TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING),
*get_values(TF_MODEL_FOR_IMAGE_CLASSIFICATION_MAPPING),
]:
inputs_dict["labels"] = tf.zeros(self.model_tester.batch_size, dtype=tf.int32)
elif model_class in get_values(TF_MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING):
inputs_dict["next_sentence_label"] = tf.zeros(self.model_tester.batch_size, dtype=tf.int32)
elif model_class in [
*get_values(TF_MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING),
*get_values(TF_MODEL_FOR_CAUSAL_LM_MAPPING),
*get_values(TF_MODEL_FOR_MASKED_LM_MAPPING),
*get_values(TF_MODEL_FOR_PRETRAINING_MAPPING),
*get_values(TF_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING),
*get_values(TF_MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING),
] and "labels" in dict(inspect.signature(model_class.call).parameters):
inputs_dict["labels"] = tf.zeros(
(self.model_tester.batch_size, self.model_tester.seq_length), dtype=tf.int32
)
elif model_class in get_values(TF_MODEL_FOR_MASKED_IMAGE_MODELING_MAPPING):
num_patches = self.model_tester.image_size // self.model_tester.patch_size
inputs_dict["bool_masked_pos"] = tf.zeros(
(self.model_tester.batch_size, num_patches**2), dtype=tf.int32
)
elif model_class in get_values(TF_MODEL_FOR_SEMANTIC_SEGMENTATION_MAPPING):
batch_size, num_channels, height, width = inputs_dict["pixel_values"].shape
inputs_dict["labels"] = tf.zeros((self.model_tester.batch_size, height, width), dtype=tf.int32)
elif model_class.__name__.endswith("ForCTC"):
# When we have enough CTC models for an AutoClass, we should use their mapping instead of name checks
inputs_dict["labels"] = tf.zeros(
(self.model_tester.batch_size, self.model_tester.seq_length), dtype=tf.int32
)
return inputs_dict
def test_initialization(self):
pass
def test_save_load(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
outputs = model(self._prepare_for_class(inputs_dict, model_class))
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname, saved_model=False)
# the config file (and the generation config file, if it can generate) should be saved
self.assertTrue(os.path.exists(os.path.join(tmpdirname, CONFIG_NAME)))
self.assertEqual(
model.can_generate(), os.path.exists(os.path.join(tmpdirname, GENERATION_CONFIG_NAME))
)
model = model_class.from_pretrained(tmpdirname)
after_outputs = model(self._prepare_for_class(inputs_dict, model_class))
self.assert_outputs_same(after_outputs, outputs)
def test_save_load_config(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
outputs = model(self._prepare_for_class(inputs_dict, model_class))
model_config = model.get_config()
# make sure that returned config is jsonifiable, which is required by keras
json.dumps(model_config)
new_model = model_class.from_config(model.get_config())
# make sure it also accepts a normal config
_ = model_class.from_config(model.config)
_ = new_model(self._prepare_for_class(inputs_dict, model_class)) # Build model
new_model.set_weights(model.get_weights())
after_outputs = new_model(self._prepare_for_class(inputs_dict, model_class))
self.assert_outputs_same(after_outputs, outputs)
@slow
def test_saved_model_creation(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.output_hidden_states = False
config.output_attentions = False
if hasattr(config, "use_cache"):
config.use_cache = False
model_class = self.all_model_classes[0]
class_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
model = model_class(config)
model(class_inputs_dict)
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname, saved_model=True)
saved_model_dir = os.path.join(tmpdirname, "saved_model", "1")
self.assertTrue(os.path.exists(saved_model_dir))
def test_prepare_serving_output(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.output_hidden_states = True
config.output_attentions = self.has_attentions
for model_class in self.all_model_classes:
model = model_class(config)
inputs = self._prepare_for_class(inputs_dict, model_class)
outputs = model(inputs)
serving_outputs = model.serving_output(outputs)
for k, v in serving_outputs.items():
# Check that we have one of three possible outputs: None, tuple of tensors or a tensor
if isinstance(v, tuple):
self.assertTrue(all(isinstance(elem, tf.Tensor) for elem in v))
elif v is not None:
self.assertIsInstance(v, tf.Tensor)
else:
self.assertIsNone(v)
def test_forward_signature(self):
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
signature = inspect.signature(model.call)
# signature.parameters is an OrderedDict => so arg_names order is deterministic
arg_names = [*signature.parameters.keys()]
if model.config.is_encoder_decoder:
expected_arg_names = [
"input_ids",
"attention_mask",
"decoder_input_ids",
"decoder_attention_mask",
]
expected_arg_names.extend(["decoder_position_ids"] if "decoder_position_ids" in arg_names else [])
expected_arg_names.extend(
["head_mask", "decoder_head_mask"] if "head_mask" and "decoder_head_mask" in arg_names else []
)
expected_arg_names.extend(
["cross_attn_head_mask", "encoder_outputs"]
if "cross_attn_head_mask" in arg_names
else ["encoder_outputs"]
)
self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
else:
expected_arg_names = ["input_ids"]
self.assertListEqual(arg_names[:1], expected_arg_names)
def test_onnx_compliancy(self):
if not self.test_onnx:
return
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
INTERNAL_OPS = [
"Assert",
"AssignVariableOp",
"EmptyTensorList",
"ReadVariableOp",
"ResourceGather",
"TruncatedNormal",
"VarHandleOp",
"VarIsInitializedOp",
]
onnx_ops = []
with open(os.path.join(".", "utils", "tf_ops", "onnx.json")) as f:
onnx_opsets = json.load(f)["opsets"]
for i in range(1, self.onnx_min_opset + 1):
onnx_ops.extend(onnx_opsets[str(i)])
for model_class in self.all_model_classes:
model_op_names = set()
with tf.Graph().as_default() as g:
model = model_class(config)
model(model.dummy_inputs)
for op in g.get_operations():
model_op_names.add(op.node_def.op)
model_op_names = sorted(model_op_names)
incompatible_ops = []
for op in model_op_names:
if op not in onnx_ops and op not in INTERNAL_OPS:
incompatible_ops.append(op)
self.assertEqual(len(incompatible_ops), 0, incompatible_ops)
@require_tf2onnx
@slow
def test_onnx_runtime_optimize(self):
if not self.test_onnx:
return
import onnxruntime
import tf2onnx
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
model(model.dummy_inputs)
onnx_model_proto, _ = tf2onnx.convert.from_keras(model, opset=self.onnx_min_opset)
onnxruntime.InferenceSession(onnx_model_proto.SerializeToString())
def test_keras_save_load(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
tf_main_layer_classes = {
module_member
for model_class in self.all_model_classes
for module in (import_module(model_class.__module__),)
for module_member_name in dir(module)
if module_member_name.endswith("MainLayer")
# This condition is required, since `modeling_tf_clip.py` has 3 classes whose names end with `MainLayer`.
and module_member_name[: -len("MainLayer")] == model_class.__name__[: -len("Model")]
for module_member in (getattr(module, module_member_name),)
if isinstance(module_member, type)
and tf.keras.layers.Layer in module_member.__bases__
and getattr(module_member, "_keras_serializable", False)
}
for main_layer_class in tf_main_layer_classes:
# T5MainLayer needs an embed_tokens parameter when called without the inputs_embeds parameter
if "T5" in main_layer_class.__name__:
# Take the same values than in TFT5ModelTester for this shared layer
shared = TFSharedEmbeddings(99, 32, name="shared")
config.use_cache = inputs_dict.pop("use_cache", None)
main_layer = main_layer_class(config, embed_tokens=shared)
else:
main_layer = main_layer_class(config)
symbolic_inputs = {
name: tf.keras.Input(tensor.shape[1:], dtype=tensor.dtype) for name, tensor in inputs_dict.items()
}
model = tf.keras.Model(symbolic_inputs, outputs=main_layer(symbolic_inputs))
outputs = model(inputs_dict)
with tempfile.TemporaryDirectory() as tmpdirname:
filepath = os.path.join(tmpdirname, "keras_model.h5")
model.save(filepath)
if "T5" in main_layer_class.__name__:
model = tf.keras.models.load_model(
filepath,
custom_objects={
main_layer_class.__name__: main_layer_class,
"TFSharedEmbeddings": TFSharedEmbeddings,
},
)
else:
model = tf.keras.models.load_model(
filepath, custom_objects={main_layer_class.__name__: main_layer_class}
)
assert isinstance(model, tf.keras.Model)
after_outputs = model(inputs_dict)
self.assert_outputs_same(after_outputs, outputs)
def assert_outputs_same(self, after_outputs, outputs):
# Make sure we don't have nans
if isinstance(after_outputs, tf.Tensor):
out_1 = after_outputs.numpy()
elif isinstance(after_outputs, dict):
out_1 = after_outputs[list(after_outputs.keys())[0]].numpy()
else:
out_1 = after_outputs[0].numpy()
out_2 = outputs[0].numpy()
self.assertEqual(out_1.shape, out_2.shape)
out_1 = out_1[~np.isnan(out_1)]
out_2 = out_2[~np.isnan(out_2)]
max_diff = np.amax(np.abs(out_1 - out_2))
self.assertLessEqual(max_diff, 1e-5)
# Don't copy this method to model specific test file!
# TODO: remove this method once the issues are all fixed!
def _make_attention_mask_non_null(self, inputs_dict):
"""Make sure no sequence has all zeros as attention mask"""
for k in ["attention_mask", "encoder_attention_mask", "decoder_attention_mask"]:
if k in inputs_dict:
attention_mask = inputs_dict[k]
# Make sure no all 0s attention masks - to avoid failure at this moment.
# Put `1` at the beginning of sequences to make it still work when combining causal attention masks.
# TODO: remove this line once a fix regarding large negative values for attention mask is done.
attention_mask = tf.concat(
[tf.ones_like(attention_mask[:, :1], dtype=attention_mask.dtype), attention_mask[:, 1:]], axis=-1
)
# Here we make the first sequence with all 0s as attention mask.
# Currently, this will fail for `TFWav2Vec2Model`. This is caused by the different large negative
# values, like `1e-4`, `1e-9`, `1e-30` and `-inf` for attention mask across models/frameworks.
# TODO: enable this block once the large negative values thing is cleaned up.
# (see https://github.com/huggingface/transformers/issues/14859)
# attention_mask = tf.concat(
# [
# tf.zeros_like(attention_mask[:1], dtype=tf.int32),
# tf.cast(attention_mask[1:], dtype=tf.int32)
# ],
# axis=0
# )
inputs_dict[k] = attention_mask
# Don't copy this method to model specific test file!
# TODO: remove this method once the issues are all fixed!
def _postprocessing_to_ignore_test_cases(self, tf_outputs, pt_outputs, model_class):
"""For temporarily ignoring some failed test cases (issues to be fixed)"""
tf_keys = {k for k, v in tf_outputs.items() if v is not None}
pt_keys = {k for k, v in pt_outputs.items() if v is not None}
key_differences = tf_keys.symmetric_difference(pt_keys)
if model_class.__name__ in [
"TFFlaubertWithLMHeadModel",
"TFFunnelForPreTraining",
"TFElectraForPreTraining",
"TFXLMWithLMHeadModel",
"TFTransfoXLLMHeadModel",
]:
for k in key_differences:
if k in ["loss", "losses"]:
tf_keys.discard(k)
pt_keys.discard(k)
elif model_class.__name__.startswith("TFGPT2"):
# `TFGPT2` has `past_key_values` as a tensor while `GPT2` has it as a tuple.
tf_keys.discard("past_key_values")
pt_keys.discard("past_key_values")
# create new outputs from the remaining fields
new_tf_outputs = type(tf_outputs)(**{k: tf_outputs[k] for k in tf_keys})
new_pt_outputs = type(pt_outputs)(**{k: pt_outputs[k] for k in pt_keys})
return new_tf_outputs, new_pt_outputs
def check_pt_tf_outputs(self, tf_outputs, pt_outputs, model_class, tol=1e-5, name="outputs", attributes=None):
"""Check the outputs from PyTorch and TensorFlow models are close enough. Checks are done in a recursive way.
Args:
model_class: The class of the model that is currently testing. For example, `TFBertModel`,
TFBertForMaskedLM`, `TFBertForSequenceClassification`, etc. Mainly used for providing more informative
error messages.
name (`str`): The name of the output. For example, `output.hidden_states`, `output.attentions`, etc.
attributes (`Tuple[str]`): The names of the output's element if the output is a tuple/list with each element
being a named field in the output.
"""
self.assertEqual(type(name), str)
if attributes is not None:
self.assertEqual(type(attributes), tuple, f"{name}: The argument `attributes` should be a `tuple`")
# Allow `ModelOutput` (e.g. `CLIPOutput` has `text_model_output` and `vision_model_output`).
if isinstance(tf_outputs, ModelOutput):
self.assertTrue(
isinstance(pt_outputs, ModelOutput),
f"{name}: `pt_outputs` should an instance of `ModelOutput` when `tf_outputs` is",
)
# Don't copy this block to model specific test file!
# TODO: remove this method and this line after issues are fixed
tf_outputs, pt_outputs = self._postprocessing_to_ignore_test_cases(tf_outputs, pt_outputs, model_class)
tf_keys = [k for k, v in tf_outputs.items() if v is not None]
pt_keys = [k for k, v in pt_outputs.items() if v is not None]
self.assertEqual(tf_keys, pt_keys, f"{name}: Output keys differ between TF and PyTorch")
# convert to the case of `tuple`
# appending each key to the current (string) `names`
attributes = tuple([f"{name}.{k}" for k in tf_keys])
self.check_pt_tf_outputs(
tf_outputs.to_tuple(), pt_outputs.to_tuple(), model_class, tol=tol, name=name, attributes=attributes
)
# Allow `list` (e.g. `TransfoXLModelOutput.mems` is a list of tensors.)
elif type(tf_outputs) in [tuple, list]:
self.assertEqual(type(tf_outputs), type(pt_outputs), f"{name}: Output types differ between TF and PyTorch")
self.assertEqual(len(tf_outputs), len(pt_outputs), f"{name}: Output lengths differ between TF and PyTorch")
if attributes is not None:
# case 1: each output has assigned name (e.g. a tuple form of a `ModelOutput`)
self.assertEqual(
len(attributes),
len(tf_outputs),
f"{name}: The tuple `names` should have the same length as `tf_outputs`",
)
else:
# case 2: each output has no assigned name (e.g. hidden states of each layer) -> add an index to `names`
attributes = tuple([f"{name}_{idx}" for idx in range(len(tf_outputs))])
for tf_output, pt_output, attr in zip(tf_outputs, pt_outputs, attributes):
self.check_pt_tf_outputs(tf_output, pt_output, model_class, tol=tol, name=attr)
elif isinstance(tf_outputs, tf.Tensor):
self.assertTrue(
isinstance(pt_outputs, torch.Tensor), f"{name}: `pt_outputs` should a tensor when `tf_outputs` is"
)
tf_outputs = tf_outputs.numpy()
pt_outputs = pt_outputs.detach().to("cpu").numpy()
self.assertEqual(
tf_outputs.shape, pt_outputs.shape, f"{name}: Output shapes differ between TF and PyTorch"
)
# deal with NumPy's scalars to make replacing nan values by 0 work.
if np.isscalar(tf_outputs):
tf_outputs = np.array([tf_outputs])
pt_outputs = np.array([pt_outputs])
tf_nans = np.isnan(tf_outputs)
pt_nans = np.isnan(pt_outputs)
pt_outputs[tf_nans] = 0
tf_outputs[tf_nans] = 0
pt_outputs[pt_nans] = 0
tf_outputs[pt_nans] = 0
max_diff = np.amax(np.abs(tf_outputs - pt_outputs))
self.assertLessEqual(max_diff, tol, f"{name}: Difference between torch and tf is {max_diff} (>= {tol}).")
else:
raise ValueError(
"`tf_outputs` should be an instance of `tf.Tensor`, a `tuple`, or an instance of `tf.Tensor`. Got"
f" {type(tf_outputs)} instead."
)
def prepare_pt_inputs_from_tf_inputs(self, tf_inputs_dict):
pt_inputs_dict = {}
for name, key in tf_inputs_dict.items():
if type(key) == bool:
pt_inputs_dict[name] = key
elif name == "input_values":
pt_inputs_dict[name] = torch.from_numpy(key.numpy()).to(torch.float32)
elif name == "pixel_values":
pt_inputs_dict[name] = torch.from_numpy(key.numpy()).to(torch.float32)
elif name == "input_features":
pt_inputs_dict[name] = torch.from_numpy(key.numpy()).to(torch.float32)
# other general float inputs
elif tf_inputs_dict[name].dtype.is_floating:
pt_inputs_dict[name] = torch.from_numpy(key.numpy()).to(torch.float32)
else:
pt_inputs_dict[name] = torch.from_numpy(key.numpy()).to(torch.long)
return pt_inputs_dict
def check_pt_tf_models(self, tf_model, pt_model, tf_inputs_dict):
pt_inputs_dict = self.prepare_pt_inputs_from_tf_inputs(tf_inputs_dict)
# send pytorch inputs to the correct device
pt_inputs_dict = {
k: v.to(device=torch_device) if isinstance(v, torch.Tensor) else v for k, v in pt_inputs_dict.items()
}
# send pytorch model to the correct device
pt_model.to(torch_device)
# Check predictions on first output (logits/hidden-states) are close enough given low-level computational differences
pt_model.eval()
with torch.no_grad():
pt_outputs = pt_model(**pt_inputs_dict)
tf_outputs = tf_model(tf_inputs_dict)
# tf models returned loss is usually a tensor rather than a scalar.
# (see `hf_compute_loss`: it uses `tf.keras.losses.Reduction.NONE`)
# Change it here to a scalar to match PyTorch models' loss
tf_loss = getattr(tf_outputs, "loss", None)
if tf_loss is not None:
tf_outputs.loss = tf.math.reduce_mean(tf_loss)
self.check_pt_tf_outputs(tf_outputs, pt_outputs, type(tf_model))
@is_pt_tf_cross_test
def test_pt_tf_model_equivalence(self, allow_missing_keys=False):
import transformers
for model_class in self.all_model_classes:
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
# Output all for aggressive testing
config.output_hidden_states = True
config.output_attentions = self.has_attentions
# Make sure no sequence has all zeros as attention mask, otherwise some tests fail due to the inconsistency
# of the usage `1e-4`, `1e-9`, `1e-30`, `-inf`.
# TODO: Use a uniform value for all models, make sure all tests pass without this processing, and remove it.
self._make_attention_mask_non_null(inputs_dict)
pt_model_class_name = model_class.__name__[2:] # Skip the "TF" at the beginning
pt_model_class = getattr(transformers, pt_model_class_name)
tf_model = model_class(config)
pt_model = pt_model_class(config)
tf_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
tf_inputs_dict_with_labels = self._prepare_for_class(
inputs_dict,
model_class,
# Not all models accept "labels" in the forward pass (yet :) )
return_labels=True if "labels" in inspect.signature(model_class.call).parameters.keys() else False,
)
# For some models (e.g. base models), there is no label returned.
# Set the input dict to `None` to avoid check outputs twice for the same input dicts.
if not set(tf_inputs_dict_with_labels.keys()).symmetric_difference(tf_inputs_dict.keys()):
tf_inputs_dict_with_labels = None
# Check we can load pt model in tf and vice-versa with model => model functions
tf_model = transformers.load_pytorch_model_in_tf2_model(
tf_model, pt_model, tf_inputs=tf_inputs_dict, allow_missing_keys=allow_missing_keys
)
pt_model = transformers.load_tf2_model_in_pytorch_model(
pt_model, tf_model, allow_missing_keys=allow_missing_keys
)
# Original test: check without `labels`
self.check_pt_tf_models(tf_model, pt_model, tf_inputs_dict)
# check with `labels`
if tf_inputs_dict_with_labels:
self.check_pt_tf_models(tf_model, pt_model, tf_inputs_dict_with_labels)
# Check we can load pt model in tf and vice-versa with checkpoint => model functions
with tempfile.TemporaryDirectory() as tmpdirname:
pt_checkpoint_path = os.path.join(tmpdirname, "pt_model.bin")
torch.save(pt_model.state_dict(), pt_checkpoint_path)
tf_model = transformers.load_pytorch_checkpoint_in_tf2_model(
tf_model, pt_checkpoint_path, allow_missing_keys=allow_missing_keys
)
tf_checkpoint_path = os.path.join(tmpdirname, "tf_model.h5")
tf_model.save_weights(tf_checkpoint_path)
pt_model = transformers.load_tf2_checkpoint_in_pytorch_model(
pt_model, tf_checkpoint_path, allow_missing_keys=allow_missing_keys
)
# Original test: check without `labels`
self.check_pt_tf_models(tf_model, pt_model, tf_inputs_dict)
# check with `labels`
if tf_inputs_dict_with_labels:
self.check_pt_tf_models(tf_model, pt_model, tf_inputs_dict_with_labels)
def test_compile_tf_model(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
max_input = getattr(self.model_tester, "max_position_embeddings", 512)
optimizer = tf.keras.optimizers.Adam(learning_rate=3e-5, epsilon=1e-08, clipnorm=1.0)
loss = tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True)
metric = tf.keras.metrics.SparseCategoricalAccuracy("accuracy")
for model_class in self.all_model_classes:
if model_class.__name__ in ["TFSpeech2TextModel", "TFSpeech2TextForConditionalGeneration"]:
inputs = {
"decoder_input_ids": tf.keras.Input(
batch_shape=(2, max_input),
name="decoder_input_ids",
dtype="int32",
),
"input_features": tf.keras.Input(
batch_shape=(
2,
max_input,
self.model_tester.input_feat_per_channel * self.model_tester.input_channels,
),
name="input_features",
dtype="float32",
),
}
elif model_class.__name__ in ["TFWhisperModel", "TFWhisperForConditionalGeneration"]:
inputs = {
"decoder_input_ids": tf.keras.Input(
batch_shape=(2, max_input),
name="decoder_input_ids",
dtype="int32",
),
"input_features": tf.keras.Input(
batch_shape=(
2,
self.model_tester.num_mel_bins,
self.model_tester.seq_length,
),
name="input_features",
dtype="float32",
),
}
elif self.is_encoder_decoder:
inputs = {
"decoder_input_ids": tf.keras.Input(
batch_shape=(2, max_input),
name="decoder_input_ids",
dtype="int32",
),
"input_ids": tf.keras.Input(batch_shape=(2, max_input), name="input_ids", dtype="int32"),
}
# `pixel_values` implies that the input is an image
elif model_class.main_input_name == "pixel_values":
inputs = tf.keras.Input(
batch_shape=(
3,
self.model_tester.num_channels,
self.model_tester.image_size,
self.model_tester.image_size,
),
name="pixel_values",
dtype="float32",
)
elif model_class.__name__ in ["TFCLIPModel", "TFGroupViTModel", "TFBlipModel"]:
inputs = {
"input_ids": tf.keras.Input(batch_shape=(3, max_input), name="input_ids", dtype="int32"),
"pixel_values": tf.keras.Input(
batch_shape=(
3,
self.model_tester.vision_model_tester.num_channels,
self.model_tester.vision_model_tester.image_size,
self.model_tester.vision_model_tester.image_size,
),
name="pixel_values",
dtype="float32",
),
}
elif model_class in get_values(TF_MODEL_FOR_MULTIPLE_CHOICE_MAPPING):
inputs = tf.keras.Input(batch_shape=(4, 2, max_input), name="input_ids", dtype="int32")
else:
inputs = tf.keras.Input(batch_shape=(2, max_input), name="input_ids", dtype="int32")
# Prepare our model
model = model_class(config)
model(self._prepare_for_class(inputs_dict, model_class)) # Model must be called before saving.
# Let's load it from the disk to be sure we can use pretrained weights
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname, saved_model=False)
model = model_class.from_pretrained(tmpdirname)
outputs_dict = model(inputs)
hidden_states = outputs_dict[0]
# Add a dense layer on top to test integration with other keras modules
outputs = tf.keras.layers.Dense(2, activation="softmax", name="outputs")(hidden_states)
# Compile extended model
extended_model = tf.keras.Model(inputs=[inputs], outputs=[outputs])
extended_model.compile(optimizer=optimizer, loss=loss, metrics=[metric])
def test_keyword_and_dict_args(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
inputs = self._prepare_for_class(inputs_dict, model_class)
outputs_dict = model(inputs)
inputs_keywords = copy.deepcopy(self._prepare_for_class(inputs_dict, model_class))
outputs_keywords = model(**inputs_keywords)
output_dict = outputs_dict[0].numpy()
output_keywords = outputs_keywords[0].numpy()
self.assertLess(np.sum(np.abs(output_dict - output_keywords)), 1e-6)
def test_attention_outputs(self):
if not self.has_attentions:
self.skipTest(reason="Model does not output attentions")
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.return_dict = True
decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", self.model_tester.seq_length)
encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", self.model_tester.seq_length)
decoder_key_length = getattr(self.model_tester, "key_length", decoder_seq_length)
encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
def check_decoder_attentions_output(outputs):
out_len = len(outputs)
self.assertEqual(min(out_len % 2, out_len % 5), 0) # differentiation due to newly added cross_attentions
decoder_attentions = outputs.decoder_attentions
self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(decoder_attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, decoder_seq_length, decoder_key_length],
)
def check_encoder_attentions_output(outputs):
attentions = [
t.numpy() for t in (outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions)
]
self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
self.assertListEqual(
list(attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
)
for model_class in self.all_model_classes:
inputs_dict["output_attentions"] = True
config.output_hidden_states = False
model = model_class(config)
outputs = model(self._prepare_for_class(inputs_dict, model_class))
out_len = len(outputs)
self.assertEqual(config.output_hidden_states, False)
check_encoder_attentions_output(outputs)
if self.is_encoder_decoder:
model = model_class(config)
outputs = model(self._prepare_for_class(inputs_dict, model_class))
self.assertEqual(config.output_hidden_states, False)
check_decoder_attentions_output(outputs)
# Check that output attentions can also be changed via the config
del inputs_dict["output_attentions"]
config.output_attentions = True
model = model_class(config)
outputs = model(self._prepare_for_class(inputs_dict, model_class))
self.assertEqual(config.output_hidden_states, False)
check_encoder_attentions_output(outputs)
# Check attention is always last and order is fine
inputs_dict["output_attentions"] = True
config.output_hidden_states = True
model = model_class(config)
outputs = model(self._prepare_for_class(inputs_dict, model_class))
self.assertEqual(out_len + (2 if self.is_encoder_decoder else 1), len(outputs))
self.assertEqual(model.config.output_hidden_states, True)
check_encoder_attentions_output(outputs)
def test_headmasking(self):
if not self.test_head_masking:
return
random.Random().seed(42)
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
random.Random().seed()
inputs_dict["output_attentions"] = True
config.output_hidden_states = True
configs_no_init = _config_zero_init(config) # To be sure we have no Nan
for model_class in self.all_model_classes:
model = model_class(config=configs_no_init)
# Prepare head_mask
def prepare_layer_head_mask(i, attention_heads, num_hidden_layers):
if i == 0:
return tf.concat(
(tf.zeros(1, dtype=tf.float32), tf.ones(attention_heads - 1, dtype=tf.float32)), 0
)
elif i == num_hidden_layers - 1:
return tf.concat(
(tf.zeros(attention_heads - 1, dtype=tf.float32), tf.ones(1, dtype=tf.float32)), 0
)
else:
return tf.ones(attention_heads, dtype=tf.float32)
head_mask = tf.stack(
[
prepare_layer_head_mask(i, config.num_attention_heads, config.num_hidden_layers)
for i in range(config.num_hidden_layers)
],
0,
)
inputs = self._prepare_for_class(inputs_dict, model_class).copy()
inputs["head_mask"] = head_mask
if model.config.is_encoder_decoder:
signature = inspect.signature(model.call)
arg_names = [*signature.parameters.keys()]
if "decoder_head_mask" in arg_names: # necessary diferentiation because of T5 model
inputs["decoder_head_mask"] = head_mask
if "cross_attn_head_mask" in arg_names:
inputs["cross_attn_head_mask"] = head_mask
outputs = model(**inputs, return_dict=True)
def check_attentions_validity(attentions):
# Remove Nan
for t in attentions:
self.assertLess(
(tf.math.reduce_sum(tf.cast(tf.math.is_nan(t), tf.float32))).numpy(), (tf.size(t) / 4).numpy()
) # Check we don't have more than 25% nans (arbitrary)
attentions = [
tf.where(tf.math.is_nan(t), 0.0, t) for t in attentions
] # remove them (the test is less complete)
self.assertAlmostEqual(tf.math.reduce_sum(attentions[0][..., 0, :, :]).numpy(), 0.0)
self.assertNotEqual(tf.math.reduce_sum(attentions[0][..., -1, :, :]).numpy(), 0.0)
if len(attentions) > 2: # encoder-decodere models have only 2 layers in each modules
self.assertNotEqual(tf.math.reduce_sum(attentions[1][..., 0, :, :]).numpy(), 0.0)
self.assertAlmostEqual(tf.math.reduce_sum(attentions[-1][..., -2, :, :]).numpy(), 0.0)
self.assertNotEqual(tf.math.reduce_sum(attentions[-1][..., -1, :, :]).numpy(), 0.0)
if model.config.is_encoder_decoder:
check_attentions_validity(outputs.encoder_attentions)
check_attentions_validity(outputs.decoder_attentions)
if "cross_attn_head_mask" in arg_names:
check_attentions_validity(outputs.cross_attentions)
else:
check_attentions_validity(outputs.attentions)
def test_hidden_states_output(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
def check_hidden_states_output(config, inputs_dict, model_class):
model = model_class(config)
outputs = model(self._prepare_for_class(inputs_dict, model_class))
expected_num_layers = getattr(
self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1
)
if model.config.is_encoder_decoder:
encoder_hidden_states = outputs.encoder_hidden_states
decoder_hidden_states = outputs.decoder_hidden_states
self.assertEqual(config.output_attentions, False)
self.assertEqual(len(encoder_hidden_states), expected_num_layers)
self.assertListEqual(
list(encoder_hidden_states[0].shape[-2:]),
[self.model_tester.seq_length, self.model_tester.hidden_size],
)
self.assertEqual(len(decoder_hidden_states), expected_num_layers)
self.assertListEqual(
list(decoder_hidden_states[0].shape[-2:]),
[self.model_tester.seq_length, self.model_tester.hidden_size],
)
else:
hidden_states = outputs.hidden_states
self.assertEqual(config.output_attentions, False)
self.assertEqual(len(hidden_states), expected_num_layers)
self.assertListEqual(
list(hidden_states[0].shape[-2:]),
[self.model_tester.seq_length, self.model_tester.hidden_size],
)
for model_class in self.all_model_classes:
inputs_dict["output_hidden_states"] = True
check_hidden_states_output(config, inputs_dict, model_class)
del inputs_dict["output_hidden_states"]
config.output_hidden_states = True
check_hidden_states_output(config, inputs_dict, model_class)
def test_model_common_attributes(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
text_in_text_out_models = (
get_values(TF_MODEL_FOR_CAUSAL_LM_MAPPING)
+ get_values(TF_MODEL_FOR_MASKED_LM_MAPPING)
+ get_values(TF_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING)
)
speech_in_text_out_models = get_values(TF_MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING)
for model_class in self.all_model_classes:
model = model_class(config)
assert isinstance(model.get_input_embeddings(), tf.keras.layers.Layer)
if model_class in text_in_text_out_models:
x = model.get_output_embeddings()
assert isinstance(x, tf.keras.layers.Layer)
name = model.get_bias()
assert isinstance(name, dict)
for k, v in name.items():
assert isinstance(v, tf.Variable)
elif model_class in speech_in_text_out_models:
x = model.get_output_embeddings()
assert isinstance(x, tf.keras.layers.Layer)
name = model.get_bias()
assert name is None
else:
x = model.get_output_embeddings()
assert x is None
name = model.get_bias()
assert name is None
def test_determinism(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
first, second = (
model(self._prepare_for_class(inputs_dict, model_class), training=False)[0],
model(self._prepare_for_class(inputs_dict, model_class), training=False)[0],
)
out_1 = first.numpy()
out_2 = second.numpy()
out_1 = out_1[~np.isnan(out_1)]
out_2 = out_2[~np.isnan(out_2)]
max_diff = np.amax(np.abs(out_1 - out_2))
self.assertLessEqual(max_diff, 1e-5)
def test_model_outputs_equivalence(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
def check_equivalence(model, tuple_inputs, dict_inputs, additional_kwargs={}):
tuple_output = model(tuple_inputs, return_dict=False, **additional_kwargs)
dict_output = model(dict_inputs, return_dict=True, **additional_kwargs).to_tuple()
def recursive_check(tuple_object, dict_object):
if isinstance(tuple_object, (List, Tuple)):
for tuple_iterable_value, dict_iterable_value in zip(tuple_object, dict_object):
recursive_check(tuple_iterable_value, dict_iterable_value)
elif tuple_object is None:
return
else:
self.assertTrue(
all(tf.equal(tuple_object, dict_object)),
msg=(
"Tuple and dict output are not equal. Difference:"
f" {tf.math.reduce_max(tf.abs(tuple_object - dict_object))}"
),
)
recursive_check(tuple_output, dict_output)
for model_class in self.all_model_classes:
model = model_class(config)
tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
dict_inputs = self._prepare_for_class(inputs_dict, model_class)
check_equivalence(model, tuple_inputs, dict_inputs)
tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
dict_inputs = self._prepare_for_class(inputs_dict, model_class)
check_equivalence(model, tuple_inputs, dict_inputs, {"output_hidden_states": True})
if self.has_attentions:
tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
dict_inputs = self._prepare_for_class(inputs_dict, model_class)
check_equivalence(model, tuple_inputs, dict_inputs, {"output_attentions": True})
# Not all models accept "labels" in the forward pass (yet :) )
if "labels" in inspect.signature(model.call).parameters.keys():
tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
check_equivalence(model, tuple_inputs, dict_inputs)
tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
check_equivalence(model, tuple_inputs, dict_inputs, {"output_hidden_states": True})
if self.has_attentions:
tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
check_equivalence(model, tuple_inputs, dict_inputs, {"output_attentions": True})
tuple_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
dict_inputs = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
check_equivalence(
model, tuple_inputs, dict_inputs, {"output_hidden_states": True, "output_attentions": True}
)
def test_inputs_embeds(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
inputs = copy.deepcopy(inputs_dict)
if not self.is_encoder_decoder:
input_ids = inputs["input_ids"]
del inputs["input_ids"]
else:
encoder_input_ids = inputs["input_ids"]
decoder_input_ids = inputs.get("decoder_input_ids", encoder_input_ids)
del inputs["input_ids"]
inputs.pop("decoder_input_ids", None)
if not self.is_encoder_decoder:
inputs["inputs_embeds"] = model.get_input_embeddings()(input_ids)
else:
inputs["inputs_embeds"] = model.get_input_embeddings()(encoder_input_ids)
inputs["decoder_inputs_embeds"] = model.get_input_embeddings()(decoder_input_ids)
inputs = self._prepare_for_class(inputs, model_class)
model(inputs)
def test_numpy_arrays_inputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
def prepare_numpy_arrays(inputs_dict):
inputs_np_dict = {}
for k, v in inputs_dict.items():
if tf.is_tensor(v):
inputs_np_dict[k] = v.numpy()
else:
inputs_np_dict[k] = np.array(k)
return inputs_np_dict
for model_class in self.all_model_classes:
model = model_class(config)
inputs = self._prepare_for_class(inputs_dict, model_class)
inputs_np = prepare_numpy_arrays(inputs)
output_for_dict_input = model(inputs_np)
output_for_kw_input = model(**inputs_np)
self.assert_outputs_same(output_for_dict_input, output_for_kw_input)
def test_resize_token_embeddings(self):
# TODO (joao): after the embeddings refactor is complete, rework this test so as to rely exclusively on
# tf.keras.layers.Embedding
if not self.test_resize_embeddings:
return
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
def _get_word_embedding_weight(model, embedding_layer):
if isinstance(embedding_layer, tf.keras.layers.Embedding):
# builds the embeddings layer
model(model.dummy_inputs)
return embedding_layer.embeddings
else:
return model._get_word_embedding_weight(embedding_layer)
for model_class in self.all_model_classes:
for size in [config.vocab_size - 10, config.vocab_size + 10, None]:
# build the embeddings
model = model_class(config=copy.deepcopy(config)) # `resize_token_embeddings` mutates `config`
old_input_embeddings = _get_word_embedding_weight(model, model.get_input_embeddings())
old_bias = model.get_bias()
old_output_embeddings = _get_word_embedding_weight(model, model.get_output_embeddings())
# reshape the embeddings
model.resize_token_embeddings(size)
new_input_embeddings = _get_word_embedding_weight(model, model.get_input_embeddings())
new_bias = model.get_bias()
new_output_embeddings = _get_word_embedding_weight(model, model.get_output_embeddings())
# check that the resized embeddings size matches the desired size.
assert_size = size if size is not None else config.vocab_size
self.assertEqual(new_input_embeddings.shape[0], assert_size)
# check that weights remain the same after resizing
models_equal = True
for p1, p2 in zip(old_input_embeddings.value(), new_input_embeddings.value()):
if tf.math.reduce_sum(tf.math.abs(p1 - p2)) > 0:
models_equal = False
self.assertTrue(models_equal)
if old_bias is not None and new_bias is not None:
for old_weight, new_weight in zip(old_bias.values(), new_bias.values()):
self.assertEqual(new_weight.shape[-1], assert_size)
models_equal = True
for p1, p2 in zip(tf.squeeze(old_weight), tf.squeeze(new_weight)):
if tf.math.reduce_sum(tf.math.abs(p1 - p2)) > 0:
models_equal = False
self.assertTrue(models_equal)
if old_output_embeddings is not None and new_output_embeddings is not None:
self.assertEqual(new_output_embeddings.shape[0], assert_size)
self.assertEqual(new_output_embeddings.shape[1], old_output_embeddings.shape[1])
models_equal = True
for p1, p2 in zip(old_output_embeddings.value(), new_output_embeddings.value()):
if tf.math.reduce_sum(tf.math.abs(p1 - p2)) > 0:
models_equal = False
self.assertTrue(models_equal)
# TODO (Joao): this test is not slow, but it's tagged as such to keep track of failures on the scheduled CI runs,
# while passing push CI. Fix the underlying issues and remove the tag.
@slow
def test_save_load_after_resize_token_embeddings(self):
if not self.test_resize_embeddings:
return
config, original_inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
# create a model with resized (expended) embeddings
new_tokens_size = 10
old_total_size = config.vocab_size
new_total_size = old_total_size + new_tokens_size
model = model_class(config=copy.deepcopy(config)) # `resize_token_embeddings` mutates `config`
model(model.dummy_inputs) # builds the embeddings layer
model.resize_token_embeddings(new_total_size)
# fetch the output for an input exclusively made of new members of the vocabulary
inputs_dict = copy.deepcopy(original_inputs_dict)
ids_feat_name = None
if "input_ids" in inputs_dict:
ids_feat_name = "input_ids"
elif "decoder_input_ids" in inputs_dict:
ids_feat_name = "decoder_input_ids"
else:
assert False, "No input ids feature found in the inputs dict"
new_vocab_input_ids = ids_tensor(inputs_dict[ids_feat_name].shape, new_tokens_size)
new_vocab_input_ids += old_total_size
inputs_dict[ids_feat_name] = new_vocab_input_ids
if "input_ids" in inputs_dict:
inputs_dict["input_ids"] = new_vocab_input_ids
if "decoder_input_ids" in inputs_dict:
inputs_dict["decoder_input_ids"] = new_vocab_input_ids
prepared_inputs = self._prepare_for_class(inputs_dict, model_class)
outputs = model(**prepared_inputs)
# save and load the model
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname, saved_model=False)
model = model_class.from_pretrained(tmpdirname)
restored_model_outputs = model(**prepared_inputs)
# check that the output for the restored model is the same
self.assert_outputs_same(restored_model_outputs, outputs)
@unittest.skipIf(
not is_tf_available() or len(tf.config.list_physical_devices("GPU")) == 0,
reason="This test always passes on CPU.",
)
def test_embeddings_out_of_bounds_raise_exception(self):
# TF embeddings layers don't raise an exception when an index is out of bounds on GPU, so we manually raise it.
# This test should only fail on GPU for models where we haven't added the safety check.
if not self.test_resize_embeddings:
return
config, original_inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config=config)
inputs_dict = copy.deepcopy(original_inputs_dict)
if "input_ids" in inputs_dict:
inputs_dict["input_ids"] = inputs_dict["input_ids"] * int(1e9)
if "decoder_input_ids" in inputs_dict:
inputs_dict["decoder_input_ids"] = inputs_dict["decoder_input_ids"] * int(1e9)
prepared_inputs = self._prepare_for_class(inputs_dict, model_class)
with self.assertRaises(tf.errors.InvalidArgumentError):
model(**prepared_inputs)
def test_lm_head_model_random_no_beam_search_generate(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
input_ids = inputs_dict.get("input_ids", None)
# iterate over all generative models
for model_class in self.all_generative_model_classes:
model = model_class(config)
if config.bos_token_id is None:
# if bos token id is not defined model needs input_ids
with self.assertRaises(ValueError):
model.generate(do_sample=True, max_length=5)
# num_return_sequences = 1
self._check_generated_ids(model.generate(input_ids, do_sample=True))
elif model_class.__name__ not in ["TFSpeech2TextForConditionalGeneration"]:
# Models with non-text inputs won't work here; num_return_sequences = 1
self._check_generated_ids(model.generate(do_sample=True, max_length=5))
with self.assertRaises(ValueError):
# generating multiple sequences when no beam search generation
# is not allowed as it would always generate the same sequences
model.generate(input_ids, do_sample=False, num_return_sequences=2)
# num_return_sequences > 1, sample
self._check_generated_ids(model.generate(input_ids, do_sample=True, num_return_sequences=2))
# check bad words tokens language generation
# create list of 1-seq bad token and list of 2-seq of bad tokens
bad_words_ids = [self._generate_random_bad_tokens(1, model), self._generate_random_bad_tokens(2, model)]
output_tokens = model.generate(
input_ids, do_sample=True, bad_words_ids=bad_words_ids, num_return_sequences=2
)
# only count generated tokens
generated_ids = output_tokens[:, input_ids.shape[-1] :]
self.assertFalse(self._check_match_tokens(generated_ids.numpy().tolist(), bad_words_ids))
def test_lm_head_model_no_beam_search_generate_dict_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
input_ids = inputs_dict.get("input_ids", None)
if input_ids is None:
input_ids = inputs_dict.get("input_features", None)
# iterate over all generative models
for model_class in self.all_generative_model_classes:
model = model_class(config)
output_greedy = model.generate(
input_ids,
do_sample=False,
output_scores=True,
output_hidden_states=True,
output_attentions=True,
return_dict_in_generate=True,
)
output_sample = model.generate(
input_ids,
do_sample=True,
output_scores=True,
output_hidden_states=True,
output_attentions=True,
return_dict_in_generate=True,
)
if model.config.is_encoder_decoder:
self.assertIsInstance(output_greedy, TFGreedySearchEncoderDecoderOutput)
self.assertIsInstance(output_sample, TFSampleEncoderDecoderOutput)
else:
self.assertIsInstance(output_greedy, TFGreedySearchDecoderOnlyOutput)
self.assertIsInstance(output_sample, TFSampleDecoderOnlyOutput)
def test_lm_head_model_random_beam_search_generate(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
input_ids = inputs_dict.get("input_ids", None)
for model_class in self.all_generative_model_classes:
model = model_class(config)
if config.bos_token_id is None:
# if bos token id is not defined model needs input_ids, num_return_sequences = 1
self._check_generated_ids(model.generate(input_ids, do_sample=True, num_beams=2))
else:
# num_return_sequences = 1
self._check_generated_ids(model.generate(do_sample=True, max_length=5, num_beams=2))
with self.assertRaises(ValueError):
# generating more sequences than having beams leads is not possible
model.generate(input_ids, do_sample=False, num_return_sequences=3, num_beams=2)
# num_return_sequences > 1, sample
self._check_generated_ids(
model.generate(
input_ids,
do_sample=True,
num_beams=2,
num_return_sequences=2,
)
)
# num_return_sequences > 1, greedy
self._check_generated_ids(model.generate(input_ids, do_sample=False, num_beams=2, num_return_sequences=2))
# check bad words tokens language generation
# create list of 1-seq bad token and list of 2-seq of bad tokens
bad_words_ids = [self._generate_random_bad_tokens(1, model), self._generate_random_bad_tokens(2, model)]
output_tokens = model.generate(
input_ids, do_sample=False, bad_words_ids=bad_words_ids, num_beams=2, num_return_sequences=2
)
# only count generated tokens
generated_ids = output_tokens[:, input_ids.shape[-1] :]
self.assertFalse(self._check_match_tokens(generated_ids.numpy().tolist(), bad_words_ids))
def test_lm_head_model_beam_search_generate_dict_outputs(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
input_ids = inputs_dict.get("input_ids", None)
if input_ids is None:
input_ids = inputs_dict.get("input_features", None)
# iterate over all generative models
for model_class in self.all_generative_model_classes:
model = model_class(config)
output_beam_search = model.generate(
input_ids,
num_beams=2,
do_sample=False,
output_scores=True,
output_hidden_states=True,
output_attentions=True,
return_dict_in_generate=True,
)
output_beam_sample = model.generate(
input_ids,
num_beams=2,
do_sample=True,
output_scores=True,
output_hidden_states=True,
output_attentions=True,
return_dict_in_generate=True,
)
if model.config.is_encoder_decoder:
self.assertIsInstance(output_beam_search, TFBeamSearchEncoderDecoderOutput)
self.assertIsInstance(output_beam_sample, TFBeamSampleEncoderDecoderOutput)
else:
self.assertIsInstance(output_beam_search, TFBeamSearchDecoderOnlyOutput)
self.assertIsInstance(output_beam_sample, TFBeamSampleDecoderOnlyOutput)
def test_loss_computation(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
if not getattr(model, "hf_compute_loss", None) and not _return_type_has_loss(model):
continue
# The number of elements in the loss should be the same as the number of elements in the label
prepared_for_class = self._prepare_for_class(inputs_dict.copy(), model_class, return_labels=True)
added_label_names = sorted(prepared_for_class.keys() - inputs_dict.keys(), reverse=True)
if not added_label_names:
continue # This test is only for models with easily-separable labels
added_label = prepared_for_class[added_label_names[0]]
expected_loss_size = added_label.shape.as_list()[:1]
# Test that model correctly compute the loss with kwargs
prepared_for_class = self._prepare_for_class(inputs_dict.copy(), model_class, return_labels=True)
possible_input_names = {"input_ids", "pixel_values", "input_features", "input_values"}
input_name = possible_input_names.intersection(set(prepared_for_class)).pop()
model_input = prepared_for_class.pop(input_name)
loss = model(model_input, **prepared_for_class)[0]
self.assertTrue(loss.shape.as_list() == expected_loss_size or loss.shape.as_list() == [1])
# Test that model correctly compute the loss when we mask some positions
prepared_for_class = self._prepare_for_class(inputs_dict.copy(), model_class, return_labels=True)
possible_input_names = {"input_ids", "pixel_values", "input_features", "input_values"}
input_name = possible_input_names.intersection(set(prepared_for_class)).pop()
model_input = prepared_for_class.pop(input_name)
if "labels" in prepared_for_class:
labels = prepared_for_class["labels"].numpy()
if len(labels.shape) > 1 and labels.shape[1] != 1:
labels[0] = -100
prepared_for_class["labels"] = tf.convert_to_tensor(labels)
loss = model(model_input, **prepared_for_class)[0]
self.assertTrue(loss.shape.as_list() == expected_loss_size or loss.shape.as_list() == [1])
self.assertTrue(not np.any(np.isnan(loss.numpy())))
# Test that model correctly compute the loss with a dict
prepared_for_class = self._prepare_for_class(inputs_dict.copy(), model_class, return_labels=True)
loss = model(prepared_for_class)[0]
self.assertTrue(loss.shape.as_list() == expected_loss_size or loss.shape.as_list() == [1])
# Test that model correctly compute the loss with a tuple
prepared_for_class = self._prepare_for_class(inputs_dict.copy(), model_class, return_labels=True)
# Get keys that were added with the _prepare_for_class function
label_keys = prepared_for_class.keys() - inputs_dict.keys()
signature = inspect.signature(model.call).parameters
signature_names = list(signature.keys())
# Create a dictionary holding the location of the tensors in the tuple
tuple_index_mapping = {0: input_name}
for label_key in label_keys:
label_key_index = signature_names.index(label_key)
tuple_index_mapping[label_key_index] = label_key
sorted_tuple_index_mapping = sorted(tuple_index_mapping.items())
# Initialize a list with their default values, update the values and convert to a tuple
list_input = []
for name in signature_names:
if name != "kwargs":
list_input.append(signature[name].default)
for index, value in sorted_tuple_index_mapping:
list_input[index] = prepared_for_class[value]
tuple_input = tuple(list_input)
# Send to model
loss = model(tuple_input[:-1])[0]
self.assertTrue(loss.shape.as_list() == expected_loss_size or loss.shape.as_list() == [1])
def check_keras_fit_results(self, val_loss1, val_loss2, atol=1e-2, rtol=1e-3):
self.assertTrue(np.allclose(val_loss1, val_loss2, atol=atol, rtol=rtol))
def test_keras_fit(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
if not getattr(model, "hf_compute_loss", False) and not _return_type_has_loss(model):
continue
# Test that model correctly compute the loss with kwargs
prepared_for_class = self._prepare_for_class(inputs_dict.copy(), model_class, return_labels=True)
# Is there a better way to remove these decoder inputs?
# We also remove "return_loss" as this is covered by the train_step when using fit()
prepared_for_class = {
key: val
for key, val in prepared_for_class.items()
if key
not in ("head_mask", "decoder_head_mask", "cross_attn_head_mask", "decoder_input_ids", "return_loss")
}
accuracy_classes = [
"ForPreTraining",
"ForCausalLM",
"ForMaskedLM",
"ForQuestionAnswering",
"ForMultipleChoice",
"ForSequenceClassification",
"ForTokenClassification",
"ForNextSentencePrediction",
"LMHeadModel",
]
for accuracy_class in accuracy_classes:
if model.__class__.__name__.endswith(accuracy_class):
metrics = [tf.keras.metrics.SparseCategoricalAccuracy()]
break
else:
metrics = []
if hasattr(self.model_tester, "batch_size"):
sample_weight = tf.convert_to_tensor([0.5] * self.model_tester.batch_size, dtype=tf.float32)
else:
sample_weight = None
model(model.dummy_inputs) # Build the model so we can get some constant weights
model_weights = model.get_weights()
# Run eagerly to save some expensive compilation times
model.compile(optimizer=tf.keras.optimizers.SGD(0.0), run_eagerly=True, metrics=metrics)
# Make sure the model fits without crashing regardless of where we pass the labels
history1 = model.fit(
prepared_for_class,
validation_data=prepared_for_class,
sample_weight=sample_weight,
steps_per_epoch=1,
validation_steps=1,
shuffle=False,
)
val_loss1 = history1.history["val_loss"][0]
self.assertTrue(not isnan(val_loss1))
accuracy1 = {key: val[0] for key, val in history1.history.items() if key.endswith("accuracy")}
possible_label_cols = {
"labels",
"label",
"label_ids",
"start_positions",
"start_position",
"end_positions",
"end_position",
"next_sentence_label",
}
label_names = possible_label_cols.intersection(set(prepared_for_class))
if len(label_names) == 0:
# The next tests only make sense for models with separate inputs and labels, and do not make
# sense for models that don't clearly distinguish between the two (e.g. CLIP)
return
labels = {key: val for key, val in prepared_for_class.items() if key in label_names}
inputs_minus_labels = {key: val for key, val in prepared_for_class.items() if key not in label_names}
self.assertGreater(len(inputs_minus_labels), 0)
# We reinitialize the model here even though our learning rate was zero
# because BatchNorm updates weights by means other than gradient descent.
model.set_weights(model_weights)
history2 = model.fit(
inputs_minus_labels,
labels,
validation_data=(inputs_minus_labels, labels),
sample_weight=sample_weight,
steps_per_epoch=1,
validation_steps=1,
shuffle=False,
)
val_loss2 = history2.history["val_loss"][0]
self.assertTrue(not isnan(val_loss2))
accuracy2 = {key: val[0] for key, val in history2.history.items() if key.endswith("accuracy")}
self.check_keras_fit_results(val_loss1, val_loss2)
self.assertEqual(history1.history.keys(), history2.history.keys())
for key in history1.history.keys():
if not key.startswith("val_"):
self.assertTrue("val_" + key in history1.history.keys(), "Outputs differ in train/test step!")
if metrics:
self.assertTrue(len(accuracy1) == len(accuracy2) > 0, "Missing metrics!")
# Make sure fit works with tf.data.Dataset and results are consistent
dataset = tf.data.Dataset.from_tensor_slices(prepared_for_class)
if sample_weight is not None:
# Add in the sample weight
weighted_dataset = dataset.map(lambda x: (x, None, tf.convert_to_tensor(0.5, dtype=tf.float32)))
else:
weighted_dataset = dataset
# Pass in all samples as a batch to match other `fit` calls
weighted_dataset = weighted_dataset.batch(len(dataset))
dataset = dataset.batch(len(dataset))
# Reinitialize to fix batchnorm again
model.set_weights(model_weights)
# To match the other calls, don't pass sample weights in the validation data
history3 = model.fit(
weighted_dataset,
validation_data=dataset,
steps_per_epoch=1,
validation_steps=1,
shuffle=False,
)
val_loss3 = history3.history["val_loss"][0]
self.assertTrue(not isnan(val_loss3))
accuracy3 = {key: val[0] for key, val in history3.history.items() if key.endswith("accuracy")}
self.check_keras_fit_results(val_loss1, val_loss3)
self.assertEqual(history1.history.keys(), history3.history.keys())
if metrics:
self.assertTrue(len(accuracy1) == len(accuracy3) > 0, "Missing metrics!")
def test_int_support(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
prepared_for_class = self._prepare_for_class(
inputs_dict.copy(),
model_class,
return_labels=True if "labels" in inspect.signature(model_class.call).parameters.keys() else False,
)
if not any(
[tensor.dtype.is_integer for tensor in prepared_for_class.values() if isinstance(tensor, tf.Tensor)]
):
return # No integer inputs means no need for this test
prepared_for_class = {
key: tf.cast(tensor, tf.int64) if isinstance(tensor, tf.Tensor) and tensor.dtype.is_integer else tensor
for key, tensor in prepared_for_class.items()
}
model = model_class(config)
model(**prepared_for_class) # No assertion, we're just checking this doesn't throw an error
int32_prepared_for_class = {
key: tf.cast(tensor, tf.int32) if isinstance(tensor, tf.Tensor) and tensor.dtype.is_integer else tensor
for key, tensor in prepared_for_class.items()
}
model(**int32_prepared_for_class) # No assertion, we're just checking this doesn't throw an error
# After testing that the model accepts all int inputs, confirm that its dummies are int32
for key, tensor in model.dummy_inputs.items():
self.assertTrue(isinstance(tensor, tf.Tensor), "Dummy inputs should be tf.Tensor!")
if tensor.dtype.is_integer:
self.assertTrue(tensor.dtype == tf.int32, "Integer dummy inputs should be tf.int32!")
# Also confirm that the serving sig uses int32
if hasattr(model, "serving"):
serving_sig = model.serving.input_signature
for key, tensor_spec in serving_sig[0].items():
if tensor_spec.dtype.is_integer:
self.assertTrue(
tensor_spec.dtype == tf.int32, "Serving signatures should use tf.int32 for ints!"
)
def test_generate_with_headmasking(self):
attention_names = ["encoder_attentions", "decoder_attentions", "cross_attentions"]
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_generative_model_classes:
model = model_class(config)
# We want to test only encoder-decoder models
if not config.is_encoder_decoder:
continue
head_masking = {
"head_mask": tf.zeros((config.encoder_layers, config.encoder_attention_heads)),
"decoder_head_mask": tf.zeros((config.decoder_layers, config.decoder_attention_heads)),
"cross_attn_head_mask": tf.zeros((config.decoder_layers, config.decoder_attention_heads)),
}
signature = inspect.signature(model.call)
if set(head_masking.keys()) < {*signature.parameters.keys()}:
continue
for attn_name, (name, mask) in zip(attention_names, head_masking.items()):
out = model.generate(
inputs_dict["input_ids"],
num_beams=1,
max_length=inputs_dict["input_ids"] + 5,
output_attentions=True,
return_dict_in_generate=True,
**{name: mask},
)
# We check the state of decoder_attentions and cross_attentions just from the last step
attn_weights = out[attn_name] if attn_name == attention_names[0] else out[attn_name][-1]
self.assertEqual(sum([tf.reduce_sum(w).numpy() for w in attn_weights]), 0.0)
def test_load_with_mismatched_shapes(self):
if not self.test_mismatched_shapes:
return
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
if model_class not in get_values(TF_MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING):
continue
with self.subTest(msg=f"Testing {model_class}"):
with tempfile.TemporaryDirectory() as tmp_dir:
model = model_class(config)
inputs = self._prepare_for_class(inputs_dict, model_class)
_ = model(**inputs)
model.save_pretrained(tmp_dir)
# Fails when we don't set ignore_mismatched_sizes=True
with self.assertRaises(ValueError):
new_model = TFAutoModelForSequenceClassification.from_pretrained(tmp_dir, num_labels=42)
with self.assertRaises(ValueError):
new_model_without_prefix = TFAutoModel.from_pretrained(tmp_dir, vocab_size=10)
logger = logging.get_logger("transformers.modeling_tf_utils")
with CaptureLogger(logger) as cl:
new_model = TFAutoModelForSequenceClassification.from_pretrained(
tmp_dir, num_labels=42, ignore_mismatched_sizes=True
)
self.assertIn("the shapes did not match", cl.out)
logits = new_model(**inputs).logits
self.assertEqual(logits.shape[1], 42)
with CaptureLogger(logger) as cl:
new_model_without_prefix = TFAutoModel.from_pretrained(
tmp_dir, vocab_size=10, ignore_mismatched_sizes=True
)
self.assertIn("the shapes did not match", cl.out)
# Although Tf models always have a prefix pointing to `MainLayer`,
# we still add this "without prefix" test to keep a consistency between tf and pt tests.
input_ids = ids_tensor((2, 8), 10)
if self.is_encoder_decoder:
new_model_without_prefix(input_ids, decoder_input_ids=input_ids)
else:
new_model_without_prefix(input_ids)
def test_model_main_input_name(self):
for model_class in self.all_model_classes:
model_signature = inspect.signature(getattr(model_class, "call"))
# The main input is the name of the argument after `self`
observed_main_input_name = list(model_signature.parameters.keys())[1]
self.assertEqual(model_class.main_input_name, observed_main_input_name)
def test_dataset_conversion(self):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
for model_class in self.all_model_classes:
model = model_class(config)
tf_inputs_dict = self._prepare_for_class(inputs_dict, model_class, return_labels=False)
if "labels" in tf_inputs_dict:
return # This is some kinda funky decoder model that needs labels in its forward pass
tf_inputs_dict = {
key: val
for key, val in tf_inputs_dict.items()
if "head_mask" not in key and isinstance(val, tf.Tensor)
}
tf_inputs_dict["extra_unwanted_column"] = list(tf_inputs_dict.values())[0] # Use a random other tensor
input_dataset = Dataset.from_dict(tf_inputs_dict)
tf_dataset = model.prepare_tf_dataset(
input_dataset, batch_size=len(input_dataset), drop_remainder=False, shuffle=False
)
test_batch = next(iter(tf_dataset))
if isinstance(test_batch, tf.Tensor):
self.assertEqual(len(test_batch), len(input_dataset)) # Assert we didn't lose any data
elif isinstance(test_batch, dict):
# Assert we discarded the unwanted extra column but kept everything else
self.assertEqual(len(test_batch), len(input_dataset.features) - 1)
self.assertNotIn("extra_unwanted_column", test_batch)
for tensor in test_batch.values():
self.assertTrue(isinstance(tensor, tf.Tensor))
self.assertEqual(len(tensor), len(input_dataset)) # Assert we didn't lose any data
model(test_batch, training=False)
if "labels" in inspect.signature(model_class.call).parameters.keys():
tf_inputs_dict = self._prepare_for_class(inputs_dict, model_class, return_labels=True)
if "labels" not in tf_inputs_dict:
return # This model isn't giving us labels after all, don't try training with it
tf_inputs_dict = {key: val for key, val in tf_inputs_dict.items() if "head_mask" not in key}
tf_inputs_dict["extra_unwanted_column"] = list(tf_inputs_dict.values())[0] # Use a random other tensor
input_dataset = Dataset.from_dict(tf_inputs_dict)
tf_dataset = model.prepare_tf_dataset(
input_dataset, batch_size=len(input_dataset), drop_remainder=False, shuffle=False
)
test_batch, test_batch_labels = next(iter(tf_dataset))
self.assertGreater(len(test_batch_labels), 0) # Assert the labels are present
feature_columns = 1 if isinstance(test_batch, tf.Tensor) else len(test_batch)
label_columns = 1 if isinstance(test_batch_labels, tf.Tensor) else len(test_batch_labels)
# Assert we discarded the unwanted extra column but kept everything else
self.assertEqual(feature_columns + label_columns, len(input_dataset.features) - 1)
if isinstance(test_batch, dict):
self.assertNotIn("extra_unwanted_column", test_batch)
if isinstance(test_batch_labels, dict):
self.assertNotIn("extra_unwanted_column", test_batch_labels)
model.compile(optimizer="sgd", run_eagerly=True)
model.train_on_batch(test_batch, test_batch_labels)
def _test_xla_generate(self, **generate_kwargs):
def _generate_and_check_results(model, inputs_dict):
if "input_ids" in inputs_dict:
inputs = inputs_dict["input_ids"]
# make sure there are no pad tokens in prompt, which may trigger unwanted behavior
if model.generation_config.pad_token_id is not None:
if config.pad_token_id == 0:
new_pad_token = model.generation_config.pad_token_id + 1
else:
new_pad_token = model.generation_config.pad_token_id - 1
else:
new_pad_token = None
inputs = tf.where(inputs != model.generation_config.pad_token_id, inputs, new_pad_token)
elif "input_features" in inputs_dict:
inputs = inputs_dict["input_features"]
else:
raise ValueError("No valid generate input found in inputs_dict")
generated = model.generate(inputs, **generate_kwargs).numpy()
generate_xla = tf.function(model.generate, jit_compile=True)
generated_xla = generate_xla(inputs, **generate_kwargs).numpy()
self.assertListEqual(generated.tolist(), generated_xla.tolist())
for model_class in self.all_generative_model_classes:
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.eos_token_id = None # Generate until max length
config.do_sample = False
# fix config for models with additional sequence-length limiting settings
for var_name in ["max_position_embeddings", "max_target_positions"]:
attr = getattr(config, var_name, None)
if attr is not None and attr < generate_kwargs["max_new_tokens"]:
try:
setattr(config, var_name, generate_kwargs["max_new_tokens"])
except NotImplementedError:
# xlnet will raise an exception when trying to set
# max_position_embeddings.
pass
model = model_class(config)
if model.supports_xla_generation:
_generate_and_check_results(model, inputs_dict)
else:
with self.assertRaises(ValueError):
_generate_and_check_results(model, inputs_dict)
def test_xla_generate_fast(self):
"""
Basic quick test for generate-compatible classes that confirms that XLA-generated tokens are the same as their
non XLA counterparts.
Either the model supports XLA generation and passes the inner test, or it raises an appropriate exception
"""
self._test_xla_generate(num_beams=1, num_return_sequences=1, max_new_tokens=3)
@slow
def test_xla_generate_contrastive(self):
"""
Slow and challenging version of `test_xla_generate_fast` for contrastive search -- contrastive search directly
manipulates the model cache and other outputs, and this test ensures that they are in a valid format that is
also supported by XLA.
Either the model supports XLA generation and passes the inner test, or it raises an appropriate exception
"""
self._test_xla_generate(num_beams=1, num_return_sequences=1, max_new_tokens=16, penalty_alpha=0.5, top_k=4)
@slow
def test_xla_generate_slow(self):
"""
Slow and challenging version of `test_xla_generate_fast` -- this test asks for several long sequences using
beam search, with and without XLA. The two outputs should match, and a failure in this test indicates that the
model may need further analysis if it is to be used for XLA generation.
Either the model supports XLA generation and passes the inner test, or it raises an appropriate exception
"""
self._test_xla_generate(num_beams=8, num_return_sequences=2, max_new_tokens=128)
def _generate_random_bad_tokens(self, num_bad_tokens, model):
# special tokens cannot be bad tokens
special_tokens = []
if model.config.bos_token_id is not None:
special_tokens.append(model.config.bos_token_id)
if model.config.pad_token_id is not None:
special_tokens.append(model.config.pad_token_id)
if model.config.eos_token_id is not None:
special_tokens.append(model.config.eos_token_id)
# create random bad tokens that are not special tokens
bad_tokens = []
while len(bad_tokens) < num_bad_tokens:
token = tf.squeeze(ids_tensor((1, 1), self.model_tester.vocab_size), 0).numpy()[0]
if token not in special_tokens:
bad_tokens.append(token)
return bad_tokens
def _check_generated_ids(self, output_ids):
for token_id in output_ids[0].numpy().tolist():
self.assertGreaterEqual(token_id, 0)
self.assertLess(token_id, self.model_tester.vocab_size)
def _check_match_tokens(self, generated_ids, bad_words_ids):
# for all bad word tokens
for bad_word_ids in bad_words_ids:
# for all slices in batch
for generated_ids_slice in generated_ids:
# for all word idx
for i in range(len(bad_word_ids), len(generated_ids_slice)):
# if tokens match
if generated_ids_slice[i - len(bad_word_ids) : i] == bad_word_ids:
return True
return False
def ids_tensor(shape, vocab_size, rng=None, name=None, dtype=None):
"""Creates a random int32 tensor of the shape within the vocab size."""
if rng is None:
rng = random.Random()
total_dims = 1
for dim in shape:
total_dims *= dim
values = []
for _ in range(total_dims):
values.append(rng.randint(0, vocab_size - 1))
output = tf.constant(values, shape=shape, dtype=dtype if dtype is not None else tf.int32)
return output
def random_attention_mask(shape, rng=None, name=None, dtype=None):
attn_mask = ids_tensor(shape, vocab_size=2, rng=None, name=None, dtype=dtype)
# make sure that at least one token is attended to for each batch
attn_mask = tf.concat([attn_mask[:, :-1], tf.ones_like(attn_mask[:, -1:], dtype=dtype)], axis=-1)
return attn_mask
def floats_tensor(shape, scale=1.0, rng=None, name=None, dtype=None):
"""Creates a random float32 tensor"""
if rng is None:
rng = random.Random()
total_dims = 1
for dim in shape:
total_dims *= dim
values = []
for _ in range(total_dims):
values.append(rng.random() * scale)
return tf.reshape(tf.constant(values, dtype=dtype if dtype is not None else tf.float32), shape=shape)
@require_tf
class UtilsFunctionsTest(unittest.TestCase):
def test_cached_files_are_used_when_internet_is_down(self):
# A mock response for an HTTP head request to emulate server down
response_mock = mock.Mock()
response_mock.status_code = 500
response_mock.headers = {}
response_mock.raise_for_status.side_effect = HTTPError
response_mock.json.return_value = {}
# Download this model to make sure it's in the cache.
_ = TFBertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
# Under the mock environment we get a 500 error when trying to reach the model.
with mock.patch("requests.request", return_value=response_mock) as mock_head:
_ = TFBertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
# This check we did call the fake head request
mock_head.assert_called()
def test_load_from_one_file(self):
try:
tmp_file = tempfile.mktemp()
with open(tmp_file, "wb") as f:
http_get("https://huggingface.co/hf-internal-testing/tiny-random-bert/resolve/main/tf_model.h5", f)
config = BertConfig.from_pretrained("hf-internal-testing/tiny-random-bert")
_ = TFBertModel.from_pretrained(tmp_file, config=config)
finally:
os.remove(tmp_file)
def test_legacy_load_from_url(self):
# This test is for deprecated behavior and can be removed in v5
config = BertConfig.from_pretrained("hf-internal-testing/tiny-random-bert")
_ = TFBertModel.from_pretrained(
"https://huggingface.co/hf-internal-testing/tiny-random-bert/resolve/main/tf_model.h5", config=config
)
# tests whether the unpack_inputs function behaves as expected
def test_unpack_inputs(self):
class DummyModel:
def __init__(self):
config_kwargs = {"output_attentions": False, "output_hidden_states": False, "return_dict": False}
self.config = PretrainedConfig(**config_kwargs)
self.main_input_name = "input_ids"
@unpack_inputs
def call(
self,
input_ids=None,
past_key_values=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
return input_ids, past_key_values, output_attentions, output_hidden_states, return_dict
@unpack_inputs
def foo(self, pixel_values, output_attentions=None, output_hidden_states=None, return_dict=None):
return pixel_values, output_attentions, output_hidden_states, return_dict
dummy_model = DummyModel()
input_ids = tf.constant([0, 1, 2, 3], dtype=tf.int32)
past_key_values = tf.constant([4, 5, 6, 7], dtype=tf.int32)
pixel_values = tf.constant([8, 9, 10, 11], dtype=tf.int32)
# test case 1: Pass inputs as keyword arguments; Booleans are inherited from the config.
output = dummy_model.call(input_ids=input_ids, past_key_values=past_key_values)
tf.debugging.assert_equal(output[0], input_ids)
tf.debugging.assert_equal(output[1], past_key_values)
self.assertFalse(output[2])
self.assertFalse(output[3])
self.assertFalse(output[4])
# test case 2: Same as above, but with positional arguments.
output = dummy_model.call(input_ids, past_key_values)
tf.debugging.assert_equal(output[0], input_ids)
tf.debugging.assert_equal(output[1], past_key_values)
self.assertFalse(output[2])
self.assertFalse(output[3])
self.assertFalse(output[4])
# test case 3: We can also pack everything in the first input.
output = dummy_model.call(input_ids={"input_ids": input_ids, "past_key_values": past_key_values})
tf.debugging.assert_equal(output[0], input_ids)
tf.debugging.assert_equal(output[1], past_key_values)
self.assertFalse(output[2])
self.assertFalse(output[3])
self.assertFalse(output[4])
# test case 4: Explicit boolean arguments should override the config.
output = dummy_model.call(
input_ids=input_ids, past_key_values=past_key_values, output_attentions=False, return_dict=True
)
tf.debugging.assert_equal(output[0], input_ids)
tf.debugging.assert_equal(output[1], past_key_values)
self.assertFalse(output[2])
self.assertFalse(output[3])
self.assertTrue(output[4])
# test case 5: Unexpected arguments should raise an exception.
with self.assertRaises(ValueError):
output = dummy_model.call(input_ids=input_ids, past_key_values=past_key_values, foo="bar")
# test case 6: the decorator is independent from `main_input_name` -- it treats the first argument of the
# decorated function as its main input.
output = dummy_model.foo(pixel_values=pixel_values)
tf.debugging.assert_equal(output[0], pixel_values)
self.assertFalse(output[1])
self.assertFalse(output[2])
self.assertFalse(output[3])
# Tests whether the stable softmax is stable on CPU, with and without XLA
def test_xla_stable_softmax(self):
large_penalty = -1e9
n_tokens = 10
batch_size = 8
def masked_softmax(x, boolean_mask):
numerical_mask = (1.0 - tf.cast(boolean_mask, dtype=tf.float32)) * large_penalty
masked_x = x + numerical_mask
return stable_softmax(masked_x)
xla_masked_softmax = tf.function(masked_softmax, jit_compile=True)
xla_stable_softmax = tf.function(stable_softmax, jit_compile=True)
x = tf.random.normal((batch_size, n_tokens))
# Same outcome regardless of the boolean mask here
masked_tokens = random.randint(0, n_tokens)
boolean_mask = tf.convert_to_tensor([[1] * (n_tokens - masked_tokens) + [0] * masked_tokens], dtype=tf.int32)
# We can randomly mask a random numerical input OUTSIDE XLA
numerical_mask = (1.0 - tf.cast(boolean_mask, dtype=tf.float32)) * large_penalty
masked_x = x + numerical_mask
xla_out = xla_stable_softmax(masked_x)
out = stable_softmax(masked_x)
assert tf.experimental.numpy.allclose(xla_out, out)
# The stable softmax has the same output as the original softmax
unstable_out = tf.nn.softmax(masked_x)
assert tf.experimental.numpy.allclose(unstable_out, out)
# We can randomly mask a random numerical input INSIDE XLA
xla_out = xla_masked_softmax(x, boolean_mask)
out = masked_softmax(x, boolean_mask)
assert tf.experimental.numpy.allclose(xla_out, out)
def test_checkpoint_sharding_from_hub(self):
model = TFBertModel.from_pretrained("ArthurZ/tiny-random-bert-sharded")
# the model above is the same as the model below, just a sharded version.
ref_model = TFBertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
for p1, p2 in zip(model.weights, ref_model.weights):
assert np.allclose(p1.numpy(), p2.numpy())
def test_sharded_checkpoint_with_prefix(self):
model = TFBertModel.from_pretrained("hf-internal-testing/tiny-random-bert", load_weight_prefix="a/b")
sharded_model = TFBertModel.from_pretrained("ArthurZ/tiny-random-bert-sharded", load_weight_prefix="a/b")
for p1, p2 in zip(model.weights, sharded_model.weights):
self.assertTrue(np.allclose(p1.numpy(), p2.numpy()))
self.assertTrue(p1.name.startswith("a/b/"))
self.assertTrue(p2.name.startswith("a/b/"))
def test_sharded_checkpoint_transfer(self):
# If this doesn't throw an error then the test passes
TFBertForSequenceClassification.from_pretrained("ArthurZ/tiny-random-bert-sharded")
@is_pt_tf_cross_test
def test_checkpoint_sharding_local_from_pt(self):
with tempfile.TemporaryDirectory() as tmp_dir:
_ = Repository(local_dir=tmp_dir, clone_from="hf-internal-testing/tiny-random-bert-sharded")
model = TFBertModel.from_pretrained(tmp_dir, from_pt=True)
# the model above is the same as the model below, just a sharded pytorch version.
ref_model = TFBertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
for p1, p2 in zip(model.weights, ref_model.weights):
assert np.allclose(p1.numpy(), p2.numpy())
@is_pt_tf_cross_test
def test_checkpoint_loading_with_prefix_from_pt(self):
model = TFBertModel.from_pretrained(
"hf-internal-testing/tiny-random-bert", from_pt=True, load_weight_prefix="a/b"
)
ref_model = TFBertModel.from_pretrained("hf-internal-testing/tiny-random-bert", from_pt=True)
for p1, p2 in zip(model.weights, ref_model.weights):
self.assertTrue(np.allclose(p1.numpy(), p2.numpy()))
self.assertTrue(p1.name.startswith("a/b/"))
@is_pt_tf_cross_test
def test_checkpoint_sharding_hub_from_pt(self):
model = TFBertModel.from_pretrained("hf-internal-testing/tiny-random-bert-sharded", from_pt=True)
# the model above is the same as the model below, just a sharded pytorch version.
ref_model = TFBertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
for p1, p2 in zip(model.weights, ref_model.weights):
assert np.allclose(p1.numpy(), p2.numpy())
def test_shard_checkpoint(self):
# This is the model we will use, total size 340,000 bytes.
model = tf.keras.Sequential(
[
tf.keras.layers.Dense(200, use_bias=False), # size 80,000
tf.keras.layers.Dense(200, use_bias=False), # size 160,000
tf.keras.layers.Dense(100, use_bias=False), # size 80,000
tf.keras.layers.Dense(50, use_bias=False), # size 20,000
]
)
inputs = tf.zeros((1, 100), dtype=tf.float32)
model(inputs)
weights = model.weights
weights_dict = {w.name: w for w in weights}
with self.subTest("No shard when max size is bigger than model size"):
shards, index = tf_shard_checkpoint(weights)
self.assertIsNone(index)
self.assertDictEqual(shards, {TF2_WEIGHTS_NAME: weights})
with self.subTest("Test sharding, no weights bigger than max size"):
shards, index = tf_shard_checkpoint(weights, max_shard_size="300kB")
# Split is first two layers then last two.
self.assertDictEqual(
index,
{
"metadata": {"total_size": 340000},
"weight_map": {
"dense/kernel:0": "tf_model-00001-of-00002.h5",
"dense_1/kernel:0": "tf_model-00001-of-00002.h5",
"dense_2/kernel:0": "tf_model-00002-of-00002.h5",
"dense_3/kernel:0": "tf_model-00002-of-00002.h5",
},
},
)
shard1 = [weights_dict["dense/kernel:0"], weights_dict["dense_1/kernel:0"]]
shard2 = [weights_dict["dense_2/kernel:0"], weights_dict["dense_3/kernel:0"]]
self.assertDictEqual(shards, {"tf_model-00001-of-00002.h5": shard1, "tf_model-00002-of-00002.h5": shard2})
with self.subTest("Test sharding with weights bigger than max size"):
shards, index = tf_shard_checkpoint(weights, max_shard_size="100kB")
# Split is first layer, second layer then last 2.
self.assertDictEqual(
index,
{
"metadata": {"total_size": 340000},
"weight_map": {
"dense/kernel:0": "tf_model-00001-of-00003.h5",
"dense_1/kernel:0": "tf_model-00002-of-00003.h5",
"dense_2/kernel:0": "tf_model-00003-of-00003.h5",
"dense_3/kernel:0": "tf_model-00003-of-00003.h5",
},
},
)
shard1 = [weights_dict["dense/kernel:0"]]
shard2 = [weights_dict["dense_1/kernel:0"]]
shard3 = [weights_dict["dense_2/kernel:0"], weights_dict["dense_3/kernel:0"]]
self.assertDictEqual(
shards,
{
"tf_model-00001-of-00003.h5": shard1,
"tf_model-00002-of-00003.h5": shard2,
"tf_model-00003-of-00003.h5": shard3,
},
)
@slow
def test_special_layer_name_sharding(self):
retriever = RagRetriever.from_pretrained("facebook/rag-token-nq", index_name="exact", use_dummy_dataset=True)
model = TFRagModel.from_pretrained("facebook/rag-token-nq", retriever=retriever)
with tempfile.TemporaryDirectory() as tmp_dir:
for max_size in ["150kB", "150kiB", "200kB", "200kiB"]:
model.save_pretrained(tmp_dir, max_shard_size=max_size)
ref_model = TFRagModel.from_pretrained(tmp_dir, retriever=retriever)
for p1, p2 in zip(model.weights, ref_model.weights):
assert np.allclose(p1.numpy(), p2.numpy())
def test_checkpoint_sharding_local(self):
model = TFBertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
with tempfile.TemporaryDirectory() as tmp_dir:
# We use the same folder for various sizes to make sure a new save erases the old checkpoint.
for max_size in ["150kB", "150kiB", "200kB", "200kiB"]:
model.save_pretrained(tmp_dir, max_shard_size=max_size)
# Get each shard file and its size
shard_to_size = {}
for shard in os.listdir(tmp_dir):
if shard.endswith(".h5"):
shard_file = os.path.join(tmp_dir, shard)
shard_to_size[shard_file] = os.path.getsize(shard_file)
index_file = os.path.join(tmp_dir, TF2_WEIGHTS_INDEX_NAME)
# Check there is an index but no regular weight file
self.assertTrue(os.path.isfile(index_file))
self.assertFalse(os.path.isfile(os.path.join(tmp_dir, TF2_WEIGHTS_NAME)))
# Check a file is bigger than max_size only when it has a single weight
for shard_file, size in shard_to_size.items():
if max_size.endswith("kiB"):
max_size_int = int(max_size[:-3]) * 2**10
else:
max_size_int = int(max_size[:-2]) * 10**3
# Note: pickle adds some junk so the weight of the file can end up being slightly bigger than
# the size asked for (since we count parameters)
if size >= max_size_int + 50000:
with h5py.File(shard_file, "r") as state_file:
self.assertEqual(len(state_file), 1)
# Check the index and the shard files found match
with open(index_file, "r", encoding="utf-8") as f:
index = json.loads(f.read())
all_shards = set(index["weight_map"].values())
shards_found = {f for f in os.listdir(tmp_dir) if f.endswith(".h5")}
self.assertSetEqual(all_shards, shards_found)
# Finally, check the model can be reloaded
new_model = TFBertModel.from_pretrained(tmp_dir)
model(model.dummy_inputs)
new_model(model.dummy_inputs)
for p1, p2 in zip(model.weights, new_model.weights):
self.assertTrue(np.allclose(p1.numpy(), p2.numpy()))
@slow
def test_save_pretrained_signatures(self):
model = TFBertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
# Short custom TF signature function.
# `input_signature` is specific to BERT.
@tf.function(
input_signature=[
[
tf.TensorSpec([None, None], tf.int32, name="input_ids"),
tf.TensorSpec([None, None], tf.int32, name="token_type_ids"),
tf.TensorSpec([None, None], tf.int32, name="attention_mask"),
]
]
)
def serving_fn(input):
return model(input)
# Using default signature (default behavior) overrides 'serving_default'
with tempfile.TemporaryDirectory() as tmp_dir:
model.save_pretrained(tmp_dir, saved_model=True, signatures=None)
model_loaded = tf.keras.models.load_model(f"{tmp_dir}/saved_model/1")
self.assertTrue("serving_default" in list(model_loaded.signatures.keys()))
# Providing custom signature function
with tempfile.TemporaryDirectory() as tmp_dir:
model.save_pretrained(tmp_dir, saved_model=True, signatures={"custom_signature": serving_fn})
model_loaded = tf.keras.models.load_model(f"{tmp_dir}/saved_model/1")
self.assertTrue("custom_signature" in list(model_loaded.signatures.keys()))
# Providing multiple custom signature function
with tempfile.TemporaryDirectory() as tmp_dir:
model.save_pretrained(
tmp_dir,
saved_model=True,
signatures={"custom_signature_1": serving_fn, "custom_signature_2": serving_fn},
)
model_loaded = tf.keras.models.load_model(f"{tmp_dir}/saved_model/1")
self.assertTrue("custom_signature_1" in list(model_loaded.signatures.keys()))
self.assertTrue("custom_signature_2" in list(model_loaded.signatures.keys()))
@require_safetensors
def test_safetensors_save_and_load(self):
model = TFBertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
with tempfile.TemporaryDirectory() as tmp_dir:
model.save_pretrained(tmp_dir, safe_serialization=True)
# No tf_model.h5 file, only a model.safetensors
self.assertTrue(os.path.isfile(os.path.join(tmp_dir, SAFE_WEIGHTS_NAME)))
self.assertFalse(os.path.isfile(os.path.join(tmp_dir, TF2_WEIGHTS_NAME)))
new_model = TFBertModel.from_pretrained(tmp_dir)
# Check models are equal
for p1, p2 in zip(model.weights, new_model.weights):
self.assertTrue(np.allclose(p1.numpy(), p2.numpy()))
@is_pt_tf_cross_test
def test_safetensors_save_and_load_pt_to_tf(self):
model = TFBertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
pt_model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
with tempfile.TemporaryDirectory() as tmp_dir:
pt_model.save_pretrained(tmp_dir, safe_serialization=True)
# Check we have a model.safetensors file
self.assertTrue(os.path.isfile(os.path.join(tmp_dir, SAFE_WEIGHTS_NAME)))
new_model = TFBertModel.from_pretrained(tmp_dir)
# Check models are equal
for p1, p2 in zip(model.weights, new_model.weights):
self.assertTrue(np.allclose(p1.numpy(), p2.numpy()))
@require_safetensors
def test_safetensors_load_from_hub(self):
tf_model = TFBertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
# Can load from the TF-formatted checkpoint
safetensors_model = TFBertModel.from_pretrained("hf-internal-testing/tiny-random-bert-safetensors-tf")
# Check models are equal
for p1, p2 in zip(safetensors_model.weights, tf_model.weights):
self.assertTrue(np.allclose(p1.numpy(), p2.numpy()))
# Can load from the PyTorch-formatted checkpoint
safetensors_model = TFBertModel.from_pretrained("hf-internal-testing/tiny-random-bert-safetensors")
# Check models are equal
for p1, p2 in zip(safetensors_model.weights, tf_model.weights):
self.assertTrue(np.allclose(p1.numpy(), p2.numpy()))
@require_tf
@is_staging_test
class TFModelPushToHubTester(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls._token = TOKEN
HfFolder.save_token(TOKEN)
@classmethod
def tearDownClass(cls):
try:
delete_repo(token=cls._token, repo_id="test-model-tf")
except HTTPError:
pass
try:
delete_repo(token=cls._token, repo_id="test-model-tf-callback")
except HTTPError:
pass
try:
delete_repo(token=cls._token, repo_id="valid_org/test-model-tf-org")
except HTTPError:
pass
def test_push_to_hub(self):
config = BertConfig(
vocab_size=99, hidden_size=32, num_hidden_layers=5, num_attention_heads=4, intermediate_size=37
)
model = TFBertModel(config)
# Make sure model is properly initialized
_ = model(model.dummy_inputs)
logging.set_verbosity_info()
logger = logging.get_logger("transformers.utils.hub")
with CaptureLogger(logger) as cl:
model.push_to_hub("test-model-tf", use_auth_token=self._token)
logging.set_verbosity_warning()
# Check the model card was created and uploaded.
self.assertIn("Uploading the following files to __DUMMY_TRANSFORMERS_USER__/test-model-tf", cl.out)
new_model = TFBertModel.from_pretrained(f"{USER}/test-model-tf")
models_equal = True
for p1, p2 in zip(model.weights, new_model.weights):
if not tf.math.reduce_all(p1 == p2):
models_equal = False
break
self.assertTrue(models_equal)
# Reset repo
delete_repo(token=self._token, repo_id="test-model-tf")
# Push to hub via save_pretrained
with tempfile.TemporaryDirectory() as tmp_dir:
model.save_pretrained(tmp_dir, repo_id="test-model-tf", push_to_hub=True, use_auth_token=self._token)
new_model = TFBertModel.from_pretrained(f"{USER}/test-model-tf")
models_equal = True
for p1, p2 in zip(model.weights, new_model.weights):
if not tf.math.reduce_all(p1 == p2):
models_equal = False
break
self.assertTrue(models_equal)
@is_pt_tf_cross_test
def test_push_to_hub_callback(self):
config = BertConfig(
vocab_size=99, hidden_size=32, num_hidden_layers=5, num_attention_heads=4, intermediate_size=37
)
model = TFBertForMaskedLM(config)
model.compile()
with tempfile.TemporaryDirectory() as tmp_dir:
push_to_hub_callback = PushToHubCallback(
output_dir=tmp_dir,
hub_model_id="test-model-tf-callback",
hub_token=self._token,
)
model.fit(model.dummy_inputs, model.dummy_inputs, epochs=1, callbacks=[push_to_hub_callback])
new_model = TFBertForMaskedLM.from_pretrained(f"{USER}/test-model-tf-callback")
models_equal = True
for p1, p2 in zip(model.weights, new_model.weights):
if not tf.math.reduce_all(p1 == p2):
models_equal = False
break
self.assertTrue(models_equal)
tf_push_to_hub_params = dict(inspect.signature(TFPreTrainedModel.push_to_hub).parameters)
tf_push_to_hub_params.pop("base_model_card_args")
pt_push_to_hub_params = dict(inspect.signature(PreTrainedModel.push_to_hub).parameters)
pt_push_to_hub_params.pop("deprecated_kwargs")
self.assertDictEaual(tf_push_to_hub_params, pt_push_to_hub_params)
def test_push_to_hub_in_organization(self):
config = BertConfig(
vocab_size=99, hidden_size=32, num_hidden_layers=5, num_attention_heads=4, intermediate_size=37
)
model = TFBertModel(config)
# Make sure model is properly initialized
_ = model(model.dummy_inputs)
model.push_to_hub("valid_org/test-model-tf-org", use_auth_token=self._token)
new_model = TFBertModel.from_pretrained("valid_org/test-model-tf-org")
models_equal = True
for p1, p2 in zip(model.weights, new_model.weights):
if not tf.math.reduce_all(p1 == p2):
models_equal = False
break
self.assertTrue(models_equal)
# Reset repo
delete_repo(token=self._token, repo_id="valid_org/test-model-tf-org")
# Push to hub via save_pretrained
with tempfile.TemporaryDirectory() as tmp_dir:
model.save_pretrained(
tmp_dir, push_to_hub=True, use_auth_token=self._token, repo_id="valid_org/test-model-tf-org"
)
new_model = TFBertModel.from_pretrained("valid_org/test-model-tf-org")
models_equal = True
for p1, p2 in zip(model.weights, new_model.weights):
if not tf.math.reduce_all(p1 == p2):
models_equal = False
break
self.assertTrue(models_equal)
|