Spaces:
Paused
Paused
File size: 127,529 Bytes
ee6e328 |
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 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752 2753 2754 2755 2756 2757 2758 2759 2760 2761 2762 2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799 2800 2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819 2820 2821 2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881 2882 2883 2884 2885 2886 2887 |
# coding=utf-8
# Copyright 2020 The HuggingFace Team 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 clone 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 inspect
import unittest
import warnings
import numpy as np
from transformers import is_torch_available, pipeline
from transformers.testing_utils import require_torch, slow, torch_device
from ..test_modeling_common import floats_tensor, ids_tensor
from .test_framework_agnostic import GenerationIntegrationTestsMixin
if is_torch_available():
import torch
from transformers import (
AutoModelForCausalLM,
AutoModelForSeq2SeqLM,
AutoModelForSpeechSeq2Seq,
AutoModelForVision2Seq,
AutoTokenizer,
BartForConditionalGeneration,
BartTokenizer,
GPT2LMHeadModel,
GPT2Tokenizer,
ImageGPTForCausalImageModeling,
SpeechEncoderDecoderModel,
top_k_top_p_filtering,
)
from transformers.generation import (
BeamSampleDecoderOnlyOutput,
BeamSampleEncoderDecoderOutput,
BeamSearchDecoderOnlyOutput,
BeamSearchEncoderDecoderOutput,
BeamSearchScorer,
ConstrainedBeamSearchScorer,
DisjunctiveConstraint,
ForcedBOSTokenLogitsProcessor,
ForcedEOSTokenLogitsProcessor,
GreedySearchDecoderOnlyOutput,
GreedySearchEncoderDecoderOutput,
HammingDiversityLogitsProcessor,
InfNanRemoveLogitsProcessor,
LogitsProcessorList,
MaxLengthCriteria,
MinLengthLogitsProcessor,
NoBadWordsLogitsProcessor,
NoRepeatNGramLogitsProcessor,
PhrasalConstraint,
RepetitionPenaltyLogitsProcessor,
SampleDecoderOnlyOutput,
SampleEncoderDecoderOutput,
StoppingCriteria,
StoppingCriteriaList,
TemperatureLogitsWarper,
TopKLogitsWarper,
TopPLogitsWarper,
)
class GenerationTesterMixin:
model_tester = None
all_generative_model_classes = ()
input_name = "input_ids"
def _get_input_ids_and_config(self, batch_size=2):
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
input_ids = inputs_dict[self.input_name]
# cut to half length & take max batch_size 3
sequence_length = input_ids.shape[-1] // 2
input_ids = input_ids[:batch_size, :sequence_length]
# generate max 3 tokens
max_length = input_ids.shape[-1] + 3
if config.eos_token_id is not None and config.pad_token_id is None:
# hack to allow generate for models such as GPT2 as is done in `generate()`
if isinstance(config.eos_token_id, int):
config.eos_token_id = [config.eos_token_id]
config.pad_token_id = config.eos_token_id[0]
# TransfoXL has no attention mask
if "transfoxl" in config.__class__.__name__.lower():
attention_mask = None
else:
attention_mask = torch.ones_like(input_ids, dtype=torch.long)[:batch_size, :sequence_length]
return config, input_ids, attention_mask, max_length
@staticmethod
def _get_logits_processor_and_kwargs(
input_length,
eos_token_id,
forced_bos_token_id=None,
forced_eos_token_id=None,
max_length=None,
diversity_penalty=None,
):
process_kwargs = {
"min_length": input_length + 1 if max_length is None else max_length - 1,
"bad_words_ids": [[1, 0]],
"no_repeat_ngram_size": 2,
"repetition_penalty": 1.2,
}
logits_processor = LogitsProcessorList(
(
[
HammingDiversityLogitsProcessor(diversity_penalty, num_beams=2, num_beam_groups=2),
]
if diversity_penalty is not None
else []
)
+ (
[
MinLengthLogitsProcessor(process_kwargs["min_length"], eos_token_id),
]
if eos_token_id is not None
else []
)
+ (
[
ForcedBOSTokenLogitsProcessor(forced_bos_token_id),
]
if forced_bos_token_id is not None
else []
)
+ (
[ForcedEOSTokenLogitsProcessor(max_length, forced_eos_token_id)]
if forced_eos_token_id is not None
else []
)
+ [
NoBadWordsLogitsProcessor(process_kwargs["bad_words_ids"], eos_token_id),
NoRepeatNGramLogitsProcessor(process_kwargs["no_repeat_ngram_size"]),
RepetitionPenaltyLogitsProcessor(process_kwargs["repetition_penalty"]),
]
)
return process_kwargs, logits_processor
@staticmethod
def _get_warper_and_kwargs(num_beams):
warp_kwargs = {"top_k": 10, "top_p": 0.7, "temperature": 0.7}
logits_warper = LogitsProcessorList(
[
TemperatureLogitsWarper(warp_kwargs["temperature"]),
TopKLogitsWarper(top_k=warp_kwargs["top_k"], min_tokens_to_keep=(2 if num_beams > 1 else 1)),
TopPLogitsWarper(top_p=warp_kwargs["top_p"], min_tokens_to_keep=(2 if num_beams > 1 else 1)),
]
)
return warp_kwargs, logits_warper
@staticmethod
def _get_beam_scorer_and_kwargs(batch_size, max_length, num_return_sequences=1):
beam_kwargs = {
"early_stopping": False,
"length_penalty": 2.0,
"num_beams": 2,
"num_return_sequences": num_return_sequences,
}
beam_scorer = BeamSearchScorer(
batch_size=batch_size,
num_beams=beam_kwargs["num_beams"],
device=torch_device,
length_penalty=beam_kwargs["length_penalty"],
do_early_stopping=beam_kwargs["early_stopping"],
num_beam_hyps_to_keep=num_return_sequences,
)
return beam_kwargs, beam_scorer
@staticmethod
def _get_diverse_beam_scorer_and_kwargs(batch_size, max_length, num_return_sequences=1):
beam_kwargs = {
"early_stopping": False,
"length_penalty": 2.0,
"num_beams": 2,
"num_return_sequences": num_return_sequences,
"num_beam_groups": 2, # one beam per group
"diversity_penalty": 2.0,
}
beam_scorer = BeamSearchScorer(
batch_size=batch_size,
num_beams=beam_kwargs["num_beams"],
device=torch_device,
length_penalty=beam_kwargs["length_penalty"],
do_early_stopping=beam_kwargs["early_stopping"],
num_beam_hyps_to_keep=num_return_sequences,
num_beam_groups=beam_kwargs["num_beam_groups"],
)
return beam_kwargs, beam_scorer
@staticmethod
def _get_constrained_beam_scorer_and_kwargs(batch_size, max_length, constraints, num_return_sequences=1):
beam_kwargs = {
"early_stopping": False,
"length_penalty": 2.0,
"num_beams": num_return_sequences * 4,
"num_return_sequences": num_return_sequences,
}
beam_scorer = ConstrainedBeamSearchScorer(
batch_size=batch_size,
constraints=constraints,
num_beams=beam_kwargs["num_beams"],
device=torch_device,
length_penalty=beam_kwargs["length_penalty"],
do_early_stopping=beam_kwargs["early_stopping"],
num_beam_hyps_to_keep=num_return_sequences,
)
return beam_kwargs, beam_scorer
@staticmethod
def _get_encoder_outputs(
model, input_ids, attention_mask, output_attentions=None, output_hidden_states=None, num_interleave=1
):
encoder = model.get_encoder()
encoder_outputs = encoder(
input_ids,
attention_mask=attention_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
encoder_outputs["last_hidden_state"] = encoder_outputs.last_hidden_state.repeat_interleave(
num_interleave, dim=0
)
input_ids = torch.zeros_like(input_ids[:, :1]) + model._get_decoder_start_token_id()
attention_mask = None
return encoder_outputs, input_ids, attention_mask
def _greedy_generate(
self,
model,
input_ids,
attention_mask,
max_length,
output_scores=False,
output_attentions=False,
output_hidden_states=False,
return_dict_in_generate=False,
):
if model.config.is_encoder_decoder:
max_length = 4
logits_process_kwargs, logits_processor = self._get_logits_processor_and_kwargs(
input_ids.shape[-1],
eos_token_id=model.config.eos_token_id,
forced_bos_token_id=model.config.forced_bos_token_id,
forced_eos_token_id=model.config.forced_eos_token_id,
max_length=max_length,
)
kwargs = {}
model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
output_generate = model.generate(
input_ids,
do_sample=False,
num_beams=1,
max_length=max_length,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
output_scores=output_scores,
return_dict_in_generate=return_dict_in_generate,
remove_invalid_values=True,
**logits_process_kwargs,
**model_kwargs,
)
if model.config.is_encoder_decoder:
encoder_outputs, input_ids, attention_mask = self._get_encoder_outputs(
model,
input_ids,
attention_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
kwargs["encoder_outputs"] = encoder_outputs
with torch.no_grad():
model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
output_greedy = model.greedy_search(
input_ids,
max_length=max_length,
logits_processor=logits_processor,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
output_scores=output_scores,
return_dict_in_generate=return_dict_in_generate,
**kwargs,
**model_kwargs,
)
return output_greedy, output_generate
def _sample_generate(
self,
model,
input_ids,
attention_mask,
max_length,
num_return_sequences,
logits_processor,
logits_warper,
logits_warper_kwargs,
process_kwargs,
output_scores=False,
output_attentions=False,
output_hidden_states=False,
return_dict_in_generate=False,
):
torch.manual_seed(0)
model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
output_generate = model.generate(
input_ids,
do_sample=True,
num_beams=1,
max_length=max_length,
num_return_sequences=num_return_sequences,
output_scores=output_scores,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict_in_generate=return_dict_in_generate,
remove_invalid_values=True,
**logits_warper_kwargs,
**process_kwargs,
**model_kwargs,
)
torch.manual_seed(0)
kwargs = {}
if model.config.is_encoder_decoder:
encoder_outputs, input_ids, attention_mask = self._get_encoder_outputs(
model,
input_ids,
attention_mask,
num_interleave=num_return_sequences,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
kwargs["encoder_outputs"] = encoder_outputs
elif attention_mask is not None:
attention_mask = attention_mask.repeat_interleave(num_return_sequences, dim=0)
# prevent flaky generation test failures
logits_processor.append(InfNanRemoveLogitsProcessor())
with torch.no_grad():
model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
output_sample = model.sample(
input_ids.repeat_interleave(num_return_sequences, dim=0),
max_length=max_length,
logits_processor=logits_processor,
logits_warper=logits_warper,
output_scores=output_scores,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict_in_generate=return_dict_in_generate,
**kwargs,
**model_kwargs,
)
return output_sample, output_generate
def _beam_search_generate(
self,
model,
input_ids,
attention_mask,
max_length,
beam_scorer,
beam_kwargs,
logits_processor,
logits_process_kwargs,
output_scores=False,
output_attentions=False,
output_hidden_states=False,
return_dict_in_generate=False,
):
model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
output_generate = model.generate(
input_ids,
do_sample=False,
max_length=max_length,
output_scores=output_scores,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict_in_generate=return_dict_in_generate,
remove_invalid_values=True,
**beam_kwargs,
**logits_process_kwargs,
**model_kwargs,
)
# beam_search does not automatically interleave `batch_size` dim for `num_beams`
kwargs = {}
if model.config.is_encoder_decoder:
encoder_outputs, input_ids, attention_mask = self._get_encoder_outputs(
model,
input_ids,
attention_mask,
num_interleave=beam_scorer.num_beams,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
kwargs["encoder_outputs"] = encoder_outputs
elif attention_mask is not None:
attention_mask = attention_mask.repeat_interleave(beam_scorer.num_beams, dim=0)
with torch.no_grad():
model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
output_beam_search = model.beam_search(
input_ids.repeat_interleave(beam_scorer.num_beams, dim=0),
beam_scorer,
max_length=max_length,
logits_processor=logits_processor,
output_scores=output_scores,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict_in_generate=return_dict_in_generate,
**kwargs,
**model_kwargs,
)
return output_generate, output_beam_search
def _beam_sample_generate(
self,
model,
input_ids,
attention_mask,
max_length,
beam_scorer,
beam_kwargs,
logits_warper,
logits_warper_kwargs,
output_scores=False,
output_attentions=False,
output_hidden_states=False,
return_dict_in_generate=False,
):
torch.manual_seed(0)
model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
output_generate = model.generate(
input_ids,
do_sample=True,
max_length=max_length,
output_scores=output_scores,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict_in_generate=return_dict_in_generate,
remove_invalid_values=True,
**beam_kwargs,
**logits_warper_kwargs,
**model_kwargs,
)
# beam_search does not automatically interleave `batch_size` dim for `num_beams`
torch.manual_seed(0)
kwargs = {}
if model.config.is_encoder_decoder:
encoder_outputs, input_ids, attention_mask = self._get_encoder_outputs(
model,
input_ids,
attention_mask,
num_interleave=beam_scorer.num_beams,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
kwargs["encoder_outputs"] = encoder_outputs
elif attention_mask is not None:
attention_mask = attention_mask.repeat_interleave(beam_scorer.num_beams, dim=0)
# prevent flaky generation test failures
logits_processor = LogitsProcessorList()
logits_processor.append(InfNanRemoveLogitsProcessor())
with torch.no_grad():
model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
output_beam_sample = model.beam_sample(
input_ids.repeat_interleave(beam_scorer.num_beams, dim=0),
beam_scorer,
max_length=max_length,
logits_warper=logits_warper,
logits_processor=logits_processor,
output_scores=output_scores,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict_in_generate=return_dict_in_generate,
**kwargs,
**model_kwargs,
)
return output_generate, output_beam_sample
def _group_beam_search_generate(
self,
model,
input_ids,
attention_mask,
max_length,
beam_scorer,
beam_kwargs,
logits_processor,
logits_process_kwargs,
output_scores=False,
output_attentions=False,
output_hidden_states=False,
return_dict_in_generate=False,
):
model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
output_generate = model.generate(
input_ids,
do_sample=False,
max_length=max_length,
output_scores=output_scores,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict_in_generate=return_dict_in_generate,
remove_invalid_values=True,
**beam_kwargs,
**logits_process_kwargs,
**model_kwargs,
)
# group_beam_search does not automatically interleave `batch_size` dim for `num_beams`
kwargs = {}
if model.config.is_encoder_decoder:
encoder_outputs, input_ids, attention_mask = self._get_encoder_outputs(
model,
input_ids,
attention_mask,
num_interleave=beam_scorer.num_beams,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
kwargs["encoder_outputs"] = encoder_outputs
elif attention_mask is not None:
attention_mask = attention_mask.repeat_interleave(beam_scorer.num_beams, dim=0)
with torch.no_grad():
model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
output_group_beam_search = model.group_beam_search(
input_ids.repeat_interleave(beam_scorer.num_beams, dim=0),
beam_scorer,
max_length=max_length,
logits_processor=logits_processor,
output_scores=output_scores,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict_in_generate=return_dict_in_generate,
**kwargs,
**model_kwargs,
)
return output_generate, output_group_beam_search
def _constrained_beam_search_generate(
self,
model,
input_ids,
attention_mask,
max_length,
constrained_beam_scorer,
constraints,
beam_kwargs,
logits_processor,
logits_process_kwargs,
output_scores=False,
output_attentions=False,
output_hidden_states=False,
return_dict_in_generate=False,
):
model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
output_generate = model.generate(
input_ids,
do_sample=False,
max_length=max_length,
output_scores=output_scores,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict_in_generate=return_dict_in_generate,
remove_invalid_values=True,
constraints=constraints,
**beam_kwargs,
**logits_process_kwargs,
**model_kwargs,
)
# group_beam_search does not automatically interleave `batch_size` dim for `num_beams`
kwargs = {}
if model.config.is_encoder_decoder:
encoder_outputs, input_ids, attention_mask = self._get_encoder_outputs(
model,
input_ids,
attention_mask,
num_interleave=constrained_beam_scorer.num_beams,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
kwargs["encoder_outputs"] = encoder_outputs
elif attention_mask is not None:
attention_mask = attention_mask.repeat_interleave(constrained_beam_scorer.num_beams, dim=0)
with torch.no_grad():
model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
output_group_beam_search = model.constrained_beam_search(
input_ids.repeat_interleave(constrained_beam_scorer.num_beams, dim=0),
constrained_beam_scorer,
max_length=max_length,
logits_processor=logits_processor,
output_scores=output_scores,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict_in_generate=return_dict_in_generate,
**kwargs,
**model_kwargs,
)
return output_generate, output_group_beam_search
def _contrastive_generate(
self,
model,
input_ids,
attention_mask,
max_length,
output_scores=False,
output_attentions=False,
output_hidden_states=False,
return_dict_in_generate=False,
):
contrastive_search_kwargs = {
"penalty_alpha": 0.6,
"top_k": 5,
}
if model.config.is_encoder_decoder:
max_length = 4
logits_process_kwargs, logits_processor = self._get_logits_processor_and_kwargs(
input_ids.shape[-1],
eos_token_id=model.config.eos_token_id,
forced_bos_token_id=model.config.forced_bos_token_id,
forced_eos_token_id=model.config.forced_eos_token_id,
max_length=max_length,
)
kwargs = {}
model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
output_generate = model.generate(
input_ids,
do_sample=False,
num_beams=1,
max_length=max_length,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
output_scores=output_scores,
return_dict_in_generate=return_dict_in_generate,
remove_invalid_values=True,
**logits_process_kwargs,
**model_kwargs,
**contrastive_search_kwargs,
)
if model.config.is_encoder_decoder:
encoder_outputs, input_ids, attention_mask = self._get_encoder_outputs(
model,
input_ids,
attention_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
kwargs["encoder_outputs"] = encoder_outputs
with torch.no_grad():
model_kwargs = {"attention_mask": attention_mask} if attention_mask is not None else {}
stopping_criteria = StoppingCriteriaList([MaxLengthCriteria(max_length=max_length)])
output_contrastive = model.contrastive_search(
input_ids,
stopping_criteria=stopping_criteria,
logits_processor=logits_processor,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
output_scores=output_scores,
return_dict_in_generate=return_dict_in_generate,
**kwargs,
**model_kwargs,
**contrastive_search_kwargs,
)
return output_contrastive, output_generate
def test_greedy_generate(self):
# check `generate()` and `greedy_search()` are equal
for model_class in self.all_generative_model_classes:
config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
# test old generation output for backwards compatibility
model = model_class(config).to(torch_device).eval()
output_greedy, output_generate = self._greedy_generate(
model=model, input_ids=input_ids, attention_mask=attention_mask, max_length=max_length
)
self.assertListEqual(output_greedy.tolist(), output_generate.tolist())
def test_greedy_generate_dict_outputs(self):
for model_class in self.all_generative_model_classes:
# disable cache
config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
config.use_cache = False
model = model_class(config).to(torch_device).eval()
output_greedy, output_generate = self._greedy_generate(
model=model,
input_ids=input_ids,
attention_mask=attention_mask,
max_length=max_length,
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, GreedySearchEncoderDecoderOutput)
self.assertIsInstance(output_generate, GreedySearchEncoderDecoderOutput)
else:
self.assertIsInstance(output_greedy, GreedySearchDecoderOnlyOutput)
self.assertIsInstance(output_generate, GreedySearchDecoderOnlyOutput)
self.assertListEqual(output_generate.sequences.tolist(), output_greedy.sequences.tolist())
for output in (output_greedy, output_generate):
self._check_outputs(output, input_ids, model.config)
def test_greedy_generate_dict_outputs_use_cache(self):
for model_class in self.all_generative_model_classes:
# enable cache
config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
if not hasattr(config, "use_cache"):
# only relevant if model has "use_cache"
return
config.use_cache = True
config.is_decoder = True
model = model_class(config).to(torch_device).eval()
output_greedy, output_generate = self._greedy_generate(
model=model,
input_ids=input_ids,
attention_mask=attention_mask,
max_length=max_length,
output_scores=True,
output_hidden_states=True,
output_attentions=True,
return_dict_in_generate=True,
)
self.assertListEqual(output_generate.sequences.tolist(), output_greedy.sequences.tolist())
for output in (output_greedy, output_generate):
self._check_outputs(output, input_ids, model.config, use_cache=True)
def test_sample_generate(self):
for model_class in self.all_generative_model_classes:
config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
model = model_class(config).to(torch_device).eval()
if model.config.is_encoder_decoder:
max_length = 4
process_kwargs, logits_processor = self._get_logits_processor_and_kwargs(
input_ids.shape[-1],
model.config.eos_token_id,
forced_bos_token_id=model.config.forced_bos_token_id,
forced_eos_token_id=model.config.forced_eos_token_id,
max_length=max_length,
)
logits_warper_kwargs, logits_warper = self._get_warper_and_kwargs(num_beams=2)
# check `generate()` and `sample()` are equal
output_sample, output_generate = self._sample_generate(
model=model,
input_ids=input_ids,
attention_mask=attention_mask,
max_length=max_length,
num_return_sequences=1,
logits_processor=logits_processor,
logits_warper=logits_warper,
logits_warper_kwargs=logits_warper_kwargs,
process_kwargs=process_kwargs,
)
self.assertListEqual(output_sample.tolist(), output_generate.tolist())
# check `generate()` and `sample()` yield equal results for `num_return_sequences`
output_sample, output_generate = self._sample_generate(
model=model,
input_ids=input_ids,
attention_mask=attention_mask,
max_length=max_length,
num_return_sequences=3,
logits_processor=logits_processor,
logits_warper=logits_warper,
logits_warper_kwargs=logits_warper_kwargs,
process_kwargs=process_kwargs,
)
self.assertListEqual(output_sample.tolist(), output_generate.tolist())
def test_sample_generate_dict_output(self):
for model_class in self.all_generative_model_classes:
# disable cache
config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
config.use_cache = False
model = model_class(config).to(torch_device).eval()
if model.config.is_encoder_decoder:
max_length = 4
process_kwargs, logits_processor = self._get_logits_processor_and_kwargs(
input_ids.shape[-1],
model.config.eos_token_id,
forced_bos_token_id=model.config.forced_bos_token_id,
forced_eos_token_id=model.config.forced_eos_token_id,
max_length=max_length,
)
logits_warper_kwargs, logits_warper = self._get_warper_and_kwargs(num_beams=1)
output_sample, output_generate = self._sample_generate(
model=model,
input_ids=input_ids,
attention_mask=attention_mask,
max_length=max_length,
num_return_sequences=2,
logits_processor=logits_processor,
logits_warper=logits_warper,
logits_warper_kwargs=logits_warper_kwargs,
process_kwargs=process_kwargs,
output_scores=True,
output_hidden_states=True,
output_attentions=True,
return_dict_in_generate=True,
)
if model.config.is_encoder_decoder:
self.assertIsInstance(output_sample, SampleEncoderDecoderOutput)
self.assertIsInstance(output_generate, SampleEncoderDecoderOutput)
else:
self.assertIsInstance(output_sample, SampleDecoderOnlyOutput)
self.assertIsInstance(output_generate, SampleDecoderOnlyOutput)
self.assertListEqual(output_generate.sequences.tolist(), output_sample.sequences.tolist())
for output in (output_sample, output_generate):
self._check_outputs(output, input_ids, model.config, num_return_sequences=2)
def test_beam_search_generate(self):
for model_class in self.all_generative_model_classes:
config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
# It is important set set the eos_token_id to None to ensure that no sequences
# shorter than `max_length` can be generated which could lead to flaky circle ci
# failures if the top `num_return_sequences` beams are all shorter than the longest beam
config.eos_token_id = None
config.forced_eos_token_id = None
model = model_class(config).to(torch_device).eval()
if model.config.is_encoder_decoder:
max_length = 4
logits_process_kwargs, logits_processor = self._get_logits_processor_and_kwargs(
input_ids.shape[-1],
config.eos_token_id,
config.forced_bos_token_id,
config.forced_eos_token_id,
max_length,
)
beam_kwargs, beam_scorer = self._get_beam_scorer_and_kwargs(input_ids.shape[0], max_length)
# check `generate()` and `beam_search()` are equal
output_generate, output_beam_search = self._beam_search_generate(
model=model,
input_ids=input_ids,
attention_mask=attention_mask,
max_length=max_length,
beam_scorer=beam_scorer,
beam_kwargs=beam_kwargs,
logits_process_kwargs=logits_process_kwargs,
logits_processor=logits_processor,
)
self.assertListEqual(output_generate.tolist(), output_beam_search.tolist())
if model.config.is_encoder_decoder:
max_length = 4
beam_kwargs, beam_scorer = self._get_beam_scorer_and_kwargs(input_ids.shape[0], max_length)
output_generate, output_beam_search = self._beam_search_generate(
model=model,
input_ids=input_ids,
attention_mask=attention_mask,
max_length=max_length,
beam_scorer=beam_scorer,
beam_kwargs=beam_kwargs,
logits_process_kwargs=logits_process_kwargs,
logits_processor=logits_processor,
)
self.assertListEqual(output_generate.tolist(), output_beam_search.tolist())
def test_beam_search_generate_dict_output(self):
for model_class in self.all_generative_model_classes:
config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
# disable cache
config.use_cache = False
# It is important set set the eos_token_id to None to ensure that no sequences
# shorter than `max_length` can be generated which could lead to flaky circle ci
# failures if the top `num_return_sequences` beams are all shorter than the longest beam
config.eos_token_id = None
config.forced_eos_token_id = None
model = model_class(config).to(torch_device).eval()
if model.config.is_encoder_decoder:
max_length = 4
logits_process_kwargs, logits_processor = self._get_logits_processor_and_kwargs(
input_ids.shape[-1],
config.eos_token_id,
config.forced_bos_token_id,
config.forced_eos_token_id,
max_length,
)
beam_kwargs, beam_scorer = self._get_beam_scorer_and_kwargs(input_ids.shape[0], max_length)
output_generate, output_beam_search = self._beam_search_generate(
model=model,
input_ids=input_ids,
attention_mask=attention_mask,
max_length=max_length,
beam_scorer=beam_scorer,
beam_kwargs=beam_kwargs,
logits_process_kwargs=logits_process_kwargs,
logits_processor=logits_processor,
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, BeamSearchEncoderDecoderOutput)
self.assertIsInstance(output_generate, BeamSearchEncoderDecoderOutput)
else:
self.assertIsInstance(output_beam_search, BeamSearchDecoderOnlyOutput)
self.assertIsInstance(output_generate, BeamSearchDecoderOnlyOutput)
self.assertListEqual(output_generate.sequences.tolist(), output_beam_search.sequences.tolist())
self.assertTrue(
torch.allclose(output_generate["sequences_scores"], output_beam_search["sequences_scores"], atol=1e-3)
)
self.assertTrue(output_generate["sequences_scores"].shape == (output_generate["sequences"].shape[0],))
self.assertTrue((output_generate["sequences_scores"] < 0).all().item())
for output in (output_beam_search, output_generate):
self._check_outputs(output, input_ids, model.config, num_return_sequences=beam_scorer.num_beams)
def test_beam_search_generate_dict_outputs_use_cache(self):
for model_class in self.all_generative_model_classes:
# enable cache
config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
# It is important set set the eos_token_id to None to ensure that no sequences
# shorter than `max_length` can be generated which could lead to flaky circle ci
# failures if the top `num_return_sequences` beams are all shorter than the longest beam
config.eos_token_id = None
config.forced_eos_token_id = None
if not hasattr(config, "use_cache"):
# only relevant if model has "use_cache"
return
model = model_class(config).to(torch_device).eval()
if model.config.is_encoder_decoder:
max_length = 4
logits_process_kwargs, logits_processor = self._get_logits_processor_and_kwargs(
input_ids.shape[-1],
config.eos_token_id,
config.forced_bos_token_id,
config.forced_eos_token_id,
max_length,
)
beam_kwargs, beam_scorer = self._get_beam_scorer_and_kwargs(input_ids.shape[0], max_length)
config.use_cache = True
config.is_decoder = True
model = model_class(config).to(torch_device).eval()
output_beam, output_generate = self._beam_search_generate(
model=model,
input_ids=input_ids,
attention_mask=attention_mask,
max_length=max_length,
beam_scorer=beam_scorer,
beam_kwargs=beam_kwargs,
logits_process_kwargs=logits_process_kwargs,
logits_processor=logits_processor,
output_scores=True,
output_hidden_states=True,
output_attentions=True,
return_dict_in_generate=True,
)
self.assertListEqual(output_generate.sequences.tolist(), output_beam.sequences.tolist())
for output in (output_beam, output_generate):
self._check_outputs(
output, input_ids, model.config, use_cache=True, num_return_sequences=beam_scorer.num_beams
)
def test_beam_sample_generate(self):
for model_class in self.all_generative_model_classes:
config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
# It is important set set the eos_token_id to None to ensure that no sequences
# shorter than `max_length` can be generated which could lead to flaky circle ci
# failures if the top `num_return_sequences` beams are all shorter than the longest beam
config.eos_token_id = None
config.forced_eos_token_id = None
logits_warper_kwargs, logits_warper = self._get_warper_and_kwargs(num_beams=1)
model = model_class(config).to(torch_device).eval()
# check `generate()` and `beam_search()` are equal
if model.config.is_encoder_decoder:
max_length = 4
beam_kwargs, beam_scorer = self._get_beam_scorer_and_kwargs(input_ids.shape[0], max_length)
output_generate, output_beam_sample = self._beam_sample_generate(
model=model,
input_ids=input_ids,
attention_mask=attention_mask,
max_length=max_length,
beam_scorer=beam_scorer,
beam_kwargs=beam_kwargs,
logits_warper=logits_warper,
logits_warper_kwargs=logits_warper_kwargs,
)
self.assertListEqual(output_generate.tolist(), output_beam_sample.tolist())
def test_beam_sample_generate_dict_output(self):
for model_class in self.all_generative_model_classes:
config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
# disable cache
config.use_cache = False
# It is important set set the eos_token_id to None to ensure that no sequences
# shorter than `max_length` can be generated which could lead to flaky circle ci
# failures if the top `num_return_sequences` beams are all shorter than the longest beam
config.eos_token_id = None
config.forced_eos_token_id = None
model = model_class(config).to(torch_device).eval()
logits_warper_kwargs, logits_warper = self._get_warper_and_kwargs(num_beams=1)
if model.config.is_encoder_decoder:
max_length = 4
beam_kwargs, beam_scorer = self._get_beam_scorer_and_kwargs(input_ids.shape[0], max_length)
output_beam_sample, output_generate = self._beam_sample_generate(
model=model,
input_ids=input_ids,
attention_mask=attention_mask,
max_length=max_length,
beam_scorer=beam_scorer,
beam_kwargs=beam_kwargs,
logits_warper=logits_warper,
logits_warper_kwargs=logits_warper_kwargs,
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_sample, BeamSampleEncoderDecoderOutput)
self.assertIsInstance(output_generate, BeamSampleEncoderDecoderOutput)
else:
self.assertIsInstance(output_beam_sample, BeamSampleDecoderOnlyOutput)
self.assertIsInstance(output_generate, BeamSampleDecoderOnlyOutput)
self.assertListEqual(output_generate.sequences.tolist(), output_beam_sample.sequences.tolist())
self.assertTrue(
torch.allclose(output_generate["sequences_scores"], output_beam_sample["sequences_scores"], atol=1e-3)
)
self.assertTrue(output_generate["sequences_scores"].shape == (output_generate["sequences"].shape[0],))
self.assertTrue((output_generate["sequences_scores"] < 0).all().item())
for output in (output_beam_sample, output_generate):
self._check_outputs(output, input_ids, model.config, num_return_sequences=beam_scorer.num_beams)
def test_generate_without_input_ids(self):
config, _, _, max_length = self._get_input_ids_and_config()
# if no bos token id => cannot generate from None
if config.bos_token_id is None:
return
for model_class in self.all_generative_model_classes:
model = model_class(config).to(torch_device)
model.eval()
output_ids_generate = model.generate(do_sample=False, max_length=max_length, remove_invalid_values=True)
self.assertIsNotNone(output_ids_generate)
def test_group_beam_search_generate(self):
for model_class in self.all_generative_model_classes:
config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
# It is important set set the eos_token_id to None to ensure that no sequences
# shorter than `max_length` can be generated which could lead to flaky circle ci
# failures if the top `num_return_sequences` beams are all shorter than the longest beam
config.eos_token_id = None
config.forced_eos_token_id = None
model = model_class(config).to(torch_device).eval()
if model.config.is_encoder_decoder:
max_length = 4
logits_process_kwargs, logits_processor = self._get_logits_processor_and_kwargs(
input_ids.shape[-1],
config.eos_token_id,
config.forced_bos_token_id,
config.forced_eos_token_id,
max_length,
diversity_penalty=2.0,
)
# check `generate()` and `group_beam_search()` are equal
beam_kwargs, beam_scorer = self._get_diverse_beam_scorer_and_kwargs(input_ids.shape[0], max_length)
output_generate, output_group_beam_search = self._group_beam_search_generate(
model=model,
input_ids=input_ids,
attention_mask=attention_mask,
max_length=max_length,
beam_scorer=beam_scorer,
beam_kwargs=beam_kwargs,
logits_processor=logits_processor,
logits_process_kwargs=logits_process_kwargs,
)
self.assertListEqual(output_generate.tolist(), output_group_beam_search.tolist())
# check `generate()` and `group_beam_search()` are equal for `num_return_sequences`
num_return_sequences = 2
if model.config.is_encoder_decoder:
max_length = 4
beam_kwargs, beam_scorer = self._get_diverse_beam_scorer_and_kwargs(
input_ids.shape[0], max_length, num_return_sequences=num_return_sequences
)
output_generate, output_group_beam_search = self._group_beam_search_generate(
model=model,
input_ids=input_ids,
attention_mask=attention_mask,
max_length=max_length,
beam_scorer=beam_scorer,
beam_kwargs=beam_kwargs,
logits_processor=logits_processor,
logits_process_kwargs=logits_process_kwargs,
)
self.assertListEqual(output_generate.tolist(), output_group_beam_search.tolist())
def test_group_beam_search_generate_dict_output(self):
for model_class in self.all_generative_model_classes:
config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
config.use_cache = False
# It is important set set the eos_token_id to None to ensure that no sequences
# shorter than `max_length` can be generated which could lead to flaky circle ci
# failures if the top `num_return_sequences` beams are all shorter than the longest beam
config.eos_token_id = None
config.forced_eos_token_id = None
model = model_class(config).to(torch_device).eval()
if model.config.is_encoder_decoder:
max_length = 4
logits_process_kwargs, logits_processor = self._get_logits_processor_and_kwargs(
input_ids.shape[-1],
config.eos_token_id,
config.forced_bos_token_id,
config.forced_eos_token_id,
max_length,
diversity_penalty=2.0,
)
num_return_sequences = 1
beam_kwargs, beam_scorer = self._get_diverse_beam_scorer_and_kwargs(
input_ids.shape[0], max_length, num_return_sequences=num_return_sequences
)
output_generate, output_group_beam_search = self._group_beam_search_generate(
model=model,
input_ids=input_ids,
attention_mask=attention_mask,
max_length=max_length,
beam_scorer=beam_scorer,
beam_kwargs=beam_kwargs,
logits_processor=logits_processor,
logits_process_kwargs=logits_process_kwargs,
output_scores=True,
output_hidden_states=True,
output_attentions=True,
return_dict_in_generate=True,
)
if model.config.is_encoder_decoder:
self.assertIsInstance(output_group_beam_search, BeamSearchEncoderDecoderOutput)
self.assertIsInstance(output_generate, BeamSearchEncoderDecoderOutput)
else:
self.assertIsInstance(output_group_beam_search, BeamSearchDecoderOnlyOutput)
self.assertIsInstance(output_generate, BeamSearchDecoderOnlyOutput)
self.assertListEqual(output_generate.sequences.tolist(), output_group_beam_search.sequences.tolist())
self.assertTrue(
torch.allclose(
output_generate["sequences_scores"], output_group_beam_search["sequences_scores"], atol=1e-3
)
)
self.assertTrue(output_generate["sequences_scores"].shape == (output_generate["sequences"].shape[0],))
self.assertTrue((output_generate["sequences_scores"] < 0).all().item())
for output in (output_group_beam_search, output_generate):
self._check_outputs(
output, input_ids, model.config, num_return_sequences=num_return_sequences * beam_scorer.num_beams
)
def test_constrained_beam_search_generate(self):
for model_class in self.all_generative_model_classes:
config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
# It is important set set the eos_token_id to None to ensure that no sequences
# shorter than `max_length` can be generated which could lead to flaky circle ci
# failures if the top `num_return_sequences` beams are all shorter than the longest beam
config.eos_token_id = None
config.forced_eos_token_id = None
model = model_class(config).to(torch_device).eval()
max_length = 20
logits_process_kwargs, logits_processor = self._get_logits_processor_and_kwargs(
input_ids.shape[-1],
config.eos_token_id,
config.forced_bos_token_id,
config.forced_eos_token_id,
max_length,
)
# check `generate()` and `constrained_beam_search()` are equal
# Sample constraints
if not input_ids.dtype == torch.float32:
min_id = torch.min(input_ids) + 3
max_id = torch.max(input_ids)
else:
# otherwise this throws an error for Speech2TextModel since its inputs are floating points
min_id = 3
max_id = 100
force_tokens = torch.randint(min_id, max_id, (1, 2)).tolist()[0]
constraints = [
PhrasalConstraint(force_tokens),
]
beam_kwargs, beam_scorer = self._get_constrained_beam_scorer_and_kwargs(
input_ids.shape[0], max_length, constraints, num_return_sequences=1
)
output_generate, output_beam_search = self._constrained_beam_search_generate(
model=model,
input_ids=input_ids,
attention_mask=attention_mask,
max_length=max_length,
constrained_beam_scorer=beam_scorer,
constraints=constraints,
beam_kwargs=beam_kwargs,
logits_processor=logits_processor,
logits_process_kwargs=logits_process_kwargs,
)
self.assertListEqual(output_generate.tolist(), output_beam_search.tolist())
for generation_output in output_generate:
self._check_sequence_inside_sequence(force_tokens, generation_output)
# check `generate()` and `constrained_beam_search()` are equal for `num_return_sequences`
# Sample constraints
force_tokens = torch.randint(min_id, max_id, (1, 2)).tolist()[0]
constraints = [
PhrasalConstraint(force_tokens),
]
num_return_sequences = 2
max_length = 20
beam_kwargs, beam_scorer = self._get_constrained_beam_scorer_and_kwargs(
input_ids.shape[0], max_length, constraints, num_return_sequences=num_return_sequences
)
output_generate, output_beam_search = self._constrained_beam_search_generate(
model=model,
input_ids=input_ids,
attention_mask=attention_mask,
max_length=max_length,
constrained_beam_scorer=beam_scorer,
constraints=constraints,
beam_kwargs=beam_kwargs,
logits_processor=logits_processor,
logits_process_kwargs=logits_process_kwargs,
)
self.assertListEqual(output_generate.tolist(), output_beam_search.tolist())
for generation_output in output_generate:
self._check_sequence_inside_sequence(force_tokens, generation_output)
def test_constrained_beam_search_generate_dict_output(self):
for model_class in self.all_generative_model_classes:
config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
# disable cache
config.use_cache = False
# It is important set set the eos_token_id to None to ensure that no sequences
# shorter than `max_length` can be generated which could lead to flaky circle ci
# failures if the top `num_return_sequences` beams are all shorter than the longest beam
config.eos_token_id = None
config.forced_eos_token_id = None
model = model_class(config).to(torch_device).eval()
if model.config.is_encoder_decoder:
max_length = 20
logits_process_kwargs, logits_processor = self._get_logits_processor_and_kwargs(
input_ids.shape[-1],
config.eos_token_id,
config.forced_bos_token_id,
config.forced_eos_token_id,
max_length,
)
# Sample constraints
min_id = 3
max_id = model.config.vocab_size
force_tokens = torch.randint(min_id, max_id, (1, 2)).tolist()[0]
constraints = [
PhrasalConstraint(force_tokens),
]
beam_kwargs, beam_scorer = self._get_constrained_beam_scorer_and_kwargs(
input_ids.shape[0], max_length, constraints, num_return_sequences=1
)
output_generate, output_beam_search = self._constrained_beam_search_generate(
model=model,
input_ids=input_ids,
attention_mask=attention_mask,
max_length=max_length,
constrained_beam_scorer=beam_scorer,
constraints=constraints,
beam_kwargs=beam_kwargs,
logits_processor=logits_processor,
logits_process_kwargs=logits_process_kwargs,
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, BeamSearchEncoderDecoderOutput)
self.assertIsInstance(output_generate, BeamSearchEncoderDecoderOutput)
else:
self.assertIsInstance(output_beam_search, BeamSearchDecoderOnlyOutput)
self.assertIsInstance(output_generate, BeamSearchDecoderOnlyOutput)
self.assertListEqual(output_generate.sequences.tolist(), output_beam_search.sequences.tolist())
self.assertTrue(
torch.allclose(output_generate["sequences_scores"], output_beam_search["sequences_scores"], atol=1e-3)
)
self.assertTrue(output_generate["sequences_scores"].shape == (output_generate["sequences"].shape[0],))
self.assertTrue((output_generate["sequences_scores"] < 0).all().item())
for output in (output_beam_search, output_generate):
self._check_outputs(output, input_ids, model.config, num_return_sequences=beam_scorer.num_beams)
def test_contrastive_generate(self):
# check `generate()` and `contrastive_search()` are equal
for model_class in self.all_generative_model_classes:
# won't fix: FSMT and Reformer have a different cache variable type (and format).
if any(model_name in model_class.__name__.lower() for model_name in ["fsmt", "reformer"]):
return
config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
# NOTE: contrastive search only works with cache on at the moment.
if not hasattr(config, "use_cache"):
return
config.use_cache = True
config.is_decoder = True
# test old generation output for backwards compatibility
model = model_class(config).to(torch_device).eval()
output_contrastive, output_generate = self._contrastive_generate(
model=model, input_ids=input_ids, attention_mask=attention_mask, max_length=max_length
)
self.assertListEqual(output_contrastive.tolist(), output_generate.tolist())
def test_contrastive_generate_dict_outputs_use_cache(self):
for model_class in self.all_generative_model_classes:
# won't fix: FSMT and Reformer have a different cache variable type (and format).
if any(model_name in model_class.__name__.lower() for model_name in ["fsmt", "reformer"]):
return
# enable cache
config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
# NOTE: contrastive search only works with cache on at the moment.
if not hasattr(config, "use_cache"):
return
config.use_cache = True
config.is_decoder = True
model = model_class(config).to(torch_device).eval()
output_contrastive, output_generate = self._contrastive_generate(
model=model,
input_ids=input_ids,
attention_mask=attention_mask,
max_length=max_length,
output_scores=True,
output_hidden_states=True,
output_attentions=True,
return_dict_in_generate=True,
)
self.assertListEqual(output_generate.sequences.tolist(), output_contrastive.sequences.tolist())
for output in (output_contrastive, output_generate):
self._check_outputs(output, input_ids, model.config, use_cache=True)
def test_contrastive_generate_low_memory(self):
# Check that choosing 'low_memory' does not change the model output
for model_class in self.all_generative_model_classes:
# won't fix: FSMT, Reformer, gptbigcode, and speech2text have a different cache variable type (and format).
if any(
model_name in model_class.__name__.lower()
for model_name in ["fsmt", "reformer", "gptbigcode", "speech2text"]
):
return
config, input_ids, attention_mask, max_length = self._get_input_ids_and_config(batch_size=1)
# NOTE: contrastive search only works with cache on at the moment.
if not hasattr(config, "use_cache"):
return
config.use_cache = True
config.is_decoder = True
# test output equality of low versus high memory
model = model_class(config).to(torch_device).eval()
low_output = model.generate(
input_ids,
top_k=4,
penalty_alpha=0.6,
low_memory=True,
max_length=max_length,
attention_mask=attention_mask,
)
high_output = model.generate(
input_ids,
top_k=4,
penalty_alpha=0.6,
low_memory=False,
max_length=max_length,
attention_mask=attention_mask,
)
self.assertListEqual(low_output.tolist(), high_output.tolist())
return
@slow # TODO(Joao): remove this. Some models (e.g. data2vec, xcom, roberta) have an error rate between 1 and 10%.
def test_assisted_decoding_matches_greedy_search(self):
# This test ensures that the assisted generation does not introduce output changes over greedy search.
# It breaks the pattern in the tests above, for multiple reasons:
# - assisted_decoding, contrarily to the other methods, can't be called on its own (e.g. needs to
# prepare the assistant encoder outputs in the main generate body);
# - assisted_decoding does not support `use_cache = False`
# - assisted_decoding does not support `batch_size > 1`
for model_class in self.all_generative_model_classes:
# won't fix: FSMT and Reformer have a different cache variable type (and format).
if any(model_name in model_class.__name__.lower() for model_name in ["fsmt", "reformer"]):
return
# may fix in the future: the following models fail with assisted decoding, and need model-specific fixes
if any(
model_name in model_class.__name__.lower()
for model_name in ["bigbirdpegasus", "led", "mega", "speech2text", "git", "prophetnet"]
):
return
# This for loop is a naive and temporary effort to make the test less flaky.
failed = 0
for i in range(10):
# enable cache
config, input_ids, attention_mask, max_length = self._get_input_ids_and_config(batch_size=1)
# NOTE: assisted generation only works with cache on at the moment.
if not hasattr(config, "use_cache"):
return
config.use_cache = True
config.is_decoder = True
model = model_class(config).to(torch_device).eval()
output_greedy = model.generate(
input_ids,
attention_mask=attention_mask,
max_length=max_length,
num_beams=1,
do_sample=False,
output_scores=True,
output_hidden_states=True,
output_attentions=True,
return_dict_in_generate=True,
)
# Note: with assisted generate, if the same model is used as assistant, then all assistant tokens will
# be correct
output_assisted = model.generate(
input_ids,
attention_mask=attention_mask,
max_length=max_length,
num_beams=1,
do_sample=False,
assistant_model=model,
output_scores=True,
output_hidden_states=True,
output_attentions=True,
return_dict_in_generate=True,
)
try:
self.assertListEqual(output_greedy.sequences.tolist(), output_assisted.sequences.tolist())
for output in (output_greedy, output_assisted):
self._check_outputs(output, input_ids, model.config, use_cache=True)
except AssertionError:
failed += 1
if failed > 1:
self.assertListEqual(output_greedy.sequences.tolist(), output_assisted.sequences.tolist())
for output in (output_greedy, output_assisted):
self._check_outputs(output, input_ids, model.config, use_cache=True)
def test_assisted_decoding_sample(self):
# Seeded assisted decoding will not match sample for the same seed, as the forward pass does not return the
# exact same logits (the forward pass of the main model, now with several tokens at once, has causal masking).
for model_class in self.all_generative_model_classes:
# won't fix: FSMT and Reformer have a different cache variable type (and format).
if any(model_name in model_class.__name__.lower() for model_name in ["fsmt", "reformer"]):
return
# may fix in the future: the following models fail with assisted decoding, and need model-specific fixes
if any(
model_name in model_class.__name__.lower()
for model_name in ["bigbirdpegasus", "led", "mega", "speech2text", "git", "prophetnet"]
):
return
# enable cache
config, input_ids, attention_mask, max_length = self._get_input_ids_and_config(batch_size=1)
# NOTE: assisted generation only works with cache on at the moment.
if not hasattr(config, "use_cache"):
return
config.use_cache = True
config.is_decoder = True
model = model_class(config).to(torch_device).eval()
output_assisted = model.generate(
input_ids,
attention_mask=attention_mask,
max_length=max_length,
num_beams=1,
do_sample=True,
assistant_model=model, # triggers assisted decoding
output_scores=True,
output_hidden_states=True,
output_attentions=True,
return_dict_in_generate=True,
)
self._check_outputs(output_assisted, input_ids, model.config, use_cache=True)
def test_generate_with_head_masking(self):
"""Test designed for encoder-decoder models to ensure the attention head masking is used."""
attention_names = ["encoder_attentions", "decoder_attentions", "cross_attentions"]
for model_class in self.all_generative_model_classes:
config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
# We want to test only encoder-decoder models
if not config.is_encoder_decoder:
continue
model = model_class(config).to(torch_device)
head_masking = {
"head_mask": torch.zeros(config.encoder_layers, config.encoder_attention_heads, device=torch_device),
"decoder_head_mask": torch.zeros(
config.decoder_layers, config.decoder_attention_heads, device=torch_device
),
"cross_attn_head_mask": torch.zeros(
config.decoder_layers, config.decoder_attention_heads, device=torch_device
),
}
signature = inspect.signature(model.forward)
# We want to test only models where encoder/decoder head masking is implemented
if not set(head_masking.keys()) < {*signature.parameters.keys()}:
continue
for attn_name, (name, mask) in zip(attention_names, head_masking.items()):
out = model.generate(
input_ids,
attention_mask=attention_mask,
num_beams=1,
output_attentions=True,
return_dict_in_generate=True,
remove_invalid_values=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([w.sum().item() for w in attn_weights]), 0.0)
def test_left_padding_compatibility(self):
# The check done in this test is fairly difficult -- depending on the model architecture, passing the right
# position index for the position embeddings can still result in a different output, due to numerical masking.
# On the other hand, for some types of position embeddings, an incorrect position index can have a minimal
# impact on the output.
# There are two tricks employed to check whether left-padding compatibility is in place:
# 1 - To reduce the negative impact of the numerical attention mask on a correct position index, we set the
# padding size to 1.
# 2 - To reduce the chance of false positives (i.e. passing when it should be failing), we run the check
# multiple times with random inputs, and it has to pass with all of them.
# NOTE: because of 2), there is some chance of false positives in this test.
for model_class in self.all_generative_model_classes:
config, _, _, _ = self._get_input_ids_and_config()
if config.is_encoder_decoder:
continue # skip for encoder-decoder models -- they don't need left-padding compatibility
model = model_class(config).to(torch_device).eval()
signature = inspect.signature(model.forward).parameters.keys()
no_failures = True
for _ in range(10): # there may be false positives with 10 runs, we rely on the CI to catch the flakiness
_, input_ids, attention_mask, _ = self._get_input_ids_and_config()
model_kwargs = {"input_ids": input_ids, "attention_mask": attention_mask}
if "position_ids" in signature:
position_ids = torch.cumsum(attention_mask, dim=-1) - 1
position_ids.masked_fill_(attention_mask == 0, 1)
model_kwargs["position_ids"] = position_ids
next_logits_wo_padding = model(**model_kwargs).logits[:, -1, :]
pad_size = (input_ids.shape[0], 1)
padding = torch.ones(pad_size, dtype=input_ids.dtype, device=torch_device) * config.pad_token_id
padded_input_ids = torch.cat((padding, input_ids), dim=1)
padded_attention_mask = torch.cat((torch.zeros_like(padding), attention_mask), dim=1)
model_kwargs = {"input_ids": padded_input_ids, "attention_mask": padded_attention_mask}
if "position_ids" in signature:
position_ids = torch.cumsum(padded_attention_mask, dim=-1) - 1
position_ids.masked_fill_(padded_attention_mask == 0, 1)
model_kwargs["position_ids"] = position_ids
next_logits_with_padding = model(**model_kwargs).logits[:, -1, :]
if not torch.allclose(next_logits_wo_padding, next_logits_with_padding, atol=1e-7):
no_failures = False
break
self.assertTrue(no_failures)
def test_past_key_values_format(self):
# Test that the KV cache is formatted correctly. Exceptions need to explicitly overwrite this test. Having a
# standard KV cache format is important for a consistent API (and for advanced generation methods).
for model_class in self.all_generative_model_classes:
config, inputs = self.model_tester.prepare_config_and_inputs_for_common()
# If it doesn't support cache, pass the test
if not hasattr(config, "use_cache"):
return
model = model_class(config).to(torch_device)
if "use_cache" not in inputs:
inputs["use_cache"] = True
outputs = model(**inputs)
# If "past_key_values" is not returned, pass the test (e.g. RWKV uses a different cache name and format)
if "past_key_values" not in outputs:
return
num_hidden_layers = (
getattr(config, "decoder_layers", None)
or getattr(config, "num_decoder_layers", None)
or config.num_hidden_layers
)
num_attention_heads = getattr(config, "decoder_attention_heads", config.num_attention_heads)
embed_dim = getattr(config, "d_model", config.hidden_size)
per_head_embed_dim = embed_dim // num_attention_heads
past_kv = outputs["past_key_values"]
self.assertEqual(len(past_kv), num_hidden_layers)
# Encoder-Decoder checks
if config.is_encoder_decoder:
encoder_num_attention_heads = config.encoder_attention_heads
encoder_per_head_embed_dim = embed_dim // encoder_num_attention_heads
batch_size, seq_length = inputs["decoder_input_ids"].shape
for i in range(num_hidden_layers):
self.assertEqual(len(past_kv[i]), 4) # K V for the decoder + K V for the encoder = 4
self.assertEqual(
past_kv[i][0].shape, (batch_size, num_attention_heads, seq_length, per_head_embed_dim)
)
self.assertEqual(
past_kv[i][1].shape, (batch_size, num_attention_heads, seq_length, per_head_embed_dim)
)
# The sequence length for the encoder K V depends on the model. Since it is not manipulated in
# autoregressive generation, I'm keeping the test general and not checking the 3rd dim
self.assertEqual(
(past_kv[i][2].shape[0], past_kv[i][2].shape[1], past_kv[i][2].shape[3]),
(batch_size, encoder_num_attention_heads, encoder_per_head_embed_dim),
)
self.assertEqual(
(past_kv[i][3].shape[0], past_kv[i][3].shape[1], past_kv[i][3].shape[3]),
(batch_size, encoder_num_attention_heads, encoder_per_head_embed_dim),
)
# Decoder-only checks
else:
# TODO: this line is only needed because of imagegpt, where "pixel_values" = "input_ids". Fix the
# tests in imagegpt such that `prepare_config_and_inputs_for_common` returns the later (and the other
# tests use it)
key = "input_ids" if "input_ids" in inputs else "pixel_values"
batch_size, seq_length = inputs[key].shape
for i in range(num_hidden_layers):
self.assertEqual(len(past_kv[0]), 2) # K V for the decoder = 2
self.assertEqual(
past_kv[i][0].shape, (batch_size, num_attention_heads, seq_length, per_head_embed_dim)
)
self.assertEqual(
past_kv[i][1].shape, (batch_size, num_attention_heads, seq_length, per_head_embed_dim)
)
def test_generate_from_inputs_embeds_decoder_only(self):
# When supported, tests that the decoder model can generate from `inputs_embeds` instead of `input_ids`
# if fails, you should probably update the `prepare_inputs_for_generation` function
for model_class in self.all_generative_model_classes:
config, input_ids, _, _ = self._get_input_ids_and_config()
# Ignore:
# a) eos (to always output 20 tokens) and pad (so we don't try to infer the attn mask from the input_ids,
# which would cause a mismatch),
config.pad_token_id = config.eos_token_id = -1
# b) embedding scaling, the scaling factor applied after embeding from input_ids (requires knowledge of the
# variable that holds the scaling factor, which is model-dependent)
if hasattr(config, "scale_embedding"):
config.scale_embedding = False
# This test is for decoder-only models (encoder-decoder models have native input embeddings support in the
# decoder)
if config.is_encoder_decoder:
continue
# Skip models without explicit support
model = model_class(config).to(torch_device).eval()
if "inputs_embeds" not in inspect.signature(model.prepare_inputs_for_generation).parameters.keys():
continue
# Traditional way of generating text
outputs_from_ids = model.generate(input_ids)
self.assertEqual(outputs_from_ids.shape, (2, 20))
# Same thing, but from input embeddings (`input_ids` is passed so the prompt is present in the output)
inputs_embeds = model.get_input_embeddings()(input_ids)
outputs_from_embeds = model.generate(input_ids, inputs_embeds=inputs_embeds)
self.assertListEqual(outputs_from_ids.tolist(), outputs_from_embeds.tolist())
# But if we pass different inputs_embeds, we should get different outputs
torch.manual_seed(0)
random_embeds = torch.rand_like(inputs_embeds)
outputs_from_rand_embeds = model.generate(input_ids, inputs_embeds=random_embeds)
with self.assertRaises(AssertionError):
self.assertListEqual(outputs_from_rand_embeds.tolist(), outputs_from_embeds.tolist())
# input_ids is not a required input -- if we don't pass it, the newly generated tokens will be the same
outputs_from_embeds_wo_ids = model.generate(
inputs_embeds=inputs_embeds, max_new_tokens=20 - inputs_embeds.shape[1]
)
self.assertListEqual(
outputs_from_embeds[:, inputs_embeds.shape[1] :].tolist(),
outputs_from_embeds_wo_ids[:, 1:].tolist(),
)
def _check_outputs(self, output, input_ids, config, use_cache=False, num_return_sequences=1):
batch_size, seq_length = input_ids.shape
num_sequences_in_output = batch_size * num_return_sequences
gen_len = (
output.sequences.shape[-1] - 1 if config.is_encoder_decoder else output.sequences.shape[-1] - seq_length
)
# scores
self._check_scores(num_sequences_in_output, output.scores, length=gen_len, config=config)
# Attentions
if config.is_encoder_decoder:
# encoder
self._check_encoder_attention_for_generate(output.encoder_attentions, batch_size, config, seq_length)
# decoder
self._check_attentions_for_generate(
num_sequences_in_output,
output.decoder_attentions,
min_length=1,
max_length=output.sequences.shape[-1],
config=config,
use_cache=use_cache,
)
else:
# if use_cache first input is equal to no use_cache, so skip here
attentions = output.attentions if not use_cache else output.attentions[1:]
min_length = seq_length if not use_cache else seq_length + 1
self._check_attentions_for_generate(
num_sequences_in_output,
attentions=attentions,
min_length=min_length,
max_length=output.sequences.shape[-1],
config=config,
use_cache=use_cache,
)
# Hidden States
if config.is_encoder_decoder:
# encoder
self._check_encoder_hidden_states_for_generate(
output.encoder_hidden_states, batch_size, config, seq_length
)
# decoder
self._check_hidden_states_for_generate(
num_sequences_in_output,
output.decoder_hidden_states,
min_length=1,
max_length=output.sequences.shape[-1],
config=config,
use_cache=use_cache,
)
else:
# if use_cache first input is equal to no use_cache, so skip here
hidden_states = output.hidden_states if not use_cache else output.hidden_states[1:]
min_length = seq_length if not use_cache else seq_length + 1
self._check_hidden_states_for_generate(
num_sequences_in_output,
hidden_states,
min_length=min_length,
max_length=output.sequences.shape[-1],
config=config,
use_cache=use_cache,
)
def _check_scores(self, batch_size, scores, length, config):
expected_shape = (batch_size, config.vocab_size)
self.assertIsInstance(scores, tuple)
self.assertEqual(len(scores), length)
self.assertListEqual([iter_scores.shape for iter_scores in scores], [expected_shape] * len(scores))
def _check_attentions_for_generate(
self, batch_size, attentions, min_length, max_length, config, use_cache=False, num_beam_groups=1
):
self.assertIsInstance(attentions, tuple)
self.assertListEqual(
[isinstance(iter_attentions, tuple) for iter_attentions in attentions], [True] * len(attentions)
)
self.assertEqual(len(attentions), (max_length - min_length) * num_beam_groups)
for idx, iter_attentions in enumerate(attentions):
tgt_len = min_length + idx if not use_cache else 1
src_len = min_length + idx
expected_shape = (
batch_size * num_beam_groups,
config.num_attention_heads,
tgt_len,
src_len,
)
# check attn size
self.assertListEqual(
[layer_attention.shape for layer_attention in iter_attentions], [expected_shape] * len(iter_attentions)
)
def _check_encoder_attention_for_generate(self, attentions, batch_size, config, seq_length):
encoder_expected_shape = (batch_size, config.num_attention_heads, seq_length, seq_length)
self.assertIsInstance(attentions, tuple)
self.assertListEqual(
[layer_attentions.shape for layer_attentions in attentions],
[encoder_expected_shape] * len(attentions),
)
def _check_hidden_states_for_generate(
self, batch_size, hidden_states, min_length, max_length, config, use_cache=False, num_beam_groups=1
):
self.assertIsInstance(hidden_states, tuple)
self.assertListEqual(
[isinstance(iter_hidden_states, tuple) for iter_hidden_states in hidden_states],
[True] * len(hidden_states),
)
self.assertEqual(len(hidden_states), (max_length - min_length) * num_beam_groups)
for idx, iter_hidden_states in enumerate(hidden_states):
seq_len = min_length + idx if not use_cache else 1
expected_shape = (batch_size * num_beam_groups, seq_len, config.hidden_size)
# check hidden size
self.assertListEqual(
[layer_hidden_states.shape for layer_hidden_states in iter_hidden_states],
[expected_shape] * len(iter_hidden_states),
)
def _check_encoder_hidden_states_for_generate(self, hidden_states, batch_size, config, seq_length):
encoder_expected_shape = (batch_size, seq_length, config.hidden_size)
self.assertIsInstance(hidden_states, tuple)
self.assertListEqual(
[layer_hidden_states.shape for layer_hidden_states in hidden_states],
[encoder_expected_shape] * len(hidden_states),
)
def _check_sequence_inside_sequence(self, tensor_1, tensor_2):
# check if tensor_1 inside tensor_2 or tensor_2 inside tensor_1.
# set to same device. we don't care what device.
if not isinstance(tensor_1, list):
tensor_1 = tensor_1.cpu().tolist()
if not isinstance(tensor_2, list):
tensor_2 = tensor_2.cpu().tolist()
in_order = len(tensor_1) <= len(tensor_2)
longer = tensor_2 if in_order else tensor_1
shorter = tensor_1 if in_order else tensor_2
flag = False
chunk_size = len(shorter)
for chunk_idx in range(len(longer) - chunk_size + 1):
subseq = longer[chunk_idx : chunk_idx + chunk_size]
if subseq == shorter:
flag = True
break
self.assertTrue(flag)
@require_torch
class UtilsFunctionsTest(unittest.TestCase):
# tests whether the top_k_top_p function behaves as expected
def test_top_k_top_p_filtering(self):
logits = torch.tensor(
[
[
8.2220991, # 3rd highest value; idx. 0
-0.5620044,
5.23229752,
4.0386393,
-6.8798378,
-0.54785802,
-3.2012153,
2.92777176,
1.88171953,
7.35341276,
8.43207833, # 2nd highest value; idx. 10
-9.85711836,
-5.96209236,
-1.13039161,
-7.1115294,
-0.8369633,
-5.3186408,
7.06427407,
0.81369344,
-0.82023817,
-5.9179796,
0.58813443,
-6.99778438,
4.71551189,
-0.18771637,
7.44020759, # 4th highest value; idx. 25
9.38450987, # 1st highest value; idx. 26
2.12662941,
-9.32562038,
2.35652522,
], # cummulative prob of 4 highest values <= 0.6
[
0.58425518,
4.53139238,
-5.57510464,
-6.28030699,
-7.19529503,
-4.02122551,
1.39337037,
-6.06707057,
1.59480517,
-9.643119,
0.03907799,
0.67231762,
-8.88206726,
6.27115922, # 4th highest value; idx. 13
2.28520723,
4.82767506,
4.30421368,
8.8275313, # 2nd highest value; idx. 17
5.44029958,
-4.4735794,
7.38579536, # 3rd highest value; idx. 20
-2.91051663,
2.61946077,
-2.5674762,
-9.48959302,
-4.02922645,
-1.35416918,
9.67702323, # 1st highest value; idx. 27
-5.89478553,
1.85370467,
], # cummulative prob of 4 highest values <= 0.6
],
dtype=torch.float,
device=torch_device,
)
non_inf_expected_idx = torch.tensor(
[[0, 0], [0, 10], [0, 25], [0, 26], [1, 13], [1, 17], [1, 20], [1, 27]],
dtype=torch.long,
device=torch_device,
) # expected non filtered idx as noted above
non_inf_expected_output = torch.tensor(
[
8.2221,
8.4321,
7.4402,
9.3845,
6.2712,
8.8275,
7.3858,
9.6770,
], # expected non filtered values as noted above
dtype=torch.float,
device=torch_device,
)
output = top_k_top_p_filtering(logits, top_k=10, top_p=0.6, min_tokens_to_keep=4)
non_inf_output = output[output != -float("inf")].to(device=torch_device)
non_inf_idx = (output != -float("inf")).nonzero().to(device=torch_device)
self.assertTrue(torch.allclose(non_inf_expected_output, non_inf_output, atol=1e-12))
self.assertTrue(torch.all(torch.eq(non_inf_expected_idx, non_inf_idx)))
# tests whether the function uses filter_value instead of default -inf
def test_top_k_top_p_filtering_with_filter_value(self):
logits = torch.tensor(
[
[
1,
1,
1,
0.99, # get filtered by top-p filtering
0.98, # get filtered by top-k filtering
]
],
dtype=torch.float,
device=torch_device,
)
expected_output = torch.tensor(
[[1, 1, 1, 0, 0]],
dtype=torch.float,
device=torch_device,
)
output = top_k_top_p_filtering(logits, top_k=4, top_p=0.5, filter_value=0.0)
self.assertTrue(torch.allclose(expected_output, output, atol=1e-12))
@require_torch
class GenerationIntegrationTests(unittest.TestCase, GenerationIntegrationTestsMixin):
# setting framework_dependent_parameters needs to be gated, just like its contents' imports
if is_torch_available():
framework_dependent_parameters = {
"AutoModelForCausalLM": AutoModelForCausalLM,
"AutoModelForSpeechSeq2Seq": AutoModelForSpeechSeq2Seq,
"AutoModelForSeq2SeqLM": AutoModelForSeq2SeqLM,
"AutoModelForVision2Seq": AutoModelForVision2Seq,
"LogitsProcessorList": LogitsProcessorList,
"MinLengthLogitsProcessor": MinLengthLogitsProcessor,
"create_tensor_fn": torch.tensor,
"floats_tensor": floats_tensor,
"return_tensors": "pt",
}
@slow
def test_diverse_beam_search(self):
# PT-only test: TF doesn't have a diverse beam search implementation
article = """Justin Timberlake and Jessica Biel, welcome to parenthood.
The celebrity couple announced the arrival of their son, Silas Randall Timberlake, in statements to People.
"Silas was the middle name of Timberlake's maternal grandfather Bill Bomar, who died in 2012, while Randall is the musician's own middle name, as well as his father's first," People reports.
The couple announced the pregnancy in January, with an Instagram post. It is the first baby for both."""
bart_tokenizer = BartTokenizer.from_pretrained("facebook/bart-large-cnn")
bart_model = BartForConditionalGeneration.from_pretrained("facebook/bart-large-cnn").to(torch_device)
input_ids = bart_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)
outputs = bart_model.generate(
input_ids,
num_beams=4,
num_return_sequences=2,
num_beam_groups=4,
diversity_penalty=2.0,
remove_invalid_values=True,
)
generated_text = bart_tokenizer.batch_decode(outputs, skip_special_tokens=True)
self.assertListEqual(
generated_text,
[
"The couple announced the birth of their son, Silas Randall Timberlake, in a statement. Silas was the"
" middle name of Timberlake's maternal grandfather Bill Bomar. Randall is the musician's own middle"
" name, as well as his father's first. It is the first baby for both of them.",
"Justin Timberlake and Jessica Biel have a son. The baby is named Silas Randall Timberlake. It is the"
" first child for both. The couple announced the pregnancy in January. The name Silas is the middle"
" name of Timberlake's maternal grandfather. It's also his own middle name.",
],
)
def test_max_length_backward_compat_greedy(self):
# PT-only test: TF doesn't have StoppingCriteria
article = """Justin Timberlake and Jessica Biel, welcome to parenthood."""
bart_tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
bart_model = BartForConditionalGeneration.from_pretrained("hf-internal-testing/tiny-random-bart").to(
torch_device
)
input_ids = bart_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)
max_length = 20
input_ids = input_ids.expand(2, -1)
model_kwargs = bart_model._prepare_encoder_decoder_kwargs_for_generation(input_ids, {})
input_ids, model_kwargs = bart_model._prepare_decoder_input_ids_for_generation(
batch_size=input_ids.shape[0],
model_input_name=bart_model.main_input_name,
model_kwargs=model_kwargs,
decoder_start_token_id=bart_model.config.decoder_start_token_id,
bos_token_id=bart_model.config.bos_token_id,
)
with self.assertWarns(UserWarning):
bart_model.greedy_search(
input_ids,
max_length=max_length,
pad_token_id=bart_model.config.pad_token_id,
eos_token_id=bart_model.config.eos_token_id,
**model_kwargs,
)
def test_max_length_backward_compat_sample(self):
# PT-only test: TF doesn't have StoppingCriteria
article = """Justin Timberlake and Jessica Biel, welcome to parenthood."""
bart_tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
bart_model = BartForConditionalGeneration.from_pretrained("hf-internal-testing/tiny-random-bart").to(
torch_device
)
input_ids = bart_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)
max_length = 20
input_ids = input_ids.expand(2, -1)
model_kwargs = bart_model._prepare_encoder_decoder_kwargs_for_generation(input_ids, {})
input_ids, model_kwargs = bart_model._prepare_decoder_input_ids_for_generation(
batch_size=input_ids.shape[0],
model_input_name=bart_model.main_input_name,
model_kwargs=model_kwargs,
decoder_start_token_id=bart_model.config.decoder_start_token_id,
bos_token_id=bart_model.config.bos_token_id,
)
with torch.no_grad():
with self.assertWarns(UserWarning):
bart_model.sample(
input_ids,
max_length=max_length,
pad_token_id=bart_model.config.pad_token_id,
eos_token_id=bart_model.config.eos_token_id,
**model_kwargs,
)
def test_max_length_backward_compat_beam_search(self):
# PT-only test: TF doesn't have StoppingCriteria
article = """Justin Timberlake and Jessica Biel, welcome to parenthood."""
bart_tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
bart_model = BartForConditionalGeneration.from_pretrained("hf-internal-testing/tiny-random-bart").to(
torch_device
)
input_ids = bart_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)
batch_size = 1
max_length = 20
num_beams = 2
input_ids = input_ids.expand(2, -1)
model_kwargs = bart_model._prepare_encoder_decoder_kwargs_for_generation(input_ids, {})
input_ids, model_kwargs = bart_model._prepare_decoder_input_ids_for_generation(
batch_size=input_ids.shape[0],
model_input_name=bart_model.main_input_name,
model_kwargs=model_kwargs,
decoder_start_token_id=bart_model.config.decoder_start_token_id,
bos_token_id=bart_model.config.bos_token_id,
)
beam_scorer = BeamSearchScorer(
batch_size=batch_size,
num_beams=num_beams,
device=torch_device,
)
with self.assertWarns(UserWarning):
_ = bart_model.beam_search(
input_ids, num_beams=num_beams, max_length=max_length, beam_scorer=beam_scorer, **model_kwargs
)
def test_max_length_backward_compat_group_beam_search(self):
# PT-only test: TF doesn't have StoppingCriteria & group beam search
article = """Justin Timberlake and Jessica Biel, welcome to parenthood."""
bart_tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
bart_model = BartForConditionalGeneration.from_pretrained("hf-internal-testing/tiny-random-bart").to(
torch_device
)
input_ids = bart_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)
batch_size = 1
max_length = 20
num_beams = 6
num_beam_groups = 3
num_return_sequences = num_beams * batch_size
input_ids = input_ids.expand(6, -1)
model_kwargs = bart_model._prepare_encoder_decoder_kwargs_for_generation(input_ids, {})
input_ids, model_kwargs = bart_model._prepare_decoder_input_ids_for_generation(
batch_size=input_ids.shape[0],
model_input_name=bart_model.main_input_name,
model_kwargs=model_kwargs,
decoder_start_token_id=bart_model.config.decoder_start_token_id,
bos_token_id=bart_model.config.bos_token_id,
)
diverse_beam_scorer = BeamSearchScorer(
batch_size=batch_size,
num_beams=num_beams,
device=torch_device,
num_beam_hyps_to_keep=num_return_sequences,
num_beam_groups=num_beam_groups,
)
with self.assertWarns(UserWarning):
bart_model.group_beam_search(
input_ids, diverse_beam_scorer, num_beams=num_beams, max_length=max_length, **model_kwargs
)
def test_max_length_warning_if_different(self):
# PT-only test: TF doesn't have StoppingCriteria
article = """Justin Timberlake and Jessica Biel, welcome to parenthood."""
bart_tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
bart_model = BartForConditionalGeneration.from_pretrained("hf-internal-testing/tiny-random-bart").to(
torch_device
)
input_ids = bart_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)
batch_size = 1
max_length = 20
num_beams = 6
num_beam_groups = 3
num_return_sequences = num_beams * batch_size
stopping_criteria_max_length = 18
stopping_criteria = StoppingCriteriaList([MaxLengthCriteria(max_length=stopping_criteria_max_length)])
# Greedy
input_ids = input_ids.expand(6, -1)
model_kwargs = bart_model._prepare_encoder_decoder_kwargs_for_generation(input_ids, {})
input_ids, model_kwargs = bart_model._prepare_decoder_input_ids_for_generation(
batch_size=input_ids.shape[0],
model_input_name=bart_model.main_input_name,
model_kwargs=model_kwargs,
decoder_start_token_id=bart_model.config.decoder_start_token_id,
bos_token_id=bart_model.config.bos_token_id,
)
with self.assertWarns(UserWarning):
bart_model.greedy_search(
input_ids,
max_length=max_length,
pad_token_id=bart_model.config.pad_token_id,
stopping_criteria=stopping_criteria,
eos_token_id=bart_model.config.eos_token_id,
**model_kwargs,
)
# Sample
with self.assertWarns(UserWarning):
with torch.no_grad():
bart_model.sample(
input_ids,
max_length=max_length,
stopping_criteria=stopping_criteria,
pad_token_id=bart_model.config.pad_token_id,
eos_token_id=bart_model.config.eos_token_id,
**model_kwargs,
)
# Beam
beam_scorer = BeamSearchScorer(
batch_size=batch_size,
num_beams=num_beams,
device=torch_device,
)
with self.assertWarns(UserWarning):
with torch.no_grad():
bart_model.beam_search(
input_ids,
num_beams=num_beams,
stopping_criteria=stopping_criteria,
max_length=max_length,
beam_scorer=beam_scorer,
**model_kwargs,
)
# Grouped beam search
diverse_beam_scorer = BeamSearchScorer(
batch_size=batch_size,
num_beams=num_beams,
device=torch_device,
num_beam_hyps_to_keep=num_return_sequences,
num_beam_groups=num_beam_groups,
)
with self.assertWarns(UserWarning):
bart_model.group_beam_search(
input_ids,
diverse_beam_scorer,
stopping_criteria=stopping_criteria,
num_beams=num_beams,
max_length=max_length,
**model_kwargs,
)
def test_custom_stopping_criteria_overload_error(self):
# PT-only test: TF doesn't have StoppingCriteria
article = """Justin Timberlake and Jessica Biel, welcome to parenthood."""
bart_tokenizer = BartTokenizer.from_pretrained("sshleifer/bart-tiny-random")
bart_model = BartForConditionalGeneration.from_pretrained("sshleifer/bart-tiny-random").to(torch_device)
input_ids = bart_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)
stopping_criteria = StoppingCriteriaList()
stopping_criteria.append(MaxLengthCriteria(max_length=42))
with self.assertRaises(ValueError):
bart_model.generate(input_ids, stopping_criteria=stopping_criteria)
with self.assertRaises(ValueError):
bart_model.generate(input_ids, stopping_criteria=stopping_criteria, max_length=32)
def test_custom_stopping_criteria(self):
# PT-only test: TF doesn't have StoppingCriteria
article = """Justin Timberlake and Jessica Biel, welcome to parenthood."""
bart_tokenizer = BartTokenizer.from_pretrained("sshleifer/bart-tiny-random")
bart_model = BartForConditionalGeneration.from_pretrained("sshleifer/bart-tiny-random").to(torch_device)
input_ids = bart_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)
class DummyCriteria(StoppingCriteria):
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
return input_ids.shape[-1] >= 20
stopping_criteria = StoppingCriteriaList()
stopping_criteria.append(DummyCriteria())
self.assertEqual(
list(bart_model.generate(input_ids, stopping_criteria=stopping_criteria, max_length=22).shape),
[1, 20],
)
self.assertEqual(
list(bart_model.generate(input_ids, stopping_criteria=stopping_criteria, max_length=18).shape),
[1, 18],
)
def test_stop_sequence_stopping_criteria(self):
# PT-only test: TF doesn't have StoppingCriteria
prompt = """Hello I believe in"""
generator = pipeline("text-generation", model="hf-internal-testing/tiny-random-bart")
output = generator(prompt)
self.assertEqual(
output,
[
{
"generated_text": (
"Hello I believe in in in number number number number number number number number number"
)
}
],
)
output = generator(prompt, stop_sequence=" number")
self.assertEqual(output, [{"generated_text": "Hello I believe in in in number"}])
def test_generate_non_nlp_input_ids_as_kwarg(self):
# PT-only test: AFAIK there's no non-NLP model architecture in TF that supports `input_ids` as its only input
model = ImageGPTForCausalImageModeling.from_pretrained(
"hf-internal-testing/tiny-random-imagegpt", max_length=10
).to(torch_device)
input_ids = ids_tensor((3, 5), vocab_size=10)
output_sequences_kwargs = model.generate(input_ids=input_ids).cpu()
output_sequences = model.generate(input_ids).cpu()
self.assertListEqual(output_sequences.tolist(), output_sequences_kwargs.tolist())
self.assertEqual(output_sequences.shape, (3, 10))
def test_generate_input_values_as_encoder_kwarg(self):
# PT-only test: AFAIK there's no generate-capable architecture in TF that supports `input_values` as its input
input_values = floats_tensor((2, 250))
model = SpeechEncoderDecoderModel.from_pretrained("hf-internal-testing/tiny-random-speech-encoder-decoder")
model = model.to(torch_device)
output_sequences_kwargs = model.generate(input_values=input_values, max_length=5).cpu()
output_sequences = model.generate(input_values, max_length=5).cpu()
self.assertListEqual(output_sequences.tolist(), output_sequences_kwargs.tolist())
self.assertEqual(output_sequences.shape, (2, 5))
def test_transition_scores_group_beam_search_encoder_decoder(self):
# PT-only test: TF doesn't have group beam search
articles = [
"Justin Timberlake and Jessica Biel, welcome to parenthood.",
"Michael Phelps is arguably the most decorated Olympian of all time.",
]
tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
model = BartForConditionalGeneration.from_pretrained(
"hf-internal-testing/tiny-random-bart",
max_length=10,
num_beams=2,
num_beam_groups=2,
num_return_sequences=2,
diversity_penalty=1.0,
eos_token_id=None,
return_dict_in_generate=True,
output_scores=True,
length_penalty=0.0,
)
model = model.to(torch_device)
input_ids = tokenizer(articles, return_tensors="pt", padding=True).input_ids.to(torch_device)
outputs = model.generate(input_ids=input_ids)
transition_scores = model.compute_transition_scores(outputs.sequences, outputs.scores, outputs.beam_indices)
transition_scores_sum = transition_scores.sum(-1)
self.assertTrue(torch.allclose(transition_scores_sum, outputs.sequences_scores, atol=1e-3))
@slow
def test_beam_search_example_integration(self):
# PT-only test: TF doesn't have a BeamSearchScorer
# exactly the example provided in the docstrings of beam search, which previously
# failed after directly copying from it. Refer to PR #15555
tokenizer = AutoTokenizer.from_pretrained("t5-base")
model = AutoModelForSeq2SeqLM.from_pretrained("t5-base")
encoder_input_str = "translate English to German: How old are you?"
encoder_input_ids = tokenizer(encoder_input_str, return_tensors="pt").input_ids
# lets run beam search using 3 beams
num_beams = 3
# define decoder start token ids
input_ids = torch.ones((num_beams, 1), device=model.device, dtype=torch.long)
input_ids = input_ids * model.config.decoder_start_token_id
# add encoder_outputs to model keyword arguments
model_kwargs = {
"encoder_outputs": model.get_encoder()(
encoder_input_ids.repeat_interleave(num_beams, dim=0), return_dict=True
)
}
# instantiate beam scorer
beam_scorer = BeamSearchScorer(
batch_size=1,
num_beams=num_beams,
device=model.device,
)
# instantiate logits processors
logits_processor = LogitsProcessorList(
[
MinLengthLogitsProcessor(5, eos_token_id=model.config.eos_token_id),
]
)
outputs = model.beam_search(input_ids, beam_scorer, logits_processor=logits_processor, **model_kwargs)
outputs = tokenizer.batch_decode(outputs, skip_special_tokens=True)
self.assertListEqual(outputs, ["Wie alt bist du?"])
@slow
def test_constrained_beam_search(self):
# PT-only test: TF doesn't have constrained beam search
model = GPT2LMHeadModel.from_pretrained("gpt2").to(torch_device)
tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
force_tokens = tokenizer("scared", add_prefix_space=True, add_special_tokens=False).input_ids
force_tokens_2 = tokenizer("big weapons", add_prefix_space=True, add_special_tokens=False).input_ids
constraints = [
PhrasalConstraint(force_tokens),
PhrasalConstraint(force_tokens_2),
]
starting_text = ["The soldiers were not prepared and"]
input_ids = tokenizer(starting_text, return_tensors="pt").input_ids.to(torch_device)
outputs = model.generate(
input_ids,
constraints=constraints,
num_beams=10,
num_return_sequences=1,
no_repeat_ngram_size=1,
max_length=30,
remove_invalid_values=True,
)
generated_text = tokenizer.batch_decode(outputs, skip_special_tokens=True)
self.assertListEqual(
generated_text,
[
"The soldiers were not prepared and didn't know what to do. They had no idea how they would react if"
" the enemy attacked them, big weapons scared"
],
)
@slow
def test_constrained_beam_search_mixed(self):
# PT-only test: TF doesn't have constrained beam search
model = GPT2LMHeadModel.from_pretrained("gpt2").to(torch_device)
tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
force_phrase = tokenizer("scared", add_prefix_space=True, add_special_tokens=False).input_ids
flexible_phrases = tokenizer(
["scream", "screams", "screaming", "screamed"], add_prefix_space=True, add_special_tokens=False
).input_ids
constraints = [
PhrasalConstraint(force_phrase),
DisjunctiveConstraint(flexible_phrases),
]
starting_text = ["The soldiers", "The child"]
input_ids = tokenizer(starting_text, return_tensors="pt").input_ids.to(torch_device)
outputs = model.generate(
input_ids,
constraints=constraints,
num_beams=10,
num_return_sequences=1,
no_repeat_ngram_size=1,
# max_length=20,
remove_invalid_values=True,
)
generated_text = tokenizer.batch_decode(outputs, skip_special_tokens=True)
self.assertListEqual(
generated_text,
[
"The soldiers, who had been stationed at the base for more than a year before being evacuated"
" screaming scared",
"The child was taken to a local hospital where he died.\n 'I don't think screaming scared",
],
)
@slow
def test_constrained_beam_search_mixed_mixin(self):
# PT-only test: TF doesn't have constrained beam search
model = GPT2LMHeadModel.from_pretrained("gpt2").to(torch_device)
tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
force_word = "scared"
force_flexible = ["scream", "screams", "screaming", "screamed"]
force_words_ids = [
tokenizer([force_word], add_prefix_space=True, add_special_tokens=False).input_ids,
tokenizer(force_flexible, add_prefix_space=True, add_special_tokens=False).input_ids,
]
starting_text = ["The soldiers", "The child"]
input_ids = tokenizer(starting_text, return_tensors="pt").input_ids.to(torch_device)
outputs = model.generate(
input_ids,
force_words_ids=force_words_ids,
num_beams=10,
num_return_sequences=1,
no_repeat_ngram_size=1,
remove_invalid_values=True,
)
generated_text = tokenizer.batch_decode(outputs, skip_special_tokens=True)
self.assertListEqual(
generated_text,
[
"The soldiers, who had been stationed at the base for more than a year before being evacuated"
" screaming scared",
"The child was taken to a local hospital where he died.\n 'I don't think screaming scared",
],
)
@slow
def test_cfg_mixin(self):
model = GPT2LMHeadModel.from_pretrained("gpt2").to(torch_device)
tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
input = tokenizer(["The dragon flew over Paris,"], return_tensors="pt", return_attention_mask=True)
input["input_ids"] = input["input_ids"].to(torch_device)
input["attention_mask"] = input["attention_mask"].to(torch_device)
outputs = model.generate(**input, max_new_tokens=32, guidance_scale=1.5)
generated_text = tokenizer.batch_decode(outputs, skip_special_tokens=True)
self.assertListEqual(
generated_text,
[
"The dragon flew over Paris, landing in the Rue de la Bastille. The crowd was so excited "
'that they had to leave the city.\n\n"We\'re going to Paris!"\n'
],
)
neg = tokenizer(["France,"], return_tensors="pt", return_attention_mask=True)
neg["input_ids"] = neg["input_ids"].to(torch_device)
neg["attention_mask"] = neg["attention_mask"].to(torch_device)
outputs = model.generate(
**input,
max_new_tokens=32,
guidance_scale=1.5,
negative_prompt_ids=neg["input_ids"],
negative_prompt_attention_mask=neg["attention_mask"],
)
generated_text = tokenizer.batch_decode(outputs, skip_special_tokens=True)
self.assertListEqual(
generated_text,
[
'The dragon flew over Paris, landing on the pavement.\n\n"Paris!"\n\n"Paris!"\n\n"'
'Paris!"\n\n"Paris!"\n\n"Paris!"\n\n'
],
)
@slow
def test_constrained_beam_search_example_translation_mixin(self):
# PT-only test: TF doesn't have constrained beam search
tokenizer = AutoTokenizer.from_pretrained("t5-base")
model = AutoModelForSeq2SeqLM.from_pretrained("t5-base")
encoder_input_str = "translate English to German: How old are you?"
force_words = ["sind"]
input_ids = tokenizer(encoder_input_str, return_tensors="pt").input_ids
force_words_ids = tokenizer(force_words, add_special_tokens=False).input_ids
outputs = model.generate(
input_ids,
force_words_ids=force_words_ids,
num_beams=10,
num_return_sequences=1,
no_repeat_ngram_size=1,
remove_invalid_values=True,
)
outputs = tokenizer.batch_decode(outputs, skip_special_tokens=True)
self.assertListEqual(outputs, ["Wie alt sind Sie?"])
@slow
def test_constrained_beam_search_example_integration(self):
# PT-only test: TF doesn't have constrained beam search
tokenizer = AutoTokenizer.from_pretrained("t5-base")
model = AutoModelForSeq2SeqLM.from_pretrained("t5-base")
encoder_input_str = "translate English to German: How old are you?"
encoder_input_ids = tokenizer(encoder_input_str, return_tensors="pt").input_ids
# lets run beam search using 5 beams
num_beams = 5
# define decoder start token ids
input_ids = torch.ones((num_beams, 1), device=model.device, dtype=torch.long)
input_ids = input_ids * model.config.decoder_start_token_id
# add encoder_outputs to model keyword arguments
model_kwargs = {
"encoder_outputs": model.get_encoder()(
encoder_input_ids.repeat_interleave(num_beams, dim=0), return_dict=True
)
}
constraint_str = "sind"
constraint_token_ids = tokenizer.encode(constraint_str)[:-1] # remove eos token
constraints = [PhrasalConstraint(token_ids=constraint_token_ids)]
# instantiate beam scorer
beam_scorer = ConstrainedBeamSearchScorer(
batch_size=1, num_beams=num_beams, device=model.device, constraints=constraints
)
# instantiate logits processors
logits_processor = LogitsProcessorList(
[
MinLengthLogitsProcessor(5, eos_token_id=model.config.eos_token_id),
]
)
outputs = model.constrained_beam_search(
input_ids, beam_scorer, constraints=constraints, logits_processor=logits_processor, **model_kwargs
)
outputs = tokenizer.batch_decode(outputs, skip_special_tokens=True)
self.assertListEqual(outputs, ["Wie alt sind Sie?"])
def test_constrained_beam_search_mixin_type_checks(self):
# PT-only test: TF doesn't have constrained beam search
tokenizer = AutoTokenizer.from_pretrained("patrickvonplaten/t5-tiny-random")
model = AutoModelForSeq2SeqLM.from_pretrained("patrickvonplaten/t5-tiny-random")
encoder_input_str = "translate English to German: How old are you?"
input_ids = tokenizer(encoder_input_str, return_tensors="pt").input_ids
with self.assertRaises(ValueError):
force_words = ["sind"]
force_words_ids = tokenizer(force_words, return_tensors="pt").input_ids
model.generate(
input_ids,
force_words_ids=force_words_ids,
num_beams=10,
num_return_sequences=1,
no_repeat_ngram_size=1,
remove_invalid_values=True,
)
with self.assertRaises(ValueError):
force_words = ["sind"]
force_words_ids = [tokenizer(force_words, return_tensors="pt").input_ids]
model.generate(
input_ids,
force_words_ids=force_words_ids,
num_beams=10,
num_return_sequences=1,
no_repeat_ngram_size=1,
remove_invalid_values=True,
)
with self.assertRaises(ValueError):
model.generate(input_ids, force_words_ids=[])
with self.assertRaises(ValueError):
model.generate(input_ids, force_words_ids=[[-1]])
with self.assertRaises(ValueError):
model.generate(input_ids, force_words_ids=[[[-1]]])
def test_contrastive_search_batched(self):
# PT-only test: TF doesn't have constrained beam search
# Tests that contrastive search works with batched inputs (i.e. has the same output as for non-batched inputs)
articles = ["Foo", "Bar Baz"]
tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
model = BartForConditionalGeneration.from_pretrained("hf-internal-testing/tiny-random-bart").to(torch_device)
model.config.eos_token_id = None
input_ids_batched = tokenizer(articles, padding=True, return_tensors="pt").input_ids.to(torch_device)
input_ids = tokenizer(articles[1], return_tensors="pt").input_ids.to(torch_device)
output_sequences_batched = model.generate(
input_ids=input_ids_batched, penalty_alpha=0.6, top_k=4, return_dict_in_generate=True, output_scores=True
)
output_sequences = model.generate(
input_ids=input_ids, penalty_alpha=0.6, top_k=4, return_dict_in_generate=True, output_scores=True
)
batched_out = tokenizer.decode(output_sequences_batched.sequences[1], skip_special_tokens=True)
out = tokenizer.decode(output_sequences.sequences[0], skip_special_tokens=True)
self.assertEqual(batched_out, out)
# output_sequences_batched.scores[0][1] -> 1st set of logits, 2nd sequence
max_score_diff = (output_sequences_batched.scores[0][1] - output_sequences.scores[0][0]).abs().max()
self.assertTrue(max_score_diff < 1e-5)
def test_eos_token_id_int_and_list_top_k_top_sampling(self):
# Has TF equivalent: this test relies on random sampling
generation_kwargs = {
"do_sample": True,
"num_beams": 1,
"top_p": 0.7,
"top_k": 10,
"temperature": 0.7,
}
expectation = 20
tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-gpt2")
text = """Hello, my dog is cute and"""
tokens = tokenizer(text, return_tensors="pt").to(torch_device)
model = AutoModelForCausalLM.from_pretrained("hf-internal-testing/tiny-random-gpt2").to(torch_device)
# Only some seeds will work both on CPU/GPU for a fixed `expectation` value.
# The selected seed is not guaranteed to work on all torch versions.
torch.manual_seed(1)
eos_token_id = 846
generated_tokens = model.generate(**tokens, eos_token_id=eos_token_id, **generation_kwargs)
self.assertTrue(expectation == len(generated_tokens[0]))
torch.manual_seed(1)
eos_token_id = [846, 198]
generated_tokens = model.generate(**tokens, eos_token_id=eos_token_id, **generation_kwargs)
self.assertTrue(expectation == len(generated_tokens[0]))
def test_model_kwarg_encoder_signature_filtering(self):
# Has TF equivalent: ample use of framework-specific code
bart_tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
article = """Hugging Face is a technology company based in New York and Paris."""
input_ids = bart_tokenizer(article, return_tensors="pt").input_ids.to(torch_device)
bart_model = BartForConditionalGeneration.from_pretrained("hf-internal-testing/tiny-random-bart").to(
torch_device
)
output = bart_model.generate(input_ids).cpu().numpy()
# Let's create a fake model that has a different signature. In particular, this fake model accepts "foo" as an
# argument. Because "foo" is not in the encoder signature and doesn't start with "decoder_", it will be part of
# the encoder kwargs prior to signature filtering, which would lead to an exception. But filtering kicks in and
# saves the day.
class FakeBart(BartForConditionalGeneration):
def forward(self, input_ids, foo=None, **kwargs):
return super().forward(input_ids, **kwargs)
bart_model = FakeBart.from_pretrained("hf-internal-testing/tiny-random-bart").to(torch_device)
fake_output = bart_model.generate(input_ids, foo="bar").cpu().numpy()
self.assertTrue(np.array_equal(output, fake_output))
# Encoder signature filtering only kicks in if it doesn't accept wildcard kwargs. The following test will fail
# because it doesn't do signature filtering.
class FakeEncoder(bart_model.model.encoder.__class__):
def forward(self, input_ids, **kwargs):
return super().forward(input_ids, **kwargs)
fake_encoder = FakeEncoder(bart_model.config, bart_model.model.shared).to(torch_device)
bart_model.model.encoder = fake_encoder
# Normal generation still works (the output will be different because the encoder weights are different)
fake_output = bart_model.generate(input_ids).cpu().numpy()
with self.assertRaises(TypeError):
# FakeEncoder.forward() accepts **kwargs -> no filtering -> type error due to unexpected input "foo"
bart_model.generate(input_ids, foo="bar")
def test_default_max_length_warning(self):
model = AutoModelForCausalLM.from_pretrained("hf-internal-testing/tiny-random-gpt2").to(torch_device)
tokenizer = AutoTokenizer.from_pretrained("hf-internal-testing/tiny-random-gpt2")
model.config.pad_token_id = tokenizer.eos_token_id
text = "Hello world"
tokenized_inputs = tokenizer([text], return_tensors="pt")
input_ids = tokenized_inputs.input_ids.to(torch_device)
# Default generation config value of 20 -> emits warning
with self.assertWarns(UserWarning):
model.generate(input_ids)
# Explicitly setting max_length to 20 -> no warning
with warnings.catch_warnings(record=True) as warning_list:
model.generate(input_ids, max_length=20)
self.assertEqual(len(warning_list), 0)
# Generation config max_length != 20 -> no warning
with warnings.catch_warnings(record=True) as warning_list:
# generation_config is modified -> legacy mode is disabled = generation_config takes precedence
model.generation_config.max_length = 10
model.generate(input_ids)
self.assertEqual(len(warning_list), 0)
|