File size: 66,765 Bytes
ad16788 |
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 |
from abc import ABC
from abc import abstractmethod
import argparse
from dataclasses import dataclass
from distutils.version import LooseVersion
import functools
import logging
import os
from pathlib import Path
import sys
from typing import Any
from typing import Callable
from typing import Dict
from typing import List
from typing import Optional
from typing import Sequence
from typing import Tuple
from typing import Union
import humanfriendly
import numpy as np
import torch
import torch.multiprocessing
import torch.nn
import torch.optim
from torch.utils.data import DataLoader
from typeguard import check_argument_types
from typeguard import check_return_type
import wandb
import yaml
from espnet import __version__
from espnet.utils.cli_utils import get_commandline_args
from espnet2.iterators.abs_iter_factory import AbsIterFactory
from espnet2.iterators.chunk_iter_factory import ChunkIterFactory
from espnet2.iterators.multiple_iter_factory import MultipleIterFactory
from espnet2.iterators.sequence_iter_factory import SequenceIterFactory
from espnet2.main_funcs.collect_stats import collect_stats
from espnet2.optimizers.sgd import SGD
from espnet2.samplers.build_batch_sampler import BATCH_TYPES
from espnet2.samplers.build_batch_sampler import build_batch_sampler
from espnet2.samplers.unsorted_batch_sampler import UnsortedBatchSampler
from espnet2.schedulers.noam_lr import NoamLR
from espnet2.schedulers.warmup_lr import WarmupLR
from espnet2.torch_utils.load_pretrained_model import load_pretrained_model
from espnet2.torch_utils.model_summary import model_summary
from espnet2.torch_utils.pytorch_version import pytorch_cudnn_version
from espnet2.torch_utils.set_all_random_seed import set_all_random_seed
from espnet2.train.abs_espnet_model import AbsESPnetModel
from espnet2.train.class_choices import ClassChoices
from espnet2.train.dataset import AbsDataset
from espnet2.train.dataset import DATA_TYPES
from espnet2.train.dataset import ESPnetDataset
from espnet2.train.distributed_utils import DistributedOption
from espnet2.train.distributed_utils import free_port
from espnet2.train.distributed_utils import get_master_port
from espnet2.train.distributed_utils import get_node_rank
from espnet2.train.distributed_utils import get_num_nodes
from espnet2.train.distributed_utils import resolve_distributed_mode
from espnet2.train.iterable_dataset import IterableESPnetDataset
from espnet2.train.trainer import Trainer
from espnet2.utils.build_dataclass import build_dataclass
from espnet2.utils import config_argparse
from espnet2.utils.get_default_kwargs import get_default_kwargs
from espnet2.utils.nested_dict_action import NestedDictAction
from espnet2.utils.types import humanfriendly_parse_size_or_none
from espnet2.utils.types import int_or_none
from espnet2.utils.types import str2bool
from espnet2.utils.types import str2triple_str
from espnet2.utils.types import str_or_int
from espnet2.utils.types import str_or_none
from espnet2.utils.yaml_no_alias_safe_dump import yaml_no_alias_safe_dump
if LooseVersion(torch.__version__) >= LooseVersion("1.5.0"):
from torch.multiprocessing.spawn import ProcessContext
else:
from torch.multiprocessing.spawn import SpawnContext as ProcessContext
optim_classes = dict(
adam=torch.optim.Adam,
sgd=SGD,
adadelta=torch.optim.Adadelta,
adagrad=torch.optim.Adagrad,
adamax=torch.optim.Adamax,
asgd=torch.optim.ASGD,
lbfgs=torch.optim.LBFGS,
rmsprop=torch.optim.RMSprop,
rprop=torch.optim.Rprop,
)
if LooseVersion(torch.__version__) >= LooseVersion("1.2.0"):
optim_classes["adamw"] = torch.optim.AdamW
try:
import torch_optimizer
optim_classes.update(
accagd=torch_optimizer.AccSGD,
adabound=torch_optimizer.AdaBound,
adamod=torch_optimizer.AdaMod,
diffgrad=torch_optimizer.DiffGrad,
lamb=torch_optimizer.Lamb,
novograd=torch_optimizer.NovoGrad,
pid=torch_optimizer.PID,
# torch_optimizer<=0.0.1a10 doesn't support
# qhadam=torch_optimizer.QHAdam,
qhm=torch_optimizer.QHM,
radam=torch_optimizer.RAdam,
sgdw=torch_optimizer.SGDW,
yogi=torch_optimizer.Yogi,
)
del torch_optimizer
except ImportError:
pass
try:
import apex
optim_classes.update(
fusedadam=apex.optimizers.FusedAdam,
fusedlamb=apex.optimizers.FusedLAMB,
fusednovograd=apex.optimizers.FusedNovoGrad,
fusedsgd=apex.optimizers.FusedSGD,
)
del apex
except ImportError:
pass
try:
import fairscale
except ImportError:
fairscale = None
scheduler_classes = dict(
ReduceLROnPlateau=torch.optim.lr_scheduler.ReduceLROnPlateau,
lambdalr=torch.optim.lr_scheduler.LambdaLR,
steplr=torch.optim.lr_scheduler.StepLR,
multisteplr=torch.optim.lr_scheduler.MultiStepLR,
exponentiallr=torch.optim.lr_scheduler.ExponentialLR,
CosineAnnealingLR=torch.optim.lr_scheduler.CosineAnnealingLR,
)
if LooseVersion(torch.__version__) >= LooseVersion("1.1.0"):
scheduler_classes.update(
noamlr=NoamLR,
warmuplr=WarmupLR,
)
if LooseVersion(torch.__version__) >= LooseVersion("1.3.0"):
CosineAnnealingWarmRestarts = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts
scheduler_classes.update(
cycliclr=torch.optim.lr_scheduler.CyclicLR,
onecyclelr=torch.optim.lr_scheduler.OneCycleLR,
CosineAnnealingWarmRestarts=CosineAnnealingWarmRestarts,
)
# To lower keys
optim_classes = {k.lower(): v for k, v in optim_classes.items()}
scheduler_classes = {k.lower(): v for k, v in scheduler_classes.items()}
@dataclass
class IteratorOptions:
preprocess_fn: callable
collate_fn: callable
data_path_and_name_and_type: list
shape_files: list
batch_size: int
batch_bins: int
batch_type: str
max_cache_size: float
max_cache_fd: int
distributed: bool
num_batches: Optional[int]
num_iters_per_epoch: Optional[int]
train: bool
class AbsTask(ABC):
# Use @staticmethod, or @classmethod,
# instead of instance method to avoid God classes
# If you need more than one optimizers, change this value in inheritance
num_optimizers: int = 1
trainer = Trainer
class_choices_list: List[ClassChoices] = []
def __init__(self):
raise RuntimeError("This class can't be instantiated.")
@classmethod
@abstractmethod
def add_task_arguments(cls, parser: argparse.ArgumentParser):
pass
@classmethod
@abstractmethod
def build_collate_fn(
cls, args: argparse.Namespace, train: bool
) -> Callable[[Sequence[Dict[str, np.ndarray]]], Dict[str, torch.Tensor]]:
"""Return "collate_fn", which is a callable object and given to DataLoader.
>>> from torch.utils.data import DataLoader
>>> loader = DataLoader(collate_fn=cls.build_collate_fn(args, train=True), ...)
In many cases, you can use our common collate_fn.
"""
raise NotImplementedError
@classmethod
@abstractmethod
def build_preprocess_fn(
cls, args: argparse.Namespace, train: bool
) -> Optional[Callable[[str, Dict[str, np.array]], Dict[str, np.ndarray]]]:
raise NotImplementedError
@classmethod
@abstractmethod
def required_data_names(
cls, train: bool = True, inference: bool = False
) -> Tuple[str, ...]:
"""Define the required names by Task
This function is used by
>>> cls.check_task_requirements()
If your model is defined as following,
>>> from espnet2.train.abs_espnet_model import AbsESPnetModel
>>> class Model(AbsESPnetModel):
... def forward(self, input, output, opt=None): pass
then "required_data_names" should be as
>>> required_data_names = ('input', 'output')
"""
raise NotImplementedError
@classmethod
@abstractmethod
def optional_data_names(
cls, train: bool = True, inference: bool = False
) -> Tuple[str, ...]:
"""Define the optional names by Task
This function is used by
>>> cls.check_task_requirements()
If your model is defined as follows,
>>> from espnet2.train.abs_espnet_model import AbsESPnetModel
>>> class Model(AbsESPnetModel):
... def forward(self, input, output, opt=None): pass
then "optional_data_names" should be as
>>> optional_data_names = ('opt',)
"""
raise NotImplementedError
@classmethod
@abstractmethod
def build_model(cls, args: argparse.Namespace) -> AbsESPnetModel:
raise NotImplementedError
@classmethod
def get_parser(cls) -> config_argparse.ArgumentParser:
assert check_argument_types()
class ArgumentDefaultsRawTextHelpFormatter(
argparse.RawTextHelpFormatter,
argparse.ArgumentDefaultsHelpFormatter,
):
pass
parser = config_argparse.ArgumentParser(
description="base parser",
formatter_class=ArgumentDefaultsRawTextHelpFormatter,
)
# NOTE(kamo): Use '_' instead of '-' to avoid confusion.
# I think '-' looks really confusing if it's written in yaml.
# NOTE(kamo): add_arguments(..., required=True) can't be used
# to provide --print_config mode. Instead of it, do as
parser.set_defaults(required=["output_dir"])
group = parser.add_argument_group("Common configuration")
group.add_argument(
"--print_config",
action="store_true",
help="Print the config file and exit",
)
group.add_argument(
"--log_level",
type=lambda x: x.upper(),
default="INFO",
choices=("ERROR", "WARNING", "INFO", "DEBUG", "NOTSET"),
help="The verbose level of logging",
)
group.add_argument(
"--dry_run",
type=str2bool,
default=False,
help="Perform process without training",
)
group.add_argument(
"--iterator_type",
type=str,
choices=["sequence", "chunk", "task", "none"],
default="sequence",
help="Specify iterator type",
)
group.add_argument("--output_dir", type=str_or_none, default=None)
group.add_argument(
"--ngpu",
type=int,
default=0,
help="The number of gpus. 0 indicates CPU mode",
)
group.add_argument("--seed", type=int, default=0, help="Random seed")
group.add_argument(
"--num_workers",
type=int,
default=1,
help="The number of workers used for DataLoader",
)
group.add_argument(
"--num_att_plot",
type=int,
default=3,
help="The number images to plot the outputs from attention. "
"This option makes sense only when attention-based model",
)
group = parser.add_argument_group("distributed training related")
group.add_argument(
"--dist_backend",
default="nccl",
type=str,
help="distributed backend",
)
group.add_argument(
"--dist_init_method",
type=str,
default="env://",
help='if init_method="env://", env values of "MASTER_PORT", "MASTER_ADDR", '
'"WORLD_SIZE", and "RANK" are referred.',
)
group.add_argument(
"--dist_world_size",
default=None,
type=int_or_none,
help="number of nodes for distributed training",
)
group.add_argument(
"--dist_rank",
type=int_or_none,
default=None,
help="node rank for distributed training",
)
group.add_argument(
# Not starting with "dist_" for compatibility to launch.py
"--local_rank",
type=int_or_none,
default=None,
help="local rank for distributed training. This option is used if "
"--multiprocessing_distributed=false",
)
group.add_argument(
"--dist_master_addr",
default=None,
type=str_or_none,
help="The master address for distributed training. "
"This value is used when dist_init_method == 'env://'",
)
group.add_argument(
"--dist_master_port",
default=None,
type=int_or_none,
help="The master port for distributed training"
"This value is used when dist_init_method == 'env://'",
)
group.add_argument(
"--dist_launcher",
default=None,
type=str_or_none,
choices=["slurm", "mpi", None],
help="The launcher type for distributed training",
)
group.add_argument(
"--multiprocessing_distributed",
default=False,
type=str2bool,
help="Use multi-processing distributed training to launch "
"N processes per node, which has N GPUs. This is the "
"fastest way to use PyTorch for either single node or "
"multi node data parallel training",
)
group.add_argument(
"--unused_parameters",
type=str2bool,
default=False,
help="Whether to use the find_unused_parameters in "
"torch.nn.parallel.DistributedDataParallel ",
)
group.add_argument(
"--sharded_ddp",
default=False,
type=str2bool,
help="Enable sharded training provided by fairscale",
)
group = parser.add_argument_group("cudnn mode related")
group.add_argument(
"--cudnn_enabled",
type=str2bool,
default=torch.backends.cudnn.enabled,
help="Enable CUDNN",
)
group.add_argument(
"--cudnn_benchmark",
type=str2bool,
default=torch.backends.cudnn.benchmark,
help="Enable cudnn-benchmark mode",
)
group.add_argument(
"--cudnn_deterministic",
type=str2bool,
default=True,
help="Enable cudnn-deterministic mode",
)
group = parser.add_argument_group("collect stats mode related")
group.add_argument(
"--collect_stats",
type=str2bool,
default=False,
help='Perform on "collect stats" mode',
)
group.add_argument(
"--write_collected_feats",
type=str2bool,
default=False,
help='Write the output features from the model when "collect stats" mode',
)
group = parser.add_argument_group("Trainer related")
group.add_argument(
"--max_epoch",
type=int,
default=40,
help="The maximum number epoch to train",
)
group.add_argument(
"--patience",
type=int_or_none,
default=None,
help="Number of epochs to wait without improvement "
"before stopping the training",
)
group.add_argument(
"--val_scheduler_criterion",
type=str,
nargs=2,
default=("valid", "loss"),
help="The criterion used for the value given to the lr scheduler. "
'Give a pair referring the phase, "train" or "valid",'
'and the criterion name. The mode specifying "min" or "max" can '
"be changed by --scheduler_conf",
)
group.add_argument(
"--early_stopping_criterion",
type=str,
nargs=3,
default=("valid", "loss", "min"),
help="The criterion used for judging of early stopping. "
'Give a pair referring the phase, "train" or "valid",'
'the criterion name and the mode, "min" or "max", e.g. "acc,max".',
)
group.add_argument(
"--best_model_criterion",
type=str2triple_str,
nargs="+",
default=[
("train", "loss", "min"),
("valid", "loss", "min"),
("train", "acc", "max"),
("valid", "acc", "max"),
],
help="The criterion used for judging of the best model. "
'Give a pair referring the phase, "train" or "valid",'
'the criterion name, and the mode, "min" or "max", e.g. "acc,max".',
)
group.add_argument(
"--keep_nbest_models",
type=int,
nargs="+",
default=[10],
help="Remove previous snapshots excluding the n-best scored epochs",
)
group.add_argument(
"--grad_clip",
type=float,
default=5.0,
help="Gradient norm threshold to clip",
)
group.add_argument(
"--grad_clip_type",
type=float,
default=2.0,
help="The type of the used p-norm for gradient clip. Can be inf",
)
group.add_argument(
"--grad_noise",
type=str2bool,
default=False,
help="The flag to switch to use noise injection to "
"gradients during training",
)
group.add_argument(
"--accum_grad",
type=int,
default=1,
help="The number of gradient accumulation",
)
group.add_argument(
"--no_forward_run",
type=str2bool,
default=False,
help="Just only iterating data loading without "
"model forwarding and training",
)
group.add_argument(
"--resume",
type=str2bool,
default=False,
help="Enable resuming if checkpoint is existing",
)
group.add_argument(
"--train_dtype",
default="float32",
choices=["float16", "float32", "float64"],
help="Data type for training.",
)
group.add_argument(
"--use_amp",
type=str2bool,
default=False,
help="Enable Automatic Mixed Precision. This feature requires pytorch>=1.6",
)
group.add_argument(
"--log_interval",
type=int_or_none,
default=None,
help="Show the logs every the number iterations in each epochs at the "
"training phase. If None is given, it is decided according the number "
"of training samples automatically .",
)
group.add_argument(
"--use_tensorboard",
type=str2bool,
default=True,
help="Enable tensorboard logging",
)
group.add_argument(
"--use_wandb",
type=str2bool,
default=False,
help="Enable wandb logging",
)
group.add_argument(
"--wandb_project",
type=str,
default=None,
help="Specify wandb project",
)
group.add_argument(
"--wandb_id",
type=str,
default=None,
help="Specify wandb id",
)
group.add_argument(
"--detect_anomaly",
type=str2bool,
default=False,
help="Set torch.autograd.set_detect_anomaly",
)
group = parser.add_argument_group("Pretraining model related")
group.add_argument("--pretrain_path", help="This option is obsoleted")
group.add_argument(
"--init_param",
type=str,
default=[],
nargs="*",
help="Specify the file path used for initialization of parameters. "
"The format is '<file_path>:<src_key>:<dst_key>:<exclude_keys>', "
"where file_path is the model file path, "
"src_key specifies the key of model states to be used in the model file, "
"dst_key specifies the attribute of the model to be initialized, "
"and exclude_keys excludes keys of model states for the initialization."
"e.g.\n"
" # Load all parameters"
" --init_param some/where/model.pth\n"
" # Load only decoder parameters"
" --init_param some/where/model.pth:decoder:decoder\n"
" # Load only decoder parameters excluding decoder.embed"
" --init_param some/where/model.pth:decoder:decoder:decoder.embed\n"
" --init_param some/where/model.pth:decoder:decoder:decoder.embed\n",
)
group.add_argument(
"--freeze_param",
type=str,
default=[],
nargs="*",
help="Freeze parameters",
)
group = parser.add_argument_group("BatchSampler related")
group.add_argument(
"--num_iters_per_epoch",
type=int_or_none,
default=None,
help="Restrict the number of iterations for training per epoch",
)
group.add_argument(
"--batch_size",
type=int,
default=20,
help="The mini-batch size used for training. Used if batch_type='unsorted',"
" 'sorted', or 'folded'.",
)
group.add_argument(
"--valid_batch_size",
type=int_or_none,
default=None,
help="If not given, the value of --batch_size is used",
)
group.add_argument(
"--batch_bins",
type=int,
default=1000000,
help="The number of batch bins. Used if batch_type='length' or 'numel'",
)
group.add_argument(
"--valid_batch_bins",
type=int_or_none,
default=None,
help="If not given, the value of --batch_bins is used",
)
group.add_argument("--train_shape_file", type=str, action="append", default=[])
group.add_argument("--valid_shape_file", type=str, action="append", default=[])
group = parser.add_argument_group("Sequence iterator related")
_batch_type_help = ""
for key, value in BATCH_TYPES.items():
_batch_type_help += f'"{key}":\n{value}\n'
group.add_argument(
"--batch_type",
type=str,
default="folded",
choices=list(BATCH_TYPES),
help=_batch_type_help,
)
group.add_argument(
"--valid_batch_type",
type=str_or_none,
default=None,
choices=list(BATCH_TYPES) + [None],
help="If not given, the value of --batch_type is used",
)
group.add_argument("--fold_length", type=int, action="append", default=[])
group.add_argument(
"--sort_in_batch",
type=str,
default="descending",
choices=["descending", "ascending"],
help="Sort the samples in each mini-batches by the sample "
'lengths. To enable this, "shape_file" must have the length information.',
)
group.add_argument(
"--sort_batch",
type=str,
default="descending",
choices=["descending", "ascending"],
help="Sort mini-batches by the sample lengths",
)
group.add_argument(
"--multiple_iterator",
type=str2bool,
default=False,
help="Use multiple iterator mode",
)
group = parser.add_argument_group("Chunk iterator related")
group.add_argument(
"--chunk_length",
type=str_or_int,
default=500,
help="Specify chunk length. e.g. '300', '300,400,500', or '300-400'."
"If multiple numbers separated by command are given, "
"one of them is selected randomly for each samples. "
"If two numbers are given with '-', it indicates the range of the choices. "
"Note that if the sequence length is shorter than the all chunk_lengths, "
"the sample is discarded. ",
)
group.add_argument(
"--chunk_shift_ratio",
type=float,
default=0.5,
help="Specify the shift width of chunks. If it's less than 1, "
"allows the overlapping and if bigger than 1, there are some gaps "
"between each chunk.",
)
group.add_argument(
"--num_cache_chunks",
type=int,
default=1024,
help="Shuffle in the specified number of chunks and generate mini-batches "
"More larger this value, more randomness can be obtained.",
)
group = parser.add_argument_group("Dataset related")
_data_path_and_name_and_type_help = (
"Give three words splitted by comma. It's used for the training data. "
"e.g. '--train_data_path_and_name_and_type some/path/a.scp,foo,sound'. "
"The first value, some/path/a.scp, indicates the file path, "
"and the second, foo, is the key name used for the mini-batch data, "
"and the last, sound, decides the file type. "
"This option is repeatable, so you can input any number of features "
"for your task. Supported file types are as follows:\n\n"
)
for key, dic in DATA_TYPES.items():
_data_path_and_name_and_type_help += f'"{key}":\n{dic["help"]}\n\n'
group.add_argument(
"--train_data_path_and_name_and_type",
type=str2triple_str,
action="append",
default=[],
help=_data_path_and_name_and_type_help,
)
group.add_argument(
"--valid_data_path_and_name_and_type",
type=str2triple_str,
action="append",
default=[],
)
group.add_argument(
"--allow_variable_data_keys",
type=str2bool,
default=False,
help="Allow the arbitrary keys for mini-batch with ignoring "
"the task requirements",
)
group.add_argument(
"--max_cache_size",
type=humanfriendly.parse_size,
default=0.0,
help="The maximum cache size for data loader. e.g. 10MB, 20GB.",
)
group.add_argument(
"--max_cache_fd",
type=int,
default=32,
help="The maximum number of file descriptors to be kept "
"as opened for ark files. "
"This feature is only valid when data type is 'kaldi_ark'.",
)
group.add_argument(
"--valid_max_cache_size",
type=humanfriendly_parse_size_or_none,
default=None,
help="The maximum cache size for validation data loader. e.g. 10MB, 20GB. "
"If None, the 5 percent size of --max_cache_size",
)
group = parser.add_argument_group("Optimizer related")
for i in range(1, cls.num_optimizers + 1):
suf = "" if i == 1 else str(i)
group.add_argument(
f"--optim{suf}",
type=lambda x: x.lower(),
default="adadelta",
choices=list(optim_classes),
help="The optimizer type",
)
group.add_argument(
f"--optim{suf}_conf",
action=NestedDictAction,
default=dict(),
help="The keyword arguments for optimizer",
)
group.add_argument(
f"--scheduler{suf}",
type=lambda x: str_or_none(x.lower()),
default=None,
choices=list(scheduler_classes) + [None],
help="The lr scheduler type",
)
group.add_argument(
f"--scheduler{suf}_conf",
action=NestedDictAction,
default=dict(),
help="The keyword arguments for lr scheduler",
)
cls.trainer.add_arguments(parser)
cls.add_task_arguments(parser)
assert check_return_type(parser)
return parser
@classmethod
def build_optimizers(
cls,
args: argparse.Namespace,
model: torch.nn.Module,
) -> List[torch.optim.Optimizer]:
if cls.num_optimizers != 1:
raise RuntimeError(
"build_optimizers() must be overridden if num_optimizers != 1"
)
optim_class = optim_classes.get(args.optim)
if optim_class is None:
raise ValueError(f"must be one of {list(optim_classes)}: {args.optim}")
if args.sharded_ddp:
if fairscale is None:
raise RuntimeError("Requiring fairscale. Do 'pip install fairscale'")
optim = fairscale.optim.oss.OSS(
params=model.parameters(), optim=optim_class, **args.optim_conf
)
else:
optim = optim_class(model.parameters(), **args.optim_conf)
optimizers = [optim]
return optimizers
@classmethod
def exclude_opts(cls) -> Tuple[str, ...]:
"""The options not to be shown by --print_config"""
return "required", "print_config", "config", "ngpu"
@classmethod
def get_default_config(cls) -> Dict[str, Any]:
"""Return the configuration as dict.
This method is used by print_config()
"""
def get_class_type(name: str, classes: dict):
_cls = classes.get(name)
if _cls is None:
raise ValueError(f"must be one of {list(classes)}: {name}")
return _cls
# This method is used only for --print_config
assert check_argument_types()
parser = cls.get_parser()
args, _ = parser.parse_known_args()
config = vars(args)
# Excludes the options not to be shown
for k in AbsTask.exclude_opts():
config.pop(k)
for i in range(1, cls.num_optimizers + 1):
suf = "" if i == 1 else str(i)
name = config[f"optim{suf}"]
optim_class = get_class_type(name, optim_classes)
conf = get_default_kwargs(optim_class)
# Overwrite the default by the arguments,
conf.update(config[f"optim{suf}_conf"])
# and set it again
config[f"optim{suf}_conf"] = conf
name = config[f"scheduler{suf}"]
if name is not None:
scheduler_class = get_class_type(name, scheduler_classes)
conf = get_default_kwargs(scheduler_class)
# Overwrite the default by the arguments,
conf.update(config[f"scheduler{suf}_conf"])
# and set it again
config[f"scheduler{suf}_conf"] = conf
for class_choices in cls.class_choices_list:
if getattr(args, class_choices.name) is not None:
class_obj = class_choices.get_class(getattr(args, class_choices.name))
conf = get_default_kwargs(class_obj)
name = class_choices.name
# Overwrite the default by the arguments,
conf.update(config[f"{name}_conf"])
# and set it again
config[f"{name}_conf"] = conf
return config
@classmethod
def check_required_command_args(cls, args: argparse.Namespace):
assert check_argument_types()
for k in vars(args):
if "-" in k:
raise RuntimeError(f'Use "_" instead of "-": parser.get_parser("{k}")')
required = ", ".join(
f"--{a}" for a in args.required if getattr(args, a) is None
)
if len(required) != 0:
parser = cls.get_parser()
parser.print_help(file=sys.stderr)
p = Path(sys.argv[0]).name
print(file=sys.stderr)
print(
f"{p}: error: the following arguments are required: " f"{required}",
file=sys.stderr,
)
sys.exit(2)
@classmethod
def check_task_requirements(
cls,
dataset: Union[AbsDataset, IterableESPnetDataset],
allow_variable_data_keys: bool,
train: bool,
inference: bool = False,
) -> None:
"""Check if the dataset satisfy the requirement of current Task"""
assert check_argument_types()
mes = (
f"If you intend to use an additional input, modify "
f'"{cls.__name__}.required_data_names()" or '
f'"{cls.__name__}.optional_data_names()". '
f"Otherwise you need to set --allow_variable_data_keys true "
)
for k in cls.required_data_names(train, inference):
if not dataset.has_name(k):
raise RuntimeError(
f'"{cls.required_data_names(train, inference)}" are required for'
f' {cls.__name__}. but "{dataset.names()}" are input.\n{mes}'
)
if not allow_variable_data_keys:
task_keys = cls.required_data_names(
train, inference
) + cls.optional_data_names(train, inference)
for k in dataset.names():
if k not in task_keys:
raise RuntimeError(
f"The data-name must be one of {task_keys} "
f'for {cls.__name__}: "{k}" is not allowed.\n{mes}'
)
@classmethod
def print_config(cls, file=sys.stdout) -> None:
assert check_argument_types()
# Shows the config: e.g. python train.py asr --print_config
config = cls.get_default_config()
file.write(yaml_no_alias_safe_dump(config, indent=4, sort_keys=False))
@classmethod
def main(cls, args: argparse.Namespace = None, cmd: Sequence[str] = None):
assert check_argument_types()
print(get_commandline_args(), file=sys.stderr)
if args is None:
parser = cls.get_parser()
args = parser.parse_args(cmd)
args.version = __version__
if args.pretrain_path is not None:
raise RuntimeError("--pretrain_path is deprecated. Use --init_param")
if args.print_config:
cls.print_config()
sys.exit(0)
cls.check_required_command_args(args)
# "distributed" is decided using the other command args
resolve_distributed_mode(args)
if not args.distributed or not args.multiprocessing_distributed:
cls.main_worker(args)
else:
assert args.ngpu > 1, args.ngpu
# Multi-processing distributed mode: e.g. 2node-4process-4GPU
# | Host1 | Host2 |
# | Process1 | Process2 | <= Spawn processes
# |Child1|Child2|Child1|Child2|
# |GPU1 |GPU2 |GPU1 |GPU2 |
# See also the following usage of --multiprocessing-distributed:
# https://github.com/pytorch/examples/blob/master/imagenet/main.py
num_nodes = get_num_nodes(args.dist_world_size, args.dist_launcher)
if num_nodes == 1:
args.dist_master_addr = "localhost"
args.dist_rank = 0
# Single node distributed training with multi-GPUs
if (
args.dist_init_method == "env://"
and get_master_port(args.dist_master_port) is None
):
# Get the unused port
args.dist_master_port = free_port()
# Assume that nodes use same number of GPUs each other
args.dist_world_size = args.ngpu * num_nodes
node_rank = get_node_rank(args.dist_rank, args.dist_launcher)
# The following block is copied from:
# https://github.com/pytorch/pytorch/blob/master/torch/multiprocessing/spawn.py
error_queues = []
processes = []
mp = torch.multiprocessing.get_context("spawn")
for i in range(args.ngpu):
# Copy args
local_args = argparse.Namespace(**vars(args))
local_args.local_rank = i
local_args.dist_rank = args.ngpu * node_rank + i
local_args.ngpu = 1
process = mp.Process(
target=cls.main_worker,
args=(local_args,),
daemon=False,
)
process.start()
processes.append(process)
error_queues.append(mp.SimpleQueue())
# Loop on join until it returns True or raises an exception.
while not ProcessContext(processes, error_queues).join():
pass
@classmethod
def main_worker(cls, args: argparse.Namespace):
assert check_argument_types()
# 0. Init distributed process
distributed_option = build_dataclass(DistributedOption, args)
# Setting distributed_option.dist_rank, etc.
distributed_option.init_options()
# NOTE(kamo): Don't use logging before invoking logging.basicConfig()
if not distributed_option.distributed or distributed_option.dist_rank == 0:
if not distributed_option.distributed:
_rank = ""
else:
_rank = (
f":{distributed_option.dist_rank}/"
f"{distributed_option.dist_world_size}"
)
# NOTE(kamo):
# logging.basicConfig() is invoked in main_worker() instead of main()
# because it can be invoked only once in a process.
# FIXME(kamo): Should we use logging.getLogger()?
logging.basicConfig(
level=args.log_level,
format=f"[{os.uname()[1].split('.')[0]}{_rank}]"
f" %(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
)
else:
# Suppress logging if RANK != 0
logging.basicConfig(
level="ERROR",
format=f"[{os.uname()[1].split('.')[0]}"
f":{distributed_option.dist_rank}/{distributed_option.dist_world_size}]"
f" %(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
)
# Invoking torch.distributed.init_process_group
distributed_option.init_torch_distributed()
# 1. Set random-seed
set_all_random_seed(args.seed)
torch.backends.cudnn.enabled = args.cudnn_enabled
torch.backends.cudnn.benchmark = args.cudnn_benchmark
torch.backends.cudnn.deterministic = args.cudnn_deterministic
if args.detect_anomaly:
logging.info("Invoking torch.autograd.set_detect_anomaly(True)")
torch.autograd.set_detect_anomaly(args.detect_anomaly)
# 2. Build model
model = cls.build_model(args=args)
if not isinstance(model, AbsESPnetModel):
raise RuntimeError(
f"model must inherit {AbsESPnetModel.__name__}, but got {type(model)}"
)
model = model.to(
dtype=getattr(torch, args.train_dtype),
device="cuda" if args.ngpu > 0 else "cpu",
)
for t in args.freeze_param:
for k, p in model.named_parameters():
if k.startswith(t + ".") or k == t:
logging.info(f"Setting {k}.requires_grad = False")
p.requires_grad = False
# 3. Build optimizer
optimizers = cls.build_optimizers(args, model=model)
# 4. Build schedulers
schedulers = []
for i, optim in enumerate(optimizers, 1):
suf = "" if i == 1 else str(i)
name = getattr(args, f"scheduler{suf}")
conf = getattr(args, f"scheduler{suf}_conf")
if name is not None:
cls_ = scheduler_classes.get(name)
if cls_ is None:
raise ValueError(
f"must be one of {list(scheduler_classes)}: {name}"
)
scheduler = cls_(optim, **conf)
else:
scheduler = None
schedulers.append(scheduler)
logging.info(pytorch_cudnn_version())
logging.info(model_summary(model))
for i, (o, s) in enumerate(zip(optimizers, schedulers), 1):
suf = "" if i == 1 else str(i)
logging.info(f"Optimizer{suf}:\n{o}")
logging.info(f"Scheduler{suf}: {s}")
# 5. Dump "args" to config.yaml
# NOTE(kamo): "args" should be saved after object-buildings are done
# because they are allowed to modify "args".
output_dir = Path(args.output_dir)
if not distributed_option.distributed or distributed_option.dist_rank == 0:
output_dir.mkdir(parents=True, exist_ok=True)
with (output_dir / "config.yaml").open("w", encoding="utf-8") as f:
logging.info(
f'Saving the configuration in {output_dir / "config.yaml"}'
)
yaml_no_alias_safe_dump(vars(args), f, indent=4, sort_keys=False)
# 6. Loads pre-trained model
for p in args.init_param:
logging.info(f"Loading pretrained params from {p}")
load_pretrained_model(
model=model,
init_param=p,
# NOTE(kamo): "cuda" for torch.load always indicates cuda:0
# in PyTorch<=1.4
map_location=f"cuda:{torch.cuda.current_device()}"
if args.ngpu > 0
else "cpu",
)
if args.dry_run:
pass
elif args.collect_stats:
# Perform on collect_stats mode. This mode has two roles
# - Derive the length and dimension of all input data
# - Accumulate feats, square values, and the length for whitening
logging.info(args)
if args.valid_batch_size is None:
args.valid_batch_size = args.batch_size
if len(args.train_shape_file) != 0:
train_key_file = args.train_shape_file[0]
else:
train_key_file = None
if len(args.valid_shape_file) != 0:
valid_key_file = args.valid_shape_file[0]
else:
valid_key_file = None
collect_stats(
model=model,
train_iter=cls.build_streaming_iterator(
data_path_and_name_and_type=args.train_data_path_and_name_and_type,
key_file=train_key_file,
batch_size=args.batch_size,
dtype=args.train_dtype,
num_workers=args.num_workers,
allow_variable_data_keys=args.allow_variable_data_keys,
ngpu=args.ngpu,
preprocess_fn=cls.build_preprocess_fn(args, train=False),
collate_fn=cls.build_collate_fn(args, train=False),
),
valid_iter=cls.build_streaming_iterator(
data_path_and_name_and_type=args.valid_data_path_and_name_and_type,
key_file=valid_key_file,
batch_size=args.valid_batch_size,
dtype=args.train_dtype,
num_workers=args.num_workers,
allow_variable_data_keys=args.allow_variable_data_keys,
ngpu=args.ngpu,
preprocess_fn=cls.build_preprocess_fn(args, train=False),
collate_fn=cls.build_collate_fn(args, train=False),
),
output_dir=output_dir,
ngpu=args.ngpu,
log_interval=args.log_interval,
write_collected_feats=args.write_collected_feats,
)
else:
# 7. Build iterator factories
if args.multiple_iterator:
train_iter_factory = cls.build_multiple_iter_factory(
args=args,
distributed_option=distributed_option,
mode="train",
)
else:
train_iter_factory = cls.build_iter_factory(
args=args,
distributed_option=distributed_option,
mode="train",
)
valid_iter_factory = cls.build_iter_factory(
args=args,
distributed_option=distributed_option,
mode="valid",
)
if args.num_att_plot != 0:
plot_attention_iter_factory = cls.build_iter_factory(
args=args,
distributed_option=distributed_option,
mode="plot_att",
)
else:
plot_attention_iter_factory = None
# 8. Start training
if args.use_wandb:
if (
not distributed_option.distributed
or distributed_option.dist_rank == 0
):
if args.wandb_project is None:
project = (
"ESPnet_"
+ cls.__name__
+ str(Path(".").resolve()).replace("/", "_")
)
else:
project = args.wandb_project
if args.wandb_id is None:
wandb_id = str(output_dir).replace("/", "_")
else:
wandb_id = args.wandb_id
wandb.init(
project=project,
dir=output_dir,
id=wandb_id,
resume="allow",
)
wandb.config.update(args)
else:
# wandb also supports grouping for distributed training,
# but we only logs aggregated data,
# so it's enough to perform on rank0 node.
args.use_wandb = False
# Don't give args to trainer.run() directly!!!
# Instead of it, define "Options" object and build here.
trainer_options = cls.trainer.build_options(args)
cls.trainer.run(
model=model,
optimizers=optimizers,
schedulers=schedulers,
train_iter_factory=train_iter_factory,
valid_iter_factory=valid_iter_factory,
plot_attention_iter_factory=plot_attention_iter_factory,
trainer_options=trainer_options,
distributed_option=distributed_option,
)
@classmethod
def build_iter_options(
cls,
args: argparse.Namespace,
distributed_option: DistributedOption,
mode: str,
):
if mode == "train":
preprocess_fn = cls.build_preprocess_fn(args, train=True)
collate_fn = cls.build_collate_fn(args, train=True)
data_path_and_name_and_type = args.train_data_path_and_name_and_type
shape_files = args.train_shape_file
batch_size = args.batch_size
batch_bins = args.batch_bins
batch_type = args.batch_type
max_cache_size = args.max_cache_size
max_cache_fd = args.max_cache_fd
distributed = distributed_option.distributed
num_batches = None
num_iters_per_epoch = args.num_iters_per_epoch
train = True
elif mode == "valid":
preprocess_fn = cls.build_preprocess_fn(args, train=False)
collate_fn = cls.build_collate_fn(args, train=False)
data_path_and_name_and_type = args.valid_data_path_and_name_and_type
shape_files = args.valid_shape_file
if args.valid_batch_type is None:
batch_type = args.batch_type
else:
batch_type = args.valid_batch_type
if args.valid_batch_size is None:
batch_size = args.batch_size
else:
batch_size = args.valid_batch_size
if args.valid_batch_bins is None:
batch_bins = args.batch_bins
else:
batch_bins = args.valid_batch_bins
if args.valid_max_cache_size is None:
# Cache 5% of maximum size for validation loader
max_cache_size = 0.05 * args.max_cache_size
else:
max_cache_size = args.valid_max_cache_size
max_cache_fd = args.max_cache_fd
distributed = distributed_option.distributed
num_batches = None
num_iters_per_epoch = None
train = False
elif mode == "plot_att":
preprocess_fn = cls.build_preprocess_fn(args, train=False)
collate_fn = cls.build_collate_fn(args, train=False)
data_path_and_name_and_type = args.valid_data_path_and_name_and_type
shape_files = args.valid_shape_file
batch_type = "unsorted"
batch_size = 1
batch_bins = 0
num_batches = args.num_att_plot
max_cache_fd = args.max_cache_fd
# num_att_plot should be a few sample ~ 3, so cache all data.
max_cache_size = np.inf if args.max_cache_size != 0.0 else 0.0
# always False because plot_attention performs on RANK0
distributed = False
num_iters_per_epoch = None
train = False
else:
raise NotImplementedError(f"mode={mode}")
return IteratorOptions(
preprocess_fn=preprocess_fn,
collate_fn=collate_fn,
data_path_and_name_and_type=data_path_and_name_and_type,
shape_files=shape_files,
batch_type=batch_type,
batch_size=batch_size,
batch_bins=batch_bins,
num_batches=num_batches,
max_cache_size=max_cache_size,
max_cache_fd=max_cache_fd,
distributed=distributed,
num_iters_per_epoch=num_iters_per_epoch,
train=train,
)
@classmethod
def build_iter_factory(
cls,
args: argparse.Namespace,
distributed_option: DistributedOption,
mode: str,
kwargs: dict = None,
) -> AbsIterFactory:
"""Build a factory object of mini-batch iterator.
This object is invoked at every epochs to build the iterator for each epoch
as following:
>>> iter_factory = cls.build_iter_factory(...)
>>> for epoch in range(1, max_epoch):
... for keys, batch in iter_fatory.build_iter(epoch):
... model(**batch)
The mini-batches for each epochs are fully controlled by this class.
Note that the random seed used for shuffling is decided as "seed + epoch" and
the generated mini-batches can be reproduces when resuming.
Note that the definition of "epoch" doesn't always indicate
to run out of the whole training corpus.
"--num_iters_per_epoch" option restricts the number of iterations for each epoch
and the rest of samples for the originally epoch are left for the next epoch.
e.g. If The number of mini-batches equals to 4, the following two are same:
- 1 epoch without "--num_iters_per_epoch"
- 4 epoch with "--num_iters_per_epoch" == 4
"""
assert check_argument_types()
iter_options = cls.build_iter_options(args, distributed_option, mode)
# Overwrite iter_options if any kwargs is given
if kwargs is not None:
for k, v in kwargs.items():
setattr(iter_options, k, v)
if args.iterator_type == "sequence":
return cls.build_sequence_iter_factory(
args=args,
iter_options=iter_options,
mode=mode,
)
elif args.iterator_type == "chunk":
return cls.build_chunk_iter_factory(
args=args,
iter_options=iter_options,
mode=mode,
)
elif args.iterator_type == "task":
return cls.build_task_iter_factory(
args=args,
iter_options=iter_options,
mode=mode,
)
else:
raise RuntimeError(f"Not supported: iterator_type={args.iterator_type}")
@classmethod
def build_sequence_iter_factory(
cls, args: argparse.Namespace, iter_options: IteratorOptions, mode: str
) -> AbsIterFactory:
assert check_argument_types()
dataset = ESPnetDataset(
iter_options.data_path_and_name_and_type,
float_dtype=args.train_dtype,
preprocess=iter_options.preprocess_fn,
max_cache_size=iter_options.max_cache_size,
max_cache_fd=iter_options.max_cache_fd,
)
cls.check_task_requirements(
dataset, args.allow_variable_data_keys, train=iter_options.train
)
if Path(
Path(iter_options.data_path_and_name_and_type[0][0]).parent, "utt2category"
).exists():
utt2category_file = str(
Path(
Path(iter_options.data_path_and_name_and_type[0][0]).parent,
"utt2category",
)
)
else:
utt2category_file = None
batch_sampler = build_batch_sampler(
type=iter_options.batch_type,
shape_files=iter_options.shape_files,
fold_lengths=args.fold_length,
batch_size=iter_options.batch_size,
batch_bins=iter_options.batch_bins,
sort_in_batch=args.sort_in_batch,
sort_batch=args.sort_batch,
drop_last=False,
min_batch_size=torch.distributed.get_world_size()
if iter_options.distributed
else 1,
utt2category_file=utt2category_file,
)
batches = list(batch_sampler)
if iter_options.num_batches is not None:
batches = batches[: iter_options.num_batches]
bs_list = [len(batch) for batch in batches]
logging.info(f"[{mode}] dataset:\n{dataset}")
logging.info(f"[{mode}] Batch sampler: {batch_sampler}")
logging.info(
f"[{mode}] mini-batch sizes summary: N-batch={len(bs_list)}, "
f"mean={np.mean(bs_list):.1f}, min={np.min(bs_list)}, max={np.max(bs_list)}"
)
if iter_options.distributed:
world_size = torch.distributed.get_world_size()
rank = torch.distributed.get_rank()
for batch in batches:
if len(batch) < world_size:
raise RuntimeError(
f"The batch-size must be equal or more than world_size: "
f"{len(batch)} < {world_size}"
)
batches = [batch[rank::world_size] for batch in batches]
return SequenceIterFactory(
dataset=dataset,
batches=batches,
seed=args.seed,
num_iters_per_epoch=iter_options.num_iters_per_epoch,
shuffle=iter_options.train,
num_workers=args.num_workers,
collate_fn=iter_options.collate_fn,
pin_memory=args.ngpu > 0,
)
@classmethod
def build_chunk_iter_factory(
cls,
args: argparse.Namespace,
iter_options: IteratorOptions,
mode: str,
) -> AbsIterFactory:
assert check_argument_types()
dataset = ESPnetDataset(
iter_options.data_path_and_name_and_type,
float_dtype=args.train_dtype,
preprocess=iter_options.preprocess_fn,
max_cache_size=iter_options.max_cache_size,
max_cache_fd=iter_options.max_cache_fd,
)
cls.check_task_requirements(
dataset, args.allow_variable_data_keys, train=iter_options.train
)
if len(iter_options.shape_files) == 0:
key_file = iter_options.data_path_and_name_and_type[0][0]
else:
key_file = iter_options.shape_files[0]
batch_sampler = UnsortedBatchSampler(batch_size=1, key_file=key_file)
batches = list(batch_sampler)
if iter_options.num_batches is not None:
batches = batches[: iter_options.num_batches]
logging.info(f"[{mode}] dataset:\n{dataset}")
if iter_options.distributed:
world_size = torch.distributed.get_world_size()
rank = torch.distributed.get_rank()
if len(batches) < world_size:
raise RuntimeError("Number of samples is smaller than world_size")
if iter_options.batch_size < world_size:
raise RuntimeError("batch_size must be equal or more than world_size")
if rank < iter_options.batch_size % world_size:
batch_size = iter_options.batch_size // world_size + 1
else:
batch_size = iter_options.batch_size // world_size
num_cache_chunks = args.num_cache_chunks // world_size
# NOTE(kamo): Split whole corpus by sample numbers without considering
# each of the lengths, therefore the number of iteration counts are not
# always equal to each other and the iterations are limitted
# by the fewest iterations.
# i.e. the samples over the counts are discarded.
batches = batches[rank::world_size]
else:
batch_size = iter_options.batch_size
num_cache_chunks = args.num_cache_chunks
return ChunkIterFactory(
dataset=dataset,
batches=batches,
seed=args.seed,
batch_size=batch_size,
# For chunk iterator,
# --num_iters_per_epoch doesn't indicate the number of iterations,
# but indicates the number of samples.
num_samples_per_epoch=iter_options.num_iters_per_epoch,
shuffle=iter_options.train,
num_workers=args.num_workers,
collate_fn=iter_options.collate_fn,
pin_memory=args.ngpu > 0,
chunk_length=args.chunk_length,
chunk_shift_ratio=args.chunk_shift_ratio,
num_cache_chunks=num_cache_chunks,
)
# NOTE(kamo): Not abstract class
@classmethod
def build_task_iter_factory(
cls,
args: argparse.Namespace,
iter_options: IteratorOptions,
mode: str,
) -> AbsIterFactory:
"""Build task specific iterator factory
Example:
>>> class YourTask(AbsTask):
... @classmethod
... def add_task_arguments(cls, parser: argparse.ArgumentParser):
... parser.set_defaults(iterator_type="task")
...
... @classmethod
... def build_task_iter_factory(
... cls,
... args: argparse.Namespace,
... iter_options: IteratorOptions,
... mode: str,
... ):
... return FooIterFactory(...)
...
... @classmethod
... def build_iter_options(
.... args: argparse.Namespace,
... distributed_option: DistributedOption,
... mode: str
... ):
... # if you need to customize options object
"""
raise NotImplementedError
@classmethod
def build_multiple_iter_factory(
cls, args: argparse.Namespace, distributed_option: DistributedOption, mode: str
):
assert check_argument_types()
iter_options = cls.build_iter_options(args, distributed_option, mode)
assert len(iter_options.data_path_and_name_and_type) > 0, len(
iter_options.data_path_and_name_and_type
)
# 1. Sanity check
num_splits = None
for path in [
path for path, _, _ in iter_options.data_path_and_name_and_type
] + list(iter_options.shape_files):
if not Path(path).is_dir():
raise RuntimeError(f"{path} is not a directory")
p = Path(path) / "num_splits"
if not p.exists():
raise FileNotFoundError(f"{p} is not found")
with p.open() as f:
_num_splits = int(f.read())
if num_splits is not None and num_splits != _num_splits:
raise RuntimeError(
f"Number of splits are mismathed: "
f"{iter_options.data_path_and_name_and_type[0][0]} and {path}"
)
num_splits = _num_splits
for i in range(num_splits):
p = Path(path) / f"split.{i}"
if not p.exists():
raise FileNotFoundError(f"{p} is not found")
# 2. Create functions to build an iter factory for each splits
data_path_and_name_and_type_list = [
[
(str(Path(p) / f"split.{i}"), n, t)
for p, n, t in iter_options.data_path_and_name_and_type
]
for i in range(num_splits)
]
shape_files_list = [
[str(Path(s) / f"split.{i}") for s in iter_options.shape_files]
for i in range(num_splits)
]
num_iters_per_epoch_list = [
(iter_options.num_iters_per_epoch + i) // num_splits
if iter_options.num_iters_per_epoch is not None
else None
for i in range(num_splits)
]
max_cache_size = iter_options.max_cache_size / num_splits
# Note that iter-factories are built for each epoch at runtime lazily.
build_funcs = [
functools.partial(
cls.build_iter_factory,
args,
distributed_option,
mode,
kwargs=dict(
data_path_and_name_and_type=_data_path_and_name_and_type,
shape_files=_shape_files,
num_iters_per_epoch=_num_iters_per_epoch,
max_cache_size=max_cache_size,
),
)
for (
_data_path_and_name_and_type,
_shape_files,
_num_iters_per_epoch,
) in zip(
data_path_and_name_and_type_list,
shape_files_list,
num_iters_per_epoch_list,
)
]
# 3. Build MultipleIterFactory
return MultipleIterFactory(
build_funcs=build_funcs, shuffle=iter_options.train, seed=args.seed
)
@classmethod
def build_streaming_iterator(
cls,
data_path_and_name_and_type,
preprocess_fn,
collate_fn,
key_file: str = None,
batch_size: int = 1,
dtype: str = np.float32,
num_workers: int = 1,
allow_variable_data_keys: bool = False,
ngpu: int = 0,
inference: bool = False,
) -> DataLoader:
"""Build DataLoader using iterable dataset"""
assert check_argument_types()
# For backward compatibility for pytorch DataLoader
if collate_fn is not None:
kwargs = dict(collate_fn=collate_fn)
else:
kwargs = {}
# IterableDataset is supported from pytorch=1.2
if LooseVersion(torch.__version__) >= LooseVersion("1.2"):
dataset = IterableESPnetDataset(
data_path_and_name_and_type,
float_dtype=dtype,
preprocess=preprocess_fn,
key_file=key_file,
)
if dataset.apply_utt2category:
kwargs.update(batch_size=1)
else:
kwargs.update(batch_size=batch_size)
else:
dataset = ESPnetDataset(
data_path_and_name_and_type,
float_dtype=dtype,
preprocess=preprocess_fn,
)
if key_file is None:
key_file = data_path_and_name_and_type[0][0]
batch_sampler = UnsortedBatchSampler(
batch_size=batch_size,
key_file=key_file,
drop_last=False,
)
kwargs.update(batch_sampler=batch_sampler)
cls.check_task_requirements(
dataset, allow_variable_data_keys, train=False, inference=inference
)
return DataLoader(
dataset=dataset,
pin_memory=ngpu > 0,
num_workers=num_workers,
**kwargs,
)
# ~~~~~~~~~ The methods below are mainly used for inference ~~~~~~~~~
@classmethod
def build_model_from_file(
cls,
config_file: Union[Path, str],
model_file: Union[Path, str] = None,
device: str = "cpu",
) -> Tuple[AbsESPnetModel, argparse.Namespace]:
"""This method is used for inference or fine-tuning.
Args:
config_file: The yaml file saved when training.
model_file: The model file saved when training.
device:
"""
assert check_argument_types()
config_file = Path(config_file)
with config_file.open("r", encoding="utf-8") as f:
args = yaml.safe_load(f)
args = argparse.Namespace(**args)
model = cls.build_model(args)
if not isinstance(model, AbsESPnetModel):
raise RuntimeError(
f"model must inherit {AbsESPnetModel.__name__}, but got {type(model)}"
)
model.to(device)
if model_file is not None:
if device == "cuda":
# NOTE(kamo): "cuda" for torch.load always indicates cuda:0
# in PyTorch<=1.4
device = f"cuda:{torch.cuda.current_device()}"
model.load_state_dict(torch.load(model_file, map_location=device), strict=False) # TC Marker
return model, args
|