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
| #!/usr/bin/env python3
|
| import argparse
| import logging
| import os
| import sys
| from io import BytesIO
|
| import torch
|
| from funasr.build_utils.build_args import build_args
| from funasr.build_utils.build_dataloader import build_dataloader
| from funasr.build_utils.build_distributed import build_distributed
| from funasr.build_utils.build_model import build_model
| from funasr.build_utils.build_optimizer import build_optimizer
| from funasr.build_utils.build_scheduler import build_scheduler
| from funasr.build_utils.build_trainer import build_trainer
| from funasr.text.phoneme_tokenizer import g2p_choices
| from funasr.torch_utils.model_summary import model_summary
| from funasr.torch_utils.pytorch_version import pytorch_cudnn_version
| from funasr.torch_utils.set_all_random_seed import set_all_random_seed
| from funasr.utils.nested_dict_action import NestedDictAction
| from funasr.utils.prepare_data import prepare_data
| from funasr.utils.types import int_or_none
| from funasr.utils.types import str2bool
| from funasr.utils.types import str_or_none
| from funasr.utils.yaml_no_alias_safe_dump import yaml_no_alias_safe_dump
|
|
| def get_parser():
| parser = argparse.ArgumentParser(
| description="FunASR Common Training Parser",
| )
|
| # common configuration
| parser.add_argument("--output_dir", help="model save path")
| parser.add_argument(
| "--ngpu",
| type=int,
| default=0,
| help="The number of gpus. 0 indicates CPU mode",
| )
| parser.add_argument("--seed", type=int, default=0, help="Random seed")
| parser.add_argument("--task_name", type=str, default="asr", help="Name for different tasks")
|
| # ddp related
| parser.add_argument(
| "--dist_backend",
| default="nccl",
| type=str,
| help="distributed backend",
| )
| parser.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.',
| )
| parser.add_argument(
| "--dist_world_size",
| default=1,
| help="number of nodes for distributed training",
| )
| parser.add_argument(
| "--dist_rank",
| default=None,
| help="node rank for distributed training",
| )
| parser.add_argument(
| "--local_rank",
| default=None,
| help="local rank for distributed training",
| )
| parser.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://'",
| )
| parser.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://'",
| )
| parser.add_argument(
| "--dist_launcher",
| default=None,
| type=str_or_none,
| choices=["slurm", "mpi", None],
| help="The launcher type for distributed training",
| )
| parser.add_argument(
| "--multiprocessing_distributed",
| default=True,
| 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",
| )
| parser.add_argument(
| "--unused_parameters",
| type=str2bool,
| default=False,
| help="Whether to use the find_unused_parameters in "
| "torch.nn.parallel.DistributedDataParallel ",
| )
| parser.add_argument(
| "--gpu_id",
| type=int,
| default=0,
| help="local gpu id.",
| )
|
| # cudnn related
| parser.add_argument(
| "--cudnn_enabled",
| type=str2bool,
| default=torch.backends.cudnn.enabled,
| help="Enable CUDNN",
| )
| parser.add_argument(
| "--cudnn_benchmark",
| type=str2bool,
| default=torch.backends.cudnn.benchmark,
| help="Enable cudnn-benchmark mode",
| )
| parser.add_argument(
| "--cudnn_deterministic",
| type=str2bool,
| default=True,
| help="Enable cudnn-deterministic mode",
| )
|
| # trainer related
| parser.add_argument(
| "--max_epoch",
| type=int,
| default=40,
| help="The maximum number epoch to train",
| )
| parser.add_argument(
| "--max_update",
| type=int,
| default=sys.maxsize,
| help="The maximum number update step to train",
| )
| parser.add_argument(
| "--batch_interval",
| type=int,
| default=10000,
| help="The batch interval for saving model.",
| )
| parser.add_argument(
| "--patience",
| default=None,
| help="Number of epochs to wait without improvement "
| "before stopping the training",
| )
| parser.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",
| )
| parser.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".',
| )
| parser.add_argument(
| "--best_model_criterion",
| 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".',
| )
| parser.add_argument(
| "--keep_nbest_models",
| type=int,
| nargs="+",
| default=[10],
| help="Remove previous snapshots excluding the n-best scored epochs",
| )
| parser.add_argument(
| "--nbest_averaging_interval",
| type=int,
| default=0,
| help="The epoch interval to apply model averaging and save nbest models",
| )
| parser.add_argument(
| "--grad_clip",
| type=float,
| default=5.0,
| help="Gradient norm threshold to clip",
| )
| parser.add_argument(
| "--grad_clip_type",
| type=float,
| default=2.0,
| help="The type of the used p-norm for gradient clip. Can be inf",
| )
| parser.add_argument(
| "--grad_noise",
| type=str2bool,
| default=False,
| help="The flag to switch to use noise injection to "
| "gradients during training",
| )
| parser.add_argument(
| "--accum_grad",
| type=int,
| default=1,
| help="The number of gradient accumulation",
| )
| parser.add_argument(
| "--resume",
| type=str2bool,
| default=False,
| help="Enable resuming if checkpoint is existing",
| )
| parser.add_argument(
| "--use_amp",
| type=str2bool,
| default=False,
| help="Enable Automatic Mixed Precision. This feature requires pytorch>=1.6",
| )
| parser.add_argument(
| "--log_interval",
| 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 .",
| )
|
| # pretrained model related
| parser.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.pb\n"
| " # Load only decoder parameters"
| " --init_param some/where/model.pb:decoder:decoder\n"
| " # Load only decoder parameters excluding decoder.embed"
| " --init_param some/where/model.pb:decoder:decoder:decoder.embed\n"
| " --init_param some/where/model.pb:decoder:decoder:decoder.embed\n",
| )
| parser.add_argument(
| "--ignore_init_mismatch",
| type=str2bool,
| default=False,
| help="Ignore size mismatch when loading pre-trained model",
| )
| parser.add_argument(
| "--freeze_param",
| type=str,
| default=[],
| nargs="*",
| help="Freeze parameters",
| )
|
| # dataset related
| parser.add_argument(
| "--dataset_type",
| type=str,
| default="small",
| help="whether to use dataloader for large dataset",
| )
| parser.add_argument(
| "--train_data_file",
| type=str,
| default=None,
| help="train_list for large dataset",
| )
| parser.add_argument(
| "--valid_data_file",
| type=str,
| default=None,
| help="valid_list for large dataset",
| )
| parser.add_argument(
| "--train_data_path_and_name_and_type",
| action="append",
| default=[],
| help="e.g. '--train_data_path_and_name_and_type some/path/a.scp,foo,sound'. ",
| )
| parser.add_argument(
| "--valid_data_path_and_name_and_type",
| action="append",
| default=[],
| )
| parser.add_argument(
| "--train_shape_file",
| type=str, action="append",
| default=[],
| )
| parser.add_argument(
| "--valid_shape_file",
| type=str,
| action="append",
| default=[],
| )
| parser.add_argument(
| "--use_preprocessor",
| type=str2bool,
| default=True,
| help="Apply preprocessing to data or not",
| )
|
| # optimization related
| parser.add_argument(
| "--optim",
| type=lambda x: x.lower(),
| default="adam",
| help="The optimizer type",
| )
| parser.add_argument(
| "--optim_conf",
| action=NestedDictAction,
| default=dict(),
| help="The keyword arguments for optimizer",
| )
| parser.add_argument(
| "--scheduler",
| type=lambda x: str_or_none(x.lower()),
| default=None,
| help="The lr scheduler type",
| )
| parser.add_argument(
| "--scheduler_conf",
| action=NestedDictAction,
| default=dict(),
| help="The keyword arguments for lr scheduler",
| )
|
| # most task related
| parser.add_argument(
| "--init",
| type=lambda x: str_or_none(x.lower()),
| default=None,
| help="The initialization method",
| choices=[
| "chainer",
| "xavier_uniform",
| "xavier_normal",
| "kaiming_uniform",
| "kaiming_normal",
| None,
| ],
| )
| parser.add_argument(
| "--token_list",
| type=str_or_none,
| default=None,
| help="A text mapping int-id to token",
| )
| parser.add_argument(
| "--token_type",
| type=str,
| default="bpe",
| choices=["bpe", "char", "word"],
| help="",
| )
| parser.add_argument(
| "--bpemodel",
| type=str_or_none,
| default=None,
| help="The model file fo sentencepiece",
| )
| parser.add_argument(
| "--cleaner",
| type=str_or_none,
| choices=[None, "tacotron", "jaconv", "vietnamese"],
| default=None,
| help="Apply text cleaning",
| )
| parser.add_argument(
| "--g2p",
| type=str_or_none,
| choices=g2p_choices,
| default=None,
| help="Specify g2p method if --token_type=phn",
| )
|
| # pai related
| parser.add_argument(
| "--use_pai",
| type=str2bool,
| default=False,
| help="flag to indicate whether training on PAI",
| )
| parser.add_argument(
| "--simple_ddp",
| type=str2bool,
| default=False,
| )
| parser.add_argument(
| "--num_worker_count",
| type=int,
| default=1,
| help="The number of machines on PAI.",
| )
| parser.add_argument(
| "--access_key_id",
| type=str,
| default=None,
| help="The username for oss.",
| )
| parser.add_argument(
| "--access_key_secret",
| type=str,
| default=None,
| help="The password for oss.",
| )
| parser.add_argument(
| "--endpoint",
| type=str,
| default=None,
| help="The endpoint for oss.",
| )
| parser.add_argument(
| "--bucket_name",
| type=str,
| default=None,
| help="The bucket name for oss.",
| )
| parser.add_argument(
| "--oss_bucket",
| default=None,
| help="oss bucket.",
| )
|
| return parser
|
|
| if __name__ == '__main__':
| parser = get_parser()
| args, extra_task_params = parser.parse_known_args()
| if extra_task_params:
| args = build_args(args, parser, extra_task_params)
| # args = argparse.Namespace(**vars(args), **vars(task_args))
|
| # 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
|
| # ddp init
| os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu_id)
| args.distributed = args.ngpu > 1 or args.dist_world_size > 1
| distributed_option = build_distributed(args)
|
| # for logging
| if not distributed_option.distributed or distributed_option.dist_rank == 0:
| logging.basicConfig(
| level="INFO",
| format=f"[{os.uname()[1].split('.')[0]}]"
| f" %(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
| )
| else:
| logging.basicConfig(
| level="ERROR",
| format=f"[{os.uname()[1].split('.')[0]}]"
| f" %(asctime)s (%(module)s:%(lineno)d) %(levelname)s: %(message)s",
| )
|
| # prepare files for dataloader
| prepare_data(args, distributed_option)
|
| model = build_model(args)
| optimizers = build_optimizer(args, model=model)
| schedulers = build_scheduler(args, optimizers)
|
| logging.info("world size: {}, rank: {}, local_rank: {}".format(distributed_option.dist_world_size,
| distributed_option.dist_rank,
| distributed_option.local_rank))
| logging.info(pytorch_cudnn_version())
| logging.info(model_summary(model))
| logging.info("Optimizer: {}".format(optimizers))
| logging.info("Scheduler: {}".format(schedulers))
|
| # dump args to config.yaml
| if not distributed_option.distributed or distributed_option.dist_rank == 0:
| os.makedirs(args.output_dir, exist_ok=True)
| with open(os.path.join(args.output_dir, "config.yaml"), "w") as f:
| logging.info("Saving the configuration in {}/{}".format(args.output_dir, "config.yaml"))
| if args.use_pai:
| buffer = BytesIO()
| torch.save({"config": vars(args)}, buffer)
| args.oss_bucket.put_object(os.path.join(args.output_dir, "config.dict"), buffer.getvalue())
| else:
| yaml_no_alias_safe_dump(vars(args), f, indent=4, sort_keys=False)
|
| # dataloader for training/validation
| train_dataloader, valid_dataloader = build_dataloader(args)
|
| # Trainer, including model, optimizers, etc.
| trainer = build_trainer(
| args=args,
| model=model,
| optimizers=optimizers,
| schedulers=schedulers,
| train_dataloader=train_dataloader,
| valid_dataloader=valid_dataloader,
| distributed_option=distributed_option
| )
|
| trainer.run()
|
|