File size: 31,428 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
import argparse
from contextlib import contextmanager
import dataclasses
from dataclasses import is_dataclass
from distutils.version import LooseVersion
import logging
from pathlib import Path
import time
from typing import Dict
from typing import Iterable
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.nn
import torch.optim
from typeguard import check_argument_types

from espnet2.iterators.abs_iter_factory import AbsIterFactory
from espnet2.main_funcs.average_nbest_models import average_nbest_models
from espnet2.main_funcs.calculate_all_attentions import calculate_all_attentions
from espnet2.schedulers.abs_scheduler import AbsBatchStepScheduler
from espnet2.schedulers.abs_scheduler import AbsEpochStepScheduler
from espnet2.schedulers.abs_scheduler import AbsScheduler
from espnet2.schedulers.abs_scheduler import AbsValEpochStepScheduler
from espnet2.torch_utils.add_gradient_noise import add_gradient_noise
from espnet2.torch_utils.device_funcs import to_device
from espnet2.torch_utils.recursive_op import recursive_average
from espnet2.torch_utils.set_all_random_seed import set_all_random_seed
from espnet2.train.abs_espnet_model import AbsESPnetModel
from espnet2.train.distributed_utils import DistributedOption
from espnet2.train.reporter import Reporter
from espnet2.train.reporter import SubReporter
from espnet2.utils.build_dataclass import build_dataclass

if LooseVersion(torch.__version__) >= LooseVersion("1.1.0"):
    from torch.utils.tensorboard import SummaryWriter
else:
    from tensorboardX import SummaryWriter
if torch.distributed.is_available():
    if LooseVersion(torch.__version__) > LooseVersion("1.0.1"):
        from torch.distributed import ReduceOp
    else:
        from torch.distributed import reduce_op as ReduceOp
else:
    ReduceOp = None

if LooseVersion(torch.__version__) >= LooseVersion("1.6.0"):
    from torch.cuda.amp import autocast
    from torch.cuda.amp import GradScaler
else:
    # Nothing to do if torch<1.6.0
    @contextmanager
    def autocast(enabled=True):
        yield

    GradScaler = None

try:
    import fairscale
except ImportError:
    fairscale = None


@dataclasses.dataclass
class TrainerOptions:
    ngpu: int
    resume: bool
    use_amp: bool
    train_dtype: str
    grad_noise: bool
    accum_grad: int
    grad_clip: float
    grad_clip_type: float
    log_interval: Optional[int]
    no_forward_run: bool
    use_tensorboard: bool
    use_wandb: bool
    output_dir: Union[Path, str]
    max_epoch: int
    seed: int
    sharded_ddp: bool
    patience: Optional[int]
    keep_nbest_models: Union[int, List[int]]
    early_stopping_criterion: Sequence[str]
    best_model_criterion: Sequence[Sequence[str]]
    val_scheduler_criterion: Sequence[str]
    unused_parameters: bool


class Trainer:
    """Trainer having a optimizer.

    If you'd like to use multiple optimizers, then inherit this class
    and override the methods if necessary - at least "train_one_epoch()"

    >>> class TwoOptimizerTrainer(Trainer):
    ...     @classmethod
    ...     def add_arguments(cls, parser):
    ...         ...
    ...
    ...     @classmethod
    ...     def train_one_epoch(cls, model, optimizers, ...):
    ...         loss1 = model.model1(...)
    ...         loss1.backward()
    ...         optimizers[0].step()
    ...
    ...         loss2 = model.model2(...)
    ...         loss2.backward()
    ...         optimizers[1].step()

    """

    def __init__(self):
        raise RuntimeError("This class can't be instantiated.")

    @classmethod
    def build_options(cls, args: argparse.Namespace) -> TrainerOptions:
        """Build options consumed by train(), eval(), and plot_attention()"""
        assert check_argument_types()
        return build_dataclass(TrainerOptions, args)

    @classmethod
    def add_arguments(cls, parser: argparse.ArgumentParser):
        """Reserved for future development of another Trainer"""
        pass

    @staticmethod
    def resume(
        checkpoint: Union[str, Path],
        model: torch.nn.Module,
        reporter: Reporter,
        optimizers: Sequence[torch.optim.Optimizer],
        schedulers: Sequence[Optional[AbsScheduler]],
        scaler: Optional[GradScaler],
        ngpu: int = 0,
    ):
        states = torch.load(
            checkpoint,
            map_location=f"cuda:{torch.cuda.current_device()}" if ngpu > 0 else "cpu",
        )
        model.load_state_dict(states["model"])
        reporter.load_state_dict(states["reporter"])
        for optimizer, state in zip(optimizers, states["optimizers"]):
            optimizer.load_state_dict(state)
        for scheduler, state in zip(schedulers, states["schedulers"]):
            if scheduler is not None:
                scheduler.load_state_dict(state)
        if scaler is not None:
            if states["scaler"] is None:
                logging.warning("scaler state is not found")
            else:
                scaler.load_state_dict(states["scaler"])

        logging.info(f"The training was resumed using {checkpoint}")

    @classmethod
    def run(
        cls,
        model: AbsESPnetModel,
        optimizers: Sequence[torch.optim.Optimizer],
        schedulers: Sequence[Optional[AbsScheduler]],
        train_iter_factory: AbsIterFactory,
        valid_iter_factory: AbsIterFactory,
        plot_attention_iter_factory: Optional[AbsIterFactory],
        trainer_options,
        distributed_option: DistributedOption,
    ) -> None:
        """Perform training. This method performs the main process of training."""
        assert check_argument_types()
        # NOTE(kamo): Don't check the type more strictly as far trainer_options
        assert is_dataclass(trainer_options), type(trainer_options)
        assert len(optimizers) == len(schedulers), (len(optimizers), len(schedulers))

        if isinstance(trainer_options.keep_nbest_models, int):
            keep_nbest_models = trainer_options.keep_nbest_models
        else:
            if len(trainer_options.keep_nbest_models) == 0:
                logging.warning("No keep_nbest_models is given. Change to [1]")
                trainer_options.keep_nbest_models = [1]
            keep_nbest_models = max(trainer_options.keep_nbest_models)

        output_dir = Path(trainer_options.output_dir)
        reporter = Reporter()
        if trainer_options.use_amp:
            if LooseVersion(torch.__version__) < LooseVersion("1.6.0"):
                raise RuntimeError(
                    "Require torch>=1.6.0 for  Automatic Mixed Precision"
                )
            if trainer_options.sharded_ddp:
                if fairscale is None:
                    raise RuntimeError(
                        "Requiring fairscale. Do 'pip install fairscale'"
                    )
                scaler = fairscale.optim.grad_scaler.ShardedGradScaler()
            else:
                scaler = GradScaler()
        else:
            scaler = None

        if trainer_options.resume and (output_dir / "checkpoint.pth").exists():
            cls.resume(
                checkpoint=output_dir / "checkpoint.pth",
                model=model,
                optimizers=optimizers,
                schedulers=schedulers,
                reporter=reporter,
                scaler=scaler,
                ngpu=trainer_options.ngpu,
            )

        start_epoch = reporter.get_epoch() + 1
        if start_epoch == trainer_options.max_epoch + 1:
            logging.warning(
                f"The training has already reached at max_epoch: {start_epoch}"
            )

        if distributed_option.distributed:
            if trainer_options.sharded_ddp:
                dp_model = fairscale.nn.data_parallel.ShardedDataParallel(
                    module=model,
                    sharded_optimizer=optimizers,
                )
            else:
                dp_model = torch.nn.parallel.DistributedDataParallel(
                    model,
                    device_ids=(
                        # Perform multi-Process with multi-GPUs
                        [torch.cuda.current_device()]
                        if distributed_option.ngpu == 1
                        # Perform single-Process with multi-GPUs
                        else None
                    ),
                    output_device=(
                        torch.cuda.current_device()
                        if distributed_option.ngpu == 1
                        else None
                    ),
                    find_unused_parameters=trainer_options.unused_parameters,
                )
        elif distributed_option.ngpu > 1:
            dp_model = torch.nn.parallel.DataParallel(
                model,
                device_ids=list(range(distributed_option.ngpu)),
            )
        else:
            # NOTE(kamo): DataParallel also should work with ngpu=1,
            # but for debuggability it's better to keep this block.
            dp_model = model

        if trainer_options.use_tensorboard and (
            not distributed_option.distributed or distributed_option.dist_rank == 0
        ):
            summary_writer = SummaryWriter(str(output_dir / "tensorboard"))
        else:
            summary_writer = None

        start_time = time.perf_counter()
        for iepoch in range(start_epoch, trainer_options.max_epoch + 1):
            if iepoch != start_epoch:
                logging.info(
                    "{}/{}epoch started. Estimated time to finish: {}".format(
                        iepoch,
                        trainer_options.max_epoch,
                        humanfriendly.format_timespan(
                            (time.perf_counter() - start_time)
                            / (iepoch - start_epoch)
                            * (trainer_options.max_epoch - iepoch + 1)
                        ),
                    )
                )
            else:
                logging.info(f"{iepoch}/{trainer_options.max_epoch}epoch started")
            set_all_random_seed(trainer_options.seed + iepoch)

            reporter.set_epoch(iepoch)
            # 1. Train and validation for one-epoch
            with reporter.observe("train") as sub_reporter:
                all_steps_are_invalid = cls.train_one_epoch(
                    model=dp_model,
                    optimizers=optimizers,
                    schedulers=schedulers,
                    iterator=train_iter_factory.build_iter(iepoch),
                    reporter=sub_reporter,
                    scaler=scaler,
                    summary_writer=summary_writer,
                    options=trainer_options,
                    distributed_option=distributed_option,
                )

            with reporter.observe("valid") as sub_reporter:
                cls.validate_one_epoch(
                    model=dp_model,
                    iterator=valid_iter_factory.build_iter(iepoch),
                    reporter=sub_reporter,
                    options=trainer_options,
                    distributed_option=distributed_option,
                )

            if not distributed_option.distributed or distributed_option.dist_rank == 0:
                # att_plot doesn't support distributed
                if plot_attention_iter_factory is not None:
                    with reporter.observe("att_plot") as sub_reporter:
                        cls.plot_attention(
                            model=model,
                            output_dir=output_dir / "att_ws",
                            summary_writer=summary_writer,
                            iterator=plot_attention_iter_factory.build_iter(iepoch),
                            reporter=sub_reporter,
                            options=trainer_options,
                        )

            # 2. LR Scheduler step
            for scheduler in schedulers:
                if isinstance(scheduler, AbsValEpochStepScheduler):
                    scheduler.step(
                        reporter.get_value(*trainer_options.val_scheduler_criterion)
                    )
                elif isinstance(scheduler, AbsEpochStepScheduler):
                    scheduler.step()
            if trainer_options.sharded_ddp:
                for optimizer in optimizers:
                    if isinstance(optimizer, fairscale.optim.oss.OSS):
                        optimizer.consolidate_state_dict()

            if not distributed_option.distributed or distributed_option.dist_rank == 0:
                # 3. Report the results
                logging.info(reporter.log_message())
                reporter.matplotlib_plot(output_dir / "images")
                if summary_writer is not None:
                    reporter.tensorboard_add_scalar(summary_writer)
                if trainer_options.use_wandb:
                    reporter.wandb_log()

                # 4. Save/Update the checkpoint
                torch.save(
                    {
                        "model": model.state_dict(),
                        "reporter": reporter.state_dict(),
                        "optimizers": [o.state_dict() for o in optimizers],
                        "schedulers": [
                            s.state_dict() if s is not None else None
                            for s in schedulers
                        ],
                        "scaler": scaler.state_dict() if scaler is not None else None,
                    },
                    output_dir / "checkpoint.pth",
                )

                # 5. Save the model and update the link to the best model
                torch.save(model.state_dict(), output_dir / f"{iepoch}epoch.pth")

                # Creates a sym link latest.pth -> {iepoch}epoch.pth
                p = output_dir / "latest.pth"
                if p.is_symlink() or p.exists():
                    p.unlink()
                p.symlink_to(f"{iepoch}epoch.pth")

                _improved = []
                for _phase, k, _mode in trainer_options.best_model_criterion:
                    # e.g. _phase, k, _mode = "train", "loss", "min"
                    if reporter.has(_phase, k):
                        best_epoch = reporter.get_best_epoch(_phase, k, _mode)
                        # Creates sym links if it's the best result
                        if best_epoch == iepoch:
                            p = output_dir / f"{_phase}.{k}.best.pth"
                            if p.is_symlink() or p.exists():
                                p.unlink()
                            p.symlink_to(f"{iepoch}epoch.pth")
                            _improved.append(f"{_phase}.{k}")
                if len(_improved) == 0:
                    logging.info("There are no improvements in this epoch")
                else:
                    logging.info(
                        "The best model has been updated: " + ", ".join(_improved)
                    )

                # 6. Remove the model files excluding n-best epoch and latest epoch
                _removed = []
                # Get the union set of the n-best among multiple criterion
                nbests = set().union(
                    *[
                        set(reporter.sort_epochs(ph, k, m)[:keep_nbest_models])
                        for ph, k, m in trainer_options.best_model_criterion
                        if reporter.has(ph, k)
                    ]
                )
                for e in range(1, iepoch):
                    p = output_dir / f"{e}epoch.pth"
                    if p.exists() and e not in nbests:
                        p.unlink()
                        _removed.append(str(p))
                if len(_removed) != 0:
                    logging.info("The model files were removed: " + ", ".join(_removed))

            # 7. If any updating haven't happened, stops the training
            if all_steps_are_invalid:
                logging.warning(
                    f"The gradients at all steps are invalid in this epoch. "
                    f"Something seems wrong. This training was stopped at {iepoch}epoch"
                )
                break

            # 8. Check early stopping
            if trainer_options.patience is not None:
                if reporter.check_early_stopping(
                    trainer_options.patience, *trainer_options.early_stopping_criterion
                ):
                    break

        else:
            logging.info(
                f"The training was finished at {trainer_options.max_epoch} epochs "
            )

        if not distributed_option.distributed or distributed_option.dist_rank == 0:
            # Generated n-best averaged model
            average_nbest_models(
                reporter=reporter,
                output_dir=output_dir,
                best_model_criterion=trainer_options.best_model_criterion,
                nbest=keep_nbest_models,
            )

    @classmethod
    def train_one_epoch(
        cls,
        model: torch.nn.Module,
        iterator: Iterable[Tuple[List[str], Dict[str, torch.Tensor]]],
        optimizers: Sequence[torch.optim.Optimizer],
        schedulers: Sequence[Optional[AbsScheduler]],
        scaler: Optional[GradScaler],
        reporter: SubReporter,
        summary_writer: Optional[SummaryWriter],
        options: TrainerOptions,
        distributed_option: DistributedOption,
    ) -> bool:
        assert check_argument_types()

        grad_noise = options.grad_noise
        accum_grad = options.accum_grad
        grad_clip = options.grad_clip
        grad_clip_type = options.grad_clip_type
        log_interval = options.log_interval
        no_forward_run = options.no_forward_run
        ngpu = options.ngpu
        use_wandb = options.use_wandb
        distributed = distributed_option.distributed

        if log_interval is None:
            try:
                log_interval = max(len(iterator) // 20, 10)
            except TypeError:
                log_interval = 100

        model.train()
        all_steps_are_invalid = True
        # [For distributed] Because iteration counts are not always equals between
        # processes, send stop-flag to the other processes if iterator is finished
        iterator_stop = torch.tensor(0).to("cuda" if ngpu > 0 else "cpu")

        start_time = time.perf_counter()
        for iiter, (_, batch) in enumerate(
            reporter.measure_iter_time(iterator, "iter_time"), 1
        ):
            assert isinstance(batch, dict), type(batch)

            if distributed:
                torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)
                if iterator_stop > 0:
                    break

            batch = to_device(batch, "cuda" if ngpu > 0 else "cpu")
            if no_forward_run:
                all_steps_are_invalid = False
                continue

            with autocast(scaler is not None):
                with reporter.measure_time("forward_time"):
                    retval = model(**batch)

                    # Note(kamo):
                    # Supporting two patterns for the returned value from the model
                    #   a. dict type
                    if isinstance(retval, dict):
                        loss = retval["loss"]
                        stats = retval["stats"]
                        weight = retval["weight"]
                        optim_idx = retval.get("optim_idx")
                        if optim_idx is not None and not isinstance(optim_idx, int):
                            if not isinstance(optim_idx, torch.Tensor):
                                raise RuntimeError(
                                    "optim_idx must be int or 1dim torch.Tensor, "
                                    f"but got {type(optim_idx)}"
                                )
                            if optim_idx.dim() >= 2:
                                raise RuntimeError(
                                    "optim_idx must be int or 1dim torch.Tensor, "
                                    f"but got {optim_idx.dim()}dim tensor"
                                )
                            if optim_idx.dim() == 1:
                                for v in optim_idx:
                                    if v != optim_idx[0]:
                                        raise RuntimeError(
                                            "optim_idx must be 1dim tensor "
                                            "having same values for all entries"
                                        )
                                optim_idx = optim_idx[0].item()
                            else:
                                optim_idx = optim_idx.item()

                    #   b. tuple or list type
                    else:
                        loss, stats, weight = retval
                        optim_idx = None

                stats = {k: v for k, v in stats.items() if v is not None}
                if ngpu > 1 or distributed:
                    # Apply weighted averaging for loss and stats
                    loss = (loss * weight.type(loss.dtype)).sum()

                    # if distributed, this method can also apply all_reduce()
                    stats, weight = recursive_average(stats, weight, distributed)

                    # Now weight is summation over all workers
                    loss /= weight
                if distributed:
                    # NOTE(kamo): Multiply world_size because DistributedDataParallel
                    # automatically normalizes the gradient by world_size.
                    loss *= torch.distributed.get_world_size()

                loss /= accum_grad

            reporter.register(stats, weight)

            with reporter.measure_time("backward_time"):
                if scaler is not None:
                    # Scales loss.  Calls backward() on scaled loss
                    # to create scaled gradients.
                    # Backward passes under autocast are not recommended.
                    # Backward ops run in the same dtype autocast chose
                    # for corresponding forward ops.
                    scaler.scale(loss).backward()
                else:
                    loss.backward()

            if iiter % accum_grad == 0:
                if scaler is not None:
                    # Unscales the gradients of optimizer's assigned params in-place
                    for iopt, optimizer in enumerate(optimizers):
                        if optim_idx is not None and iopt != optim_idx:
                            continue
                        scaler.unscale_(optimizer)

                # gradient noise injection
                if grad_noise:
                    add_gradient_noise(
                        model,
                        reporter.get_total_count(),
                        duration=100,
                        eta=1.0,
                        scale_factor=0.55,
                    )

                # compute the gradient norm to check if it is normal or not
                grad_norm = torch.nn.utils.clip_grad_norm_(
                    model.parameters(),
                    max_norm=grad_clip,
                    norm_type=grad_clip_type,
                )
                # PyTorch<=1.4, clip_grad_norm_ returns float value
                if not isinstance(grad_norm, torch.Tensor):
                    grad_norm = torch.tensor(grad_norm)

                if not torch.isfinite(grad_norm):
                    logging.warning(
                        f"The grad norm is {grad_norm}. Skipping updating the model."
                    )

                    # Must invoke scaler.update() if unscale_() is used in the iteration
                    # to avoid the following error:
                    #   RuntimeError: unscale_() has already been called
                    #   on this optimizer since the last update().
                    # Note that if the gradient has inf/nan values,
                    # scaler.step skips optimizer.step().
                    if scaler is not None:
                        for iopt, optimizer in enumerate(optimizers):
                            if optim_idx is not None and iopt != optim_idx:
                                continue
                            scaler.step(optimizer)
                            scaler.update()

                else:
                    all_steps_are_invalid = False
                    with reporter.measure_time("optim_step_time"):
                        for iopt, (optimizer, scheduler) in enumerate(
                            zip(optimizers, schedulers)
                        ):
                            if optim_idx is not None and iopt != optim_idx:
                                continue
                            if scaler is not None:
                                # scaler.step() first unscales the gradients of
                                # the optimizer's assigned params.
                                scaler.step(optimizer)
                                # Updates the scale for next iteration.
                                scaler.update()
                            else:
                                optimizer.step()
                            if isinstance(scheduler, AbsBatchStepScheduler):
                                scheduler.step()
                            optimizer.zero_grad()

                # Register lr and train/load time[sec/step],
                # where step refers to accum_grad * mini-batch
                reporter.register(
                    dict(
                        {
                            f"optim{i}_lr{j}": pg["lr"]
                            for i, optimizer in enumerate(optimizers)
                            for j, pg in enumerate(optimizer.param_groups)
                            if "lr" in pg
                        },
                        train_time=time.perf_counter() - start_time,
                    ),
                )
                start_time = time.perf_counter()

            # NOTE(kamo): Call log_message() after next()
            reporter.next()
            if iiter % log_interval == 0:
                logging.info(reporter.log_message(-log_interval))
                if summary_writer is not None:
                    reporter.tensorboard_add_scalar(summary_writer, -log_interval)
                if use_wandb:
                    reporter.wandb_log()

        else:
            if distributed:
                iterator_stop.fill_(1)
                torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)

        return all_steps_are_invalid

    @classmethod
    @torch.no_grad()
    def validate_one_epoch(
        cls,
        model: torch.nn.Module,
        iterator: Iterable[Dict[str, torch.Tensor]],
        reporter: SubReporter,
        options: TrainerOptions,
        distributed_option: DistributedOption,
    ) -> None:
        assert check_argument_types()
        ngpu = options.ngpu
        no_forward_run = options.no_forward_run
        distributed = distributed_option.distributed

        model.eval()

        # [For distributed] Because iteration counts are not always equals between
        # processes, send stop-flag to the other processes if iterator is finished
        iterator_stop = torch.tensor(0).to("cuda" if ngpu > 0 else "cpu")
        for (_, batch) in iterator:
            assert isinstance(batch, dict), type(batch)
            if distributed:
                torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)
                if iterator_stop > 0:
                    break

            batch = to_device(batch, "cuda" if ngpu > 0 else "cpu")
            if no_forward_run:
                continue

            retval = model(**batch)
            if isinstance(retval, dict):
                stats = retval["stats"]
                weight = retval["weight"]
            else:
                _, stats, weight = retval
            if ngpu > 1 or distributed:
                # Apply weighted averaging for stats.
                # if distributed, this method can also apply all_reduce()
                stats, weight = recursive_average(stats, weight, distributed)

            reporter.register(stats, weight)
            reporter.next()

        else:
            if distributed:
                iterator_stop.fill_(1)
                torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)

    @classmethod
    @torch.no_grad()
    def plot_attention(
        cls,
        model: torch.nn.Module,
        output_dir: Optional[Path],
        summary_writer: Optional[SummaryWriter],
        iterator: Iterable[Tuple[List[str], Dict[str, torch.Tensor]]],
        reporter: SubReporter,
        options: TrainerOptions,
    ) -> None:
        assert check_argument_types()
        import matplotlib

        ngpu = options.ngpu
        no_forward_run = options.no_forward_run

        matplotlib.use("Agg")
        import matplotlib.pyplot as plt
        from matplotlib.ticker import MaxNLocator

        model.eval()
        for ids, batch in iterator:
            assert isinstance(batch, dict), type(batch)
            assert len(next(iter(batch.values()))) == len(ids), (
                len(next(iter(batch.values()))),
                len(ids),
            )
            batch = to_device(batch, "cuda" if ngpu > 0 else "cpu")
            if no_forward_run:
                continue

            # 1. Forwarding model and gathering all attentions
            #    calculate_all_attentions() uses single gpu only.
            att_dict = calculate_all_attentions(model, batch)

            # 2. Plot attentions: This part is slow due to matplotlib
            for k, att_list in att_dict.items():
                assert len(att_list) == len(ids), (len(att_list), len(ids))
                for id_, att_w in zip(ids, att_list):

                    if isinstance(att_w, torch.Tensor):
                        att_w = att_w.detach().cpu().numpy()

                    if att_w.ndim == 2:
                        att_w = att_w[None]
                    elif att_w.ndim > 3 or att_w.ndim == 1:
                        raise RuntimeError(f"Must be 2 or 3 dimension: {att_w.ndim}")

                    w, h = plt.figaspect(1.0 / len(att_w))
                    fig = plt.Figure(figsize=(w * 1.3, h * 1.3))
                    axes = fig.subplots(1, len(att_w))
                    if len(att_w) == 1:
                        axes = [axes]

                    for ax, aw in zip(axes, att_w):
                        ax.imshow(aw.astype(np.float32), aspect="auto")
                        ax.set_title(f"{k}_{id_}")
                        ax.set_xlabel("Input")
                        ax.set_ylabel("Output")
                        ax.xaxis.set_major_locator(MaxNLocator(integer=True))
                        ax.yaxis.set_major_locator(MaxNLocator(integer=True))

                    if output_dir is not None:
                        p = output_dir / id_ / f"{k}.{reporter.get_epoch()}ep.png"
                        p.parent.mkdir(parents=True, exist_ok=True)
                        fig.savefig(p)

                    if summary_writer is not None:
                        summary_writer.add_figure(
                            f"{k}_{id_}", fig, reporter.get_epoch()
                        )
            reporter.next()