-
Notifications
You must be signed in to change notification settings - Fork 3.4k
Expand file tree
/
Copy pathrnnt_models.py
More file actions
1125 lines (936 loc) · 49.5 KB
/
rnnt_models.py
File metadata and controls
1125 lines (936 loc) · 49.5 KB
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
# Copyright (c) 2020, NVIDIA CORPORATION. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import copy
import os
from math import ceil
from typing import Any, Dict, List, Optional, Union
import numpy as np
import torch
from lightning.pytorch import Trainer
from omegaconf import DictConfig, OmegaConf, open_dict
from torch.utils.data import DataLoader
from nemo.collections.asr.data import audio_to_text_dataset
from nemo.collections.asr.data.audio_to_text import _AudioTextDataset
from nemo.collections.asr.data.audio_to_text_dali import AudioToCharDALIDataset, DALIOutputs
from nemo.collections.asr.data.audio_to_text_lhotse import LhotseSpeechToTextBpeDataset
from nemo.collections.asr.losses.rnnt import RNNTLoss, resolve_rnnt_default_loss_name
from nemo.collections.asr.metrics.wer import WER
from nemo.collections.asr.models.asr_model import ASRModel, ExportableEncDecModel
from nemo.collections.asr.modules.rnnt import RNNTDecoderJoint
from nemo.collections.asr.parts.mixins import (
ASRModuleMixin,
ASRTranscriptionMixin,
TranscribeConfig,
TranscriptionReturnType,
)
from nemo.collections.asr.parts.preprocessing.segment import ChannelSelectorType
from nemo.collections.asr.parts.submodules.rnnt_decoding import RNNTDecoding, RNNTDecodingConfig
from nemo.collections.asr.parts.utils.asr_batching import get_semi_sorted_batch_sampler
from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis
from nemo.collections.asr.parts.utils.timestamp_utils import process_timestamp_outputs
from nemo.collections.common.data.lhotse import get_lhotse_dataloader_from_config
from nemo.collections.common.parts.preprocessing.parsers import make_parser
from nemo.core.classes.common import PretrainedModelInfo, typecheck
from nemo.core.classes.mixins import AccessMixin
from nemo.core.neural_types import AcousticEncodedRepresentation, AudioSignal, LengthsType, NeuralType, SpectrogramType
from nemo.utils import logging
class EncDecRNNTModel(ASRModel, ASRModuleMixin, ExportableEncDecModel, ASRTranscriptionMixin):
"""Base class for encoder decoder RNNT-based models."""
def __init__(self, cfg: DictConfig, trainer: Trainer = None):
# Get global rank and total number of GPU workers for IterableDataset partitioning, if applicable
# Global_rank and local_rank is set by LightningModule in Lightning 1.2.0
self.world_size = 1
if trainer is not None:
self.world_size = trainer.world_size
super().__init__(cfg=cfg, trainer=trainer)
# Initialize components
self.preprocessor = EncDecRNNTModel.from_config_dict(self.cfg.preprocessor)
self.encoder = EncDecRNNTModel.from_config_dict(self.cfg.encoder)
# Update config values required by components dynamically
with open_dict(self.cfg.decoder):
self.cfg.decoder.vocab_size = len(self.cfg.labels)
with open_dict(self.cfg.joint):
self.cfg.joint.num_classes = len(self.cfg.labels)
self.cfg.joint.vocabulary = self.cfg.labels
self.cfg.joint.jointnet.encoder_hidden = self.cfg.model_defaults.enc_hidden
self.cfg.joint.jointnet.pred_hidden = self.cfg.model_defaults.pred_hidden
self.decoder = EncDecRNNTModel.from_config_dict(self.cfg.decoder)
self.joint = EncDecRNNTModel.from_config_dict(self.cfg.joint)
# Setup RNNT Loss
loss_name, loss_kwargs = self.extract_rnnt_loss_cfg(self.cfg.get("loss", None))
num_classes = self.joint.num_classes_with_blank - 1 # for standard RNNT and multi-blank
if loss_name == 'tdt':
num_classes = num_classes - self.joint.num_extra_outputs
self.loss = RNNTLoss(
num_classes=num_classes,
loss_name=loss_name,
loss_kwargs=loss_kwargs,
reduction=self.cfg.get("rnnt_reduction", "mean_batch"),
)
if hasattr(self.cfg, 'spec_augment') and self._cfg.spec_augment is not None:
self.spec_augmentation = EncDecRNNTModel.from_config_dict(self.cfg.spec_augment)
else:
self.spec_augmentation = None
self.cfg.decoding = self.set_decoding_type_according_to_loss(self.cfg.decoding)
# Setup decoding objects
self.decoding = RNNTDecoding(
decoding_cfg=self.cfg.decoding,
decoder=self.decoder,
joint=self.joint,
vocabulary=self.joint.vocabulary,
)
# Setup WER calculation
self.wer = WER(
decoding=self.decoding,
batch_dim_index=0,
use_cer=self._cfg.get('use_cer', False),
log_prediction=self._cfg.get('log_prediction', True),
dist_sync_on_step=True,
)
# Whether to compute loss during evaluation
if 'compute_eval_loss' in self.cfg:
self.compute_eval_loss = self.cfg.compute_eval_loss
else:
self.compute_eval_loss = True
# Setup fused Joint step if flag is set
if self.joint.fuse_loss_wer or (
self.decoding.joint_fused_batch_size is not None and self.decoding.joint_fused_batch_size > 0
):
self.joint.set_loss(self.loss)
self.joint.set_wer(self.wer)
# Setup optimization normalization (if provided in config)
self.setup_optim_normalization()
# Setup optional Optimization flags
self.setup_optimization_flags()
# Setup encoder adapters (from ASRAdapterModelMixin)
self.setup_adapters()
def setup_optim_normalization(self):
"""
Helper method to setup normalization of certain parts of the model prior to the optimization step.
Supported pre-optimization normalizations are as follows:
.. code-block:: yaml
# Variation Noise injection
model:
variational_noise:
std: 0.0
start_step: 0
# Joint - Length normalization
model:
normalize_joint_txu: false
# Encoder Network - gradient normalization
model:
normalize_encoder_norm: false
# Decoder / Prediction Network - gradient normalization
model:
normalize_decoder_norm: false
# Joint - gradient normalization
model:
normalize_joint_norm: false
"""
# setting up the variational noise for the decoder
if hasattr(self.cfg, 'variational_noise'):
self._optim_variational_noise_std = self.cfg['variational_noise'].get('std', 0)
self._optim_variational_noise_start = self.cfg['variational_noise'].get('start_step', 0)
else:
self._optim_variational_noise_std = 0
self._optim_variational_noise_start = 0
# Setup normalized gradients for model joint by T x U scaling factor (joint length normalization)
self._optim_normalize_joint_txu = self.cfg.get('normalize_joint_txu', False)
self._optim_normalize_txu = None
# Setup normalized encoder norm for model
self._optim_normalize_encoder_norm = self.cfg.get('normalize_encoder_norm', False)
# Setup normalized decoder norm for model
self._optim_normalize_decoder_norm = self.cfg.get('normalize_decoder_norm', False)
# Setup normalized joint norm for model
self._optim_normalize_joint_norm = self.cfg.get('normalize_joint_norm', False)
def extract_rnnt_loss_cfg(self, cfg: Optional[DictConfig]):
"""
Helper method to extract the rnnt loss name, and potentially its kwargs
to be passed.
Args:
cfg: Should contain `loss_name` as a string which is resolved to a RNNT loss name.
If the default should be used, then `default` can be used.
Optionally, one can pass additional kwargs to the loss function. The subdict
should have a keyname as follows : `{loss_name}_kwargs`.
Note that whichever loss_name is selected, that corresponding kwargs will be
selected. For the "default" case, the "{resolved_default}_kwargs" will be used.
Examples:
.. code-block:: yaml
loss_name: "default"
warprnnt_numba_kwargs:
kwargs2: some_other_val
Returns:
A tuple, the resolved loss name as well as its kwargs (if found).
"""
if cfg is None:
cfg = DictConfig({})
loss_name = cfg.get("loss_name", "default")
if loss_name == "default":
loss_name = resolve_rnnt_default_loss_name()
loss_kwargs = cfg.get(f"{loss_name}_kwargs", None)
logging.info(f"Using RNNT Loss : {loss_name}\n" f"Loss {loss_name}_kwargs: {loss_kwargs}")
return loss_name, loss_kwargs
def set_decoding_type_according_to_loss(self, decoding_cfg):
loss_name, loss_kwargs = self.extract_rnnt_loss_cfg(self.cfg.get("loss", None))
if loss_name == 'tdt':
decoding_cfg.durations = loss_kwargs.durations
elif loss_name == 'multiblank_rnnt':
decoding_cfg.big_blank_durations = loss_kwargs.big_blank_durations
return decoding_cfg
@torch.no_grad()
def transcribe(
self,
audio: Union[str, List[str], np.ndarray, DataLoader],
use_lhotse: bool = True,
batch_size: int = 4,
return_hypotheses: bool = False,
partial_hypothesis: Optional[List['Hypothesis']] = None,
num_workers: int = 0,
channel_selector: Optional[ChannelSelectorType] = None,
augmentor: DictConfig = None,
verbose: bool = True,
timestamps: Optional[bool] = None,
override_config: Optional[TranscribeConfig] = None,
) -> TranscriptionReturnType:
"""
Uses greedy decoding to transcribe audio files. Use this method for debugging and prototyping.
Args:
audio: (a single or list) of paths to audio files or a np.ndarray/tensor audio array or path
to a manifest file.
Can also be a dataloader object that provides values that can be consumed by the model.
Recommended length per file is between 5 and 25 seconds. \
But it is possible to pass a few hours long file if enough GPU memory is available.
use_lhotse: (bool) If audio is not a dataloder, defines whether to create a lhotse dataloader or a
non-lhotse dataloader.
batch_size: (int) batch size to use during inference. \
Bigger will result in better throughput performance but would use more memory.
return_hypotheses: (bool) Either return hypotheses or text
With hypotheses can do some postprocessing like getting timestamp or rescoring
partial_hypothesis: Optional[List['Hypothesis']] - A list of partial hypotheses to be used during rnnt
decoding. This is useful for streaming rnnt decoding. If this is not None, then the length of this
list should be equal to the length of the audio list.
num_workers: (int) number of workers for DataLoader
channel_selector (int | Iterable[int] | str): select a single channel or a subset of channels
from multi-channel audio. If set to `'average'`, it performs averaging across channels.
Disabled if set to `None`. Defaults to `None`. Uses zero-based indexing.
augmentor: (DictConfig): Augment audio samples during transcription if augmentor is applied.
verbose: (bool) whether to display tqdm progress bar
timestamps: Optional(Bool): timestamps will be returned if set to True as part of hypothesis object
(output.timestep['segment']/output.timestep['word']). Refer to `Hypothesis` class for more details.
Default is None and would retain the previous state set by using self.change_decoding_strategy().
override_config: (Optional[TranscribeConfig]) override transcription config pre-defined by the user.
**Note**: All other arguments in the function will be ignored if override_config is passed.
You should call this argument as `model.transcribe(audio, override_config=TranscribeConfig(...))`.
Returns:
Returns a tuple of 2 items -
* A list of greedy transcript texts / Hypothesis
* An optional list of beam search transcript texts / Hypothesis / NBestHypothesis.
"""
timestamps = timestamps or (override_config.timestamps if override_config is not None else None)
if timestamps is not None:
if timestamps or (override_config is not None and override_config.timestamps):
logging.info(
"Timestamps requested, setting decoding timestamps to True. Capture them in Hypothesis object, \
with output[0][idx].timestep['word'/'segment'/'char']"
)
return_hypotheses = True
with open_dict(self.cfg.decoding):
self.cfg.decoding.compute_timestamps = True
else:
return_hypotheses = False
with open_dict(self.cfg.decoding):
self.cfg.decoding.compute_timestamps = False
self.change_decoding_strategy(self.cfg.decoding, verbose=False)
return super().transcribe(
audio=audio,
use_lhotse=use_lhotse,
batch_size=batch_size,
return_hypotheses=return_hypotheses,
num_workers=num_workers,
channel_selector=channel_selector,
augmentor=augmentor,
verbose=verbose,
timestamps=timestamps,
override_config=override_config,
# Additional arguments
partial_hypothesis=partial_hypothesis,
)
def change_vocabulary(self, new_vocabulary: List[str], decoding_cfg: Optional[DictConfig] = None):
"""
Changes vocabulary used during RNNT decoding process. Use this method when fine-tuning a
pre-trained model. This method changes only decoder and leaves encoder and pre-processing
modules unchanged. For example, you would use it if you want to use pretrained encoder when
fine-tuning on data in another language, or when you'd need model to learn capitalization,
punctuation and/or special characters.
Args:
new_vocabulary: list with new vocabulary. Must contain at least 2 elements. Typically, \
this is target alphabet.
decoding_cfg: A config for the decoder, which is optional. If the decoding type
needs to be changed (from say Greedy to Beam decoding etc), the config can be passed here.
Returns: None
"""
if self.joint.vocabulary == new_vocabulary:
logging.warning(f"Old {self.joint.vocabulary} and new {new_vocabulary} match. Not changing anything.")
else:
if new_vocabulary is None or len(new_vocabulary) == 0:
raise ValueError(f'New vocabulary must be non-empty list of chars. But I got: {new_vocabulary}')
joint_config = self.joint.to_config_dict()
new_joint_config = copy.deepcopy(joint_config)
new_joint_config['vocabulary'] = new_vocabulary
new_joint_config['num_classes'] = len(new_vocabulary)
del self.joint
self.joint = EncDecRNNTModel.from_config_dict(new_joint_config)
decoder_config = self.decoder.to_config_dict()
new_decoder_config = copy.deepcopy(decoder_config)
new_decoder_config.vocab_size = len(new_vocabulary)
del self.decoder
self.decoder = EncDecRNNTModel.from_config_dict(new_decoder_config)
del self.loss
loss_name, loss_kwargs = self.extract_rnnt_loss_cfg(self.cfg.get('loss', None))
self.loss = RNNTLoss(
num_classes=self.joint.num_classes_with_blank - 1, loss_name=loss_name, loss_kwargs=loss_kwargs
)
if decoding_cfg is None:
# Assume same decoding config as before
decoding_cfg = self.cfg.decoding
# Assert the decoding config with all hyper parameters
decoding_cls = OmegaConf.structured(RNNTDecodingConfig)
decoding_cls = OmegaConf.create(OmegaConf.to_container(decoding_cls))
decoding_cfg = OmegaConf.merge(decoding_cls, decoding_cfg)
decoding_cfg = self.set_decoding_type_according_to_loss(decoding_cfg)
self.decoding = RNNTDecoding(
decoding_cfg=decoding_cfg,
decoder=self.decoder,
joint=self.joint,
vocabulary=self.joint.vocabulary,
)
self.wer = WER(
decoding=self.decoding,
batch_dim_index=self.wer.batch_dim_index,
use_cer=self.wer.use_cer,
log_prediction=self.wer.log_prediction,
dist_sync_on_step=True,
)
# Setup fused Joint step
if self.joint.fuse_loss_wer or (
self.decoding.joint_fused_batch_size is not None and self.decoding.joint_fused_batch_size > 0
):
self.joint.set_loss(self.loss)
self.joint.set_wer(self.wer)
# Update config
with open_dict(self.cfg.joint):
self.cfg.joint = new_joint_config
with open_dict(self.cfg.decoder):
self.cfg.decoder = new_decoder_config
with open_dict(self.cfg.decoding):
self.cfg.decoding = decoding_cfg
ds_keys = ['train_ds', 'validation_ds', 'test_ds']
for key in ds_keys:
if key in self.cfg:
with open_dict(self.cfg[key]):
self.cfg[key]['labels'] = OmegaConf.create(new_vocabulary)
logging.info(f"Changed decoder to output to {self.joint.vocabulary} vocabulary.")
def change_decoding_strategy(self, decoding_cfg: DictConfig, verbose=True):
"""
Changes decoding strategy used during RNNT decoding process.
Args:
decoding_cfg: A config for the decoder, which is optional. If the decoding type
needs to be changed (from say Greedy to Beam decoding etc), the config can be passed here.
verbose: (bool) whether to display logging information
"""
if decoding_cfg is None:
# Assume same decoding config as before
logging.info("No `decoding_cfg` passed when changing decoding strategy, using internal config")
decoding_cfg = self.cfg.decoding
# Assert the decoding config with all hyper parameters
decoding_cls = OmegaConf.structured(RNNTDecodingConfig)
decoding_cls = OmegaConf.create(OmegaConf.to_container(decoding_cls))
decoding_cfg = OmegaConf.merge(decoding_cls, decoding_cfg)
decoding_cfg = self.set_decoding_type_according_to_loss(decoding_cfg)
self.decoding = RNNTDecoding(
decoding_cfg=decoding_cfg,
decoder=self.decoder,
joint=self.joint,
vocabulary=self.joint.vocabulary,
)
self.wer = WER(
decoding=self.decoding,
batch_dim_index=self.wer.batch_dim_index,
use_cer=self.wer.use_cer,
log_prediction=self.wer.log_prediction,
dist_sync_on_step=True,
)
# Setup fused Joint step
if self.joint.fuse_loss_wer or (
self.decoding.joint_fused_batch_size is not None and self.decoding.joint_fused_batch_size > 0
):
self.joint.set_loss(self.loss)
self.joint.set_wer(self.wer)
self.joint.temperature = decoding_cfg.get('temperature', 1.0)
# Update config
with open_dict(self.cfg.decoding):
self.cfg.decoding = decoding_cfg
if verbose:
logging.info(f"Changed decoding strategy to \n{OmegaConf.to_yaml(self.cfg.decoding)}")
def _setup_dataloader_from_config(self, config: Optional[Dict]):
# Automatically inject args from model config to dataloader config
audio_to_text_dataset.inject_dataloader_value_from_model_config(self.cfg, config, key='sample_rate')
audio_to_text_dataset.inject_dataloader_value_from_model_config(self.cfg, config, key='labels')
if config.get("use_lhotse"):
return get_lhotse_dataloader_from_config(
config,
# During transcription, the model is initially loaded on the CPU.
# To ensure the correct global_rank and world_size are set,
# these values must be passed from the configuration.
global_rank=self.global_rank if not config.get("do_transcribe", False) else config.get("global_rank"),
world_size=self.world_size if not config.get("do_transcribe", False) else config.get("world_size"),
dataset=LhotseSpeechToTextBpeDataset(
tokenizer=make_parser(
labels=config.get('labels', None),
name=config.get('parser', 'en'),
unk_id=config.get('unk_index', -1),
blank_id=config.get('blank_index', -1),
do_normalize=config.get('normalize_transcripts', False),
),
return_cuts=config.get("do_transcribe", False),
),
)
dataset = audio_to_text_dataset.get_audio_to_text_char_dataset_from_config(
config=config,
local_rank=self.local_rank,
global_rank=self.global_rank,
world_size=self.world_size,
preprocessor_cfg=self._cfg.get("preprocessor", None),
)
if dataset is None:
return None
if isinstance(dataset, AudioToCharDALIDataset):
# DALI Dataset implements dataloader interface
return dataset
shuffle = config['shuffle']
if isinstance(dataset, torch.utils.data.IterableDataset):
shuffle = False
if hasattr(dataset, 'collate_fn'):
collate_fn = dataset.collate_fn
elif hasattr(dataset.datasets[0], 'collate_fn'):
# support datasets that are lists of entries
collate_fn = dataset.datasets[0].collate_fn
else:
# support datasets that are lists of lists
collate_fn = dataset.datasets[0].datasets[0].collate_fn
batch_sampler = None
if config.get('use_semi_sorted_batching', False):
if not isinstance(dataset, _AudioTextDataset):
raise RuntimeError(
"Semi Sorted Batch sampler can be used with AudioToCharDataset or AudioToBPEDataset "
f"but found dataset of type {type(dataset)}"
)
# set batch_size and batch_sampler to None to disable automatic batching
batch_sampler = get_semi_sorted_batch_sampler(self, dataset, config)
config['batch_size'] = None
config['drop_last'] = False
shuffle = False
return torch.utils.data.DataLoader(
dataset=dataset,
batch_size=config['batch_size'],
sampler=batch_sampler,
batch_sampler=None,
collate_fn=collate_fn,
drop_last=config.get('drop_last', False),
shuffle=shuffle,
num_workers=config.get('num_workers', 0),
pin_memory=config.get('pin_memory', False),
)
def setup_training_data(self, train_data_config: Optional[Union[DictConfig, Dict]]):
"""
Sets up the training data loader via a Dict-like object.
Args:
train_data_config: A config that contains the information regarding construction
of an ASR Training dataset.
Supported Datasets:
- :class:`~nemo.collections.asr.data.audio_to_text.AudioToCharDataset`
- :class:`~nemo.collections.asr.data.audio_to_text.AudioToBPEDataset`
- :class:`~nemo.collections.asr.data.audio_to_text.TarredAudioToCharDataset`
- :class:`~nemo.collections.asr.data.audio_to_text.TarredAudioToBPEDataset`
- :class:`~nemo.collections.asr.data.audio_to_text_dali.AudioToCharDALIDataset`
"""
if 'shuffle' not in train_data_config:
train_data_config['shuffle'] = True
# preserve config
self._update_dataset_config(dataset_name='train', config=train_data_config)
self._train_dl = self._setup_dataloader_from_config(config=train_data_config)
# Need to set this because if using an IterableDataset, the length of the dataloader is the total number
# of samples rather than the number of batches, and this messes up the tqdm progress bar.
# So we set the number of steps manually (to the correct number) to fix this.
if (
self._train_dl is not None
and hasattr(self._train_dl, 'dataset')
and isinstance(self._train_dl.dataset, torch.utils.data.IterableDataset)
):
# We also need to check if limit_train_batches is already set.
# If it's an int, we assume that the user has set it to something sane, i.e. <= # training batches,
# and don't change it. Otherwise, adjust batches accordingly if it's a float (including 1.0).
if self._trainer is not None and isinstance(self._trainer.limit_train_batches, float):
self._trainer.limit_train_batches = int(
self._trainer.limit_train_batches
* ceil((len(self._train_dl.dataset) / self.world_size) / train_data_config['batch_size'])
)
elif self._trainer is None:
logging.warning(
"Model Trainer was not set before constructing the dataset, incorrect number of "
"training batches will be used. Please set the trainer and rebuild the dataset."
)
def setup_validation_data(self, val_data_config: Optional[Union[DictConfig, Dict]]):
"""
Sets up the validation data loader via a Dict-like object.
Args:
val_data_config: A config that contains the information regarding construction
of an ASR Training dataset.
Supported Datasets:
- :class:`~nemo.collections.asr.data.audio_to_text.AudioToCharDataset`
- :class:`~nemo.collections.asr.data.audio_to_text.AudioToBPEDataset`
- :class:`~nemo.collections.asr.data.audio_to_text.TarredAudioToCharDataset`
- :class:`~nemo.collections.asr.data.audio_to_text.TarredAudioToBPEDataset`
- :class:`~nemo.collections.asr.data.audio_to_text_dali.AudioToCharDALIDataset`
"""
if 'shuffle' not in val_data_config:
val_data_config['shuffle'] = False
# preserve config
self._update_dataset_config(dataset_name='validation', config=val_data_config)
self._validation_dl = self._setup_dataloader_from_config(config=val_data_config)
def setup_test_data(self, test_data_config: Optional[Union[DictConfig, Dict]]):
"""
Sets up the test data loader via a Dict-like object.
Args:
test_data_config: A config that contains the information regarding construction
of an ASR Training dataset.
Supported Datasets:
- :class:`~nemo.collections.asr.data.audio_to_text.AudioToCharDataset`
- :class:`~nemo.collections.asr.data.audio_to_text.AudioToBPEDataset`
- :class:`~nemo.collections.asr.data.audio_to_text.TarredAudioToCharDataset`
- :class:`~nemo.collections.asr.data.audio_to_text.TarredAudioToBPEDataset`
- :class:`~nemo.collections.asr.data.audio_to_text_dali.AudioToCharDALIDataset`
"""
if 'shuffle' not in test_data_config:
test_data_config['shuffle'] = False
# preserve config
self._update_dataset_config(dataset_name='test', config=test_data_config)
self._test_dl = self._setup_dataloader_from_config(config=test_data_config)
@property
def input_types(self) -> Optional[Dict[str, NeuralType]]:
if hasattr(self.preprocessor, '_sample_rate'):
input_signal_eltype = AudioSignal(freq=self.preprocessor._sample_rate)
else:
input_signal_eltype = AudioSignal()
return {
"input_signal": NeuralType(('B', 'T'), input_signal_eltype, optional=True),
"input_signal_length": NeuralType(tuple('B'), LengthsType(), optional=True),
"processed_signal": NeuralType(('B', 'D', 'T'), SpectrogramType(), optional=True),
"processed_signal_length": NeuralType(tuple('B'), LengthsType(), optional=True),
}
@property
def output_types(self) -> Optional[Dict[str, NeuralType]]:
return {
"outputs": NeuralType(('B', 'D', 'T'), AcousticEncodedRepresentation()),
"encoded_lengths": NeuralType(tuple('B'), LengthsType()),
}
@typecheck()
def forward(
self, input_signal=None, input_signal_length=None, processed_signal=None, processed_signal_length=None
):
"""
Forward pass of the model. Note that for RNNT Models, the forward pass of the model is a 3 step process,
and this method only performs the first step - forward of the acoustic model.
Please refer to the `training_step` in order to see the full `forward` step for training - which
performs the forward of the acoustic model, the prediction network and then the joint network.
Finally, it computes the loss and possibly compute the detokenized text via the `decoding` step.
Please refer to the `validation_step` in order to see the full `forward` step for inference - which
performs the forward of the acoustic model, the prediction network and then the joint network.
Finally, it computes the decoded tokens via the `decoding` step and possibly compute the batch metrics.
Args:
input_signal: Tensor that represents a batch of raw audio signals,
of shape [B, T]. T here represents timesteps, with 1 second of audio represented as
`self.sample_rate` number of floating point values.
input_signal_length: Vector of length B, that contains the individual lengths of the audio
sequences.
processed_signal: Tensor that represents a batch of processed audio signals,
of shape (B, D, T) that has undergone processing via some DALI preprocessor.
processed_signal_length: Vector of length B, that contains the individual lengths of the
processed audio sequences.
Returns:
A tuple of 2 elements -
1) The log probabilities tensor of shape [B, T, D].
2) The lengths of the acoustic sequence after propagation through the encoder, of shape [B].
"""
has_input_signal = input_signal is not None and input_signal_length is not None
has_processed_signal = processed_signal is not None and processed_signal_length is not None
if (has_input_signal ^ has_processed_signal) is False:
raise ValueError(
f"{self} Arguments ``input_signal`` and ``input_signal_length`` are mutually exclusive "
" with ``processed_signal`` and ``processed_signal_len`` arguments."
)
if not has_processed_signal:
processed_signal, processed_signal_length = self.preprocessor(
input_signal=input_signal,
length=input_signal_length,
)
# Spec augment is not applied during evaluation/testing
if self.spec_augmentation is not None and self.training:
processed_signal = self.spec_augmentation(input_spec=processed_signal, length=processed_signal_length)
encoded, encoded_len = self.encoder(audio_signal=processed_signal, length=processed_signal_length)
return encoded, encoded_len
# PTL-specific methods
def training_step(self, batch, batch_nb):
# Reset access registry
if AccessMixin.is_access_enabled(self.model_guid):
AccessMixin.reset_registry(self)
signal, signal_len, transcript, transcript_len = batch
# forward() only performs encoder forward
if isinstance(batch, DALIOutputs) and batch.has_processed_signal:
encoded, encoded_len = self.forward(processed_signal=signal, processed_signal_length=signal_len)
else:
encoded, encoded_len = self.forward(input_signal=signal, input_signal_length=signal_len)
del signal
# During training, loss must be computed, so decoder forward is necessary
decoder, target_length, states = self.decoder(targets=transcript, target_length=transcript_len)
if hasattr(self, '_trainer') and self._trainer is not None:
log_every_n_steps = self._trainer.log_every_n_steps
sample_id = self._trainer.global_step
else:
log_every_n_steps = 1
sample_id = batch_nb
# If experimental fused Joint-Loss-WER is not used
if not self.joint.fuse_loss_wer:
# Compute full joint and loss
joint = self.joint(encoder_outputs=encoded, decoder_outputs=decoder)
loss_value = self.loss(
log_probs=joint, targets=transcript, input_lengths=encoded_len, target_lengths=target_length
)
# Add auxiliary losses, if registered
loss_value = self.add_auxiliary_losses(loss_value)
# Reset access registry
if AccessMixin.is_access_enabled(self.model_guid):
AccessMixin.reset_registry(self)
tensorboard_logs = {
'train_loss': loss_value,
'learning_rate': self._optimizer.param_groups[0]['lr'],
'global_step': torch.tensor(self.trainer.global_step, dtype=torch.float32),
}
if (sample_id + 1) % log_every_n_steps == 0:
self.wer.update(
predictions=encoded,
predictions_lengths=encoded_len,
targets=transcript,
targets_lengths=transcript_len,
)
_, scores, words = self.wer.compute()
self.wer.reset()
tensorboard_logs.update({'training_batch_wer': scores.float() / words})
else:
# If experimental fused Joint-Loss-WER is used
if (sample_id + 1) % log_every_n_steps == 0:
compute_wer = True
else:
compute_wer = False
# Fused joint step
loss_value, wer, _, _ = self.joint(
encoder_outputs=encoded,
decoder_outputs=decoder,
encoder_lengths=encoded_len,
transcripts=transcript,
transcript_lengths=transcript_len,
compute_wer=compute_wer,
)
# Add auxiliary losses, if registered
loss_value = self.add_auxiliary_losses(loss_value)
# Reset access registry
if AccessMixin.is_access_enabled(self.model_guid):
AccessMixin.reset_registry(self)
tensorboard_logs = {
'train_loss': loss_value,
'learning_rate': self._optimizer.param_groups[0]['lr'],
'global_step': torch.tensor(self.trainer.global_step, dtype=torch.float32),
}
if compute_wer:
tensorboard_logs.update({'training_batch_wer': wer})
# Log items
self.log_dict(tensorboard_logs)
# Preserve batch acoustic model T and language model U parameters if normalizing
if self._optim_normalize_joint_txu:
self._optim_normalize_txu = [encoded_len.max(), transcript_len.max()]
return {'loss': loss_value}
def predict_step(self, batch, batch_idx, dataloader_idx=0):
signal, signal_len, transcript, transcript_len, sample_id = batch
# forward() only performs encoder forward
if isinstance(batch, DALIOutputs) and batch.has_processed_signal:
encoded, encoded_len = self.forward(processed_signal=signal, processed_signal_length=signal_len)
else:
encoded, encoded_len = self.forward(input_signal=signal, input_signal_length=signal_len)
del signal
best_hyp_text = self.decoding.rnnt_decoder_predictions_tensor(
encoder_output=encoded, encoded_lengths=encoded_len, return_hypotheses=True
)
if isinstance(sample_id, torch.Tensor):
sample_id = sample_id.cpu().detach().numpy()
return list(zip(sample_id, best_hyp_text))
def validation_pass(self, batch, batch_idx, dataloader_idx=0):
signal, signal_len, transcript, transcript_len = batch
# forward() only performs encoder forward
if isinstance(batch, DALIOutputs) and batch.has_processed_signal:
encoded, encoded_len = self.forward(processed_signal=signal, processed_signal_length=signal_len)
else:
encoded, encoded_len = self.forward(input_signal=signal, input_signal_length=signal_len)
del signal
tensorboard_logs = {}
# If experimental fused Joint-Loss-WER is not used
if not self.joint.fuse_loss_wer:
if self.compute_eval_loss:
decoder, target_length, states = self.decoder(targets=transcript, target_length=transcript_len)
joint = self.joint(encoder_outputs=encoded, decoder_outputs=decoder)
loss_value = self.loss(
log_probs=joint, targets=transcript, input_lengths=encoded_len, target_lengths=target_length
)
tensorboard_logs['val_loss'] = loss_value
self.wer.update(
predictions=encoded,
predictions_lengths=encoded_len,
targets=transcript,
targets_lengths=transcript_len,
)
wer, wer_num, wer_denom = self.wer.compute()
self.wer.reset()
tensorboard_logs['val_wer_num'] = wer_num
tensorboard_logs['val_wer_denom'] = wer_denom
tensorboard_logs['val_wer'] = wer
else:
# If experimental fused Joint-Loss-WER is used
compute_wer = True
if self.compute_eval_loss:
decoded, target_len, states = self.decoder(targets=transcript, target_length=transcript_len)
else:
decoded = None
target_len = transcript_len
# Fused joint step
loss_value, wer, wer_num, wer_denom = self.joint(
encoder_outputs=encoded,
decoder_outputs=decoded,
encoder_lengths=encoded_len,
transcripts=transcript,
transcript_lengths=target_len,
compute_wer=compute_wer,
)
if loss_value is not None:
tensorboard_logs['val_loss'] = loss_value
tensorboard_logs['val_wer_num'] = wer_num
tensorboard_logs['val_wer_denom'] = wer_denom
tensorboard_logs['val_wer'] = wer
self.log('global_step', torch.tensor(self.trainer.global_step, dtype=torch.float32))
return tensorboard_logs
def validation_step(self, batch, batch_idx, dataloader_idx=0):
metrics = self.validation_pass(batch, batch_idx, dataloader_idx)
if type(self.trainer.val_dataloaders) == list and len(self.trainer.val_dataloaders) > 1:
self.validation_step_outputs[dataloader_idx].append(metrics)
else:
self.validation_step_outputs.append(metrics)
return metrics
def test_step(self, batch, batch_idx, dataloader_idx=0):
logs = self.validation_pass(batch, batch_idx, dataloader_idx=dataloader_idx)
test_logs = {name.replace("val_", "test_"): value for name, value in logs.items()}
if type(self.trainer.test_dataloaders) == list and len(self.trainer.test_dataloaders) > 1:
self.test_step_outputs[dataloader_idx].append(test_logs)
else:
self.test_step_outputs.append(test_logs)
return test_logs
def multi_validation_epoch_end(self, outputs, dataloader_idx: int = 0):
if self.compute_eval_loss:
val_loss_mean = torch.stack([x['val_loss'] for x in outputs]).mean()
val_loss_log = {'val_loss': val_loss_mean}
else:
val_loss_log = {}
wer_num = torch.stack([x['val_wer_num'] for x in outputs]).sum()
wer_denom = torch.stack([x['val_wer_denom'] for x in outputs]).sum()
tensorboard_logs = {**val_loss_log, 'val_wer': wer_num.float() / wer_denom}
return {**val_loss_log, 'log': tensorboard_logs}
def multi_test_epoch_end(self, outputs, dataloader_idx: int = 0):
if self.compute_eval_loss:
test_loss_mean = torch.stack([x['test_loss'] for x in outputs]).mean()
test_loss_log = {'test_loss': test_loss_mean}
else:
test_loss_log = {}
wer_num = torch.stack([x['test_wer_num'] for x in outputs]).sum()
wer_denom = torch.stack([x['test_wer_denom'] for x in outputs]).sum()
tensorboard_logs = {**test_loss_log, 'test_wer': wer_num.float() / wer_denom}
return {**test_loss_log, 'log': tensorboard_logs}
""" Transcription related methods """
def _transcribe_forward(self, batch: Any, trcfg: TranscribeConfig):
encoded, encoded_len = self.forward(input_signal=batch[0], input_signal_length=batch[1])
output = dict(encoded=encoded, encoded_len=encoded_len)
return output
def _transcribe_output_processing(
self, outputs, trcfg: TranscribeConfig
) -> Union[List['Hypothesis'], List[List['Hypothesis']]]:
encoded = outputs.pop('encoded')
encoded_len = outputs.pop('encoded_len')
hyp = self.decoding.rnnt_decoder_predictions_tensor(
encoded,
encoded_len,
return_hypotheses=trcfg.return_hypotheses,
partial_hypotheses=trcfg.partial_hypothesis,
)
# cleanup memory
del encoded, encoded_len
if trcfg.timestamps:
hyp = process_timestamp_outputs(
hyp, self.encoder.subsampling_factor, self.cfg['preprocessor']['window_stride']
)
return hyp
def _setup_transcribe_dataloader(self, config: Dict) -> 'torch.utils.data.DataLoader':
"""
Setup function for a temporary data loader which wraps the provided audio file.
Args:
config: A python dictionary which contains the following keys:
paths2audio_files: (a list) of paths to audio files. The files should be relatively short fragments. \
Recommended length per file is between 5 and 25 seconds.
batch_size: (int) batch size to use during inference. \
Bigger will result in better throughput performance but would use more memory.
temp_dir: (str) A temporary directory where the audio manifest is temporarily
stored.
Returns:
A pytorch DataLoader for the given audio file(s).
"""
if 'manifest_filepath' in config:
manifest_filepath = config['manifest_filepath']
batch_size = config['batch_size']
else:
manifest_filepath = os.path.join(config['temp_dir'], 'manifest.json')
batch_size = min(config['batch_size'], len(config['paths2audio_files']))
dl_config = {
'manifest_filepath': manifest_filepath,
'sample_rate': self.preprocessor._sample_rate,
'labels': self.joint.vocabulary,
'batch_size': batch_size,
'trim_silence': False,
'shuffle': False,
'num_workers': config.get('num_workers', min(batch_size, os.cpu_count() - 1)),
'pin_memory': True,
}
if config.get("augmentor"):
dl_config['augmentor'] = config.get("augmentor")