-
Notifications
You must be signed in to change notification settings - Fork 303
/
run_distillation.py
1811 lines (1607 loc) · 80.2 KB
/
run_distillation.py
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
#!/usr/bin/env python
# coding=utf-8
# Copyright 2023 The HuggingFace Inc. team. 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.
"""
Training the Whisper model for sequence to sequence speech recognition via teacher-student distillation.
"""
# You can also adapt this script for your own distillation tasks. Pointers for this are left as comments.
import logging
import os
import re
import shutil
import sys
import time
from dataclasses import dataclass, field
from functools import partial
from pathlib import Path
from typing import Any, Dict, List, Optional, Union
import datasets
import evaluate
import numpy as np
import torch
import torch.nn as nn
import transformers
from accelerate import Accelerator
from accelerate.logging import get_logger
from accelerate.utils import set_seed
from datasets import (
DatasetDict,
IterableDataset,
IterableDatasetDict,
concatenate_datasets,
interleave_datasets,
load_dataset,
)
from huggingface_hub import create_repo, get_full_repo_name, upload_folder
from torch.utils.data import DataLoader
from tqdm import tqdm
from transformers import (
AddedToken,
HfArgumentParser,
Seq2SeqTrainingArguments,
WhisperConfig,
WhisperFeatureExtractor,
WhisperForConditionalGeneration,
WhisperProcessor,
WhisperTokenizerFast,
get_scheduler
)
from transformers.modeling_outputs import BaseModelOutput
from transformers.models.whisper.english_normalizer import BasicTextNormalizer, EnglishTextNormalizer
from transformers.utils import check_min_version
from transformers.utils.versions import require_version
# Will error if the minimal version of Transformers is not installed. Remove at your own risks.
check_min_version("4.34.0.dev0")
require_version("datasets>=2.14.6", "To fix: `pip install --upgrade datasets`")
logger = get_logger(__name__)
@dataclass
class ModelArguments:
"""
Arguments pertaining to which model/config/tokenizer we are going to distill from.
"""
model_name_or_path: str = field(
metadata={"help": "Path to pretrained Whisper model or model identifier from huggingface.co/models"}
)
teacher_model_name_or_path: str = field(
metadata={"help": "Path to pretrained teacher model or model identifier from huggingface.co/models"}
)
config_name: Optional[str] = field(
default=None,
metadata={"help": "Pretrained config name or path if not the same as model_name"},
)
tokenizer_name: Optional[str] = field(
default=None,
metadata={"help": "Pretrained tokenizer name or path if not the same as model_name"},
)
feature_extractor_name: Optional[str] = field(
default=None,
metadata={"help": "feature extractor name or path if not the same as model_name"},
)
cache_dir: Optional[str] = field(
default=None,
metadata={"help": "Where to store the pretrained models downloaded from huggingface.co"},
)
use_fast_tokenizer: bool = field(
default=True,
metadata={"help": "Whether to use one of the fast tokenizer (backed by the tokenizers library) or not."},
)
model_revision: str = field(
default="main",
metadata={"help": "The specific model version to use (can be a branch name, tag name or commit id)."},
)
subfolder: str = field(
default="",
metadata={
"help": "In case the relevant files are located inside a subfolder of the model repo on huggingface.co, you can"
"specify the folder name here."
},
)
token: str = field(
default=None,
metadata={
"help": (
"The token to use as HTTP bearer authorization for remote files. If not specified, will use the token "
"generated when running `huggingface-cli login` (stored in `~/.huggingface`)."
)
},
)
attn_implementation: Optional[str] = field(
default=None,
metadata={
"help": (
"Which attention implementation to use in the encoder and decoder attention layers. Can be one of:\n"
"1. `eager` or `None`: default Transformers attention implementation.\n"
"2. `sdpa`: Flash Attention through PyTorch SDPA. Requires `torch>=2.1`. Recommended for hardware where Flash Attention 2 is not supported, e.g. Turing GPUs, (T4, RTX 2080).\n"
"3. `flash_attn_2`: Flash Attention 2 through the Flash Attention package https://github.com/Dao-AILab/flash-attention. **Always** recommended on supported hardware (Ampere, Ada, or Hopper GPUs, e.g., A100, RTX 3090, RTX 4090, H100)."
)
},
)
def __post_init__(self):
if self.attn_implementation not in [None, "eager", "sdpa", "flash_attention_2"]:
raise ValueError(
f"Got `--attn_implementation={self.attn_implementation}`, which is an invalid attention type. Should be one of:\n"
"1. `eager` or `None`: default Transformers attention implementation.\n"
"2. `sdpa`: Flash Attention through PyTorch SDPA. Requires `torch>=2.1`. Recommended for hardware where Flash Attention 2 is not supported, e.g. Turing GPUs, (T4, RTX 2080).\n"
"3. `flash_attn_2`: Flash Attention 2 through the Flash Attention package https://github.com/Dao-AILab/flash-attention. **Always** recommended on supported hardware (Ampere, Ada, or Hopper GPUs, e.g., A100, RTX 3090, RTX 4090, H100)."
)
@dataclass
class DataTrainingArguments:
"""
Arguments pertaining to what data we are going to input our model for training and eval.
"""
train_dataset_name: str = field(
default=None,
metadata={
"help": "The name of the training dataset to use (via the datasets library). Load and combine "
"multiple datasets by separating dataset ids by a '+' symbol. For example, to load LibriSpeech "
"and Common Voice, set `train_dataset_name='librispeech_asr+common_voice'`."
},
)
train_dataset_config_name: Optional[str] = field(
default=None,
metadata={
"help": "The configuration name of the training dataset to use (via the datasets library). Load and combine "
"multiple datasets by separating dataset configs by a '+' symbol. Note that the order of the configs should "
"match the order of the datasets."
},
)
train_dataset_samples: str = field(
default=None,
metadata={
"help": "Number of samples in each dataset when loading multiple datasets with streaming mode. "
"Not required when using one dataset or non-streaming mode. The sample values provide the sampling "
"probability for each dataset. Setting them equal to the number of sample values ensures that every "
"sample from every dataset is used once per epoch."
},
)
eval_dataset_name: str = field(
default=None,
metadata={
"help": "The name of the evaluation dataset to use (via the datasets library). Defaults to the training "
"dataset name if unspecified. Load multiple evaluation datasets by separating dataset "
"ids by a '+' symbol."
},
)
eval_dataset_config_name: Optional[str] = field(
default=None,
metadata={
"help": "The configuration name of the evaluation dataset to use (via the datasets library). Defaults to the "
"training dataset config name if unspecified."
},
)
dataset_cache_dir: Optional[str] = field(
default=None,
metadata={"help": "Path to cache directory for saving and loading datasets"},
)
overwrite_cache: bool = field(
default=False,
metadata={"help": "Overwrite the cached training and evaluation sets"},
)
preprocessing_num_workers: Optional[int] = field(
default=None,
metadata={"help": "The number of processes to use for the preprocessing if using non-streaming mode."},
)
preprocessing_batch_size: Optional[int] = field(
default=256,
metadata={"help": "Number of examples per batch provided to the `prepare_dataset` function."},
)
max_train_samples: Optional[int] = field(
default=None,
metadata={
"help": (
"For debugging purposes or quicker training, truncate the number of training examples to this value if set."
)
},
)
max_eval_samples: Optional[int] = field(
default=None,
metadata={
"help": (
"For debugging purposes or quicker training, truncate the number of evaluation examples to this value if set."
)
},
)
audio_column_name: str = field(
default="audio",
metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
)
text_column_name: str = field(
default=None,
metadata={"help": "The name of the dataset column containing the text data in the training set."},
)
eval_text_column_name: str = field(
default="text",
metadata={"help": ("The name of the dataset column containing the text data in the evaluation set.")},
)
max_duration_in_seconds: float = field(
default=30.0,
metadata={"help": "Filter audio files that are longer than `max_duration_in_seconds` seconds"},
)
min_duration_in_seconds: float = field(
default=0.0,
metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"},
)
max_label_length: int = field(
default=448,
metadata={"help": "Truncate transcriptions that are longer `max_label_length` tokens."},
)
pad_target_to_multiple_of: Optional[int] = field(
default=None,
metadata={
"help": (
"If set will pad the target sequence to a multiple of the provided"
" value. This is important to avoid triggering recompilations on TPU."
" If unspecified, will default to padding the targets to max length."
)
},
)
preprocessing_only: bool = field(
default=False,
metadata={
"help": (
"Whether to only do data preprocessing and skip training. This is"
" especially useful when data preprocessing errors out in distributed"
" training due to timeout. In this case, one should run the"
" preprocessing in a non-distributed setup with"
" `preprocessing_only=True` so that the cached datasets can"
" consequently be loaded in distributed training"
)
},
)
train_split_name: str = field(
default="train",
metadata={
"help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
},
)
eval_split_name: str = field(
default="validation",
metadata={
"help": (
"The name of the evaluation data set split to use (via the datasets library). Defaults to 'validation'"
)
},
)
streaming: bool = field(
default=True,
metadata={"help": "Whether to use Datasets' streaming mode to load and pre-process the data."},
)
wer_threshold: float = field(
default=None,
metadata={
"help": "Filter training data with Whisper transcriptions that have greater than `wer_threshold` "
"WER with the normalised transcriptions. This only takes effect if training on pseudo-labels targets."
"If `--use_pseudo_labels=False`, then no WER filtering is performed, since we train directly on the text"
"transcriptions."
},
)
use_pseudo_labels: bool = field(
default=True,
metadata={
"help": "Whether or not to use pseudo-label transcriptions as the targets. If True, the pseudo-labels "
"must be in the dataset column `whisper_transcript` from the previous pseudo-labelling step. This is "
"not currently yet configurable."
},
)
timestamp_probability: float = field(
default=0.2, metadata={"help": "Probability for training on timestamped tokens if the data contains it."}
)
condition_on_prev_probability: float = field(
default=0.2, metadata={"help": "Probability for conditioning on the previous text example."}
)
return_timestamps: bool = field(
default=False, metadata={"help": "Whether or not to predict timestamps in the generation step."}
)
language: str = field(
default=None,
metadata={
"help": (
"Language for multilingual distillation. This argument should be set for multilingual distillation "
"only. For English speech recognition, it should be left as `None`."
)
},
)
task: str = field(
default="transcribe",
metadata={
"help": "Task, either `transcribe` for speech recognition or `translate` for speech translation."
"This argument should be set for multilingual distillation only. For English speech recognition, it should be left as `None`."
},
)
wandb_project: str = field(
default="distil-whisper",
metadata={"help": "The name of the wandb project."},
)
wandb_name: str = field(
default=None,
metadata={"help": "The name of the wandb run."},
)
wandb_dir: str = field(
default="./wandb",
metadata={"help": "The dir where wandb metadata will be stored."},
)
@dataclass
class DistillationTrainingArguments(Seq2SeqTrainingArguments):
freeze_encoder: Optional[bool] = field(
default=False,
metadata={
"help": (
"Whether to freeze the entire encoder model. Only recommended when the entire encoder has been "
"copied from the teacher model."
)
},
)
freeze_decoder: Optional[bool] = field(
default=False,
metadata={
"help": (
"Whether to freeze the entire decoder model. Note that the decoder input embeddings are **not** frozen, since they are tied to the LM head."
)
},
)
freeze_embed_positions: Optional[bool] = field(
default=False,
metadata={"help": "Whether to freeze the decoder embedding positions."},
)
temperature: Optional[float] = field(
default=2.0, metadata={"help": "Temperature to anneal the logits when computing the softmax."}
)
kl_weight: Optional[float] = field(
default=1.0,
metadata={
"help": (
"Weighting assigned to the MSE loss in the KD formulation. MSE loss is "
"computed between the teacher-student hidden states and attentions."
)
},
)
dtype: Optional[str] = field(
default="float32",
metadata={
"help": (
"The data type (dtype) in which to run training. One of `float32` (full-precision), "
"`float16` or `bfloat16` (both half-precision)."
)
},
)
save_best_total_limit: Optional[int] = field(
default=1,
metadata={
"help": (
"Number of best models to be saved."
)
}
)
@dataclass
class DataCollatorSpeechSeq2SeqWithPadding:
"""
Data collator that will dynamically pad the inputs received.
Args:
processor ([`Wav2Vec2Processor`])
The processor used for proccessing the data.
decoder_start_token_id (:obj: `int`)
The start-of-sequence token id of the decoder.
decoder_prev_token_id (:obj: `int`)
The start-of-prompt token id of the decoder
input_padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
Select a strategy to pad the returned input sequences (according to the model's padding side and padding index)
among:
* :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
sequence if provided).
* :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
maximum acceptable input length for the model if that argument is not provided.
* :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
different lengths).
target_padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
Select a strategy to pad the returned target sequences (according to the model's padding side and padding index).
See above for details.
max_target_length (:obj:`int`, `optional`):
Maximum length of the ``labels`` of the returned list and optionally padding length (see above).
"""
processor: Any
decoder_start_token_id: int
decoder_prev_token_id: int
input_padding: Union[bool, str] = "max_length"
target_padding: Union[bool, str] = "max_length"
max_target_length: Optional[int] = None
def __call__(self, features: List[Dict[str, Union[List[int], np.ndarray]]]) -> Dict[str, np.ndarray]:
# split inputs and labels since they have to be of different lengths and need
# different padding methods
# dataloader returns a list of features which we convert to a dict
input_features = {"input_features": [feature["input_features"] for feature in features]}
label_features = {"input_ids": [feature["labels"] for feature in features]}
# reformat list to dict and set to pytorch format
batch = self.processor.feature_extractor.pad(
input_features,
padding=self.input_padding,
return_tensors="pt",
)
labels_batch = self.processor.tokenizer.pad(
label_features,
max_length=self.max_target_length,
padding=self.target_padding,
return_tensors="pt",
)
# shift labels to the right to get decoder input ids
labels = labels_batch["input_ids"]
decoder_input_ids = labels[:, :-1]
labels = labels[:, 1:]
labels_mask = labels_batch.attention_mask[:, 1:]
# replace padding with -100 to ignore correctly when computing the loss
labels = labels.masked_fill(labels_mask.ne(1), -100)
# replace initial prompt tokens with -100 to ignore correctly when computing the loss
bos_index = torch.argmax((labels == self.decoder_start_token_id).long(), dim=1)
bos_index = torch.where(bos_index > 0, bos_index + 1, bos_index)
prompt_mask = torch.arange(labels.shape[1]) < bos_index[:, None]
labels = torch.where(prompt_mask, -100, labels)
batch["labels"] = labels
batch["decoder_input_ids"] = decoder_input_ids
return batch
def log_metric(
accelerator,
metrics: Dict,
train_time: float,
step: int,
epoch: int,
learning_rate: float = None,
prefix: str = "train",
):
"""Helper function to log all training/evaluation metrics with the correct prefixes and styling."""
log_metrics = {}
for k, v in metrics.items():
log_metrics[f"{prefix}/{k}"] = v
log_metrics[f"{prefix}/time"] = train_time
log_metrics[f"{prefix}/epoch"] = epoch
if learning_rate is not None:
log_metrics[f"{prefix}/learning_rate"] = learning_rate
accelerator.log(log_metrics, step=step)
def log_pred(
accelerator,
pred_str: List[str],
label_str: List[str],
norm_pred_str: List[str],
norm_label_str: List[str],
step: int,
prefix: str = "eval",
num_lines: int = 200000,
):
"""Helper function to log target/predicted transcriptions to weights and biases (wandb)."""
if accelerator.is_main_process:
wandb_tracker = accelerator.get_tracker("wandb")
# pretty name for current step: step 50000 -> step 50k
cur_step_pretty = f"{int(step // 1000)}k" if step > 1000 else step
prefix_pretty = prefix.replace("/", "-")
# convert str data to a wandb compatible format
str_data = [[label_str[i], pred_str[i], norm_label_str[i], norm_pred_str[i]] for i in range(len(pred_str))]
# log as a table with the appropriate headers
wandb_tracker.log_table(
table_name=f"predictions/{prefix_pretty}-step-{cur_step_pretty}",
columns=["Target", "Pred", "Norm Target", "Norm Pred"],
data=str_data[:num_lines],
step=step,
)
# log incorrect normalised predictions
str_data = np.asarray(str_data)
str_data_incorrect = str_data[str_data[:, -2] != str_data[:, -1]]
# log as a table with the appropriate headers
wandb_tracker.log_table(
table_name=f"incorrect_predictions/{prefix_pretty}-step-{cur_step_pretty}",
columns=["Target", "Pred", "Norm Target", "Norm Pred"],
data=str_data_incorrect[:num_lines],
step=step,
)
def convert_dataset_str_to_list(
dataset_names,
dataset_config_names,
splits=None,
text_column_names=None,
dataset_samples=None,
default_split="train",
) -> List[Dict]:
"""
Given three lists of dataset names, configs and splits, this function groups the corresponding
names/configs/splits. Each dataset is assigned a unique dictionary with these metadata values, and the
function returns a list of dictionaries, one for each dataset.
"""
if isinstance(dataset_names, str):
dataset_names = dataset_names.split("+")
dataset_config_names = dataset_config_names.split("+") if dataset_config_names is not None else None
splits = splits.split("+") if splits is not None else None
text_column_names = text_column_names.split("+") if text_column_names is not None else None
dataset_samples = dataset_samples.split("+") if dataset_samples is not None else None
# basic checks to ensure we've got the right number of datasets/configs/splits/columns/probs
if dataset_config_names is not None and len(dataset_names) != len(dataset_config_names):
raise ValueError(
f"Ensure one config is passed for each dataset, got {len(dataset_names)} datasets and"
f" {len(dataset_config_names)} configs."
)
if splits is not None and len(splits) != len(dataset_names):
raise ValueError(
f"Ensure one split is passed for each dataset, got {len(dataset_names)} datasets and {len(splits)} splits."
)
if text_column_names is not None and len(text_column_names) != len(dataset_names):
raise ValueError(
f"Ensure one text column name is passed for each dataset, got {len(dataset_names)} datasets and"
f" {len(text_column_names)} text column names."
)
if dataset_samples is not None:
if len(dataset_samples) != len(dataset_names):
raise ValueError(
f"Ensure one sample is passed for each dataset, got {len(dataset_names)} datasets and "
f"{len(dataset_samples)} samples."
)
dataset_samples = [float(ds_sample) for ds_sample in dataset_samples]
else:
dataset_samples = [None] * len(dataset_names)
dataset_config_names = (
dataset_config_names if dataset_config_names is not None else ["default" for _ in range(len(dataset_names))]
)
text_column_names = (
text_column_names if text_column_names is not None else ["text" for _ in range(len(dataset_names))]
)
splits = splits if splits is not None else [default_split for _ in range(len(dataset_names))]
dataset_names_dict = []
for i, ds_name in enumerate(dataset_names):
dataset_names_dict.append(
{
"name": ds_name,
"config": dataset_config_names[i],
"split": splits[i],
"text_column_name": text_column_names[i],
"samples": dataset_samples[i],
}
)
return dataset_names_dict
def load_multiple_datasets(
dataset_names: Union[List, str],
dataset_config_names: Union[List, str],
splits: Optional[Union[List, str]] = None,
text_column_names: Optional[List] = None,
sampling_rate: Optional[int] = 16000,
stopping_strategy: Optional[str] = "first_exhausted",
dataset_samples: Optional[Union[List, np.array]] = None,
streaming: Optional[bool] = True,
seed: Optional[int] = None,
accelerator: Optional[Accelerator] = None,
use_pseudo_labels: float = None,
**kwargs,
) -> IterableDataset:
dataset_names_dict = convert_dataset_str_to_list(
dataset_names, dataset_config_names, splits, text_column_names, dataset_samples
)
if dataset_samples is not None:
dataset_samples = [ds_dict["samples"] for ds_dict in dataset_names_dict]
probabilities = np.array(dataset_samples) / np.sum(dataset_samples)
else:
probabilities = None
all_datasets = []
# iterate over the datasets we want to interleave
for dataset_dict in tqdm(
dataset_names_dict,
desc="Combining datasets...",
disable=not accelerator.is_local_main_process if accelerator is not None else False,
):
dataset = load_dataset(
dataset_dict["name"],
dataset_dict["config"],
split=dataset_dict["split"],
streaming=streaming,
**kwargs,
)
# resample to specified sampling rate
dataset = dataset.cast_column("audio", datasets.features.Audio(sampling_rate))
dataset_features = dataset.features.keys()
columns_to_keep = {"audio", "text"}
if dataset_dict["text_column_name"] not in dataset_features:
raise ValueError(
f"Text column name {dataset_dict['text_column_name']} not found in dataset"
f" '{dataset_dict['name']}'. Make sure to set `--text_column_name` to the"
f" correct text column - one of {', '.join(dataset_features)}."
)
# blanket renaming of all transcription columns to text
if dataset_dict["text_column_name"] != "text":
dataset = dataset.rename_column(dataset_dict["text_column_name"], "text")
if use_pseudo_labels:
if "whisper_transcript" not in dataset_features:
raise ValueError(
f"Pseudo-label column `whisper_transcript` not found in dataset {dataset_dict['name']}. Ensure"
"pseudo-labels are present in the dataset under this column name, or train directly on the text "
"labels by setting `--use_pseudo_labels=False` and defining the appropriate `--text_column_name`."
)
columns_to_keep.add("whisper_transcript")
if "condition_on_prev" in dataset_features:
columns_to_keep.add("condition_on_prev")
dataset_features = dataset.features.keys()
dataset = dataset.remove_columns(set(dataset_features - columns_to_keep))
all_datasets.append(dataset)
if len(all_datasets) == 1:
# we have a single dataset so just return it as is
return all_datasets[0]
if streaming:
interleaved_dataset = interleave_datasets(
all_datasets,
stopping_strategy=stopping_strategy,
probabilities=probabilities,
seed=seed,
)
else:
interleaved_dataset = concatenate_datasets(all_datasets)
return interleaved_dataset
def sorted_checkpoints(output_dir=None, checkpoint_prefix="checkpoint") -> List[str]:
"""Helper function to sort saved checkpoints from oldest to newest."""
ordering_and_checkpoint_path = []
glob_checkpoints = [str(x) for x in Path(output_dir).glob(f"{checkpoint_prefix}-*") if os.path.isdir(x)]
glob_checkpoints = [path for path in glob_checkpoints if "val-wer" not in path] # filter out best model checkpoints
for path in glob_checkpoints:
regex_match = re.match(f".*{checkpoint_prefix}-([0-9]+)", path)
if regex_match is not None and regex_match.groups() is not None:
ordering_and_checkpoint_path.append((int(regex_match.groups()[0]), path))
checkpoints_sorted = sorted(ordering_and_checkpoint_path)
checkpoints_sorted = [checkpoint[1] for checkpoint in checkpoints_sorted]
return checkpoints_sorted
def sorted_best_checkpoints(output_dir=None, checkpoint_prefix="checkpoint"):
"""Helper function to sort saved best checkpoints."""
ordering_and_checkpoint_path = []
glob_checkpoints = [str(x) for x in Path(output_dir).glob(f"{checkpoint_prefix}-*") if os.path.isdir(x)]
for path in glob_checkpoints:
regex_match = re.search(r"val-wer-([0-9]+\.[0-9]+)", path)
if regex_match is not None and regex_match.groups() is not None:
ordering_and_checkpoint_path.append((float(regex_match.groups(1)[0]), path))
checkpoints_sorted = sorted(ordering_and_checkpoint_path, reverse=True)
checkpoints_sorted = [checkpoint[1] for checkpoint in checkpoints_sorted]
return checkpoints_sorted
def rotate_checkpoints(save_total_limit=None, output_dir=None, checkpoint_prefix="checkpoint", sorting_fn=sorted_checkpoints) -> None:
"""Helper function to delete old checkpoints."""
if save_total_limit is None or save_total_limit <= 0:
return
# Check if we should delete older checkpoint(s)
checkpoints_sorted = sorting_fn(output_dir=output_dir, checkpoint_prefix=checkpoint_prefix)
if len(checkpoints_sorted) <= save_total_limit:
return
number_of_checkpoints_to_delete = max(0, len(checkpoints_sorted) - save_total_limit)
checkpoints_to_be_deleted = checkpoints_sorted[:number_of_checkpoints_to_delete]
for checkpoint in checkpoints_to_be_deleted:
logger.info(f"Deleting older checkpoint [{checkpoint}].")
shutil.rmtree(checkpoint, ignore_errors=True)
_RE_CHECKPOINT = re.compile(r"^checkpoint-(\d+)-epoch-(\d+)$")
def get_last_checkpoint(folder):
content = os.listdir(folder)
checkpoints = [
path
for path in content
if _RE_CHECKPOINT.search(path) is not None and os.path.isdir(os.path.join(folder, path))
]
if len(checkpoints) == 0:
return
return os.path.join(folder, max(checkpoints, key=lambda x: int(_RE_CHECKPOINT.search(x).groups()[0])))
def get_parameter_names(model, forbidden_layer_types, forbidden_module=None):
"""
Returns the names of the model parameters that are not inside a forbidden layer or forbidden module.
Can be used to get a subset of parameter names for decay masks, or to exclude parameters from an optimiser
(e.g. if the module is frozen).
"""
result = []
for name, child in model.named_children():
result += [
f"{name}.{n}"
for n in get_parameter_names(child, forbidden_layer_types, forbidden_module)
if not (
isinstance(child, tuple(forbidden_layer_types))
or (child in tuple(forbidden_module) if forbidden_module is not None else False)
)
]
# Add model specific parameters (defined with nn.Parameter) since they are not in any child.
result += list(model._parameters.keys())
return result
def main():
# 1. Parse input arguments
# We keep distinct sets of args, for cleaner separation of model/data/training related args
parser = HfArgumentParser((ModelArguments, DataTrainingArguments, DistillationTrainingArguments))
if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
# If we pass only one argument to the script and it's the path to a json file,
# let's parse it to get our arguments.
model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
else:
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
# 2. Initialize the accelerator
# We will let the accelerator handle device placement for us in this example
# We simply have to specify the training precision and any trackers being used
# We'll use the same dtype arguments as our JAX/Flax training script and convert
# it to accelerate format
if training_args.dtype == "float16":
mixed_precision = "fp16"
teacher_dtype = torch.float16
elif training_args.dtype == "bfloat16":
mixed_precision = "bf16"
teacher_dtype = torch.bfloat16
else:
mixed_precision = "no"
teacher_dtype = torch.float32
accelerator = Accelerator(
gradient_accumulation_steps=training_args.gradient_accumulation_steps,
mixed_precision=mixed_precision,
log_with=training_args.report_to,
project_dir=training_args.output_dir,
)
accelerator.init_trackers(
project_name=data_args.wandb_project,
init_kwargs={
"wandb": {"name": data_args.wandb_name,
"dir": data_args.wandb_dir}
}
)
# 3. Set-up basic logging
# Create one log on every process with the configuration for debugging
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
# Log a small summary on each proces
logger.warning(
f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}, "
f"distributed training: {training_args.parallel_mode.value == 'distributed'}, 16-bits training: {training_args.fp16}"
)
# Set the verbosity to info of the Transformers logger (on main process only)
if accelerator.is_local_main_process:
datasets.utils.logging.set_verbosity_warning()
transformers.utils.logging.set_verbosity_info()
else:
datasets.utils.logging.set_verbosity_error()
transformers.utils.logging.set_verbosity_error()
logger.info("Training/evaluation parameters %s", training_args)
# 4. Detecting last checkpoint and eventually continue from last checkpoint
last_checkpoint = None
if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
last_checkpoint = get_last_checkpoint(training_args.output_dir)
if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
raise ValueError(
f"Output directory ({training_args.output_dir}) already exists and is not empty. "
"Use --overwrite_output_dir to overcome."
)
elif last_checkpoint is not None and training_args.resume_from_checkpoint is None:
logger.info(
f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
"the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
)
# 5. Handle the repository creation
if accelerator.is_main_process:
if training_args.push_to_hub:
if training_args.hub_model_id is None:
repo_name = get_full_repo_name(
Path(training_args.output_dir).absolute().name,
token=training_args.hub_token,
)
else:
repo_name = training_args.hub_model_id
create_repo(repo_name, exist_ok=True, token=training_args.hub_token)
with open(os.path.join(training_args.output_dir, ".gitignore"), "w+") as gitignore:
if "wandb" not in gitignore:
gitignore.write("wandb\n")
elif training_args.output_dir is not None:
os.makedirs(training_args.output_dir, exist_ok=True)
accelerator.wait_for_everyone()
# 6. Load dataset - either streaming or non-streaming (offline)
raw_datasets = IterableDatasetDict() if data_args.streaming else DatasetDict()
# set seed for determinism
set_seed(training_args.seed)
if training_args.do_train:
raw_datasets["train"] = load_multiple_datasets(
data_args.train_dataset_name,
data_args.train_dataset_config_name,
splits=data_args.train_split_name,
text_column_names=data_args.text_column_name,
use_pseudo_labels=data_args.use_pseudo_labels,
streaming=data_args.streaming,
dataset_samples=data_args.train_dataset_samples,
seed=training_args.seed,
accelerator=accelerator,
cache_dir=data_args.dataset_cache_dir,
token=model_args.token,
)
raw_datasets_train_features = list(raw_datasets["train"].features.keys())
if training_args.do_eval:
dataset_names_dict = convert_dataset_str_to_list(
data_args.eval_dataset_name if data_args.eval_dataset_name else data_args.train_dataset_name,
(
data_args.eval_dataset_config_name
if data_args.eval_dataset_config_name
else data_args.train_dataset_config_name
),
splits=data_args.eval_split_name,
text_column_names=data_args.eval_text_column_name,
)
all_eval_splits = []
if len(dataset_names_dict) == 1:
# load a single eval set
dataset_dict = dataset_names_dict[0]
all_eval_splits.append("eval")
raw_datasets["eval"] = load_dataset(
dataset_dict["name"],
dataset_dict["config"],
split=dataset_dict["split"],
cache_dir=data_args.dataset_cache_dir,
token=model_args.token,
streaming=data_args.streaming,
)
if data_args.eval_text_column_name != "text":
raw_datasets["eval"] = raw_datasets["eval"].rename_column(data_args.eval_text_column_name, "text")
else:
# load multiple eval sets
for dataset_dict in dataset_names_dict:
if dataset_dict["name"] == "esb/diagnostic-dataset":
# for the ESB diagnostic dataset, the dataset name is effectively the config
pretty_name = f"{dataset_dict['config']}-diagnostic/{dataset_dict['split']}"
else:
pretty_name = f"{dataset_dict['name'].split('/')[-1]}/{dataset_dict['split'].replace('.', '-')}"
all_eval_splits.append(pretty_name)
raw_datasets[pretty_name] = load_dataset(
dataset_dict["name"],
dataset_dict["config"],
split=dataset_dict["split"],
cache_dir=data_args.dataset_cache_dir,
token=model_args.token,
streaming=data_args.streaming,
)
# make column names consistent (text, audio)
if dataset_dict["text_column_name"] != "text":
raw_datasets[pretty_name] = raw_datasets[pretty_name].rename_column(
dataset_dict["text_column_name"], "text"
)
raw_datasets[pretty_name] = raw_datasets[pretty_name].remove_columns(
set(raw_datasets[pretty_name].features.keys()) - {"audio", "text"}
)
if not training_args.do_train and not training_args.do_eval:
raise ValueError(
"Cannot not train and not do evaluation. At least one of training or evaluation has to be performed."
)
# 7. Load pretrained model, tokenizer, and feature extractor
config = WhisperConfig.from_pretrained(
(model_args.config_name if model_args.config_name else model_args.model_name_or_path),
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
token=model_args.token,
)
feature_extractor = WhisperFeatureExtractor.from_pretrained(
(model_args.feature_extractor_name if model_args.feature_extractor_name else model_args.model_name_or_path),
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
token=model_args.token,
)
tokenizer = WhisperTokenizerFast.from_pretrained(
(model_args.tokenizer_name if model_args.tokenizer_name else model_args.model_name_or_path),
cache_dir=model_args.cache_dir,
use_fast=model_args.use_fast_tokenizer,
revision=model_args.model_revision,
token=model_args.token,
)
# override timestamp tokens until tokenizer issues are fixed in transformers
timestamps = [AddedToken("<|%.2f|>" % (i * 0.02), lstrip=False, rstrip=False) for i in range(1500 + 1)]
tokenizer.add_tokens(timestamps)
# The teacher model can safely be cast to the dtype of training since we don't
# update the params
teacher_model = WhisperForConditionalGeneration.from_pretrained(
model_args.teacher_model_name_or_path,
cache_dir=model_args.cache_dir,
token=model_args.token,
low_cpu_mem_usage=True,
torch_dtype=teacher_dtype,
attn_implementation=model_args.attn_implementation,
)
student_model = WhisperForConditionalGeneration.from_pretrained(
model_args.model_name_or_path,
config=config,
cache_dir=model_args.cache_dir,
revision=model_args.model_revision,
subfolder=model_args.subfolder,