anuragshas commited on
Commit
3f8af43
1 Parent(s): 3ad4a92

Training in progress, step 400

Browse files
.gitignore ADDED
@@ -0,0 +1 @@
 
 
1
+ checkpoint-*/
.ipynb_checkpoints/run-checkpoint.sh ADDED
@@ -0,0 +1,36 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ python run_speech_recognition_ctc.py \
2
+ --dataset_name="mozilla-foundation/common_voice_9_0" \
3
+ --model_name_or_path="facebook/wav2vec2-xls-r-300m" \
4
+ --dataset_config_name="mr" \
5
+ --output_dir="./" \
6
+ --overwrite_output_dir \
7
+ --max_steps 6124 \
8
+ --per_device_train_batch_size="64" \
9
+ --per_device_eval_batch_size="64" \
10
+ --gradient_accumulation_steps="2" \
11
+ --learning_rate="7.5e-5" \
12
+ --warmup_ratio="0.1" \
13
+ --length_column_name="input_length" \
14
+ --evaluation_strategy="steps" \
15
+ --text_column_name="sentence" \
16
+ --chars_to_ignore , ? . ! \- \; \: \" “ % ‘ ” � — ’ … – \' । ॥ \॔ \॓ \
17
+ --save_steps="400" \
18
+ --eval_steps="400" \
19
+ --logging_steps="100" \
20
+ --layerdrop="0.0" \
21
+ --activation_dropout="0.1" \
22
+ --save_total_limit="1" \
23
+ --freeze_feature_encoder \
24
+ --feat_proj_dropout="0.0" \
25
+ --mask_time_prob="0.75" \
26
+ --mask_time_length="10" \
27
+ --mask_feature_prob="0.25" \
28
+ --mask_feature_length="64" \
29
+ --seed="42" \
30
+ --gradient_checkpointing \
31
+ --use_auth_token \
32
+ --fp16 \
33
+ --group_by_length \
34
+ --do_train --do_eval \
35
+ --bnb --tristage_sched \
36
+ --push_to_hub
.ipynb_checkpoints/run_speech_recognition_ctc-checkpoint.py ADDED
@@ -0,0 +1,932 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python
2
+ # coding=utf-8
3
+ # Copyright 2021 The HuggingFace Inc. team. All rights reserved.
4
+ #
5
+ # Licensed under the Apache License, Version 2.0 (the "License");
6
+ # you may not use this file except in compliance with the License.
7
+ # You may obtain a copy of the License at
8
+ #
9
+ # http://www.apache.org/licenses/LICENSE-2.0
10
+ #
11
+ # Unless required by applicable law or agreed to in writing, software
12
+ # distributed under the License is distributed on an "AS IS" BASIS,
13
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14
+ # See the License for the specific language governing permissions and
15
+
16
+ """ Fine-tuning a 🤗 Transformers CTC model for automatic speech recognition"""
17
+
18
+ import functools
19
+ import json
20
+ import logging
21
+ import os
22
+ import re
23
+ import sys
24
+ import unicodedata
25
+ import warnings
26
+ from dataclasses import dataclass, field
27
+ from typing import Dict, List, Optional, Union
28
+
29
+ import datasets
30
+ import numpy as np
31
+ import torch
32
+ from torch.optim.lr_scheduler import LambdaLR
33
+ from datasets import DatasetDict, load_dataset, load_metric, load_from_disk
34
+
35
+ import bitsandbytes as bnb
36
+ import transformers
37
+ from transformers import (
38
+ AutoConfig,
39
+ AutoFeatureExtractor,
40
+ AutoModelForCTC,
41
+ AutoProcessor,
42
+ AutoTokenizer,
43
+ HfArgumentParser,
44
+ Trainer,
45
+ TrainingArguments,
46
+ Wav2Vec2Processor,
47
+ set_seed,
48
+ )
49
+ from transformers.trainer_pt_utils import get_parameter_names
50
+ from transformers.trainer_utils import get_last_checkpoint, is_main_process
51
+ from transformers.utils import check_min_version
52
+ from transformers.utils.versions import require_version
53
+
54
+
55
+ # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
56
+ check_min_version("4.16.0.dev0")
57
+
58
+ require_version(
59
+ "datasets>=1.13.3",
60
+ "To fix: pip install -r examples/pytorch/text-classification/requirements.txt",
61
+ )
62
+
63
+
64
+ logger = logging.getLogger(__name__)
65
+
66
+
67
+ def list_field(default=None, metadata=None):
68
+ return field(default_factory=lambda: default, metadata=metadata)
69
+
70
+
71
+ @dataclass
72
+ class ModelArguments:
73
+ """
74
+ Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
75
+ """
76
+
77
+ model_name_or_path: str = field(
78
+ metadata={
79
+ "help": "Path to pretrained model or model identifier from huggingface.co/models"
80
+ }
81
+ )
82
+ tokenizer_name_or_path: Optional[str] = field(
83
+ default=None,
84
+ metadata={
85
+ "help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"
86
+ },
87
+ )
88
+ cache_dir: Optional[str] = field(
89
+ default=None,
90
+ metadata={
91
+ "help": "Where do you want to store the pretrained models downloaded from huggingface.co"
92
+ },
93
+ )
94
+ freeze_feature_encoder: bool = field(
95
+ default=True,
96
+ metadata={"help": "Whether to freeze the feature encoder layers of the model."},
97
+ )
98
+ attention_dropout: float = field(
99
+ default=0.0,
100
+ metadata={"help": "The dropout ratio for the attention probabilities."},
101
+ )
102
+ activation_dropout: float = field(
103
+ default=0.0,
104
+ metadata={
105
+ "help": "The dropout ratio for activations inside the fully connected layer."
106
+ },
107
+ )
108
+ feat_proj_dropout: float = field(
109
+ default=0.0, metadata={"help": "The dropout ratio for the projected features."}
110
+ )
111
+ hidden_dropout: float = field(
112
+ default=0.0,
113
+ metadata={
114
+ "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
115
+ },
116
+ )
117
+ final_dropout: float = field(
118
+ default=0.0,
119
+ metadata={"help": "The dropout probability for the final projection layer."},
120
+ )
121
+ mask_time_prob: float = field(
122
+ default=0.05,
123
+ metadata={
124
+ "help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
125
+ "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
126
+ "vectors will be masked along the time axis."
127
+ },
128
+ )
129
+ mask_time_length: int = field(
130
+ default=10,
131
+ metadata={"help": "Length of vector span to mask along the time axis."},
132
+ )
133
+ mask_feature_prob: float = field(
134
+ default=0.0,
135
+ metadata={
136
+ "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
137
+ "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
138
+ },
139
+ )
140
+ mask_feature_length: int = field(
141
+ default=10,
142
+ metadata={"help": "Length of vector span to mask along the feature axis."},
143
+ )
144
+ layerdrop: float = field(
145
+ default=0.0, metadata={"help": "The LayerDrop probability."}
146
+ )
147
+ ctc_loss_reduction: Optional[str] = field(
148
+ default="mean",
149
+ metadata={
150
+ "help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."
151
+ },
152
+ )
153
+
154
+
155
+ @dataclass
156
+ class DataTrainingArguments:
157
+ """
158
+ Arguments pertaining to what data we are going to input our model for training and eval.
159
+
160
+ Using `HfArgumentParser` we can turn this class
161
+ into argparse arguments to be able to specify them on
162
+ the command line.
163
+ """
164
+
165
+ dataset_name: str = field(
166
+ metadata={
167
+ "help": "The configuration name of the dataset to use (via the datasets library)."
168
+ }
169
+ )
170
+ dataset_config_name: str = field(
171
+ default=None,
172
+ metadata={
173
+ "help": "The configuration name of the dataset to use (via the datasets library)."
174
+ },
175
+ )
176
+ train_split_name: str = field(
177
+ default="train+validation",
178
+ metadata={
179
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
180
+ },
181
+ )
182
+ eval_split_name: str = field(
183
+ default="test",
184
+ metadata={
185
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
186
+ },
187
+ )
188
+ audio_column_name: str = field(
189
+ default="audio",
190
+ metadata={
191
+ "help": "The name of the dataset column containing the audio data. Defaults to 'audio'"
192
+ },
193
+ )
194
+ text_column_name: str = field(
195
+ default="text",
196
+ metadata={
197
+ "help": "The name of the dataset column containing the text data. Defaults to 'text'"
198
+ },
199
+ )
200
+ overwrite_cache: bool = field(
201
+ default=False,
202
+ metadata={"help": "Overwrite the cached preprocessed datasets or not."},
203
+ )
204
+ preprocessing_num_workers: Optional[int] = field(
205
+ default=None,
206
+ metadata={"help": "The number of processes to use for the preprocessing."},
207
+ )
208
+ max_train_samples: Optional[int] = field(
209
+ default=None,
210
+ metadata={
211
+ "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
212
+ "value if set."
213
+ },
214
+ )
215
+ max_eval_samples: Optional[int] = field(
216
+ default=None,
217
+ metadata={
218
+ "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
219
+ "value if set."
220
+ },
221
+ )
222
+ chars_to_ignore: Optional[List[str]] = list_field(
223
+ default=None,
224
+ metadata={"help": "A list of characters to remove from the transcripts."},
225
+ )
226
+ eval_metrics: List[str] = list_field(
227
+ default=["wer", "cer"],
228
+ metadata={
229
+ "help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"
230
+ },
231
+ )
232
+ max_duration_in_seconds: float = field(
233
+ default=20.0,
234
+ metadata={
235
+ "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
236
+ },
237
+ )
238
+ min_duration_in_seconds: float = field(
239
+ default=0.0,
240
+ metadata={
241
+ "help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"
242
+ },
243
+ )
244
+ preprocessing_only: bool = field(
245
+ default=False,
246
+ metadata={
247
+ "help": "Whether to only do data preprocessing and skip training. "
248
+ "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
249
+ "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
250
+ "so that the cached datasets can consequently be loaded in distributed training"
251
+ },
252
+ )
253
+ use_auth_token: bool = field(
254
+ default=False,
255
+ metadata={
256
+ "help": "If :obj:`True`, will use the token generated when running"
257
+ ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
258
+ },
259
+ )
260
+ unk_token: str = field(
261
+ default="[UNK]",
262
+ metadata={"help": "The unk token for the tokenizer"},
263
+ )
264
+ pad_token: str = field(
265
+ default="[PAD]",
266
+ metadata={"help": "The padding token for the tokenizer"},
267
+ )
268
+ word_delimiter_token: str = field(
269
+ default="|",
270
+ metadata={"help": "The word delimiter token for the tokenizer"},
271
+ )
272
+ phoneme_language: Optional[str] = field(
273
+ default=None,
274
+ metadata={
275
+ "help": "The target language that should be used be"
276
+ " passed to the tokenizer for tokenization. Note that"
277
+ " this is only relevant if the model classifies the"
278
+ " input audio to a sequence of phoneme sequences."
279
+ },
280
+ )
281
+
282
+
283
+ @dataclass
284
+ class ExtraArguments:
285
+ "Additional training arguments"
286
+ bnb: bool = field(default=False, metadata={"help": "If true uses 8bit Adam"})
287
+ tristage_sched: bool = field(
288
+ default=False,
289
+ metadata={"help": "If true uses tristage LR scheduler (refer to XLS-R paper)"},
290
+ )
291
+ wandb_project: str = field(
292
+ default="", metadata={"help": "Name of wandb project to log into"}
293
+ )
294
+
295
+
296
+ @dataclass
297
+ class DataCollatorCTCWithPadding:
298
+ """
299
+ Data collator that will dynamically pad the inputs received.
300
+ Args:
301
+ processor (:class:`~transformers.AutoProcessor`)
302
+ The processor used for proccessing the data.
303
+ padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
304
+ Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
305
+ among:
306
+ * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
307
+ sequence if provided).
308
+ * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
309
+ maximum acceptable input length for the model if that argument is not provided.
310
+ * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
311
+ different lengths).
312
+ max_length (:obj:`int`, `optional`):
313
+ Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
314
+ max_length_labels (:obj:`int`, `optional`):
315
+ Maximum length of the ``labels`` returned list and optionally padding length (see above).
316
+ pad_to_multiple_of (:obj:`int`, `optional`):
317
+ If set will pad the sequence to a multiple of the provided value.
318
+ This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
319
+ 7.5 (Volta).
320
+ """
321
+
322
+ processor: AutoProcessor
323
+ padding: Union[bool, str] = "longest"
324
+ pad_to_multiple_of: Optional[int] = None
325
+ pad_to_multiple_of_labels: Optional[int] = None
326
+
327
+ def __call__(
328
+ self, features: List[Dict[str, Union[List[int], torch.Tensor]]]
329
+ ) -> Dict[str, torch.Tensor]:
330
+ # split inputs and labels since they have to be of different lenghts and need
331
+ # different padding methods
332
+ input_features = [
333
+ {"input_values": feature["input_values"]} for feature in features
334
+ ]
335
+ label_features = [{"input_ids": feature["labels"]} for feature in features]
336
+
337
+ batch = self.processor.pad(
338
+ input_features,
339
+ padding=self.padding,
340
+ pad_to_multiple_of=self.pad_to_multiple_of,
341
+ return_tensors="pt",
342
+ )
343
+
344
+ with self.processor.as_target_processor():
345
+ labels_batch = self.processor.pad(
346
+ label_features,
347
+ padding=self.padding,
348
+ pad_to_multiple_of=self.pad_to_multiple_of_labels,
349
+ return_tensors="pt",
350
+ )
351
+
352
+ # replace padding with -100 to ignore loss correctly
353
+ labels = labels_batch["input_ids"].masked_fill(
354
+ labels_batch.attention_mask.ne(1), -100
355
+ )
356
+
357
+ batch["labels"] = labels
358
+
359
+ return batch
360
+
361
+
362
+ def get_tri_stage_schedule(
363
+ optimizer,
364
+ num_training_steps,
365
+ ratios=[0.1, 0.4, 0.5],
366
+ num_warmup_steps=None,
367
+ num_hold_steps=None,
368
+ start_ratio=0.01,
369
+ end_ratio=0.05,
370
+ ):
371
+ assert (num_warmup_steps is None) == (num_hold_steps is None)
372
+ if num_warmup_steps is None:
373
+ num_warmup_steps = int(ratios[0] * num_training_steps)
374
+ num_hold_steps = int(ratios[1] * num_training_steps)
375
+ start_decay_step = num_warmup_steps + num_hold_steps
376
+ a_w, b_w = (1 - start_ratio) / num_warmup_steps, start_ratio
377
+ num_decay_steps = num_training_steps - start_decay_step
378
+ a_d, b_d = (end_ratio - 1) / num_decay_steps, 1.0
379
+
380
+ def lr_lambda(current_step):
381
+ if current_step < num_warmup_steps:
382
+ return a_w * float(current_step) + b_w
383
+ if current_step < start_decay_step:
384
+ return 1.0
385
+ return max(end_ratio, a_d * float(current_step - start_decay_step) + b_d)
386
+
387
+ return LambdaLR(optimizer, lr_lambda)
388
+
389
+
390
+ def create_vocabulary_from_data(
391
+ datasets: DatasetDict,
392
+ word_delimiter_token: Optional[str] = None,
393
+ unk_token: Optional[str] = None,
394
+ pad_token: Optional[str] = None,
395
+ ):
396
+ # Given training and test labels create vocabulary
397
+ def extract_all_chars(batch):
398
+ all_text = " ".join(batch["target_text"])
399
+ vocab = list(set(all_text))
400
+ return {"vocab": [vocab], "all_text": [all_text]}
401
+
402
+ vocabs = datasets.map(
403
+ extract_all_chars,
404
+ batched=True,
405
+ batch_size=-1,
406
+ keep_in_memory=True,
407
+ remove_columns=datasets["train"].column_names,
408
+ )
409
+
410
+ # take union of all unique characters in each dataset
411
+ vocab_set = functools.reduce(
412
+ lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]),
413
+ vocabs.values(),
414
+ )
415
+
416
+ vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
417
+
418
+ # replace white space with delimiter token
419
+ if word_delimiter_token is not None:
420
+ vocab_dict[word_delimiter_token] = vocab_dict[" "]
421
+ del vocab_dict[" "]
422
+
423
+ # add unk and pad token
424
+ if unk_token is not None:
425
+ vocab_dict[unk_token] = len(vocab_dict)
426
+
427
+ if pad_token is not None:
428
+ vocab_dict[pad_token] = len(vocab_dict)
429
+
430
+ return vocab_dict
431
+
432
+
433
+ def main():
434
+ # See all possible arguments in src/transformers/training_args.py
435
+ # or by passing the --help flag to this script.
436
+ # We now keep distinct sets of args, for a cleaner separation of concerns.
437
+
438
+ parser = HfArgumentParser(
439
+ (ModelArguments, DataTrainingArguments, TrainingArguments, ExtraArguments)
440
+ )
441
+ if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
442
+ # If we pass only one argument to the script and it's the path to a json file,
443
+ # let's parse it to get our arguments.
444
+ model_args, data_args, training_args, extra_args = parser.parse_json_file(
445
+ json_file=os.path.abspath(sys.argv[1])
446
+ )
447
+ else:
448
+ (
449
+ model_args,
450
+ data_args,
451
+ training_args,
452
+ extra_args,
453
+ ) = parser.parse_args_into_dataclasses()
454
+
455
+ # Detecting last checkpoint.
456
+ last_checkpoint = None
457
+ if (
458
+ os.path.isdir(training_args.output_dir)
459
+ and training_args.do_train
460
+ and not training_args.overwrite_output_dir
461
+ ):
462
+ last_checkpoint = get_last_checkpoint(training_args.output_dir)
463
+ if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
464
+ raise ValueError(
465
+ f"Output directory ({training_args.output_dir}) already exists and is not empty. "
466
+ "Use --overwrite_output_dir to overcome."
467
+ )
468
+ elif last_checkpoint is not None:
469
+ logger.info(
470
+ f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
471
+ "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
472
+ )
473
+
474
+ # Setup logging
475
+ logging.basicConfig(
476
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
477
+ datefmt="%m/%d/%Y %H:%M:%S",
478
+ handlers=[logging.StreamHandler(sys.stdout)],
479
+ )
480
+ logger.setLevel(
481
+ logging.INFO if is_main_process(training_args.local_rank) else logging.WARN
482
+ )
483
+
484
+ # Log on each process the small summary:
485
+ logger.warning(
486
+ f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
487
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
488
+ )
489
+ # Set the verbosity to info of the Transformers logger (on main process only):
490
+ if is_main_process(training_args.local_rank):
491
+ transformers.utils.logging.set_verbosity_info()
492
+ logger.info("Training/evaluation parameters %s", training_args)
493
+
494
+ # Set seed before initializing model.
495
+ set_seed(training_args.seed)
496
+
497
+ # configure wandb run
498
+ os.environ["WANDB_PROJECT"] = extra_args.wandb_project
499
+
500
+ # 1. First, let's load the dataset
501
+ raw_datasets = DatasetDict()
502
+
503
+ if training_args.do_train:
504
+ if os.path.isdir(data_args.dataset_name):
505
+ raw_datasets["train"] = load_from_disk(
506
+ f"{data_args.dataset_name}/{data_args.train_split_name}"
507
+ )
508
+ else:
509
+ raw_datasets["train"] = load_dataset(
510
+ data_args.dataset_name,
511
+ data_args.dataset_config_name,
512
+ split=data_args.train_split_name,
513
+ use_auth_token=data_args.use_auth_token,
514
+ )
515
+
516
+ if data_args.audio_column_name not in raw_datasets["train"].column_names:
517
+ raise ValueError(
518
+ f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
519
+ "Make sure to set `--audio_column_name` to the correct audio column - one of "
520
+ f"{', '.join(raw_datasets['train'].column_names)}."
521
+ )
522
+
523
+ if data_args.text_column_name not in raw_datasets["train"].column_names:
524
+ raise ValueError(
525
+ f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
526
+ "Make sure to set `--text_column_name` to the correct text column - one of "
527
+ f"{', '.join(raw_datasets['train'].column_names)}."
528
+ )
529
+
530
+ if data_args.max_train_samples is not None:
531
+ raw_datasets["train"] = raw_datasets["train"].select(
532
+ range(data_args.max_train_samples)
533
+ )
534
+
535
+ if training_args.do_eval:
536
+ if os.path.isdir(data_args.dataset_name):
537
+ raw_datasets["eval"] = load_from_disk(
538
+ f"{data_args.dataset_name}/{data_args.eval_split_name}"
539
+ )
540
+ else:
541
+ raw_datasets["eval"] = load_dataset(
542
+ data_args.dataset_name,
543
+ data_args.dataset_config_name,
544
+ split=data_args.eval_split_name,
545
+ use_auth_token=data_args.use_auth_token,
546
+ )
547
+
548
+ if data_args.max_eval_samples is not None:
549
+ raw_datasets["eval"] = raw_datasets["eval"].select(
550
+ range(data_args.max_eval_samples)
551
+ )
552
+
553
+ # 2. We remove some special characters from the datasets
554
+ # that make training complicated and do not help in transcribing the speech
555
+ # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
556
+ # that could be easily picked up by the model
557
+ chars_to_ignore_regex = (
558
+ f'[{"".join(data_args.chars_to_ignore)}]'
559
+ if data_args.chars_to_ignore is not None
560
+ else None
561
+ )
562
+ text_column_name = data_args.text_column_name
563
+
564
+ def remove_special_characters(batch):
565
+ if chars_to_ignore_regex is not None:
566
+ batch["target_text"] = (
567
+ re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
568
+ )
569
+ else:
570
+ batch["target_text"] = batch[text_column_name].lower() + " "
571
+ # Unicode Normalization
572
+ batch["target_text"] = unicodedata.normalize("NFKC", batch["target_text"])
573
+ return batch
574
+
575
+ with training_args.main_process_first(
576
+ desc="dataset map special characters removal"
577
+ ):
578
+ raw_datasets = raw_datasets.map(
579
+ remove_special_characters,
580
+ remove_columns=[text_column_name],
581
+ desc="remove special characters from datasets",
582
+ )
583
+
584
+ # save special tokens for tokenizer
585
+ word_delimiter_token = data_args.word_delimiter_token
586
+ unk_token = data_args.unk_token
587
+ pad_token = data_args.pad_token
588
+
589
+ # 3. Next, let's load the config as we might need it to create
590
+ # the tokenizer
591
+ # load config
592
+ config = AutoConfig.from_pretrained(
593
+ model_args.model_name_or_path,
594
+ cache_dir=model_args.cache_dir,
595
+ use_auth_token=data_args.use_auth_token,
596
+ )
597
+
598
+ # 4. Next, if no tokenizer file is defined,
599
+ # we create the vocabulary of the model by extracting all unique characters from
600
+ # the training and evaluation datasets
601
+ # We need to make sure that only first rank saves vocabulary
602
+ # make sure all processes wait until vocab is created
603
+ tokenizer_name_or_path = model_args.tokenizer_name_or_path
604
+ tokenizer_kwargs = {}
605
+ if tokenizer_name_or_path is None:
606
+ # save vocab in training output dir
607
+ tokenizer_name_or_path = training_args.output_dir
608
+
609
+ vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
610
+
611
+ with training_args.main_process_first():
612
+ if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
613
+ os.remove(vocab_file)
614
+
615
+ with training_args.main_process_first(desc="dataset map vocabulary creation"):
616
+ if not os.path.isfile(vocab_file):
617
+ os.makedirs(tokenizer_name_or_path, exist_ok=True)
618
+ vocab_dict = create_vocabulary_from_data(
619
+ raw_datasets,
620
+ word_delimiter_token=word_delimiter_token,
621
+ unk_token=unk_token,
622
+ pad_token=pad_token,
623
+ )
624
+
625
+ # save vocab dict to be loaded into tokenizer
626
+ with open(vocab_file, "w") as file:
627
+ json.dump(vocab_dict, file)
628
+
629
+ # if tokenizer has just been created
630
+ # it is defined by `tokenizer_class` if present in config else by `model_type`
631
+ tokenizer_kwargs = {
632
+ "config": config if config.tokenizer_class is not None else None,
633
+ "tokenizer_type": config.model_type
634
+ if config.tokenizer_class is None
635
+ else None,
636
+ "unk_token": unk_token,
637
+ "pad_token": pad_token,
638
+ "word_delimiter_token": word_delimiter_token,
639
+ }
640
+
641
+ # 5. Now we can instantiate the feature extractor, tokenizer and model
642
+ # Note for distributed training, the .from_pretrained methods guarantee that only
643
+ # one local process can concurrently download model & vocab.
644
+
645
+ # load feature_extractor and tokenizer
646
+ tokenizer = AutoTokenizer.from_pretrained(
647
+ tokenizer_name_or_path,
648
+ use_auth_token=data_args.use_auth_token,
649
+ **tokenizer_kwargs,
650
+ )
651
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
652
+ model_args.model_name_or_path,
653
+ cache_dir=model_args.cache_dir,
654
+ use_auth_token=data_args.use_auth_token,
655
+ )
656
+
657
+ # adapt config
658
+ config.update(
659
+ {
660
+ "feat_proj_dropout": model_args.feat_proj_dropout,
661
+ "attention_dropout": model_args.attention_dropout,
662
+ "hidden_dropout": model_args.hidden_dropout,
663
+ "final_dropout": model_args.final_dropout,
664
+ "mask_time_prob": model_args.mask_time_prob,
665
+ "mask_time_length": model_args.mask_time_length,
666
+ "mask_feature_prob": model_args.mask_feature_prob,
667
+ "mask_feature_length": model_args.mask_feature_length,
668
+ "gradient_checkpointing": training_args.gradient_checkpointing,
669
+ "layerdrop": model_args.layerdrop,
670
+ "ctc_loss_reduction": model_args.ctc_loss_reduction,
671
+ "pad_token_id": tokenizer.pad_token_id,
672
+ "vocab_size": len(tokenizer),
673
+ "activation_dropout": model_args.activation_dropout,
674
+ }
675
+ )
676
+
677
+ # create model
678
+ model = AutoModelForCTC.from_pretrained(
679
+ model_args.model_name_or_path,
680
+ cache_dir=model_args.cache_dir,
681
+ config=config,
682
+ use_auth_token=data_args.use_auth_token,
683
+ )
684
+
685
+ # freeze encoder
686
+ if model_args.freeze_feature_encoder:
687
+ model.freeze_feature_encoder()
688
+
689
+ # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
690
+ # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
691
+ # so that we just need to set the correct target sampling rate and normalize the input
692
+ # via the `feature_extractor`
693
+
694
+ # make sure that dataset decodes audio with correct sampling rate
695
+ dataset_sampling_rate = (
696
+ next(iter(raw_datasets.values()))
697
+ .features[data_args.audio_column_name]
698
+ .sampling_rate
699
+ )
700
+ if dataset_sampling_rate != feature_extractor.sampling_rate:
701
+ raw_datasets = raw_datasets.cast_column(
702
+ data_args.audio_column_name,
703
+ datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate),
704
+ )
705
+
706
+ # derive max & min input length for sample rate & max duration
707
+ max_input_length = (
708
+ data_args.max_duration_in_seconds * feature_extractor.sampling_rate
709
+ )
710
+ min_input_length = (
711
+ data_args.min_duration_in_seconds * feature_extractor.sampling_rate
712
+ )
713
+ audio_column_name = data_args.audio_column_name
714
+ num_workers = data_args.preprocessing_num_workers
715
+
716
+ # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
717
+ phoneme_language = data_args.phoneme_language
718
+
719
+ # Preprocessing the datasets.
720
+ # We need to read the audio files as arrays and tokenize the targets.
721
+ def prepare_dataset(batch):
722
+ # load audio
723
+ sample = batch[audio_column_name]
724
+
725
+ inputs = feature_extractor(
726
+ sample["array"], sampling_rate=sample["sampling_rate"]
727
+ )
728
+ batch["input_values"] = inputs.input_values[0]
729
+ batch["length"] = len(batch["input_values"])
730
+
731
+ # encode targets
732
+ additional_kwargs = {}
733
+ if phoneme_language is not None:
734
+ additional_kwargs["phonemizer_lang"] = phoneme_language
735
+
736
+ batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
737
+ return batch
738
+
739
+ with training_args.main_process_first(desc="dataset map preprocessing"):
740
+ vectorized_datasets = raw_datasets.map(
741
+ prepare_dataset,
742
+ remove_columns=next(iter(raw_datasets.values())).column_names,
743
+ num_proc=num_workers,
744
+ desc="preprocess datasets",
745
+ )
746
+
747
+ def is_audio_in_length_range(length):
748
+ return length > min_input_length and length < max_input_length
749
+
750
+ # filter data that is shorter than min_input_length
751
+ vectorized_datasets = vectorized_datasets.filter(
752
+ is_audio_in_length_range,
753
+ num_proc=num_workers,
754
+ input_columns=["length"],
755
+ )
756
+
757
+ # 7. Next, we can prepare the training.
758
+ # Let's use word error rate (WER) as our evaluation metric,
759
+ # instantiate a data collator and the trainer
760
+
761
+ # Define evaluation metrics during training, *i.e.* word error rate, character error rate
762
+ eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
763
+
764
+ # for large datasets it is advised to run the preprocessing on a
765
+ # single machine first with ``args.preprocessing_only`` since there will mostly likely
766
+ # be a timeout when running the script in distributed mode.
767
+ # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
768
+ # cached dataset
769
+ if data_args.preprocessing_only:
770
+ logger.info(
771
+ f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}"
772
+ )
773
+ return
774
+
775
+ def compute_metrics(pred):
776
+ pred_logits = pred.predictions
777
+ pred_ids = np.argmax(pred_logits, axis=-1)
778
+
779
+ pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
780
+
781
+ pred_str = tokenizer.batch_decode(pred_ids)
782
+ # we do not want to group tokens when computing the metrics
783
+ label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
784
+
785
+ metrics = {
786
+ k: v.compute(predictions=pred_str, references=label_str)
787
+ for k, v in eval_metrics.items()
788
+ }
789
+
790
+ return metrics
791
+
792
+ # Now save everything to be able to create a single processor later
793
+ if is_main_process(training_args.local_rank):
794
+ # save feature extractor, tokenizer and config
795
+ feature_extractor.save_pretrained(training_args.output_dir)
796
+ tokenizer.save_pretrained(training_args.output_dir)
797
+ config.save_pretrained(training_args.output_dir)
798
+
799
+ try:
800
+ processor = AutoProcessor.from_pretrained(training_args.output_dir)
801
+ except (OSError, KeyError):
802
+ warnings.warn(
803
+ "Loading a processor from a feature extractor config that does not"
804
+ " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
805
+ " attribute to your `preprocessor_config.json` file to suppress this warning: "
806
+ " `'processor_class': 'Wav2Vec2Processor'`",
807
+ FutureWarning,
808
+ )
809
+ processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
810
+
811
+ # Instantiate custom data collator
812
+ data_collator = DataCollatorCTCWithPadding(processor=processor)
813
+
814
+ decay_parameters = get_parameter_names(model, [torch.nn.LayerNorm])
815
+ decay_parameters = [name for name in decay_parameters if "bias" not in name]
816
+ optimizer_grouped_parameters = [
817
+ {
818
+ "params": [p for n, p in model.named_parameters() if n in decay_parameters],
819
+ "weight_decay": training_args.weight_decay,
820
+ },
821
+ {
822
+ "params": [
823
+ p for n, p in model.named_parameters() if n not in decay_parameters
824
+ ],
825
+ "weight_decay": 0.0,
826
+ },
827
+ ]
828
+ if extra_args.bnb:
829
+ optimizer = bnb.optim.Adam8bit(
830
+ params=optimizer_grouped_parameters,
831
+ lr=training_args.learning_rate,
832
+ betas=(training_args.adam_beta1, training_args.adam_beta2),
833
+ eps=training_args.adam_epsilon,
834
+ )
835
+ else:
836
+ optimizer = torch.optim.AdamW(
837
+ params=optimizer_grouped_parameters,
838
+ lr=training_args.learning_rate,
839
+ betas=(training_args.adam_beta1, training_args.adam_beta2),
840
+ eps=training_args.adam_epsilon,
841
+ )
842
+ if extra_args.tristage_sched:
843
+ scheduler = get_tri_stage_schedule(optimizer, training_args.max_steps)
844
+ else:
845
+ scheduler = None
846
+ optimizers = (optimizer, scheduler)
847
+
848
+ # Initialize Trainer
849
+ trainer = Trainer(
850
+ model=model,
851
+ data_collator=data_collator,
852
+ args=training_args,
853
+ compute_metrics=compute_metrics,
854
+ train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
855
+ eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
856
+ tokenizer=feature_extractor,
857
+ optimizers=optimizers,
858
+ )
859
+
860
+ # 8. Finally, we can start training
861
+
862
+ # Training
863
+ if training_args.do_train:
864
+
865
+ # use last checkpoint if exist
866
+ if last_checkpoint is not None:
867
+ checkpoint = last_checkpoint
868
+ elif os.path.isdir(model_args.model_name_or_path):
869
+ checkpoint = model_args.model_name_or_path
870
+ else:
871
+ checkpoint = None
872
+
873
+ train_result = trainer.train(resume_from_checkpoint=checkpoint)
874
+ trainer.save_model()
875
+
876
+ metrics = train_result.metrics
877
+ max_train_samples = (
878
+ data_args.max_train_samples
879
+ if data_args.max_train_samples is not None
880
+ else len(vectorized_datasets["train"])
881
+ )
882
+ metrics["train_samples"] = min(
883
+ max_train_samples, len(vectorized_datasets["train"])
884
+ )
885
+
886
+ trainer.log_metrics("train", metrics)
887
+ trainer.save_metrics("train", metrics)
888
+ trainer.save_state()
889
+
890
+ # Evaluation
891
+ results = {}
892
+ if training_args.do_eval:
893
+ logger.info("*** Evaluate ***")
894
+ metrics = trainer.evaluate()
895
+ max_eval_samples = (
896
+ data_args.max_eval_samples
897
+ if data_args.max_eval_samples is not None
898
+ else len(vectorized_datasets["eval"])
899
+ )
900
+ metrics["eval_samples"] = min(
901
+ max_eval_samples, len(vectorized_datasets["eval"])
902
+ )
903
+
904
+ trainer.log_metrics("eval", metrics)
905
+ trainer.save_metrics("eval", metrics)
906
+
907
+ # Write model card and (optionally) push to hub
908
+ config_name = (
909
+ data_args.dataset_config_name
910
+ if data_args.dataset_config_name is not None
911
+ else "na"
912
+ )
913
+ kwargs = {
914
+ "finetuned_from": model_args.model_name_or_path,
915
+ "tasks": "speech-recognition",
916
+ "tags": ["automatic-speech-recognition", data_args.dataset_name],
917
+ "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
918
+ "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
919
+ }
920
+ if "common_voice" in data_args.dataset_name:
921
+ kwargs["language"] = config_name
922
+
923
+ if training_args.push_to_hub:
924
+ trainer.push_to_hub(**kwargs)
925
+ else:
926
+ trainer.create_model_card(**kwargs)
927
+
928
+ return results
929
+
930
+
931
+ if __name__ == "__main__":
932
+ main()
.ipynb_checkpoints/vocab-checkpoint.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"ँ": 1, "ं": 2, "ः": 3, "अ": 4, "आ": 5, "इ": 6, "ई": 7, "उ": 8, "ऊ": 9, "ऋ": 10, "ऌ": 11, "ऍ": 12, "ऎ": 13, "ए": 14, "ऐ": 15, "ऑ": 16, "ओ": 17, "औ": 18, "क": 19, "ख": 20, "ग": 21, "घ": 22, "ङ": 23, "च": 24, "छ": 25, "ज": 26, "झ": 27, "ञ": 28, "ट": 29, "ठ": 30, "ड": 31, "ढ": 32, "ण": 33, "त": 34, "थ": 35, "द": 36, "ध": 37, "न": 38, "प": 39, "फ": 40, "ब": 41, "भ": 42, "म": 43, "य": 44, "र": 45, "ऱ": 46, "ल": 47, "ळ": 48, "व": 49, "श": 50, "ष": 51, "स": 52, "ह": 53, "़": 54, "ऽ": 55, "ा": 56, "ि": 57, "ी": 58, "ु": 59, "ू": 60, "ृ": 61, "ॄ": 62, "ॅ": 63, "ॆ": 64, "े": 65, "ै": 66, "ॉ": 67, "ॊ": 68, "ो": 69, "ौ": 70, "्": 71, "ॢ": 72, "ॲ": 73, "|": 0, "[UNK]": 74, "[PAD]": 75}
added_tokens.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"<s>": 76, "</s>": 77}
config.json ADDED
@@ -0,0 +1,107 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "facebook/wav2vec2-xls-r-300m",
3
+ "activation_dropout": 0.1,
4
+ "adapter_kernel_size": 3,
5
+ "adapter_stride": 2,
6
+ "add_adapter": false,
7
+ "apply_spec_augment": true,
8
+ "architectures": [
9
+ "Wav2Vec2ForCTC"
10
+ ],
11
+ "attention_dropout": 0.0,
12
+ "bos_token_id": 1,
13
+ "classifier_proj_size": 256,
14
+ "codevector_dim": 768,
15
+ "contrastive_logits_temperature": 0.1,
16
+ "conv_bias": true,
17
+ "conv_dim": [
18
+ 512,
19
+ 512,
20
+ 512,
21
+ 512,
22
+ 512,
23
+ 512,
24
+ 512
25
+ ],
26
+ "conv_kernel": [
27
+ 10,
28
+ 3,
29
+ 3,
30
+ 3,
31
+ 3,
32
+ 2,
33
+ 2
34
+ ],
35
+ "conv_stride": [
36
+ 5,
37
+ 2,
38
+ 2,
39
+ 2,
40
+ 2,
41
+ 2,
42
+ 2
43
+ ],
44
+ "ctc_loss_reduction": "mean",
45
+ "ctc_zero_infinity": false,
46
+ "diversity_loss_weight": 0.1,
47
+ "do_stable_layer_norm": true,
48
+ "eos_token_id": 2,
49
+ "feat_extract_activation": "gelu",
50
+ "feat_extract_dropout": 0.0,
51
+ "feat_extract_norm": "layer",
52
+ "feat_proj_dropout": 0.0,
53
+ "feat_quantizer_dropout": 0.0,
54
+ "final_dropout": 0.0,
55
+ "hidden_act": "gelu",
56
+ "hidden_dropout": 0.0,
57
+ "hidden_size": 1024,
58
+ "initializer_range": 0.02,
59
+ "intermediate_size": 4096,
60
+ "layer_norm_eps": 1e-05,
61
+ "layerdrop": 0.0,
62
+ "mask_feature_length": 64,
63
+ "mask_feature_min_masks": 0,
64
+ "mask_feature_prob": 0.25,
65
+ "mask_time_length": 10,
66
+ "mask_time_min_masks": 2,
67
+ "mask_time_prob": 0.75,
68
+ "model_type": "wav2vec2",
69
+ "num_adapter_layers": 3,
70
+ "num_attention_heads": 16,
71
+ "num_codevector_groups": 2,
72
+ "num_codevectors_per_group": 320,
73
+ "num_conv_pos_embedding_groups": 16,
74
+ "num_conv_pos_embeddings": 128,
75
+ "num_feat_extract_layers": 7,
76
+ "num_hidden_layers": 24,
77
+ "num_negatives": 100,
78
+ "output_hidden_size": 1024,
79
+ "pad_token_id": 75,
80
+ "proj_codevector_dim": 768,
81
+ "tdnn_dilation": [
82
+ 1,
83
+ 2,
84
+ 3,
85
+ 1,
86
+ 1
87
+ ],
88
+ "tdnn_dim": [
89
+ 512,
90
+ 512,
91
+ 512,
92
+ 512,
93
+ 1500
94
+ ],
95
+ "tdnn_kernel": [
96
+ 5,
97
+ 3,
98
+ 3,
99
+ 1,
100
+ 1
101
+ ],
102
+ "torch_dtype": "float32",
103
+ "transformers_version": "4.19.0.dev0",
104
+ "use_weighted_layer_sum": false,
105
+ "vocab_size": 78,
106
+ "xvector_output_dim": 512
107
+ }
preprocessor_config.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "do_normalize": true,
3
+ "feature_extractor_type": "Wav2Vec2FeatureExtractor",
4
+ "feature_size": 1,
5
+ "padding_side": "right",
6
+ "padding_value": 0,
7
+ "return_attention_mask": true,
8
+ "sampling_rate": 16000
9
+ }
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:132a9fbe20bae448fa993466829bb2ba06f190a79c76436d5975df0ed6e5edbd
3
+ size 1262218481
run.sh ADDED
@@ -0,0 +1,36 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ python run_speech_recognition_ctc.py \
2
+ --dataset_name="mozilla-foundation/common_voice_9_0" \
3
+ --model_name_or_path="facebook/wav2vec2-xls-r-300m" \
4
+ --dataset_config_name="mr" \
5
+ --output_dir="./" \
6
+ --overwrite_output_dir \
7
+ --max_steps 6124 \
8
+ --per_device_train_batch_size="64" \
9
+ --per_device_eval_batch_size="64" \
10
+ --gradient_accumulation_steps="2" \
11
+ --learning_rate="7.5e-5" \
12
+ --warmup_ratio="0.1" \
13
+ --length_column_name="input_length" \
14
+ --evaluation_strategy="steps" \
15
+ --text_column_name="sentence" \
16
+ --chars_to_ignore , ? . ! \- \; \: \" “ % ‘ ” � — ’ … – \' । ॥ \॔ \॓ \
17
+ --save_steps="400" \
18
+ --eval_steps="400" \
19
+ --logging_steps="100" \
20
+ --layerdrop="0.0" \
21
+ --activation_dropout="0.1" \
22
+ --save_total_limit="1" \
23
+ --freeze_feature_encoder \
24
+ --feat_proj_dropout="0.0" \
25
+ --mask_time_prob="0.75" \
26
+ --mask_time_length="10" \
27
+ --mask_feature_prob="0.25" \
28
+ --mask_feature_length="64" \
29
+ --seed="42" \
30
+ --gradient_checkpointing \
31
+ --use_auth_token \
32
+ --fp16 \
33
+ --group_by_length \
34
+ --do_train --do_eval \
35
+ --bnb --tristage_sched \
36
+ --push_to_hub
run_speech_recognition_ctc.py ADDED
@@ -0,0 +1,932 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python
2
+ # coding=utf-8
3
+ # Copyright 2021 The HuggingFace Inc. team. All rights reserved.
4
+ #
5
+ # Licensed under the Apache License, Version 2.0 (the "License");
6
+ # you may not use this file except in compliance with the License.
7
+ # You may obtain a copy of the License at
8
+ #
9
+ # http://www.apache.org/licenses/LICENSE-2.0
10
+ #
11
+ # Unless required by applicable law or agreed to in writing, software
12
+ # distributed under the License is distributed on an "AS IS" BASIS,
13
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14
+ # See the License for the specific language governing permissions and
15
+
16
+ """ Fine-tuning a 🤗 Transformers CTC model for automatic speech recognition"""
17
+
18
+ import functools
19
+ import json
20
+ import logging
21
+ import os
22
+ import re
23
+ import sys
24
+ import unicodedata
25
+ import warnings
26
+ from dataclasses import dataclass, field
27
+ from typing import Dict, List, Optional, Union
28
+
29
+ import datasets
30
+ import numpy as np
31
+ import torch
32
+ from torch.optim.lr_scheduler import LambdaLR
33
+ from datasets import DatasetDict, load_dataset, load_metric, load_from_disk
34
+
35
+ import bitsandbytes as bnb
36
+ import transformers
37
+ from transformers import (
38
+ AutoConfig,
39
+ AutoFeatureExtractor,
40
+ AutoModelForCTC,
41
+ AutoProcessor,
42
+ AutoTokenizer,
43
+ HfArgumentParser,
44
+ Trainer,
45
+ TrainingArguments,
46
+ Wav2Vec2Processor,
47
+ set_seed,
48
+ )
49
+ from transformers.trainer_pt_utils import get_parameter_names
50
+ from transformers.trainer_utils import get_last_checkpoint, is_main_process
51
+ from transformers.utils import check_min_version
52
+ from transformers.utils.versions import require_version
53
+
54
+
55
+ # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
56
+ check_min_version("4.16.0.dev0")
57
+
58
+ require_version(
59
+ "datasets>=1.13.3",
60
+ "To fix: pip install -r examples/pytorch/text-classification/requirements.txt",
61
+ )
62
+
63
+
64
+ logger = logging.getLogger(__name__)
65
+
66
+
67
+ def list_field(default=None, metadata=None):
68
+ return field(default_factory=lambda: default, metadata=metadata)
69
+
70
+
71
+ @dataclass
72
+ class ModelArguments:
73
+ """
74
+ Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
75
+ """
76
+
77
+ model_name_or_path: str = field(
78
+ metadata={
79
+ "help": "Path to pretrained model or model identifier from huggingface.co/models"
80
+ }
81
+ )
82
+ tokenizer_name_or_path: Optional[str] = field(
83
+ default=None,
84
+ metadata={
85
+ "help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"
86
+ },
87
+ )
88
+ cache_dir: Optional[str] = field(
89
+ default=None,
90
+ metadata={
91
+ "help": "Where do you want to store the pretrained models downloaded from huggingface.co"
92
+ },
93
+ )
94
+ freeze_feature_encoder: bool = field(
95
+ default=True,
96
+ metadata={"help": "Whether to freeze the feature encoder layers of the model."},
97
+ )
98
+ attention_dropout: float = field(
99
+ default=0.0,
100
+ metadata={"help": "The dropout ratio for the attention probabilities."},
101
+ )
102
+ activation_dropout: float = field(
103
+ default=0.0,
104
+ metadata={
105
+ "help": "The dropout ratio for activations inside the fully connected layer."
106
+ },
107
+ )
108
+ feat_proj_dropout: float = field(
109
+ default=0.0, metadata={"help": "The dropout ratio for the projected features."}
110
+ )
111
+ hidden_dropout: float = field(
112
+ default=0.0,
113
+ metadata={
114
+ "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
115
+ },
116
+ )
117
+ final_dropout: float = field(
118
+ default=0.0,
119
+ metadata={"help": "The dropout probability for the final projection layer."},
120
+ )
121
+ mask_time_prob: float = field(
122
+ default=0.05,
123
+ metadata={
124
+ "help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
125
+ "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
126
+ "vectors will be masked along the time axis."
127
+ },
128
+ )
129
+ mask_time_length: int = field(
130
+ default=10,
131
+ metadata={"help": "Length of vector span to mask along the time axis."},
132
+ )
133
+ mask_feature_prob: float = field(
134
+ default=0.0,
135
+ metadata={
136
+ "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
137
+ "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
138
+ },
139
+ )
140
+ mask_feature_length: int = field(
141
+ default=10,
142
+ metadata={"help": "Length of vector span to mask along the feature axis."},
143
+ )
144
+ layerdrop: float = field(
145
+ default=0.0, metadata={"help": "The LayerDrop probability."}
146
+ )
147
+ ctc_loss_reduction: Optional[str] = field(
148
+ default="mean",
149
+ metadata={
150
+ "help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."
151
+ },
152
+ )
153
+
154
+
155
+ @dataclass
156
+ class DataTrainingArguments:
157
+ """
158
+ Arguments pertaining to what data we are going to input our model for training and eval.
159
+
160
+ Using `HfArgumentParser` we can turn this class
161
+ into argparse arguments to be able to specify them on
162
+ the command line.
163
+ """
164
+
165
+ dataset_name: str = field(
166
+ metadata={
167
+ "help": "The configuration name of the dataset to use (via the datasets library)."
168
+ }
169
+ )
170
+ dataset_config_name: str = field(
171
+ default=None,
172
+ metadata={
173
+ "help": "The configuration name of the dataset to use (via the datasets library)."
174
+ },
175
+ )
176
+ train_split_name: str = field(
177
+ default="train+validation",
178
+ metadata={
179
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
180
+ },
181
+ )
182
+ eval_split_name: str = field(
183
+ default="test",
184
+ metadata={
185
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
186
+ },
187
+ )
188
+ audio_column_name: str = field(
189
+ default="audio",
190
+ metadata={
191
+ "help": "The name of the dataset column containing the audio data. Defaults to 'audio'"
192
+ },
193
+ )
194
+ text_column_name: str = field(
195
+ default="text",
196
+ metadata={
197
+ "help": "The name of the dataset column containing the text data. Defaults to 'text'"
198
+ },
199
+ )
200
+ overwrite_cache: bool = field(
201
+ default=False,
202
+ metadata={"help": "Overwrite the cached preprocessed datasets or not."},
203
+ )
204
+ preprocessing_num_workers: Optional[int] = field(
205
+ default=None,
206
+ metadata={"help": "The number of processes to use for the preprocessing."},
207
+ )
208
+ max_train_samples: Optional[int] = field(
209
+ default=None,
210
+ metadata={
211
+ "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
212
+ "value if set."
213
+ },
214
+ )
215
+ max_eval_samples: Optional[int] = field(
216
+ default=None,
217
+ metadata={
218
+ "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
219
+ "value if set."
220
+ },
221
+ )
222
+ chars_to_ignore: Optional[List[str]] = list_field(
223
+ default=None,
224
+ metadata={"help": "A list of characters to remove from the transcripts."},
225
+ )
226
+ eval_metrics: List[str] = list_field(
227
+ default=["wer", "cer"],
228
+ metadata={
229
+ "help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"
230
+ },
231
+ )
232
+ max_duration_in_seconds: float = field(
233
+ default=20.0,
234
+ metadata={
235
+ "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
236
+ },
237
+ )
238
+ min_duration_in_seconds: float = field(
239
+ default=0.0,
240
+ metadata={
241
+ "help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"
242
+ },
243
+ )
244
+ preprocessing_only: bool = field(
245
+ default=False,
246
+ metadata={
247
+ "help": "Whether to only do data preprocessing and skip training. "
248
+ "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
249
+ "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
250
+ "so that the cached datasets can consequently be loaded in distributed training"
251
+ },
252
+ )
253
+ use_auth_token: bool = field(
254
+ default=False,
255
+ metadata={
256
+ "help": "If :obj:`True`, will use the token generated when running"
257
+ ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
258
+ },
259
+ )
260
+ unk_token: str = field(
261
+ default="[UNK]",
262
+ metadata={"help": "The unk token for the tokenizer"},
263
+ )
264
+ pad_token: str = field(
265
+ default="[PAD]",
266
+ metadata={"help": "The padding token for the tokenizer"},
267
+ )
268
+ word_delimiter_token: str = field(
269
+ default="|",
270
+ metadata={"help": "The word delimiter token for the tokenizer"},
271
+ )
272
+ phoneme_language: Optional[str] = field(
273
+ default=None,
274
+ metadata={
275
+ "help": "The target language that should be used be"
276
+ " passed to the tokenizer for tokenization. Note that"
277
+ " this is only relevant if the model classifies the"
278
+ " input audio to a sequence of phoneme sequences."
279
+ },
280
+ )
281
+
282
+
283
+ @dataclass
284
+ class ExtraArguments:
285
+ "Additional training arguments"
286
+ bnb: bool = field(default=False, metadata={"help": "If true uses 8bit Adam"})
287
+ tristage_sched: bool = field(
288
+ default=False,
289
+ metadata={"help": "If true uses tristage LR scheduler (refer to XLS-R paper)"},
290
+ )
291
+ wandb_project: str = field(
292
+ default="", metadata={"help": "Name of wandb project to log into"}
293
+ )
294
+
295
+
296
+ @dataclass
297
+ class DataCollatorCTCWithPadding:
298
+ """
299
+ Data collator that will dynamically pad the inputs received.
300
+ Args:
301
+ processor (:class:`~transformers.AutoProcessor`)
302
+ The processor used for proccessing the data.
303
+ padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
304
+ Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
305
+ among:
306
+ * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
307
+ sequence if provided).
308
+ * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
309
+ maximum acceptable input length for the model if that argument is not provided.
310
+ * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
311
+ different lengths).
312
+ max_length (:obj:`int`, `optional`):
313
+ Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
314
+ max_length_labels (:obj:`int`, `optional`):
315
+ Maximum length of the ``labels`` returned list and optionally padding length (see above).
316
+ pad_to_multiple_of (:obj:`int`, `optional`):
317
+ If set will pad the sequence to a multiple of the provided value.
318
+ This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
319
+ 7.5 (Volta).
320
+ """
321
+
322
+ processor: AutoProcessor
323
+ padding: Union[bool, str] = "longest"
324
+ pad_to_multiple_of: Optional[int] = None
325
+ pad_to_multiple_of_labels: Optional[int] = None
326
+
327
+ def __call__(
328
+ self, features: List[Dict[str, Union[List[int], torch.Tensor]]]
329
+ ) -> Dict[str, torch.Tensor]:
330
+ # split inputs and labels since they have to be of different lenghts and need
331
+ # different padding methods
332
+ input_features = [
333
+ {"input_values": feature["input_values"]} for feature in features
334
+ ]
335
+ label_features = [{"input_ids": feature["labels"]} for feature in features]
336
+
337
+ batch = self.processor.pad(
338
+ input_features,
339
+ padding=self.padding,
340
+ pad_to_multiple_of=self.pad_to_multiple_of,
341
+ return_tensors="pt",
342
+ )
343
+
344
+ with self.processor.as_target_processor():
345
+ labels_batch = self.processor.pad(
346
+ label_features,
347
+ padding=self.padding,
348
+ pad_to_multiple_of=self.pad_to_multiple_of_labels,
349
+ return_tensors="pt",
350
+ )
351
+
352
+ # replace padding with -100 to ignore loss correctly
353
+ labels = labels_batch["input_ids"].masked_fill(
354
+ labels_batch.attention_mask.ne(1), -100
355
+ )
356
+
357
+ batch["labels"] = labels
358
+
359
+ return batch
360
+
361
+
362
+ def get_tri_stage_schedule(
363
+ optimizer,
364
+ num_training_steps,
365
+ ratios=[0.1, 0.4, 0.5],
366
+ num_warmup_steps=None,
367
+ num_hold_steps=None,
368
+ start_ratio=0.01,
369
+ end_ratio=0.05,
370
+ ):
371
+ assert (num_warmup_steps is None) == (num_hold_steps is None)
372
+ if num_warmup_steps is None:
373
+ num_warmup_steps = int(ratios[0] * num_training_steps)
374
+ num_hold_steps = int(ratios[1] * num_training_steps)
375
+ start_decay_step = num_warmup_steps + num_hold_steps
376
+ a_w, b_w = (1 - start_ratio) / num_warmup_steps, start_ratio
377
+ num_decay_steps = num_training_steps - start_decay_step
378
+ a_d, b_d = (end_ratio - 1) / num_decay_steps, 1.0
379
+
380
+ def lr_lambda(current_step):
381
+ if current_step < num_warmup_steps:
382
+ return a_w * float(current_step) + b_w
383
+ if current_step < start_decay_step:
384
+ return 1.0
385
+ return max(end_ratio, a_d * float(current_step - start_decay_step) + b_d)
386
+
387
+ return LambdaLR(optimizer, lr_lambda)
388
+
389
+
390
+ def create_vocabulary_from_data(
391
+ datasets: DatasetDict,
392
+ word_delimiter_token: Optional[str] = None,
393
+ unk_token: Optional[str] = None,
394
+ pad_token: Optional[str] = None,
395
+ ):
396
+ # Given training and test labels create vocabulary
397
+ def extract_all_chars(batch):
398
+ all_text = " ".join(batch["target_text"])
399
+ vocab = list(set(all_text))
400
+ return {"vocab": [vocab], "all_text": [all_text]}
401
+
402
+ vocabs = datasets.map(
403
+ extract_all_chars,
404
+ batched=True,
405
+ batch_size=-1,
406
+ keep_in_memory=True,
407
+ remove_columns=datasets["train"].column_names,
408
+ )
409
+
410
+ # take union of all unique characters in each dataset
411
+ vocab_set = functools.reduce(
412
+ lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]),
413
+ vocabs.values(),
414
+ )
415
+
416
+ vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
417
+
418
+ # replace white space with delimiter token
419
+ if word_delimiter_token is not None:
420
+ vocab_dict[word_delimiter_token] = vocab_dict[" "]
421
+ del vocab_dict[" "]
422
+
423
+ # add unk and pad token
424
+ if unk_token is not None:
425
+ vocab_dict[unk_token] = len(vocab_dict)
426
+
427
+ if pad_token is not None:
428
+ vocab_dict[pad_token] = len(vocab_dict)
429
+
430
+ return vocab_dict
431
+
432
+
433
+ def main():
434
+ # See all possible arguments in src/transformers/training_args.py
435
+ # or by passing the --help flag to this script.
436
+ # We now keep distinct sets of args, for a cleaner separation of concerns.
437
+
438
+ parser = HfArgumentParser(
439
+ (ModelArguments, DataTrainingArguments, TrainingArguments, ExtraArguments)
440
+ )
441
+ if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
442
+ # If we pass only one argument to the script and it's the path to a json file,
443
+ # let's parse it to get our arguments.
444
+ model_args, data_args, training_args, extra_args = parser.parse_json_file(
445
+ json_file=os.path.abspath(sys.argv[1])
446
+ )
447
+ else:
448
+ (
449
+ model_args,
450
+ data_args,
451
+ training_args,
452
+ extra_args,
453
+ ) = parser.parse_args_into_dataclasses()
454
+
455
+ # Detecting last checkpoint.
456
+ last_checkpoint = None
457
+ if (
458
+ os.path.isdir(training_args.output_dir)
459
+ and training_args.do_train
460
+ and not training_args.overwrite_output_dir
461
+ ):
462
+ last_checkpoint = get_last_checkpoint(training_args.output_dir)
463
+ if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
464
+ raise ValueError(
465
+ f"Output directory ({training_args.output_dir}) already exists and is not empty. "
466
+ "Use --overwrite_output_dir to overcome."
467
+ )
468
+ elif last_checkpoint is not None:
469
+ logger.info(
470
+ f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
471
+ "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
472
+ )
473
+
474
+ # Setup logging
475
+ logging.basicConfig(
476
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
477
+ datefmt="%m/%d/%Y %H:%M:%S",
478
+ handlers=[logging.StreamHandler(sys.stdout)],
479
+ )
480
+ logger.setLevel(
481
+ logging.INFO if is_main_process(training_args.local_rank) else logging.WARN
482
+ )
483
+
484
+ # Log on each process the small summary:
485
+ logger.warning(
486
+ f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
487
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
488
+ )
489
+ # Set the verbosity to info of the Transformers logger (on main process only):
490
+ if is_main_process(training_args.local_rank):
491
+ transformers.utils.logging.set_verbosity_info()
492
+ logger.info("Training/evaluation parameters %s", training_args)
493
+
494
+ # Set seed before initializing model.
495
+ set_seed(training_args.seed)
496
+
497
+ # configure wandb run
498
+ os.environ["WANDB_PROJECT"] = extra_args.wandb_project
499
+
500
+ # 1. First, let's load the dataset
501
+ raw_datasets = DatasetDict()
502
+
503
+ if training_args.do_train:
504
+ if os.path.isdir(data_args.dataset_name):
505
+ raw_datasets["train"] = load_from_disk(
506
+ f"{data_args.dataset_name}/{data_args.train_split_name}"
507
+ )
508
+ else:
509
+ raw_datasets["train"] = load_dataset(
510
+ data_args.dataset_name,
511
+ data_args.dataset_config_name,
512
+ split=data_args.train_split_name,
513
+ use_auth_token=data_args.use_auth_token,
514
+ )
515
+
516
+ if data_args.audio_column_name not in raw_datasets["train"].column_names:
517
+ raise ValueError(
518
+ f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
519
+ "Make sure to set `--audio_column_name` to the correct audio column - one of "
520
+ f"{', '.join(raw_datasets['train'].column_names)}."
521
+ )
522
+
523
+ if data_args.text_column_name not in raw_datasets["train"].column_names:
524
+ raise ValueError(
525
+ f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
526
+ "Make sure to set `--text_column_name` to the correct text column - one of "
527
+ f"{', '.join(raw_datasets['train'].column_names)}."
528
+ )
529
+
530
+ if data_args.max_train_samples is not None:
531
+ raw_datasets["train"] = raw_datasets["train"].select(
532
+ range(data_args.max_train_samples)
533
+ )
534
+
535
+ if training_args.do_eval:
536
+ if os.path.isdir(data_args.dataset_name):
537
+ raw_datasets["eval"] = load_from_disk(
538
+ f"{data_args.dataset_name}/{data_args.eval_split_name}"
539
+ )
540
+ else:
541
+ raw_datasets["eval"] = load_dataset(
542
+ data_args.dataset_name,
543
+ data_args.dataset_config_name,
544
+ split=data_args.eval_split_name,
545
+ use_auth_token=data_args.use_auth_token,
546
+ )
547
+
548
+ if data_args.max_eval_samples is not None:
549
+ raw_datasets["eval"] = raw_datasets["eval"].select(
550
+ range(data_args.max_eval_samples)
551
+ )
552
+
553
+ # 2. We remove some special characters from the datasets
554
+ # that make training complicated and do not help in transcribing the speech
555
+ # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
556
+ # that could be easily picked up by the model
557
+ chars_to_ignore_regex = (
558
+ f'[{"".join(data_args.chars_to_ignore)}]'
559
+ if data_args.chars_to_ignore is not None
560
+ else None
561
+ )
562
+ text_column_name = data_args.text_column_name
563
+
564
+ def remove_special_characters(batch):
565
+ if chars_to_ignore_regex is not None:
566
+ batch["target_text"] = (
567
+ re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
568
+ )
569
+ else:
570
+ batch["target_text"] = batch[text_column_name].lower() + " "
571
+ # Unicode Normalization
572
+ batch["target_text"] = unicodedata.normalize("NFKC", batch["target_text"])
573
+ return batch
574
+
575
+ with training_args.main_process_first(
576
+ desc="dataset map special characters removal"
577
+ ):
578
+ raw_datasets = raw_datasets.map(
579
+ remove_special_characters,
580
+ remove_columns=[text_column_name],
581
+ desc="remove special characters from datasets",
582
+ )
583
+
584
+ # save special tokens for tokenizer
585
+ word_delimiter_token = data_args.word_delimiter_token
586
+ unk_token = data_args.unk_token
587
+ pad_token = data_args.pad_token
588
+
589
+ # 3. Next, let's load the config as we might need it to create
590
+ # the tokenizer
591
+ # load config
592
+ config = AutoConfig.from_pretrained(
593
+ model_args.model_name_or_path,
594
+ cache_dir=model_args.cache_dir,
595
+ use_auth_token=data_args.use_auth_token,
596
+ )
597
+
598
+ # 4. Next, if no tokenizer file is defined,
599
+ # we create the vocabulary of the model by extracting all unique characters from
600
+ # the training and evaluation datasets
601
+ # We need to make sure that only first rank saves vocabulary
602
+ # make sure all processes wait until vocab is created
603
+ tokenizer_name_or_path = model_args.tokenizer_name_or_path
604
+ tokenizer_kwargs = {}
605
+ if tokenizer_name_or_path is None:
606
+ # save vocab in training output dir
607
+ tokenizer_name_or_path = training_args.output_dir
608
+
609
+ vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
610
+
611
+ with training_args.main_process_first():
612
+ if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
613
+ os.remove(vocab_file)
614
+
615
+ with training_args.main_process_first(desc="dataset map vocabulary creation"):
616
+ if not os.path.isfile(vocab_file):
617
+ os.makedirs(tokenizer_name_or_path, exist_ok=True)
618
+ vocab_dict = create_vocabulary_from_data(
619
+ raw_datasets,
620
+ word_delimiter_token=word_delimiter_token,
621
+ unk_token=unk_token,
622
+ pad_token=pad_token,
623
+ )
624
+
625
+ # save vocab dict to be loaded into tokenizer
626
+ with open(vocab_file, "w") as file:
627
+ json.dump(vocab_dict, file)
628
+
629
+ # if tokenizer has just been created
630
+ # it is defined by `tokenizer_class` if present in config else by `model_type`
631
+ tokenizer_kwargs = {
632
+ "config": config if config.tokenizer_class is not None else None,
633
+ "tokenizer_type": config.model_type
634
+ if config.tokenizer_class is None
635
+ else None,
636
+ "unk_token": unk_token,
637
+ "pad_token": pad_token,
638
+ "word_delimiter_token": word_delimiter_token,
639
+ }
640
+
641
+ # 5. Now we can instantiate the feature extractor, tokenizer and model
642
+ # Note for distributed training, the .from_pretrained methods guarantee that only
643
+ # one local process can concurrently download model & vocab.
644
+
645
+ # load feature_extractor and tokenizer
646
+ tokenizer = AutoTokenizer.from_pretrained(
647
+ tokenizer_name_or_path,
648
+ use_auth_token=data_args.use_auth_token,
649
+ **tokenizer_kwargs,
650
+ )
651
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
652
+ model_args.model_name_or_path,
653
+ cache_dir=model_args.cache_dir,
654
+ use_auth_token=data_args.use_auth_token,
655
+ )
656
+
657
+ # adapt config
658
+ config.update(
659
+ {
660
+ "feat_proj_dropout": model_args.feat_proj_dropout,
661
+ "attention_dropout": model_args.attention_dropout,
662
+ "hidden_dropout": model_args.hidden_dropout,
663
+ "final_dropout": model_args.final_dropout,
664
+ "mask_time_prob": model_args.mask_time_prob,
665
+ "mask_time_length": model_args.mask_time_length,
666
+ "mask_feature_prob": model_args.mask_feature_prob,
667
+ "mask_feature_length": model_args.mask_feature_length,
668
+ "gradient_checkpointing": training_args.gradient_checkpointing,
669
+ "layerdrop": model_args.layerdrop,
670
+ "ctc_loss_reduction": model_args.ctc_loss_reduction,
671
+ "pad_token_id": tokenizer.pad_token_id,
672
+ "vocab_size": len(tokenizer),
673
+ "activation_dropout": model_args.activation_dropout,
674
+ }
675
+ )
676
+
677
+ # create model
678
+ model = AutoModelForCTC.from_pretrained(
679
+ model_args.model_name_or_path,
680
+ cache_dir=model_args.cache_dir,
681
+ config=config,
682
+ use_auth_token=data_args.use_auth_token,
683
+ )
684
+
685
+ # freeze encoder
686
+ if model_args.freeze_feature_encoder:
687
+ model.freeze_feature_encoder()
688
+
689
+ # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
690
+ # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
691
+ # so that we just need to set the correct target sampling rate and normalize the input
692
+ # via the `feature_extractor`
693
+
694
+ # make sure that dataset decodes audio with correct sampling rate
695
+ dataset_sampling_rate = (
696
+ next(iter(raw_datasets.values()))
697
+ .features[data_args.audio_column_name]
698
+ .sampling_rate
699
+ )
700
+ if dataset_sampling_rate != feature_extractor.sampling_rate:
701
+ raw_datasets = raw_datasets.cast_column(
702
+ data_args.audio_column_name,
703
+ datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate),
704
+ )
705
+
706
+ # derive max & min input length for sample rate & max duration
707
+ max_input_length = (
708
+ data_args.max_duration_in_seconds * feature_extractor.sampling_rate
709
+ )
710
+ min_input_length = (
711
+ data_args.min_duration_in_seconds * feature_extractor.sampling_rate
712
+ )
713
+ audio_column_name = data_args.audio_column_name
714
+ num_workers = data_args.preprocessing_num_workers
715
+
716
+ # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
717
+ phoneme_language = data_args.phoneme_language
718
+
719
+ # Preprocessing the datasets.
720
+ # We need to read the audio files as arrays and tokenize the targets.
721
+ def prepare_dataset(batch):
722
+ # load audio
723
+ sample = batch[audio_column_name]
724
+
725
+ inputs = feature_extractor(
726
+ sample["array"], sampling_rate=sample["sampling_rate"]
727
+ )
728
+ batch["input_values"] = inputs.input_values[0]
729
+ batch["length"] = len(batch["input_values"])
730
+
731
+ # encode targets
732
+ additional_kwargs = {}
733
+ if phoneme_language is not None:
734
+ additional_kwargs["phonemizer_lang"] = phoneme_language
735
+
736
+ batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
737
+ return batch
738
+
739
+ with training_args.main_process_first(desc="dataset map preprocessing"):
740
+ vectorized_datasets = raw_datasets.map(
741
+ prepare_dataset,
742
+ remove_columns=next(iter(raw_datasets.values())).column_names,
743
+ num_proc=num_workers,
744
+ desc="preprocess datasets",
745
+ )
746
+
747
+ def is_audio_in_length_range(length):
748
+ return length > min_input_length and length < max_input_length
749
+
750
+ # filter data that is shorter than min_input_length
751
+ vectorized_datasets = vectorized_datasets.filter(
752
+ is_audio_in_length_range,
753
+ num_proc=num_workers,
754
+ input_columns=["length"],
755
+ )
756
+
757
+ # 7. Next, we can prepare the training.
758
+ # Let's use word error rate (WER) as our evaluation metric,
759
+ # instantiate a data collator and the trainer
760
+
761
+ # Define evaluation metrics during training, *i.e.* word error rate, character error rate
762
+ eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
763
+
764
+ # for large datasets it is advised to run the preprocessing on a
765
+ # single machine first with ``args.preprocessing_only`` since there will mostly likely
766
+ # be a timeout when running the script in distributed mode.
767
+ # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
768
+ # cached dataset
769
+ if data_args.preprocessing_only:
770
+ logger.info(
771
+ f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}"
772
+ )
773
+ return
774
+
775
+ def compute_metrics(pred):
776
+ pred_logits = pred.predictions
777
+ pred_ids = np.argmax(pred_logits, axis=-1)
778
+
779
+ pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
780
+
781
+ pred_str = tokenizer.batch_decode(pred_ids)
782
+ # we do not want to group tokens when computing the metrics
783
+ label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
784
+
785
+ metrics = {
786
+ k: v.compute(predictions=pred_str, references=label_str)
787
+ for k, v in eval_metrics.items()
788
+ }
789
+
790
+ return metrics
791
+
792
+ # Now save everything to be able to create a single processor later
793
+ if is_main_process(training_args.local_rank):
794
+ # save feature extractor, tokenizer and config
795
+ feature_extractor.save_pretrained(training_args.output_dir)
796
+ tokenizer.save_pretrained(training_args.output_dir)
797
+ config.save_pretrained(training_args.output_dir)
798
+
799
+ try:
800
+ processor = AutoProcessor.from_pretrained(training_args.output_dir)
801
+ except (OSError, KeyError):
802
+ warnings.warn(
803
+ "Loading a processor from a feature extractor config that does not"
804
+ " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
805
+ " attribute to your `preprocessor_config.json` file to suppress this warning: "
806
+ " `'processor_class': 'Wav2Vec2Processor'`",
807
+ FutureWarning,
808
+ )
809
+ processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
810
+
811
+ # Instantiate custom data collator
812
+ data_collator = DataCollatorCTCWithPadding(processor=processor)
813
+
814
+ decay_parameters = get_parameter_names(model, [torch.nn.LayerNorm])
815
+ decay_parameters = [name for name in decay_parameters if "bias" not in name]
816
+ optimizer_grouped_parameters = [
817
+ {
818
+ "params": [p for n, p in model.named_parameters() if n in decay_parameters],
819
+ "weight_decay": training_args.weight_decay,
820
+ },
821
+ {
822
+ "params": [
823
+ p for n, p in model.named_parameters() if n not in decay_parameters
824
+ ],
825
+ "weight_decay": 0.0,
826
+ },
827
+ ]
828
+ if extra_args.bnb:
829
+ optimizer = bnb.optim.Adam8bit(
830
+ params=optimizer_grouped_parameters,
831
+ lr=training_args.learning_rate,
832
+ betas=(training_args.adam_beta1, training_args.adam_beta2),
833
+ eps=training_args.adam_epsilon,
834
+ )
835
+ else:
836
+ optimizer = torch.optim.AdamW(
837
+ params=optimizer_grouped_parameters,
838
+ lr=training_args.learning_rate,
839
+ betas=(training_args.adam_beta1, training_args.adam_beta2),
840
+ eps=training_args.adam_epsilon,
841
+ )
842
+ if extra_args.tristage_sched:
843
+ scheduler = get_tri_stage_schedule(optimizer, training_args.max_steps)
844
+ else:
845
+ scheduler = None
846
+ optimizers = (optimizer, scheduler)
847
+
848
+ # Initialize Trainer
849
+ trainer = Trainer(
850
+ model=model,
851
+ data_collator=data_collator,
852
+ args=training_args,
853
+ compute_metrics=compute_metrics,
854
+ train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
855
+ eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
856
+ tokenizer=feature_extractor,
857
+ optimizers=optimizers,
858
+ )
859
+
860
+ # 8. Finally, we can start training
861
+
862
+ # Training
863
+ if training_args.do_train:
864
+
865
+ # use last checkpoint if exist
866
+ if last_checkpoint is not None:
867
+ checkpoint = last_checkpoint
868
+ elif os.path.isdir(model_args.model_name_or_path):
869
+ checkpoint = model_args.model_name_or_path
870
+ else:
871
+ checkpoint = None
872
+
873
+ train_result = trainer.train(resume_from_checkpoint=checkpoint)
874
+ trainer.save_model()
875
+
876
+ metrics = train_result.metrics
877
+ max_train_samples = (
878
+ data_args.max_train_samples
879
+ if data_args.max_train_samples is not None
880
+ else len(vectorized_datasets["train"])
881
+ )
882
+ metrics["train_samples"] = min(
883
+ max_train_samples, len(vectorized_datasets["train"])
884
+ )
885
+
886
+ trainer.log_metrics("train", metrics)
887
+ trainer.save_metrics("train", metrics)
888
+ trainer.save_state()
889
+
890
+ # Evaluation
891
+ results = {}
892
+ if training_args.do_eval:
893
+ logger.info("*** Evaluate ***")
894
+ metrics = trainer.evaluate()
895
+ max_eval_samples = (
896
+ data_args.max_eval_samples
897
+ if data_args.max_eval_samples is not None
898
+ else len(vectorized_datasets["eval"])
899
+ )
900
+ metrics["eval_samples"] = min(
901
+ max_eval_samples, len(vectorized_datasets["eval"])
902
+ )
903
+
904
+ trainer.log_metrics("eval", metrics)
905
+ trainer.save_metrics("eval", metrics)
906
+
907
+ # Write model card and (optionally) push to hub
908
+ config_name = (
909
+ data_args.dataset_config_name
910
+ if data_args.dataset_config_name is not None
911
+ else "na"
912
+ )
913
+ kwargs = {
914
+ "finetuned_from": model_args.model_name_or_path,
915
+ "tasks": "speech-recognition",
916
+ "tags": ["automatic-speech-recognition", data_args.dataset_name],
917
+ "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
918
+ "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
919
+ }
920
+ if "common_voice" in data_args.dataset_name:
921
+ kwargs["language"] = config_name
922
+
923
+ if training_args.push_to_hub:
924
+ trainer.push_to_hub(**kwargs)
925
+ else:
926
+ trainer.create_model_card(**kwargs)
927
+
928
+ return results
929
+
930
+
931
+ if __name__ == "__main__":
932
+ main()
special_tokens_map.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"bos_token": "<s>", "eos_token": "</s>", "unk_token": "[UNK]", "pad_token": "[PAD]", "additional_special_tokens": [{"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}]}
tokenizer_config.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"unk_token": "[UNK]", "bos_token": "<s>", "eos_token": "</s>", "pad_token": "[PAD]", "do_lower_case": false, "word_delimiter_token": "|", "replace_word_delimiter_char": " ", "special_tokens_map_file": null, "name_or_path": "./", "tokenizer_class": "Wav2Vec2CTCTokenizer"}
training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:00e1cda6a71202a91b8f4962bfa088ea5f49af157712a4bdf3479ce459f75bc8
3
+ size 3119
vocab.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"ँ": 1, "ं": 2, "ः": 3, "अ": 4, "आ": 5, "इ": 6, "ई": 7, "उ": 8, "ऊ": 9, "ऋ": 10, "ऌ": 11, "ऍ": 12, "ऎ": 13, "ए": 14, "ऐ": 15, "ऑ": 16, "ओ": 17, "औ": 18, "क": 19, "ख": 20, "ग": 21, "घ": 22, "ङ": 23, "च": 24, "छ": 25, "ज": 26, "झ": 27, "ञ": 28, "ट": 29, "ठ": 30, "ड": 31, "ढ": 32, "ण": 33, "त": 34, "थ": 35, "द": 36, "ध": 37, "न": 38, "प": 39, "फ": 40, "ब": 41, "भ": 42, "म": 43, "य": 44, "र": 45, "ऱ": 46, "ल": 47, "ळ": 48, "व": 49, "श": 50, "ष": 51, "स": 52, "ह": 53, "़": 54, "ऽ": 55, "ा": 56, "ि": 57, "ी": 58, "ु": 59, "ू": 60, "ृ": 61, "ॄ": 62, "ॅ": 63, "ॆ": 64, "े": 65, "ै": 66, "ॉ": 67, "ॊ": 68, "ो": 69, "ौ": 70, "्": 71, "ॢ": 72, "ॲ": 73, "|": 0, "[UNK]": 74, "[PAD]": 75}