anuragshas commited on
Commit
44af273
1 Parent(s): 591f7ea

Training in progress, step 400

Browse files
.gitignore ADDED
@@ -0,0 +1 @@
 
 
1
+ checkpoint-*/
.ipynb_checkpoints/run-checkpoint.sh ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ python run_speech_recognition_ctc.py \
2
+ --dataset_name="mozilla-foundation/common_voice_7_0" \
3
+ --model_name_or_path="facebook/wav2vec2-xls-r-1b" \
4
+ --dataset_config_name="hi" \
5
+ --output_dir="./" \
6
+ --overwrite_output_dir \
7
+ --num_train_epochs="100" \
8
+ --per_device_train_batch_size="8" \
9
+ --per_device_eval_batch_size="16" \
10
+ --gradient_accumulation_steps="4" \
11
+ --learning_rate="7.5e-5" \
12
+ --warmup_steps="2000" \
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
+ --push_to_hub
.ipynb_checkpoints/run_speech_recognition_ctc-checkpoint.py ADDED
@@ -0,0 +1,765 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 warnings
25
+ import unicodedata
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 datasets import DatasetDict, load_dataset, load_metric
33
+
34
+ import bitsandbytes as bnb
35
+ import transformers
36
+ from transformers import (
37
+ AutoConfig,
38
+ AutoFeatureExtractor,
39
+ AutoModelForCTC,
40
+ AutoProcessor,
41
+ AutoTokenizer,
42
+ HfArgumentParser,
43
+ Trainer,
44
+ TrainingArguments,
45
+ Wav2Vec2Processor,
46
+ set_seed,
47
+ )
48
+ from transformers.trainer_pt_utils import get_parameter_names
49
+ from transformers.trainer_utils import get_last_checkpoint, is_main_process
50
+ from transformers.utils import check_min_version
51
+ from transformers.utils.versions import require_version
52
+
53
+
54
+ # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
55
+ check_min_version("4.16.0.dev0")
56
+
57
+ require_version("datasets>=1.13.3", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
58
+
59
+
60
+ logger = logging.getLogger(__name__)
61
+
62
+
63
+ def list_field(default=None, metadata=None):
64
+ return field(default_factory=lambda: default, metadata=metadata)
65
+
66
+
67
+ @dataclass
68
+ class ModelArguments:
69
+ """
70
+ Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
71
+ """
72
+
73
+ model_name_or_path: str = field(
74
+ metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
75
+ )
76
+ tokenizer_name_or_path: Optional[str] = field(
77
+ default=None,
78
+ metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
79
+ )
80
+ cache_dir: Optional[str] = field(
81
+ default=None,
82
+ metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
83
+ )
84
+ freeze_feature_encoder: bool = field(
85
+ default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
86
+ )
87
+ attention_dropout: float = field(
88
+ default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
89
+ )
90
+ activation_dropout: float = field(
91
+ default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
92
+ )
93
+ feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
94
+ hidden_dropout: float = field(
95
+ default=0.0,
96
+ metadata={
97
+ "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
98
+ },
99
+ )
100
+ final_dropout: float = field(
101
+ default=0.0,
102
+ metadata={"help": "The dropout probability for the final projection layer."},
103
+ )
104
+ mask_time_prob: float = field(
105
+ default=0.05,
106
+ metadata={
107
+ "help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
108
+ "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
109
+ "vectors will be masked along the time axis."
110
+ },
111
+ )
112
+ mask_time_length: int = field(
113
+ default=10,
114
+ metadata={"help": "Length of vector span to mask along the time axis."},
115
+ )
116
+ mask_feature_prob: float = field(
117
+ default=0.0,
118
+ metadata={
119
+ "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
120
+ "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
121
+ },
122
+ )
123
+ mask_feature_length: int = field(
124
+ default=10,
125
+ metadata={"help": "Length of vector span to mask along the feature axis."},
126
+ )
127
+ layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
128
+ ctc_loss_reduction: Optional[str] = field(
129
+ default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
130
+ )
131
+
132
+
133
+ @dataclass
134
+ class DataTrainingArguments:
135
+ """
136
+ Arguments pertaining to what data we are going to input our model for training and eval.
137
+
138
+ Using `HfArgumentParser` we can turn this class
139
+ into argparse arguments to be able to specify them on
140
+ the command line.
141
+ """
142
+
143
+ dataset_name: str = field(
144
+ metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
145
+ )
146
+ dataset_config_name: str = field(
147
+ default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
148
+ )
149
+ train_split_name: str = field(
150
+ default="train+validation",
151
+ metadata={
152
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
153
+ },
154
+ )
155
+ eval_split_name: str = field(
156
+ default="test",
157
+ metadata={
158
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
159
+ },
160
+ )
161
+ audio_column_name: str = field(
162
+ default="audio",
163
+ metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
164
+ )
165
+ text_column_name: str = field(
166
+ default="text",
167
+ metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
168
+ )
169
+ overwrite_cache: bool = field(
170
+ default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
171
+ )
172
+ preprocessing_num_workers: Optional[int] = field(
173
+ default=None,
174
+ metadata={"help": "The number of processes to use for the preprocessing."},
175
+ )
176
+ max_train_samples: Optional[int] = field(
177
+ default=None,
178
+ metadata={
179
+ "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
180
+ "value if set."
181
+ },
182
+ )
183
+ max_eval_samples: Optional[int] = field(
184
+ default=None,
185
+ metadata={
186
+ "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
187
+ "value if set."
188
+ },
189
+ )
190
+ chars_to_ignore: Optional[List[str]] = list_field(
191
+ default=None,
192
+ metadata={"help": "A list of characters to remove from the transcripts."},
193
+ )
194
+ eval_metrics: List[str] = list_field(
195
+ default=["wer"],
196
+ metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
197
+ )
198
+ max_duration_in_seconds: float = field(
199
+ default=20.0,
200
+ metadata={
201
+ "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
202
+ },
203
+ )
204
+ min_duration_in_seconds: float = field(
205
+ default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
206
+ )
207
+ preprocessing_only: bool = field(
208
+ default=False,
209
+ metadata={
210
+ "help": "Whether to only do data preprocessing and skip training. "
211
+ "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
212
+ "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
213
+ "so that the cached datasets can consequently be loaded in distributed training"
214
+ },
215
+ )
216
+ use_auth_token: bool = field(
217
+ default=False,
218
+ metadata={
219
+ "help": "If :obj:`True`, will use the token generated when running"
220
+ ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
221
+ },
222
+ )
223
+ unk_token: str = field(
224
+ default="[UNK]",
225
+ metadata={"help": "The unk token for the tokenizer"},
226
+ )
227
+ pad_token: str = field(
228
+ default="[PAD]",
229
+ metadata={"help": "The padding token for the tokenizer"},
230
+ )
231
+ word_delimiter_token: str = field(
232
+ default="|",
233
+ metadata={"help": "The word delimiter token for the tokenizer"},
234
+ )
235
+ phoneme_language: Optional[str] = field(
236
+ default=None,
237
+ metadata={
238
+ "help": "The target language that should be used be"
239
+ " passed to the tokenizer for tokenization. Note that"
240
+ " this is only relevant if the model classifies the"
241
+ " input audio to a sequence of phoneme sequences."
242
+ },
243
+ )
244
+
245
+
246
+ @dataclass
247
+ class DataCollatorCTCWithPadding:
248
+ """
249
+ Data collator that will dynamically pad the inputs received.
250
+ Args:
251
+ processor (:class:`~transformers.AutoProcessor`)
252
+ The processor used for proccessing the data.
253
+ padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
254
+ Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
255
+ among:
256
+ * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
257
+ sequence if provided).
258
+ * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
259
+ maximum acceptable input length for the model if that argument is not provided.
260
+ * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
261
+ different lengths).
262
+ max_length (:obj:`int`, `optional`):
263
+ Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
264
+ max_length_labels (:obj:`int`, `optional`):
265
+ Maximum length of the ``labels`` returned list and optionally padding length (see above).
266
+ pad_to_multiple_of (:obj:`int`, `optional`):
267
+ If set will pad the sequence to a multiple of the provided value.
268
+ This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
269
+ 7.5 (Volta).
270
+ """
271
+
272
+ processor: AutoProcessor
273
+ padding: Union[bool, str] = "longest"
274
+ pad_to_multiple_of: Optional[int] = None
275
+ pad_to_multiple_of_labels: Optional[int] = None
276
+
277
+ def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
278
+ # split inputs and labels since they have to be of different lenghts and need
279
+ # different padding methods
280
+ input_features = [{"input_values": feature["input_values"]} for feature in features]
281
+ label_features = [{"input_ids": feature["labels"]} for feature in features]
282
+
283
+ batch = self.processor.pad(
284
+ input_features,
285
+ padding=self.padding,
286
+ pad_to_multiple_of=self.pad_to_multiple_of,
287
+ return_tensors="pt",
288
+ )
289
+
290
+ with self.processor.as_target_processor():
291
+ labels_batch = self.processor.pad(
292
+ label_features,
293
+ padding=self.padding,
294
+ pad_to_multiple_of=self.pad_to_multiple_of_labels,
295
+ return_tensors="pt",
296
+ )
297
+
298
+ # replace padding with -100 to ignore loss correctly
299
+ labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
300
+
301
+ batch["labels"] = labels
302
+
303
+ return batch
304
+
305
+
306
+ def create_vocabulary_from_data(
307
+ datasets: DatasetDict,
308
+ word_delimiter_token: Optional[str] = None,
309
+ unk_token: Optional[str] = None,
310
+ pad_token: Optional[str] = None,
311
+ ):
312
+ # Given training and test labels create vocabulary
313
+ def extract_all_chars(batch):
314
+ all_text = " ".join(batch["target_text"])
315
+ vocab = list(set(all_text))
316
+ return {"vocab": [vocab], "all_text": [all_text]}
317
+
318
+ vocabs = datasets.map(
319
+ extract_all_chars,
320
+ batched=True,
321
+ batch_size=-1,
322
+ keep_in_memory=True,
323
+ remove_columns=datasets["train"].column_names,
324
+ )
325
+
326
+ # take union of all unique characters in each dataset
327
+ vocab_set = functools.reduce(
328
+ lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
329
+ )
330
+
331
+ vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
332
+
333
+ # replace white space with delimiter token
334
+ if word_delimiter_token is not None:
335
+ vocab_dict[word_delimiter_token] = vocab_dict[" "]
336
+ del vocab_dict[" "]
337
+
338
+ # add unk and pad token
339
+ if unk_token is not None:
340
+ vocab_dict[unk_token] = len(vocab_dict)
341
+
342
+ if pad_token is not None:
343
+ vocab_dict[pad_token] = len(vocab_dict)
344
+
345
+ return vocab_dict
346
+
347
+
348
+ def main():
349
+ # See all possible arguments in src/transformers/training_args.py
350
+ # or by passing the --help flag to this script.
351
+ # We now keep distinct sets of args, for a cleaner separation of concerns.
352
+
353
+ parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
354
+ if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
355
+ # If we pass only one argument to the script and it's the path to a json file,
356
+ # let's parse it to get our arguments.
357
+ model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
358
+ else:
359
+ model_args, data_args, training_args = parser.parse_args_into_dataclasses()
360
+
361
+ # Detecting last checkpoint.
362
+ last_checkpoint = None
363
+ if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
364
+ last_checkpoint = get_last_checkpoint(training_args.output_dir)
365
+ if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
366
+ raise ValueError(
367
+ f"Output directory ({training_args.output_dir}) already exists and is not empty. "
368
+ "Use --overwrite_output_dir to overcome."
369
+ )
370
+ elif last_checkpoint is not None:
371
+ logger.info(
372
+ f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
373
+ "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
374
+ )
375
+
376
+ # Setup logging
377
+ logging.basicConfig(
378
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
379
+ datefmt="%m/%d/%Y %H:%M:%S",
380
+ handlers=[logging.StreamHandler(sys.stdout)],
381
+ )
382
+ logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
383
+
384
+ # Log on each process the small summary:
385
+ logger.warning(
386
+ f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
387
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
388
+ )
389
+ # Set the verbosity to info of the Transformers logger (on main process only):
390
+ if is_main_process(training_args.local_rank):
391
+ transformers.utils.logging.set_verbosity_info()
392
+ logger.info("Training/evaluation parameters %s", training_args)
393
+
394
+ # Set seed before initializing model.
395
+ set_seed(training_args.seed)
396
+
397
+ # 1. First, let's load the dataset
398
+ raw_datasets = DatasetDict()
399
+
400
+ if training_args.do_train:
401
+ raw_datasets["train"] = load_dataset(
402
+ data_args.dataset_name,
403
+ data_args.dataset_config_name,
404
+ split=data_args.train_split_name,
405
+ use_auth_token=data_args.use_auth_token,
406
+ )
407
+
408
+ if data_args.audio_column_name not in raw_datasets["train"].column_names:
409
+ raise ValueError(
410
+ f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
411
+ "Make sure to set `--audio_column_name` to the correct audio column - one of "
412
+ f"{', '.join(raw_datasets['train'].column_names)}."
413
+ )
414
+
415
+ if data_args.text_column_name not in raw_datasets["train"].column_names:
416
+ raise ValueError(
417
+ f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
418
+ "Make sure to set `--text_column_name` to the correct text column - one of "
419
+ f"{', '.join(raw_datasets['train'].column_names)}."
420
+ )
421
+
422
+ if data_args.max_train_samples is not None:
423
+ raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
424
+
425
+ if training_args.do_eval:
426
+ raw_datasets["eval"] = load_dataset(
427
+ data_args.dataset_name,
428
+ data_args.dataset_config_name,
429
+ split=data_args.eval_split_name,
430
+ use_auth_token=data_args.use_auth_token,
431
+ )
432
+
433
+ if data_args.max_eval_samples is not None:
434
+ raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
435
+
436
+ # 2. We remove some special characters from the datasets
437
+ # that make training complicated and do not help in transcribing the speech
438
+ # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
439
+ # that could be easily picked up by the model
440
+ chars_to_ignore_regex = (
441
+ f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
442
+ )
443
+ text_column_name = data_args.text_column_name
444
+
445
+ def remove_special_characters(batch):
446
+ if chars_to_ignore_regex is not None:
447
+ batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower()
448
+ else:
449
+ batch["target_text"] = batch[text_column_name].lower()
450
+ # Remove english characters
451
+ batch["target_text"] = re.sub(r'[a-z]', '', batch["target_text"])
452
+ # Unicode Normalization
453
+ batch["target_text"] = unicodedata.normalize('NFKC', batch["target_text"])
454
+ return batch
455
+
456
+ with training_args.main_process_first(desc="dataset map special characters removal"):
457
+ raw_datasets = raw_datasets.map(
458
+ remove_special_characters,
459
+ remove_columns=[text_column_name],
460
+ desc="remove special characters from datasets",
461
+ )
462
+
463
+ # save special tokens for tokenizer
464
+ word_delimiter_token = data_args.word_delimiter_token
465
+ unk_token = data_args.unk_token
466
+ pad_token = data_args.pad_token
467
+
468
+ # 3. Next, let's load the config as we might need it to create
469
+ # the tokenizer
470
+ # load config
471
+ config = AutoConfig.from_pretrained(
472
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
473
+ )
474
+
475
+ # 4. Next, if no tokenizer file is defined,
476
+ # we create the vocabulary of the model by extracting all unique characters from
477
+ # the training and evaluation datasets
478
+ # We need to make sure that only first rank saves vocabulary
479
+ # make sure all processes wait until vocab is created
480
+ tokenizer_name_or_path = model_args.tokenizer_name_or_path
481
+ tokenizer_kwargs = {}
482
+ if tokenizer_name_or_path is None:
483
+ # save vocab in training output dir
484
+ tokenizer_name_or_path = training_args.output_dir
485
+
486
+ vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
487
+
488
+ with training_args.main_process_first():
489
+ if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
490
+ os.remove(vocab_file)
491
+
492
+ with training_args.main_process_first(desc="dataset map vocabulary creation"):
493
+ if not os.path.isfile(vocab_file):
494
+ os.makedirs(tokenizer_name_or_path, exist_ok=True)
495
+ vocab_dict = create_vocabulary_from_data(
496
+ raw_datasets,
497
+ word_delimiter_token=word_delimiter_token,
498
+ unk_token=unk_token,
499
+ pad_token=pad_token,
500
+ )
501
+
502
+ # save vocab dict to be loaded into tokenizer
503
+ with open(vocab_file, "w") as file:
504
+ json.dump(vocab_dict, file)
505
+
506
+ # if tokenizer has just been created
507
+ # it is defined by `tokenizer_class` if present in config else by `model_type`
508
+ tokenizer_kwargs = {
509
+ "config": config if config.tokenizer_class is not None else None,
510
+ "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
511
+ "unk_token": unk_token,
512
+ "pad_token": pad_token,
513
+ "word_delimiter_token": word_delimiter_token,
514
+ }
515
+
516
+ # 5. Now we can instantiate the feature extractor, tokenizer and model
517
+ # Note for distributed training, the .from_pretrained methods guarantee that only
518
+ # one local process can concurrently download model & vocab.
519
+
520
+ # load feature_extractor and tokenizer
521
+ tokenizer = AutoTokenizer.from_pretrained(
522
+ tokenizer_name_or_path,
523
+ use_auth_token=data_args.use_auth_token,
524
+ **tokenizer_kwargs,
525
+ )
526
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
527
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
528
+ )
529
+
530
+ # adapt config
531
+ config.update(
532
+ {
533
+ "feat_proj_dropout": model_args.feat_proj_dropout,
534
+ "attention_dropout": model_args.attention_dropout,
535
+ "hidden_dropout": model_args.hidden_dropout,
536
+ "final_dropout": model_args.final_dropout,
537
+ "mask_time_prob": model_args.mask_time_prob,
538
+ "mask_time_length": model_args.mask_time_length,
539
+ "mask_feature_prob": model_args.mask_feature_prob,
540
+ "mask_feature_length": model_args.mask_feature_length,
541
+ "gradient_checkpointing": training_args.gradient_checkpointing,
542
+ "layerdrop": model_args.layerdrop,
543
+ "ctc_loss_reduction": model_args.ctc_loss_reduction,
544
+ "pad_token_id": tokenizer.pad_token_id,
545
+ "vocab_size": len(tokenizer),
546
+ "activation_dropout": model_args.activation_dropout,
547
+ }
548
+ )
549
+
550
+ # create model
551
+ model = AutoModelForCTC.from_pretrained(
552
+ model_args.model_name_or_path,
553
+ cache_dir=model_args.cache_dir,
554
+ config=config,
555
+ use_auth_token=data_args.use_auth_token,
556
+ )
557
+
558
+ # freeze encoder
559
+ if model_args.freeze_feature_encoder:
560
+ model.freeze_feature_encoder()
561
+
562
+ # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
563
+ # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
564
+ # so that we just need to set the correct target sampling rate and normalize the input
565
+ # via the `feature_extractor`
566
+
567
+ # make sure that dataset decodes audio with correct sampling rate
568
+ dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
569
+ if dataset_sampling_rate != feature_extractor.sampling_rate:
570
+ raw_datasets = raw_datasets.cast_column(
571
+ data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
572
+ )
573
+
574
+ # derive max & min input length for sample rate & max duration
575
+ max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
576
+ min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
577
+ audio_column_name = data_args.audio_column_name
578
+ num_workers = data_args.preprocessing_num_workers
579
+
580
+ # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
581
+ phoneme_language = data_args.phoneme_language
582
+
583
+ # Preprocessing the datasets.
584
+ # We need to read the audio files as arrays and tokenize the targets.
585
+ def prepare_dataset(batch):
586
+ # load audio
587
+ sample = batch[audio_column_name]
588
+
589
+ inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
590
+ batch["input_values"] = inputs.input_values[0]
591
+ batch["input_length"] = len(batch["input_values"])
592
+
593
+ # encode targets
594
+ additional_kwargs = {}
595
+ if phoneme_language is not None:
596
+ additional_kwargs["phonemizer_lang"] = phoneme_language
597
+
598
+ batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
599
+ return batch
600
+
601
+ with training_args.main_process_first(desc="dataset map preprocessing"):
602
+ vectorized_datasets = raw_datasets.map(
603
+ prepare_dataset,
604
+ remove_columns=next(iter(raw_datasets.values())).column_names,
605
+ num_proc=num_workers,
606
+ desc="preprocess datasets",
607
+ )
608
+
609
+ def is_audio_in_length_range(length):
610
+ return length > min_input_length and length < max_input_length
611
+
612
+ # filter data that is shorter than min_input_length
613
+ vectorized_datasets = vectorized_datasets.filter(
614
+ is_audio_in_length_range,
615
+ num_proc=num_workers,
616
+ input_columns=["input_length"],
617
+ )
618
+
619
+ # 7. Next, we can prepare the training.
620
+ # Let's use word error rate (WER) as our evaluation metric,
621
+ # instantiate a data collator and the trainer
622
+
623
+ # Define evaluation metrics during training, *i.e.* word error rate, character error rate
624
+ eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
625
+
626
+ # for large datasets it is advised to run the preprocessing on a
627
+ # single machine first with ``args.preprocessing_only`` since there will mostly likely
628
+ # be a timeout when running the script in distributed mode.
629
+ # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
630
+ # cached dataset
631
+ if data_args.preprocessing_only:
632
+ logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
633
+ return
634
+
635
+ def compute_metrics(pred):
636
+ pred_logits = pred.predictions
637
+ pred_ids = np.argmax(pred_logits, axis=-1)
638
+
639
+ pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
640
+
641
+ pred_str = tokenizer.batch_decode(pred_ids)
642
+ # we do not want to group tokens when computing the metrics
643
+ label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
644
+
645
+ metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
646
+
647
+ return metrics
648
+
649
+ # Now save everything to be able to create a single processor later
650
+ if is_main_process(training_args.local_rank):
651
+ # save feature extractor, tokenizer and config
652
+ feature_extractor.save_pretrained(training_args.output_dir)
653
+ tokenizer.save_pretrained(training_args.output_dir)
654
+ config.save_pretrained(training_args.output_dir)
655
+
656
+ try:
657
+ processor = AutoProcessor.from_pretrained(training_args.output_dir)
658
+ except (OSError, KeyError):
659
+ warnings.warn(
660
+ "Loading a processor from a feature extractor config that does not"
661
+ " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
662
+ " attribute to your `preprocessor_config.json` file to suppress this warning: "
663
+ " `'processor_class': 'Wav2Vec2Processor'`",
664
+ FutureWarning,
665
+ )
666
+ processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
667
+
668
+ # Instantiate custom data collator
669
+ data_collator = DataCollatorCTCWithPadding(processor=processor)
670
+
671
+ # decay_parameters = get_parameter_names(model, [torch.nn.LayerNorm])
672
+ # decay_parameters = [name for name in decay_parameters if "bias" not in name]
673
+ # optimizer_grouped_parameters = [
674
+ # {
675
+ # "params": [p for n, p in model.named_parameters() if n in decay_parameters],
676
+ # "weight_decay": training_args.weight_decay,
677
+ # },
678
+ # {
679
+ # "params": [p for n, p in model.named_parameters() if n not in decay_parameters],
680
+ # "weight_decay": 0.0,
681
+ # },
682
+ # ]
683
+ # optimizer = bnb.optim.Adam8bit(
684
+ # params=optimizer_grouped_parameters,
685
+ # betas=(training_args.adam_beta1, training_args.adam_beta2),
686
+ # eps=training_args.adam_epsilon,
687
+ # )
688
+
689
+ # optimizers = (optimizer, None)
690
+
691
+ # Initialize Trainer
692
+ trainer = Trainer(
693
+ model=model,
694
+ data_collator=data_collator,
695
+ args=training_args,
696
+ compute_metrics=compute_metrics,
697
+ train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
698
+ eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
699
+ tokenizer=feature_extractor,
700
+ # optimizers=optimizers,
701
+ )
702
+
703
+ # 8. Finally, we can start training
704
+
705
+ # Training
706
+ if training_args.do_train:
707
+
708
+ # use last checkpoint if exist
709
+ if last_checkpoint is not None:
710
+ checkpoint = last_checkpoint
711
+ elif os.path.isdir(model_args.model_name_or_path):
712
+ checkpoint = model_args.model_name_or_path
713
+ else:
714
+ checkpoint = None
715
+
716
+ train_result = trainer.train(resume_from_checkpoint=checkpoint)
717
+ trainer.save_model()
718
+
719
+ metrics = train_result.metrics
720
+ max_train_samples = (
721
+ data_args.max_train_samples
722
+ if data_args.max_train_samples is not None
723
+ else len(vectorized_datasets["train"])
724
+ )
725
+ metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
726
+
727
+ trainer.log_metrics("train", metrics)
728
+ trainer.save_metrics("train", metrics)
729
+ trainer.save_state()
730
+
731
+ # Evaluation
732
+ results = {}
733
+ if training_args.do_eval:
734
+ logger.info("*** Evaluate ***")
735
+ metrics = trainer.evaluate()
736
+ max_eval_samples = (
737
+ data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
738
+ )
739
+ metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
740
+
741
+ trainer.log_metrics("eval", metrics)
742
+ trainer.save_metrics("eval", metrics)
743
+
744
+ # Write model card and (optionally) push to hub
745
+ config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
746
+ kwargs = {
747
+ "finetuned_from": model_args.model_name_or_path,
748
+ "tasks": "speech-recognition",
749
+ "tags": ["automatic-speech-recognition", data_args.dataset_name],
750
+ "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
751
+ "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
752
+ }
753
+ if "common_voice" in data_args.dataset_name:
754
+ kwargs["language"] = config_name
755
+
756
+ if training_args.push_to_hub:
757
+ trainer.push_to_hub(**kwargs)
758
+ else:
759
+ trainer.create_model_card(**kwargs)
760
+
761
+ return results
762
+
763
+
764
+ if __name__ == "__main__":
765
+ main()
added_tokens.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"<s>": 64, "</s>": 65}
config.json ADDED
@@ -0,0 +1,107 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "facebook/wav2vec2-xls-r-1b",
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": 1024,
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": 1280,
58
+ "initializer_range": 0.02,
59
+ "intermediate_size": 5120,
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": 48,
77
+ "num_negatives": 100,
78
+ "output_hidden_size": 1280,
79
+ "pad_token_id": 63,
80
+ "proj_codevector_dim": 1024,
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": "4.16.0.dev0",
104
+ "use_weighted_layer_sum": false,
105
+ "vocab_size": 66,
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:b7f40567db47ff1bd7f405615252671059a9080ca2dc89787098fb7b42ec5bf8
3
+ size 3850650929
run.sh ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ python run_speech_recognition_ctc.py \
2
+ --dataset_name="mozilla-foundation/common_voice_7_0" \
3
+ --model_name_or_path="facebook/wav2vec2-xls-r-1b" \
4
+ --dataset_config_name="hi" \
5
+ --output_dir="./" \
6
+ --overwrite_output_dir \
7
+ --num_train_epochs="100" \
8
+ --per_device_train_batch_size="8" \
9
+ --per_device_eval_batch_size="16" \
10
+ --gradient_accumulation_steps="4" \
11
+ --learning_rate="7.5e-5" \
12
+ --warmup_steps="2000" \
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
+ --push_to_hub
run_speech_recognition_ctc.py ADDED
@@ -0,0 +1,765 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 warnings
25
+ import unicodedata
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 datasets import DatasetDict, load_dataset, load_metric
33
+
34
+ import bitsandbytes as bnb
35
+ import transformers
36
+ from transformers import (
37
+ AutoConfig,
38
+ AutoFeatureExtractor,
39
+ AutoModelForCTC,
40
+ AutoProcessor,
41
+ AutoTokenizer,
42
+ HfArgumentParser,
43
+ Trainer,
44
+ TrainingArguments,
45
+ Wav2Vec2Processor,
46
+ set_seed,
47
+ )
48
+ from transformers.trainer_pt_utils import get_parameter_names
49
+ from transformers.trainer_utils import get_last_checkpoint, is_main_process
50
+ from transformers.utils import check_min_version
51
+ from transformers.utils.versions import require_version
52
+
53
+
54
+ # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
55
+ check_min_version("4.16.0.dev0")
56
+
57
+ require_version("datasets>=1.13.3", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
58
+
59
+
60
+ logger = logging.getLogger(__name__)
61
+
62
+
63
+ def list_field(default=None, metadata=None):
64
+ return field(default_factory=lambda: default, metadata=metadata)
65
+
66
+
67
+ @dataclass
68
+ class ModelArguments:
69
+ """
70
+ Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
71
+ """
72
+
73
+ model_name_or_path: str = field(
74
+ metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
75
+ )
76
+ tokenizer_name_or_path: Optional[str] = field(
77
+ default=None,
78
+ metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
79
+ )
80
+ cache_dir: Optional[str] = field(
81
+ default=None,
82
+ metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
83
+ )
84
+ freeze_feature_encoder: bool = field(
85
+ default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
86
+ )
87
+ attention_dropout: float = field(
88
+ default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
89
+ )
90
+ activation_dropout: float = field(
91
+ default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
92
+ )
93
+ feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
94
+ hidden_dropout: float = field(
95
+ default=0.0,
96
+ metadata={
97
+ "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
98
+ },
99
+ )
100
+ final_dropout: float = field(
101
+ default=0.0,
102
+ metadata={"help": "The dropout probability for the final projection layer."},
103
+ )
104
+ mask_time_prob: float = field(
105
+ default=0.05,
106
+ metadata={
107
+ "help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
108
+ "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
109
+ "vectors will be masked along the time axis."
110
+ },
111
+ )
112
+ mask_time_length: int = field(
113
+ default=10,
114
+ metadata={"help": "Length of vector span to mask along the time axis."},
115
+ )
116
+ mask_feature_prob: float = field(
117
+ default=0.0,
118
+ metadata={
119
+ "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
120
+ "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
121
+ },
122
+ )
123
+ mask_feature_length: int = field(
124
+ default=10,
125
+ metadata={"help": "Length of vector span to mask along the feature axis."},
126
+ )
127
+ layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
128
+ ctc_loss_reduction: Optional[str] = field(
129
+ default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
130
+ )
131
+
132
+
133
+ @dataclass
134
+ class DataTrainingArguments:
135
+ """
136
+ Arguments pertaining to what data we are going to input our model for training and eval.
137
+
138
+ Using `HfArgumentParser` we can turn this class
139
+ into argparse arguments to be able to specify them on
140
+ the command line.
141
+ """
142
+
143
+ dataset_name: str = field(
144
+ metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
145
+ )
146
+ dataset_config_name: str = field(
147
+ default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
148
+ )
149
+ train_split_name: str = field(
150
+ default="train+validation",
151
+ metadata={
152
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
153
+ },
154
+ )
155
+ eval_split_name: str = field(
156
+ default="test",
157
+ metadata={
158
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
159
+ },
160
+ )
161
+ audio_column_name: str = field(
162
+ default="audio",
163
+ metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
164
+ )
165
+ text_column_name: str = field(
166
+ default="text",
167
+ metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
168
+ )
169
+ overwrite_cache: bool = field(
170
+ default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
171
+ )
172
+ preprocessing_num_workers: Optional[int] = field(
173
+ default=None,
174
+ metadata={"help": "The number of processes to use for the preprocessing."},
175
+ )
176
+ max_train_samples: Optional[int] = field(
177
+ default=None,
178
+ metadata={
179
+ "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
180
+ "value if set."
181
+ },
182
+ )
183
+ max_eval_samples: Optional[int] = field(
184
+ default=None,
185
+ metadata={
186
+ "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
187
+ "value if set."
188
+ },
189
+ )
190
+ chars_to_ignore: Optional[List[str]] = list_field(
191
+ default=None,
192
+ metadata={"help": "A list of characters to remove from the transcripts."},
193
+ )
194
+ eval_metrics: List[str] = list_field(
195
+ default=["wer"],
196
+ metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
197
+ )
198
+ max_duration_in_seconds: float = field(
199
+ default=20.0,
200
+ metadata={
201
+ "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
202
+ },
203
+ )
204
+ min_duration_in_seconds: float = field(
205
+ default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
206
+ )
207
+ preprocessing_only: bool = field(
208
+ default=False,
209
+ metadata={
210
+ "help": "Whether to only do data preprocessing and skip training. "
211
+ "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
212
+ "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
213
+ "so that the cached datasets can consequently be loaded in distributed training"
214
+ },
215
+ )
216
+ use_auth_token: bool = field(
217
+ default=False,
218
+ metadata={
219
+ "help": "If :obj:`True`, will use the token generated when running"
220
+ ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
221
+ },
222
+ )
223
+ unk_token: str = field(
224
+ default="[UNK]",
225
+ metadata={"help": "The unk token for the tokenizer"},
226
+ )
227
+ pad_token: str = field(
228
+ default="[PAD]",
229
+ metadata={"help": "The padding token for the tokenizer"},
230
+ )
231
+ word_delimiter_token: str = field(
232
+ default="|",
233
+ metadata={"help": "The word delimiter token for the tokenizer"},
234
+ )
235
+ phoneme_language: Optional[str] = field(
236
+ default=None,
237
+ metadata={
238
+ "help": "The target language that should be used be"
239
+ " passed to the tokenizer for tokenization. Note that"
240
+ " this is only relevant if the model classifies the"
241
+ " input audio to a sequence of phoneme sequences."
242
+ },
243
+ )
244
+
245
+
246
+ @dataclass
247
+ class DataCollatorCTCWithPadding:
248
+ """
249
+ Data collator that will dynamically pad the inputs received.
250
+ Args:
251
+ processor (:class:`~transformers.AutoProcessor`)
252
+ The processor used for proccessing the data.
253
+ padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
254
+ Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
255
+ among:
256
+ * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
257
+ sequence if provided).
258
+ * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
259
+ maximum acceptable input length for the model if that argument is not provided.
260
+ * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
261
+ different lengths).
262
+ max_length (:obj:`int`, `optional`):
263
+ Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
264
+ max_length_labels (:obj:`int`, `optional`):
265
+ Maximum length of the ``labels`` returned list and optionally padding length (see above).
266
+ pad_to_multiple_of (:obj:`int`, `optional`):
267
+ If set will pad the sequence to a multiple of the provided value.
268
+ This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
269
+ 7.5 (Volta).
270
+ """
271
+
272
+ processor: AutoProcessor
273
+ padding: Union[bool, str] = "longest"
274
+ pad_to_multiple_of: Optional[int] = None
275
+ pad_to_multiple_of_labels: Optional[int] = None
276
+
277
+ def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
278
+ # split inputs and labels since they have to be of different lenghts and need
279
+ # different padding methods
280
+ input_features = [{"input_values": feature["input_values"]} for feature in features]
281
+ label_features = [{"input_ids": feature["labels"]} for feature in features]
282
+
283
+ batch = self.processor.pad(
284
+ input_features,
285
+ padding=self.padding,
286
+ pad_to_multiple_of=self.pad_to_multiple_of,
287
+ return_tensors="pt",
288
+ )
289
+
290
+ with self.processor.as_target_processor():
291
+ labels_batch = self.processor.pad(
292
+ label_features,
293
+ padding=self.padding,
294
+ pad_to_multiple_of=self.pad_to_multiple_of_labels,
295
+ return_tensors="pt",
296
+ )
297
+
298
+ # replace padding with -100 to ignore loss correctly
299
+ labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
300
+
301
+ batch["labels"] = labels
302
+
303
+ return batch
304
+
305
+
306
+ def create_vocabulary_from_data(
307
+ datasets: DatasetDict,
308
+ word_delimiter_token: Optional[str] = None,
309
+ unk_token: Optional[str] = None,
310
+ pad_token: Optional[str] = None,
311
+ ):
312
+ # Given training and test labels create vocabulary
313
+ def extract_all_chars(batch):
314
+ all_text = " ".join(batch["target_text"])
315
+ vocab = list(set(all_text))
316
+ return {"vocab": [vocab], "all_text": [all_text]}
317
+
318
+ vocabs = datasets.map(
319
+ extract_all_chars,
320
+ batched=True,
321
+ batch_size=-1,
322
+ keep_in_memory=True,
323
+ remove_columns=datasets["train"].column_names,
324
+ )
325
+
326
+ # take union of all unique characters in each dataset
327
+ vocab_set = functools.reduce(
328
+ lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
329
+ )
330
+
331
+ vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
332
+
333
+ # replace white space with delimiter token
334
+ if word_delimiter_token is not None:
335
+ vocab_dict[word_delimiter_token] = vocab_dict[" "]
336
+ del vocab_dict[" "]
337
+
338
+ # add unk and pad token
339
+ if unk_token is not None:
340
+ vocab_dict[unk_token] = len(vocab_dict)
341
+
342
+ if pad_token is not None:
343
+ vocab_dict[pad_token] = len(vocab_dict)
344
+
345
+ return vocab_dict
346
+
347
+
348
+ def main():
349
+ # See all possible arguments in src/transformers/training_args.py
350
+ # or by passing the --help flag to this script.
351
+ # We now keep distinct sets of args, for a cleaner separation of concerns.
352
+
353
+ parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
354
+ if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
355
+ # If we pass only one argument to the script and it's the path to a json file,
356
+ # let's parse it to get our arguments.
357
+ model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
358
+ else:
359
+ model_args, data_args, training_args = parser.parse_args_into_dataclasses()
360
+
361
+ # Detecting last checkpoint.
362
+ last_checkpoint = None
363
+ if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
364
+ last_checkpoint = get_last_checkpoint(training_args.output_dir)
365
+ if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
366
+ raise ValueError(
367
+ f"Output directory ({training_args.output_dir}) already exists and is not empty. "
368
+ "Use --overwrite_output_dir to overcome."
369
+ )
370
+ elif last_checkpoint is not None:
371
+ logger.info(
372
+ f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
373
+ "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
374
+ )
375
+
376
+ # Setup logging
377
+ logging.basicConfig(
378
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
379
+ datefmt="%m/%d/%Y %H:%M:%S",
380
+ handlers=[logging.StreamHandler(sys.stdout)],
381
+ )
382
+ logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
383
+
384
+ # Log on each process the small summary:
385
+ logger.warning(
386
+ f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
387
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
388
+ )
389
+ # Set the verbosity to info of the Transformers logger (on main process only):
390
+ if is_main_process(training_args.local_rank):
391
+ transformers.utils.logging.set_verbosity_info()
392
+ logger.info("Training/evaluation parameters %s", training_args)
393
+
394
+ # Set seed before initializing model.
395
+ set_seed(training_args.seed)
396
+
397
+ # 1. First, let's load the dataset
398
+ raw_datasets = DatasetDict()
399
+
400
+ if training_args.do_train:
401
+ raw_datasets["train"] = load_dataset(
402
+ data_args.dataset_name,
403
+ data_args.dataset_config_name,
404
+ split=data_args.train_split_name,
405
+ use_auth_token=data_args.use_auth_token,
406
+ )
407
+
408
+ if data_args.audio_column_name not in raw_datasets["train"].column_names:
409
+ raise ValueError(
410
+ f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
411
+ "Make sure to set `--audio_column_name` to the correct audio column - one of "
412
+ f"{', '.join(raw_datasets['train'].column_names)}."
413
+ )
414
+
415
+ if data_args.text_column_name not in raw_datasets["train"].column_names:
416
+ raise ValueError(
417
+ f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
418
+ "Make sure to set `--text_column_name` to the correct text column - one of "
419
+ f"{', '.join(raw_datasets['train'].column_names)}."
420
+ )
421
+
422
+ if data_args.max_train_samples is not None:
423
+ raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
424
+
425
+ if training_args.do_eval:
426
+ raw_datasets["eval"] = load_dataset(
427
+ data_args.dataset_name,
428
+ data_args.dataset_config_name,
429
+ split=data_args.eval_split_name,
430
+ use_auth_token=data_args.use_auth_token,
431
+ )
432
+
433
+ if data_args.max_eval_samples is not None:
434
+ raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
435
+
436
+ # 2. We remove some special characters from the datasets
437
+ # that make training complicated and do not help in transcribing the speech
438
+ # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
439
+ # that could be easily picked up by the model
440
+ chars_to_ignore_regex = (
441
+ f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
442
+ )
443
+ text_column_name = data_args.text_column_name
444
+
445
+ def remove_special_characters(batch):
446
+ if chars_to_ignore_regex is not None:
447
+ batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower()
448
+ else:
449
+ batch["target_text"] = batch[text_column_name].lower()
450
+ # Remove english characters
451
+ batch["target_text"] = re.sub(r'[a-z]', '', batch["target_text"])
452
+ # Unicode Normalization
453
+ batch["target_text"] = unicodedata.normalize('NFKC', batch["target_text"])
454
+ return batch
455
+
456
+ with training_args.main_process_first(desc="dataset map special characters removal"):
457
+ raw_datasets = raw_datasets.map(
458
+ remove_special_characters,
459
+ remove_columns=[text_column_name],
460
+ desc="remove special characters from datasets",
461
+ )
462
+
463
+ # save special tokens for tokenizer
464
+ word_delimiter_token = data_args.word_delimiter_token
465
+ unk_token = data_args.unk_token
466
+ pad_token = data_args.pad_token
467
+
468
+ # 3. Next, let's load the config as we might need it to create
469
+ # the tokenizer
470
+ # load config
471
+ config = AutoConfig.from_pretrained(
472
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
473
+ )
474
+
475
+ # 4. Next, if no tokenizer file is defined,
476
+ # we create the vocabulary of the model by extracting all unique characters from
477
+ # the training and evaluation datasets
478
+ # We need to make sure that only first rank saves vocabulary
479
+ # make sure all processes wait until vocab is created
480
+ tokenizer_name_or_path = model_args.tokenizer_name_or_path
481
+ tokenizer_kwargs = {}
482
+ if tokenizer_name_or_path is None:
483
+ # save vocab in training output dir
484
+ tokenizer_name_or_path = training_args.output_dir
485
+
486
+ vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
487
+
488
+ with training_args.main_process_first():
489
+ if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
490
+ os.remove(vocab_file)
491
+
492
+ with training_args.main_process_first(desc="dataset map vocabulary creation"):
493
+ if not os.path.isfile(vocab_file):
494
+ os.makedirs(tokenizer_name_or_path, exist_ok=True)
495
+ vocab_dict = create_vocabulary_from_data(
496
+ raw_datasets,
497
+ word_delimiter_token=word_delimiter_token,
498
+ unk_token=unk_token,
499
+ pad_token=pad_token,
500
+ )
501
+
502
+ # save vocab dict to be loaded into tokenizer
503
+ with open(vocab_file, "w") as file:
504
+ json.dump(vocab_dict, file)
505
+
506
+ # if tokenizer has just been created
507
+ # it is defined by `tokenizer_class` if present in config else by `model_type`
508
+ tokenizer_kwargs = {
509
+ "config": config if config.tokenizer_class is not None else None,
510
+ "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
511
+ "unk_token": unk_token,
512
+ "pad_token": pad_token,
513
+ "word_delimiter_token": word_delimiter_token,
514
+ }
515
+
516
+ # 5. Now we can instantiate the feature extractor, tokenizer and model
517
+ # Note for distributed training, the .from_pretrained methods guarantee that only
518
+ # one local process can concurrently download model & vocab.
519
+
520
+ # load feature_extractor and tokenizer
521
+ tokenizer = AutoTokenizer.from_pretrained(
522
+ tokenizer_name_or_path,
523
+ use_auth_token=data_args.use_auth_token,
524
+ **tokenizer_kwargs,
525
+ )
526
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
527
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
528
+ )
529
+
530
+ # adapt config
531
+ config.update(
532
+ {
533
+ "feat_proj_dropout": model_args.feat_proj_dropout,
534
+ "attention_dropout": model_args.attention_dropout,
535
+ "hidden_dropout": model_args.hidden_dropout,
536
+ "final_dropout": model_args.final_dropout,
537
+ "mask_time_prob": model_args.mask_time_prob,
538
+ "mask_time_length": model_args.mask_time_length,
539
+ "mask_feature_prob": model_args.mask_feature_prob,
540
+ "mask_feature_length": model_args.mask_feature_length,
541
+ "gradient_checkpointing": training_args.gradient_checkpointing,
542
+ "layerdrop": model_args.layerdrop,
543
+ "ctc_loss_reduction": model_args.ctc_loss_reduction,
544
+ "pad_token_id": tokenizer.pad_token_id,
545
+ "vocab_size": len(tokenizer),
546
+ "activation_dropout": model_args.activation_dropout,
547
+ }
548
+ )
549
+
550
+ # create model
551
+ model = AutoModelForCTC.from_pretrained(
552
+ model_args.model_name_or_path,
553
+ cache_dir=model_args.cache_dir,
554
+ config=config,
555
+ use_auth_token=data_args.use_auth_token,
556
+ )
557
+
558
+ # freeze encoder
559
+ if model_args.freeze_feature_encoder:
560
+ model.freeze_feature_encoder()
561
+
562
+ # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
563
+ # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
564
+ # so that we just need to set the correct target sampling rate and normalize the input
565
+ # via the `feature_extractor`
566
+
567
+ # make sure that dataset decodes audio with correct sampling rate
568
+ dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
569
+ if dataset_sampling_rate != feature_extractor.sampling_rate:
570
+ raw_datasets = raw_datasets.cast_column(
571
+ data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
572
+ )
573
+
574
+ # derive max & min input length for sample rate & max duration
575
+ max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
576
+ min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
577
+ audio_column_name = data_args.audio_column_name
578
+ num_workers = data_args.preprocessing_num_workers
579
+
580
+ # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
581
+ phoneme_language = data_args.phoneme_language
582
+
583
+ # Preprocessing the datasets.
584
+ # We need to read the audio files as arrays and tokenize the targets.
585
+ def prepare_dataset(batch):
586
+ # load audio
587
+ sample = batch[audio_column_name]
588
+
589
+ inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
590
+ batch["input_values"] = inputs.input_values[0]
591
+ batch["input_length"] = len(batch["input_values"])
592
+
593
+ # encode targets
594
+ additional_kwargs = {}
595
+ if phoneme_language is not None:
596
+ additional_kwargs["phonemizer_lang"] = phoneme_language
597
+
598
+ batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
599
+ return batch
600
+
601
+ with training_args.main_process_first(desc="dataset map preprocessing"):
602
+ vectorized_datasets = raw_datasets.map(
603
+ prepare_dataset,
604
+ remove_columns=next(iter(raw_datasets.values())).column_names,
605
+ num_proc=num_workers,
606
+ desc="preprocess datasets",
607
+ )
608
+
609
+ def is_audio_in_length_range(length):
610
+ return length > min_input_length and length < max_input_length
611
+
612
+ # filter data that is shorter than min_input_length
613
+ vectorized_datasets = vectorized_datasets.filter(
614
+ is_audio_in_length_range,
615
+ num_proc=num_workers,
616
+ input_columns=["input_length"],
617
+ )
618
+
619
+ # 7. Next, we can prepare the training.
620
+ # Let's use word error rate (WER) as our evaluation metric,
621
+ # instantiate a data collator and the trainer
622
+
623
+ # Define evaluation metrics during training, *i.e.* word error rate, character error rate
624
+ eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
625
+
626
+ # for large datasets it is advised to run the preprocessing on a
627
+ # single machine first with ``args.preprocessing_only`` since there will mostly likely
628
+ # be a timeout when running the script in distributed mode.
629
+ # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
630
+ # cached dataset
631
+ if data_args.preprocessing_only:
632
+ logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
633
+ return
634
+
635
+ def compute_metrics(pred):
636
+ pred_logits = pred.predictions
637
+ pred_ids = np.argmax(pred_logits, axis=-1)
638
+
639
+ pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
640
+
641
+ pred_str = tokenizer.batch_decode(pred_ids)
642
+ # we do not want to group tokens when computing the metrics
643
+ label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
644
+
645
+ metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
646
+
647
+ return metrics
648
+
649
+ # Now save everything to be able to create a single processor later
650
+ if is_main_process(training_args.local_rank):
651
+ # save feature extractor, tokenizer and config
652
+ feature_extractor.save_pretrained(training_args.output_dir)
653
+ tokenizer.save_pretrained(training_args.output_dir)
654
+ config.save_pretrained(training_args.output_dir)
655
+
656
+ try:
657
+ processor = AutoProcessor.from_pretrained(training_args.output_dir)
658
+ except (OSError, KeyError):
659
+ warnings.warn(
660
+ "Loading a processor from a feature extractor config that does not"
661
+ " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
662
+ " attribute to your `preprocessor_config.json` file to suppress this warning: "
663
+ " `'processor_class': 'Wav2Vec2Processor'`",
664
+ FutureWarning,
665
+ )
666
+ processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
667
+
668
+ # Instantiate custom data collator
669
+ data_collator = DataCollatorCTCWithPadding(processor=processor)
670
+
671
+ # decay_parameters = get_parameter_names(model, [torch.nn.LayerNorm])
672
+ # decay_parameters = [name for name in decay_parameters if "bias" not in name]
673
+ # optimizer_grouped_parameters = [
674
+ # {
675
+ # "params": [p for n, p in model.named_parameters() if n in decay_parameters],
676
+ # "weight_decay": training_args.weight_decay,
677
+ # },
678
+ # {
679
+ # "params": [p for n, p in model.named_parameters() if n not in decay_parameters],
680
+ # "weight_decay": 0.0,
681
+ # },
682
+ # ]
683
+ # optimizer = bnb.optim.Adam8bit(
684
+ # params=optimizer_grouped_parameters,
685
+ # betas=(training_args.adam_beta1, training_args.adam_beta2),
686
+ # eps=training_args.adam_epsilon,
687
+ # )
688
+
689
+ # optimizers = (optimizer, None)
690
+
691
+ # Initialize Trainer
692
+ trainer = Trainer(
693
+ model=model,
694
+ data_collator=data_collator,
695
+ args=training_args,
696
+ compute_metrics=compute_metrics,
697
+ train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
698
+ eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
699
+ tokenizer=feature_extractor,
700
+ # optimizers=optimizers,
701
+ )
702
+
703
+ # 8. Finally, we can start training
704
+
705
+ # Training
706
+ if training_args.do_train:
707
+
708
+ # use last checkpoint if exist
709
+ if last_checkpoint is not None:
710
+ checkpoint = last_checkpoint
711
+ elif os.path.isdir(model_args.model_name_or_path):
712
+ checkpoint = model_args.model_name_or_path
713
+ else:
714
+ checkpoint = None
715
+
716
+ train_result = trainer.train(resume_from_checkpoint=checkpoint)
717
+ trainer.save_model()
718
+
719
+ metrics = train_result.metrics
720
+ max_train_samples = (
721
+ data_args.max_train_samples
722
+ if data_args.max_train_samples is not None
723
+ else len(vectorized_datasets["train"])
724
+ )
725
+ metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
726
+
727
+ trainer.log_metrics("train", metrics)
728
+ trainer.save_metrics("train", metrics)
729
+ trainer.save_state()
730
+
731
+ # Evaluation
732
+ results = {}
733
+ if training_args.do_eval:
734
+ logger.info("*** Evaluate ***")
735
+ metrics = trainer.evaluate()
736
+ max_eval_samples = (
737
+ data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
738
+ )
739
+ metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
740
+
741
+ trainer.log_metrics("eval", metrics)
742
+ trainer.save_metrics("eval", metrics)
743
+
744
+ # Write model card and (optionally) push to hub
745
+ config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
746
+ kwargs = {
747
+ "finetuned_from": model_args.model_name_or_path,
748
+ "tasks": "speech-recognition",
749
+ "tags": ["automatic-speech-recognition", data_args.dataset_name],
750
+ "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
751
+ "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
752
+ }
753
+ if "common_voice" in data_args.dataset_name:
754
+ kwargs["language"] = config_name
755
+
756
+ if training_args.push_to_hub:
757
+ trainer.push_to_hub(**kwargs)
758
+ else:
759
+ trainer.create_model_card(**kwargs)
760
+
761
+ return results
762
+
763
+
764
+ if __name__ == "__main__":
765
+ 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}, {"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"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": "|", "special_tokens_map_file": null, "tokenizer_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 sha256:62db44acc41bdd822bfe0559ab15f498d5f04d4e0479bc655a0c230ee0c788c9
3
+ size 2991
vocab.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"ँ": 1, "ं": 2, "ः": 3, "अ": 4, "आ": 5, "इ": 6, "ई": 7, "उ": 8, "ऊ": 9, "ऋ": 10, "ए": 11, "ऐ": 12, "ऑ": 13, "ओ": 14, "औ": 15, "क": 16, "ख": 17, "ग": 18, "घ": 19, "च": 20, "छ": 21, "ज": 22, "झ": 23, "ञ": 24, "ट": 25, "ठ": 26, "ड": 27, "ढ": 28, "ण": 29, "त": 30, "थ": 31, "द": 32, "ध": 33, "न": 34, "प": 35, "फ": 36, "ब": 37, "भ": 38, "म": 39, "य": 40, "र": 41, "ल": 42, "व": 43, "श": 44, "ष": 45, "स": 46, "ह": 47, "़": 48, "ा": 49, "ि": 50, "ी": 51, "ु": 52, "ू": 53, "ृ": 54, "ॅ": 55, "े": 56, "ै": 57, "ॉ": 58, "ो": 59, "ौ": 60, "्": 61, "|": 0, "[UNK]": 62, "[PAD]": 63}