g8a9 commited on
Commit
756d60d
1 Parent(s): f4b9bb8

Add runner script and bash

Browse files
run_speech_recognition_ctc_bnb.py ADDED
@@ -0,0 +1,764 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ from dotenv import load_dotenv
19
+ load_dotenv() # take environment variables from .env.
20
+
21
+ import comet_ml
22
+ import functools
23
+ import json
24
+ import logging
25
+ import os
26
+ import re
27
+ import sys
28
+ import warnings
29
+ from dataclasses import dataclass, field
30
+ from typing import Dict, List, Optional, Union
31
+
32
+ import datasets
33
+ import numpy as np
34
+ import torch
35
+ from datasets import DatasetDict, load_dataset, load_metric
36
+
37
+ import bitsandbytes as bnb
38
+ import transformers
39
+ from transformers import (
40
+ AutoConfig,
41
+ AutoFeatureExtractor,
42
+ AutoModelForCTC,
43
+ AutoProcessor,
44
+ AutoTokenizer,
45
+ HfArgumentParser,
46
+ Trainer,
47
+ TrainingArguments,
48
+ Wav2Vec2Processor,
49
+ set_seed,
50
+ )
51
+ from transformers.trainer_pt_utils import get_parameter_names
52
+ from transformers.trainer_utils import get_last_checkpoint, is_main_process
53
+ from transformers.utils import check_min_version
54
+ from transformers.utils.versions import require_version
55
+
56
+
57
+ # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
58
+ check_min_version("4.16.0.dev0")
59
+
60
+ require_version("datasets>=1.13.3", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
61
+
62
+
63
+ logger = logging.getLogger(__name__)
64
+
65
+
66
+ def list_field(default=None, metadata=None):
67
+ return field(default_factory=lambda: default, metadata=metadata)
68
+
69
+
70
+ @dataclass
71
+ class ModelArguments:
72
+ """
73
+ Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
74
+ """
75
+
76
+ model_name_or_path: str = field(
77
+ metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
78
+ )
79
+ tokenizer_name_or_path: Optional[str] = field(
80
+ default=None,
81
+ metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
82
+ )
83
+ cache_dir: Optional[str] = field(
84
+ default=None,
85
+ metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
86
+ )
87
+ freeze_feature_encoder: bool = field(
88
+ default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
89
+ )
90
+ attention_dropout: float = field(
91
+ default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
92
+ )
93
+ activation_dropout: float = field(
94
+ default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
95
+ )
96
+ feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
97
+ hidden_dropout: float = field(
98
+ default=0.0,
99
+ metadata={
100
+ "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
101
+ },
102
+ )
103
+ final_dropout: float = field(
104
+ default=0.0,
105
+ metadata={"help": "The dropout probability for the final projection layer."},
106
+ )
107
+ mask_time_prob: float = field(
108
+ default=0.05,
109
+ metadata={
110
+ "help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
111
+ "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
112
+ "vectors will be masked along the time axis."
113
+ },
114
+ )
115
+ mask_time_length: int = field(
116
+ default=10,
117
+ metadata={"help": "Length of vector span to mask along the time axis."},
118
+ )
119
+ mask_feature_prob: float = field(
120
+ default=0.0,
121
+ metadata={
122
+ "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
123
+ "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
124
+ },
125
+ )
126
+ mask_feature_length: int = field(
127
+ default=10,
128
+ metadata={"help": "Length of vector span to mask along the feature axis."},
129
+ )
130
+ layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
131
+ ctc_loss_reduction: Optional[str] = field(
132
+ default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
133
+ )
134
+
135
+
136
+ @dataclass
137
+ class DataTrainingArguments:
138
+ """
139
+ Arguments pertaining to what data we are going to input our model for training and eval.
140
+
141
+ Using `HfArgumentParser` we can turn this class
142
+ into argparse arguments to be able to specify them on
143
+ the command line.
144
+ """
145
+
146
+ dataset_name: str = field(
147
+ metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
148
+ )
149
+ dataset_config_name: str = field(
150
+ default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
151
+ )
152
+ train_split_name: str = field(
153
+ default="train+validation",
154
+ metadata={
155
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
156
+ },
157
+ )
158
+ eval_split_name: str = field(
159
+ default="test",
160
+ metadata={
161
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
162
+ },
163
+ )
164
+ audio_column_name: str = field(
165
+ default="audio",
166
+ metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
167
+ )
168
+ text_column_name: str = field(
169
+ default="text",
170
+ metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
171
+ )
172
+ overwrite_cache: bool = field(
173
+ default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
174
+ )
175
+ preprocessing_num_workers: Optional[int] = field(
176
+ default=None,
177
+ metadata={"help": "The number of processes to use for the preprocessing."},
178
+ )
179
+ max_train_samples: Optional[int] = field(
180
+ default=None,
181
+ metadata={
182
+ "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
183
+ "value if set."
184
+ },
185
+ )
186
+ max_eval_samples: Optional[int] = field(
187
+ default=None,
188
+ metadata={
189
+ "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
190
+ "value if set."
191
+ },
192
+ )
193
+ chars_to_ignore: Optional[List[str]] = list_field(
194
+ default=None,
195
+ metadata={"help": "A list of characters to remove from the transcripts."},
196
+ )
197
+ eval_metrics: List[str] = list_field(
198
+ default=["wer"],
199
+ metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
200
+ )
201
+ max_duration_in_seconds: float = field(
202
+ default=20.0,
203
+ metadata={
204
+ "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
205
+ },
206
+ )
207
+ min_duration_in_seconds: float = field(
208
+ default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
209
+ )
210
+ preprocessing_only: bool = field(
211
+ default=False,
212
+ metadata={
213
+ "help": "Whether to only do data preprocessing and skip training. "
214
+ "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
215
+ "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
216
+ "so that the cached datasets can consequently be loaded in distributed training"
217
+ },
218
+ )
219
+ use_auth_token: bool = field(
220
+ default=False,
221
+ metadata={
222
+ "help": "If :obj:`True`, will use the token generated when running"
223
+ ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
224
+ },
225
+ )
226
+ unk_token: str = field(
227
+ default="[UNK]",
228
+ metadata={"help": "The unk token for the tokenizer"},
229
+ )
230
+ pad_token: str = field(
231
+ default="[PAD]",
232
+ metadata={"help": "The padding token for the tokenizer"},
233
+ )
234
+ word_delimiter_token: str = field(
235
+ default="|",
236
+ metadata={"help": "The word delimiter token for the tokenizer"},
237
+ )
238
+ phoneme_language: Optional[str] = field(
239
+ default=None,
240
+ metadata={
241
+ "help": "The target language that should be used be"
242
+ " passed to the tokenizer for tokenization. Note that"
243
+ " this is only relevant if the model classifies the"
244
+ " input audio to a sequence of phoneme sequences."
245
+ },
246
+ )
247
+
248
+
249
+ @dataclass
250
+ class DataCollatorCTCWithPadding:
251
+ """
252
+ Data collator that will dynamically pad the inputs received.
253
+ Args:
254
+ processor (:class:`~transformers.AutoProcessor`)
255
+ The processor used for proccessing the data.
256
+ padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
257
+ Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
258
+ among:
259
+ * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
260
+ sequence if provided).
261
+ * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
262
+ maximum acceptable input length for the model if that argument is not provided.
263
+ * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
264
+ different lengths).
265
+ max_length (:obj:`int`, `optional`):
266
+ Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
267
+ max_length_labels (:obj:`int`, `optional`):
268
+ Maximum length of the ``labels`` returned list and optionally padding length (see above).
269
+ pad_to_multiple_of (:obj:`int`, `optional`):
270
+ If set will pad the sequence to a multiple of the provided value.
271
+ This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
272
+ 7.5 (Volta).
273
+ """
274
+
275
+ processor: AutoProcessor
276
+ padding: Union[bool, str] = "longest"
277
+ pad_to_multiple_of: Optional[int] = None
278
+ pad_to_multiple_of_labels: Optional[int] = None
279
+
280
+ def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
281
+ # split inputs and labels since they have to be of different lenghts and need
282
+ # different padding methods
283
+ input_features = [{"input_values": feature["input_values"]} for feature in features]
284
+ label_features = [{"input_ids": feature["labels"]} for feature in features]
285
+
286
+ batch = self.processor.pad(
287
+ input_features,
288
+ padding=self.padding,
289
+ pad_to_multiple_of=self.pad_to_multiple_of,
290
+ return_tensors="pt",
291
+ )
292
+
293
+ with self.processor.as_target_processor():
294
+ labels_batch = self.processor.pad(
295
+ label_features,
296
+ padding=self.padding,
297
+ pad_to_multiple_of=self.pad_to_multiple_of_labels,
298
+ return_tensors="pt",
299
+ )
300
+
301
+ # replace padding with -100 to ignore loss correctly
302
+ labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
303
+
304
+ batch["labels"] = labels
305
+
306
+ return batch
307
+
308
+
309
+ def create_vocabulary_from_data(
310
+ datasets: DatasetDict,
311
+ word_delimiter_token: Optional[str] = None,
312
+ unk_token: Optional[str] = None,
313
+ pad_token: Optional[str] = None,
314
+ ):
315
+ # Given training and test labels create vocabulary
316
+ def extract_all_chars(batch):
317
+ all_text = " ".join(batch["target_text"])
318
+ vocab = list(set(all_text))
319
+ return {"vocab": [vocab], "all_text": [all_text]}
320
+
321
+ vocabs = datasets.map(
322
+ extract_all_chars,
323
+ batched=True,
324
+ batch_size=-1,
325
+ keep_in_memory=True,
326
+ remove_columns=datasets["train"].column_names,
327
+ )
328
+
329
+ # take union of all unique characters in each dataset
330
+ vocab_set = functools.reduce(
331
+ lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
332
+ )
333
+
334
+ vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
335
+
336
+ # replace white space with delimiter token
337
+ if word_delimiter_token is not None:
338
+ vocab_dict[word_delimiter_token] = vocab_dict[" "]
339
+ del vocab_dict[" "]
340
+
341
+ # add unk and pad token
342
+ if unk_token is not None:
343
+ vocab_dict[unk_token] = len(vocab_dict)
344
+
345
+ if pad_token is not None:
346
+ vocab_dict[pad_token] = len(vocab_dict)
347
+
348
+ return vocab_dict
349
+
350
+
351
+ def main():
352
+ # See all possible arguments in src/transformers/training_args.py
353
+ # or by passing the --help flag to this script.
354
+ # We now keep distinct sets of args, for a cleaner separation of concerns.
355
+
356
+ parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
357
+ if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
358
+ # If we pass only one argument to the script and it's the path to a json file,
359
+ # let's parse it to get our arguments.
360
+ model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
361
+ else:
362
+ model_args, data_args, training_args = parser.parse_args_into_dataclasses()
363
+
364
+ # Detecting last checkpoint.
365
+ last_checkpoint = None
366
+ if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
367
+ last_checkpoint = get_last_checkpoint(training_args.output_dir)
368
+ if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
369
+ raise ValueError(
370
+ f"Output directory ({training_args.output_dir}) already exists and is not empty. "
371
+ "Use --overwrite_output_dir to overcome."
372
+ )
373
+ elif last_checkpoint is not None:
374
+ logger.info(
375
+ f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
376
+ "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
377
+ )
378
+
379
+ # Setup logging
380
+ logging.basicConfig(
381
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
382
+ datefmt="%m/%d/%Y %H:%M:%S",
383
+ handlers=[logging.StreamHandler(sys.stdout)],
384
+ )
385
+ logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
386
+
387
+ # Log on each process the small summary:
388
+ logger.warning(
389
+ f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
390
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
391
+ )
392
+ # Set the verbosity to info of the Transformers logger (on main process only):
393
+ if is_main_process(training_args.local_rank):
394
+ transformers.utils.logging.set_verbosity_info()
395
+ logger.info("Training/evaluation parameters %s", training_args)
396
+
397
+ # Set seed before initializing model.
398
+ set_seed(training_args.seed)
399
+
400
+ # 1. First, let's load the dataset
401
+ raw_datasets = DatasetDict()
402
+
403
+ if training_args.do_train:
404
+ raw_datasets["train"] = load_dataset(
405
+ data_args.dataset_name,
406
+ data_args.dataset_config_name,
407
+ split=data_args.train_split_name,
408
+ use_auth_token=data_args.use_auth_token,
409
+ )
410
+
411
+ if data_args.audio_column_name not in raw_datasets["train"].column_names:
412
+ raise ValueError(
413
+ f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
414
+ "Make sure to set `--audio_column_name` to the correct audio column - one of "
415
+ f"{', '.join(raw_datasets['train'].column_names)}."
416
+ )
417
+
418
+ if data_args.text_column_name not in raw_datasets["train"].column_names:
419
+ raise ValueError(
420
+ f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
421
+ "Make sure to set `--text_column_name` to the correct text column - one of "
422
+ f"{', '.join(raw_datasets['train'].column_names)}."
423
+ )
424
+
425
+ if data_args.max_train_samples is not None:
426
+ raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
427
+
428
+ if training_args.do_eval:
429
+ raw_datasets["eval"] = load_dataset(
430
+ data_args.dataset_name,
431
+ data_args.dataset_config_name,
432
+ split=data_args.eval_split_name,
433
+ use_auth_token=data_args.use_auth_token,
434
+ )
435
+
436
+ if data_args.max_eval_samples is not None:
437
+ raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
438
+
439
+ # 2. We remove some special characters from the datasets
440
+ # that make training complicated and do not help in transcribing the speech
441
+ # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
442
+ # that could be easily picked up by the model
443
+ chars_to_ignore_regex = (
444
+ f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
445
+ )
446
+ text_column_name = data_args.text_column_name
447
+
448
+ def remove_special_characters(batch):
449
+ if chars_to_ignore_regex is not None:
450
+ batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
451
+ else:
452
+ batch["target_text"] = batch[text_column_name].lower() + " "
453
+ return batch
454
+
455
+ with training_args.main_process_first(desc="dataset map special characters removal"):
456
+ raw_datasets = raw_datasets.map(
457
+ remove_special_characters,
458
+ remove_columns=[text_column_name],
459
+ desc="remove special characters from datasets",
460
+ )
461
+
462
+ # save special tokens for tokenizer
463
+ word_delimiter_token = data_args.word_delimiter_token
464
+ unk_token = data_args.unk_token
465
+ pad_token = data_args.pad_token
466
+
467
+ # 3. Next, let's load the config as we might need it to create
468
+ # the tokenizer
469
+ # load config
470
+ config = AutoConfig.from_pretrained(
471
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
472
+ )
473
+
474
+ # 4. Next, if no tokenizer file is defined,
475
+ # we create the vocabulary of the model by extracting all unique characters from
476
+ # the training and evaluation datasets
477
+ # We need to make sure that only first rank saves vocabulary
478
+ # make sure all processes wait until vocab is created
479
+ tokenizer_name_or_path = model_args.tokenizer_name_or_path
480
+ tokenizer_kwargs = {}
481
+ if tokenizer_name_or_path is None:
482
+ # save vocab in training output dir
483
+ tokenizer_name_or_path = training_args.output_dir
484
+
485
+ vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
486
+
487
+ with training_args.main_process_first():
488
+ if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
489
+ os.remove(vocab_file)
490
+
491
+ with training_args.main_process_first(desc="dataset map vocabulary creation"):
492
+ if not os.path.isfile(vocab_file):
493
+ os.makedirs(tokenizer_name_or_path, exist_ok=True)
494
+ vocab_dict = create_vocabulary_from_data(
495
+ raw_datasets,
496
+ word_delimiter_token=word_delimiter_token,
497
+ unk_token=unk_token,
498
+ pad_token=pad_token,
499
+ )
500
+
501
+ # save vocab dict to be loaded into tokenizer
502
+ with open(vocab_file, "w") as file:
503
+ json.dump(vocab_dict, file)
504
+
505
+ # if tokenizer has just been created
506
+ # it is defined by `tokenizer_class` if present in config else by `model_type`
507
+ tokenizer_kwargs = {
508
+ "config": config if config.tokenizer_class is not None else None,
509
+ "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
510
+ "unk_token": unk_token,
511
+ "pad_token": pad_token,
512
+ "word_delimiter_token": word_delimiter_token,
513
+ }
514
+
515
+ # 5. Now we can instantiate the feature extractor, tokenizer and model
516
+ # Note for distributed training, the .from_pretrained methods guarantee that only
517
+ # one local process can concurrently download model & vocab.
518
+
519
+ # load feature_extractor and tokenizer
520
+ tokenizer = AutoTokenizer.from_pretrained(
521
+ tokenizer_name_or_path,
522
+ use_auth_token=data_args.use_auth_token,
523
+ **tokenizer_kwargs,
524
+ )
525
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
526
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
527
+ )
528
+
529
+ # adapt config
530
+ config.update(
531
+ {
532
+ "feat_proj_dropout": model_args.feat_proj_dropout,
533
+ "attention_dropout": model_args.attention_dropout,
534
+ "hidden_dropout": model_args.hidden_dropout,
535
+ "final_dropout": model_args.final_dropout,
536
+ "mask_time_prob": model_args.mask_time_prob,
537
+ "mask_time_length": model_args.mask_time_length,
538
+ "mask_feature_prob": model_args.mask_feature_prob,
539
+ "mask_feature_length": model_args.mask_feature_length,
540
+ "gradient_checkpointing": training_args.gradient_checkpointing,
541
+ "layerdrop": model_args.layerdrop,
542
+ "ctc_loss_reduction": model_args.ctc_loss_reduction,
543
+ "pad_token_id": tokenizer.pad_token_id,
544
+ "vocab_size": len(tokenizer),
545
+ "activation_dropout": model_args.activation_dropout,
546
+ }
547
+ )
548
+
549
+ # create model
550
+ model = AutoModelForCTC.from_pretrained(
551
+ model_args.model_name_or_path,
552
+ cache_dir=model_args.cache_dir,
553
+ config=config,
554
+ use_auth_token=data_args.use_auth_token,
555
+ )
556
+
557
+ # freeze encoder
558
+ if model_args.freeze_feature_encoder:
559
+ model.freeze_feature_encoder()
560
+
561
+ # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
562
+ # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
563
+ # so that we just need to set the correct target sampling rate and normalize the input
564
+ # via the `feature_extractor`
565
+
566
+ # make sure that dataset decodes audio with correct sampling rate
567
+ dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
568
+ if dataset_sampling_rate != feature_extractor.sampling_rate:
569
+ raw_datasets = raw_datasets.cast_column(
570
+ data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
571
+ )
572
+
573
+ # derive max & min input length for sample rate & max duration
574
+ max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
575
+ min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
576
+ audio_column_name = data_args.audio_column_name
577
+ num_workers = data_args.preprocessing_num_workers
578
+
579
+ # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
580
+ phoneme_language = data_args.phoneme_language
581
+
582
+ # Preprocessing the datasets.
583
+ # We need to read the audio files as arrays and tokenize the targets.
584
+ def prepare_dataset(batch):
585
+ # load audio
586
+ sample = batch[audio_column_name]
587
+
588
+ inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
589
+ batch["input_values"] = inputs.input_values[0]
590
+ batch["input_length"] = len(batch["input_values"])
591
+
592
+ # encode targets
593
+ additional_kwargs = {}
594
+ if phoneme_language is not None:
595
+ additional_kwargs["phonemizer_lang"] = phoneme_language
596
+
597
+ batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
598
+ return batch
599
+
600
+ with training_args.main_process_first(desc="dataset map preprocessing"):
601
+ vectorized_datasets = raw_datasets.map(
602
+ prepare_dataset,
603
+ remove_columns=next(iter(raw_datasets.values())).column_names,
604
+ num_proc=num_workers,
605
+ desc="preprocess datasets",
606
+ )
607
+
608
+ def is_audio_in_length_range(length):
609
+ return length > min_input_length and length < max_input_length
610
+
611
+ # filter data that is shorter than min_input_length
612
+ vectorized_datasets = vectorized_datasets.filter(
613
+ is_audio_in_length_range,
614
+ num_proc=num_workers,
615
+ input_columns=["input_length"],
616
+ )
617
+
618
+ # 7. Next, we can prepare the training.
619
+ # Let's use word error rate (WER) as our evaluation metric,
620
+ # instantiate a data collator and the trainer
621
+
622
+ # Define evaluation metrics during training, *i.e.* word error rate, character error rate
623
+ eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
624
+
625
+ # for large datasets it is advised to run the preprocessing on a
626
+ # single machine first with ``args.preprocessing_only`` since there will mostly likely
627
+ # be a timeout when running the script in distributed mode.
628
+ # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
629
+ # cached dataset
630
+ if data_args.preprocessing_only:
631
+ logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
632
+ return
633
+
634
+ def compute_metrics(pred):
635
+ pred_logits = pred.predictions
636
+ pred_ids = np.argmax(pred_logits, axis=-1)
637
+
638
+ pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
639
+
640
+ pred_str = tokenizer.batch_decode(pred_ids)
641
+ # we do not want to group tokens when computing the metrics
642
+ label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
643
+
644
+ metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
645
+
646
+ return metrics
647
+
648
+ # Now save everything to be able to create a single processor later
649
+ if is_main_process(training_args.local_rank):
650
+ # save feature extractor, tokenizer and config
651
+ feature_extractor.save_pretrained(training_args.output_dir)
652
+ tokenizer.save_pretrained(training_args.output_dir)
653
+ config.save_pretrained(training_args.output_dir)
654
+
655
+ try:
656
+ processor = AutoProcessor.from_pretrained(training_args.output_dir)
657
+ except (OSError, KeyError):
658
+ warnings.warn(
659
+ "Loading a processor from a feature extractor config that does not"
660
+ " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
661
+ " attribute to your `preprocessor_config.json` file to suppress this warning: "
662
+ " `'processor_class': 'Wav2Vec2Processor'`",
663
+ FutureWarning,
664
+ )
665
+ processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
666
+
667
+ # Instantiate custom data collator
668
+ data_collator = DataCollatorCTCWithPadding(processor=processor)
669
+
670
+ decay_parameters = get_parameter_names(model, [torch.nn.LayerNorm])
671
+ decay_parameters = [name for name in decay_parameters if "bias" not in name]
672
+ optimizer_grouped_parameters = [
673
+ {
674
+ "params": [p for n, p in model.named_parameters() if n in decay_parameters],
675
+ "weight_decay": training_args.weight_decay,
676
+ },
677
+ {
678
+ "params": [p for n, p in model.named_parameters() if n not in decay_parameters],
679
+ "weight_decay": 0.0,
680
+ },
681
+ ]
682
+ optimizer = bnb.optim.Adam8bit(
683
+ params=optimizer_grouped_parameters,
684
+ betas=(training_args.adam_beta1, training_args.adam_beta2),
685
+ eps=training_args.adam_epsilon,
686
+ )
687
+
688
+ optimizers = (optimizer, None)
689
+
690
+ # Initialize Trainer
691
+ trainer = Trainer(
692
+ model=model,
693
+ data_collator=data_collator,
694
+ args=training_args,
695
+ compute_metrics=compute_metrics,
696
+ train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
697
+ eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
698
+ tokenizer=feature_extractor,
699
+ optimizers=optimizers,
700
+ )
701
+
702
+ # 8. Finally, we can start training
703
+
704
+ # Training
705
+ if training_args.do_train:
706
+
707
+ # use last checkpoint if exist
708
+ if last_checkpoint is not None:
709
+ checkpoint = last_checkpoint
710
+ elif os.path.isdir(model_args.model_name_or_path):
711
+ checkpoint = model_args.model_name_or_path
712
+ else:
713
+ checkpoint = None
714
+
715
+ train_result = trainer.train(resume_from_checkpoint=checkpoint)
716
+ trainer.save_model()
717
+
718
+ metrics = train_result.metrics
719
+ max_train_samples = (
720
+ data_args.max_train_samples
721
+ if data_args.max_train_samples is not None
722
+ else len(vectorized_datasets["train"])
723
+ )
724
+ metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
725
+
726
+ trainer.log_metrics("train", metrics)
727
+ trainer.save_metrics("train", metrics)
728
+ trainer.save_state()
729
+
730
+ # Evaluation
731
+ results = {}
732
+ if training_args.do_eval:
733
+ logger.info("*** Evaluate ***")
734
+ metrics = trainer.evaluate()
735
+ max_eval_samples = (
736
+ data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
737
+ )
738
+ metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
739
+
740
+ trainer.log_metrics("eval", metrics)
741
+ trainer.save_metrics("eval", metrics)
742
+
743
+ # Write model card and (optionally) push to hub
744
+ config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
745
+ kwargs = {
746
+ "finetuned_from": model_args.model_name_or_path,
747
+ "tasks": "speech-recognition",
748
+ "tags": ["automatic-speech-recognition", data_args.dataset_name],
749
+ "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
750
+ "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
751
+ }
752
+ if "common_voice" in data_args.dataset_name:
753
+ kwargs["language"] = config_name
754
+
755
+ if training_args.push_to_hub:
756
+ trainer.push_to_hub(**kwargs)
757
+ else:
758
+ trainer.create_model_card(**kwargs)
759
+
760
+ return results
761
+
762
+
763
+ if __name__ == "__main__":
764
+ main()
run_xlsr_300_common_voice.sh ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/bin/bash
2
+
3
+ python run_speech_recognition_ctc_bnb.py \
4
+ --dataset_name="mozilla-foundation/common_voice_7_0" \
5
+ --model_name_or_path="facebook/wav2vec2-xls-r-300m" \
6
+ --dataset_config_name="it" \
7
+ --output_dir="../wav2vec2-xls-r-300m-italian" \
8
+ --overwrite_output_dir \
9
+ --num_train_epochs="5" \
10
+ --per_device_train_batch_size="64" \
11
+ --learning_rate="3e-4" \
12
+ --warmup_steps="500" \
13
+ --evaluation_strategy="steps" \
14
+ --text_column_name="sentence" \
15
+ --length_column_name="input_length" \
16
+ --save_steps="400" \
17
+ --eval_steps="100" \
18
+ --layerdrop="0.0" \
19
+ --save_total_limit="3" \
20
+ --freeze_feature_encoder \
21
+ --gradient_checkpointing \
22
+ --chars_to_ignore , ? . ! - \; \: \" “ % ‘ ” � \
23
+ --fp16 \
24
+ --group_by_length \
25
+ --push_to_hub \
26
+ --use_auth_token \
27
+ --hub_model_id="dbdmg/wav2vec2-xls-r-300m-italian" \
28
+ --do_train --do_eval