versae commited on
Commit
b3540f4
1 Parent(s): b4a3c9a

Preparing for training

Browse files
Files changed (4) hide show
  1. .gitattributes +1 -0
  2. .gitignore +1 -0
  3. run.sh +40 -0
  4. run_speech_recognition_ctc.py +792 -0
.gitattributes CHANGED
@@ -25,3 +25,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
25
  *.zip filter=lfs diff=lfs merge=lfs -text
26
  *.zstandard filter=lfs diff=lfs merge=lfs -text
27
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
25
  *.zip filter=lfs diff=lfs merge=lfs -text
26
  *.zstandard filter=lfs diff=lfs merge=lfs -text
27
  *tfevents* filter=lfs diff=lfs merge=lfs -text
28
+ *.wandb filter=lfs diff=lfs merge=lfs -text
.gitignore ADDED
@@ -0,0 +1 @@
 
 
1
+ checkpoint-*/
run.sh ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ WANDB_ENTITY=NbAiLab WANDB_PROJECT=wav2vec2 python run_speech_recognition_ctc.py \
2
+ --dataset_name="NbAiLab/NPSC" \
3
+ --model_name_or_path="KBLab/wav2vec2-large-voxrex" \
4
+ --hub_model_id="NbAiLab/wav2vec2-large-voxrex-npsc" \
5
+ --dataset_config_name="16K_mp3" \
6
+ --output_dir="./" \
7
+ --overwrite_output_dir \
8
+ --num_train_epochs="15" \
9
+ --per_device_train_batch_size="16" \
10
+ --per_device_eval_batch_size="16" \
11
+ --gradient_accumulation_steps="2" \
12
+ --learning_rate="1e-4" \
13
+ --warmup_steps="2000" \
14
+ --length_column_name="input_length" \
15
+ --evaluation_strategy="steps" \
16
+ --text_column_name="text" \
17
+ --save_steps="500" \
18
+ --eval_steps="500" \
19
+ --logging_steps="100" \
20
+ --layerdrop="0.041" \
21
+ --attention_dropout="0.094" \
22
+ --activation_dropout="0.055" \
23
+ --hidden_dropout="0.047" \
24
+ --save_total_limit="3" \
25
+ --freeze_feature_encoder \
26
+ --feat_proj_dropout="0.04" \
27
+ --mask_time_prob="0.082" \
28
+ --mask_time_length="10" \
29
+ --mask_feature_prob="0.25" \
30
+ --mask_feature_length="64" \
31
+ --gradient_checkpointing \
32
+ --min_duration_in_seconds="0.5" \
33
+ --max_duration_in_seconds="30.0" \
34
+ --use_auth_token \
35
+ --seed="42" \
36
+ --fp16 \
37
+ --group_by_length \
38
+ --do_train --do_eval \
39
+ --push_to_hub \
40
+ --preprocessing_num_workers="32"
run_speech_recognition_ctc.py ADDED
@@ -0,0 +1,792 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+ from dataclasses import dataclass, field
26
+ from typing import Dict, List, Optional, Union
27
+
28
+ import datasets
29
+ import numpy as np
30
+ import torch
31
+ from datasets import DatasetDict, load_dataset, load_metric
32
+
33
+ import transformers
34
+ from transformers import (
35
+ AutoConfig,
36
+ AutoFeatureExtractor,
37
+ AutoModelForCTC,
38
+ AutoProcessor,
39
+ AutoTokenizer,
40
+ HfArgumentParser,
41
+ Trainer,
42
+ TrainingArguments,
43
+ Wav2Vec2Processor,
44
+ set_seed,
45
+ )
46
+ from transformers.trainer_utils import get_last_checkpoint, is_main_process
47
+ from transformers.utils import check_min_version
48
+ from transformers.utils.versions import require_version
49
+
50
+
51
+ # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
52
+ check_min_version("4.16.0.dev0")
53
+
54
+ require_version("datasets>=1.13.3", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
55
+
56
+
57
+ logger = logging.getLogger(__name__)
58
+
59
+
60
+ def list_field(default=None, metadata=None):
61
+ return field(default_factory=lambda: default, metadata=metadata)
62
+
63
+
64
+ @dataclass
65
+ class ModelArguments:
66
+ """
67
+ Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
68
+ """
69
+
70
+ model_name_or_path: str = field(
71
+ metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
72
+ )
73
+ tokenizer_name_or_path: Optional[str] = field(
74
+ default=None,
75
+ metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
76
+ )
77
+ cache_dir: Optional[str] = field(
78
+ default=None,
79
+ metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
80
+ )
81
+ freeze_feature_encoder: bool = field(
82
+ default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
83
+ )
84
+ attention_dropout: float = field(
85
+ default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
86
+ )
87
+ activation_dropout: float = field(
88
+ default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
89
+ )
90
+ feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
91
+ hidden_dropout: float = field(
92
+ default=0.0,
93
+ metadata={
94
+ "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
95
+ },
96
+ )
97
+ final_dropout: float = field(
98
+ default=0.0,
99
+ metadata={"help": "The dropout probability for the final projection layer."},
100
+ )
101
+ mask_time_prob: float = field(
102
+ default=0.05,
103
+ metadata={
104
+ "help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
105
+ "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
106
+ "vectors will be masked along the time axis."
107
+ },
108
+ )
109
+ mask_time_length: int = field(
110
+ default=10,
111
+ metadata={"help": "Length of vector span to mask along the time axis."},
112
+ )
113
+ mask_feature_prob: float = field(
114
+ default=0.0,
115
+ metadata={
116
+ "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
117
+ "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
118
+ },
119
+ )
120
+ mask_feature_length: int = field(
121
+ default=10,
122
+ metadata={"help": "Length of vector span to mask along the feature axis."},
123
+ )
124
+ layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
125
+ ctc_loss_reduction: Optional[str] = field(
126
+ default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
127
+ )
128
+ ctc_zero_infinity: Optional[bool] = field(
129
+ default=False, metadata={"help": "If True, will try yo aboud the CTC loss goinf to infinity."}
130
+ )
131
+
132
+ @dataclass
133
+ class DataTrainingArguments:
134
+ """
135
+ Arguments pertaining to what data we are going to input our model for training and eval.
136
+
137
+ Using `HfArgumentParser` we can turn this class
138
+ into argparse arguments to be able to specify them on
139
+ the command line.
140
+ """
141
+
142
+ dataset_name: str = field(
143
+ metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
144
+ )
145
+ dataset_config_name: str = field(
146
+ default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
147
+ )
148
+ train_split_name: str = field(
149
+ default="train+validation",
150
+ metadata={
151
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
152
+ },
153
+ )
154
+ eval_split_name: str = field(
155
+ default="test",
156
+ metadata={
157
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
158
+ },
159
+ )
160
+ audio_column_name: str = field(
161
+ default="audio",
162
+ metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
163
+ )
164
+ text_column_name: str = field(
165
+ default="text",
166
+ metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
167
+ )
168
+ overwrite_cache: bool = field(
169
+ default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
170
+ )
171
+ preprocessing_num_workers: Optional[int] = field(
172
+ default=None,
173
+ metadata={"help": "The number of processes to use for the preprocessing."},
174
+ )
175
+ max_train_samples: Optional[int] = field(
176
+ default=None,
177
+ metadata={
178
+ "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
179
+ "value if set."
180
+ },
181
+ )
182
+ max_eval_samples: Optional[int] = field(
183
+ default=None,
184
+ metadata={
185
+ "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
186
+ "value if set."
187
+ },
188
+ )
189
+ chars_to_ignore: Optional[List[str]] = list_field(
190
+ default=None,
191
+ metadata={"help": "A list of characters to remove from the transcripts."},
192
+ )
193
+ eval_metrics: List[str] = list_field(
194
+ default=["wer"],
195
+ metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
196
+ )
197
+ max_duration_in_seconds: float = field(
198
+ default=20.0,
199
+ metadata={
200
+ "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
201
+ },
202
+ )
203
+ min_duration_in_seconds: float = field(
204
+ default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
205
+ )
206
+ preprocessing_only: bool = field(
207
+ default=False,
208
+ metadata={
209
+ "help": "Whether to only do data preprocessing and skip training. "
210
+ "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
211
+ "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
212
+ "so that the cached datasets can consequently be loaded in distributed training"
213
+ },
214
+ )
215
+ use_auth_token: bool = field(
216
+ default=False,
217
+ metadata={
218
+ "help": "If :obj:`True`, will use the token generated when running"
219
+ ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
220
+ },
221
+ )
222
+ unk_token: str = field(
223
+ default="[UNK]",
224
+ metadata={"help": "The unk token for the tokenizer"},
225
+ )
226
+ pad_token: str = field(
227
+ default="[PAD]",
228
+ metadata={"help": "The padding token for the tokenizer"},
229
+ )
230
+ word_delimiter_token: str = field(
231
+ default="|",
232
+ metadata={"help": "The word delimiter token for the tokenizer"},
233
+ )
234
+ phoneme_language: Optional[str] = field(
235
+ default=None,
236
+ metadata={
237
+ "help": "The target language that should be used be"
238
+ " passed to the tokenizer for tokenization. Note that"
239
+ " this is only relevant if the model classifies the"
240
+ " input audio to a sequence of phoneme sequences."
241
+ },
242
+ )
243
+
244
+
245
+ @dataclass
246
+ class DataCollatorCTCWithPadding:
247
+ """
248
+ Data collator that will dynamically pad the inputs received.
249
+ Args:
250
+ processor (:class:`~transformers.AutoProcessor`)
251
+ The processor used for proccessing the data.
252
+ padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
253
+ Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
254
+ among:
255
+ * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
256
+ sequence if provided).
257
+ * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
258
+ maximum acceptable input length for the model if that argument is not provided.
259
+ * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
260
+ different lengths).
261
+ max_length (:obj:`int`, `optional`):
262
+ Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
263
+ max_length_labels (:obj:`int`, `optional`):
264
+ Maximum length of the ``labels`` returned list and optionally padding length (see above).
265
+ pad_to_multiple_of (:obj:`int`, `optional`):
266
+ If set will pad the sequence to a multiple of the provided value.
267
+ This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
268
+ 7.5 (Volta).
269
+ """
270
+
271
+ processor: AutoProcessor
272
+ padding: Union[bool, str] = "longest"
273
+ pad_to_multiple_of: Optional[int] = None
274
+ pad_to_multiple_of_labels: Optional[int] = None
275
+
276
+ def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
277
+ # split inputs and labels since they have to be of different lenghts and need
278
+ # different padding methods
279
+ input_features = [{"input_values": feature["input_values"]} for feature in features]
280
+ label_features = [{"input_ids": feature["labels"]} for feature in features]
281
+
282
+ batch = self.processor.pad(
283
+ input_features,
284
+ padding=self.padding,
285
+ pad_to_multiple_of=self.pad_to_multiple_of,
286
+ return_tensors="pt",
287
+ )
288
+
289
+ with self.processor.as_target_processor():
290
+ labels_batch = self.processor.pad(
291
+ label_features,
292
+ padding=self.padding,
293
+ pad_to_multiple_of=self.pad_to_multiple_of_labels,
294
+ return_tensors="pt",
295
+ )
296
+
297
+ # replace padding with -100 to ignore loss correctly
298
+ labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
299
+
300
+ batch["labels"] = labels
301
+
302
+ return batch
303
+
304
+
305
+ def create_vocabulary_from_data(
306
+ datasets: DatasetDict,
307
+ word_delimiter_token: Optional[str] = None,
308
+ unk_token: Optional[str] = None,
309
+ pad_token: Optional[str] = None,
310
+ ):
311
+ # Given training and test labels create vocabulary
312
+ def extract_all_chars(batch):
313
+ all_text = " ".join(batch["target_text"])
314
+ vocab = list(set(all_text))
315
+ return {"vocab": [vocab], "all_text": [all_text]}
316
+
317
+ vocabs = datasets.map(
318
+ extract_all_chars,
319
+ batched=True,
320
+ batch_size=-1,
321
+ keep_in_memory=True,
322
+ remove_columns=datasets["train"].column_names,
323
+ )
324
+
325
+ # take union of all unique characters in each dataset
326
+ vocab_set = functools.reduce(
327
+ lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
328
+ )
329
+
330
+ vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
331
+
332
+ # replace white space with delimiter token
333
+ if word_delimiter_token is not None:
334
+ vocab_dict[word_delimiter_token] = vocab_dict[" "]
335
+ del vocab_dict[" "]
336
+
337
+ # add unk and pad token
338
+ if unk_token is not None:
339
+ vocab_dict[unk_token] = len(vocab_dict)
340
+
341
+ if pad_token is not None:
342
+ vocab_dict[pad_token] = len(vocab_dict)
343
+
344
+ return vocab_dict
345
+
346
+
347
+ def main():
348
+ # See all possible arguments in src/transformers/training_args.py
349
+ # or by passing the --help flag to this script.
350
+ # We now keep distinct sets of args, for a cleaner separation of concerns.
351
+
352
+ parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
353
+ if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
354
+ # If we pass only one argument to the script and it's the path to a json file,
355
+ # let's parse it to get our arguments.
356
+ model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
357
+ else:
358
+ model_args, data_args, training_args = parser.parse_args_into_dataclasses()
359
+
360
+ # Detecting last checkpoint.
361
+ last_checkpoint = None
362
+ if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
363
+ last_checkpoint = get_last_checkpoint(training_args.output_dir)
364
+ if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
365
+ raise ValueError(
366
+ f"Output directory ({training_args.output_dir}) already exists and is not empty. "
367
+ "Use --overwrite_output_dir to overcome."
368
+ )
369
+ elif last_checkpoint is not None:
370
+ logger.info(
371
+ f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
372
+ "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
373
+ )
374
+
375
+ # Setup logging
376
+ logging.basicConfig(
377
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
378
+ datefmt="%m/%d/%Y %H:%M:%S",
379
+ handlers=[logging.StreamHandler(sys.stdout)],
380
+ )
381
+ logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
382
+
383
+ # Log on each process the small summary:
384
+ logger.warning(
385
+ f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
386
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
387
+ )
388
+ # Set the verbosity to info of the Transformers logger (on main process only):
389
+ if is_main_process(training_args.local_rank):
390
+ transformers.utils.logging.set_verbosity_info()
391
+ logger.info("Training/evaluation parameters %s", training_args)
392
+
393
+ # Set seed before initializing model.
394
+ set_seed(training_args.seed)
395
+
396
+ # Pre-processing dataset
397
+ import re
398
+ def filter_numeric(entry):
399
+ return (
400
+ "0" not in entry["text"]
401
+ and "1" not in entry["text"]
402
+ and "2" not in entry["text"]
403
+ and "3" not in entry["text"]
404
+ and "4" not in entry["text"]
405
+ and "5" not in entry["text"]
406
+ and "6" not in entry["text"]
407
+ and "7" not in entry["text"]
408
+ and "8" not in entry["text"]
409
+ and "9" not in entry["text"]
410
+ )
411
+
412
+ def filter_inaudible(entry):
413
+ return not re.search("\d|<inaudible>", entry["text"], flags=re.IGNORECASE)
414
+
415
+ def filter_nynorsk(entry):
416
+ return re.search("nb-no", entry["sentence_language_code"], flags=re.IGNORECASE)
417
+
418
+ def filter_tooshort(entry):
419
+ #print(f"The audio sample ({entry["audio"]["path"]}) is too small, and has been omitted. "
420
+ return (len(entry["text"]) <= len(entry["audio"]["array"]) // 320) and (len(entry["text"].strip()) >= 3)
421
+
422
+ def map_dataset(entry):
423
+ batch = {"text": entry["text"].lower()}
424
+ batch["text"] = re.sub('[áàâ]', 'a', batch["text"])
425
+ batch["text"] = re.sub('[ä]', 'æ', batch["text"])
426
+ batch["text"] = re.sub('[éèëê]', 'e', batch["text"])
427
+ batch["text"] = re.sub('[íìïî]', 'i', batch["text"])
428
+ batch["text"] = re.sub('[óòöô]', 'o', batch["text"])
429
+ batch["text"] = re.sub('[ö]', 'ø', batch["text"])
430
+ batch["text"] = re.sub('[ç]', 'c', batch["text"])
431
+ batch["text"] = re.sub('[úùüû]', 'u', batch["text"])
432
+ batch["text"] = re.sub('\s', ' ', batch["text"])
433
+ batch["text"] = re.sub('<ee>', 'eee', batch["text"])
434
+ batch["text"] = re.sub('<qq>', 'qqq', batch["text"])
435
+ batch["text"] = re.sub('<mm>', 'mmm', batch["text"])
436
+ # batch["text"] = re.sub('<inaudible>', '?', batch["text"])
437
+ if "<" in batch["text"]:
438
+ raise ValueError(batch["text"])
439
+ return batch
440
+
441
+ # 1. First, let's load the dataset
442
+ raw_datasets = DatasetDict()
443
+
444
+ if training_args.do_train:
445
+ raw_datasets["train"] = load_dataset(
446
+ data_args.dataset_name,
447
+ data_args.dataset_config_name,
448
+ split=data_args.train_split_name,
449
+ use_auth_token=data_args.use_auth_token,
450
+ ).shuffle()
451
+ raw_datasets["train"] = raw_datasets["train"].filter(filter_numeric).filter(filter_inaudible).filter(filter_nynorsk).filter(filter_tooshort)
452
+ raw_datasets["train"] = raw_datasets["train"].map(map_dataset)
453
+
454
+ if data_args.audio_column_name not in raw_datasets["train"].column_names:
455
+ raise ValueError(
456
+ f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
457
+ "Make sure to set `--audio_column_name` to the correct audio column - one of "
458
+ f"{', '.join(raw_datasets['train'].column_names)}."
459
+ )
460
+
461
+ if data_args.text_column_name not in raw_datasets["train"].column_names:
462
+ raise ValueError(
463
+ f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
464
+ "Make sure to set `--text_column_name` to the correct text column - one of "
465
+ f"{', '.join(raw_datasets['train'].column_names)}."
466
+ )
467
+
468
+ if data_args.max_train_samples is not None:
469
+ raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
470
+
471
+ if training_args.do_eval:
472
+ raw_datasets["eval"] = load_dataset(
473
+ data_args.dataset_name,
474
+ data_args.dataset_config_name,
475
+ split=data_args.eval_split_name,
476
+ use_auth_token=data_args.use_auth_token,
477
+ ).shuffle()
478
+ raw_datasets["eval"] = raw_datasets["eval"].filter(filter_numeric).filter(filter_inaudible).filter(filter_nynorsk).filter(filter_tooshort)
479
+ raw_datasets["eval"] = raw_datasets["eval"].map(map_dataset)
480
+
481
+ if data_args.max_eval_samples is not None:
482
+ raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
483
+
484
+
485
+ # 2. We remove some special characters from the datasets
486
+ # that make training complicated and do not help in transcribing the speech
487
+ # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
488
+ # that could be easily picked up by the model
489
+ #chars_to_ignore_regex = (
490
+ # f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
491
+ #)
492
+ chars_to_ignore_regex = '[\,\?\.\!\-\;\:\"\“\%\‘\”\�\'\–\_\\\+\#\/]'
493
+
494
+ text_column_name = data_args.text_column_name
495
+
496
+ def remove_special_characters(batch):
497
+ if chars_to_ignore_regex is not None:
498
+ batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
499
+ else:
500
+ batch["target_text"] = batch[text_column_name].lower() + " "
501
+ return batch
502
+
503
+ with training_args.main_process_first(desc="dataset map special characters removal"):
504
+ raw_datasets = raw_datasets.map(
505
+ remove_special_characters,
506
+ remove_columns=[text_column_name],
507
+ desc="remove special characters from datasets",
508
+ )
509
+
510
+ # save special tokens for tokenizer
511
+ word_delimiter_token = data_args.word_delimiter_token
512
+ unk_token = data_args.unk_token
513
+ pad_token = data_args.pad_token
514
+
515
+ # 3. Next, let's load the config as we might need it to create
516
+ # the tokenizer
517
+ # load config
518
+ config = AutoConfig.from_pretrained(
519
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
520
+ )
521
+
522
+ # 4. Next, if no tokenizer file is defined,
523
+ # we create the vocabulary of the model by extracting all unique characters from
524
+ # the training and evaluation datasets
525
+ # We need to make sure that only first rank saves vocabulary
526
+ # make sure all processes wait until vocab is created
527
+ tokenizer_name_or_path = model_args.tokenizer_name_or_path
528
+ tokenizer_kwargs = {}
529
+ if tokenizer_name_or_path is None:
530
+ # save vocab in training output dir
531
+ tokenizer_name_or_path = training_args.output_dir
532
+
533
+ vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
534
+
535
+ with training_args.main_process_first():
536
+ if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
537
+ os.remove(vocab_file)
538
+
539
+ with training_args.main_process_first(desc="dataset map vocabulary creation"):
540
+ if not os.path.isfile(vocab_file):
541
+ os.makedirs(tokenizer_name_or_path, exist_ok=True)
542
+ vocab_dict = create_vocabulary_from_data(
543
+ raw_datasets,
544
+ word_delimiter_token=word_delimiter_token,
545
+ unk_token=unk_token,
546
+ pad_token=pad_token,
547
+ )
548
+
549
+ # save vocab dict to be loaded into tokenizer
550
+ with open(vocab_file, "w") as file:
551
+ json.dump(vocab_dict, file)
552
+
553
+ # if tokenizer has just been created
554
+ # it is defined by `tokenizer_class` if present in config else by `model_type`
555
+ tokenizer_kwargs = {
556
+ "config": config if config.tokenizer_class is not None else None,
557
+ "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
558
+ "unk_token": unk_token,
559
+ "pad_token": pad_token,
560
+ "word_delimiter_token": word_delimiter_token,
561
+ }
562
+
563
+ # 5. Now we can instantiate the feature extractor, tokenizer and model
564
+ # Note for distributed training, the .from_pretrained methods guarantee that only
565
+ # one local process can concurrently download model & vocab.
566
+
567
+ # load feature_extractor and tokenizer
568
+ tokenizer = AutoTokenizer.from_pretrained(
569
+ tokenizer_name_or_path,
570
+ use_auth_token=data_args.use_auth_token,
571
+ **tokenizer_kwargs,
572
+ )
573
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
574
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
575
+ )
576
+
577
+ # adapt config
578
+ config.update(
579
+ {
580
+ "feat_proj_dropout": model_args.feat_proj_dropout,
581
+ "attention_dropout": model_args.attention_dropout,
582
+ "hidden_dropout": model_args.hidden_dropout,
583
+ "final_dropout": model_args.final_dropout,
584
+ "mask_time_prob": model_args.mask_time_prob,
585
+ "mask_time_length": model_args.mask_time_length,
586
+ "mask_feature_prob": model_args.mask_feature_prob,
587
+ "mask_feature_length": model_args.mask_feature_length,
588
+ "gradient_checkpointing": training_args.gradient_checkpointing,
589
+ "layerdrop": model_args.layerdrop,
590
+ "ctc_loss_reduction": model_args.ctc_loss_reduction,
591
+ "ctc_zero_infinity": model_args.ctc_zero_infinity,
592
+ "pad_token_id": tokenizer.pad_token_id,
593
+ "vocab_size": len(tokenizer),
594
+ "activation_dropout": model_args.activation_dropout,
595
+ }
596
+ )
597
+
598
+ # create model
599
+ model = AutoModelForCTC.from_pretrained(
600
+ model_args.model_name_or_path,
601
+ cache_dir=model_args.cache_dir,
602
+ config=config,
603
+ use_auth_token=data_args.use_auth_token,
604
+ )
605
+
606
+ # freeze encoder
607
+ if model_args.freeze_feature_encoder:
608
+ model.freeze_feature_encoder()
609
+
610
+ # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
611
+ # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
612
+ # so that we just need to set the correct target sampling rate and normalize the input
613
+ # via the `feature_extractor`
614
+
615
+ # make sure that dataset decodes audio with correct sampling rate
616
+ dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
617
+ if dataset_sampling_rate != feature_extractor.sampling_rate:
618
+ raw_datasets = raw_datasets.cast_column(
619
+ data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
620
+ )
621
+
622
+ # derive max & min input length for sample rate & max duration
623
+ max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
624
+ min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
625
+ audio_column_name = data_args.audio_column_name
626
+ num_workers = data_args.preprocessing_num_workers
627
+
628
+ # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
629
+ phoneme_language = data_args.phoneme_language
630
+
631
+ # Preprocessing the datasets.
632
+ # We need to read the audio files as arrays and tokenize the targets.
633
+ def prepare_dataset(batch):
634
+ # load audio
635
+ sample = batch[audio_column_name]
636
+
637
+ inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
638
+ batch["input_values"] = inputs.input_values[0]
639
+ batch["input_length"] = len(batch["input_values"])
640
+
641
+ # encode targets
642
+ additional_kwargs = {}
643
+ if phoneme_language is not None:
644
+ additional_kwargs["phonemizer_lang"] = phoneme_language
645
+
646
+ batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
647
+ return batch
648
+
649
+ with training_args.main_process_first(desc="dataset map preprocessing"):
650
+ vectorized_datasets = raw_datasets.map(
651
+ prepare_dataset,
652
+ remove_columns=next(iter(raw_datasets.values())).column_names,
653
+ num_proc=num_workers,
654
+ desc="preprocess datasets",
655
+ )
656
+
657
+ def is_audio_in_length_range(length):
658
+ return length > min_input_length and length < max_input_length
659
+
660
+ # filter data that is shorter than min_input_length
661
+ vectorized_datasets = vectorized_datasets.filter(
662
+ is_audio_in_length_range,
663
+ num_proc=num_workers,
664
+ input_columns=["input_length"],
665
+ )
666
+
667
+ # 7. Next, we can prepare the training.
668
+ # Let's use word error rate (WER) as our evaluation metric,
669
+ # instantiate a data collator and the trainer
670
+
671
+ # Define evaluation metrics during training, *i.e.* word error rate, character error rate
672
+ eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
673
+
674
+ # for large datasets it is advised to run the preprocessing on a
675
+ # single machine first with ``args.preprocessing_only`` since there will mostly likely
676
+ # be a timeout when running the script in distributed mode.
677
+ # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
678
+ # cached dataset
679
+ if data_args.preprocessing_only:
680
+ logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
681
+ return
682
+
683
+ def compute_metrics(pred):
684
+ pred_logits = pred.predictions
685
+ pred_ids = np.argmax(pred_logits, axis=-1)
686
+
687
+ pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
688
+
689
+ pred_str = tokenizer.batch_decode(pred_ids)
690
+ # we do not want to group tokens when computing the metrics
691
+ label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
692
+
693
+ metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
694
+
695
+ return metrics
696
+
697
+ # Now save everything to be able to create a single processor later
698
+ if is_main_process(training_args.local_rank):
699
+ # save feature extractor, tokenizer and config
700
+ feature_extractor.save_pretrained(training_args.output_dir)
701
+ tokenizer.save_pretrained(training_args.output_dir)
702
+ config.save_pretrained(training_args.output_dir)
703
+
704
+ try:
705
+ processor = AutoProcessor.from_pretrained(training_args.output_dir)
706
+ except (OSError, KeyError):
707
+ warnings.warn(
708
+ "Loading a processor from a feature extractor config that does not"
709
+ " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
710
+ " attribute to your `preprocessor_config.json` file to suppress this warning: "
711
+ " `'processor_class': 'Wav2Vec2Processor'`",
712
+ FutureWarning,
713
+ )
714
+ processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
715
+
716
+ # Instantiate custom data collator
717
+ data_collator = DataCollatorCTCWithPadding(processor=processor)
718
+
719
+ # Initialize Trainer
720
+ trainer = Trainer(
721
+ model=model,
722
+ data_collator=data_collator,
723
+ args=training_args,
724
+ compute_metrics=compute_metrics,
725
+ train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
726
+ eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
727
+ tokenizer=feature_extractor,
728
+ )
729
+
730
+ # 8. Finally, we can start training
731
+
732
+ # Training
733
+ if training_args.do_train:
734
+
735
+ # use last checkpoint if exist
736
+ if last_checkpoint is not None:
737
+ checkpoint = last_checkpoint
738
+ elif os.path.isdir(model_args.model_name_or_path):
739
+ checkpoint = model_args.model_name_or_path
740
+ else:
741
+ checkpoint = None
742
+
743
+ train_result = trainer.train(resume_from_checkpoint=checkpoint)
744
+ trainer.save_model()
745
+
746
+ metrics = train_result.metrics
747
+ max_train_samples = (
748
+ data_args.max_train_samples
749
+ if data_args.max_train_samples is not None
750
+ else len(vectorized_datasets["train"])
751
+ )
752
+ metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
753
+
754
+ trainer.log_metrics("train", metrics)
755
+ trainer.save_metrics("train", metrics)
756
+ trainer.save_state()
757
+
758
+ # Evaluation
759
+ results = {}
760
+ if training_args.do_eval:
761
+ logger.info("*** Evaluate ***")
762
+ metrics = trainer.evaluate()
763
+ max_eval_samples = (
764
+ data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
765
+ )
766
+ metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
767
+
768
+ trainer.log_metrics("eval", metrics)
769
+ trainer.save_metrics("eval", metrics)
770
+
771
+ # Write model card and (optionally) push to hub
772
+ config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
773
+ kwargs = {
774
+ "finetuned_from": model_args.model_name_or_path,
775
+ "tasks": "speech-recognition",
776
+ "tags": ["automatic-speech-recognition", data_args.dataset_name],
777
+ "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
778
+ "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
779
+ }
780
+ if "common_voice" in data_args.dataset_name:
781
+ kwargs["language"] = config_name
782
+
783
+ if training_args.push_to_hub:
784
+ trainer.push_to_hub(**kwargs)
785
+ else:
786
+ trainer.create_model_card(**kwargs)
787
+
788
+ return results
789
+
790
+
791
+ if __name__ == "__main__":
792
+ main()