anuragshas commited on
Commit
c9c2211
1 Parent(s): 4869634

Training in progress, step 400

Browse files
.gitignore ADDED
@@ -0,0 +1 @@
 
 
1
+ checkpoint-*/
.ipynb_checkpoints/run-checkpoint.sh ADDED
@@ -0,0 +1,36 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ python run_speech_recognition_ctc.py \
2
+ --dataset_name="mozilla-foundation/common_voice_9_0" \
3
+ --model_name_or_path="facebook/wav2vec2-xls-r-300m" \
4
+ --dataset_config_name="ur" \
5
+ --output_dir="./" \
6
+ --overwrite_output_dir \
7
+ --max_steps="5108" \
8
+ --per_device_train_batch_size="64" \
9
+ --per_device_eval_batch_size="64" \
10
+ --gradient_accumulation_steps="2" \
11
+ --learning_rate="7.5e-5" \
12
+ --warmup_ratio="0.1" \
13
+ --length_column_name="input_length" \
14
+ --evaluation_strategy="steps" \
15
+ --text_column_name="sentence" \
16
+ --chars_to_ignore , ? . ! \- \; \: \" “ % ‘ ” � — ’ … – \| \& ؛ ، ٫ ؟ ۔ ٪ \` \
17
+ --save_steps="400" \
18
+ --eval_steps="400" \
19
+ --logging_steps="100" \
20
+ --layerdrop="0.0" \
21
+ --activation_dropout="0.1" \
22
+ --save_total_limit="1" \
23
+ --freeze_feature_encoder \
24
+ --feat_proj_dropout="0.0" \
25
+ --mask_time_prob="0.75" \
26
+ --mask_time_length="10" \
27
+ --mask_feature_prob="0.25" \
28
+ --mask_feature_length="64" \
29
+ --seed="42" \
30
+ --gradient_checkpointing \
31
+ --use_auth_token \
32
+ --fp16 \
33
+ --group_by_length \
34
+ --do_train --do_eval \
35
+ --bnb --tristage_sched \
36
+ --push_to_hub
.ipynb_checkpoints/run_speech_recognition_ctc-checkpoint.py ADDED
@@ -0,0 +1,822 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python
2
+ # coding=utf-8
3
+ # Copyright 2021 The HuggingFace Inc. team. All rights reserved.
4
+ #
5
+ # Licensed under the Apache License, Version 2.0 (the "License");
6
+ # you may not use this file except in compliance with the License.
7
+ # You may obtain a copy of the License at
8
+ #
9
+ # http://www.apache.org/licenses/LICENSE-2.0
10
+ #
11
+ # Unless required by applicable law or agreed to in writing, software
12
+ # distributed under the License is distributed on an "AS IS" BASIS,
13
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14
+ # See the License for the specific language governing permissions and
15
+
16
+ """ Fine-tuning a 🤗 Transformers CTC model for automatic speech recognition"""
17
+
18
+ import functools
19
+ import json
20
+ import logging
21
+ import os
22
+ import re
23
+ import sys
24
+ import unicodedata
25
+ import warnings
26
+ from dataclasses import dataclass, field
27
+ from typing import Dict, List, Optional, Union
28
+
29
+ import datasets
30
+ import numpy as np
31
+ import torch
32
+ from torch.optim.lr_scheduler import LambdaLR
33
+ from datasets import DatasetDict, load_dataset, load_metric, load_from_disk
34
+
35
+ import bitsandbytes as bnb
36
+ import transformers
37
+ from transformers import (
38
+ AutoConfig,
39
+ AutoFeatureExtractor,
40
+ AutoModelForCTC,
41
+ AutoProcessor,
42
+ AutoTokenizer,
43
+ HfArgumentParser,
44
+ Trainer,
45
+ TrainingArguments,
46
+ Wav2Vec2Processor,
47
+ set_seed,
48
+ )
49
+ from transformers.trainer_pt_utils import get_parameter_names
50
+ from transformers.trainer_utils import get_last_checkpoint, is_main_process
51
+ from transformers.utils import check_min_version
52
+ from transformers.utils.versions import require_version
53
+
54
+
55
+ # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
56
+ check_min_version("4.16.0.dev0")
57
+
58
+ require_version("datasets>=1.13.3", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
59
+
60
+
61
+ logger = logging.getLogger(__name__)
62
+
63
+
64
+ def list_field(default=None, metadata=None):
65
+ return field(default_factory=lambda: default, metadata=metadata)
66
+
67
+
68
+ @dataclass
69
+ class ModelArguments:
70
+ """
71
+ Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
72
+ """
73
+
74
+ model_name_or_path: str = field(
75
+ metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
76
+ )
77
+ tokenizer_name_or_path: Optional[str] = field(
78
+ default=None,
79
+ metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
80
+ )
81
+ cache_dir: Optional[str] = field(
82
+ default=None,
83
+ metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
84
+ )
85
+ freeze_feature_encoder: bool = field(
86
+ default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
87
+ )
88
+ attention_dropout: float = field(
89
+ default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
90
+ )
91
+ activation_dropout: float = field(
92
+ default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
93
+ )
94
+ feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
95
+ hidden_dropout: float = field(
96
+ default=0.0,
97
+ metadata={
98
+ "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
99
+ },
100
+ )
101
+ final_dropout: float = field(
102
+ default=0.0,
103
+ metadata={"help": "The dropout probability for the final projection layer."},
104
+ )
105
+ mask_time_prob: float = field(
106
+ default=0.05,
107
+ metadata={
108
+ "help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
109
+ "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
110
+ "vectors will be masked along the time axis."
111
+ },
112
+ )
113
+ mask_time_length: int = field(
114
+ default=10,
115
+ metadata={"help": "Length of vector span to mask along the time axis."},
116
+ )
117
+ mask_feature_prob: float = field(
118
+ default=0.0,
119
+ metadata={
120
+ "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
121
+ "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
122
+ },
123
+ )
124
+ mask_feature_length: int = field(
125
+ default=10,
126
+ metadata={"help": "Length of vector span to mask along the feature axis."},
127
+ )
128
+ layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
129
+ ctc_loss_reduction: Optional[str] = field(
130
+ default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
131
+ )
132
+
133
+
134
+ @dataclass
135
+ class DataTrainingArguments:
136
+ """
137
+ Arguments pertaining to what data we are going to input our model for training and eval.
138
+
139
+ Using `HfArgumentParser` we can turn this class
140
+ into argparse arguments to be able to specify them on
141
+ the command line.
142
+ """
143
+
144
+ dataset_name: str = field(
145
+ metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
146
+ )
147
+ dataset_config_name: str = field(
148
+ default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
149
+ )
150
+ train_split_name: str = field(
151
+ default="train+validation",
152
+ metadata={
153
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
154
+ },
155
+ )
156
+ eval_split_name: str = field(
157
+ default="test",
158
+ metadata={
159
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
160
+ },
161
+ )
162
+ audio_column_name: str = field(
163
+ default="audio",
164
+ metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
165
+ )
166
+ text_column_name: str = field(
167
+ default="text",
168
+ metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
169
+ )
170
+ overwrite_cache: bool = field(
171
+ default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
172
+ )
173
+ preprocessing_num_workers: Optional[int] = field(
174
+ default=None,
175
+ metadata={"help": "The number of processes to use for the preprocessing."},
176
+ )
177
+ max_train_samples: Optional[int] = field(
178
+ default=None,
179
+ metadata={
180
+ "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
181
+ "value if set."
182
+ },
183
+ )
184
+ max_eval_samples: Optional[int] = field(
185
+ default=None,
186
+ metadata={
187
+ "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
188
+ "value if set."
189
+ },
190
+ )
191
+ chars_to_ignore: Optional[List[str]] = list_field(
192
+ default=None,
193
+ metadata={"help": "A list of characters to remove from the transcripts."},
194
+ )
195
+ eval_metrics: List[str] = list_field(
196
+ default=["wer", "cer"],
197
+ metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
198
+ )
199
+ max_duration_in_seconds: float = field(
200
+ default=20.0,
201
+ metadata={
202
+ "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
203
+ },
204
+ )
205
+ min_duration_in_seconds: float = field(
206
+ default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
207
+ )
208
+ preprocessing_only: bool = field(
209
+ default=False,
210
+ metadata={
211
+ "help": "Whether to only do data preprocessing and skip training. "
212
+ "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
213
+ "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
214
+ "so that the cached datasets can consequently be loaded in distributed training"
215
+ },
216
+ )
217
+ use_auth_token: bool = field(
218
+ default=False,
219
+ metadata={
220
+ "help": "If :obj:`True`, will use the token generated when running"
221
+ ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
222
+ },
223
+ )
224
+ unk_token: str = field(
225
+ default="[UNK]",
226
+ metadata={"help": "The unk token for the tokenizer"},
227
+ )
228
+ pad_token: str = field(
229
+ default="[PAD]",
230
+ metadata={"help": "The padding token for the tokenizer"},
231
+ )
232
+ word_delimiter_token: str = field(
233
+ default="|",
234
+ metadata={"help": "The word delimiter token for the tokenizer"},
235
+ )
236
+ phoneme_language: Optional[str] = field(
237
+ default=None,
238
+ metadata={
239
+ "help": "The target language that should be used be"
240
+ " passed to the tokenizer for tokenization. Note that"
241
+ " this is only relevant if the model classifies the"
242
+ " input audio to a sequence of phoneme sequences."
243
+ },
244
+ )
245
+
246
+ @dataclass
247
+ class ExtraArguments:
248
+ "Additional training arguments"
249
+ bnb: bool = field(
250
+ default=False,
251
+ metadata = {"help":"If true uses 8bit Adam"}
252
+ )
253
+ tristage_sched: bool = field(
254
+ default=False,
255
+ metadata = {"help":"If true uses tristage LR scheduler (refer to XLS-R paper)"}
256
+ )
257
+ wandb_project: str = field(
258
+ default="",
259
+ metadata = {"help":"Name of wandb project to log into"}
260
+ )
261
+
262
+
263
+ @dataclass
264
+ class DataCollatorCTCWithPadding:
265
+ """
266
+ Data collator that will dynamically pad the inputs received.
267
+ Args:
268
+ processor (:class:`~transformers.AutoProcessor`)
269
+ The processor used for proccessing the data.
270
+ padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
271
+ Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
272
+ among:
273
+ * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
274
+ sequence if provided).
275
+ * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
276
+ maximum acceptable input length for the model if that argument is not provided.
277
+ * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
278
+ different lengths).
279
+ max_length (:obj:`int`, `optional`):
280
+ Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
281
+ max_length_labels (:obj:`int`, `optional`):
282
+ Maximum length of the ``labels`` returned list and optionally padding length (see above).
283
+ pad_to_multiple_of (:obj:`int`, `optional`):
284
+ If set will pad the sequence to a multiple of the provided value.
285
+ This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
286
+ 7.5 (Volta).
287
+ """
288
+
289
+ processor: AutoProcessor
290
+ padding: Union[bool, str] = "longest"
291
+ pad_to_multiple_of: Optional[int] = None
292
+ pad_to_multiple_of_labels: Optional[int] = None
293
+
294
+ def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
295
+ # split inputs and labels since they have to be of different lenghts and need
296
+ # different padding methods
297
+ input_features = [{"input_values": feature["input_values"]} for feature in features]
298
+ label_features = [{"input_ids": feature["labels"]} for feature in features]
299
+
300
+ batch = self.processor.pad(
301
+ input_features,
302
+ padding=self.padding,
303
+ pad_to_multiple_of=self.pad_to_multiple_of,
304
+ return_tensors="pt",
305
+ )
306
+
307
+ with self.processor.as_target_processor():
308
+ labels_batch = self.processor.pad(
309
+ label_features,
310
+ padding=self.padding,
311
+ pad_to_multiple_of=self.pad_to_multiple_of_labels,
312
+ return_tensors="pt",
313
+ )
314
+
315
+ # replace padding with -100 to ignore loss correctly
316
+ labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
317
+
318
+ batch["labels"] = labels
319
+
320
+ return batch
321
+
322
+
323
+ def get_tri_stage_schedule(
324
+ optimizer, num_training_steps, ratios=[0.1, 0.4, 0.5], num_warmup_steps=None, num_hold_steps=None, start_ratio=0.01, end_ratio=0.05
325
+ ):
326
+ assert (num_warmup_steps is None) == (num_hold_steps is None)
327
+ if num_warmup_steps is None:
328
+ num_warmup_steps = int(ratios[0]*num_training_steps)
329
+ num_hold_steps = int(ratios[1]*num_training_steps)
330
+ start_decay_step = num_warmup_steps + num_hold_steps
331
+ a_w, b_w = (1-start_ratio)/num_warmup_steps, start_ratio
332
+ num_decay_steps = num_training_steps - start_decay_step
333
+ a_d, b_d = (end_ratio-1)/num_decay_steps, 1.
334
+
335
+ def lr_lambda(current_step):
336
+ if current_step < num_warmup_steps:
337
+ return a_w * float(current_step) + b_w
338
+ if current_step < start_decay_step:
339
+ return 1.
340
+ return max(end_ratio, a_d * float(current_step - start_decay_step) + b_d )
341
+
342
+ return LambdaLR(optimizer, lr_lambda)
343
+
344
+ def create_vocabulary_from_data(
345
+ datasets: DatasetDict,
346
+ word_delimiter_token: Optional[str] = None,
347
+ unk_token: Optional[str] = None,
348
+ pad_token: Optional[str] = None,
349
+ ):
350
+ # Given training and test labels create vocabulary
351
+ def extract_all_chars(batch):
352
+ all_text = " ".join(batch["target_text"])
353
+ vocab = list(set(all_text))
354
+ return {"vocab": [vocab], "all_text": [all_text]}
355
+
356
+ vocabs = datasets.map(
357
+ extract_all_chars,
358
+ batched=True,
359
+ batch_size=-1,
360
+ keep_in_memory=True,
361
+ remove_columns=datasets["train"].column_names,
362
+ )
363
+
364
+ # take union of all unique characters in each dataset
365
+ vocab_set = functools.reduce(
366
+ lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
367
+ )
368
+
369
+ vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
370
+
371
+ # replace white space with delimiter token
372
+ if word_delimiter_token is not None:
373
+ vocab_dict[word_delimiter_token] = vocab_dict[" "]
374
+ del vocab_dict[" "]
375
+
376
+ # add unk and pad token
377
+ if unk_token is not None:
378
+ vocab_dict[unk_token] = len(vocab_dict)
379
+
380
+ if pad_token is not None:
381
+ vocab_dict[pad_token] = len(vocab_dict)
382
+
383
+ return vocab_dict
384
+
385
+
386
+ def main():
387
+ # See all possible arguments in src/transformers/training_args.py
388
+ # or by passing the --help flag to this script.
389
+ # We now keep distinct sets of args, for a cleaner separation of concerns.
390
+
391
+ parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments, ExtraArguments))
392
+ if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
393
+ # If we pass only one argument to the script and it's the path to a json file,
394
+ # let's parse it to get our arguments.
395
+ model_args, data_args, training_args, extra_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
396
+ else:
397
+ model_args, data_args, training_args, extra_args = parser.parse_args_into_dataclasses()
398
+
399
+ # Detecting last checkpoint.
400
+ last_checkpoint = None
401
+ if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
402
+ last_checkpoint = get_last_checkpoint(training_args.output_dir)
403
+ if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
404
+ raise ValueError(
405
+ f"Output directory ({training_args.output_dir}) already exists and is not empty. "
406
+ "Use --overwrite_output_dir to overcome."
407
+ )
408
+ elif last_checkpoint is not None:
409
+ logger.info(
410
+ f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
411
+ "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
412
+ )
413
+
414
+ # Setup logging
415
+ logging.basicConfig(
416
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
417
+ datefmt="%m/%d/%Y %H:%M:%S",
418
+ handlers=[logging.StreamHandler(sys.stdout)],
419
+ )
420
+ logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
421
+
422
+ # Log on each process the small summary:
423
+ logger.warning(
424
+ f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
425
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
426
+ )
427
+ # Set the verbosity to info of the Transformers logger (on main process only):
428
+ if is_main_process(training_args.local_rank):
429
+ transformers.utils.logging.set_verbosity_info()
430
+ logger.info("Training/evaluation parameters %s", training_args)
431
+
432
+ # Set seed before initializing model.
433
+ set_seed(training_args.seed)
434
+
435
+ # configure wandb run
436
+ os.environ["WANDB_PROJECT"] = extra_args.wandb_project
437
+
438
+ # 1. First, let's load the dataset
439
+ raw_datasets = DatasetDict()
440
+
441
+ if training_args.do_train:
442
+ if os.path.isdir(data_args.dataset_name):
443
+ raw_datasets["train"] = load_from_disk(f"{data_args.dataset_name}/{data_args.train_split_name}")
444
+ else:
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
+ )
451
+
452
+ if data_args.audio_column_name not in raw_datasets["train"].column_names:
453
+ raise ValueError(
454
+ f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
455
+ "Make sure to set `--audio_column_name` to the correct audio column - one of "
456
+ f"{', '.join(raw_datasets['train'].column_names)}."
457
+ )
458
+
459
+ if data_args.text_column_name not in raw_datasets["train"].column_names:
460
+ raise ValueError(
461
+ f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
462
+ "Make sure to set `--text_column_name` to the correct text column - one of "
463
+ f"{', '.join(raw_datasets['train'].column_names)}."
464
+ )
465
+
466
+ if data_args.max_train_samples is not None:
467
+ raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
468
+
469
+ if training_args.do_eval:
470
+ if os.path.isdir(data_args.dataset_name):
471
+ raw_datasets["eval"] = load_from_disk(f"{data_args.dataset_name}/{data_args.eval_split_name}")
472
+ else:
473
+ raw_datasets["eval"] = load_dataset(
474
+ data_args.dataset_name,
475
+ data_args.dataset_config_name,
476
+ split=data_args.eval_split_name,
477
+ use_auth_token=data_args.use_auth_token,
478
+ )
479
+
480
+ if data_args.max_eval_samples is not None:
481
+ raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
482
+
483
+ # 2. We remove some special characters from the datasets
484
+ # that make training complicated and do not help in transcribing the speech
485
+ # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
486
+ # that could be easily picked up by the model
487
+ chars_to_ignore_regex = (
488
+ f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
489
+ )
490
+ text_column_name = data_args.text_column_name
491
+
492
+ def remove_special_characters(batch):
493
+ if chars_to_ignore_regex is not None:
494
+ batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
495
+ else:
496
+ batch["target_text"] = batch[text_column_name].lower() + " "
497
+ # Unicode Normalization
498
+ batch["target_text"] = unicodedata.normalize('NFKC', batch["target_text"])
499
+ return batch
500
+
501
+ with training_args.main_process_first(desc="dataset map special characters removal"):
502
+ raw_datasets = raw_datasets.map(
503
+ remove_special_characters,
504
+ remove_columns=[text_column_name],
505
+ desc="remove special characters from datasets",
506
+ )
507
+
508
+ # save special tokens for tokenizer
509
+ word_delimiter_token = data_args.word_delimiter_token
510
+ unk_token = data_args.unk_token
511
+ pad_token = data_args.pad_token
512
+
513
+ # 3. Next, let's load the config as we might need it to create
514
+ # the tokenizer
515
+ # load config
516
+ config = AutoConfig.from_pretrained(
517
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
518
+ )
519
+
520
+ # 4. Next, if no tokenizer file is defined,
521
+ # we create the vocabulary of the model by extracting all unique characters from
522
+ # the training and evaluation datasets
523
+ # We need to make sure that only first rank saves vocabulary
524
+ # make sure all processes wait until vocab is created
525
+ tokenizer_name_or_path = model_args.tokenizer_name_or_path
526
+ tokenizer_kwargs = {}
527
+ if tokenizer_name_or_path is None:
528
+ # save vocab in training output dir
529
+ tokenizer_name_or_path = training_args.output_dir
530
+
531
+ vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
532
+
533
+ with training_args.main_process_first():
534
+ if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
535
+ os.remove(vocab_file)
536
+
537
+ with training_args.main_process_first(desc="dataset map vocabulary creation"):
538
+ if not os.path.isfile(vocab_file):
539
+ os.makedirs(tokenizer_name_or_path, exist_ok=True)
540
+ vocab_dict = create_vocabulary_from_data(
541
+ raw_datasets,
542
+ word_delimiter_token=word_delimiter_token,
543
+ unk_token=unk_token,
544
+ pad_token=pad_token,
545
+ )
546
+
547
+ # save vocab dict to be loaded into tokenizer
548
+ with open(vocab_file, "w") as file:
549
+ json.dump(vocab_dict, file)
550
+
551
+ # if tokenizer has just been created
552
+ # it is defined by `tokenizer_class` if present in config else by `model_type`
553
+ tokenizer_kwargs = {
554
+ "config": config if config.tokenizer_class is not None else None,
555
+ "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
556
+ "unk_token": unk_token,
557
+ "pad_token": pad_token,
558
+ "word_delimiter_token": word_delimiter_token,
559
+ }
560
+
561
+ # 5. Now we can instantiate the feature extractor, tokenizer and model
562
+ # Note for distributed training, the .from_pretrained methods guarantee that only
563
+ # one local process can concurrently download model & vocab.
564
+
565
+ # load feature_extractor and tokenizer
566
+ tokenizer = AutoTokenizer.from_pretrained(
567
+ tokenizer_name_or_path,
568
+ use_auth_token=data_args.use_auth_token,
569
+ **tokenizer_kwargs,
570
+ )
571
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
572
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
573
+ )
574
+
575
+ # adapt config
576
+ config.update(
577
+ {
578
+ "feat_proj_dropout": model_args.feat_proj_dropout,
579
+ "attention_dropout": model_args.attention_dropout,
580
+ "hidden_dropout": model_args.hidden_dropout,
581
+ "final_dropout": model_args.final_dropout,
582
+ "mask_time_prob": model_args.mask_time_prob,
583
+ "mask_time_length": model_args.mask_time_length,
584
+ "mask_feature_prob": model_args.mask_feature_prob,
585
+ "mask_feature_length": model_args.mask_feature_length,
586
+ "gradient_checkpointing": training_args.gradient_checkpointing,
587
+ "layerdrop": model_args.layerdrop,
588
+ "ctc_loss_reduction": model_args.ctc_loss_reduction,
589
+ "pad_token_id": tokenizer.pad_token_id,
590
+ "vocab_size": len(tokenizer),
591
+ "activation_dropout": model_args.activation_dropout,
592
+ }
593
+ )
594
+
595
+ # create model
596
+ model = AutoModelForCTC.from_pretrained(
597
+ model_args.model_name_or_path,
598
+ cache_dir=model_args.cache_dir,
599
+ config=config,
600
+ use_auth_token=data_args.use_auth_token,
601
+ )
602
+
603
+ # freeze encoder
604
+ if model_args.freeze_feature_encoder:
605
+ model.freeze_feature_encoder()
606
+
607
+ # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
608
+ # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
609
+ # so that we just need to set the correct target sampling rate and normalize the input
610
+ # via the `feature_extractor`
611
+
612
+ # make sure that dataset decodes audio with correct sampling rate
613
+ dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
614
+ if dataset_sampling_rate != feature_extractor.sampling_rate:
615
+ raw_datasets = raw_datasets.cast_column(
616
+ data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
617
+ )
618
+
619
+ # derive max & min input length for sample rate & max duration
620
+ max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
621
+ min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
622
+ audio_column_name = data_args.audio_column_name
623
+ num_workers = data_args.preprocessing_num_workers
624
+
625
+ # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
626
+ phoneme_language = data_args.phoneme_language
627
+
628
+ # Preprocessing the datasets.
629
+ # We need to read the audio files as arrays and tokenize the targets.
630
+ def prepare_dataset(batch):
631
+ # load audio
632
+ sample = batch[audio_column_name]
633
+
634
+ inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
635
+ batch["input_values"] = inputs.input_values[0]
636
+ batch["length"] = len(batch["input_values"])
637
+
638
+ # encode targets
639
+ additional_kwargs = {}
640
+ if phoneme_language is not None:
641
+ additional_kwargs["phonemizer_lang"] = phoneme_language
642
+
643
+ batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
644
+ return batch
645
+
646
+ with training_args.main_process_first(desc="dataset map preprocessing"):
647
+ vectorized_datasets = raw_datasets.map(
648
+ prepare_dataset,
649
+ remove_columns=next(iter(raw_datasets.values())).column_names,
650
+ num_proc=num_workers,
651
+ desc="preprocess datasets",
652
+ )
653
+
654
+ def is_audio_in_length_range(length):
655
+ return length > min_input_length and length < max_input_length
656
+
657
+ # filter data that is shorter than min_input_length
658
+ vectorized_datasets = vectorized_datasets.filter(
659
+ is_audio_in_length_range,
660
+ num_proc=num_workers,
661
+ input_columns=["length"],
662
+ )
663
+
664
+ # 7. Next, we can prepare the training.
665
+ # Let's use word error rate (WER) as our evaluation metric,
666
+ # instantiate a data collator and the trainer
667
+
668
+ # Define evaluation metrics during training, *i.e.* word error rate, character error rate
669
+ eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
670
+
671
+ # for large datasets it is advised to run the preprocessing on a
672
+ # single machine first with ``args.preprocessing_only`` since there will mostly likely
673
+ # be a timeout when running the script in distributed mode.
674
+ # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
675
+ # cached dataset
676
+ if data_args.preprocessing_only:
677
+ logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
678
+ return
679
+
680
+ def compute_metrics(pred):
681
+ pred_logits = pred.predictions
682
+ pred_ids = np.argmax(pred_logits, axis=-1)
683
+
684
+ pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
685
+
686
+ pred_str = tokenizer.batch_decode(pred_ids)
687
+ # we do not want to group tokens when computing the metrics
688
+ label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
689
+
690
+ metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
691
+
692
+ return metrics
693
+
694
+ # Now save everything to be able to create a single processor later
695
+ if is_main_process(training_args.local_rank):
696
+ # save feature extractor, tokenizer and config
697
+ feature_extractor.save_pretrained(training_args.output_dir)
698
+ tokenizer.save_pretrained(training_args.output_dir)
699
+ config.save_pretrained(training_args.output_dir)
700
+
701
+ try:
702
+ processor = AutoProcessor.from_pretrained(training_args.output_dir)
703
+ except (OSError, KeyError):
704
+ warnings.warn(
705
+ "Loading a processor from a feature extractor config that does not"
706
+ " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
707
+ " attribute to your `preprocessor_config.json` file to suppress this warning: "
708
+ " `'processor_class': 'Wav2Vec2Processor'`",
709
+ FutureWarning,
710
+ )
711
+ processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
712
+
713
+ # Instantiate custom data collator
714
+ data_collator = DataCollatorCTCWithPadding(processor=processor)
715
+
716
+ decay_parameters = get_parameter_names(model, [torch.nn.LayerNorm])
717
+ decay_parameters = [name for name in decay_parameters if "bias" not in name]
718
+ optimizer_grouped_parameters = [
719
+ {
720
+ "params": [p for n, p in model.named_parameters() if n in decay_parameters],
721
+ "weight_decay": training_args.weight_decay,
722
+ },
723
+ {
724
+ "params": [p for n, p in model.named_parameters() if n not in decay_parameters],
725
+ "weight_decay": 0.0,
726
+ },
727
+ ]
728
+ if extra_args.bnb:
729
+ optimizer = bnb.optim.Adam8bit(
730
+ params=optimizer_grouped_parameters,
731
+ lr=training_args.learning_rate,
732
+ betas=(training_args.adam_beta1, training_args.adam_beta2),
733
+ eps=training_args.adam_epsilon,
734
+ )
735
+ else:
736
+ optimizer = torch.optim.AdamW(
737
+ params=optimizer_grouped_parameters,
738
+ lr=training_args.learning_rate,
739
+ betas=(training_args.adam_beta1, training_args.adam_beta2),
740
+ eps=training_args.adam_epsilon,
741
+ )
742
+ if extra_args.tristage_sched:
743
+ scheduler = get_tri_stage_schedule(optimizer, training_args.max_steps)
744
+ else:
745
+ scheduler = None
746
+ optimizers = (optimizer, scheduler)
747
+
748
+ # Initialize Trainer
749
+ trainer = Trainer(
750
+ model=model,
751
+ data_collator=data_collator,
752
+ args=training_args,
753
+ compute_metrics=compute_metrics,
754
+ train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
755
+ eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
756
+ tokenizer=feature_extractor,
757
+ optimizers=optimizers,
758
+ )
759
+
760
+ # 8. Finally, we can start training
761
+
762
+ # Training
763
+ if training_args.do_train:
764
+
765
+ # use last checkpoint if exist
766
+ if last_checkpoint is not None:
767
+ checkpoint = last_checkpoint
768
+ elif os.path.isdir(model_args.model_name_or_path):
769
+ checkpoint = model_args.model_name_or_path
770
+ else:
771
+ checkpoint = None
772
+
773
+ train_result = trainer.train(resume_from_checkpoint=checkpoint)
774
+ trainer.save_model()
775
+
776
+ metrics = train_result.metrics
777
+ max_train_samples = (
778
+ data_args.max_train_samples
779
+ if data_args.max_train_samples is not None
780
+ else len(vectorized_datasets["train"])
781
+ )
782
+ metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
783
+
784
+ trainer.log_metrics("train", metrics)
785
+ trainer.save_metrics("train", metrics)
786
+ trainer.save_state()
787
+
788
+ # Evaluation
789
+ results = {}
790
+ if training_args.do_eval:
791
+ logger.info("*** Evaluate ***")
792
+ metrics = trainer.evaluate()
793
+ max_eval_samples = (
794
+ data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
795
+ )
796
+ metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
797
+
798
+ trainer.log_metrics("eval", metrics)
799
+ trainer.save_metrics("eval", metrics)
800
+
801
+ # Write model card and (optionally) push to hub
802
+ config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
803
+ kwargs = {
804
+ "finetuned_from": model_args.model_name_or_path,
805
+ "tasks": "speech-recognition",
806
+ "tags": ["automatic-speech-recognition", data_args.dataset_name],
807
+ "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
808
+ "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
809
+ }
810
+ if "common_voice" in data_args.dataset_name:
811
+ kwargs["language"] = config_name
812
+
813
+ if training_args.push_to_hub:
814
+ trainer.push_to_hub(**kwargs)
815
+ else:
816
+ trainer.create_model_card(**kwargs)
817
+
818
+ return results
819
+
820
+
821
+ if __name__ == "__main__":
822
+ main()
.ipynb_checkpoints/vocab-checkpoint.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"\u0611": 1, "\u0613": 2, "\u0614": 3, "\u0621": 4, "\u0622": 5, "\u0623": 6, "\u0624": 7, "\u0626": 8, "\u0627": 9, "\u0628": 10, "\u062a": 11, "\u062b": 12, "\u062c": 13, "\u062d": 14, "\u062e": 15, "\u062f": 16, "\u0630": 17, "\u0631": 18, "\u0632": 19, "\u0633": 20, "\u0634": 21, "\u0635": 22, "\u0636": 23, "\u0637": 24, "\u0638": 25, "\u0639": 26, "\u063a": 27, "\u0641": 28, "\u0642": 29, "\u0643": 30, "\u0644": 31, "\u0645": 32, "\u0646": 33, "\u0647": 34, "\u0648": 35, "\u0649": 36, "\u064a": 37, "\u064b": 38, "\u064e": 39, "\u064f": 40, "\u0650": 41, "\u0651": 42, "\u0652": 43, "\u0653": 44, "\u0654": 45, "\u0657": 46, "\u0670": 47, "\u0679": 48, "\u067e": 49, "\u0686": 50, "\u0688": 51, "\u0691": 52, "\u0698": 53, "\u06a9": 54, "\u06af": 55, "\u06ba": 56, "\u06be": 57, "\u06c1": 58, "\u06c2": 59, "\u06c3": 60, "\u06cc": 61, "\u06d2": 62, "\u06d3": 63, "\ufb68": 64, "\ufbad": 65, "\ufbaf": 66, "\ufbfe": 67, "\ufdf2": 68, "\ufdfa": 69, "\ufe97": 70, "\ufe98": 71, "\ufea9": 72, "\ufeb2": 73, "\ufee7": 74, "\ufeee": 75, "|": 0, "[UNK]": 76, "[PAD]": 77}
added_tokens.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"<s>": 65, "</s>": 66}
config.json ADDED
@@ -0,0 +1,107 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "facebook/wav2vec2-xls-r-300m",
3
+ "activation_dropout": 0.1,
4
+ "adapter_kernel_size": 3,
5
+ "adapter_stride": 2,
6
+ "add_adapter": false,
7
+ "apply_spec_augment": true,
8
+ "architectures": [
9
+ "Wav2Vec2ForCTC"
10
+ ],
11
+ "attention_dropout": 0.0,
12
+ "bos_token_id": 1,
13
+ "classifier_proj_size": 256,
14
+ "codevector_dim": 768,
15
+ "contrastive_logits_temperature": 0.1,
16
+ "conv_bias": true,
17
+ "conv_dim": [
18
+ 512,
19
+ 512,
20
+ 512,
21
+ 512,
22
+ 512,
23
+ 512,
24
+ 512
25
+ ],
26
+ "conv_kernel": [
27
+ 10,
28
+ 3,
29
+ 3,
30
+ 3,
31
+ 3,
32
+ 2,
33
+ 2
34
+ ],
35
+ "conv_stride": [
36
+ 5,
37
+ 2,
38
+ 2,
39
+ 2,
40
+ 2,
41
+ 2,
42
+ 2
43
+ ],
44
+ "ctc_loss_reduction": "mean",
45
+ "ctc_zero_infinity": false,
46
+ "diversity_loss_weight": 0.1,
47
+ "do_stable_layer_norm": true,
48
+ "eos_token_id": 2,
49
+ "feat_extract_activation": "gelu",
50
+ "feat_extract_dropout": 0.0,
51
+ "feat_extract_norm": "layer",
52
+ "feat_proj_dropout": 0.0,
53
+ "feat_quantizer_dropout": 0.0,
54
+ "final_dropout": 0.0,
55
+ "hidden_act": "gelu",
56
+ "hidden_dropout": 0.0,
57
+ "hidden_size": 1024,
58
+ "initializer_range": 0.02,
59
+ "intermediate_size": 4096,
60
+ "layer_norm_eps": 1e-05,
61
+ "layerdrop": 0.0,
62
+ "mask_feature_length": 64,
63
+ "mask_feature_min_masks": 0,
64
+ "mask_feature_prob": 0.25,
65
+ "mask_time_length": 10,
66
+ "mask_time_min_masks": 2,
67
+ "mask_time_prob": 0.75,
68
+ "model_type": "wav2vec2",
69
+ "num_adapter_layers": 3,
70
+ "num_attention_heads": 16,
71
+ "num_codevector_groups": 2,
72
+ "num_codevectors_per_group": 320,
73
+ "num_conv_pos_embedding_groups": 16,
74
+ "num_conv_pos_embeddings": 128,
75
+ "num_feat_extract_layers": 7,
76
+ "num_hidden_layers": 24,
77
+ "num_negatives": 100,
78
+ "output_hidden_size": 1024,
79
+ "pad_token_id": 64,
80
+ "proj_codevector_dim": 768,
81
+ "tdnn_dilation": [
82
+ 1,
83
+ 2,
84
+ 3,
85
+ 1,
86
+ 1
87
+ ],
88
+ "tdnn_dim": [
89
+ 512,
90
+ 512,
91
+ 512,
92
+ 512,
93
+ 1500
94
+ ],
95
+ "tdnn_kernel": [
96
+ 5,
97
+ 3,
98
+ 3,
99
+ 1,
100
+ 1
101
+ ],
102
+ "torch_dtype": "float32",
103
+ "transformers_version": "4.19.0.dev0",
104
+ "use_weighted_layer_sum": false,
105
+ "vocab_size": 67,
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 sha256:258a01d5578febdaf102bf5f3074ee153965e6713eba74d60a09ba58f339a111
3
+ size 1262173425
run.sh ADDED
@@ -0,0 +1,36 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ python run_speech_recognition_ctc.py \
2
+ --dataset_name="mozilla-foundation/common_voice_9_0" \
3
+ --model_name_or_path="facebook/wav2vec2-xls-r-300m" \
4
+ --dataset_config_name="ur" \
5
+ --output_dir="./" \
6
+ --overwrite_output_dir \
7
+ --max_steps="5108" \
8
+ --per_device_train_batch_size="64" \
9
+ --per_device_eval_batch_size="64" \
10
+ --gradient_accumulation_steps="2" \
11
+ --learning_rate="7.5e-5" \
12
+ --warmup_ratio="0.1" \
13
+ --length_column_name="input_length" \
14
+ --evaluation_strategy="steps" \
15
+ --text_column_name="sentence" \
16
+ --chars_to_ignore , ? . ! \- \; \: \" “ % ‘ ” � — ’ … – \| \& ؛ ، ٫ ؟ ۔ ٪ \` \
17
+ --save_steps="400" \
18
+ --eval_steps="400" \
19
+ --logging_steps="100" \
20
+ --layerdrop="0.0" \
21
+ --activation_dropout="0.1" \
22
+ --save_total_limit="1" \
23
+ --freeze_feature_encoder \
24
+ --feat_proj_dropout="0.0" \
25
+ --mask_time_prob="0.75" \
26
+ --mask_time_length="10" \
27
+ --mask_feature_prob="0.25" \
28
+ --mask_feature_length="64" \
29
+ --seed="42" \
30
+ --gradient_checkpointing \
31
+ --use_auth_token \
32
+ --fp16 \
33
+ --group_by_length \
34
+ --do_train --do_eval \
35
+ --bnb --tristage_sched \
36
+ --push_to_hub
run_speech_recognition_ctc.py ADDED
@@ -0,0 +1,822 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python
2
+ # coding=utf-8
3
+ # Copyright 2021 The HuggingFace Inc. team. All rights reserved.
4
+ #
5
+ # Licensed under the Apache License, Version 2.0 (the "License");
6
+ # you may not use this file except in compliance with the License.
7
+ # You may obtain a copy of the License at
8
+ #
9
+ # http://www.apache.org/licenses/LICENSE-2.0
10
+ #
11
+ # Unless required by applicable law or agreed to in writing, software
12
+ # distributed under the License is distributed on an "AS IS" BASIS,
13
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14
+ # See the License for the specific language governing permissions and
15
+
16
+ """ Fine-tuning a 🤗 Transformers CTC model for automatic speech recognition"""
17
+
18
+ import functools
19
+ import json
20
+ import logging
21
+ import os
22
+ import re
23
+ import sys
24
+ import unicodedata
25
+ import warnings
26
+ from dataclasses import dataclass, field
27
+ from typing import Dict, List, Optional, Union
28
+
29
+ import datasets
30
+ import numpy as np
31
+ import torch
32
+ from torch.optim.lr_scheduler import LambdaLR
33
+ from datasets import DatasetDict, load_dataset, load_metric, load_from_disk
34
+
35
+ import bitsandbytes as bnb
36
+ import transformers
37
+ from transformers import (
38
+ AutoConfig,
39
+ AutoFeatureExtractor,
40
+ AutoModelForCTC,
41
+ AutoProcessor,
42
+ AutoTokenizer,
43
+ HfArgumentParser,
44
+ Trainer,
45
+ TrainingArguments,
46
+ Wav2Vec2Processor,
47
+ set_seed,
48
+ )
49
+ from transformers.trainer_pt_utils import get_parameter_names
50
+ from transformers.trainer_utils import get_last_checkpoint, is_main_process
51
+ from transformers.utils import check_min_version
52
+ from transformers.utils.versions import require_version
53
+
54
+
55
+ # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
56
+ check_min_version("4.16.0.dev0")
57
+
58
+ require_version("datasets>=1.13.3", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
59
+
60
+
61
+ logger = logging.getLogger(__name__)
62
+
63
+
64
+ def list_field(default=None, metadata=None):
65
+ return field(default_factory=lambda: default, metadata=metadata)
66
+
67
+
68
+ @dataclass
69
+ class ModelArguments:
70
+ """
71
+ Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
72
+ """
73
+
74
+ model_name_or_path: str = field(
75
+ metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
76
+ )
77
+ tokenizer_name_or_path: Optional[str] = field(
78
+ default=None,
79
+ metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
80
+ )
81
+ cache_dir: Optional[str] = field(
82
+ default=None,
83
+ metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
84
+ )
85
+ freeze_feature_encoder: bool = field(
86
+ default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
87
+ )
88
+ attention_dropout: float = field(
89
+ default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
90
+ )
91
+ activation_dropout: float = field(
92
+ default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
93
+ )
94
+ feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
95
+ hidden_dropout: float = field(
96
+ default=0.0,
97
+ metadata={
98
+ "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
99
+ },
100
+ )
101
+ final_dropout: float = field(
102
+ default=0.0,
103
+ metadata={"help": "The dropout probability for the final projection layer."},
104
+ )
105
+ mask_time_prob: float = field(
106
+ default=0.05,
107
+ metadata={
108
+ "help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
109
+ "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
110
+ "vectors will be masked along the time axis."
111
+ },
112
+ )
113
+ mask_time_length: int = field(
114
+ default=10,
115
+ metadata={"help": "Length of vector span to mask along the time axis."},
116
+ )
117
+ mask_feature_prob: float = field(
118
+ default=0.0,
119
+ metadata={
120
+ "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
121
+ "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
122
+ },
123
+ )
124
+ mask_feature_length: int = field(
125
+ default=10,
126
+ metadata={"help": "Length of vector span to mask along the feature axis."},
127
+ )
128
+ layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
129
+ ctc_loss_reduction: Optional[str] = field(
130
+ default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
131
+ )
132
+
133
+
134
+ @dataclass
135
+ class DataTrainingArguments:
136
+ """
137
+ Arguments pertaining to what data we are going to input our model for training and eval.
138
+
139
+ Using `HfArgumentParser` we can turn this class
140
+ into argparse arguments to be able to specify them on
141
+ the command line.
142
+ """
143
+
144
+ dataset_name: str = field(
145
+ metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
146
+ )
147
+ dataset_config_name: str = field(
148
+ default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
149
+ )
150
+ train_split_name: str = field(
151
+ default="train+validation",
152
+ metadata={
153
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
154
+ },
155
+ )
156
+ eval_split_name: str = field(
157
+ default="test",
158
+ metadata={
159
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
160
+ },
161
+ )
162
+ audio_column_name: str = field(
163
+ default="audio",
164
+ metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
165
+ )
166
+ text_column_name: str = field(
167
+ default="text",
168
+ metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
169
+ )
170
+ overwrite_cache: bool = field(
171
+ default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
172
+ )
173
+ preprocessing_num_workers: Optional[int] = field(
174
+ default=None,
175
+ metadata={"help": "The number of processes to use for the preprocessing."},
176
+ )
177
+ max_train_samples: Optional[int] = field(
178
+ default=None,
179
+ metadata={
180
+ "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
181
+ "value if set."
182
+ },
183
+ )
184
+ max_eval_samples: Optional[int] = field(
185
+ default=None,
186
+ metadata={
187
+ "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
188
+ "value if set."
189
+ },
190
+ )
191
+ chars_to_ignore: Optional[List[str]] = list_field(
192
+ default=None,
193
+ metadata={"help": "A list of characters to remove from the transcripts."},
194
+ )
195
+ eval_metrics: List[str] = list_field(
196
+ default=["wer", "cer"],
197
+ metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
198
+ )
199
+ max_duration_in_seconds: float = field(
200
+ default=20.0,
201
+ metadata={
202
+ "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
203
+ },
204
+ )
205
+ min_duration_in_seconds: float = field(
206
+ default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
207
+ )
208
+ preprocessing_only: bool = field(
209
+ default=False,
210
+ metadata={
211
+ "help": "Whether to only do data preprocessing and skip training. "
212
+ "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
213
+ "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
214
+ "so that the cached datasets can consequently be loaded in distributed training"
215
+ },
216
+ )
217
+ use_auth_token: bool = field(
218
+ default=False,
219
+ metadata={
220
+ "help": "If :obj:`True`, will use the token generated when running"
221
+ ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
222
+ },
223
+ )
224
+ unk_token: str = field(
225
+ default="[UNK]",
226
+ metadata={"help": "The unk token for the tokenizer"},
227
+ )
228
+ pad_token: str = field(
229
+ default="[PAD]",
230
+ metadata={"help": "The padding token for the tokenizer"},
231
+ )
232
+ word_delimiter_token: str = field(
233
+ default="|",
234
+ metadata={"help": "The word delimiter token for the tokenizer"},
235
+ )
236
+ phoneme_language: Optional[str] = field(
237
+ default=None,
238
+ metadata={
239
+ "help": "The target language that should be used be"
240
+ " passed to the tokenizer for tokenization. Note that"
241
+ " this is only relevant if the model classifies the"
242
+ " input audio to a sequence of phoneme sequences."
243
+ },
244
+ )
245
+
246
+ @dataclass
247
+ class ExtraArguments:
248
+ "Additional training arguments"
249
+ bnb: bool = field(
250
+ default=False,
251
+ metadata = {"help":"If true uses 8bit Adam"}
252
+ )
253
+ tristage_sched: bool = field(
254
+ default=False,
255
+ metadata = {"help":"If true uses tristage LR scheduler (refer to XLS-R paper)"}
256
+ )
257
+ wandb_project: str = field(
258
+ default="",
259
+ metadata = {"help":"Name of wandb project to log into"}
260
+ )
261
+
262
+
263
+ @dataclass
264
+ class DataCollatorCTCWithPadding:
265
+ """
266
+ Data collator that will dynamically pad the inputs received.
267
+ Args:
268
+ processor (:class:`~transformers.AutoProcessor`)
269
+ The processor used for proccessing the data.
270
+ padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
271
+ Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
272
+ among:
273
+ * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
274
+ sequence if provided).
275
+ * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
276
+ maximum acceptable input length for the model if that argument is not provided.
277
+ * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
278
+ different lengths).
279
+ max_length (:obj:`int`, `optional`):
280
+ Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
281
+ max_length_labels (:obj:`int`, `optional`):
282
+ Maximum length of the ``labels`` returned list and optionally padding length (see above).
283
+ pad_to_multiple_of (:obj:`int`, `optional`):
284
+ If set will pad the sequence to a multiple of the provided value.
285
+ This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
286
+ 7.5 (Volta).
287
+ """
288
+
289
+ processor: AutoProcessor
290
+ padding: Union[bool, str] = "longest"
291
+ pad_to_multiple_of: Optional[int] = None
292
+ pad_to_multiple_of_labels: Optional[int] = None
293
+
294
+ def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
295
+ # split inputs and labels since they have to be of different lenghts and need
296
+ # different padding methods
297
+ input_features = [{"input_values": feature["input_values"]} for feature in features]
298
+ label_features = [{"input_ids": feature["labels"]} for feature in features]
299
+
300
+ batch = self.processor.pad(
301
+ input_features,
302
+ padding=self.padding,
303
+ pad_to_multiple_of=self.pad_to_multiple_of,
304
+ return_tensors="pt",
305
+ )
306
+
307
+ with self.processor.as_target_processor():
308
+ labels_batch = self.processor.pad(
309
+ label_features,
310
+ padding=self.padding,
311
+ pad_to_multiple_of=self.pad_to_multiple_of_labels,
312
+ return_tensors="pt",
313
+ )
314
+
315
+ # replace padding with -100 to ignore loss correctly
316
+ labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
317
+
318
+ batch["labels"] = labels
319
+
320
+ return batch
321
+
322
+
323
+ def get_tri_stage_schedule(
324
+ optimizer, num_training_steps, ratios=[0.1, 0.4, 0.5], num_warmup_steps=None, num_hold_steps=None, start_ratio=0.01, end_ratio=0.05
325
+ ):
326
+ assert (num_warmup_steps is None) == (num_hold_steps is None)
327
+ if num_warmup_steps is None:
328
+ num_warmup_steps = int(ratios[0]*num_training_steps)
329
+ num_hold_steps = int(ratios[1]*num_training_steps)
330
+ start_decay_step = num_warmup_steps + num_hold_steps
331
+ a_w, b_w = (1-start_ratio)/num_warmup_steps, start_ratio
332
+ num_decay_steps = num_training_steps - start_decay_step
333
+ a_d, b_d = (end_ratio-1)/num_decay_steps, 1.
334
+
335
+ def lr_lambda(current_step):
336
+ if current_step < num_warmup_steps:
337
+ return a_w * float(current_step) + b_w
338
+ if current_step < start_decay_step:
339
+ return 1.
340
+ return max(end_ratio, a_d * float(current_step - start_decay_step) + b_d )
341
+
342
+ return LambdaLR(optimizer, lr_lambda)
343
+
344
+ def create_vocabulary_from_data(
345
+ datasets: DatasetDict,
346
+ word_delimiter_token: Optional[str] = None,
347
+ unk_token: Optional[str] = None,
348
+ pad_token: Optional[str] = None,
349
+ ):
350
+ # Given training and test labels create vocabulary
351
+ def extract_all_chars(batch):
352
+ all_text = " ".join(batch["target_text"])
353
+ vocab = list(set(all_text))
354
+ return {"vocab": [vocab], "all_text": [all_text]}
355
+
356
+ vocabs = datasets.map(
357
+ extract_all_chars,
358
+ batched=True,
359
+ batch_size=-1,
360
+ keep_in_memory=True,
361
+ remove_columns=datasets["train"].column_names,
362
+ )
363
+
364
+ # take union of all unique characters in each dataset
365
+ vocab_set = functools.reduce(
366
+ lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
367
+ )
368
+
369
+ vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
370
+
371
+ # replace white space with delimiter token
372
+ if word_delimiter_token is not None:
373
+ vocab_dict[word_delimiter_token] = vocab_dict[" "]
374
+ del vocab_dict[" "]
375
+
376
+ # add unk and pad token
377
+ if unk_token is not None:
378
+ vocab_dict[unk_token] = len(vocab_dict)
379
+
380
+ if pad_token is not None:
381
+ vocab_dict[pad_token] = len(vocab_dict)
382
+
383
+ return vocab_dict
384
+
385
+
386
+ def main():
387
+ # See all possible arguments in src/transformers/training_args.py
388
+ # or by passing the --help flag to this script.
389
+ # We now keep distinct sets of args, for a cleaner separation of concerns.
390
+
391
+ parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments, ExtraArguments))
392
+ if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
393
+ # If we pass only one argument to the script and it's the path to a json file,
394
+ # let's parse it to get our arguments.
395
+ model_args, data_args, training_args, extra_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
396
+ else:
397
+ model_args, data_args, training_args, extra_args = parser.parse_args_into_dataclasses()
398
+
399
+ # Detecting last checkpoint.
400
+ last_checkpoint = None
401
+ if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
402
+ last_checkpoint = get_last_checkpoint(training_args.output_dir)
403
+ if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
404
+ raise ValueError(
405
+ f"Output directory ({training_args.output_dir}) already exists and is not empty. "
406
+ "Use --overwrite_output_dir to overcome."
407
+ )
408
+ elif last_checkpoint is not None:
409
+ logger.info(
410
+ f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
411
+ "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
412
+ )
413
+
414
+ # Setup logging
415
+ logging.basicConfig(
416
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
417
+ datefmt="%m/%d/%Y %H:%M:%S",
418
+ handlers=[logging.StreamHandler(sys.stdout)],
419
+ )
420
+ logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
421
+
422
+ # Log on each process the small summary:
423
+ logger.warning(
424
+ f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
425
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
426
+ )
427
+ # Set the verbosity to info of the Transformers logger (on main process only):
428
+ if is_main_process(training_args.local_rank):
429
+ transformers.utils.logging.set_verbosity_info()
430
+ logger.info("Training/evaluation parameters %s", training_args)
431
+
432
+ # Set seed before initializing model.
433
+ set_seed(training_args.seed)
434
+
435
+ # configure wandb run
436
+ os.environ["WANDB_PROJECT"] = extra_args.wandb_project
437
+
438
+ # 1. First, let's load the dataset
439
+ raw_datasets = DatasetDict()
440
+
441
+ if training_args.do_train:
442
+ if os.path.isdir(data_args.dataset_name):
443
+ raw_datasets["train"] = load_from_disk(f"{data_args.dataset_name}/{data_args.train_split_name}")
444
+ else:
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
+ )
451
+
452
+ if data_args.audio_column_name not in raw_datasets["train"].column_names:
453
+ raise ValueError(
454
+ f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
455
+ "Make sure to set `--audio_column_name` to the correct audio column - one of "
456
+ f"{', '.join(raw_datasets['train'].column_names)}."
457
+ )
458
+
459
+ if data_args.text_column_name not in raw_datasets["train"].column_names:
460
+ raise ValueError(
461
+ f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
462
+ "Make sure to set `--text_column_name` to the correct text column - one of "
463
+ f"{', '.join(raw_datasets['train'].column_names)}."
464
+ )
465
+
466
+ if data_args.max_train_samples is not None:
467
+ raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
468
+
469
+ if training_args.do_eval:
470
+ if os.path.isdir(data_args.dataset_name):
471
+ raw_datasets["eval"] = load_from_disk(f"{data_args.dataset_name}/{data_args.eval_split_name}")
472
+ else:
473
+ raw_datasets["eval"] = load_dataset(
474
+ data_args.dataset_name,
475
+ data_args.dataset_config_name,
476
+ split=data_args.eval_split_name,
477
+ use_auth_token=data_args.use_auth_token,
478
+ )
479
+
480
+ if data_args.max_eval_samples is not None:
481
+ raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
482
+
483
+ # 2. We remove some special characters from the datasets
484
+ # that make training complicated and do not help in transcribing the speech
485
+ # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
486
+ # that could be easily picked up by the model
487
+ chars_to_ignore_regex = (
488
+ f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
489
+ )
490
+ text_column_name = data_args.text_column_name
491
+
492
+ def remove_special_characters(batch):
493
+ if chars_to_ignore_regex is not None:
494
+ batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
495
+ else:
496
+ batch["target_text"] = batch[text_column_name].lower() + " "
497
+ # Unicode Normalization
498
+ batch["target_text"] = unicodedata.normalize('NFKC', batch["target_text"])
499
+ return batch
500
+
501
+ with training_args.main_process_first(desc="dataset map special characters removal"):
502
+ raw_datasets = raw_datasets.map(
503
+ remove_special_characters,
504
+ remove_columns=[text_column_name],
505
+ desc="remove special characters from datasets",
506
+ )
507
+
508
+ # save special tokens for tokenizer
509
+ word_delimiter_token = data_args.word_delimiter_token
510
+ unk_token = data_args.unk_token
511
+ pad_token = data_args.pad_token
512
+
513
+ # 3. Next, let's load the config as we might need it to create
514
+ # the tokenizer
515
+ # load config
516
+ config = AutoConfig.from_pretrained(
517
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
518
+ )
519
+
520
+ # 4. Next, if no tokenizer file is defined,
521
+ # we create the vocabulary of the model by extracting all unique characters from
522
+ # the training and evaluation datasets
523
+ # We need to make sure that only first rank saves vocabulary
524
+ # make sure all processes wait until vocab is created
525
+ tokenizer_name_or_path = model_args.tokenizer_name_or_path
526
+ tokenizer_kwargs = {}
527
+ if tokenizer_name_or_path is None:
528
+ # save vocab in training output dir
529
+ tokenizer_name_or_path = training_args.output_dir
530
+
531
+ vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
532
+
533
+ with training_args.main_process_first():
534
+ if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
535
+ os.remove(vocab_file)
536
+
537
+ with training_args.main_process_first(desc="dataset map vocabulary creation"):
538
+ if not os.path.isfile(vocab_file):
539
+ os.makedirs(tokenizer_name_or_path, exist_ok=True)
540
+ vocab_dict = create_vocabulary_from_data(
541
+ raw_datasets,
542
+ word_delimiter_token=word_delimiter_token,
543
+ unk_token=unk_token,
544
+ pad_token=pad_token,
545
+ )
546
+
547
+ # save vocab dict to be loaded into tokenizer
548
+ with open(vocab_file, "w") as file:
549
+ json.dump(vocab_dict, file)
550
+
551
+ # if tokenizer has just been created
552
+ # it is defined by `tokenizer_class` if present in config else by `model_type`
553
+ tokenizer_kwargs = {
554
+ "config": config if config.tokenizer_class is not None else None,
555
+ "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
556
+ "unk_token": unk_token,
557
+ "pad_token": pad_token,
558
+ "word_delimiter_token": word_delimiter_token,
559
+ }
560
+
561
+ # 5. Now we can instantiate the feature extractor, tokenizer and model
562
+ # Note for distributed training, the .from_pretrained methods guarantee that only
563
+ # one local process can concurrently download model & vocab.
564
+
565
+ # load feature_extractor and tokenizer
566
+ tokenizer = AutoTokenizer.from_pretrained(
567
+ tokenizer_name_or_path,
568
+ use_auth_token=data_args.use_auth_token,
569
+ **tokenizer_kwargs,
570
+ )
571
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
572
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
573
+ )
574
+
575
+ # adapt config
576
+ config.update(
577
+ {
578
+ "feat_proj_dropout": model_args.feat_proj_dropout,
579
+ "attention_dropout": model_args.attention_dropout,
580
+ "hidden_dropout": model_args.hidden_dropout,
581
+ "final_dropout": model_args.final_dropout,
582
+ "mask_time_prob": model_args.mask_time_prob,
583
+ "mask_time_length": model_args.mask_time_length,
584
+ "mask_feature_prob": model_args.mask_feature_prob,
585
+ "mask_feature_length": model_args.mask_feature_length,
586
+ "gradient_checkpointing": training_args.gradient_checkpointing,
587
+ "layerdrop": model_args.layerdrop,
588
+ "ctc_loss_reduction": model_args.ctc_loss_reduction,
589
+ "pad_token_id": tokenizer.pad_token_id,
590
+ "vocab_size": len(tokenizer),
591
+ "activation_dropout": model_args.activation_dropout,
592
+ }
593
+ )
594
+
595
+ # create model
596
+ model = AutoModelForCTC.from_pretrained(
597
+ model_args.model_name_or_path,
598
+ cache_dir=model_args.cache_dir,
599
+ config=config,
600
+ use_auth_token=data_args.use_auth_token,
601
+ )
602
+
603
+ # freeze encoder
604
+ if model_args.freeze_feature_encoder:
605
+ model.freeze_feature_encoder()
606
+
607
+ # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
608
+ # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
609
+ # so that we just need to set the correct target sampling rate and normalize the input
610
+ # via the `feature_extractor`
611
+
612
+ # make sure that dataset decodes audio with correct sampling rate
613
+ dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
614
+ if dataset_sampling_rate != feature_extractor.sampling_rate:
615
+ raw_datasets = raw_datasets.cast_column(
616
+ data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
617
+ )
618
+
619
+ # derive max & min input length for sample rate & max duration
620
+ max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
621
+ min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
622
+ audio_column_name = data_args.audio_column_name
623
+ num_workers = data_args.preprocessing_num_workers
624
+
625
+ # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
626
+ phoneme_language = data_args.phoneme_language
627
+
628
+ # Preprocessing the datasets.
629
+ # We need to read the audio files as arrays and tokenize the targets.
630
+ def prepare_dataset(batch):
631
+ # load audio
632
+ sample = batch[audio_column_name]
633
+
634
+ inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
635
+ batch["input_values"] = inputs.input_values[0]
636
+ batch["length"] = len(batch["input_values"])
637
+
638
+ # encode targets
639
+ additional_kwargs = {}
640
+ if phoneme_language is not None:
641
+ additional_kwargs["phonemizer_lang"] = phoneme_language
642
+
643
+ batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
644
+ return batch
645
+
646
+ with training_args.main_process_first(desc="dataset map preprocessing"):
647
+ vectorized_datasets = raw_datasets.map(
648
+ prepare_dataset,
649
+ remove_columns=next(iter(raw_datasets.values())).column_names,
650
+ num_proc=num_workers,
651
+ desc="preprocess datasets",
652
+ )
653
+
654
+ def is_audio_in_length_range(length):
655
+ return length > min_input_length and length < max_input_length
656
+
657
+ # filter data that is shorter than min_input_length
658
+ vectorized_datasets = vectorized_datasets.filter(
659
+ is_audio_in_length_range,
660
+ num_proc=num_workers,
661
+ input_columns=["length"],
662
+ )
663
+
664
+ # 7. Next, we can prepare the training.
665
+ # Let's use word error rate (WER) as our evaluation metric,
666
+ # instantiate a data collator and the trainer
667
+
668
+ # Define evaluation metrics during training, *i.e.* word error rate, character error rate
669
+ eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
670
+
671
+ # for large datasets it is advised to run the preprocessing on a
672
+ # single machine first with ``args.preprocessing_only`` since there will mostly likely
673
+ # be a timeout when running the script in distributed mode.
674
+ # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
675
+ # cached dataset
676
+ if data_args.preprocessing_only:
677
+ logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
678
+ return
679
+
680
+ def compute_metrics(pred):
681
+ pred_logits = pred.predictions
682
+ pred_ids = np.argmax(pred_logits, axis=-1)
683
+
684
+ pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
685
+
686
+ pred_str = tokenizer.batch_decode(pred_ids)
687
+ # we do not want to group tokens when computing the metrics
688
+ label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
689
+
690
+ metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
691
+
692
+ return metrics
693
+
694
+ # Now save everything to be able to create a single processor later
695
+ if is_main_process(training_args.local_rank):
696
+ # save feature extractor, tokenizer and config
697
+ feature_extractor.save_pretrained(training_args.output_dir)
698
+ tokenizer.save_pretrained(training_args.output_dir)
699
+ config.save_pretrained(training_args.output_dir)
700
+
701
+ try:
702
+ processor = AutoProcessor.from_pretrained(training_args.output_dir)
703
+ except (OSError, KeyError):
704
+ warnings.warn(
705
+ "Loading a processor from a feature extractor config that does not"
706
+ " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
707
+ " attribute to your `preprocessor_config.json` file to suppress this warning: "
708
+ " `'processor_class': 'Wav2Vec2Processor'`",
709
+ FutureWarning,
710
+ )
711
+ processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
712
+
713
+ # Instantiate custom data collator
714
+ data_collator = DataCollatorCTCWithPadding(processor=processor)
715
+
716
+ decay_parameters = get_parameter_names(model, [torch.nn.LayerNorm])
717
+ decay_parameters = [name for name in decay_parameters if "bias" not in name]
718
+ optimizer_grouped_parameters = [
719
+ {
720
+ "params": [p for n, p in model.named_parameters() if n in decay_parameters],
721
+ "weight_decay": training_args.weight_decay,
722
+ },
723
+ {
724
+ "params": [p for n, p in model.named_parameters() if n not in decay_parameters],
725
+ "weight_decay": 0.0,
726
+ },
727
+ ]
728
+ if extra_args.bnb:
729
+ optimizer = bnb.optim.Adam8bit(
730
+ params=optimizer_grouped_parameters,
731
+ lr=training_args.learning_rate,
732
+ betas=(training_args.adam_beta1, training_args.adam_beta2),
733
+ eps=training_args.adam_epsilon,
734
+ )
735
+ else:
736
+ optimizer = torch.optim.AdamW(
737
+ params=optimizer_grouped_parameters,
738
+ lr=training_args.learning_rate,
739
+ betas=(training_args.adam_beta1, training_args.adam_beta2),
740
+ eps=training_args.adam_epsilon,
741
+ )
742
+ if extra_args.tristage_sched:
743
+ scheduler = get_tri_stage_schedule(optimizer, training_args.max_steps)
744
+ else:
745
+ scheduler = None
746
+ optimizers = (optimizer, scheduler)
747
+
748
+ # Initialize Trainer
749
+ trainer = Trainer(
750
+ model=model,
751
+ data_collator=data_collator,
752
+ args=training_args,
753
+ compute_metrics=compute_metrics,
754
+ train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
755
+ eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
756
+ tokenizer=feature_extractor,
757
+ optimizers=optimizers,
758
+ )
759
+
760
+ # 8. Finally, we can start training
761
+
762
+ # Training
763
+ if training_args.do_train:
764
+
765
+ # use last checkpoint if exist
766
+ if last_checkpoint is not None:
767
+ checkpoint = last_checkpoint
768
+ elif os.path.isdir(model_args.model_name_or_path):
769
+ checkpoint = model_args.model_name_or_path
770
+ else:
771
+ checkpoint = None
772
+
773
+ train_result = trainer.train(resume_from_checkpoint=checkpoint)
774
+ trainer.save_model()
775
+
776
+ metrics = train_result.metrics
777
+ max_train_samples = (
778
+ data_args.max_train_samples
779
+ if data_args.max_train_samples is not None
780
+ else len(vectorized_datasets["train"])
781
+ )
782
+ metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
783
+
784
+ trainer.log_metrics("train", metrics)
785
+ trainer.save_metrics("train", metrics)
786
+ trainer.save_state()
787
+
788
+ # Evaluation
789
+ results = {}
790
+ if training_args.do_eval:
791
+ logger.info("*** Evaluate ***")
792
+ metrics = trainer.evaluate()
793
+ max_eval_samples = (
794
+ data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
795
+ )
796
+ metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
797
+
798
+ trainer.log_metrics("eval", metrics)
799
+ trainer.save_metrics("eval", metrics)
800
+
801
+ # Write model card and (optionally) push to hub
802
+ config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
803
+ kwargs = {
804
+ "finetuned_from": model_args.model_name_or_path,
805
+ "tasks": "speech-recognition",
806
+ "tags": ["automatic-speech-recognition", data_args.dataset_name],
807
+ "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
808
+ "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
809
+ }
810
+ if "common_voice" in data_args.dataset_name:
811
+ kwargs["language"] = config_name
812
+
813
+ if training_args.push_to_hub:
814
+ trainer.push_to_hub(**kwargs)
815
+ else:
816
+ trainer.create_model_card(**kwargs)
817
+
818
+ return results
819
+
820
+
821
+ if __name__ == "__main__":
822
+ main()
special_tokens_map.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"bos_token": "<s>", "eos_token": "</s>", "unk_token": "[UNK]", "pad_token": "[PAD]", "additional_special_tokens": [{"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}]}
tokenizer_config.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"unk_token": "[UNK]", "bos_token": "<s>", "eos_token": "</s>", "pad_token": "[PAD]", "do_lower_case": false, "word_delimiter_token": "|", "replace_word_delimiter_char": " ", "special_tokens_map_file": null, "name_or_path": "./", "tokenizer_class": "Wav2Vec2CTCTokenizer"}
training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a1656821df254a54225c39804dffd317a7a88c13126c9ac2a3bc0800ab2f6ee6
3
+ size 3119
vocab.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"ؑ": 1, "ؓ": 2, "ؔ": 3, "ء": 4, "آ": 5, "أ": 6, "ؤ": 7, "ئ": 8, "ا": 9, "ب": 10, "ت": 11, "ث": 12, "ج": 13, "ح": 14, "خ": 15, "د": 16, "ذ": 17, "ر": 18, "ز": 19, "س": 20, "ش": 21, "ص": 22, "ض": 23, "ط": 24, "ظ": 25, "ع": 26, "غ": 27, "ف": 28, "ق": 29, "ك": 30, "ل": 31, "م": 32, "ن": 33, "ه": 34, "و": 35, "ى": 36, "ي": 37, "ً": 38, "َ": 39, "ُ": 40, "ِ": 41, "ّ": 42, "ْ": 43, "ٔ": 44, "ٗ": 45, "ٰ": 46, "ٹ": 47, "پ": 48, "چ": 49, "ڈ": 50, "ڑ": 51, "ژ": 52, "ک": 53, "گ": 54, "ں": 55, "ھ": 56, "ہ": 57, "ۂ": 58, "ۃ": 59, "ی": 60, "ے": 61, "ۓ": 62, "|": 0, "[UNK]": 63, "[PAD]": 64}