RuudVelo commited on
Commit
ecbad0d
1 Parent(s): 6dc72d1
.gitignore ADDED
@@ -0,0 +1 @@
 
 
1
+ checkpoint-*/
added_tokens.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"<s>": 50, "</s>": 51}
config.json ADDED
@@ -0,0 +1,107 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "facebook/wav2vec2-xls-r-1b",
3
+ "activation_dropout": 0.1,
4
+ "adapter_kernel_size": 3,
5
+ "adapter_stride": 2,
6
+ "add_adapter": false,
7
+ "apply_spec_augment": true,
8
+ "architectures": [
9
+ "Wav2Vec2ForCTC"
10
+ ],
11
+ "attention_dropout": 0.0,
12
+ "bos_token_id": 1,
13
+ "classifier_proj_size": 256,
14
+ "codevector_dim": 1024,
15
+ "contrastive_logits_temperature": 0.1,
16
+ "conv_bias": true,
17
+ "conv_dim": [
18
+ 512,
19
+ 512,
20
+ 512,
21
+ 512,
22
+ 512,
23
+ 512,
24
+ 512
25
+ ],
26
+ "conv_kernel": [
27
+ 10,
28
+ 3,
29
+ 3,
30
+ 3,
31
+ 3,
32
+ 2,
33
+ 2
34
+ ],
35
+ "conv_stride": [
36
+ 5,
37
+ 2,
38
+ 2,
39
+ 2,
40
+ 2,
41
+ 2,
42
+ 2
43
+ ],
44
+ "ctc_loss_reduction": "mean",
45
+ "ctc_zero_infinity": false,
46
+ "diversity_loss_weight": 0.1,
47
+ "do_stable_layer_norm": true,
48
+ "eos_token_id": 2,
49
+ "feat_extract_activation": "gelu",
50
+ "feat_extract_dropout": 0.0,
51
+ "feat_extract_norm": "layer",
52
+ "feat_proj_dropout": 0.0,
53
+ "feat_quantizer_dropout": 0.0,
54
+ "final_dropout": 0.0,
55
+ "hidden_act": "gelu",
56
+ "hidden_dropout": 0.0,
57
+ "hidden_size": 1280,
58
+ "initializer_range": 0.02,
59
+ "intermediate_size": 5120,
60
+ "layer_norm_eps": 1e-05,
61
+ "layerdrop": 0.0,
62
+ "mask_feature_length": 64,
63
+ "mask_feature_min_masks": 0,
64
+ "mask_feature_prob": 0.25,
65
+ "mask_time_length": 10,
66
+ "mask_time_min_masks": 2,
67
+ "mask_time_prob": 0.75,
68
+ "model_type": "wav2vec2",
69
+ "num_adapter_layers": 3,
70
+ "num_attention_heads": 16,
71
+ "num_codevector_groups": 2,
72
+ "num_codevectors_per_group": 320,
73
+ "num_conv_pos_embedding_groups": 16,
74
+ "num_conv_pos_embeddings": 128,
75
+ "num_feat_extract_layers": 7,
76
+ "num_hidden_layers": 48,
77
+ "num_negatives": 100,
78
+ "output_hidden_size": 1280,
79
+ "pad_token_id": 49,
80
+ "proj_codevector_dim": 1024,
81
+ "tdnn_dilation": [
82
+ 1,
83
+ 2,
84
+ 3,
85
+ 1,
86
+ 1
87
+ ],
88
+ "tdnn_dim": [
89
+ 512,
90
+ 512,
91
+ 512,
92
+ 512,
93
+ 1500
94
+ ],
95
+ "tdnn_kernel": [
96
+ 5,
97
+ 3,
98
+ 3,
99
+ 1,
100
+ 1
101
+ ],
102
+ "torch_dtype": "float32",
103
+ "transformers_version": "4.17.0.dev0",
104
+ "use_weighted_layer_sum": false,
105
+ "vocab_size": 52,
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
+ }
run_speech_recognition_ctc.py ADDED
@@ -0,0 +1,736 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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("4.17.0.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
+
129
+
130
+ @dataclass
131
+ class DataTrainingArguments:
132
+ """
133
+ Arguments pertaining to what data we are going to input our model for training and eval.
134
+ Using `HfArgumentParser` we can turn this class
135
+ into argparse arguments to be able to specify them on
136
+ the command line.
137
+ """
138
+
139
+ dataset_name: str = field(
140
+ metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
141
+ )
142
+ dataset_config_name: str = field(
143
+ default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
144
+ )
145
+ train_split_name: str = field(
146
+ default="train+validation",
147
+ metadata={
148
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
149
+ },
150
+ )
151
+ eval_split_name: str = field(
152
+ default="test",
153
+ metadata={
154
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'test'"
155
+ },
156
+ )
157
+ audio_column_name: str = field(
158
+ default="audio",
159
+ metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
160
+ )
161
+ text_column_name: str = field(
162
+ default="text",
163
+ metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
164
+ )
165
+ overwrite_cache: bool = field(
166
+ default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
167
+ )
168
+ preprocessing_num_workers: Optional[int] = field(
169
+ default=None,
170
+ metadata={"help": "The number of processes to use for the preprocessing."},
171
+ )
172
+ max_train_samples: Optional[int] = field(
173
+ default=None,
174
+ metadata={
175
+ "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
176
+ "value if set."
177
+ },
178
+ )
179
+ max_eval_samples: Optional[int] = field(
180
+ default=None,
181
+ metadata={
182
+ "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
183
+ "value if set."
184
+ },
185
+ )
186
+ chars_to_ignore: Optional[List[str]] = list_field(
187
+ default=None,
188
+ metadata={"help": "A list of characters to remove from the transcripts."},
189
+ )
190
+ eval_metrics: List[str] = list_field(
191
+ default=["wer"],
192
+ metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
193
+ )
194
+ max_duration_in_seconds: float = field(
195
+ default=20.0,
196
+ metadata={
197
+ "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
198
+ },
199
+ )
200
+ min_duration_in_seconds: float = field(
201
+ default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
202
+ )
203
+ preprocessing_only: bool = field(
204
+ default=False,
205
+ metadata={
206
+ "help": "Whether to only do data preprocessing and skip training. "
207
+ "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
208
+ "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
209
+ "so that the cached datasets can consequently be loaded in distributed training"
210
+ },
211
+ )
212
+ use_auth_token: bool = field(
213
+ default=False,
214
+ metadata={
215
+ "help": "If :obj:`True`, will use the token generated when running"
216
+ ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
217
+ },
218
+ )
219
+ unk_token: str = field(
220
+ default="[UNK]",
221
+ metadata={"help": "The unk token for the tokenizer"},
222
+ )
223
+ pad_token: str = field(
224
+ default="[PAD]",
225
+ metadata={"help": "The padding token for the tokenizer"},
226
+ )
227
+ word_delimiter_token: str = field(
228
+ default="|",
229
+ metadata={"help": "The word delimiter token for the tokenizer"},
230
+ )
231
+ phoneme_language: Optional[str] = field(
232
+ default=None,
233
+ metadata={
234
+ "help": "The target language that should be used be"
235
+ " passed to the tokenizer for tokenization. Note that"
236
+ " this is only relevant if the model classifies the"
237
+ " input audio to a sequence of phoneme sequences."
238
+ },
239
+ )
240
+
241
+
242
+ @dataclass
243
+ class DataCollatorCTCWithPadding:
244
+ """
245
+ Data collator that will dynamically pad the inputs received.
246
+ Args:
247
+ processor (:class:`~transformers.AutoProcessor`)
248
+ The processor used for proccessing the data.
249
+ padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
250
+ Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
251
+ among:
252
+ * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
253
+ sequence if provided).
254
+ * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
255
+ maximum acceptable input length for the model if that argument is not provided.
256
+ * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
257
+ different lengths).
258
+ max_length (:obj:`int`, `optional`):
259
+ Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
260
+ max_length_labels (:obj:`int`, `optional`):
261
+ Maximum length of the ``labels`` returned list and optionally padding length (see above).
262
+ pad_to_multiple_of (:obj:`int`, `optional`):
263
+ If set will pad the sequence to a multiple of the provided value.
264
+ This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
265
+ 7.5 (Volta).
266
+ """
267
+
268
+ processor: AutoProcessor
269
+ padding: Union[bool, str] = "longest"
270
+ pad_to_multiple_of: Optional[int] = None
271
+ pad_to_multiple_of_labels: Optional[int] = None
272
+
273
+ def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
274
+ # split inputs and labels since they have to be of different lenghts and need
275
+ # different padding methods
276
+ input_features = [{"input_values": feature["input_values"]} for feature in features]
277
+ label_features = [{"input_ids": feature["labels"]} for feature in features]
278
+
279
+ batch = self.processor.pad(
280
+ input_features,
281
+ padding=self.padding,
282
+ pad_to_multiple_of=self.pad_to_multiple_of,
283
+ return_tensors="pt",
284
+ )
285
+
286
+ with self.processor.as_target_processor():
287
+ labels_batch = self.processor.pad(
288
+ label_features,
289
+ padding=self.padding,
290
+ pad_to_multiple_of=self.pad_to_multiple_of_labels,
291
+ return_tensors="pt",
292
+ )
293
+
294
+ # replace padding with -100 to ignore loss correctly
295
+ labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
296
+
297
+ batch["labels"] = labels
298
+
299
+ return batch
300
+
301
+
302
+ def create_vocabulary_from_data(
303
+ datasets: DatasetDict,
304
+ word_delimiter_token: Optional[str] = None,
305
+ unk_token: Optional[str] = None,
306
+ pad_token: Optional[str] = None,
307
+ ):
308
+ # Given training and test labels create vocabulary
309
+ def extract_all_chars(batch):
310
+ all_text = " ".join(batch["target_text"])
311
+ vocab = list(set(all_text))
312
+ return {"vocab": [vocab], "all_text": [all_text]}
313
+
314
+ vocabs = datasets.map(
315
+ extract_all_chars,
316
+ batched=True,
317
+ batch_size=-1,
318
+ keep_in_memory=True,
319
+ remove_columns=datasets["train"].column_names,
320
+ )
321
+
322
+ # take union of all unique characters in each dataset
323
+ vocab_set = functools.reduce(
324
+ lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
325
+ )
326
+
327
+ vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
328
+
329
+ # replace white space with delimiter token
330
+ if word_delimiter_token is not None:
331
+ vocab_dict[word_delimiter_token] = vocab_dict[" "]
332
+ del vocab_dict[" "]
333
+
334
+ # add unk and pad token
335
+ if unk_token is not None:
336
+ vocab_dict[unk_token] = len(vocab_dict)
337
+
338
+ if pad_token is not None:
339
+ vocab_dict[pad_token] = len(vocab_dict)
340
+
341
+ return vocab_dict
342
+
343
+
344
+ def main():
345
+ # See all possible arguments in src/transformers/training_args.py
346
+ # or by passing the --help flag to this script.
347
+ # We now keep distinct sets of args, for a cleaner separation of concerns.
348
+
349
+ parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
350
+ if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
351
+ # If we pass only one argument to the script and it's the path to a json file,
352
+ # let's parse it to get our arguments.
353
+ model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
354
+ else:
355
+ model_args, data_args, training_args = parser.parse_args_into_dataclasses()
356
+
357
+ # Detecting last checkpoint.
358
+ last_checkpoint = None
359
+ if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
360
+ last_checkpoint = get_last_checkpoint(training_args.output_dir)
361
+ if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
362
+ raise ValueError(
363
+ f"Output directory ({training_args.output_dir}) already exists and is not empty. "
364
+ "Use --overwrite_output_dir to overcome."
365
+ )
366
+ elif last_checkpoint is not None:
367
+ logger.info(
368
+ f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
369
+ "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
370
+ )
371
+
372
+ # Setup logging
373
+ logging.basicConfig(
374
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
375
+ datefmt="%m/%d/%Y %H:%M:%S",
376
+ handlers=[logging.StreamHandler(sys.stdout)],
377
+ )
378
+ logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
379
+
380
+ # Log on each process the small summary:
381
+ logger.warning(
382
+ f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
383
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
384
+ )
385
+ # Set the verbosity to info of the Transformers logger (on main process only):
386
+ if is_main_process(training_args.local_rank):
387
+ transformers.utils.logging.set_verbosity_info()
388
+ logger.info("Training/evaluation parameters %s", training_args)
389
+
390
+ # Set seed before initializing model.
391
+ set_seed(training_args.seed)
392
+
393
+ # 1. First, let's load the dataset
394
+ raw_datasets = DatasetDict()
395
+
396
+ if training_args.do_train:
397
+ raw_datasets["train"] = load_dataset(
398
+ data_args.dataset_name,
399
+ data_args.dataset_config_name,
400
+ split=data_args.train_split_name,
401
+ use_auth_token=data_args.use_auth_token,
402
+ )
403
+
404
+ if data_args.audio_column_name not in raw_datasets["train"].column_names:
405
+ raise ValueError(
406
+ f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
407
+ "Make sure to set `--audio_column_name` to the correct audio column - one of "
408
+ f"{', '.join(raw_datasets['train'].column_names)}."
409
+ )
410
+
411
+ if data_args.text_column_name not in raw_datasets["train"].column_names:
412
+ raise ValueError(
413
+ f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
414
+ "Make sure to set `--text_column_name` to the correct text column - one of "
415
+ f"{', '.join(raw_datasets['train'].column_names)}."
416
+ )
417
+
418
+ if data_args.max_train_samples is not None:
419
+ raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
420
+
421
+ if training_args.do_eval:
422
+ raw_datasets["eval"] = load_dataset(
423
+ data_args.dataset_name,
424
+ data_args.dataset_config_name,
425
+ split=data_args.eval_split_name,
426
+ use_auth_token=data_args.use_auth_token,
427
+ )
428
+
429
+ if data_args.max_eval_samples is not None:
430
+ raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
431
+
432
+ # 2. We remove some special characters from the datasets
433
+ # that make training complicated and do not help in transcribing the speech
434
+ # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
435
+ # that could be easily picked up by the model
436
+ chars_to_ignore_regex = (
437
+ f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
438
+ )
439
+ text_column_name = data_args.text_column_name
440
+
441
+ def remove_special_characters(batch):
442
+ if chars_to_ignore_regex is not None:
443
+ batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
444
+ else:
445
+ batch["target_text"] = batch[text_column_name].lower() + " "
446
+ return batch
447
+
448
+ with training_args.main_process_first(desc="dataset map special characters removal"):
449
+ raw_datasets = raw_datasets.map(
450
+ remove_special_characters,
451
+ remove_columns=[text_column_name],
452
+ desc="remove special characters from datasets",
453
+ )
454
+
455
+ # save special tokens for tokenizer
456
+ word_delimiter_token = data_args.word_delimiter_token
457
+ unk_token = data_args.unk_token
458
+ pad_token = data_args.pad_token
459
+
460
+ # 3. Next, let's load the config as we might need it to create
461
+ # the tokenizer
462
+ # load config
463
+ config = AutoConfig.from_pretrained(
464
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
465
+ )
466
+
467
+ # 4. Next, if no tokenizer file is defined,
468
+ # we create the vocabulary of the model by extracting all unique characters from
469
+ # the training and evaluation datasets
470
+ # We need to make sure that only first rank saves vocabulary
471
+ # make sure all processes wait until vocab is created
472
+ tokenizer_name_or_path = model_args.tokenizer_name_or_path
473
+ tokenizer_kwargs = {}
474
+ if tokenizer_name_or_path is None:
475
+ # save vocab in training output dir
476
+ tokenizer_name_or_path = training_args.output_dir
477
+
478
+ vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
479
+
480
+ with training_args.main_process_first():
481
+ if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
482
+ os.remove(vocab_file)
483
+
484
+ with training_args.main_process_first(desc="dataset map vocabulary creation"):
485
+ if not os.path.isfile(vocab_file):
486
+ os.makedirs(tokenizer_name_or_path, exist_ok=True)
487
+ vocab_dict = create_vocabulary_from_data(
488
+ raw_datasets,
489
+ word_delimiter_token=word_delimiter_token,
490
+ unk_token=unk_token,
491
+ pad_token=pad_token,
492
+ )
493
+
494
+ # save vocab dict to be loaded into tokenizer
495
+ with open(vocab_file, "w") as file:
496
+ json.dump(vocab_dict, file)
497
+
498
+ # if tokenizer has just been created
499
+ # it is defined by `tokenizer_class` if present in config else by `model_type`
500
+ tokenizer_kwargs = {
501
+ "config": config if config.tokenizer_class is not None else None,
502
+ "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
503
+ "unk_token": unk_token,
504
+ "pad_token": pad_token,
505
+ "word_delimiter_token": word_delimiter_token,
506
+ }
507
+
508
+ # 5. Now we can instantiate the feature extractor, tokenizer and model
509
+ # Note for distributed training, the .from_pretrained methods guarantee that only
510
+ # one local process can concurrently download model & vocab.
511
+
512
+ # load feature_extractor and tokenizer
513
+ tokenizer = AutoTokenizer.from_pretrained(
514
+ tokenizer_name_or_path,
515
+ use_auth_token=data_args.use_auth_token,
516
+ **tokenizer_kwargs,
517
+ )
518
+ feature_extractor = AutoFeatureExtractor.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
+ # adapt config
523
+ config.update(
524
+ {
525
+ "feat_proj_dropout": model_args.feat_proj_dropout,
526
+ "attention_dropout": model_args.attention_dropout,
527
+ "hidden_dropout": model_args.hidden_dropout,
528
+ "final_dropout": model_args.final_dropout,
529
+ "mask_time_prob": model_args.mask_time_prob,
530
+ "mask_time_length": model_args.mask_time_length,
531
+ "mask_feature_prob": model_args.mask_feature_prob,
532
+ "mask_feature_length": model_args.mask_feature_length,
533
+ "gradient_checkpointing": training_args.gradient_checkpointing,
534
+ "layerdrop": model_args.layerdrop,
535
+ "ctc_loss_reduction": model_args.ctc_loss_reduction,
536
+ "pad_token_id": tokenizer.pad_token_id,
537
+ "vocab_size": len(tokenizer),
538
+ "activation_dropout": model_args.activation_dropout,
539
+ }
540
+ )
541
+
542
+ # create model
543
+ model = AutoModelForCTC.from_pretrained(
544
+ model_args.model_name_or_path,
545
+ cache_dir=model_args.cache_dir,
546
+ config=config,
547
+ use_auth_token=data_args.use_auth_token,
548
+ )
549
+
550
+ # freeze encoder
551
+ if model_args.freeze_feature_encoder:
552
+ model.freeze_feature_encoder()
553
+
554
+ # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
555
+ # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
556
+ # so that we just need to set the correct target sampling rate and normalize the input
557
+ # via the `feature_extractor`
558
+
559
+ # make sure that dataset decodes audio with correct sampling rate
560
+ dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
561
+ if dataset_sampling_rate != feature_extractor.sampling_rate:
562
+ raw_datasets = raw_datasets.cast_column(
563
+ data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
564
+ )
565
+
566
+ # derive max & min input length for sample rate & max duration
567
+ max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
568
+ min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
569
+ audio_column_name = data_args.audio_column_name
570
+ num_workers = data_args.preprocessing_num_workers
571
+
572
+ # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
573
+ phoneme_language = data_args.phoneme_language
574
+
575
+ # Preprocessing the datasets.
576
+ # We need to read the audio files as arrays and tokenize the targets.
577
+ def prepare_dataset(batch):
578
+ # load audio
579
+ sample = batch[audio_column_name]
580
+
581
+ inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
582
+ batch["input_values"] = inputs.input_values[0]
583
+ batch["input_length"] = len(batch["input_values"])
584
+
585
+ # encode targets
586
+ additional_kwargs = {}
587
+ if phoneme_language is not None:
588
+ additional_kwargs["phonemizer_lang"] = phoneme_language
589
+
590
+ batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
591
+ return batch
592
+
593
+ with training_args.main_process_first(desc="dataset map preprocessing"):
594
+ vectorized_datasets = raw_datasets.map(
595
+ prepare_dataset,
596
+ remove_columns=next(iter(raw_datasets.values())).column_names,
597
+ num_proc=num_workers,
598
+ desc="preprocess datasets",
599
+ )
600
+
601
+ def is_audio_in_length_range(length):
602
+ return length > min_input_length and length < max_input_length
603
+
604
+ # filter data that is shorter than min_input_length
605
+ vectorized_datasets = vectorized_datasets.filter(
606
+ is_audio_in_length_range,
607
+ num_proc=num_workers,
608
+ input_columns=["input_length"],
609
+ )
610
+
611
+ # 7. Next, we can prepare the training.
612
+ # Let's use word error rate (WER) as our evaluation metric,
613
+ # instantiate a data collator and the trainer
614
+
615
+ # Define evaluation metrics during training, *i.e.* word error rate, character error rate
616
+ eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
617
+
618
+ # for large datasets it is advised to run the preprocessing on a
619
+ # single machine first with ``args.preprocessing_only`` since there will mostly likely
620
+ # be a timeout when running the script in distributed mode.
621
+ # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
622
+ # cached dataset
623
+ if data_args.preprocessing_only:
624
+ logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
625
+ return
626
+
627
+ def compute_metrics(pred):
628
+ pred_logits = pred.predictions
629
+ pred_ids = np.argmax(pred_logits, axis=-1)
630
+
631
+ pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
632
+
633
+ pred_str = tokenizer.batch_decode(pred_ids)
634
+ # we do not want to group tokens when computing the metrics
635
+ label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
636
+
637
+ metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
638
+
639
+ return metrics
640
+
641
+ # Now save everything to be able to create a single processor later
642
+ if is_main_process(training_args.local_rank):
643
+ # save feature extractor, tokenizer and config
644
+ feature_extractor.save_pretrained(training_args.output_dir)
645
+ tokenizer.save_pretrained(training_args.output_dir)
646
+ config.save_pretrained(training_args.output_dir)
647
+
648
+ try:
649
+ processor = AutoProcessor.from_pretrained(training_args.output_dir)
650
+ except (OSError, KeyError):
651
+ warnings.warn(
652
+ "Loading a processor from a feature extractor config that does not"
653
+ " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
654
+ " attribute to your `preprocessor_config.json` file to suppress this warning: "
655
+ " `'processor_class': 'Wav2Vec2Processor'`",
656
+ FutureWarning,
657
+ )
658
+ processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
659
+
660
+ # Instantiate custom data collator
661
+ data_collator = DataCollatorCTCWithPadding(processor=processor)
662
+
663
+ # Initialize Trainer
664
+ trainer = Trainer(
665
+ model=model,
666
+ data_collator=data_collator,
667
+ args=training_args,
668
+ compute_metrics=compute_metrics,
669
+ train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
670
+ eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
671
+ tokenizer=feature_extractor,
672
+ )
673
+
674
+ # 8. Finally, we can start training
675
+
676
+ # Training
677
+ if training_args.do_train:
678
+
679
+ # use last checkpoint if exist
680
+ if last_checkpoint is not None:
681
+ checkpoint = last_checkpoint
682
+ elif os.path.isdir(model_args.model_name_or_path):
683
+ checkpoint = model_args.model_name_or_path
684
+ else:
685
+ checkpoint = None
686
+
687
+ train_result = trainer.train(resume_from_checkpoint=checkpoint)
688
+ trainer.save_model()
689
+
690
+ metrics = train_result.metrics
691
+ max_train_samples = (
692
+ data_args.max_train_samples
693
+ if data_args.max_train_samples is not None
694
+ else len(vectorized_datasets["train"])
695
+ )
696
+ metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
697
+
698
+ trainer.log_metrics("train", metrics)
699
+ trainer.save_metrics("train", metrics)
700
+ trainer.save_state()
701
+
702
+ # Evaluation
703
+ results = {}
704
+ if training_args.do_eval:
705
+ logger.info("*** Evaluate ***")
706
+ metrics = trainer.evaluate()
707
+ max_eval_samples = (
708
+ data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
709
+ )
710
+ metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
711
+
712
+ trainer.log_metrics("eval", metrics)
713
+ trainer.save_metrics("eval", metrics)
714
+
715
+ # Write model card and (optionally) push to hub
716
+ config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
717
+ kwargs = {
718
+ "finetuned_from": model_args.model_name_or_path,
719
+ "tasks": "speech-recognition",
720
+ "tags": ["automatic-speech-recognition", data_args.dataset_name],
721
+ "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
722
+ "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
723
+ }
724
+ if "common_voice" in data_args.dataset_name:
725
+ kwargs["language"] = config_name
726
+
727
+ if training_args.push_to_hub:
728
+ trainer.push_to_hub(**kwargs)
729
+ else:
730
+ trainer.create_model_card(**kwargs)
731
+
732
+ return results
733
+
734
+
735
+ if __name__ == "__main__":
736
+ 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": "|", "special_tokens_map_file": null, "tokenizer_file": null, "name_or_path": "./", "tokenizer_class": "Wav2Vec2CTCTokenizer"}
training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4ab30bd674eab3fb9a449216bf31040fe00f4446703d9ee31e356fc6710aeb4d
3
+ size 2991
vocab.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"a": 1, "b": 2, "c": 3, "d": 4, "e": 5, "f": 6, "g": 7, "h": 8, "i": 9, "j": 10, "k": 11, "l": 12, "m": 13, "n": 14, "o": 15, "p": 16, "q": 17, "r": 18, "s": 19, "t": 20, "u": 21, "v": 22, "w": 23, "x": 24, "y": 25, "z": 26, "à": 27, "á": 28, "ä": 29, "ç": 30, "è": 31, "é": 32, "ê": 33, "ë": 34, "í": 35, "î": 36, "ï": 37, "ó": 38, "ö": 39, "ú": 40, "û": 41, "ü": 42, "ă": 43, "ć": 44, "đ": 45, "š": 46, "ţ": 47, "|": 0, "[UNK]": 48, "[PAD]": 49}