DrishtiSharma commited on
Commit
cf071cb
1 Parent(s): 547a1da

Training in progress, step 200

Browse files
.gitignore ADDED
@@ -0,0 +1 @@
 
 
1
+ checkpoint-*/
.ipynb_checkpoints/run-checkpoint.sh ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ python run_speech_recognition_ctc.py \
2
+ --dataset_name="mozilla-foundation/common_voice_8_0" \
3
+ --model_name_or_path="facebook/wav2vec2-xls-r-300m" \
4
+ --dataset_config_name="kk" \
5
+ --output_dir="./" \
6
+ --overwrite_output_dir \
7
+ --num_train_epochs="150" \
8
+ --per_device_train_batch_size="16" \
9
+ --learning_rate="2.22e-4" \
10
+ --gradient_accumulation_steps="2" \
11
+ --warmup_steps="1000" \
12
+ --length_column_name="input_length" \
13
+ --evaluation_strategy="steps" \
14
+ --text_column_name="sentence" \
15
+ --save_steps="200" \
16
+ --eval_steps="200" \
17
+ --logging_steps="200" \
18
+ --layerdrop="0.0" \
19
+ --attention_dropout="0.0" \
20
+ --activation_dropout="0.0" \
21
+ --hidden_dropout="0.0" \
22
+ --feat_proj_dropout="0.0" \
23
+ --layerdrop="0.0" \
24
+ --ctc_loss_reduction="mean" \
25
+ --save_total_limit="1" \
26
+ --freeze_feature_encoder \
27
+ --feat_proj_dropout="0.0" \
28
+ --mask_time_prob="0.25" \
29
+ --chars_to_ignore , ? . ! \- \; \: \" “ % ‘ ” � — ’ … – « » „ \` _ \
30
+ --gradient_checkpointing \
31
+ --use_auth_token \
32
+ --fp16 \
33
+ --group_by_length \
34
+ --do_train --do_eval \
35
+ --push_to_hub
added_tokens.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"<s>": 41, "</s>": 42}
config.json ADDED
@@ -0,0 +1,107 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "facebook/wav2vec2-xls-r-300m",
3
+ "activation_dropout": 0.0,
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": 10,
63
+ "mask_feature_min_masks": 0,
64
+ "mask_feature_prob": 0.0,
65
+ "mask_time_length": 10,
66
+ "mask_time_min_masks": 2,
67
+ "mask_time_prob": 0.25,
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": 40,
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": "4.17.0.dev0",
104
+ "use_weighted_layer_sum": false,
105
+ "vocab_size": 43,
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 sha256:8f1cdce6bbadd5edcbc49ff49e16c03549f2f5d8e4a50ae8a9d8451c947263e2
3
+ size 1262099953
run.sh ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ python run_speech_recognition_ctc.py \
2
+ --dataset_name="mozilla-foundation/common_voice_8_0" \
3
+ --model_name_or_path="facebook/wav2vec2-xls-r-300m" \
4
+ --dataset_config_name="kk" \
5
+ --output_dir="./" \
6
+ --overwrite_output_dir \
7
+ --num_train_epochs="150" \
8
+ --per_device_train_batch_size="16" \
9
+ --learning_rate="2.22e-4" \
10
+ --gradient_accumulation_steps="2" \
11
+ --warmup_steps="1000" \
12
+ --length_column_name="input_length" \
13
+ --evaluation_strategy="steps" \
14
+ --text_column_name="sentence" \
15
+ --save_steps="200" \
16
+ --eval_steps="200" \
17
+ --logging_steps="200" \
18
+ --layerdrop="0.0" \
19
+ --attention_dropout="0.0" \
20
+ --activation_dropout="0.0" \
21
+ --hidden_dropout="0.0" \
22
+ --feat_proj_dropout="0.0" \
23
+ --layerdrop="0.0" \
24
+ --ctc_loss_reduction="mean" \
25
+ --save_total_limit="1" \
26
+ --freeze_feature_encoder \
27
+ --feat_proj_dropout="0.0" \
28
+ --mask_time_prob="0.25" \
29
+ --chars_to_ignore , ? . ! \- \; \: \" “ % ‘ ” � — ’ … – « » „ \` _ \
30
+ --gradient_checkpointing \
31
+ --use_auth_token \
32
+ --fp16 \
33
+ --group_by_length \
34
+ --do_train --do_eval \
35
+ --push_to_hub
run_speech_recognition_ctc.py ADDED
@@ -0,0 +1,738 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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>=1.18.0", "To fix: pip install -r examples/pytorch/speech-recognition/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
+
135
+ Using `HfArgumentParser` we can turn this class
136
+ into argparse arguments to be able to specify them on
137
+ the command line.
138
+ """
139
+
140
+ dataset_name: str = field(
141
+ metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
142
+ )
143
+ dataset_config_name: str = field(
144
+ default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
145
+ )
146
+ train_split_name: str = field(
147
+ default="train+validation",
148
+ metadata={
149
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to "
150
+ "'train+validation'"
151
+ },
152
+ )
153
+ eval_split_name: str = field(
154
+ default="test",
155
+ metadata={
156
+ "help": "The name of the evaluation data set split to use (via the datasets library). Defaults to 'test'"
157
+ },
158
+ )
159
+ audio_column_name: str = field(
160
+ default="audio",
161
+ metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
162
+ )
163
+ text_column_name: str = field(
164
+ default="text",
165
+ metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
166
+ )
167
+ overwrite_cache: bool = field(
168
+ default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
169
+ )
170
+ preprocessing_num_workers: Optional[int] = field(
171
+ default=None,
172
+ metadata={"help": "The number of processes to use for the preprocessing."},
173
+ )
174
+ max_train_samples: Optional[int] = field(
175
+ default=None,
176
+ metadata={
177
+ "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
178
+ "value if set."
179
+ },
180
+ )
181
+ max_eval_samples: Optional[int] = field(
182
+ default=None,
183
+ metadata={
184
+ "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
185
+ "value if set."
186
+ },
187
+ )
188
+ chars_to_ignore: Optional[List[str]] = list_field(
189
+ default=None,
190
+ metadata={"help": "A list of characters to remove from the transcripts."},
191
+ )
192
+ eval_metrics: List[str] = list_field(
193
+ default=["wer"],
194
+ metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
195
+ )
196
+ max_duration_in_seconds: float = field(
197
+ default=20.0,
198
+ metadata={
199
+ "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
200
+ },
201
+ )
202
+ min_duration_in_seconds: float = field(
203
+ default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
204
+ )
205
+ preprocessing_only: bool = field(
206
+ default=False,
207
+ metadata={
208
+ "help": "Whether to only do data preprocessing and skip training. "
209
+ "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
210
+ "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
211
+ "so that the cached datasets can consequently be loaded in distributed training"
212
+ },
213
+ )
214
+ use_auth_token: bool = field(
215
+ default=False,
216
+ metadata={
217
+ "help": "If :obj:`True`, will use the token generated when running"
218
+ ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
219
+ },
220
+ )
221
+ unk_token: str = field(
222
+ default="[UNK]",
223
+ metadata={"help": "The unk token for the tokenizer"},
224
+ )
225
+ pad_token: str = field(
226
+ default="[PAD]",
227
+ metadata={"help": "The padding token for the tokenizer"},
228
+ )
229
+ word_delimiter_token: str = field(
230
+ default="|",
231
+ metadata={"help": "The word delimiter token for the tokenizer"},
232
+ )
233
+ phoneme_language: Optional[str] = field(
234
+ default=None,
235
+ metadata={
236
+ "help": "The target language that should be used be"
237
+ " passed to the tokenizer for tokenization. Note that"
238
+ " this is only relevant if the model classifies the"
239
+ " input audio to a sequence of phoneme sequences."
240
+ },
241
+ )
242
+
243
+
244
+ @dataclass
245
+ class DataCollatorCTCWithPadding:
246
+ """
247
+ Data collator that will dynamically pad the inputs received.
248
+ Args:
249
+ processor (:class:`~transformers.AutoProcessor`)
250
+ The processor used for proccessing the data.
251
+ padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
252
+ Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
253
+ among:
254
+ * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
255
+ sequence if provided).
256
+ * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
257
+ maximum acceptable input length for the model if that argument is not provided.
258
+ * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
259
+ different lengths).
260
+ max_length (:obj:`int`, `optional`):
261
+ Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
262
+ max_length_labels (:obj:`int`, `optional`):
263
+ Maximum length of the ``labels`` returned list and optionally padding length (see above).
264
+ pad_to_multiple_of (:obj:`int`, `optional`):
265
+ If set will pad the sequence to a multiple of the provided value.
266
+ This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
267
+ 7.5 (Volta).
268
+ """
269
+
270
+ processor: AutoProcessor
271
+ padding: Union[bool, str] = "longest"
272
+ pad_to_multiple_of: Optional[int] = None
273
+ pad_to_multiple_of_labels: Optional[int] = None
274
+
275
+ def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
276
+ # split inputs and labels since they have to be of different lenghts and need
277
+ # different padding methods
278
+ input_features = [{"input_values": feature["input_values"]} for feature in features]
279
+ label_features = [{"input_ids": feature["labels"]} for feature in features]
280
+
281
+ batch = self.processor.pad(
282
+ input_features,
283
+ padding=self.padding,
284
+ pad_to_multiple_of=self.pad_to_multiple_of,
285
+ return_tensors="pt",
286
+ )
287
+
288
+ with self.processor.as_target_processor():
289
+ labels_batch = self.processor.pad(
290
+ label_features,
291
+ padding=self.padding,
292
+ pad_to_multiple_of=self.pad_to_multiple_of_labels,
293
+ return_tensors="pt",
294
+ )
295
+
296
+ # replace padding with -100 to ignore loss correctly
297
+ labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
298
+
299
+ batch["labels"] = labels
300
+
301
+ return batch
302
+
303
+
304
+ def create_vocabulary_from_data(
305
+ datasets: DatasetDict,
306
+ word_delimiter_token: Optional[str] = None,
307
+ unk_token: Optional[str] = None,
308
+ pad_token: Optional[str] = None,
309
+ ):
310
+ # Given training and test labels create vocabulary
311
+ def extract_all_chars(batch):
312
+ all_text = " ".join(batch["target_text"])
313
+ vocab = list(set(all_text))
314
+ return {"vocab": [vocab], "all_text": [all_text]}
315
+
316
+ vocabs = datasets.map(
317
+ extract_all_chars,
318
+ batched=True,
319
+ batch_size=-1,
320
+ keep_in_memory=True,
321
+ remove_columns=datasets["train"].column_names,
322
+ )
323
+
324
+ # take union of all unique characters in each dataset
325
+ vocab_set = functools.reduce(
326
+ lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
327
+ )
328
+
329
+ vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
330
+
331
+ # replace white space with delimiter token
332
+ if word_delimiter_token is not None:
333
+ vocab_dict[word_delimiter_token] = vocab_dict[" "]
334
+ del vocab_dict[" "]
335
+
336
+ # add unk and pad token
337
+ if unk_token is not None:
338
+ vocab_dict[unk_token] = len(vocab_dict)
339
+
340
+ if pad_token is not None:
341
+ vocab_dict[pad_token] = len(vocab_dict)
342
+
343
+ return vocab_dict
344
+
345
+
346
+ def main():
347
+ # See all possible arguments in src/transformers/training_args.py
348
+ # or by passing the --help flag to this script.
349
+ # We now keep distinct sets of args, for a cleaner separation of concerns.
350
+
351
+ parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
352
+ if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
353
+ # If we pass only one argument to the script and it's the path to a json file,
354
+ # let's parse it to get our arguments.
355
+ model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
356
+ else:
357
+ model_args, data_args, training_args = parser.parse_args_into_dataclasses()
358
+
359
+ # Detecting last checkpoint.
360
+ last_checkpoint = None
361
+ if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
362
+ last_checkpoint = get_last_checkpoint(training_args.output_dir)
363
+ if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
364
+ raise ValueError(
365
+ f"Output directory ({training_args.output_dir}) already exists and is not empty. "
366
+ "Use --overwrite_output_dir to overcome."
367
+ )
368
+ elif last_checkpoint is not None:
369
+ logger.info(
370
+ f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
371
+ "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
372
+ )
373
+
374
+ # Setup logging
375
+ logging.basicConfig(
376
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
377
+ datefmt="%m/%d/%Y %H:%M:%S",
378
+ handlers=[logging.StreamHandler(sys.stdout)],
379
+ )
380
+ logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
381
+
382
+ # Log on each process the small summary:
383
+ logger.warning(
384
+ f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
385
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
386
+ )
387
+ # Set the verbosity to info of the Transformers logger (on main process only):
388
+ if is_main_process(training_args.local_rank):
389
+ transformers.utils.logging.set_verbosity_info()
390
+ logger.info("Training/evaluation parameters %s", training_args)
391
+
392
+ # Set seed before initializing model.
393
+ set_seed(training_args.seed)
394
+
395
+ # 1. First, let's load the dataset
396
+ raw_datasets = DatasetDict()
397
+
398
+ if training_args.do_train:
399
+ raw_datasets["train"] = load_dataset(
400
+ data_args.dataset_name,
401
+ data_args.dataset_config_name,
402
+ split=data_args.train_split_name,
403
+ use_auth_token=data_args.use_auth_token,
404
+ )
405
+
406
+ if data_args.audio_column_name not in raw_datasets["train"].column_names:
407
+ raise ValueError(
408
+ f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
409
+ "Make sure to set `--audio_column_name` to the correct audio column - one of "
410
+ f"{', '.join(raw_datasets['train'].column_names)}."
411
+ )
412
+
413
+ if data_args.text_column_name not in raw_datasets["train"].column_names:
414
+ raise ValueError(
415
+ f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
416
+ "Make sure to set `--text_column_name` to the correct text column - one of "
417
+ f"{', '.join(raw_datasets['train'].column_names)}."
418
+ )
419
+
420
+ if data_args.max_train_samples is not None:
421
+ raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
422
+
423
+ if training_args.do_eval:
424
+ raw_datasets["eval"] = load_dataset(
425
+ data_args.dataset_name,
426
+ data_args.dataset_config_name,
427
+ split=data_args.eval_split_name,
428
+ use_auth_token=data_args.use_auth_token,
429
+ )
430
+
431
+ if data_args.max_eval_samples is not None:
432
+ raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
433
+
434
+ # 2. We remove some special characters from the datasets
435
+ # that make training complicated and do not help in transcribing the speech
436
+ # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
437
+ # that could be easily picked up by the model
438
+ chars_to_ignore_regex = (
439
+ f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
440
+ )
441
+ text_column_name = data_args.text_column_name
442
+
443
+ def remove_special_characters(batch):
444
+ if chars_to_ignore_regex is not None:
445
+ batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
446
+ else:
447
+ batch["target_text"] = batch[text_column_name].lower() + " "
448
+ return batch
449
+
450
+ with training_args.main_process_first(desc="dataset map special characters removal"):
451
+ raw_datasets = raw_datasets.map(
452
+ remove_special_characters,
453
+ remove_columns=[text_column_name],
454
+ desc="remove special characters from datasets",
455
+ )
456
+
457
+ # save special tokens for tokenizer
458
+ word_delimiter_token = data_args.word_delimiter_token
459
+ unk_token = data_args.unk_token
460
+ pad_token = data_args.pad_token
461
+
462
+ # 3. Next, let's load the config as we might need it to create
463
+ # the tokenizer
464
+ # load config
465
+ config = AutoConfig.from_pretrained(
466
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
467
+ )
468
+
469
+ # 4. Next, if no tokenizer file is defined,
470
+ # we create the vocabulary of the model by extracting all unique characters from
471
+ # the training and evaluation datasets
472
+ # We need to make sure that only first rank saves vocabulary
473
+ # make sure all processes wait until vocab is created
474
+ tokenizer_name_or_path = model_args.tokenizer_name_or_path
475
+ tokenizer_kwargs = {}
476
+ if tokenizer_name_or_path is None:
477
+ # save vocab in training output dir
478
+ tokenizer_name_or_path = training_args.output_dir
479
+
480
+ vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
481
+
482
+ with training_args.main_process_first():
483
+ if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
484
+ os.remove(vocab_file)
485
+
486
+ with training_args.main_process_first(desc="dataset map vocabulary creation"):
487
+ if not os.path.isfile(vocab_file):
488
+ os.makedirs(tokenizer_name_or_path, exist_ok=True)
489
+ vocab_dict = create_vocabulary_from_data(
490
+ raw_datasets,
491
+ word_delimiter_token=word_delimiter_token,
492
+ unk_token=unk_token,
493
+ pad_token=pad_token,
494
+ )
495
+
496
+ # save vocab dict to be loaded into tokenizer
497
+ with open(vocab_file, "w") as file:
498
+ json.dump(vocab_dict, file)
499
+
500
+ # if tokenizer has just been created
501
+ # it is defined by `tokenizer_class` if present in config else by `model_type`
502
+ tokenizer_kwargs = {
503
+ "config": config if config.tokenizer_class is not None else None,
504
+ "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
505
+ "unk_token": unk_token,
506
+ "pad_token": pad_token,
507
+ "word_delimiter_token": word_delimiter_token,
508
+ }
509
+
510
+ # 5. Now we can instantiate the feature extractor, tokenizer and model
511
+ # Note for distributed training, the .from_pretrained methods guarantee that only
512
+ # one local process can concurrently download model & vocab.
513
+
514
+ # load feature_extractor and tokenizer
515
+ tokenizer = AutoTokenizer.from_pretrained(
516
+ tokenizer_name_or_path,
517
+ use_auth_token=data_args.use_auth_token,
518
+ **tokenizer_kwargs,
519
+ )
520
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
521
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
522
+ )
523
+
524
+ # adapt config
525
+ config.update(
526
+ {
527
+ "feat_proj_dropout": model_args.feat_proj_dropout,
528
+ "attention_dropout": model_args.attention_dropout,
529
+ "hidden_dropout": model_args.hidden_dropout,
530
+ "final_dropout": model_args.final_dropout,
531
+ "mask_time_prob": model_args.mask_time_prob,
532
+ "mask_time_length": model_args.mask_time_length,
533
+ "mask_feature_prob": model_args.mask_feature_prob,
534
+ "mask_feature_length": model_args.mask_feature_length,
535
+ "gradient_checkpointing": training_args.gradient_checkpointing,
536
+ "layerdrop": model_args.layerdrop,
537
+ "ctc_loss_reduction": model_args.ctc_loss_reduction,
538
+ "pad_token_id": tokenizer.pad_token_id,
539
+ "vocab_size": len(tokenizer),
540
+ "activation_dropout": model_args.activation_dropout,
541
+ }
542
+ )
543
+
544
+ # create model
545
+ model = AutoModelForCTC.from_pretrained(
546
+ model_args.model_name_or_path,
547
+ cache_dir=model_args.cache_dir,
548
+ config=config,
549
+ use_auth_token=data_args.use_auth_token,
550
+ )
551
+
552
+ # freeze encoder
553
+ if model_args.freeze_feature_encoder:
554
+ model.freeze_feature_encoder()
555
+
556
+ # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
557
+ # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
558
+ # so that we just need to set the correct target sampling rate and normalize the input
559
+ # via the `feature_extractor`
560
+
561
+ # make sure that dataset decodes audio with correct sampling rate
562
+ dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
563
+ if dataset_sampling_rate != feature_extractor.sampling_rate:
564
+ raw_datasets = raw_datasets.cast_column(
565
+ data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
566
+ )
567
+
568
+ # derive max & min input length for sample rate & max duration
569
+ max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
570
+ min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
571
+ audio_column_name = data_args.audio_column_name
572
+ num_workers = data_args.preprocessing_num_workers
573
+
574
+ # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
575
+ phoneme_language = data_args.phoneme_language
576
+
577
+ # Preprocessing the datasets.
578
+ # We need to read the audio files as arrays and tokenize the targets.
579
+ def prepare_dataset(batch):
580
+ # load audio
581
+ sample = batch[audio_column_name]
582
+
583
+ inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
584
+ batch["input_values"] = inputs.input_values[0]
585
+ batch["input_length"] = len(batch["input_values"])
586
+
587
+ # encode targets
588
+ additional_kwargs = {}
589
+ if phoneme_language is not None:
590
+ additional_kwargs["phonemizer_lang"] = phoneme_language
591
+
592
+ batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
593
+ return batch
594
+
595
+ with training_args.main_process_first(desc="dataset map preprocessing"):
596
+ vectorized_datasets = raw_datasets.map(
597
+ prepare_dataset,
598
+ remove_columns=next(iter(raw_datasets.values())).column_names,
599
+ num_proc=num_workers,
600
+ desc="preprocess datasets",
601
+ )
602
+
603
+ def is_audio_in_length_range(length):
604
+ return length > min_input_length and length < max_input_length
605
+
606
+ # filter data that is shorter than min_input_length
607
+ vectorized_datasets = vectorized_datasets.filter(
608
+ is_audio_in_length_range,
609
+ num_proc=num_workers,
610
+ input_columns=["input_length"],
611
+ )
612
+
613
+ # 7. Next, we can prepare the training.
614
+ # Let's use word error rate (WER) as our evaluation metric,
615
+ # instantiate a data collator and the trainer
616
+
617
+ # Define evaluation metrics during training, *i.e.* word error rate, character error rate
618
+ eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
619
+
620
+ # for large datasets it is advised to run the preprocessing on a
621
+ # single machine first with ``args.preprocessing_only`` since there will mostly likely
622
+ # be a timeout when running the script in distributed mode.
623
+ # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
624
+ # cached dataset
625
+ if data_args.preprocessing_only:
626
+ logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
627
+ return
628
+
629
+ def compute_metrics(pred):
630
+ pred_logits = pred.predictions
631
+ pred_ids = np.argmax(pred_logits, axis=-1)
632
+
633
+ pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
634
+
635
+ pred_str = tokenizer.batch_decode(pred_ids)
636
+ # we do not want to group tokens when computing the metrics
637
+ label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
638
+
639
+ metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
640
+
641
+ return metrics
642
+
643
+ # Now save everything to be able to create a single processor later
644
+ if is_main_process(training_args.local_rank):
645
+ # save feature extractor, tokenizer and config
646
+ feature_extractor.save_pretrained(training_args.output_dir)
647
+ tokenizer.save_pretrained(training_args.output_dir)
648
+ config.save_pretrained(training_args.output_dir)
649
+
650
+ try:
651
+ processor = AutoProcessor.from_pretrained(training_args.output_dir)
652
+ except (OSError, KeyError):
653
+ warnings.warn(
654
+ "Loading a processor from a feature extractor config that does not"
655
+ " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
656
+ " attribute to your `preprocessor_config.json` file to suppress this warning: "
657
+ " `'processor_class': 'Wav2Vec2Processor'`",
658
+ FutureWarning,
659
+ )
660
+ processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
661
+
662
+ # Instantiate custom data collator
663
+ data_collator = DataCollatorCTCWithPadding(processor=processor)
664
+
665
+ # Initialize Trainer
666
+ trainer = Trainer(
667
+ model=model,
668
+ data_collator=data_collator,
669
+ args=training_args,
670
+ compute_metrics=compute_metrics,
671
+ train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
672
+ eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
673
+ tokenizer=feature_extractor,
674
+ )
675
+
676
+ # 8. Finally, we can start training
677
+
678
+ # Training
679
+ if training_args.do_train:
680
+
681
+ # use last checkpoint if exist
682
+ if last_checkpoint is not None:
683
+ checkpoint = last_checkpoint
684
+ elif os.path.isdir(model_args.model_name_or_path):
685
+ checkpoint = model_args.model_name_or_path
686
+ else:
687
+ checkpoint = None
688
+
689
+ train_result = trainer.train(resume_from_checkpoint=checkpoint)
690
+ trainer.save_model()
691
+
692
+ metrics = train_result.metrics
693
+ max_train_samples = (
694
+ data_args.max_train_samples
695
+ if data_args.max_train_samples is not None
696
+ else len(vectorized_datasets["train"])
697
+ )
698
+ metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
699
+
700
+ trainer.log_metrics("train", metrics)
701
+ trainer.save_metrics("train", metrics)
702
+ trainer.save_state()
703
+
704
+ # Evaluation
705
+ results = {}
706
+ if training_args.do_eval:
707
+ logger.info("*** Evaluate ***")
708
+ metrics = trainer.evaluate()
709
+ max_eval_samples = (
710
+ data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
711
+ )
712
+ metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
713
+
714
+ trainer.log_metrics("eval", metrics)
715
+ trainer.save_metrics("eval", metrics)
716
+
717
+ # Write model card and (optionally) push to hub
718
+ config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
719
+ kwargs = {
720
+ "finetuned_from": model_args.model_name_or_path,
721
+ "tasks": "speech-recognition",
722
+ "tags": ["automatic-speech-recognition", data_args.dataset_name],
723
+ "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
724
+ "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
725
+ }
726
+ if "common_voice" in data_args.dataset_name:
727
+ kwargs["language"] = config_name
728
+
729
+ if training_args.push_to_hub:
730
+ trainer.push_to_hub(**kwargs)
731
+ else:
732
+ trainer.create_model_card(**kwargs)
733
+
734
+ return results
735
+
736
+
737
+ if __name__ == "__main__":
738
+ 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:adb0105863dcce00203582f0ef97e1325605ebb083f887a9da4eb3e96f1c91bb
3
+ size 2991
vocab.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"a": 1, "c": 2, "i": 3, "k": 4, "а": 5, "б": 6, "г": 7, "д": 8, "е": 9, "ж": 10, "з": 11, "и": 12, "й": 13, "к": 14, "л": 15, "м": 16, "н": 17, "о": 18, "п": 19, "р": 20, "с": 21, "т": 22, "у": 23, "х": 24, "ш": 25, "щ": 26, "ы": 27, "ю": 28, "я": 29, "і": 30, "ғ": 31, "қ": 32, "ң": 33, "ү": 34, "ұ": 35, "һ": 36, "ә": 37, "ө": 38, "|": 0, "[UNK]": 39, "[PAD]": 40}