chmanoj commited on
Commit
630b73d
1 Parent(s): 05052ff

Training 2428 steps check-in

Browse files
.gitattributes ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ *.7z filter=lfs diff=lfs merge=lfs -text
2
+ *.arrow filter=lfs diff=lfs merge=lfs -text
3
+ *.bin filter=lfs diff=lfs merge=lfs -text
4
+ *.bin.* filter=lfs diff=lfs merge=lfs -text
5
+ *.bz2 filter=lfs diff=lfs merge=lfs -text
6
+ *.ftz filter=lfs diff=lfs merge=lfs -text
7
+ *.gz filter=lfs diff=lfs merge=lfs -text
8
+ *.h5 filter=lfs diff=lfs merge=lfs -text
9
+ *.joblib filter=lfs diff=lfs merge=lfs -text
10
+ *.lfs.* filter=lfs diff=lfs merge=lfs -text
11
+ *.model filter=lfs diff=lfs merge=lfs -text
12
+ *.msgpack filter=lfs diff=lfs merge=lfs -text
13
+ *.onnx filter=lfs diff=lfs merge=lfs -text
14
+ *.ot filter=lfs diff=lfs merge=lfs -text
15
+ *.parquet filter=lfs diff=lfs merge=lfs -text
16
+ *.pb filter=lfs diff=lfs merge=lfs -text
17
+ *.pt filter=lfs diff=lfs merge=lfs -text
18
+ *.pth filter=lfs diff=lfs merge=lfs -text
19
+ *.rar filter=lfs diff=lfs merge=lfs -text
20
+ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
21
+ *.tar.* filter=lfs diff=lfs merge=lfs -text
22
+ *.tflite filter=lfs diff=lfs merge=lfs -text
23
+ *.tgz filter=lfs diff=lfs merge=lfs -text
24
+ *.xz filter=lfs diff=lfs merge=lfs -text
25
+ *.zip filter=lfs diff=lfs merge=lfs -text
26
+ *.zstandard filter=lfs diff=lfs merge=lfs -text
27
+ *tfevents* filter=lfs diff=lfs merge=lfs -text
.gitignore ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ checkpoint-*/
2
+ .ipynb_checkpoints/*
added_tokens.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"<s>": 77, "</s>": 78}
config.json ADDED
@@ -0,0 +1,107 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "facebook/wav2vec2-xls-r-2b",
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
+ "Wav2Vec2ForPreTraining"
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": 1920,
58
+ "initializer_range": 0.02,
59
+ "intermediate_size": 7680,
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": 1920,
79
+ "pad_token_id": 76,
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": "4.16.0.dev0",
104
+ "use_weighted_layer_sum": false,
105
+ "vocab_size": 79,
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.sh ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ python run_speech_recognition_ctc.py \
2
+ --model_name_or_path="facebook/wav2vec2-xls-r-2b" \
3
+ --dataset_name="openslr_SLR66" \
4
+ --train_split_name="train" \
5
+ --preprocessing_num_workers="8" \
6
+ --output_dir="./" \
7
+ --overwrite_output_dir \
8
+ --num_train_epochs="5" \
9
+ --per_device_train_batch_size="4" \
10
+ --per_device_eval_batch_size="4" \
11
+ --gradient_accumulation_steps="8" \
12
+ --learning_rate="7.5e-5" \
13
+ --warmup_steps="2000" \
14
+ --length_column_name="input_length" \
15
+ --evaluation_strategy="steps" \
16
+ --text_column_name="sentence" \
17
+ --chars_to_ignore , ? . ! \- \; \: \" “ % ‘ ” � — ’ … – \
18
+ --save_steps="500" \
19
+ --eval_steps="500" \
20
+ --logging_steps="100" \
21
+ --layerdrop="0.0" \
22
+ --activation_dropout="0.1" \
23
+ --save_total_limit="3" \
24
+ --freeze_feature_encoder \
25
+ --feat_proj_dropout="0.0" \
26
+ --mask_time_prob="0.75" \
27
+ --mask_time_length="10" \
28
+ --mask_feature_prob="0.25" \
29
+ --mask_feature_length="64" \
30
+ --gradient_checkpointing \
31
+ --use_auth_token \
32
+ --fp16 \
33
+ --group_by_length \
34
+ --do_train --do_eval \
35
+ --push_to_hub
run_bnb.sh ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ python run_speech_recognition_ctc_bnb.py \
2
+ --model_name_or_path="facebook/wav2vec2-xls-r-2b" \
3
+ --dataset_name="openslr_SLR66" \
4
+ --train_split_name="train" \
5
+ --preprocessing_num_workers="8" \
6
+ --output_dir="./" \
7
+ --overwrite_output_dir \
8
+ --num_train_epochs="150" \
9
+ --per_device_train_batch_size="4" \
10
+ --per_device_eval_batch_size="4" \
11
+ --gradient_accumulation_steps="8" \
12
+ --learning_rate="3e-6" \
13
+ --warmup_steps="2000" \
14
+ --length_column_name="input_length" \
15
+ --evaluation_strategy="steps" \
16
+ --text_column_name="sentence" \
17
+ --chars_to_ignore , ? . ! \- \; \: \" “ % ‘ ” � — ’ … – \
18
+ --save_steps="500" \
19
+ --eval_steps="500" \
20
+ --logging_steps="10" \
21
+ --layerdrop="0.0" \
22
+ --activation_dropout="0.1" \
23
+ --save_total_limit="2" \
24
+ --freeze_feature_encoder \
25
+ --feat_proj_dropout="0.0" \
26
+ --mask_time_prob="0.4" \
27
+ --mask_time_length="10" \
28
+ --mask_feature_prob="0.25" \
29
+ --mask_feature_length="64" \
30
+ --gradient_checkpointing \
31
+ --use_auth_token \
32
+ --fp16 \
33
+ --group_by_length \
34
+ --do_train --do_eval
run_speech_recognition_ctc.py ADDED
@@ -0,0 +1,756 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python
2
+ # coding=utf-8
3
+ # Copyright 2021 The HuggingFace Inc. team. All rights reserved.
4
+ #
5
+ # Licensed under the Apache License, Version 2.0 (the "License");
6
+ # you may not use this file except in compliance with the License.
7
+ # You may obtain a copy of the License at
8
+ #
9
+ # http://www.apache.org/licenses/LICENSE-2.0
10
+ #
11
+ # Unless required by applicable law or agreed to in writing, software
12
+ # distributed under the License is distributed on an "AS IS" BASIS,
13
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14
+ # See the License for the specific language governing permissions and
15
+
16
+ """ Fine-tuning a 🤗 Transformers CTC model for automatic speech recognition"""
17
+
18
+ import functools
19
+ import json
20
+ import logging
21
+ import os
22
+ import re
23
+ import sys
24
+ import warnings
25
+ from dataclasses import dataclass, field
26
+ from typing import Dict, List, Optional, Union
27
+
28
+ import datasets
29
+ import numpy as np
30
+ import torch
31
+ from datasets import DatasetDict, load_dataset, load_metric
32
+
33
+ import transformers
34
+ from transformers import (
35
+ AutoConfig,
36
+ AutoFeatureExtractor,
37
+ AutoModelForCTC,
38
+ AutoProcessor,
39
+ AutoTokenizer,
40
+ HfArgumentParser,
41
+ Trainer,
42
+ TrainingArguments,
43
+ Wav2Vec2Processor,
44
+ set_seed,
45
+ )
46
+ from transformers.trainer_utils import get_last_checkpoint, is_main_process
47
+ from transformers.utils import check_min_version
48
+ from transformers.utils.versions import require_version
49
+
50
+
51
+ # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
52
+ check_min_version("4.16.0.dev0")
53
+
54
+ require_version("datasets>=1.13.3", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
55
+
56
+
57
+ logger = logging.getLogger(__name__)
58
+
59
+
60
+ def list_field(default=None, metadata=None):
61
+ return field(default_factory=lambda: default, metadata=metadata)
62
+
63
+ def get_telugu_dataset(validation_split=False):
64
+ dataset = load_dataset('openslr', 'SLR66')
65
+
66
+ seed=1242
67
+
68
+ if validation_split:
69
+ train_testvalid = dataset['train'].train_test_split(test_size=0.2, seed=seed)
70
+ # Split the 10% test + valid in half test, half valid
71
+ test_valid = train_testvalid['test'].train_test_split(test_size=0.33, seed=seed)
72
+ # gather everyone if you want to have a single DatasetDict
73
+ out_dataset = DatasetDict({
74
+ 'train': train_testvalid['train'],
75
+ 'test': test_valid['test'],
76
+ 'valid': test_valid['train']})
77
+ else:
78
+ train_testvalid = dataset['train'].train_test_split(test_size=0.25, seed=seed)
79
+ out_dataset = DatasetDict({
80
+ 'train': train_testvalid['train'],
81
+ 'test': train_testvalid['test']})
82
+ return out_dataset
83
+
84
+
85
+ @dataclass
86
+ class ModelArguments:
87
+ """
88
+ Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
89
+ """
90
+
91
+ model_name_or_path: str = field(
92
+ metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
93
+ )
94
+ tokenizer_name_or_path: Optional[str] = field(
95
+ default=None,
96
+ metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
97
+ )
98
+ cache_dir: Optional[str] = field(
99
+ default=None,
100
+ metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
101
+ )
102
+ freeze_feature_encoder: bool = field(
103
+ default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
104
+ )
105
+ attention_dropout: float = field(
106
+ default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
107
+ )
108
+ activation_dropout: float = field(
109
+ default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
110
+ )
111
+ feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
112
+ hidden_dropout: float = field(
113
+ default=0.0,
114
+ metadata={
115
+ "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
116
+ },
117
+ )
118
+ final_dropout: float = field(
119
+ default=0.0,
120
+ metadata={"help": "The dropout probability for the final projection layer."},
121
+ )
122
+ mask_time_prob: float = field(
123
+ default=0.05,
124
+ metadata={
125
+ "help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
126
+ "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
127
+ "vectors will be masked along the time axis."
128
+ },
129
+ )
130
+ mask_time_length: int = field(
131
+ default=10,
132
+ metadata={"help": "Length of vector span to mask along the time axis."},
133
+ )
134
+ mask_feature_prob: float = field(
135
+ default=0.0,
136
+ metadata={
137
+ "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
138
+ "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
139
+ },
140
+ )
141
+ mask_feature_length: int = field(
142
+ default=10,
143
+ metadata={"help": "Length of vector span to mask along the feature axis."},
144
+ )
145
+ layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
146
+ ctc_loss_reduction: Optional[str] = field(
147
+ default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
148
+ )
149
+
150
+
151
+ @dataclass
152
+ class DataTrainingArguments:
153
+ """
154
+ Arguments pertaining to what data we are going to input our model for training and eval.
155
+
156
+ Using `HfArgumentParser` we can turn this class
157
+ into argparse arguments to be able to specify them on
158
+ the command line.
159
+ """
160
+
161
+ dataset_name: str = field(
162
+ metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
163
+ )
164
+ dataset_config_name: str = field(
165
+ default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
166
+ )
167
+ train_split_name: str = field(
168
+ default="train+validation",
169
+ metadata={
170
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
171
+ },
172
+ )
173
+ eval_split_name: str = field(
174
+ default="test",
175
+ metadata={
176
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
177
+ },
178
+ )
179
+ audio_column_name: str = field(
180
+ default="audio",
181
+ metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
182
+ )
183
+ text_column_name: str = field(
184
+ default="text",
185
+ metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
186
+ )
187
+ overwrite_cache: bool = field(
188
+ default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
189
+ )
190
+ preprocessing_num_workers: Optional[int] = field(
191
+ default=None,
192
+ metadata={"help": "The number of processes to use for the preprocessing."},
193
+ )
194
+ max_train_samples: Optional[int] = field(
195
+ default=None,
196
+ metadata={
197
+ "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
198
+ "value if set."
199
+ },
200
+ )
201
+ max_eval_samples: Optional[int] = field(
202
+ default=None,
203
+ metadata={
204
+ "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
205
+ "value if set."
206
+ },
207
+ )
208
+ chars_to_ignore: Optional[List[str]] = list_field(
209
+ default=None,
210
+ metadata={"help": "A list of characters to remove from the transcripts."},
211
+ )
212
+ eval_metrics: List[str] = list_field(
213
+ default=["wer"],
214
+ metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
215
+ )
216
+ max_duration_in_seconds: float = field(
217
+ default=20.0,
218
+ metadata={
219
+ "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
220
+ },
221
+ )
222
+ min_duration_in_seconds: float = field(
223
+ default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
224
+ )
225
+ preprocessing_only: bool = field(
226
+ default=False,
227
+ metadata={
228
+ "help": "Whether to only do data preprocessing and skip training. "
229
+ "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
230
+ "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
231
+ "so that the cached datasets can consequently be loaded in distributed training"
232
+ },
233
+ )
234
+ use_auth_token: bool = field(
235
+ default=False,
236
+ metadata={
237
+ "help": "If :obj:`True`, will use the token generated when running"
238
+ ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
239
+ },
240
+ )
241
+ unk_token: str = field(
242
+ default="[UNK]",
243
+ metadata={"help": "The unk token for the tokenizer"},
244
+ )
245
+ pad_token: str = field(
246
+ default="[PAD]",
247
+ metadata={"help": "The padding token for the tokenizer"},
248
+ )
249
+ word_delimiter_token: str = field(
250
+ default="|",
251
+ metadata={"help": "The word delimiter token for the tokenizer"},
252
+ )
253
+ phoneme_language: Optional[str] = field(
254
+ default=None,
255
+ metadata={
256
+ "help": "The target language that should be used be"
257
+ " passed to the tokenizer for tokenization. Note that"
258
+ " this is only relevant if the model classifies the"
259
+ " input audio to a sequence of phoneme sequences."
260
+ },
261
+ )
262
+
263
+
264
+ @dataclass
265
+ class DataCollatorCTCWithPadding:
266
+ """
267
+ Data collator that will dynamically pad the inputs received.
268
+ Args:
269
+ processor (:class:`~transformers.AutoProcessor`)
270
+ The processor used for proccessing the data.
271
+ padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
272
+ Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
273
+ among:
274
+ * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
275
+ sequence if provided).
276
+ * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
277
+ maximum acceptable input length for the model if that argument is not provided.
278
+ * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
279
+ different lengths).
280
+ max_length (:obj:`int`, `optional`):
281
+ Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
282
+ max_length_labels (:obj:`int`, `optional`):
283
+ Maximum length of the ``labels`` returned list and optionally padding length (see above).
284
+ pad_to_multiple_of (:obj:`int`, `optional`):
285
+ If set will pad the sequence to a multiple of the provided value.
286
+ This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
287
+ 7.5 (Volta).
288
+ """
289
+
290
+ processor: AutoProcessor
291
+ padding: Union[bool, str] = "longest"
292
+ pad_to_multiple_of: Optional[int] = None
293
+ pad_to_multiple_of_labels: Optional[int] = None
294
+
295
+ def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
296
+ # split inputs and labels since they have to be of different lenghts and need
297
+ # different padding methods
298
+ input_features = [{"input_values": feature["input_values"]} for feature in features]
299
+ label_features = [{"input_ids": feature["labels"]} for feature in features]
300
+
301
+ batch = self.processor.pad(
302
+ input_features,
303
+ padding=self.padding,
304
+ pad_to_multiple_of=self.pad_to_multiple_of,
305
+ return_tensors="pt",
306
+ )
307
+
308
+ with self.processor.as_target_processor():
309
+ labels_batch = self.processor.pad(
310
+ label_features,
311
+ padding=self.padding,
312
+ pad_to_multiple_of=self.pad_to_multiple_of_labels,
313
+ return_tensors="pt",
314
+ )
315
+
316
+ # replace padding with -100 to ignore loss correctly
317
+ labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
318
+
319
+ batch["labels"] = labels
320
+
321
+ return batch
322
+
323
+
324
+ def create_vocabulary_from_data(
325
+ datasets: DatasetDict,
326
+ word_delimiter_token: Optional[str] = None,
327
+ unk_token: Optional[str] = None,
328
+ pad_token: Optional[str] = None,
329
+ ):
330
+ # Given training and test labels create vocabulary
331
+ def extract_all_chars(batch):
332
+ all_text = " ".join(batch["target_text"])
333
+ vocab = list(set(all_text))
334
+ return {"vocab": [vocab], "all_text": [all_text]}
335
+
336
+ vocabs = datasets.map(
337
+ extract_all_chars,
338
+ batched=True,
339
+ batch_size=-1,
340
+ keep_in_memory=True,
341
+ remove_columns=datasets["train"].column_names,
342
+ )
343
+
344
+ # take union of all unique characters in each dataset
345
+ vocab_set = functools.reduce(
346
+ lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
347
+ )
348
+
349
+ vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
350
+
351
+ # replace white space with delimiter token
352
+ if word_delimiter_token is not None:
353
+ vocab_dict[word_delimiter_token] = vocab_dict[" "]
354
+ del vocab_dict[" "]
355
+
356
+ # add unk and pad token
357
+ if unk_token is not None:
358
+ vocab_dict[unk_token] = len(vocab_dict)
359
+
360
+ if pad_token is not None:
361
+ vocab_dict[pad_token] = len(vocab_dict)
362
+
363
+ return vocab_dict
364
+
365
+
366
+ def main():
367
+ # See all possible arguments in src/transformers/training_args.py
368
+ # or by passing the --help flag to this script.
369
+ # We now keep distinct sets of args, for a cleaner separation of concerns.
370
+
371
+ parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
372
+ if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
373
+ # If we pass only one argument to the script and it's the path to a json file,
374
+ # let's parse it to get our arguments.
375
+ model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
376
+ else:
377
+ model_args, data_args, training_args = parser.parse_args_into_dataclasses()
378
+
379
+ # Setup logging
380
+ logging.basicConfig(
381
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
382
+ datefmt="%m/%d/%Y %H:%M:%S",
383
+ handlers=[logging.StreamHandler(sys.stdout)],
384
+ )
385
+ logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
386
+
387
+ # Detecting last checkpoint.
388
+ last_checkpoint = None
389
+ if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
390
+ last_checkpoint = get_last_checkpoint(training_args.output_dir)
391
+ if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
392
+ raise ValueError(
393
+ f"Output directory ({training_args.output_dir}) already exists and is not empty. "
394
+ "Use --overwrite_output_dir to overcome."
395
+ )
396
+ elif last_checkpoint is not None:
397
+ logger.info(
398
+ f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
399
+ "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
400
+ )
401
+
402
+ # Log on each process the small summary:
403
+ logger.warning(
404
+ f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
405
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
406
+ )
407
+ # Set the verbosity to info of the Transformers logger (on main process only):
408
+ if is_main_process(training_args.local_rank):
409
+ transformers.utils.logging.set_verbosity_info()
410
+ logger.info("Training/evaluation parameters %s", training_args)
411
+
412
+ # Set seed before initializing model.
413
+ set_seed(training_args.seed)
414
+
415
+ # 1. First, let's load the dataset
416
+ te_dataset = get_telugu_dataset(validation_split=False)
417
+ def load_te_dataset(split):
418
+ return te_dataset[split]
419
+
420
+ raw_datasets = DatasetDict()
421
+
422
+ if training_args.do_train:
423
+ raw_datasets["train"] = load_te_dataset(
424
+ split=data_args.train_split_name
425
+ )
426
+
427
+ if data_args.audio_column_name not in raw_datasets["train"].column_names:
428
+ raise ValueError(
429
+ f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
430
+ "Make sure to set `--audio_column_name` to the correct audio column - one of "
431
+ f"{', '.join(raw_datasets['train'].column_names)}."
432
+ )
433
+
434
+ if data_args.text_column_name not in raw_datasets["train"].column_names:
435
+ raise ValueError(
436
+ f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
437
+ "Make sure to set `--text_column_name` to the correct text column - one of "
438
+ f"{', '.join(raw_datasets['train'].column_names)}."
439
+ )
440
+
441
+ if data_args.max_train_samples is not None:
442
+ raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
443
+
444
+ if training_args.do_eval:
445
+ raw_datasets["eval"] = load_te_dataset(
446
+ split=data_args.eval_split_name
447
+ )
448
+
449
+ if data_args.max_eval_samples is not None:
450
+ raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
451
+
452
+ # 2. We remove some special characters from the datasets
453
+ # that make training complicated and do not help in transcribing the speech
454
+ # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
455
+ # that could be easily picked up by the model
456
+ chars_to_ignore_regex = (
457
+ f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
458
+ )
459
+ text_column_name = data_args.text_column_name
460
+
461
+ def remove_special_characters(batch):
462
+ if chars_to_ignore_regex is not None:
463
+ batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
464
+ else:
465
+ batch["target_text"] = batch[text_column_name].lower() + " "
466
+ return batch
467
+
468
+ with training_args.main_process_first(desc="dataset map special characters removal"):
469
+ raw_datasets = raw_datasets.map(
470
+ remove_special_characters,
471
+ remove_columns=[text_column_name],
472
+ desc="remove special characters from datasets",
473
+ )
474
+
475
+ # save special tokens for tokenizer
476
+ word_delimiter_token = data_args.word_delimiter_token
477
+ unk_token = data_args.unk_token
478
+ pad_token = data_args.pad_token
479
+
480
+ # 3. Next, let's load the config as we might need it to create
481
+ # the tokenizer
482
+ # load config
483
+ config = AutoConfig.from_pretrained(
484
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
485
+ )
486
+
487
+ # 4. Next, if no tokenizer file is defined,
488
+ # we create the vocabulary of the model by extracting all unique characters from
489
+ # the training and evaluation datasets
490
+ # We need to make sure that only first rank saves vocabulary
491
+ # make sure all processes wait until vocab is created
492
+ tokenizer_name_or_path = model_args.tokenizer_name_or_path
493
+ tokenizer_kwargs = {}
494
+ if tokenizer_name_or_path is None:
495
+ # save vocab in training output dir
496
+ tokenizer_name_or_path = training_args.output_dir
497
+
498
+ vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
499
+
500
+ with training_args.main_process_first():
501
+ if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
502
+ os.remove(vocab_file)
503
+
504
+ with training_args.main_process_first(desc="dataset map vocabulary creation"):
505
+ if not os.path.isfile(vocab_file):
506
+ os.makedirs(tokenizer_name_or_path, exist_ok=True)
507
+ vocab_dict = create_vocabulary_from_data(
508
+ raw_datasets,
509
+ word_delimiter_token=word_delimiter_token,
510
+ unk_token=unk_token,
511
+ pad_token=pad_token,
512
+ )
513
+
514
+ # save vocab dict to be loaded into tokenizer
515
+ with open(vocab_file, "w") as file:
516
+ json.dump(vocab_dict, file)
517
+
518
+ # if tokenizer has just been created
519
+ # it is defined by `tokenizer_class` if present in config else by `model_type`
520
+ tokenizer_kwargs = {
521
+ "config": config if config.tokenizer_class is not None else None,
522
+ "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
523
+ "unk_token": unk_token,
524
+ "pad_token": pad_token,
525
+ "word_delimiter_token": word_delimiter_token,
526
+ }
527
+
528
+ # 5. Now we can instantiate the feature extractor, tokenizer and model
529
+ # Note for distributed training, the .from_pretrained methods guarantee that only
530
+ # one local process can concurrently download model & vocab.
531
+
532
+ # load feature_extractor and tokenizer
533
+ tokenizer = AutoTokenizer.from_pretrained(
534
+ tokenizer_name_or_path,
535
+ use_auth_token=data_args.use_auth_token,
536
+ **tokenizer_kwargs,
537
+ )
538
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
539
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
540
+ )
541
+
542
+ # adapt config
543
+ config.update(
544
+ {
545
+ "feat_proj_dropout": model_args.feat_proj_dropout,
546
+ "attention_dropout": model_args.attention_dropout,
547
+ "hidden_dropout": model_args.hidden_dropout,
548
+ "final_dropout": model_args.final_dropout,
549
+ "mask_time_prob": model_args.mask_time_prob,
550
+ "mask_time_length": model_args.mask_time_length,
551
+ "mask_feature_prob": model_args.mask_feature_prob,
552
+ "mask_feature_length": model_args.mask_feature_length,
553
+ "gradient_checkpointing": training_args.gradient_checkpointing,
554
+ "layerdrop": model_args.layerdrop,
555
+ "ctc_loss_reduction": model_args.ctc_loss_reduction,
556
+ "pad_token_id": tokenizer.pad_token_id,
557
+ "vocab_size": len(tokenizer),
558
+ "activation_dropout": model_args.activation_dropout,
559
+ }
560
+ )
561
+
562
+ # create model
563
+ model = AutoModelForCTC.from_pretrained(
564
+ model_args.model_name_or_path,
565
+ cache_dir=model_args.cache_dir,
566
+ config=config,
567
+ use_auth_token=data_args.use_auth_token,
568
+ )
569
+
570
+ # freeze encoder
571
+ if model_args.freeze_feature_encoder:
572
+ model.freeze_feature_encoder()
573
+
574
+ # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
575
+ # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
576
+ # so that we just need to set the correct target sampling rate and normalize the input
577
+ # via the `feature_extractor`
578
+
579
+ # make sure that dataset decodes audio with correct sampling rate
580
+ dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
581
+ if dataset_sampling_rate != feature_extractor.sampling_rate:
582
+ raw_datasets = raw_datasets.cast_column(
583
+ data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
584
+ )
585
+
586
+ # derive max & min input length for sample rate & max duration
587
+ max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
588
+ min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
589
+ audio_column_name = data_args.audio_column_name
590
+ num_workers = data_args.preprocessing_num_workers
591
+
592
+ # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
593
+ phoneme_language = data_args.phoneme_language
594
+
595
+ # Preprocessing the datasets.
596
+ # We need to read the audio files as arrays and tokenize the targets.
597
+ def prepare_dataset(batch):
598
+ # load audio
599
+ sample = batch[audio_column_name]
600
+
601
+ inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
602
+ batch["input_values"] = inputs.input_values[0]
603
+ batch["input_length"] = len(batch["input_values"])
604
+
605
+ # encode targets
606
+ additional_kwargs = {}
607
+ if phoneme_language is not None:
608
+ additional_kwargs["phonemizer_lang"] = phoneme_language
609
+
610
+ batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
611
+ return batch
612
+
613
+ with training_args.main_process_first(desc="dataset map preprocessing"):
614
+ vectorized_datasets = raw_datasets.map(
615
+ prepare_dataset,
616
+ remove_columns=next(iter(raw_datasets.values())).column_names,
617
+ num_proc=num_workers,
618
+ desc="preprocess datasets",
619
+ )
620
+
621
+ def is_audio_in_length_range(length):
622
+ return length > min_input_length and length < max_input_length
623
+
624
+ # filter data that is shorter than min_input_length
625
+ vectorized_datasets = vectorized_datasets.filter(
626
+ is_audio_in_length_range,
627
+ num_proc=num_workers,
628
+ input_columns=["input_length"],
629
+ )
630
+
631
+ # 7. Next, we can prepare the training.
632
+ # Let's use word error rate (WER) as our evaluation metric,
633
+ # instantiate a data collator and the trainer
634
+
635
+ # Define evaluation metrics during training, *i.e.* word error rate, character error rate
636
+ eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
637
+
638
+ # for large datasets it is advised to run the preprocessing on a
639
+ # single machine first with ``args.preprocessing_only`` since there will mostly likely
640
+ # be a timeout when running the script in distributed mode.
641
+ # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
642
+ # cached dataset
643
+ if data_args.preprocessing_only:
644
+ logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
645
+ return
646
+
647
+ def compute_metrics(pred):
648
+ pred_logits = pred.predictions
649
+ pred_ids = np.argmax(pred_logits, axis=-1)
650
+
651
+ pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
652
+
653
+ pred_str = tokenizer.batch_decode(pred_ids)
654
+ # we do not want to group tokens when computing the metrics
655
+ label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
656
+
657
+ metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
658
+
659
+ return metrics
660
+
661
+ # Now save everything to be able to create a single processor later
662
+ if is_main_process(training_args.local_rank):
663
+ # save feature extractor, tokenizer and config
664
+ feature_extractor.save_pretrained(training_args.output_dir)
665
+ tokenizer.save_pretrained(training_args.output_dir)
666
+ config.save_pretrained(training_args.output_dir)
667
+
668
+ try:
669
+ processor = AutoProcessor.from_pretrained(training_args.output_dir)
670
+ except (OSError, KeyError):
671
+ warnings.warn(
672
+ "Loading a processor from a feature extractor config that does not"
673
+ " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
674
+ " attribute to your `preprocessor_config.json` file to suppress this warning: "
675
+ " `'processor_class': 'Wav2Vec2Processor'`",
676
+ FutureWarning,
677
+ )
678
+ processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
679
+
680
+ # Instantiate custom data collator
681
+ data_collator = DataCollatorCTCWithPadding(processor=processor)
682
+
683
+ # Initialize Trainer
684
+ trainer = Trainer(
685
+ model=model,
686
+ data_collator=data_collator,
687
+ args=training_args,
688
+ compute_metrics=compute_metrics,
689
+ train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
690
+ eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
691
+ tokenizer=feature_extractor,
692
+ )
693
+
694
+ # 8. Finally, we can start training
695
+
696
+ # Training
697
+ if training_args.do_train:
698
+
699
+ # use last checkpoint if exist
700
+ if last_checkpoint is not None:
701
+ checkpoint = last_checkpoint
702
+ elif os.path.isdir(model_args.model_name_or_path):
703
+ checkpoint = model_args.model_name_or_path
704
+ else:
705
+ checkpoint = None
706
+
707
+ train_result = trainer.train(resume_from_checkpoint=checkpoint)
708
+ trainer.save_model()
709
+
710
+ metrics = train_result.metrics
711
+ max_train_samples = (
712
+ data_args.max_train_samples
713
+ if data_args.max_train_samples is not None
714
+ else len(vectorized_datasets["train"])
715
+ )
716
+ metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
717
+
718
+ trainer.log_metrics("train", metrics)
719
+ trainer.save_metrics("train", metrics)
720
+ trainer.save_state()
721
+
722
+ # Evaluation
723
+ results = {}
724
+ if training_args.do_eval:
725
+ logger.info("*** Evaluate ***")
726
+ metrics = trainer.evaluate()
727
+ max_eval_samples = (
728
+ data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
729
+ )
730
+ metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
731
+
732
+ trainer.log_metrics("eval", metrics)
733
+ trainer.save_metrics("eval", metrics)
734
+
735
+ # Write model card and (optionally) push to hub
736
+ config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
737
+ kwargs = {
738
+ "finetuned_from": model_args.model_name_or_path,
739
+ "tasks": "speech-recognition",
740
+ "tags": ["automatic-speech-recognition", data_args.dataset_name, "robust-speech-event"],
741
+ "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
742
+ "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
743
+ }
744
+ if "common_voice" in data_args.dataset_name:
745
+ kwargs["language"] = config_name
746
+
747
+ if training_args.push_to_hub:
748
+ trainer.push_to_hub(**kwargs)
749
+ else:
750
+ trainer.create_model_card(**kwargs)
751
+
752
+ return results
753
+
754
+
755
+ if __name__ == "__main__":
756
+ main()
run_speech_recognition_ctc_bnb.py ADDED
@@ -0,0 +1,780 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 bitsandbytes as bnb
34
+ import transformers
35
+ from transformers import (
36
+ AutoConfig,
37
+ AutoFeatureExtractor,
38
+ AutoModelForCTC,
39
+ AutoProcessor,
40
+ AutoTokenizer,
41
+ HfArgumentParser,
42
+ Trainer,
43
+ TrainingArguments,
44
+ Wav2Vec2Processor,
45
+ set_seed,
46
+ )
47
+ from transformers.trainer_pt_utils import get_parameter_names
48
+ from transformers.trainer_utils import get_last_checkpoint, is_main_process
49
+ from transformers.utils import check_min_version
50
+ from transformers.utils.versions import require_version
51
+
52
+
53
+ # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
54
+ check_min_version("4.16.0.dev0")
55
+
56
+ require_version("datasets>=1.13.3", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
57
+
58
+
59
+ logger = logging.getLogger(__name__)
60
+
61
+
62
+ def list_field(default=None, metadata=None):
63
+ return field(default_factory=lambda: default, metadata=metadata)
64
+
65
+ def get_telugu_dataset(validation_split=False):
66
+ dataset = load_dataset('openslr', 'SLR66')
67
+
68
+ seed=1242
69
+
70
+ if validation_split:
71
+ train_testvalid = dataset['train'].train_test_split(test_size=0.2, seed=seed)
72
+ # Split the 10% test + valid in half test, half valid
73
+ test_valid = train_testvalid['test'].train_test_split(test_size=0.33, seed=seed)
74
+ # gather everyone if you want to have a single DatasetDict
75
+ out_dataset = DatasetDict({
76
+ 'train': train_testvalid['train'],
77
+ 'test': test_valid['test'],
78
+ 'valid': test_valid['train']})
79
+ else:
80
+ train_testvalid = dataset['train'].train_test_split(test_size=0.25, seed=seed)
81
+ out_dataset = DatasetDict({
82
+ 'train': train_testvalid['train'],
83
+ 'test': train_testvalid['test']})
84
+ return out_dataset
85
+
86
+
87
+ @dataclass
88
+ class ModelArguments:
89
+ """
90
+ Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
91
+ """
92
+
93
+ model_name_or_path: str = field(
94
+ metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
95
+ )
96
+ tokenizer_name_or_path: Optional[str] = field(
97
+ default=None,
98
+ metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
99
+ )
100
+ cache_dir: Optional[str] = field(
101
+ default=None,
102
+ metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
103
+ )
104
+ freeze_feature_encoder: bool = field(
105
+ default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
106
+ )
107
+ attention_dropout: float = field(
108
+ default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
109
+ )
110
+ activation_dropout: float = field(
111
+ default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
112
+ )
113
+ feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
114
+ hidden_dropout: float = field(
115
+ default=0.0,
116
+ metadata={
117
+ "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
118
+ },
119
+ )
120
+ final_dropout: float = field(
121
+ default=0.0,
122
+ metadata={"help": "The dropout probability for the final projection layer."},
123
+ )
124
+ mask_time_prob: float = field(
125
+ default=0.05,
126
+ metadata={
127
+ "help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
128
+ "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
129
+ "vectors will be masked along the time axis."
130
+ },
131
+ )
132
+ mask_time_length: int = field(
133
+ default=10,
134
+ metadata={"help": "Length of vector span to mask along the time axis."},
135
+ )
136
+ mask_feature_prob: float = field(
137
+ default=0.0,
138
+ metadata={
139
+ "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
140
+ "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
141
+ },
142
+ )
143
+ mask_feature_length: int = field(
144
+ default=10,
145
+ metadata={"help": "Length of vector span to mask along the feature axis."},
146
+ )
147
+ layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
148
+ ctc_loss_reduction: Optional[str] = field(
149
+ default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
150
+ )
151
+
152
+
153
+ @dataclass
154
+ class DataTrainingArguments:
155
+ """
156
+ Arguments pertaining to what data we are going to input our model for training and eval.
157
+
158
+ Using `HfArgumentParser` we can turn this class
159
+ into argparse arguments to be able to specify them on
160
+ the command line.
161
+ """
162
+
163
+ dataset_name: str = field(
164
+ metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
165
+ )
166
+ dataset_config_name: str = field(
167
+ default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
168
+ )
169
+ train_split_name: str = field(
170
+ default="train+validation",
171
+ metadata={
172
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
173
+ },
174
+ )
175
+ eval_split_name: str = field(
176
+ default="test",
177
+ metadata={
178
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
179
+ },
180
+ )
181
+ audio_column_name: str = field(
182
+ default="audio",
183
+ metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
184
+ )
185
+ text_column_name: str = field(
186
+ default="text",
187
+ metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
188
+ )
189
+ overwrite_cache: bool = field(
190
+ default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
191
+ )
192
+ preprocessing_num_workers: Optional[int] = field(
193
+ default=None,
194
+ metadata={"help": "The number of processes to use for the preprocessing."},
195
+ )
196
+ max_train_samples: Optional[int] = field(
197
+ default=None,
198
+ metadata={
199
+ "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
200
+ "value if set."
201
+ },
202
+ )
203
+ max_eval_samples: Optional[int] = field(
204
+ default=None,
205
+ metadata={
206
+ "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
207
+ "value if set."
208
+ },
209
+ )
210
+ chars_to_ignore: Optional[List[str]] = list_field(
211
+ default=None,
212
+ metadata={"help": "A list of characters to remove from the transcripts."},
213
+ )
214
+ eval_metrics: List[str] = list_field(
215
+ default=["wer"],
216
+ metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
217
+ )
218
+ max_duration_in_seconds: float = field(
219
+ default=20.0,
220
+ metadata={
221
+ "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
222
+ },
223
+ )
224
+ min_duration_in_seconds: float = field(
225
+ default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
226
+ )
227
+ preprocessing_only: bool = field(
228
+ default=False,
229
+ metadata={
230
+ "help": "Whether to only do data preprocessing and skip training. "
231
+ "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
232
+ "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
233
+ "so that the cached datasets can consequently be loaded in distributed training"
234
+ },
235
+ )
236
+ use_auth_token: bool = field(
237
+ default=False,
238
+ metadata={
239
+ "help": "If :obj:`True`, will use the token generated when running"
240
+ ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
241
+ },
242
+ )
243
+ unk_token: str = field(
244
+ default="[UNK]",
245
+ metadata={"help": "The unk token for the tokenizer"},
246
+ )
247
+ pad_token: str = field(
248
+ default="[PAD]",
249
+ metadata={"help": "The padding token for the tokenizer"},
250
+ )
251
+ word_delimiter_token: str = field(
252
+ default="|",
253
+ metadata={"help": "The word delimiter token for the tokenizer"},
254
+ )
255
+ phoneme_language: Optional[str] = field(
256
+ default=None,
257
+ metadata={
258
+ "help": "The target language that should be used be"
259
+ " passed to the tokenizer for tokenization. Note that"
260
+ " this is only relevant if the model classifies the"
261
+ " input audio to a sequence of phoneme sequences."
262
+ },
263
+ )
264
+
265
+
266
+ @dataclass
267
+ class DataCollatorCTCWithPadding:
268
+ """
269
+ Data collator that will dynamically pad the inputs received.
270
+ Args:
271
+ processor (:class:`~transformers.AutoProcessor`)
272
+ The processor used for proccessing the data.
273
+ padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
274
+ Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
275
+ among:
276
+ * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
277
+ sequence if provided).
278
+ * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
279
+ maximum acceptable input length for the model if that argument is not provided.
280
+ * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
281
+ different lengths).
282
+ max_length (:obj:`int`, `optional`):
283
+ Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
284
+ max_length_labels (:obj:`int`, `optional`):
285
+ Maximum length of the ``labels`` returned list and optionally padding length (see above).
286
+ pad_to_multiple_of (:obj:`int`, `optional`):
287
+ If set will pad the sequence to a multiple of the provided value.
288
+ This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
289
+ 7.5 (Volta).
290
+ """
291
+
292
+ processor: AutoProcessor
293
+ padding: Union[bool, str] = "longest"
294
+ pad_to_multiple_of: Optional[int] = None
295
+ pad_to_multiple_of_labels: Optional[int] = None
296
+
297
+ def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
298
+ # split inputs and labels since they have to be of different lenghts and need
299
+ # different padding methods
300
+ input_features = [{"input_values": feature["input_values"]} for feature in features]
301
+ label_features = [{"input_ids": feature["labels"]} for feature in features]
302
+
303
+ batch = self.processor.pad(
304
+ input_features,
305
+ padding=self.padding,
306
+ pad_to_multiple_of=self.pad_to_multiple_of,
307
+ return_tensors="pt",
308
+ )
309
+
310
+ with self.processor.as_target_processor():
311
+ labels_batch = self.processor.pad(
312
+ label_features,
313
+ padding=self.padding,
314
+ pad_to_multiple_of=self.pad_to_multiple_of_labels,
315
+ return_tensors="pt",
316
+ )
317
+
318
+ # replace padding with -100 to ignore loss correctly
319
+ labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
320
+
321
+ batch["labels"] = labels
322
+
323
+ return batch
324
+
325
+
326
+ def create_vocabulary_from_data(
327
+ datasets: DatasetDict,
328
+ word_delimiter_token: Optional[str] = None,
329
+ unk_token: Optional[str] = None,
330
+ pad_token: Optional[str] = None,
331
+ ):
332
+ # Given training and test labels create vocabulary
333
+ def extract_all_chars(batch):
334
+ all_text = " ".join(batch["target_text"])
335
+ vocab = list(set(all_text))
336
+ return {"vocab": [vocab], "all_text": [all_text]}
337
+
338
+ vocabs = datasets.map(
339
+ extract_all_chars,
340
+ batched=True,
341
+ batch_size=-1,
342
+ keep_in_memory=True,
343
+ remove_columns=datasets["train"].column_names,
344
+ )
345
+
346
+ # take union of all unique characters in each dataset
347
+ vocab_set = functools.reduce(
348
+ lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
349
+ )
350
+
351
+ vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
352
+
353
+ # replace white space with delimiter token
354
+ if word_delimiter_token is not None:
355
+ vocab_dict[word_delimiter_token] = vocab_dict[" "]
356
+ del vocab_dict[" "]
357
+
358
+ # add unk and pad token
359
+ if unk_token is not None:
360
+ vocab_dict[unk_token] = len(vocab_dict)
361
+
362
+ if pad_token is not None:
363
+ vocab_dict[pad_token] = len(vocab_dict)
364
+
365
+ return vocab_dict
366
+
367
+
368
+ def main():
369
+ # See all possible arguments in src/transformers/training_args.py
370
+ # or by passing the --help flag to this script.
371
+ # We now keep distinct sets of args, for a cleaner separation of concerns.
372
+
373
+ parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
374
+ if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
375
+ # If we pass only one argument to the script and it's the path to a json file,
376
+ # let's parse it to get our arguments.
377
+ model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
378
+ else:
379
+ model_args, data_args, training_args = parser.parse_args_into_dataclasses()
380
+
381
+ # Setup logging
382
+ logging.basicConfig(
383
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
384
+ datefmt="%m/%d/%Y %H:%M:%S",
385
+ handlers=[logging.StreamHandler(sys.stdout)],
386
+ )
387
+ logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
388
+
389
+ # Detecting last checkpoint.
390
+ last_checkpoint = None
391
+ if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
392
+ last_checkpoint = get_last_checkpoint(training_args.output_dir)
393
+ if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
394
+ raise ValueError(
395
+ f"Output directory ({training_args.output_dir}) already exists and is not empty. "
396
+ "Use --overwrite_output_dir to overcome."
397
+ )
398
+ elif last_checkpoint is not None:
399
+ logger.info(
400
+ f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
401
+ "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
402
+ )
403
+
404
+
405
+ # Log on each process the small summary:
406
+ logger.warning(
407
+ f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
408
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
409
+ )
410
+ # Set the verbosity to info of the Transformers logger (on main process only):
411
+ if is_main_process(training_args.local_rank):
412
+ transformers.utils.logging.set_verbosity_info()
413
+ logger.info("Training/evaluation parameters %s", training_args)
414
+
415
+ # Set seed before initializing model.
416
+ set_seed(training_args.seed)
417
+
418
+ # 1. First, let's load the dataset
419
+ te_dataset = get_telugu_dataset(validation_split=False)
420
+ def load_te_dataset(split):
421
+ return te_dataset[split]
422
+
423
+ raw_datasets = DatasetDict()
424
+
425
+ if training_args.do_train:
426
+ raw_datasets["train"] = load_te_dataset(
427
+ split=data_args.train_split_name
428
+ )
429
+
430
+ if data_args.audio_column_name not in raw_datasets["train"].column_names:
431
+ raise ValueError(
432
+ f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
433
+ "Make sure to set `--audio_column_name` to the correct audio column - one of "
434
+ f"{', '.join(raw_datasets['train'].column_names)}."
435
+ )
436
+
437
+ if data_args.text_column_name not in raw_datasets["train"].column_names:
438
+ raise ValueError(
439
+ f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
440
+ "Make sure to set `--text_column_name` to the correct text column - one of "
441
+ f"{', '.join(raw_datasets['train'].column_names)}."
442
+ )
443
+
444
+ if data_args.max_train_samples is not None:
445
+ raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
446
+
447
+ if training_args.do_eval:
448
+ raw_datasets["eval"] = load_te_dataset(
449
+ split=data_args.eval_split_name
450
+ )
451
+
452
+ if data_args.max_eval_samples is not None:
453
+ raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
454
+
455
+ # 2. We remove some special characters from the datasets
456
+ # that make training complicated and do not help in transcribing the speech
457
+ # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
458
+ # that could be easily picked up by the model
459
+ chars_to_ignore_regex = (
460
+ f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
461
+ )
462
+ text_column_name = data_args.text_column_name
463
+
464
+ def remove_special_characters(batch):
465
+ if chars_to_ignore_regex is not None:
466
+ batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
467
+ else:
468
+ batch["target_text"] = batch[text_column_name].lower() + " "
469
+ return batch
470
+
471
+ with training_args.main_process_first(desc="dataset map special characters removal"):
472
+ raw_datasets = raw_datasets.map(
473
+ remove_special_characters,
474
+ remove_columns=[text_column_name],
475
+ desc="remove special characters from datasets",
476
+ )
477
+
478
+ # save special tokens for tokenizer
479
+ word_delimiter_token = data_args.word_delimiter_token
480
+ unk_token = data_args.unk_token
481
+ pad_token = data_args.pad_token
482
+
483
+ # 3. Next, let's load the config as we might need it to create
484
+ # the tokenizer
485
+ # load config
486
+ config = AutoConfig.from_pretrained(
487
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
488
+ )
489
+
490
+ # 4. Next, if no tokenizer file is defined,
491
+ # we create the vocabulary of the model by extracting all unique characters from
492
+ # the training and evaluation datasets
493
+ # We need to make sure that only first rank saves vocabulary
494
+ # make sure all processes wait until vocab is created
495
+ tokenizer_name_or_path = model_args.tokenizer_name_or_path
496
+ tokenizer_kwargs = {}
497
+ if tokenizer_name_or_path is None:
498
+ # save vocab in training output dir
499
+ tokenizer_name_or_path = training_args.output_dir
500
+
501
+ vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
502
+
503
+ with training_args.main_process_first():
504
+ if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
505
+ os.remove(vocab_file)
506
+
507
+ with training_args.main_process_first(desc="dataset map vocabulary creation"):
508
+ if not os.path.isfile(vocab_file):
509
+ os.makedirs(tokenizer_name_or_path, exist_ok=True)
510
+ vocab_dict = create_vocabulary_from_data(
511
+ raw_datasets,
512
+ word_delimiter_token=word_delimiter_token,
513
+ unk_token=unk_token,
514
+ pad_token=pad_token,
515
+ )
516
+
517
+ # save vocab dict to be loaded into tokenizer
518
+ with open(vocab_file, "w") as file:
519
+ json.dump(vocab_dict, file)
520
+
521
+ # if tokenizer has just been created
522
+ # it is defined by `tokenizer_class` if present in config else by `model_type`
523
+ tokenizer_kwargs = {
524
+ "config": config if config.tokenizer_class is not None else None,
525
+ "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
526
+ "unk_token": unk_token,
527
+ "pad_token": pad_token,
528
+ "word_delimiter_token": word_delimiter_token,
529
+ }
530
+
531
+ # 5. Now we can instantiate the feature extractor, tokenizer and model
532
+ # Note for distributed training, the .from_pretrained methods guarantee that only
533
+ # one local process can concurrently download model & vocab.
534
+
535
+ # load feature_extractor and tokenizer
536
+ tokenizer = AutoTokenizer.from_pretrained(
537
+ tokenizer_name_or_path,
538
+ use_auth_token=data_args.use_auth_token,
539
+ **tokenizer_kwargs,
540
+ )
541
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
542
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
543
+ )
544
+
545
+ # adapt config
546
+ config.update(
547
+ {
548
+ "feat_proj_dropout": model_args.feat_proj_dropout,
549
+ "attention_dropout": model_args.attention_dropout,
550
+ "hidden_dropout": model_args.hidden_dropout,
551
+ "final_dropout": model_args.final_dropout,
552
+ "mask_time_prob": model_args.mask_time_prob,
553
+ "mask_time_length": model_args.mask_time_length,
554
+ "mask_feature_prob": model_args.mask_feature_prob,
555
+ "mask_feature_length": model_args.mask_feature_length,
556
+ "gradient_checkpointing": training_args.gradient_checkpointing,
557
+ "layerdrop": model_args.layerdrop,
558
+ "ctc_loss_reduction": model_args.ctc_loss_reduction,
559
+ "pad_token_id": tokenizer.pad_token_id,
560
+ "vocab_size": len(tokenizer),
561
+ "activation_dropout": model_args.activation_dropout,
562
+ }
563
+ )
564
+
565
+ # create model
566
+ model = AutoModelForCTC.from_pretrained(
567
+ model_args.model_name_or_path,
568
+ cache_dir=model_args.cache_dir,
569
+ config=config,
570
+ use_auth_token=data_args.use_auth_token,
571
+ )
572
+
573
+ # freeze encoder
574
+ if model_args.freeze_feature_encoder:
575
+ model.freeze_feature_encoder()
576
+
577
+ # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
578
+ # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
579
+ # so that we just need to set the correct target sampling rate and normalize the input
580
+ # via the `feature_extractor`
581
+
582
+ # make sure that dataset decodes audio with correct sampling rate
583
+ dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
584
+ if dataset_sampling_rate != feature_extractor.sampling_rate:
585
+ raw_datasets = raw_datasets.cast_column(
586
+ data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
587
+ )
588
+
589
+ # derive max & min input length for sample rate & max duration
590
+ max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
591
+ min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
592
+ audio_column_name = data_args.audio_column_name
593
+ num_workers = data_args.preprocessing_num_workers
594
+
595
+ # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
596
+ phoneme_language = data_args.phoneme_language
597
+
598
+ # Preprocessing the datasets.
599
+ # We need to read the audio files as arrays and tokenize the targets.
600
+ def prepare_dataset(batch):
601
+ # load audio
602
+ sample = batch[audio_column_name]
603
+
604
+ inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
605
+ batch["input_values"] = inputs.input_values[0]
606
+ batch["input_length"] = len(batch["input_values"])
607
+
608
+ # encode targets
609
+ additional_kwargs = {}
610
+ if phoneme_language is not None:
611
+ additional_kwargs["phonemizer_lang"] = phoneme_language
612
+
613
+ batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
614
+ return batch
615
+
616
+ with training_args.main_process_first(desc="dataset map preprocessing"):
617
+ vectorized_datasets = raw_datasets.map(
618
+ prepare_dataset,
619
+ remove_columns=next(iter(raw_datasets.values())).column_names,
620
+ num_proc=num_workers,
621
+ desc="preprocess datasets",
622
+ )
623
+
624
+ def is_audio_in_length_range(length):
625
+ return length > min_input_length and length < max_input_length
626
+
627
+ # filter data that is shorter than min_input_length
628
+ vectorized_datasets = vectorized_datasets.filter(
629
+ is_audio_in_length_range,
630
+ num_proc=num_workers,
631
+ input_columns=["input_length"],
632
+ )
633
+
634
+ # 7. Next, we can prepare the training.
635
+ # Let's use word error rate (WER) as our evaluation metric,
636
+ # instantiate a data collator and the trainer
637
+
638
+ # Define evaluation metrics during training, *i.e.* word error rate, character error rate
639
+ eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
640
+
641
+ # for large datasets it is advised to run the preprocessing on a
642
+ # single machine first with ``args.preprocessing_only`` since there will mostly likely
643
+ # be a timeout when running the script in distributed mode.
644
+ # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
645
+ # cached dataset
646
+ if data_args.preprocessing_only:
647
+ logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
648
+ return
649
+
650
+ def compute_metrics(pred):
651
+ pred_logits = pred.predictions
652
+ pred_ids = np.argmax(pred_logits, axis=-1)
653
+
654
+ pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
655
+
656
+ pred_str = tokenizer.batch_decode(pred_ids)
657
+ # we do not want to group tokens when computing the metrics
658
+ label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
659
+
660
+ metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
661
+
662
+ return metrics
663
+
664
+ # Now save everything to be able to create a single processor later
665
+ if is_main_process(training_args.local_rank):
666
+ # save feature extractor, tokenizer and config
667
+ feature_extractor.save_pretrained(training_args.output_dir)
668
+ tokenizer.save_pretrained(training_args.output_dir)
669
+ config.save_pretrained(training_args.output_dir)
670
+
671
+ try:
672
+ processor = AutoProcessor.from_pretrained(training_args.output_dir)
673
+ except (OSError, KeyError):
674
+ warnings.warn(
675
+ "Loading a processor from a feature extractor config that does not"
676
+ " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
677
+ " attribute to your `preprocessor_config.json` file to suppress this warning: "
678
+ " `'processor_class': 'Wav2Vec2Processor'`",
679
+ FutureWarning,
680
+ )
681
+ processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
682
+
683
+ # Instantiate custom data collator
684
+ data_collator = DataCollatorCTCWithPadding(processor=processor)
685
+
686
+ decay_parameters = get_parameter_names(model, [torch.nn.LayerNorm])
687
+ decay_parameters = [name for name in decay_parameters if "bias" not in name]
688
+ optimizer_grouped_parameters = [
689
+ {
690
+ "params": [p for n, p in model.named_parameters() if n in decay_parameters],
691
+ "weight_decay": training_args.weight_decay,
692
+ },
693
+ {
694
+ "params": [p for n, p in model.named_parameters() if n not in decay_parameters],
695
+ "weight_decay": 0.0,
696
+ },
697
+ ]
698
+ optimizer = bnb.optim.Adam8bit(
699
+ params=optimizer_grouped_parameters,
700
+ betas=(training_args.adam_beta1, training_args.adam_beta2),
701
+ eps=training_args.adam_epsilon,
702
+ )
703
+
704
+ optimizers = (optimizer, None)
705
+
706
+ # Initialize Trainer
707
+ trainer = Trainer(
708
+ model=model,
709
+ data_collator=data_collator,
710
+ args=training_args,
711
+ compute_metrics=compute_metrics,
712
+ train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
713
+ eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
714
+ tokenizer=feature_extractor,
715
+ optimizers=optimizers,
716
+ )
717
+
718
+ # 8. Finally, we can start training
719
+
720
+ # Training
721
+ if training_args.do_train:
722
+
723
+ # use last checkpoint if exist
724
+ if last_checkpoint is not None:
725
+ checkpoint = last_checkpoint
726
+ elif os.path.isdir(model_args.model_name_or_path):
727
+ checkpoint = model_args.model_name_or_path
728
+ else:
729
+ checkpoint = None
730
+
731
+ train_result = trainer.train(resume_from_checkpoint=checkpoint)
732
+ trainer.save_model()
733
+
734
+ metrics = train_result.metrics
735
+ max_train_samples = (
736
+ data_args.max_train_samples
737
+ if data_args.max_train_samples is not None
738
+ else len(vectorized_datasets["train"])
739
+ )
740
+ metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
741
+
742
+ trainer.log_metrics("train", metrics)
743
+ trainer.save_metrics("train", metrics)
744
+ trainer.save_state()
745
+
746
+ # Evaluation
747
+ results = {}
748
+ if training_args.do_eval:
749
+ logger.info("*** Evaluate ***")
750
+ metrics = trainer.evaluate()
751
+ max_eval_samples = (
752
+ data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
753
+ )
754
+ metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
755
+
756
+ trainer.log_metrics("eval", metrics)
757
+ trainer.save_metrics("eval", metrics)
758
+
759
+ # Write model card and (optionally) push to hub
760
+ config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
761
+ kwargs = {
762
+ "finetuned_from": model_args.model_name_or_path,
763
+ "tasks": "speech-recognition",
764
+ "tags": ["automatic-speech-recognition", data_args.dataset_name, "robust-speech-event"],
765
+ "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
766
+ "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
767
+ }
768
+ if "common_voice" in data_args.dataset_name:
769
+ kwargs["language"] = config_name
770
+
771
+ if training_args.push_to_hub:
772
+ trainer.push_to_hub(**kwargs)
773
+ else:
774
+ trainer.create_model_card(**kwargs)
775
+
776
+ return results
777
+
778
+
779
+ if __name__ == "__main__":
780
+ main()
runs/Jan29_20-11-15_job-e3562c4a-f6a3-47dc-8131-7e578f1139a3/1643487216.0064042/events.out.tfevents.1643487216.job-e3562c4a-f6a3-47dc-8131-7e578f1139a3.1679549.1 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:edbc3f757e0d95c2819e3b1b77ea60f5f0ed2f009b9403ae23d0728dd9df0f8e
3
+ size 4753
runs/Jan29_20-11-15_job-e3562c4a-f6a3-47dc-8131-7e578f1139a3/events.out.tfevents.1643487215.job-e3562c4a-f6a3-47dc-8131-7e578f1139a3.1679549.0 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4d5aa6b543bd02f1330e1c9700551139120e8dba0d67ccf5821b66da691bcf68
3
+ size 43911
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"}
vocab.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"\\": 1, "_": 2, "e": 3, "g": 4, "l": 5, "n": 6, "p": 7, "r": 8, "s": 9, "t": 10, "ఁ": 11, "ం": 12, "ః": 13, "అ": 14, "ఆ": 15, "ఇ": 16, "ఈ": 17, "ఉ": 18, "ఊ": 19, "ఋ": 20, "ఎ": 21, "ఏ": 22, "ఐ": 23, "ఒ": 24, "ఓ": 25, "ఔ": 26, "క": 27, "ఖ": 28, "గ": 29, "ఘ": 30, "ఙ": 31, "చ": 32, "ఛ": 33, "జ": 34, "ఞ": 35, "ట": 36, "ఠ": 37, "డ": 38, "ఢ": 39, "ణ": 40, "త": 41, "థ": 42, "ద": 43, "ధ": 44, "న": 45, "ప": 46, "ఫ": 47, "బ": 48, "భ": 49, "మ": 50, "య": 51, "ర": 52, "ఱ": 53, "ల": 54, "ళ": 55, "వ": 56, "శ": 57, "ష": 58, "స": 59, "హ": 60, "ా": 61, "ి": 62, "ీ": 63, "ు": 64, "ూ": 65, "ృ": 66, "ె": 67, "ే": 68, "ై": 69, "ొ": 70, "ో": 71, "ౌ": 72, "్": 73, "‌": 74, "|": 0, "[UNK]": 75, "[PAD]": 76}