infinitejoy commited on
Commit
222a2a6
1 Parent(s): 788a138

training code

Browse files
Files changed (3) hide show
  1. eval.py +137 -0
  2. run.sh +34 -0
  3. run_speech_recognition_ctc.py +737 -0
eval.py ADDED
@@ -0,0 +1,137 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ import argparse
3
+ import re
4
+ from typing import Dict
5
+
6
+ import torch
7
+ from datasets import Audio, Dataset, load_dataset, load_metric
8
+
9
+ from transformers import AutoFeatureExtractor, pipeline
10
+
11
+
12
+ def log_results(result: Dataset, args: Dict[str, str]):
13
+ """DO NOT CHANGE. This function computes and logs the result metrics."""
14
+
15
+ log_outputs = args.log_outputs
16
+ dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
17
+
18
+ # load metric
19
+ wer = load_metric("wer")
20
+ cer = load_metric("cer")
21
+
22
+ # compute metrics
23
+ wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
24
+ cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
25
+
26
+ # print & log results
27
+ result_str = f"WER: {wer_result}\n" f"CER: {cer_result}"
28
+ print(result_str)
29
+
30
+ with open(f"{dataset_id}_eval_results.txt", "w") as f:
31
+ f.write(result_str)
32
+
33
+ # log all results in text file. Possibly interesting for analysis
34
+ if log_outputs is not None:
35
+ pred_file = f"log_{dataset_id}_predictions.txt"
36
+ target_file = f"log_{dataset_id}_targets.txt"
37
+
38
+ with open(pred_file, "w") as p, open(target_file, "w") as t:
39
+
40
+ # mapping function to write output
41
+ def write_to_file(batch, i):
42
+ p.write(f"{i}" + "\n")
43
+ p.write(batch["prediction"] + "\n")
44
+ t.write(f"{i}" + "\n")
45
+ t.write(batch["target"] + "\n")
46
+
47
+ result.map(write_to_file, with_indices=True)
48
+
49
+
50
+ def normalize_text(text: str) -> str:
51
+ """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
52
+
53
+ chars_to_ignore_regex = '[,?.!\-\;\:"“%‘”�—’…–]' # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
54
+
55
+ text = re.sub(chars_to_ignore_regex, "", text.lower())
56
+
57
+ # In addition, we can normalize the target text, e.g. removing new lines characters etc...
58
+ # note that order is important here!
59
+ token_sequences_to_ignore = ["\n\n", "\n", " ", " "]
60
+
61
+ for t in token_sequences_to_ignore:
62
+ text = " ".join(text.split(t))
63
+
64
+ return text
65
+
66
+
67
+ def main(args):
68
+ # load dataset
69
+ dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
70
+
71
+ # for testing: only process the first two examples as a test
72
+ # dataset = dataset.select(range(10))
73
+
74
+ # load processor
75
+ feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
76
+ sampling_rate = feature_extractor.sampling_rate
77
+
78
+ # resample audio
79
+ dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
80
+
81
+ # load eval pipeline
82
+ if args.device is None:
83
+ args.device = 0 if torch.cuda.is_available() else -1
84
+ asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
85
+
86
+ # map function to decode audio
87
+ def map_to_pred(batch):
88
+ prediction = asr(
89
+ batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
90
+ )
91
+
92
+ batch["prediction"] = prediction["text"]
93
+ batch["target"] = normalize_text(batch["sentence"])
94
+ return batch
95
+
96
+ # run inference on all examples
97
+ result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
98
+
99
+ # compute and log_results
100
+ # do not change function below
101
+ log_results(result, args)
102
+
103
+
104
+ if __name__ == "__main__":
105
+ parser = argparse.ArgumentParser()
106
+
107
+ parser.add_argument(
108
+ "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
109
+ )
110
+ parser.add_argument(
111
+ "--dataset",
112
+ type=str,
113
+ required=True,
114
+ help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
115
+ )
116
+ parser.add_argument(
117
+ "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'` for Common Voice"
118
+ )
119
+ parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
120
+ parser.add_argument(
121
+ "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
122
+ )
123
+ parser.add_argument(
124
+ "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
125
+ )
126
+ parser.add_argument(
127
+ "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
128
+ )
129
+ parser.add_argument(
130
+ "--device",
131
+ type=int,
132
+ default=None,
133
+ help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
134
+ )
135
+ args = parser.parse_args()
136
+
137
+ main(args)
run.sh ADDED
@@ -0,0 +1,34 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ python run_speech_recognition_ctc.py \
2
+ --dataset_name="mozilla-foundation/common_voice_7_0" \
3
+ --model_name_or_path="facebook/wav2vec2-xls-r-300m" \
4
+ --dataset_config_name="ha" \
5
+ --output_dir="./wav2vec2-large-xls-r-300m-hausa" \
6
+ --overwrite_output_dir \
7
+ --num_train_epochs="100" \
8
+ --per_device_train_batch_size="32" \
9
+ --per_device_eval_batch_size="32" \
10
+ --gradient_accumulation_steps="1" \
11
+ --learning_rate="7e-5" \
12
+ --warmup_steps="500" \
13
+ --length_column_name="input_length" \
14
+ --evaluation_strategy="steps" \
15
+ --text_column_name="sentence" \
16
+ --chars_to_ignore , ? . ! \- \; \: \" “ % ‘ ” � — ’ … – \
17
+ --save_steps="500" \
18
+ --eval_steps="500" \
19
+ --logging_steps="100" \
20
+ --layerdrop="0.0" \
21
+ --activation_dropout="0.1" \
22
+ --save_total_limit="2" \
23
+ --freeze_feature_encoder \
24
+ --feat_proj_dropout="0.0" \
25
+ --mask_time_prob="0.75" \
26
+ --mask_time_length="10" \
27
+ --mask_feature_prob="0.25" \
28
+ --mask_feature_length="64" \
29
+ --gradient_checkpointing \
30
+ --use_auth_token \
31
+ --fp16 \
32
+ --group_by_length \
33
+ --do_train --do_eval \
34
+ --push_to_hub
run_speech_recognition_ctc.py ADDED
@@ -0,0 +1,737 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python
2
+ # coding=utf-8
3
+ # Copyright 2021 The HuggingFace Inc. team. All rights reserved.
4
+ #
5
+ # Licensed under the Apache License, Version 2.0 (the "License");
6
+ # you may not use this file except in compliance with the License.
7
+ # You may obtain a copy of the License at
8
+ #
9
+ # http://www.apache.org/licenses/LICENSE-2.0
10
+ #
11
+ # Unless required by applicable law or agreed to in writing, software
12
+ # distributed under the License is distributed on an "AS IS" BASIS,
13
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14
+ # See the License for the specific language governing permissions and
15
+
16
+ """ Fine-tuning a 🤗 Transformers CTC model for automatic speech recognition"""
17
+
18
+ import functools
19
+ import json
20
+ import logging
21
+ import os
22
+ import re
23
+ import sys
24
+ import warnings
25
+ from dataclasses import dataclass, field
26
+ from typing import Dict, List, Optional, Union
27
+
28
+ import datasets
29
+ import numpy as np
30
+ import torch
31
+ from datasets import DatasetDict, load_dataset, load_metric
32
+
33
+ import transformers
34
+ from transformers import (
35
+ AutoConfig,
36
+ AutoFeatureExtractor,
37
+ AutoModelForCTC,
38
+ AutoProcessor,
39
+ AutoTokenizer,
40
+ HfArgumentParser,
41
+ Trainer,
42
+ TrainingArguments,
43
+ Wav2Vec2Processor,
44
+ set_seed,
45
+ )
46
+ from transformers.trainer_utils import get_last_checkpoint, is_main_process
47
+ from transformers.utils import check_min_version
48
+ from transformers.utils.versions import require_version
49
+
50
+
51
+ # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
52
+ check_min_version("4.16.0.dev0")
53
+
54
+ require_version("datasets>=1.13.3", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
55
+
56
+
57
+ logger = logging.getLogger(__name__)
58
+
59
+
60
+ def list_field(default=None, metadata=None):
61
+ return field(default_factory=lambda: default, metadata=metadata)
62
+
63
+
64
+ @dataclass
65
+ class ModelArguments:
66
+ """
67
+ Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
68
+ """
69
+
70
+ model_name_or_path: str = field(
71
+ metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
72
+ )
73
+ tokenizer_name_or_path: Optional[str] = field(
74
+ default=None,
75
+ metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
76
+ )
77
+ cache_dir: Optional[str] = field(
78
+ default=None,
79
+ metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
80
+ )
81
+ freeze_feature_encoder: bool = field(
82
+ default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
83
+ )
84
+ attention_dropout: float = field(
85
+ default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
86
+ )
87
+ activation_dropout: float = field(
88
+ default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
89
+ )
90
+ feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
91
+ hidden_dropout: float = field(
92
+ default=0.0,
93
+ metadata={
94
+ "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
95
+ },
96
+ )
97
+ final_dropout: float = field(
98
+ default=0.0,
99
+ metadata={"help": "The dropout probability for the final projection layer."},
100
+ )
101
+ mask_time_prob: float = field(
102
+ default=0.05,
103
+ metadata={
104
+ "help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
105
+ "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
106
+ "vectors will be masked along the time axis."
107
+ },
108
+ )
109
+ mask_time_length: int = field(
110
+ default=10,
111
+ metadata={"help": "Length of vector span to mask along the time axis."},
112
+ )
113
+ mask_feature_prob: float = field(
114
+ default=0.0,
115
+ metadata={
116
+ "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
117
+ "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
118
+ },
119
+ )
120
+ mask_feature_length: int = field(
121
+ default=10,
122
+ metadata={"help": "Length of vector span to mask along the feature axis."},
123
+ )
124
+ layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
125
+ ctc_loss_reduction: Optional[str] = field(
126
+ default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
127
+ )
128
+
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 'train'"
150
+ },
151
+ )
152
+ eval_split_name: str = field(
153
+ default="test",
154
+ metadata={
155
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
156
+ },
157
+ )
158
+ audio_column_name: str = field(
159
+ default="audio",
160
+ metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
161
+ )
162
+ text_column_name: str = field(
163
+ default="text",
164
+ metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
165
+ )
166
+ overwrite_cache: bool = field(
167
+ default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
168
+ )
169
+ preprocessing_num_workers: Optional[int] = field(
170
+ default=None,
171
+ metadata={"help": "The number of processes to use for the preprocessing."},
172
+ )
173
+ max_train_samples: Optional[int] = field(
174
+ default=None,
175
+ metadata={
176
+ "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
177
+ "value if set."
178
+ },
179
+ )
180
+ max_eval_samples: Optional[int] = field(
181
+ default=None,
182
+ metadata={
183
+ "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
184
+ "value if set."
185
+ },
186
+ )
187
+ chars_to_ignore: Optional[List[str]] = list_field(
188
+ default=None,
189
+ metadata={"help": "A list of characters to remove from the transcripts."},
190
+ )
191
+ eval_metrics: List[str] = list_field(
192
+ default=["wer"],
193
+ metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
194
+ )
195
+ max_duration_in_seconds: float = field(
196
+ default=20.0,
197
+ metadata={
198
+ "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
199
+ },
200
+ )
201
+ min_duration_in_seconds: float = field(
202
+ default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
203
+ )
204
+ preprocessing_only: bool = field(
205
+ default=False,
206
+ metadata={
207
+ "help": "Whether to only do data preprocessing and skip training. "
208
+ "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
209
+ "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
210
+ "so that the cached datasets can consequently be loaded in distributed training"
211
+ },
212
+ )
213
+ use_auth_token: bool = field(
214
+ default=False,
215
+ metadata={
216
+ "help": "If :obj:`True`, will use the token generated when running"
217
+ ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
218
+ },
219
+ )
220
+ unk_token: str = field(
221
+ default="[UNK]",
222
+ metadata={"help": "The unk token for the tokenizer"},
223
+ )
224
+ pad_token: str = field(
225
+ default="[PAD]",
226
+ metadata={"help": "The padding token for the tokenizer"},
227
+ )
228
+ word_delimiter_token: str = field(
229
+ default="|",
230
+ metadata={"help": "The word delimiter token for the tokenizer"},
231
+ )
232
+ phoneme_language: Optional[str] = field(
233
+ default=None,
234
+ metadata={
235
+ "help": "The target language that should be used be"
236
+ " passed to the tokenizer for tokenization. Note that"
237
+ " this is only relevant if the model classifies the"
238
+ " input audio to a sequence of phoneme sequences."
239
+ },
240
+ )
241
+
242
+
243
+ @dataclass
244
+ class DataCollatorCTCWithPadding:
245
+ """
246
+ Data collator that will dynamically pad the inputs received.
247
+ Args:
248
+ processor (:class:`~transformers.AutoProcessor`)
249
+ The processor used for proccessing the data.
250
+ padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
251
+ Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
252
+ among:
253
+ * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
254
+ sequence if provided).
255
+ * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
256
+ maximum acceptable input length for the model if that argument is not provided.
257
+ * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
258
+ different lengths).
259
+ max_length (:obj:`int`, `optional`):
260
+ Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
261
+ max_length_labels (:obj:`int`, `optional`):
262
+ Maximum length of the ``labels`` returned list and optionally padding length (see above).
263
+ pad_to_multiple_of (:obj:`int`, `optional`):
264
+ If set will pad the sequence to a multiple of the provided value.
265
+ This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
266
+ 7.5 (Volta).
267
+ """
268
+
269
+ processor: AutoProcessor
270
+ padding: Union[bool, str] = "longest"
271
+ pad_to_multiple_of: Optional[int] = None
272
+ pad_to_multiple_of_labels: Optional[int] = None
273
+
274
+ def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
275
+ # split inputs and labels since they have to be of different lenghts and need
276
+ # different padding methods
277
+ input_features = [{"input_values": feature["input_values"]} for feature in features]
278
+ label_features = [{"input_ids": feature["labels"]} for feature in features]
279
+
280
+ batch = self.processor.pad(
281
+ input_features,
282
+ padding=self.padding,
283
+ pad_to_multiple_of=self.pad_to_multiple_of,
284
+ return_tensors="pt",
285
+ )
286
+
287
+ with self.processor.as_target_processor():
288
+ labels_batch = self.processor.pad(
289
+ label_features,
290
+ padding=self.padding,
291
+ pad_to_multiple_of=self.pad_to_multiple_of_labels,
292
+ return_tensors="pt",
293
+ )
294
+
295
+ # replace padding with -100 to ignore loss correctly
296
+ labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
297
+
298
+ batch["labels"] = labels
299
+
300
+ return batch
301
+
302
+
303
+ def create_vocabulary_from_data(
304
+ datasets: DatasetDict,
305
+ word_delimiter_token: Optional[str] = None,
306
+ unk_token: Optional[str] = None,
307
+ pad_token: Optional[str] = None,
308
+ ):
309
+ # Given training and test labels create vocabulary
310
+ def extract_all_chars(batch):
311
+ all_text = " ".join(batch["target_text"])
312
+ vocab = list(set(all_text))
313
+ return {"vocab": [vocab], "all_text": [all_text]}
314
+
315
+ vocabs = datasets.map(
316
+ extract_all_chars,
317
+ batched=True,
318
+ batch_size=-1,
319
+ keep_in_memory=True,
320
+ remove_columns=datasets["train"].column_names,
321
+ )
322
+
323
+ # take union of all unique characters in each dataset
324
+ vocab_set = functools.reduce(
325
+ lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
326
+ )
327
+
328
+ vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
329
+
330
+ # replace white space with delimiter token
331
+ if word_delimiter_token is not None:
332
+ vocab_dict[word_delimiter_token] = vocab_dict[" "]
333
+ del vocab_dict[" "]
334
+
335
+ # add unk and pad token
336
+ if unk_token is not None:
337
+ vocab_dict[unk_token] = len(vocab_dict)
338
+
339
+ if pad_token is not None:
340
+ vocab_dict[pad_token] = len(vocab_dict)
341
+
342
+ return vocab_dict
343
+
344
+
345
+ def main():
346
+ # See all possible arguments in src/transformers/training_args.py
347
+ # or by passing the --help flag to this script.
348
+ # We now keep distinct sets of args, for a cleaner separation of concerns.
349
+
350
+ parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
351
+ if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
352
+ # If we pass only one argument to the script and it's the path to a json file,
353
+ # let's parse it to get our arguments.
354
+ model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
355
+ else:
356
+ model_args, data_args, training_args = parser.parse_args_into_dataclasses()
357
+
358
+ # Detecting last checkpoint.
359
+ last_checkpoint = None
360
+ if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
361
+ last_checkpoint = get_last_checkpoint(training_args.output_dir)
362
+ if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
363
+ raise ValueError(
364
+ f"Output directory ({training_args.output_dir}) already exists and is not empty. "
365
+ "Use --overwrite_output_dir to overcome."
366
+ )
367
+ elif last_checkpoint is not None:
368
+ logger.info(
369
+ f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
370
+ "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
371
+ )
372
+
373
+ # Setup logging
374
+ logging.basicConfig(
375
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
376
+ datefmt="%m/%d/%Y %H:%M:%S",
377
+ handlers=[logging.StreamHandler(sys.stdout)],
378
+ )
379
+ logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
380
+
381
+ # Log on each process the small summary:
382
+ logger.warning(
383
+ f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
384
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
385
+ )
386
+ # Set the verbosity to info of the Transformers logger (on main process only):
387
+ if is_main_process(training_args.local_rank):
388
+ transformers.utils.logging.set_verbosity_info()
389
+ logger.info("Training/evaluation parameters %s", training_args)
390
+
391
+ # Set seed before initializing model.
392
+ set_seed(training_args.seed)
393
+
394
+ # 1. First, let's load the dataset
395
+ raw_datasets = DatasetDict()
396
+
397
+ if training_args.do_train:
398
+ raw_datasets["train"] = load_dataset(
399
+ data_args.dataset_name,
400
+ data_args.dataset_config_name,
401
+ split=data_args.train_split_name,
402
+ use_auth_token=data_args.use_auth_token,
403
+ )
404
+
405
+ if data_args.audio_column_name not in raw_datasets["train"].column_names:
406
+ raise ValueError(
407
+ f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
408
+ "Make sure to set `--audio_column_name` to the correct audio column - one of "
409
+ f"{', '.join(raw_datasets['train'].column_names)}."
410
+ )
411
+
412
+ if data_args.text_column_name not in raw_datasets["train"].column_names:
413
+ raise ValueError(
414
+ f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
415
+ "Make sure to set `--text_column_name` to the correct text column - one of "
416
+ f"{', '.join(raw_datasets['train'].column_names)}."
417
+ )
418
+
419
+ if data_args.max_train_samples is not None:
420
+ raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
421
+
422
+ if training_args.do_eval:
423
+ raw_datasets["eval"] = load_dataset(
424
+ data_args.dataset_name,
425
+ data_args.dataset_config_name,
426
+ split=data_args.eval_split_name,
427
+ use_auth_token=data_args.use_auth_token,
428
+ )
429
+
430
+ if data_args.max_eval_samples is not None:
431
+ raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
432
+
433
+ # 2. We remove some special characters from the datasets
434
+ # that make training complicated and do not help in transcribing the speech
435
+ # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
436
+ # that could be easily picked up by the model
437
+ chars_to_ignore_regex = (
438
+ f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
439
+ )
440
+ text_column_name = data_args.text_column_name
441
+
442
+ def remove_special_characters(batch):
443
+ if chars_to_ignore_regex is not None:
444
+ batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
445
+ else:
446
+ batch["target_text"] = batch[text_column_name].lower() + " "
447
+ return batch
448
+
449
+ with training_args.main_process_first(desc="dataset map special characters removal"):
450
+ raw_datasets = raw_datasets.map(
451
+ remove_special_characters,
452
+ remove_columns=[text_column_name],
453
+ desc="remove special characters from datasets",
454
+ )
455
+
456
+ # save special tokens for tokenizer
457
+ word_delimiter_token = data_args.word_delimiter_token
458
+ unk_token = data_args.unk_token
459
+ pad_token = data_args.pad_token
460
+
461
+ # 3. Next, let's load the config as we might need it to create
462
+ # the tokenizer
463
+ # load config
464
+ config = AutoConfig.from_pretrained(
465
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
466
+ )
467
+
468
+ # 4. Next, if no tokenizer file is defined,
469
+ # we create the vocabulary of the model by extracting all unique characters from
470
+ # the training and evaluation datasets
471
+ # We need to make sure that only first rank saves vocabulary
472
+ # make sure all processes wait until vocab is created
473
+ tokenizer_name_or_path = model_args.tokenizer_name_or_path
474
+ tokenizer_kwargs = {}
475
+ if tokenizer_name_or_path is None:
476
+ # save vocab in training output dir
477
+ tokenizer_name_or_path = training_args.output_dir
478
+
479
+ vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
480
+
481
+ with training_args.main_process_first():
482
+ if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
483
+ os.remove(vocab_file)
484
+
485
+ with training_args.main_process_first(desc="dataset map vocabulary creation"):
486
+ if not os.path.isfile(vocab_file):
487
+ os.makedirs(tokenizer_name_or_path, exist_ok=True)
488
+ vocab_dict = create_vocabulary_from_data(
489
+ raw_datasets,
490
+ word_delimiter_token=word_delimiter_token,
491
+ unk_token=unk_token,
492
+ pad_token=pad_token,
493
+ )
494
+
495
+ # save vocab dict to be loaded into tokenizer
496
+ with open(vocab_file, "w") as file:
497
+ json.dump(vocab_dict, file)
498
+
499
+ # if tokenizer has just been created
500
+ # it is defined by `tokenizer_class` if present in config else by `model_type`
501
+ tokenizer_kwargs = {
502
+ "config": config if config.tokenizer_class is not None else None,
503
+ "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
504
+ "unk_token": unk_token,
505
+ "pad_token": pad_token,
506
+ "word_delimiter_token": word_delimiter_token,
507
+ }
508
+
509
+ # 5. Now we can instantiate the feature extractor, tokenizer and model
510
+ # Note for distributed training, the .from_pretrained methods guarantee that only
511
+ # one local process can concurrently download model & vocab.
512
+
513
+ # load feature_extractor and tokenizer
514
+ tokenizer = AutoTokenizer.from_pretrained(
515
+ tokenizer_name_or_path,
516
+ use_auth_token=data_args.use_auth_token,
517
+ **tokenizer_kwargs,
518
+ )
519
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
520
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
521
+ )
522
+
523
+ # adapt config
524
+ config.update(
525
+ {
526
+ "feat_proj_dropout": model_args.feat_proj_dropout,
527
+ "attention_dropout": model_args.attention_dropout,
528
+ "hidden_dropout": model_args.hidden_dropout,
529
+ "final_dropout": model_args.final_dropout,
530
+ "mask_time_prob": model_args.mask_time_prob,
531
+ "mask_time_length": model_args.mask_time_length,
532
+ "mask_feature_prob": model_args.mask_feature_prob,
533
+ "mask_feature_length": model_args.mask_feature_length,
534
+ "gradient_checkpointing": training_args.gradient_checkpointing,
535
+ "layerdrop": model_args.layerdrop,
536
+ "ctc_loss_reduction": model_args.ctc_loss_reduction,
537
+ "pad_token_id": tokenizer.pad_token_id,
538
+ "vocab_size": len(tokenizer),
539
+ "activation_dropout": model_args.activation_dropout,
540
+ }
541
+ )
542
+
543
+ # create model
544
+ model = AutoModelForCTC.from_pretrained(
545
+ model_args.model_name_or_path,
546
+ cache_dir=model_args.cache_dir,
547
+ config=config,
548
+ use_auth_token=data_args.use_auth_token,
549
+ )
550
+
551
+ # freeze encoder
552
+ if model_args.freeze_feature_encoder:
553
+ model.freeze_feature_encoder()
554
+
555
+ # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
556
+ # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
557
+ # so that we just need to set the correct target sampling rate and normalize the input
558
+ # via the `feature_extractor`
559
+
560
+ # make sure that dataset decodes audio with correct sampling rate
561
+ dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
562
+ if dataset_sampling_rate != feature_extractor.sampling_rate:
563
+ raw_datasets = raw_datasets.cast_column(
564
+ data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
565
+ )
566
+
567
+ # derive max & min input length for sample rate & max duration
568
+ max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
569
+ min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
570
+ audio_column_name = data_args.audio_column_name
571
+ num_workers = data_args.preprocessing_num_workers
572
+
573
+ # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
574
+ phoneme_language = data_args.phoneme_language
575
+
576
+ # Preprocessing the datasets.
577
+ # We need to read the audio files as arrays and tokenize the targets.
578
+ def prepare_dataset(batch):
579
+ # load audio
580
+ sample = batch[audio_column_name]
581
+
582
+ inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
583
+ batch["input_values"] = inputs.input_values[0]
584
+ batch["input_length"] = len(batch["input_values"])
585
+
586
+ # encode targets
587
+ additional_kwargs = {}
588
+ if phoneme_language is not None:
589
+ additional_kwargs["phonemizer_lang"] = phoneme_language
590
+
591
+ batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
592
+ return batch
593
+
594
+ with training_args.main_process_first(desc="dataset map preprocessing"):
595
+ vectorized_datasets = raw_datasets.map(
596
+ prepare_dataset,
597
+ remove_columns=next(iter(raw_datasets.values())).column_names,
598
+ num_proc=num_workers,
599
+ desc="preprocess datasets",
600
+ )
601
+
602
+ def is_audio_in_length_range(length):
603
+ return length > min_input_length and length < max_input_length
604
+
605
+ # filter data that is shorter than min_input_length
606
+ vectorized_datasets = vectorized_datasets.filter(
607
+ is_audio_in_length_range,
608
+ num_proc=num_workers,
609
+ input_columns=["input_length"],
610
+ )
611
+
612
+ # 7. Next, we can prepare the training.
613
+ # Let's use word error rate (WER) as our evaluation metric,
614
+ # instantiate a data collator and the trainer
615
+
616
+ # Define evaluation metrics during training, *i.e.* word error rate, character error rate
617
+ eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
618
+
619
+ # for large datasets it is advised to run the preprocessing on a
620
+ # single machine first with ``args.preprocessing_only`` since there will mostly likely
621
+ # be a timeout when running the script in distributed mode.
622
+ # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
623
+ # cached dataset
624
+ if data_args.preprocessing_only:
625
+ logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
626
+ return
627
+
628
+ def compute_metrics(pred):
629
+ pred_logits = pred.predictions
630
+ pred_ids = np.argmax(pred_logits, axis=-1)
631
+
632
+ pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
633
+
634
+ pred_str = tokenizer.batch_decode(pred_ids)
635
+ # we do not want to group tokens when computing the metrics
636
+ label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
637
+
638
+ metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
639
+
640
+ return metrics
641
+
642
+ # Now save everything to be able to create a single processor later
643
+ if is_main_process(training_args.local_rank):
644
+ # save feature extractor, tokenizer and config
645
+ feature_extractor.save_pretrained(training_args.output_dir)
646
+ tokenizer.save_pretrained(training_args.output_dir)
647
+ config.save_pretrained(training_args.output_dir)
648
+
649
+ try:
650
+ processor = AutoProcessor.from_pretrained(training_args.output_dir)
651
+ except (OSError, KeyError):
652
+ warnings.warn(
653
+ "Loading a processor from a feature extractor config that does not"
654
+ " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
655
+ " attribute to your `preprocessor_config.json` file to suppress this warning: "
656
+ " `'processor_class': 'Wav2Vec2Processor'`",
657
+ FutureWarning,
658
+ )
659
+ processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
660
+
661
+ # Instantiate custom data collator
662
+ data_collator = DataCollatorCTCWithPadding(processor=processor)
663
+
664
+ # Initialize Trainer
665
+ trainer = Trainer(
666
+ model=model,
667
+ data_collator=data_collator,
668
+ args=training_args,
669
+ compute_metrics=compute_metrics,
670
+ train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
671
+ eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
672
+ tokenizer=feature_extractor,
673
+ )
674
+
675
+ # 8. Finally, we can start training
676
+
677
+ # Training
678
+ if training_args.do_train:
679
+
680
+ # use last checkpoint if exist
681
+ if last_checkpoint is not None:
682
+ checkpoint = last_checkpoint
683
+ elif os.path.isdir(model_args.model_name_or_path):
684
+ checkpoint = model_args.model_name_or_path
685
+ else:
686
+ checkpoint = None
687
+
688
+ train_result = trainer.train(resume_from_checkpoint=checkpoint)
689
+ trainer.save_model()
690
+
691
+ metrics = train_result.metrics
692
+ max_train_samples = (
693
+ data_args.max_train_samples
694
+ if data_args.max_train_samples is not None
695
+ else len(vectorized_datasets["train"])
696
+ )
697
+ metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
698
+
699
+ trainer.log_metrics("train", metrics)
700
+ trainer.save_metrics("train", metrics)
701
+ trainer.save_state()
702
+
703
+ # Evaluation
704
+ results = {}
705
+ if training_args.do_eval:
706
+ logger.info("*** Evaluate ***")
707
+ metrics = trainer.evaluate()
708
+ max_eval_samples = (
709
+ data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
710
+ )
711
+ metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
712
+
713
+ trainer.log_metrics("eval", metrics)
714
+ trainer.save_metrics("eval", metrics)
715
+
716
+ # Write model card and (optionally) push to hub
717
+ config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
718
+ kwargs = {
719
+ "finetuned_from": model_args.model_name_or_path,
720
+ "tasks": "speech-recognition",
721
+ "tags": ["automatic-speech-recognition", data_args.dataset_name],
722
+ "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
723
+ "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
724
+ }
725
+ if "common_voice" in data_args.dataset_name:
726
+ kwargs["language"] = config_name
727
+
728
+ if training_args.push_to_hub:
729
+ trainer.push_to_hub(**kwargs)
730
+ else:
731
+ trainer.create_model_card(**kwargs)
732
+
733
+ return results
734
+
735
+
736
+ if __name__ == "__main__":
737
+ main()