Plim commited on
Commit
1fb32b9
1 Parent(s): 6539a9a

Training in progress, step 1000

Browse files
.gitignore ADDED
@@ -0,0 +1 @@
 
 
1
+ checkpoint-*/
.ipynb_checkpoints/eval-checkpoint.py ADDED
@@ -0,0 +1,128 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ import argparse
3
+ import re
4
+ from typing import Dict
5
+
6
+ from datasets import Audio, Dataset, load_dataset, load_metric
7
+
8
+ from transformers import AutoFeatureExtractor, pipeline
9
+
10
+
11
+ def log_results(result: Dataset, args: Dict[str, str]):
12
+ """DO NOT CHANGE. This function computes and logs the result metrics."""
13
+
14
+ log_outputs = args.log_outputs
15
+ dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
16
+
17
+ # load metric
18
+ wer = load_metric("wer")
19
+ cer = load_metric("cer")
20
+
21
+ # compute metrics
22
+ wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
23
+ cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
24
+
25
+ # print & log results
26
+ result_str = f"WER: {wer_result}\n" f"CER: {cer_result}"
27
+ print(result_str)
28
+
29
+ with open(f"{dataset_id}_eval_results.txt", "w") as f:
30
+ f.write(result_str)
31
+
32
+ # log all results in text file. Possibly interesting for analysis
33
+ if log_outputs is not None:
34
+ pred_file = f"log_{dataset_id}_predictions.txt"
35
+ target_file = f"log_{dataset_id}_targets.txt"
36
+
37
+ with open(pred_file, "w") as p, open(target_file, "w") as t:
38
+
39
+ # mapping function to write output
40
+ def write_to_file(batch, i):
41
+ p.write(f"{i}" + "\n")
42
+ p.write(batch["prediction"] + "\n")
43
+ t.write(f"{i}" + "\n")
44
+ t.write(batch["target"] + "\n")
45
+
46
+ result.map(write_to_file, with_indices=True)
47
+
48
+
49
+ def normalize_text(text: str) -> str:
50
+ """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
51
+
52
+ chars_to_ignore_regex = '[^a-zàâäçéèêëîïôöùûüÿ\'’ ]' # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
53
+
54
+ text = re.sub(chars_to_ignore_regex, "", text.lower()).replace('’', "'")
55
+
56
+ # In addition, we can normalize the target text, e.g. removing new lines characters etc...
57
+ # note that order is important here!
58
+ token_sequences_to_ignore = ["\n\n", "\n", " ", " "]
59
+
60
+ for t in token_sequences_to_ignore:
61
+ text = " ".join(text.split(t))
62
+
63
+ return text
64
+
65
+
66
+ def main(args):
67
+ # load dataset
68
+ dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
69
+
70
+ # for testing: only process the first two examples as a test
71
+ # dataset = dataset.select(range(10))
72
+
73
+ # load processor
74
+ feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
75
+ sampling_rate = feature_extractor.sampling_rate
76
+
77
+ # resample audio
78
+ dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
79
+
80
+ # load eval pipeline
81
+ asr = pipeline("automatic-speech-recognition", model=args.model_id)
82
+
83
+ # map function to decode audio
84
+ def map_to_pred(batch):
85
+ prediction = asr(
86
+ batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
87
+ )
88
+
89
+ batch["prediction"] = prediction["text"]
90
+ batch["target"] = normalize_text(batch["sentence"])
91
+ return batch
92
+
93
+ # run inference on all examples
94
+ result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
95
+
96
+ # compute and log_results
97
+ # do not change function below
98
+ log_results(result, args)
99
+
100
+
101
+ if __name__ == "__main__":
102
+ parser = argparse.ArgumentParser()
103
+
104
+ parser.add_argument(
105
+ "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
106
+ )
107
+ parser.add_argument(
108
+ "--dataset",
109
+ type=str,
110
+ required=True,
111
+ help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
112
+ )
113
+ parser.add_argument(
114
+ "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'` for Common Voice"
115
+ )
116
+ parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
117
+ parser.add_argument(
118
+ "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
119
+ )
120
+ parser.add_argument(
121
+ "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
122
+ )
123
+ parser.add_argument(
124
+ "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
125
+ )
126
+ args = parser.parse_args()
127
+
128
+ main(args)
.ipynb_checkpoints/run-checkpoint.sh ADDED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ export WANDB_PROJECT="xls-r-300m-cv_8-fr"
2
+ python run_speech_recognition_ctc.py \
3
+ --activation_dropout="0.1" \
4
+ --dataset_name="mozilla-foundation/common_voice_8_0" \
5
+ --dataset_config_name="fr" \
6
+ --eval_steps="1000" \
7
+ --evaluation_strategy="steps" \
8
+ --feat_proj_dropout="0.0" \
9
+ --freeze_feature_encoder \
10
+ --fp16 \
11
+ --gradient_accumulation_steps="8" \
12
+ --gradient_checkpointing \
13
+ --group_by_length \
14
+ --layerdrop="0.0" \
15
+ --learning_rate="7.5e-5" \
16
+ --length_column_name="input_length" \
17
+ --load_best_model_at_end \
18
+ --logging_steps="100" \
19
+ --mask_feature_length="64" \
20
+ --mask_feature_prob="0.25" \
21
+ --mask_time_length="10" \
22
+ --mask_time_prob="0.75" \
23
+ --model_name_or_path="facebook/wav2vec2-xls-r-300m" \
24
+ --num_train_epochs="5.0" \
25
+ --output_dir="./" \
26
+ --overwrite_output_dir \
27
+ --per_device_train_batch_size="16" \
28
+ --per_device_eval_batch_size="16" \
29
+ --preprocessing_num_workers="4" \
30
+ --push_to_hub \
31
+ --report_to="wandb" \
32
+ --save_steps="1000" \
33
+ --save_total_limit="3" \
34
+ --text_column_name="sentence" \
35
+ --use_auth_token \
36
+ --warmup_steps="2000" \
37
+ --do_train --do_eval
.ipynb_checkpoints/run_speech_recognition_ctc-checkpoint.py ADDED
@@ -0,0 +1,734 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 'test'"
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 = '[^a-zàâäçéèêëîïôöùûüÿ\'’ ]'
438
+ text_column_name = data_args.text_column_name
439
+
440
+ def remove_and_replace_special_characters(batch):
441
+ batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name].lower()).replace('’', "'") + " "
442
+ return batch
443
+
444
+ with training_args.main_process_first(desc="dataset map special characters removal"):
445
+ raw_datasets = raw_datasets.map(
446
+ remove_and_replace_special_characters,
447
+ remove_columns=[text_column_name],
448
+ desc="remove special characters from datasets",
449
+ )
450
+
451
+ # save special tokens for tokenizer
452
+ word_delimiter_token = data_args.word_delimiter_token
453
+ unk_token = data_args.unk_token
454
+ pad_token = data_args.pad_token
455
+
456
+ # 3. Next, let's load the config as we might need it to create
457
+ # the tokenizer
458
+ # load config
459
+ config = AutoConfig.from_pretrained(
460
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
461
+ )
462
+
463
+ # 4. Next, if no tokenizer file is defined,
464
+ # we create the vocabulary of the model by extracting all unique characters from
465
+ # the training and evaluation datasets
466
+ # We need to make sure that only first rank saves vocabulary
467
+ # make sure all processes wait until vocab is created
468
+ tokenizer_name_or_path = model_args.tokenizer_name_or_path
469
+ tokenizer_kwargs = {}
470
+ if tokenizer_name_or_path is None:
471
+ # save vocab in training output dir
472
+ tokenizer_name_or_path = training_args.output_dir
473
+
474
+ vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
475
+
476
+ with training_args.main_process_first():
477
+ if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
478
+ os.remove(vocab_file)
479
+
480
+ with training_args.main_process_first(desc="dataset map vocabulary creation"):
481
+ if not os.path.isfile(vocab_file):
482
+ os.makedirs(tokenizer_name_or_path, exist_ok=True)
483
+ vocab_dict = create_vocabulary_from_data(
484
+ raw_datasets,
485
+ word_delimiter_token=word_delimiter_token,
486
+ unk_token=unk_token,
487
+ pad_token=pad_token,
488
+ )
489
+
490
+ # save vocab dict to be loaded into tokenizer
491
+ with open(vocab_file, "w") as file:
492
+ json.dump(vocab_dict, file)
493
+
494
+ # if tokenizer has just been created
495
+ # it is defined by `tokenizer_class` if present in config else by `model_type`
496
+ tokenizer_kwargs = {
497
+ "config": config if config.tokenizer_class is not None else None,
498
+ "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
499
+ "unk_token": unk_token,
500
+ "pad_token": pad_token,
501
+ "eos_token": None,
502
+ "bos_token": None,
503
+ "word_delimiter_token": word_delimiter_token,
504
+ }
505
+
506
+ # 5. Now we can instantiate the feature extractor, tokenizer and model
507
+ # Note for distributed training, the .from_pretrained methods guarantee that only
508
+ # one local process can concurrently download model & vocab.
509
+
510
+ # load feature_extractor and tokenizer
511
+ tokenizer = AutoTokenizer.from_pretrained(
512
+ tokenizer_name_or_path,
513
+ use_auth_token=data_args.use_auth_token,
514
+ **tokenizer_kwargs,
515
+ )
516
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
517
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
518
+ )
519
+
520
+ # adapt config
521
+ config.update(
522
+ {
523
+ "feat_proj_dropout": model_args.feat_proj_dropout,
524
+ "attention_dropout": model_args.attention_dropout,
525
+ "hidden_dropout": model_args.hidden_dropout,
526
+ "final_dropout": model_args.final_dropout,
527
+ "mask_time_prob": model_args.mask_time_prob,
528
+ "mask_time_length": model_args.mask_time_length,
529
+ "mask_feature_prob": model_args.mask_feature_prob,
530
+ "mask_feature_length": model_args.mask_feature_length,
531
+ "gradient_checkpointing": training_args.gradient_checkpointing,
532
+ "layerdrop": model_args.layerdrop,
533
+ "ctc_loss_reduction": model_args.ctc_loss_reduction,
534
+ "pad_token_id": tokenizer.pad_token_id,
535
+ "vocab_size": len(tokenizer),
536
+ "activation_dropout": model_args.activation_dropout,
537
+ }
538
+ )
539
+
540
+ # create model
541
+ model = AutoModelForCTC.from_pretrained(
542
+ model_args.model_name_or_path,
543
+ cache_dir=model_args.cache_dir,
544
+ config=config,
545
+ use_auth_token=data_args.use_auth_token,
546
+ )
547
+
548
+ # freeze encoder
549
+ if model_args.freeze_feature_encoder:
550
+ model.freeze_feature_encoder()
551
+
552
+ # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
553
+ # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
554
+ # so that we just need to set the correct target sampling rate and normalize the input
555
+ # via the `feature_extractor`
556
+
557
+ # make sure that dataset decodes audio with correct sampling rate
558
+ dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
559
+ if dataset_sampling_rate != feature_extractor.sampling_rate:
560
+ raw_datasets = raw_datasets.cast_column(
561
+ data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
562
+ )
563
+
564
+ # derive max & min input length for sample rate & max duration
565
+ max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
566
+ min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
567
+ audio_column_name = data_args.audio_column_name
568
+ num_workers = data_args.preprocessing_num_workers
569
+
570
+ # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
571
+ phoneme_language = data_args.phoneme_language
572
+
573
+ # Preprocessing the datasets.
574
+ # We need to read the audio files as arrays and tokenize the targets.
575
+ def prepare_dataset(batch):
576
+ # load audio
577
+ sample = batch[audio_column_name]
578
+
579
+ inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
580
+ batch["input_values"] = inputs.input_values[0]
581
+ batch["input_length"] = len(batch["input_values"])
582
+
583
+ # encode targets
584
+ additional_kwargs = {}
585
+ if phoneme_language is not None:
586
+ additional_kwargs["phonemizer_lang"] = phoneme_language
587
+
588
+ batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
589
+ return batch
590
+
591
+ with training_args.main_process_first(desc="dataset map preprocessing"):
592
+ vectorized_datasets = raw_datasets.map(
593
+ prepare_dataset,
594
+ remove_columns=next(iter(raw_datasets.values())).column_names,
595
+ num_proc=num_workers,
596
+ desc="preprocess datasets",
597
+ )
598
+
599
+ def is_audio_in_length_range(length):
600
+ return length > min_input_length and length < max_input_length
601
+
602
+ # filter data that is shorter than min_input_length
603
+ vectorized_datasets = vectorized_datasets.filter(
604
+ is_audio_in_length_range,
605
+ num_proc=num_workers,
606
+ input_columns=["input_length"],
607
+ )
608
+
609
+ # 7. Next, we can prepare the training.
610
+ # Let's use word error rate (WER) as our evaluation metric,
611
+ # instantiate a data collator and the trainer
612
+
613
+ # Define evaluation metrics during training, *i.e.* word error rate, character error rate
614
+ eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
615
+
616
+ # for large datasets it is advised to run the preprocessing on a
617
+ # single machine first with ``args.preprocessing_only`` since there will mostly likely
618
+ # be a timeout when running the script in distributed mode.
619
+ # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
620
+ # cached dataset
621
+ if data_args.preprocessing_only:
622
+ logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
623
+ return
624
+
625
+ def compute_metrics(pred):
626
+ pred_logits = pred.predictions
627
+ pred_ids = np.argmax(pred_logits, axis=-1)
628
+
629
+ pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
630
+
631
+ pred_str = tokenizer.batch_decode(pred_ids)
632
+ # we do not want to group tokens when computing the metrics
633
+ label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
634
+
635
+ metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
636
+
637
+ return metrics
638
+
639
+ # Now save everything to be able to create a single processor later
640
+ if is_main_process(training_args.local_rank):
641
+ # save feature extractor, tokenizer and config
642
+ feature_extractor.save_pretrained(training_args.output_dir)
643
+ tokenizer.save_pretrained(training_args.output_dir)
644
+ config.save_pretrained(training_args.output_dir)
645
+
646
+ try:
647
+ processor = AutoProcessor.from_pretrained(training_args.output_dir)
648
+ except (OSError, KeyError):
649
+ warnings.warn(
650
+ "Loading a processor from a feature extractor config that does not"
651
+ " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
652
+ " attribute to your `preprocessor_config.json` file to suppress this warning: "
653
+ " `'processor_class': 'Wav2Vec2Processor'`",
654
+ FutureWarning,
655
+ )
656
+ processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
657
+
658
+ # Instantiate custom data collator
659
+ data_collator = DataCollatorCTCWithPadding(processor=processor)
660
+
661
+ # Initialize Trainer
662
+ trainer = Trainer(
663
+ model=model,
664
+ data_collator=data_collator,
665
+ args=training_args,
666
+ compute_metrics=compute_metrics,
667
+ train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
668
+ eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
669
+ tokenizer=feature_extractor,
670
+ )
671
+
672
+ # 8. Finally, we can start training
673
+
674
+ # Training
675
+ if training_args.do_train:
676
+
677
+ # use last checkpoint if exist
678
+ if last_checkpoint is not None:
679
+ checkpoint = last_checkpoint
680
+ elif os.path.isdir(model_args.model_name_or_path):
681
+ checkpoint = model_args.model_name_or_path
682
+ else:
683
+ checkpoint = None
684
+
685
+ train_result = trainer.train(resume_from_checkpoint=checkpoint)
686
+ trainer.save_model()
687
+
688
+ metrics = train_result.metrics
689
+ max_train_samples = (
690
+ data_args.max_train_samples
691
+ if data_args.max_train_samples is not None
692
+ else len(vectorized_datasets["train"])
693
+ )
694
+ metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
695
+
696
+ trainer.log_metrics("train", metrics)
697
+ trainer.save_metrics("train", metrics)
698
+ trainer.save_state()
699
+
700
+ # Evaluation
701
+ results = {}
702
+ if training_args.do_eval:
703
+ logger.info("*** Evaluate ***")
704
+ metrics = trainer.evaluate()
705
+ max_eval_samples = (
706
+ data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
707
+ )
708
+ metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
709
+
710
+ trainer.log_metrics("eval", metrics)
711
+ trainer.save_metrics("eval", metrics)
712
+
713
+ # Write model card and (optionally) push to hub
714
+ config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
715
+ kwargs = {
716
+ "finetuned_from": model_args.model_name_or_path,
717
+ "tasks": "speech-recognition",
718
+ "tags": ["automatic-speech-recognition", data_args.dataset_name],
719
+ "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
720
+ "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
721
+ }
722
+ if "common_voice" in data_args.dataset_name:
723
+ kwargs["language"] = config_name
724
+
725
+ if training_args.push_to_hub:
726
+ trainer.push_to_hub(**kwargs)
727
+ else:
728
+ trainer.create_model_card(**kwargs)
729
+
730
+ return results
731
+
732
+
733
+ if __name__ == "__main__":
734
+ main()
.ipynb_checkpoints/vocab-checkpoint.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"'": 1, "a": 2, "b": 3, "c": 4, "d": 5, "e": 6, "f": 7, "g": 8, "h": 9, "i": 10, "j": 11, "k": 12, "l": 13, "m": 14, "n": 15, "o": 16, "p": 17, "q": 18, "r": 19, "s": 20, "t": 21, "u": 22, "v": 23, "w": 24, "x": 25, "y": 26, "z": 27, "\u00e0": 28, "\u00e2": 29, "\u00e4": 30, "\u00e7": 31, "\u00e8": 32, "\u00e9": 33, "\u00ea": 34, "\u00eb": 35, "\u00ee": 36, "\u00ef": 37, "\u00f4": 38, "\u00f6": 39, "\u00f9": 40, "\u00fb": 41, "\u00fc": 42, "\u00ff": 43, "|": 0, "[UNK]": 44, "[PAD]": 45}
config.json ADDED
@@ -0,0 +1,107 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "facebook/wav2vec2-xls-r-300m",
3
+ "activation_dropout": 0.1,
4
+ "adapter_kernel_size": 3,
5
+ "adapter_stride": 2,
6
+ "add_adapter": false,
7
+ "apply_spec_augment": true,
8
+ "architectures": [
9
+ "Wav2Vec2ForCTC"
10
+ ],
11
+ "attention_dropout": 0.0,
12
+ "bos_token_id": 1,
13
+ "classifier_proj_size": 256,
14
+ "codevector_dim": 768,
15
+ "contrastive_logits_temperature": 0.1,
16
+ "conv_bias": true,
17
+ "conv_dim": [
18
+ 512,
19
+ 512,
20
+ 512,
21
+ 512,
22
+ 512,
23
+ 512,
24
+ 512
25
+ ],
26
+ "conv_kernel": [
27
+ 10,
28
+ 3,
29
+ 3,
30
+ 3,
31
+ 3,
32
+ 2,
33
+ 2
34
+ ],
35
+ "conv_stride": [
36
+ 5,
37
+ 2,
38
+ 2,
39
+ 2,
40
+ 2,
41
+ 2,
42
+ 2
43
+ ],
44
+ "ctc_loss_reduction": "mean",
45
+ "ctc_zero_infinity": false,
46
+ "diversity_loss_weight": 0.1,
47
+ "do_stable_layer_norm": true,
48
+ "eos_token_id": 2,
49
+ "feat_extract_activation": "gelu",
50
+ "feat_extract_dropout": 0.0,
51
+ "feat_extract_norm": "layer",
52
+ "feat_proj_dropout": 0.0,
53
+ "feat_quantizer_dropout": 0.0,
54
+ "final_dropout": 0.0,
55
+ "hidden_act": "gelu",
56
+ "hidden_dropout": 0.0,
57
+ "hidden_size": 1024,
58
+ "initializer_range": 0.02,
59
+ "intermediate_size": 4096,
60
+ "layer_norm_eps": 1e-05,
61
+ "layerdrop": 0.0,
62
+ "mask_feature_length": 64,
63
+ "mask_feature_min_masks": 0,
64
+ "mask_feature_prob": 0.25,
65
+ "mask_time_length": 10,
66
+ "mask_time_min_masks": 2,
67
+ "mask_time_prob": 0.75,
68
+ "model_type": "wav2vec2",
69
+ "num_adapter_layers": 3,
70
+ "num_attention_heads": 16,
71
+ "num_codevector_groups": 2,
72
+ "num_codevectors_per_group": 320,
73
+ "num_conv_pos_embedding_groups": 16,
74
+ "num_conv_pos_embeddings": 128,
75
+ "num_feat_extract_layers": 7,
76
+ "num_hidden_layers": 24,
77
+ "num_negatives": 100,
78
+ "output_hidden_size": 1024,
79
+ "pad_token_id": 45,
80
+ "proj_codevector_dim": 768,
81
+ "tdnn_dilation": [
82
+ 1,
83
+ 2,
84
+ 3,
85
+ 1,
86
+ 1
87
+ ],
88
+ "tdnn_dim": [
89
+ 512,
90
+ 512,
91
+ 512,
92
+ 512,
93
+ 1500
94
+ ],
95
+ "tdnn_kernel": [
96
+ 5,
97
+ 3,
98
+ 3,
99
+ 1,
100
+ 1
101
+ ],
102
+ "torch_dtype": "float32",
103
+ "transformers_version": "4.17.0.dev0",
104
+ "use_weighted_layer_sum": false,
105
+ "vocab_size": 46,
106
+ "xvector_output_dim": 512
107
+ }
eval.py ADDED
@@ -0,0 +1,128 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ import argparse
3
+ import re
4
+ from typing import Dict
5
+
6
+ from datasets import Audio, Dataset, load_dataset, load_metric
7
+
8
+ from transformers import AutoFeatureExtractor, pipeline
9
+
10
+
11
+ def log_results(result: Dataset, args: Dict[str, str]):
12
+ """DO NOT CHANGE. This function computes and logs the result metrics."""
13
+
14
+ log_outputs = args.log_outputs
15
+ dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
16
+
17
+ # load metric
18
+ wer = load_metric("wer")
19
+ cer = load_metric("cer")
20
+
21
+ # compute metrics
22
+ wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
23
+ cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
24
+
25
+ # print & log results
26
+ result_str = f"WER: {wer_result}\n" f"CER: {cer_result}"
27
+ print(result_str)
28
+
29
+ with open(f"{dataset_id}_eval_results.txt", "w") as f:
30
+ f.write(result_str)
31
+
32
+ # log all results in text file. Possibly interesting for analysis
33
+ if log_outputs is not None:
34
+ pred_file = f"log_{dataset_id}_predictions.txt"
35
+ target_file = f"log_{dataset_id}_targets.txt"
36
+
37
+ with open(pred_file, "w") as p, open(target_file, "w") as t:
38
+
39
+ # mapping function to write output
40
+ def write_to_file(batch, i):
41
+ p.write(f"{i}" + "\n")
42
+ p.write(batch["prediction"] + "\n")
43
+ t.write(f"{i}" + "\n")
44
+ t.write(batch["target"] + "\n")
45
+
46
+ result.map(write_to_file, with_indices=True)
47
+
48
+
49
+ def normalize_text(text: str) -> str:
50
+ """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
51
+
52
+ chars_to_ignore_regex = '[^a-zàâäçéèêëîïôöùûüÿ\'’ ]' # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
53
+
54
+ text = re.sub(chars_to_ignore_regex, "", text.lower()).replace('’', "'")
55
+
56
+ # In addition, we can normalize the target text, e.g. removing new lines characters etc...
57
+ # note that order is important here!
58
+ token_sequences_to_ignore = ["\n\n", "\n", " ", " "]
59
+
60
+ for t in token_sequences_to_ignore:
61
+ text = " ".join(text.split(t))
62
+
63
+ return text
64
+
65
+
66
+ def main(args):
67
+ # load dataset
68
+ dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
69
+
70
+ # for testing: only process the first two examples as a test
71
+ # dataset = dataset.select(range(10))
72
+
73
+ # load processor
74
+ feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
75
+ sampling_rate = feature_extractor.sampling_rate
76
+
77
+ # resample audio
78
+ dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
79
+
80
+ # load eval pipeline
81
+ asr = pipeline("automatic-speech-recognition", model=args.model_id)
82
+
83
+ # map function to decode audio
84
+ def map_to_pred(batch):
85
+ prediction = asr(
86
+ batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
87
+ )
88
+
89
+ batch["prediction"] = prediction["text"]
90
+ batch["target"] = normalize_text(batch["sentence"])
91
+ return batch
92
+
93
+ # run inference on all examples
94
+ result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
95
+
96
+ # compute and log_results
97
+ # do not change function below
98
+ log_results(result, args)
99
+
100
+
101
+ if __name__ == "__main__":
102
+ parser = argparse.ArgumentParser()
103
+
104
+ parser.add_argument(
105
+ "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
106
+ )
107
+ parser.add_argument(
108
+ "--dataset",
109
+ type=str,
110
+ required=True,
111
+ help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
112
+ )
113
+ parser.add_argument(
114
+ "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'` for Common Voice"
115
+ )
116
+ parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
117
+ parser.add_argument(
118
+ "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
119
+ )
120
+ parser.add_argument(
121
+ "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
122
+ )
123
+ parser.add_argument(
124
+ "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
125
+ )
126
+ args = parser.parse_args()
127
+
128
+ main(args)
preprocessor_config.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "do_normalize": true,
3
+ "feature_extractor_type": "Wav2Vec2FeatureExtractor",
4
+ "feature_size": 1,
5
+ "padding_side": "right",
6
+ "padding_value": 0,
7
+ "return_attention_mask": true,
8
+ "sampling_rate": 16000
9
+ }
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:fe30ae12c53d8f58e684006bd1c1927acfc8a0a8ffea0b10a29486e0564b7ed8
3
+ size 1262112241
run.sh ADDED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ export WANDB_PROJECT="xls-r-300m-cv_8-fr"
2
+ python run_speech_recognition_ctc.py \
3
+ --activation_dropout="0.1" \
4
+ --dataset_name="mozilla-foundation/common_voice_8_0" \
5
+ --dataset_config_name="fr" \
6
+ --eval_steps="1000" \
7
+ --evaluation_strategy="steps" \
8
+ --feat_proj_dropout="0.0" \
9
+ --freeze_feature_encoder \
10
+ --fp16 \
11
+ --gradient_accumulation_steps="8" \
12
+ --gradient_checkpointing \
13
+ --group_by_length \
14
+ --layerdrop="0.0" \
15
+ --learning_rate="7.5e-5" \
16
+ --length_column_name="input_length" \
17
+ --load_best_model_at_end \
18
+ --logging_steps="100" \
19
+ --mask_feature_length="64" \
20
+ --mask_feature_prob="0.25" \
21
+ --mask_time_length="10" \
22
+ --mask_time_prob="0.75" \
23
+ --model_name_or_path="facebook/wav2vec2-xls-r-300m" \
24
+ --num_train_epochs="5.0" \
25
+ --output_dir="./" \
26
+ --overwrite_output_dir \
27
+ --per_device_train_batch_size="16" \
28
+ --per_device_eval_batch_size="16" \
29
+ --preprocessing_num_workers="4" \
30
+ --push_to_hub \
31
+ --report_to="wandb" \
32
+ --save_steps="1000" \
33
+ --save_total_limit="3" \
34
+ --text_column_name="sentence" \
35
+ --use_auth_token \
36
+ --warmup_steps="2000" \
37
+ --do_train --do_eval
run_speech_recognition_ctc.py ADDED
@@ -0,0 +1,734 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 'test'"
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 = '[^a-zàâäçéèêëîïôöùûüÿ\'’ ]'
438
+ text_column_name = data_args.text_column_name
439
+
440
+ def remove_and_replace_special_characters(batch):
441
+ batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name].lower()).replace('’', "'") + " "
442
+ return batch
443
+
444
+ with training_args.main_process_first(desc="dataset map special characters removal"):
445
+ raw_datasets = raw_datasets.map(
446
+ remove_and_replace_special_characters,
447
+ remove_columns=[text_column_name],
448
+ desc="remove special characters from datasets",
449
+ )
450
+
451
+ # save special tokens for tokenizer
452
+ word_delimiter_token = data_args.word_delimiter_token
453
+ unk_token = data_args.unk_token
454
+ pad_token = data_args.pad_token
455
+
456
+ # 3. Next, let's load the config as we might need it to create
457
+ # the tokenizer
458
+ # load config
459
+ config = AutoConfig.from_pretrained(
460
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
461
+ )
462
+
463
+ # 4. Next, if no tokenizer file is defined,
464
+ # we create the vocabulary of the model by extracting all unique characters from
465
+ # the training and evaluation datasets
466
+ # We need to make sure that only first rank saves vocabulary
467
+ # make sure all processes wait until vocab is created
468
+ tokenizer_name_or_path = model_args.tokenizer_name_or_path
469
+ tokenizer_kwargs = {}
470
+ if tokenizer_name_or_path is None:
471
+ # save vocab in training output dir
472
+ tokenizer_name_or_path = training_args.output_dir
473
+
474
+ vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
475
+
476
+ with training_args.main_process_first():
477
+ if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
478
+ os.remove(vocab_file)
479
+
480
+ with training_args.main_process_first(desc="dataset map vocabulary creation"):
481
+ if not os.path.isfile(vocab_file):
482
+ os.makedirs(tokenizer_name_or_path, exist_ok=True)
483
+ vocab_dict = create_vocabulary_from_data(
484
+ raw_datasets,
485
+ word_delimiter_token=word_delimiter_token,
486
+ unk_token=unk_token,
487
+ pad_token=pad_token,
488
+ )
489
+
490
+ # save vocab dict to be loaded into tokenizer
491
+ with open(vocab_file, "w") as file:
492
+ json.dump(vocab_dict, file)
493
+
494
+ # if tokenizer has just been created
495
+ # it is defined by `tokenizer_class` if present in config else by `model_type`
496
+ tokenizer_kwargs = {
497
+ "config": config if config.tokenizer_class is not None else None,
498
+ "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
499
+ "unk_token": unk_token,
500
+ "pad_token": pad_token,
501
+ "eos_token": None,
502
+ "bos_token": None,
503
+ "word_delimiter_token": word_delimiter_token,
504
+ }
505
+
506
+ # 5. Now we can instantiate the feature extractor, tokenizer and model
507
+ # Note for distributed training, the .from_pretrained methods guarantee that only
508
+ # one local process can concurrently download model & vocab.
509
+
510
+ # load feature_extractor and tokenizer
511
+ tokenizer = AutoTokenizer.from_pretrained(
512
+ tokenizer_name_or_path,
513
+ use_auth_token=data_args.use_auth_token,
514
+ **tokenizer_kwargs,
515
+ )
516
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
517
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
518
+ )
519
+
520
+ # adapt config
521
+ config.update(
522
+ {
523
+ "feat_proj_dropout": model_args.feat_proj_dropout,
524
+ "attention_dropout": model_args.attention_dropout,
525
+ "hidden_dropout": model_args.hidden_dropout,
526
+ "final_dropout": model_args.final_dropout,
527
+ "mask_time_prob": model_args.mask_time_prob,
528
+ "mask_time_length": model_args.mask_time_length,
529
+ "mask_feature_prob": model_args.mask_feature_prob,
530
+ "mask_feature_length": model_args.mask_feature_length,
531
+ "gradient_checkpointing": training_args.gradient_checkpointing,
532
+ "layerdrop": model_args.layerdrop,
533
+ "ctc_loss_reduction": model_args.ctc_loss_reduction,
534
+ "pad_token_id": tokenizer.pad_token_id,
535
+ "vocab_size": len(tokenizer),
536
+ "activation_dropout": model_args.activation_dropout,
537
+ }
538
+ )
539
+
540
+ # create model
541
+ model = AutoModelForCTC.from_pretrained(
542
+ model_args.model_name_or_path,
543
+ cache_dir=model_args.cache_dir,
544
+ config=config,
545
+ use_auth_token=data_args.use_auth_token,
546
+ )
547
+
548
+ # freeze encoder
549
+ if model_args.freeze_feature_encoder:
550
+ model.freeze_feature_encoder()
551
+
552
+ # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
553
+ # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
554
+ # so that we just need to set the correct target sampling rate and normalize the input
555
+ # via the `feature_extractor`
556
+
557
+ # make sure that dataset decodes audio with correct sampling rate
558
+ dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
559
+ if dataset_sampling_rate != feature_extractor.sampling_rate:
560
+ raw_datasets = raw_datasets.cast_column(
561
+ data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
562
+ )
563
+
564
+ # derive max & min input length for sample rate & max duration
565
+ max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
566
+ min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
567
+ audio_column_name = data_args.audio_column_name
568
+ num_workers = data_args.preprocessing_num_workers
569
+
570
+ # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
571
+ phoneme_language = data_args.phoneme_language
572
+
573
+ # Preprocessing the datasets.
574
+ # We need to read the audio files as arrays and tokenize the targets.
575
+ def prepare_dataset(batch):
576
+ # load audio
577
+ sample = batch[audio_column_name]
578
+
579
+ inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
580
+ batch["input_values"] = inputs.input_values[0]
581
+ batch["input_length"] = len(batch["input_values"])
582
+
583
+ # encode targets
584
+ additional_kwargs = {}
585
+ if phoneme_language is not None:
586
+ additional_kwargs["phonemizer_lang"] = phoneme_language
587
+
588
+ batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
589
+ return batch
590
+
591
+ with training_args.main_process_first(desc="dataset map preprocessing"):
592
+ vectorized_datasets = raw_datasets.map(
593
+ prepare_dataset,
594
+ remove_columns=next(iter(raw_datasets.values())).column_names,
595
+ num_proc=num_workers,
596
+ desc="preprocess datasets",
597
+ )
598
+
599
+ def is_audio_in_length_range(length):
600
+ return length > min_input_length and length < max_input_length
601
+
602
+ # filter data that is shorter than min_input_length
603
+ vectorized_datasets = vectorized_datasets.filter(
604
+ is_audio_in_length_range,
605
+ num_proc=num_workers,
606
+ input_columns=["input_length"],
607
+ )
608
+
609
+ # 7. Next, we can prepare the training.
610
+ # Let's use word error rate (WER) as our evaluation metric,
611
+ # instantiate a data collator and the trainer
612
+
613
+ # Define evaluation metrics during training, *i.e.* word error rate, character error rate
614
+ eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
615
+
616
+ # for large datasets it is advised to run the preprocessing on a
617
+ # single machine first with ``args.preprocessing_only`` since there will mostly likely
618
+ # be a timeout when running the script in distributed mode.
619
+ # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
620
+ # cached dataset
621
+ if data_args.preprocessing_only:
622
+ logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
623
+ return
624
+
625
+ def compute_metrics(pred):
626
+ pred_logits = pred.predictions
627
+ pred_ids = np.argmax(pred_logits, axis=-1)
628
+
629
+ pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
630
+
631
+ pred_str = tokenizer.batch_decode(pred_ids)
632
+ # we do not want to group tokens when computing the metrics
633
+ label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
634
+
635
+ metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
636
+
637
+ return metrics
638
+
639
+ # Now save everything to be able to create a single processor later
640
+ if is_main_process(training_args.local_rank):
641
+ # save feature extractor, tokenizer and config
642
+ feature_extractor.save_pretrained(training_args.output_dir)
643
+ tokenizer.save_pretrained(training_args.output_dir)
644
+ config.save_pretrained(training_args.output_dir)
645
+
646
+ try:
647
+ processor = AutoProcessor.from_pretrained(training_args.output_dir)
648
+ except (OSError, KeyError):
649
+ warnings.warn(
650
+ "Loading a processor from a feature extractor config that does not"
651
+ " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
652
+ " attribute to your `preprocessor_config.json` file to suppress this warning: "
653
+ " `'processor_class': 'Wav2Vec2Processor'`",
654
+ FutureWarning,
655
+ )
656
+ processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
657
+
658
+ # Instantiate custom data collator
659
+ data_collator = DataCollatorCTCWithPadding(processor=processor)
660
+
661
+ # Initialize Trainer
662
+ trainer = Trainer(
663
+ model=model,
664
+ data_collator=data_collator,
665
+ args=training_args,
666
+ compute_metrics=compute_metrics,
667
+ train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
668
+ eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
669
+ tokenizer=feature_extractor,
670
+ )
671
+
672
+ # 8. Finally, we can start training
673
+
674
+ # Training
675
+ if training_args.do_train:
676
+
677
+ # use last checkpoint if exist
678
+ if last_checkpoint is not None:
679
+ checkpoint = last_checkpoint
680
+ elif os.path.isdir(model_args.model_name_or_path):
681
+ checkpoint = model_args.model_name_or_path
682
+ else:
683
+ checkpoint = None
684
+
685
+ train_result = trainer.train(resume_from_checkpoint=checkpoint)
686
+ trainer.save_model()
687
+
688
+ metrics = train_result.metrics
689
+ max_train_samples = (
690
+ data_args.max_train_samples
691
+ if data_args.max_train_samples is not None
692
+ else len(vectorized_datasets["train"])
693
+ )
694
+ metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
695
+
696
+ trainer.log_metrics("train", metrics)
697
+ trainer.save_metrics("train", metrics)
698
+ trainer.save_state()
699
+
700
+ # Evaluation
701
+ results = {}
702
+ if training_args.do_eval:
703
+ logger.info("*** Evaluate ***")
704
+ metrics = trainer.evaluate()
705
+ max_eval_samples = (
706
+ data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
707
+ )
708
+ metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
709
+
710
+ trainer.log_metrics("eval", metrics)
711
+ trainer.save_metrics("eval", metrics)
712
+
713
+ # Write model card and (optionally) push to hub
714
+ config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
715
+ kwargs = {
716
+ "finetuned_from": model_args.model_name_or_path,
717
+ "tasks": "speech-recognition",
718
+ "tags": ["automatic-speech-recognition", data_args.dataset_name],
719
+ "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
720
+ "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
721
+ }
722
+ if "common_voice" in data_args.dataset_name:
723
+ kwargs["language"] = config_name
724
+
725
+ if training_args.push_to_hub:
726
+ trainer.push_to_hub(**kwargs)
727
+ else:
728
+ trainer.create_model_card(**kwargs)
729
+
730
+ return results
731
+
732
+
733
+ if __name__ == "__main__":
734
+ main()
special_tokens_map.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"unk_token": "[UNK]", "pad_token": "[PAD]"}
tokenizer_config.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"unk_token": "[UNK]", "bos_token": null, "eos_token": null, "pad_token": "[PAD]", "do_lower_case": false, "word_delimiter_token": "|", "special_tokens_map_file": null, "tokenizer_file": null, "name_or_path": "./", "tokenizer_class": "Wav2Vec2CTCTokenizer"}
training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:954b51e3ecb6c307d1a94d85367240ee2bf34ca857be239a02052c3fa3818e96
3
+ size 2991
vocab.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"'": 1, "a": 2, "b": 3, "c": 4, "d": 5, "e": 6, "f": 7, "g": 8, "h": 9, "i": 10, "j": 11, "k": 12, "l": 13, "m": 14, "n": 15, "o": 16, "p": 17, "q": 18, "r": 19, "s": 20, "t": 21, "u": 22, "v": 23, "w": 24, "x": 25, "y": 26, "z": 27, "à": 28, "â": 29, "ä": 30, "ç": 31, "è": 32, "é": 33, "ê": 34, "ë": 35, "î": 36, "ï": 37, "ô": 38, "ö": 39, "ù": 40, "û": 41, "ü": 42, "ÿ": 43, "|": 0, "[UNK]": 44, "[PAD]": 45}
wandb/debug-internal.log ADDED
@@ -0,0 +1 @@
 
 
1
+ run-20220203_135844-2tzexn1o/logs/debug-internal.log
wandb/debug.log ADDED
@@ -0,0 +1 @@
 
 
1
+ run-20220203_135844-2tzexn1o/logs/debug.log
wandb/latest-run ADDED
@@ -0,0 +1 @@
 
 
1
+ run-20220203_135844-2tzexn1o
wandb/run-20220203_135844-2tzexn1o/files/conda-environment.yaml ADDED
File without changes
wandb/run-20220203_135844-2tzexn1o/files/config.yaml ADDED
The diff for this file is too large to render. See raw diff
 
wandb/run-20220203_135844-2tzexn1o/files/output.log ADDED
@@ -0,0 +1,1361 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+
3
+
4
+
5
+
6
+
7
+
8
+
9
+
10
+
11
+
12
+
13
+
14
+
15
+
16
+
17
+
18
+
19
+
20
+
21
+
22
+
23
+
24
+
25
+
26
+
27
+
28
+
29
+
30
+
31
+
32
+
33
+
34
+
35
+
36
+
37
+
38
+
39
+
40
+
41
+
42
+
43
+
44
+
45
+
46
+
47
+
48
+
49
+
50
+
51
+
52
+
53
+
54
+
55
+
56
+
57
+
58
+
59
+
60
+
61
+
62
+
63
+
64
+
65
+
66
+
67
+
68
+
69
+
70
+
71
+
72
+
73
+
74
+
75
+
76
+
77
+
78
+
79
+
80
+
81
+
82
+
83
+
84
+
85
+
86
+
87
+
88
+
89
+
90
+
91
+
92
+
93
+
94
+
95
+
96
+
97
+
98
+
99
+
100
+
101
+
102
+ 1%|▊ | 100/17440 [12:17<19:56:25, 4.14s/it]
103
+
104
+
105
+
106
+
107
+
108
+
109
+
110
+
111
+
112
+
113
+
114
+
115
+
116
+
117
+
118
+
119
+
120
+
121
+
122
+
123
+
124
+
125
+
126
+
127
+
128
+
129
+
130
+
131
+
132
+
133
+
134
+
135
+
136
+
137
+
138
+
139
+
140
+
141
+
142
+
143
+
144
+
145
+
146
+
147
+
148
+
149
+
150
+
151
+
152
+
153
+
154
+
155
+
156
+
157
+
158
+
159
+
160
+
161
+
162
+
163
+
164
+
165
+
166
+
167
+
168
+
169
+
170
+
171
+
172
+
173
+
174
+
175
+
176
+
177
+
178
+
179
+
180
+
181
+
182
+
183
+
184
+
185
+
186
+
187
+
188
+
189
+
190
+
191
+
192
+
193
+
194
+
195
+
196
+
197
+
198
+
199
+
200
+
201
+
202
+ 1%|█▋ | 199/17440 [24:34<20:52:21, 4.36s/it]
203
+
204
+
205
+
206
+
207
+
208
+
209
+
210
+
211
+
212
+
213
+
214
+
215
+
216
+
217
+
218
+
219
+
220
+
221
+
222
+
223
+
224
+
225
+
226
+
227
+
228
+
229
+
230
+
231
+
232
+
233
+
234
+
235
+
236
+
237
+
238
+
239
+
240
+
241
+
242
+
243
+
244
+
245
+
246
+
247
+
248
+
249
+
250
+
251
+
252
+
253
+
254
+
255
+
256
+
257
+
258
+
259
+
260
+
261
+
262
+
263
+
264
+
265
+
266
+
267
+
268
+
269
+
270
+
271
+
272
+
273
+
274
+
275
+
276
+
277
+
278
+
279
+
280
+
281
+
282
+
283
+
284
+
285
+
286
+
287
+
288
+
289
+
290
+
291
+
292
+
293
+
294
+
295
+
296
+
297
+
298
+
299
+
300
+
301
+
302
+
303
+
304
+ 2%|██▌ | 300/17440 [37:07<20:04:05, 4.22s/it]
305
+
306
+
307
+
308
+
309
+
310
+
311
+
312
+
313
+
314
+
315
+
316
+
317
+
318
+
319
+
320
+
321
+
322
+
323
+
324
+
325
+
326
+
327
+
328
+
329
+
330
+
331
+
332
+
333
+
334
+
335
+
336
+
337
+
338
+
339
+
340
+
341
+
342
+
343
+
344
+
345
+
346
+
347
+
348
+
349
+
350
+
351
+
352
+
353
+
354
+
355
+
356
+
357
+
358
+
359
+
360
+
361
+
362
+
363
+
364
+
365
+
366
+
367
+
368
+
369
+
370
+
371
+
372
+
373
+
374
+
375
+
376
+
377
+
378
+
379
+
380
+
381
+
382
+
383
+
384
+
385
+
386
+
387
+
388
+
389
+
390
+
391
+
392
+
393
+
394
+
395
+
396
+
397
+
398
+
399
+
400
+
401
+
402
+
403
+
404
+ 2%|███▍ | 399/17440 [49:31<21:14:28, 4.49s/it]
405
+
406
+
407
+
408
+
409
+
410
+
411
+
412
+
413
+
414
+
415
+
416
+
417
+
418
+
419
+
420
+
421
+
422
+
423
+
424
+
425
+
426
+
427
+
428
+
429
+
430
+
431
+
432
+
433
+
434
+
435
+
436
+
437
+
438
+
439
+
440
+
441
+
442
+
443
+
444
+
445
+
446
+
447
+
448
+
449
+
450
+
451
+
452
+
453
+
454
+
455
+
456
+
457
+
458
+
459
+
460
+
461
+
462
+
463
+
464
+
465
+
466
+
467
+
468
+
469
+
470
+
471
+
472
+
473
+
474
+
475
+
476
+
477
+
478
+
479
+
480
+
481
+
482
+
483
+
484
+
485
+
486
+
487
+
488
+
489
+
490
+
491
+
492
+
493
+
494
+
495
+
496
+
497
+
498
+
499
+
500
+
501
+
502
+
503
+
504
+
505
+ 3%|████▏ | 499/17440 [1:01:50<20:45:51, 4.41s/it]
506
+
507
+
508
+
509
+
510
+
511
+
512
+
513
+
514
+
515
+
516
+
517
+
518
+
519
+
520
+
521
+
522
+
523
+
524
+
525
+
526
+
527
+
528
+
529
+
530
+
531
+
532
+
533
+
534
+
535
+
536
+
537
+
538
+
539
+
540
+
541
+
542
+
543
+
544
+
545
+
546
+
547
+
548
+
549
+
550
+
551
+
552
+
553
+
554
+
555
+
556
+
557
+
558
+
559
+
560
+
561
+
562
+
563
+
564
+
565
+
566
+
567
+
568
+
569
+
570
+
571
+
572
+
573
+
574
+
575
+
576
+
577
+
578
+
579
+
580
+
581
+
582
+
583
+
584
+
585
+
586
+
587
+
588
+
589
+
590
+
591
+
592
+
593
+
594
+
595
+
596
+
597
+
598
+
599
+
600
+
601
+
602
+
603
+
604
+
605
+
606
+
607
+ 3%|█████ | 600/17440 [1:14:19<19:27:18, 4.16s/it]
608
+
609
+
610
+
611
+
612
+
613
+
614
+
615
+
616
+
617
+
618
+
619
+
620
+
621
+
622
+
623
+
624
+
625
+
626
+
627
+
628
+
629
+
630
+
631
+
632
+
633
+
634
+
635
+
636
+
637
+
638
+
639
+
640
+
641
+
642
+
643
+
644
+
645
+
646
+
647
+
648
+
649
+
650
+
651
+
652
+
653
+
654
+
655
+
656
+
657
+
658
+
659
+
660
+
661
+
662
+
663
+
664
+
665
+
666
+
667
+
668
+
669
+
670
+
671
+
672
+
673
+
674
+
675
+
676
+
677
+
678
+
679
+
680
+
681
+
682
+
683
+
684
+
685
+
686
+
687
+
688
+
689
+
690
+
691
+
692
+
693
+
694
+
695
+
696
+
697
+
698
+
699
+
700
+
701
+
702
+
703
+
704
+
705
+
706
+
707
+
708
+ 4%|█████▉ | 700/17440 [1:27:12<19:52:48, 4.28s/it]
709
+
710
+
711
+
712
+
713
+
714
+
715
+
716
+
717
+
718
+
719
+
720
+
721
+
722
+
723
+
724
+
725
+
726
+
727
+
728
+
729
+
730
+
731
+
732
+
733
+
734
+
735
+
736
+
737
+
738
+
739
+
740
+
741
+
742
+
743
+
744
+
745
+
746
+
747
+
748
+
749
+
750
+
751
+
752
+
753
+
754
+
755
+
756
+
757
+
758
+
759
+
760
+
761
+
762
+
763
+
764
+
765
+
766
+
767
+
768
+
769
+
770
+
771
+
772
+
773
+
774
+
775
+
776
+
777
+
778
+
779
+
780
+
781
+
782
+
783
+
784
+
785
+
786
+
787
+
788
+
789
+
790
+
791
+
792
+
793
+
794
+
795
+
796
+
797
+
798
+
799
+
800
+
801
+
802
+
803
+
804
+
805
+
806
+
807
+
808
+ 5%|██████▊ | 799/17440 [1:39:39<20:49:39, 4.51s/it]
809
+
810
+
811
+
812
+
813
+
814
+
815
+
816
+
817
+
818
+
819
+
820
+
821
+
822
+
823
+
824
+
825
+
826
+
827
+
828
+
829
+
830
+
831
+
832
+
833
+
834
+
835
+
836
+
837
+
838
+
839
+
840
+
841
+
842
+
843
+
844
+
845
+
846
+
847
+
848
+
849
+
850
+
851
+
852
+
853
+
854
+
855
+
856
+
857
+
858
+
859
+
860
+
861
+
862
+
863
+
864
+
865
+
866
+
867
+
868
+
869
+
870
+
871
+
872
+
873
+
874
+
875
+
876
+
877
+
878
+
879
+
880
+
881
+
882
+
883
+
884
+
885
+
886
+
887
+
888
+
889
+
890
+
891
+
892
+
893
+
894
+
895
+
896
+
897
+
898
+
899
+
900
+
901
+
902
+
903
+
904
+
905
+
906
+
907
+
908
+
909
+ 5%|███████▋ | 899/17440 [1:52:23<21:09:25, 4.60s/it]
910
+
911
+
912
+
913
+
914
+
915
+
916
+
917
+
918
+
919
+
920
+
921
+
922
+
923
+
924
+
925
+
926
+
927
+
928
+
929
+
930
+
931
+
932
+
933
+
934
+
935
+
936
+
937
+
938
+
939
+
940
+
941
+
942
+
943
+
944
+
945
+
946
+
947
+
948
+
949
+
950
+
951
+
952
+
953
+
954
+
955
+
956
+
957
+
958
+
959
+
960
+
961
+
962
+
963
+
964
+
965
+
966
+
967
+
968
+
969
+
970
+
971
+
972
+
973
+
974
+
975
+
976
+
977
+
978
+
979
+
980
+
981
+
982
+
983
+
984
+
985
+
986
+
987
+
988
+
989
+
990
+
991
+
992
+
993
+
994
+
995
+
996
+
997
+
998
+
999
+
1000
+
1001
+
1002
+
1003
+
1004
+
1005
+
1006
+
1007
+
1008
+
1009
+
1010
+ 6%|████████▍ | 999/17440 [2:05:08<20:53:06, 4.57s/it]
1011
+ 6%|████████▍ | 1000/17440 [2:05:11<19:26:40, 4.26s/it]The following columns in the evaluation set don't have a corresponding argument in `Wav2Vec2ForCTC.forward` and have been ignored: input_length. If input_length are not expected by `Wav2Vec2ForCTC.forward`, you can safely ignore this message.
1012
+ ***** Running Evaluation *****
1013
+ Num examples = 16021
1014
+ Batch size = 16
1015
+
1016
+
1017
+
1018
+
1019
+
1020
+
1021
+
1022
+
1023
+
1024
+
1025
+
1026
+
1027
+
1028
+
1029
+
1030
+
1031
+
1032
+
1033
+
1034
+
1035
+
1036
+
1037
+
1038
+
1039
+
1040
+
1041
+
1042
+
1043
+
1044
+
1045
+
1046
+
1047
+
1048
+
1049
+
1050
+
1051
+
1052
+
1053
+
1054
+
1055
+
1056
+
1057
+
1058
+
1059
+
1060
+
1061
+
1062
+
1063
+
1064
+
1065
+
1066
+
1067
+
1068
+
1069
+
1070
+
1071
+
1072
+
1073
+
1074
+
1075
+
1076
+
1077
+
1078
+
1079
+
1080
+
1081
+
1082
+
1083
+
1084
+
1085
+
1086
+
1087
+
1088
+
1089
+
1090
+
1091
+
1092
+
1093
+
1094
+
1095
+
1096
+
1097
+
1098
+
1099
+
1100
+
1101
+
1102
+
1103
+
1104
+
1105
+
1106
+
1107
+
1108
+
1109
+
1110
+
1111
+
1112
+
1113
+
1114
+
1115
+
1116
+
1117
+
1118
+
1119
+
1120
+
1121
+
1122
+
1123
+
1124
+
1125
+
1126
+
1127
+
1128
+
1129
+
1130
+
1131
+
1132
+
1133
+
1134
+
1135
+
1136
+
1137
+
1138
+
1139
+
1140
+
1141
+
1142
+
1143
+
1144
+
1145
+
1146
+
1147
+
1148
+
1149
+
1150
+
1151
+
1152
+
1153
+
1154
+
1155
+
1156
+
1157
+
1158
+
1159
+
1160
+
1161
+
1162
+
1163
+
1164
+
1165
+
1166
+
1167
+
1168
+
1169
+
1170
+
1171
+
1172
+
1173
+
1174
+
1175
+
1176
+
1177
+
1178
+
1179
+
1180
+
1181
+
1182
+
1183
+
1184
+
1185
+
1186
+
1187
+
1188
+
1189
+
1190
+
1191
+
1192
+
1193
+
1194
+
1195
+
1196
+
1197
+
1198
+
1199
+
1200
+
1201
+
1202
+
1203
+
1204
+
1205
+
1206
+
1207
+
1208
+
1209
+
1210
+
1211
+
1212
+
1213
+
1214
+
1215
+
1216
+
1217
+
1218
+
1219
+
1220
+
1221
+
1222
+
1223
+
1224
+
1225
+
1226
+
1227
+
1228
+
1229
+
1230
+
1231
+
1232
+
1233
+
1234
+
1235
+
1236
+
1237
+
1238
+
1239
+
1240
+
1241
+
1242
+
1243
+
1244
+
1245
+
1246
+
1247
+
1248
+
1249
+
1250
+
1251
+
1252
+
1253
+
1254
+
1255
+
1256
+
1257
+
1258
+
1259
+
1260
+
1261
+
1262
+
1263
+
1264
+
1265
+
1266
+
1267
+
1268
+
1269
+
1270
+
1271
+
1272
+
1273
+
1274
+
1275
+
1276
+
1277
+
1278
+
1279
+
1280
+
1281
+
1282
+
1283
+
1284
+
1285
+
1286
+
1287
+
1288
+
1289
+
1290
+
1291
+
1292
+
1293
+
1294
+
1295
+
1296
+
1297
+
1298
+
1299
+
1300
+
1301
+
1302
+
1303
+
1304
+
1305
+
1306
+
1307
+
1308
+
1309
+
1310
+
1311
+
1312
+
1313
+
1314
+
1315
+
1316
+
1317
+
1318
+
1319
+
1320
+
1321
+
1322
+
1323
+
1324
+
1325
+
1326
+
1327
+
1328
+
1329
+
1330
+
1331
+
1332
+
1333
+
1334
+
1335
+
1336
+
1337
+
1338
+
1339
+
1340
+
1341
+
1342
+
1343
+
1344
+
1345
+
1346
+
1347
+
1348
+
1349
+
1350
+
1351
+
1352
+
1353
+
1354
+
1355
+
1356
+
1357
+
1358
+ Configuration saved in ./checkpoint-1000/config.json
1359
+ {'eval_loss': inf, 'eval_wer': 0.9997048122028068, 'eval_runtime': 711.5482, 'eval_samples_per_second': 22.516, 'eval_steps_per_second': 1.408, 'epoch': 0.29}
1360
+ Model weights saved in ./checkpoint-1000/pytorch_model.bin
1361
+ Configuration saved in ./checkpoint-1000/preprocessor_config.json
wandb/run-20220203_135844-2tzexn1o/files/requirements.txt ADDED
@@ -0,0 +1,178 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ aiohttp==3.8.1
2
+ aiosignal==1.2.0
3
+ analytics-python==1.4.0
4
+ anyio==3.5.0
5
+ appdirs==1.4.4
6
+ argon2-cffi-bindings==21.2.0
7
+ argon2-cffi==21.3.0
8
+ asgiref==3.5.0
9
+ asttokens==2.0.5
10
+ async-timeout==4.0.2
11
+ attrs==21.4.0
12
+ audioread==2.1.9
13
+ backcall==0.2.0
14
+ backoff==1.10.0
15
+ bcrypt==3.2.0
16
+ beautifulsoup4==4.9.3
17
+ black==22.1.0
18
+ bleach==4.1.0
19
+ brotlipy==0.7.0
20
+ certifi==2020.12.5
21
+ cffi==1.14.3
22
+ chardet==3.0.4
23
+ charset-normalizer==2.0.11
24
+ click==8.0.3
25
+ conda-build==3.21.4
26
+ conda-package-handling==1.7.2
27
+ conda==4.9.2
28
+ cryptography==3.2.1
29
+ cycler==0.11.0
30
+ datasets==1.18.3.dev0
31
+ debugpy==1.5.1
32
+ decorator==4.4.2
33
+ defusedxml==0.7.1
34
+ dill==0.3.4
35
+ dnspython==2.1.0
36
+ docker-pycreds==0.4.0
37
+ entrypoints==0.3
38
+ executing==0.8.2
39
+ fastapi==0.73.0
40
+ ffmpy==0.3.0
41
+ filelock==3.0.12
42
+ fonttools==4.29.1
43
+ frozenlist==1.3.0
44
+ fsspec==2022.1.0
45
+ gitdb==4.0.9
46
+ gitpython==3.1.26
47
+ glob2==0.7
48
+ gradio==2.7.5.2
49
+ h11==0.13.0
50
+ huggingface-hub==0.4.0
51
+ idna==2.10
52
+ importlib-resources==5.4.0
53
+ ipykernel==6.7.0
54
+ ipython-genutils==0.2.0
55
+ ipython==8.0.1
56
+ ipywidgets==7.6.3
57
+ jedi==0.17.0
58
+ jinja2==2.11.3
59
+ jiwer==2.3.0
60
+ joblib==1.1.0
61
+ json5==0.9.6
62
+ jsonschema==4.4.0
63
+ jupyter-client==7.1.2
64
+ jupyter-core==4.9.1
65
+ jupyterlab-pygments==0.1.2
66
+ jupyterlab-server==1.2.0
67
+ jupyterlab-widgets==1.0.2
68
+ jupyterlab==2.2.9
69
+ kiwisolver==1.3.2
70
+ libarchive-c==2.9
71
+ librosa==0.8.1
72
+ llvmlite==0.38.0
73
+ markdown2==2.4.2
74
+ markupsafe==1.1.1
75
+ matplotlib-inline==0.1.3
76
+ matplotlib==3.5.1
77
+ mistune==0.8.4
78
+ mkl-fft==1.3.0
79
+ mkl-random==1.1.1
80
+ mkl-service==2.3.0
81
+ monotonic==1.6
82
+ multidict==6.0.2
83
+ multiprocess==0.70.12.2
84
+ mypy-extensions==0.4.3
85
+ nano==0.10.0
86
+ nbclient==0.5.10
87
+ nbconvert==6.4.1
88
+ nbformat==5.1.3
89
+ nest-asyncio==1.5.4
90
+ notebook==6.4.8
91
+ numba==0.55.1
92
+ numpy==1.19.2
93
+ olefile==0.46
94
+ packaging==21.3
95
+ pandas==1.4.0
96
+ pandocfilters==1.5.0
97
+ paramiko==2.9.2
98
+ parso==0.8.1
99
+ pathspec==0.9.0
100
+ pathtools==0.1.2
101
+ pexpect==4.8.0
102
+ pickleshare==0.7.5
103
+ pillow==8.1.2
104
+ pip==22.0.2
105
+ pkginfo==1.7.0
106
+ platformdirs==2.4.1
107
+ pooch==1.6.0
108
+ prometheus-client==0.13.1
109
+ promise==2.3
110
+ prompt-toolkit==3.0.8
111
+ protobuf==3.19.4
112
+ psutil==5.8.0
113
+ ptyprocess==0.7.0
114
+ pure-eval==0.2.2
115
+ pyarrow==6.0.1
116
+ pycosat==0.6.3
117
+ pycparser==2.20
118
+ pycryptodome==3.14.0
119
+ pydantic==1.9.0
120
+ pydub==0.25.1
121
+ pygments==2.8.0
122
+ pynacl==1.5.0
123
+ pyopenssl==19.1.0
124
+ pyparsing==3.0.7
125
+ pyrsistent==0.18.1
126
+ pysocks==1.7.1
127
+ python-dateutil==2.8.2
128
+ python-etcd==0.4.5
129
+ python-levenshtein==0.12.2
130
+ python-multipart==0.0.5
131
+ pytz==2021.1
132
+ pyyaml==5.4.1
133
+ pyzmq==22.3.0
134
+ regex==2022.1.18
135
+ requests==2.24.0
136
+ resampy==0.2.2
137
+ ruamel-yaml==0.15.87
138
+ sacremoses==0.0.47
139
+ scikit-learn==1.0.2
140
+ scipy==1.7.3
141
+ send2trash==1.8.0
142
+ sentry-sdk==1.5.4
143
+ setuptools==50.3.1.post20201107
144
+ shortuuid==1.0.8
145
+ six==1.15.0
146
+ smmap==5.0.0
147
+ sniffio==1.2.0
148
+ soundfile==0.10.3.post1
149
+ soupsieve==2.2
150
+ stack-data==0.1.4
151
+ starlette==0.17.1
152
+ termcolor==1.1.0
153
+ terminado==0.13.1
154
+ testpath==0.5.0
155
+ threadpoolctl==3.1.0
156
+ tokenizers==0.11.4
157
+ tomli==2.0.0
158
+ torch==1.10.2
159
+ torchaudio==0.10.2
160
+ torchelastic==0.2.2
161
+ torchtext==0.9.1
162
+ torchvision==0.9.1
163
+ tornado==6.1
164
+ tqdm==4.62.3
165
+ traitlets==5.1.1
166
+ transformers==4.17.0.dev0
167
+ typing-extensions==4.0.1
168
+ urllib3==1.25.11
169
+ uvicorn==0.17.1
170
+ wandb==0.12.10
171
+ wcwidth==0.2.5
172
+ webencodings==0.5.1
173
+ wheel==0.35.1
174
+ widgetsnbextension==3.5.2
175
+ xxhash==2.0.2
176
+ yarl==1.7.2
177
+ yaspin==2.1.0
178
+ zipp==3.7.0
wandb/run-20220203_135844-2tzexn1o/files/wandb-metadata.json ADDED
@@ -0,0 +1,61 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "os": "Linux-4.15.0-151-generic-x86_64-with-glibc2.10",
3
+ "python": "3.8.8",
4
+ "heartbeatAt": "2022-02-03T13:58:46.127751",
5
+ "startedAt": "2022-02-03T13:58:44.294948",
6
+ "docker": null,
7
+ "gpu": "Tesla V100S-PCIE-32GB",
8
+ "gpu_count": 1,
9
+ "cpu_count": 60,
10
+ "cuda": null,
11
+ "args": [
12
+ "--activation_dropout=0.1",
13
+ "--dataset_name=mozilla-foundation/common_voice_8_0",
14
+ "--dataset_config_name=fr",
15
+ "--eval_steps=1000",
16
+ "--evaluation_strategy=steps",
17
+ "--feat_proj_dropout=0.0",
18
+ "--freeze_feature_encoder",
19
+ "--fp16",
20
+ "--gradient_accumulation_steps=8",
21
+ "--gradient_checkpointing",
22
+ "--group_by_length",
23
+ "--layerdrop=0.0",
24
+ "--learning_rate=7.5e-5",
25
+ "--length_column_name=input_length",
26
+ "--load_best_model_at_end",
27
+ "--logging_steps=100",
28
+ "--mask_feature_length=64",
29
+ "--mask_feature_prob=0.25",
30
+ "--mask_time_length=10",
31
+ "--mask_time_prob=0.75",
32
+ "--model_name_or_path=facebook/wav2vec2-xls-r-300m",
33
+ "--num_train_epochs=5.0",
34
+ "--output_dir=./",
35
+ "--overwrite_output_dir",
36
+ "--per_device_train_batch_size=16",
37
+ "--per_device_eval_batch_size=16",
38
+ "--preprocessing_num_workers=4",
39
+ "--push_to_hub",
40
+ "--report_to=wandb",
41
+ "--save_steps=1000",
42
+ "--save_total_limit=3",
43
+ "--text_column_name=sentence",
44
+ "--use_auth_token",
45
+ "--warmup_steps=2000",
46
+ "--do_train",
47
+ "--do_eval"
48
+ ],
49
+ "state": "running",
50
+ "program": "run_speech_recognition_ctc.py",
51
+ "codePath": "run_speech_recognition_ctc.py",
52
+ "git": {
53
+ "remote": "https://huggingface.co/Plim/xls-r-300m-cv_8-fr",
54
+ "commit": "6539a9ab1c39bdf4a299206d2a4d17ad09039c09"
55
+ },
56
+ "email": "lim.pascal93@gmail.com",
57
+ "root": "/workspace/xls-r-300m-cv_8-fr",
58
+ "host": "job-8ff6a64a-ab40-4ef5-b3ec-d0e2f05ceb9f",
59
+ "username": "ovh",
60
+ "executable": "/opt/conda/bin/python"
61
+ }
wandb/run-20220203_135844-2tzexn1o/files/wandb-summary.json ADDED
The diff for this file is too large to render. See raw diff
 
wandb/run-20220203_135844-2tzexn1o/logs/debug-internal.log ADDED
The diff for this file is too large to render. See raw diff
 
wandb/run-20220203_135844-2tzexn1o/logs/debug.log ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 2022-02-03 13:58:44,300 INFO MainThread:42729 [wandb_setup.py:_flush():75] Loading settings from /workspace/.config/wandb/settings
2
+ 2022-02-03 13:58:44,300 INFO MainThread:42729 [wandb_setup.py:_flush():75] Loading settings from /workspace/xls-r-300m-cv_8-fr/wandb/settings
3
+ 2022-02-03 13:58:44,300 INFO MainThread:42729 [wandb_setup.py:_flush():75] Loading settings from environment variables: {'project': 'xls-r-300m-cv_8-fr'}
4
+ 2022-02-03 13:58:44,300 INFO MainThread:42729 [wandb_setup.py:_flush():75] Inferring run settings from compute environment: {'program_relpath': 'run_speech_recognition_ctc.py', 'program': 'run_speech_recognition_ctc.py'}
5
+ 2022-02-03 13:58:44,301 INFO MainThread:42729 [wandb_init.py:_log_setup():386] Logging user logs to /workspace/xls-r-300m-cv_8-fr/wandb/run-20220203_135844-2tzexn1o/logs/debug.log
6
+ 2022-02-03 13:58:44,301 INFO MainThread:42729 [wandb_init.py:_log_setup():387] Logging internal logs to /workspace/xls-r-300m-cv_8-fr/wandb/run-20220203_135844-2tzexn1o/logs/debug-internal.log
7
+ 2022-02-03 13:58:44,301 INFO MainThread:42729 [wandb_init.py:init():420] calling init triggers
8
+ 2022-02-03 13:58:44,301 INFO MainThread:42729 [wandb_init.py:init():425] wandb.init called with sweep_config: {}
9
+ config: {}
10
+ 2022-02-03 13:58:44,301 INFO MainThread:42729 [wandb_init.py:init():471] starting backend
11
+ 2022-02-03 13:58:44,301 INFO MainThread:42729 [backend.py:_multiprocessing_setup():99] multiprocessing start_methods=fork,spawn,forkserver, using: spawn
12
+ 2022-02-03 13:58:44,539 INFO MainThread:42729 [backend.py:ensure_launched():219] starting backend process...
13
+ 2022-02-03 13:58:44,779 INFO MainThread:42729 [backend.py:ensure_launched():224] started backend process with pid: 3742
14
+ 2022-02-03 13:58:44,782 INFO MainThread:42729 [wandb_init.py:init():480] backend started and connected
15
+ 2022-02-03 13:58:44,791 INFO MainThread:42729 [wandb_init.py:init():550] updated telemetry
16
+ 2022-02-03 13:58:45,331 INFO MainThread:42729 [wandb_init.py:init():581] communicating current version
17
+ 2022-02-03 13:58:45,892 INFO MainThread:42729 [wandb_init.py:init():586] got version response
18
+ 2022-02-03 13:58:45,892 INFO MainThread:42729 [wandb_init.py:init():596] communicating run to backend with 30 second timeout
19
+ 2022-02-03 13:58:46,119 INFO MainThread:42729 [wandb_init.py:init():624] starting run threads in backend
20
+ 2022-02-03 13:58:46,700 INFO MainThread:42729 [wandb_run.py:_console_start():1827] atexit reg
21
+ 2022-02-03 13:58:46,701 INFO MainThread:42729 [wandb_run.py:_redirect():1701] redirect: SettingsConsole.REDIRECT
22
+ 2022-02-03 13:58:46,702 INFO MainThread:42729 [wandb_run.py:_redirect():1706] Redirecting console.
23
+ 2022-02-03 13:58:46,709 INFO MainThread:42729 [wandb_run.py:_redirect():1762] Redirects installed.
24
+ 2022-02-03 13:58:46,709 INFO MainThread:42729 [wandb_init.py:init():651] run started, returning control to user process
25
+ 2022-02-03 13:58:46,712 INFO MainThread:42729 [wandb_run.py:_config_callback():966] config_cb None None {'return_dict': True, 'output_hidden_states': False, 'output_attentions': False, 'torchscript': False, 'torch_dtype': 'float32', 'use_bfloat16': False, 'pruned_heads': {}, 'tie_word_embeddings': True, 'is_encoder_decoder': False, 'is_decoder': False, 'cross_attention_hidden_size': None, 'add_cross_attention': False, 'tie_encoder_decoder': False, 'max_length': 20, 'min_length': 0, 'do_sample': False, 'early_stopping': False, 'num_beams': 1, 'num_beam_groups': 1, 'diversity_penalty': 0.0, 'temperature': 1.0, 'top_k': 50, 'top_p': 1.0, 'repetition_penalty': 1.0, 'length_penalty': 1.0, 'no_repeat_ngram_size': 0, 'encoder_no_repeat_ngram_size': 0, 'bad_words_ids': None, 'num_return_sequences': 1, 'chunk_size_feed_forward': 0, 'output_scores': False, 'return_dict_in_generate': False, 'forced_bos_token_id': None, 'forced_eos_token_id': None, 'remove_invalid_values': False, 'architectures': ['Wav2Vec2ForPreTraining'], 'finetuning_task': None, 'id2label': {0: 'LABEL_0', 1: 'LABEL_1'}, 'label2id': {'LABEL_0': 0, 'LABEL_1': 1}, 'tokenizer_class': None, 'prefix': None, 'bos_token_id': 1, 'pad_token_id': 45, 'eos_token_id': 2, 'sep_token_id': None, 'decoder_start_token_id': None, 'task_specific_params': None, 'problem_type': None, '_name_or_path': 'facebook/wav2vec2-xls-r-300m', 'transformers_version': '4.17.0.dev0', 'feat_extract_dropout': 0.0, 'model_type': 'wav2vec2', 'num_feat_extract_layers': 7, 'hidden_size': 1024, 'feat_extract_norm': 'layer', 'feat_extract_activation': 'gelu', 'conv_dim': [512, 512, 512, 512, 512, 512, 512], 'conv_stride': [5, 2, 2, 2, 2, 2, 2], 'conv_kernel': [10, 3, 3, 3, 3, 2, 2], 'conv_bias': True, 'num_conv_pos_embeddings': 128, 'num_conv_pos_embedding_groups': 16, 'num_hidden_layers': 24, 'intermediate_size': 4096, 'hidden_act': 'gelu', 'num_attention_heads': 16, 'hidden_dropout': 0.0, 'attention_dropout': 0.0, 'activation_dropout': 0.1, 'feat_proj_dropout': 0.0, 'final_dropout': 0.0, 'layerdrop': 0.0, 'layer_norm_eps': 1e-05, 'initializer_range': 0.02, 'vocab_size': 46, 'do_stable_layer_norm': True, 'use_weighted_layer_sum': False, 'apply_spec_augment': True, 'mask_time_prob': 0.75, 'mask_time_length': 10, 'mask_time_min_masks': 2, 'mask_feature_prob': 0.25, 'mask_feature_length': 64, 'mask_feature_min_masks': 0, 'num_codevectors_per_group': 320, 'num_codevector_groups': 2, 'contrastive_logits_temperature': 0.1, 'feat_quantizer_dropout': 0.0, 'num_negatives': 100, 'codevector_dim': 768, 'proj_codevector_dim': 768, 'diversity_loss_weight': 0.1, 'ctc_loss_reduction': 'mean', 'ctc_zero_infinity': False, 'add_adapter': False, 'adapter_kernel_size': 3, 'adapter_stride': 2, 'num_adapter_layers': 3, 'output_hidden_size': 1024, 'classifier_proj_size': 256, 'tdnn_dim': [512, 512, 512, 512, 1500], 'tdnn_kernel': [5, 3, 3, 1, 1], 'tdnn_dilation': [1, 2, 3, 1, 1], 'xvector_output_dim': 512, 'output_dir': './', 'overwrite_output_dir': True, 'do_train': True, 'do_eval': True, 'do_predict': False, 'evaluation_strategy': 'steps', 'prediction_loss_only': False, 'per_device_train_batch_size': 16, 'per_device_eval_batch_size': 16, 'per_gpu_train_batch_size': 'None', 'per_gpu_eval_batch_size': 'None', 'gradient_accumulation_steps': 8, 'eval_accumulation_steps': 'None', 'learning_rate': 7.5e-05, 'weight_decay': 0.0, 'adam_beta1': 0.9, 'adam_beta2': 0.999, 'adam_epsilon': 1e-08, 'max_grad_norm': 1.0, 'num_train_epochs': 5.0, 'max_steps': -1, 'lr_scheduler_type': 'linear', 'warmup_ratio': 0.0, 'warmup_steps': 2000, 'log_level': -1, 'log_level_replica': -1, 'log_on_each_node': True, 'logging_dir': './runs/Feb03_10-26-21_job-8ff6a64a-ab40-4ef5-b3ec-d0e2f05ceb9f', 'logging_strategy': 'steps', 'logging_first_step': False, 'logging_steps': 100, 'logging_nan_inf_filter': True, 'save_strategy': 'steps', 'save_steps': 1000, 'save_total_limit': 3, 'save_on_each_node': False, 'no_cuda': False, 'seed': 42, 'bf16': False, 'fp16': True, 'fp16_opt_level': 'O1', 'half_precision_backend': 'amp', 'bf16_full_eval': False, 'fp16_full_eval': False, 'tf32': 'None', 'local_rank': -1, 'xpu_backend': 'None', 'tpu_num_cores': 'None', 'tpu_metrics_debug': False, 'debug': '[]', 'dataloader_drop_last': False, 'eval_steps': 1000, 'dataloader_num_workers': 0, 'past_index': -1, 'run_name': './', 'disable_tqdm': False, 'remove_unused_columns': True, 'label_names': 'None', 'load_best_model_at_end': True, 'metric_for_best_model': 'loss', 'greater_is_better': False, 'ignore_data_skip': False, 'sharded_ddp': '[]', 'deepspeed': 'None', 'label_smoothing_factor': 0.0, 'optim': 'adamw_hf', 'adafactor': False, 'group_by_length': True, 'length_column_name': 'input_length', 'report_to': "['wandb']", 'ddp_find_unused_parameters': 'None', 'ddp_bucket_cap_mb': 'None', 'dataloader_pin_memory': True, 'skip_memory_metrics': True, 'use_legacy_prediction_loop': False, 'push_to_hub': True, 'resume_from_checkpoint': 'None', 'hub_model_id': 'None', 'hub_strategy': 'every_save', 'hub_token': '<HUB_TOKEN>', 'gradient_checkpointing': True, 'fp16_backend': 'auto', 'push_to_hub_model_id': 'None', 'push_to_hub_organization': 'None', 'push_to_hub_token': '<PUSH_TO_HUB_TOKEN>', '_n_gpu': 1, 'mp_parameters': '', 'train_batch_size': 16, 'eval_batch_size': 16}
26
+ 2022-02-03 13:58:46,715 INFO MainThread:42729 [wandb_watch.py:watch():43] Watching
wandb/run-20220203_135844-2tzexn1o/run-2tzexn1o.wandb ADDED
Binary file (7.31 MB). View file