ihanif commited on
Commit
37ac01e
1 Parent(s): 9df476e

Training in progress, step 500

Browse files
.gitignore ADDED
@@ -0,0 +1 @@
 
 
1
+ checkpoint-*/
added_tokens.json ADDED
@@ -0,0 +1,4 @@
 
 
 
 
 
1
+ {
2
+ "</s>": 116,
3
+ "<s>": 115
4
+ }
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": 0.1,
65
+ "mask_time_length": 10,
66
+ "mask_time_min_masks": 2,
67
+ "mask_time_prob": 0.3,
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": 114,
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": "4.26.0.dev0",
104
+ "use_weighted_layer_sum": false,
105
+ "vocab_size": 117,
106
+ "xvector_output_dim": 512
107
+ }
preprocessor_config.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "do_normalize": true,
3
+ "feature_extractor_type": "Wav2Vec2FeatureExtractor",
4
+ "feature_size": 1,
5
+ "padding_side": "right",
6
+ "padding_value": 0,
7
+ "return_attention_mask": true,
8
+ "sampling_rate": 16000
9
+ }
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d47a5c0818d8f28fd464809d5257860c023102f7f43d32404fb6237358f7c36d
3
+ size 1262381549
run.sh ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ python run_speech_recognition_ctc.py \
2
+ --dataset_name="google/fleurs" \
3
+ --dataset_config_name="ps_af" \
4
+ --model_name_or_path="facebook/wav2vec2-xls-r-300m" \
5
+ --output_dir="./" \
6
+ --overwrite_output_dir \
7
+ --num_train_epochs="50" \
8
+ --per_device_train_batch_size="8" \
9
+ --per_device_eval_batch_size="8" \
10
+ --gradient_accumulation_steps="4" \
11
+ --learning_rate="7.5e-5" \
12
+ --warmup_steps="2000" \
13
+ --evaluation_strategy="steps" \
14
+ --text_column_name="transcription" \
15
+ --save_steps="500" \
16
+ --eval_steps="500" \
17
+ --logging_steps="10" \
18
+ --layerdrop="0.0" \
19
+ --activation_dropout="0.1" \
20
+ --eval_metrics wer cer \
21
+ --save_total_limit="1" \
22
+ --mask_time_prob="0.3" \
23
+ --mask_time_length="10" \
24
+ --mask_feature_prob="0.1" \
25
+ --fp16 \
26
+ --mask_feature_length="64" \
27
+ --chars_to_ignore , ? . ! - \; \: \" “ % ‘ ” � \
28
+ --group_by_length \
29
+ --push_to_hub \
30
+ --do_train --do_eval \
31
+ --gradient_checkpointing \
32
+ --use_auth_token
33
+ --freeze_feature_extractor="True"
run_speech_recognition_ctc.py ADDED
@@ -0,0 +1,772 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
32
+
33
+ import evaluate
34
+ import transformers
35
+ from transformers import (
36
+ AutoConfig,
37
+ AutoFeatureExtractor,
38
+ AutoModelForCTC,
39
+ AutoProcessor,
40
+ AutoTokenizer,
41
+ HfArgumentParser,
42
+ Trainer,
43
+ TrainingArguments,
44
+ Wav2Vec2Processor,
45
+ set_seed,
46
+ )
47
+ from transformers.trainer_utils import get_last_checkpoint, is_main_process
48
+ from transformers.utils import check_min_version, send_example_telemetry
49
+ from transformers.utils.versions import require_version
50
+
51
+
52
+ # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
53
+ check_min_version("4.26.0.dev0")
54
+
55
+ require_version("datasets>=1.18.0", "To fix: pip install -r examples/pytorch/speech-recognition/requirements.txt")
56
+
57
+
58
+ logger = logging.getLogger(__name__)
59
+
60
+
61
+ def list_field(default=None, metadata=None):
62
+ return field(default_factory=lambda: default, metadata=metadata)
63
+
64
+
65
+ @dataclass
66
+ class ModelArguments:
67
+ """
68
+ Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
69
+ """
70
+
71
+ model_name_or_path: str = field(
72
+ metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
73
+ )
74
+ tokenizer_name_or_path: Optional[str] = field(
75
+ default=None,
76
+ metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
77
+ )
78
+ cache_dir: Optional[str] = field(
79
+ default=None,
80
+ metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
81
+ )
82
+ freeze_feature_encoder: bool = field(
83
+ default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
84
+ )
85
+ attention_dropout: float = field(
86
+ default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
87
+ )
88
+ activation_dropout: float = field(
89
+ default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
90
+ )
91
+ feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
92
+ hidden_dropout: float = field(
93
+ default=0.0,
94
+ metadata={
95
+ "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
96
+ },
97
+ )
98
+ final_dropout: float = field(
99
+ default=0.0,
100
+ metadata={"help": "The dropout probability for the final projection layer."},
101
+ )
102
+ mask_time_prob: float = field(
103
+ default=0.05,
104
+ metadata={
105
+ "help": (
106
+ "Probability of each feature vector along the time axis to be chosen as the start of the vector"
107
+ "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
108
+ "vectors will be masked along the time axis."
109
+ )
110
+ },
111
+ )
112
+ mask_time_length: int = field(
113
+ default=10,
114
+ metadata={"help": "Length of vector span to mask along the time axis."},
115
+ )
116
+ mask_feature_prob: float = field(
117
+ default=0.0,
118
+ metadata={
119
+ "help": (
120
+ "Probability of each feature vector along the feature axis to be chosen as the start of the vectorspan"
121
+ " to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature"
122
+ " bins will be masked along the time axis."
123
+ )
124
+ },
125
+ )
126
+ mask_feature_length: int = field(
127
+ default=10,
128
+ metadata={"help": "Length of vector span to mask along the feature axis."},
129
+ )
130
+ layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
131
+ ctc_loss_reduction: Optional[str] = field(
132
+ default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
133
+ )
134
+
135
+
136
+ @dataclass
137
+ class DataTrainingArguments:
138
+ """
139
+ Arguments pertaining to what data we are going to input our model for training and eval.
140
+
141
+ Using `HfArgumentParser` we can turn this class
142
+ into argparse arguments to be able to specify them on
143
+ the command line.
144
+ """
145
+
146
+ dataset_name: str = field(
147
+ metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
148
+ )
149
+ dataset_config_name: str = field(
150
+ default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
151
+ )
152
+ train_split_name: str = field(
153
+ default="train+validation",
154
+ metadata={
155
+ "help": (
156
+ "The name of the training data set split to use (via the datasets library). Defaults to "
157
+ "'train+validation'"
158
+ )
159
+ },
160
+ )
161
+ eval_split_name: str = field(
162
+ default="test",
163
+ metadata={
164
+ "help": "The name of the evaluation data set split to use (via the datasets library). Defaults to 'test'"
165
+ },
166
+ )
167
+ audio_column_name: str = field(
168
+ default="audio",
169
+ metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
170
+ )
171
+ text_column_name: str = field(
172
+ default="text",
173
+ metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
174
+ )
175
+ overwrite_cache: bool = field(
176
+ default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
177
+ )
178
+ preprocessing_num_workers: Optional[int] = field(
179
+ default=None,
180
+ metadata={"help": "The number of processes to use for the preprocessing."},
181
+ )
182
+ max_train_samples: Optional[int] = field(
183
+ default=None,
184
+ metadata={
185
+ "help": (
186
+ "For debugging purposes or quicker training, truncate the number of training examples to this "
187
+ "value if set."
188
+ )
189
+ },
190
+ )
191
+ max_eval_samples: Optional[int] = field(
192
+ default=None,
193
+ metadata={
194
+ "help": (
195
+ "For debugging purposes or quicker training, truncate the number of validation examples to this "
196
+ "value if set."
197
+ )
198
+ },
199
+ )
200
+ chars_to_ignore: Optional[List[str]] = list_field(
201
+ default=None,
202
+ metadata={"help": "A list of characters to remove from the transcripts."},
203
+ )
204
+ eval_metrics: List[str] = list_field(
205
+ default=["wer"],
206
+ metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
207
+ )
208
+ max_duration_in_seconds: float = field(
209
+ default=20.0,
210
+ metadata={
211
+ "help": (
212
+ "Filter audio files that are longer than `max_duration_in_seconds` seconds to"
213
+ " 'max_duration_in_seconds`"
214
+ )
215
+ },
216
+ )
217
+ min_duration_in_seconds: float = field(
218
+ default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
219
+ )
220
+ preprocessing_only: bool = field(
221
+ default=False,
222
+ metadata={
223
+ "help": (
224
+ "Whether to only do data preprocessing and skip training. This is especially useful when data"
225
+ " preprocessing errors out in distributed training due to timeout. In this case, one should run the"
226
+ " preprocessing in a non-distributed setup with `preprocessing_only=True` so that the cached datasets"
227
+ " can consequently be loaded in distributed training"
228
+ )
229
+ },
230
+ )
231
+ use_auth_token: bool = field(
232
+ default=False,
233
+ metadata={
234
+ "help": (
235
+ "If :obj:`True`, will use the token generated when running"
236
+ ":obj:`huggingface-cli login` as HTTP bearer authorization for remote files."
237
+ )
238
+ },
239
+ )
240
+ unk_token: str = field(
241
+ default="[UNK]",
242
+ metadata={"help": "The unk token for the tokenizer"},
243
+ )
244
+ pad_token: str = field(
245
+ default="[PAD]",
246
+ metadata={"help": "The padding token for the tokenizer"},
247
+ )
248
+ word_delimiter_token: str = field(
249
+ default="|",
250
+ metadata={"help": "The word delimiter token for the tokenizer"},
251
+ )
252
+ phoneme_language: Optional[str] = field(
253
+ default=None,
254
+ metadata={
255
+ "help": (
256
+ "The target language that should be used be"
257
+ " passed to the tokenizer for tokenization. Note that"
258
+ " this is only relevant if the model classifies the"
259
+ " input audio to a sequence of phoneme sequences."
260
+ )
261
+ },
262
+ )
263
+
264
+
265
+ @dataclass
266
+ class DataCollatorCTCWithPadding:
267
+ """
268
+ Data collator that will dynamically pad the inputs received.
269
+ Args:
270
+ processor (:class:`~transformers.AutoProcessor`)
271
+ The processor used for proccessing the data.
272
+ padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
273
+ Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
274
+ among:
275
+ * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
276
+ sequence if provided).
277
+ * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
278
+ maximum acceptable input length for the model if that argument is not provided.
279
+ * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
280
+ different lengths).
281
+ max_length (:obj:`int`, `optional`):
282
+ Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
283
+ max_length_labels (:obj:`int`, `optional`):
284
+ Maximum length of the ``labels`` returned list and optionally padding length (see above).
285
+ pad_to_multiple_of (:obj:`int`, `optional`):
286
+ If set will pad the sequence to a multiple of the provided value.
287
+ This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
288
+ 7.5 (Volta).
289
+ """
290
+
291
+ processor: AutoProcessor
292
+ padding: Union[bool, str] = "longest"
293
+ pad_to_multiple_of: Optional[int] = None
294
+ pad_to_multiple_of_labels: Optional[int] = None
295
+
296
+ def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
297
+ # split inputs and labels since they have to be of different lenghts and need
298
+ # different padding methods
299
+ input_features = [{"input_values": feature["input_values"]} for feature in features]
300
+ label_features = [{"input_ids": feature["labels"]} for feature in features]
301
+
302
+ batch = self.processor.pad(
303
+ input_features,
304
+ padding=self.padding,
305
+ pad_to_multiple_of=self.pad_to_multiple_of,
306
+ return_tensors="pt",
307
+ )
308
+
309
+ labels_batch = self.processor.pad(
310
+ labels=label_features,
311
+ padding=self.padding,
312
+ pad_to_multiple_of=self.pad_to_multiple_of_labels,
313
+ return_tensors="pt",
314
+ )
315
+
316
+ # replace padding with -100 to ignore loss correctly
317
+ labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
318
+
319
+ batch["labels"] = labels
320
+ if "attention_mask" in batch:
321
+ batch["attention_mask"] = batch["attention_mask"].to(torch.long)
322
+
323
+ return batch
324
+
325
+
326
+ def create_vocabulary_from_data(
327
+ datasets: DatasetDict,
328
+ word_delimiter_token: Optional[str] = None,
329
+ unk_token: Optional[str] = None,
330
+ pad_token: Optional[str] = None,
331
+ ):
332
+ # Given training and test labels create vocabulary
333
+ def extract_all_chars(batch):
334
+ all_text = " ".join(batch["target_text"])
335
+ vocab = list(set(all_text))
336
+ return {"vocab": [vocab], "all_text": [all_text]}
337
+
338
+ vocabs = datasets.map(
339
+ extract_all_chars,
340
+ batched=True,
341
+ batch_size=-1,
342
+ keep_in_memory=True,
343
+ remove_columns=datasets["train"].column_names,
344
+ )
345
+
346
+ # take union of all unique characters in each dataset
347
+ vocab_set = functools.reduce(
348
+ lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
349
+ )
350
+
351
+ vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
352
+
353
+ # replace white space with delimiter token
354
+ if word_delimiter_token is not None:
355
+ vocab_dict[word_delimiter_token] = vocab_dict[" "]
356
+ del vocab_dict[" "]
357
+
358
+ # add unk and pad token
359
+ if unk_token is not None:
360
+ vocab_dict[unk_token] = len(vocab_dict)
361
+
362
+ if pad_token is not None:
363
+ vocab_dict[pad_token] = len(vocab_dict)
364
+
365
+ return vocab_dict
366
+
367
+
368
+ def main():
369
+ # See all possible arguments in src/transformers/training_args.py
370
+ # or by passing the --help flag to this script.
371
+ # We now keep distinct sets of args, for a cleaner separation of concerns.
372
+
373
+ parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
374
+ if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
375
+ # If we pass only one argument to the script and it's the path to a json file,
376
+ # let's parse it to get our arguments.
377
+ model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
378
+ else:
379
+ model_args, data_args, training_args = parser.parse_args_into_dataclasses()
380
+
381
+ # Sending telemetry. Tracking the example usage helps us better allocate resources to maintain them. The
382
+ # information sent is the one passed as arguments along with your Python/PyTorch versions.
383
+ send_example_telemetry("run_speech_recognition_ctc", model_args, data_args)
384
+
385
+ # Detecting last checkpoint.
386
+ last_checkpoint = None
387
+ if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
388
+ last_checkpoint = get_last_checkpoint(training_args.output_dir)
389
+ if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
390
+ raise ValueError(
391
+ f"Output directory ({training_args.output_dir}) already exists and is not empty. "
392
+ "Use --overwrite_output_dir to overcome."
393
+ )
394
+ elif last_checkpoint is not None:
395
+ logger.info(
396
+ f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
397
+ "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
398
+ )
399
+
400
+ # Setup logging
401
+ logging.basicConfig(
402
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
403
+ datefmt="%m/%d/%Y %H:%M:%S",
404
+ handlers=[logging.StreamHandler(sys.stdout)],
405
+ )
406
+ logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
407
+
408
+ # Log on each process the small summary:
409
+ logger.warning(
410
+ f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
411
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
412
+ )
413
+ # Set the verbosity to info of the Transformers logger (on main process only):
414
+ if is_main_process(training_args.local_rank):
415
+ transformers.utils.logging.set_verbosity_info()
416
+ logger.info("Training/evaluation parameters %s", training_args)
417
+
418
+ # Set seed before initializing model.
419
+ set_seed(training_args.seed)
420
+
421
+ # 1. First, let's load the dataset
422
+ raw_datasets = DatasetDict()
423
+
424
+ if training_args.do_train:
425
+ raw_datasets["train"] = load_dataset(
426
+ data_args.dataset_name,
427
+ data_args.dataset_config_name,
428
+ split=data_args.train_split_name,
429
+ use_auth_token=data_args.use_auth_token,
430
+ )
431
+
432
+ if data_args.audio_column_name not in raw_datasets["train"].column_names:
433
+ raise ValueError(
434
+ f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'."
435
+ " Make sure to set `--audio_column_name` to the correct audio column - one of"
436
+ f" {', '.join(raw_datasets['train'].column_names)}."
437
+ )
438
+
439
+ if data_args.text_column_name not in raw_datasets["train"].column_names:
440
+ raise ValueError(
441
+ f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
442
+ "Make sure to set `--text_column_name` to the correct text column - one of "
443
+ f"{', '.join(raw_datasets['train'].column_names)}."
444
+ )
445
+
446
+ if data_args.max_train_samples is not None:
447
+ raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
448
+
449
+ if training_args.do_eval:
450
+ raw_datasets["eval"] = load_dataset(
451
+ data_args.dataset_name,
452
+ data_args.dataset_config_name,
453
+ split=data_args.eval_split_name,
454
+ use_auth_token=data_args.use_auth_token,
455
+ )
456
+
457
+ if data_args.max_eval_samples is not None:
458
+ raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
459
+
460
+ # 2. We remove some special characters from the datasets
461
+ # that make training complicated and do not help in transcribing the speech
462
+ # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
463
+ # that could be easily picked up by the model
464
+ chars_to_ignore_regex = (
465
+ f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
466
+ )
467
+ text_column_name = data_args.text_column_name
468
+
469
+ def remove_special_characters(batch):
470
+ if chars_to_ignore_regex is not None:
471
+ batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
472
+ else:
473
+ batch["target_text"] = batch[text_column_name].lower() + " "
474
+ return batch
475
+
476
+ with training_args.main_process_first(desc="dataset map special characters removal"):
477
+ raw_datasets = raw_datasets.map(
478
+ remove_special_characters,
479
+ remove_columns=[text_column_name],
480
+ desc="remove special characters from datasets",
481
+ )
482
+
483
+ # save special tokens for tokenizer
484
+ word_delimiter_token = data_args.word_delimiter_token
485
+ unk_token = data_args.unk_token
486
+ pad_token = data_args.pad_token
487
+
488
+ # 3. Next, let's load the config as we might need it to create
489
+ # the tokenizer
490
+ # load config
491
+ config = AutoConfig.from_pretrained(
492
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
493
+ )
494
+
495
+ # 4. Next, if no tokenizer file is defined,
496
+ # we create the vocabulary of the model by extracting all unique characters from
497
+ # the training and evaluation datasets
498
+ # We need to make sure that only first rank saves vocabulary
499
+ # make sure all processes wait until vocab is created
500
+ tokenizer_name_or_path = model_args.tokenizer_name_or_path
501
+ tokenizer_kwargs = {}
502
+ if tokenizer_name_or_path is None:
503
+ # save vocab in training output dir
504
+ tokenizer_name_or_path = training_args.output_dir
505
+
506
+ vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
507
+
508
+ with training_args.main_process_first():
509
+ if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
510
+ try:
511
+ os.remove(vocab_file)
512
+ except OSError:
513
+ # in shared file-systems it might be the case that
514
+ # two processes try to delete the vocab file at the some time
515
+ pass
516
+
517
+ with training_args.main_process_first(desc="dataset map vocabulary creation"):
518
+ if not os.path.isfile(vocab_file):
519
+ os.makedirs(tokenizer_name_or_path, exist_ok=True)
520
+ vocab_dict = create_vocabulary_from_data(
521
+ raw_datasets,
522
+ word_delimiter_token=word_delimiter_token,
523
+ unk_token=unk_token,
524
+ pad_token=pad_token,
525
+ )
526
+
527
+ # save vocab dict to be loaded into tokenizer
528
+ with open(vocab_file, "w") as file:
529
+ json.dump(vocab_dict, file)
530
+
531
+ # if tokenizer has just been created
532
+ # it is defined by `tokenizer_class` if present in config else by `model_type`
533
+ tokenizer_kwargs = {
534
+ "config": config if config.tokenizer_class is not None else None,
535
+ "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
536
+ "unk_token": unk_token,
537
+ "pad_token": pad_token,
538
+ "word_delimiter_token": word_delimiter_token,
539
+ }
540
+
541
+ # 5. Now we can instantiate the feature extractor, tokenizer and model
542
+ # Note for distributed training, the .from_pretrained methods guarantee that only
543
+ # one local process can concurrently download model & vocab.
544
+
545
+ # load feature_extractor and tokenizer
546
+ tokenizer = AutoTokenizer.from_pretrained(
547
+ tokenizer_name_or_path,
548
+ use_auth_token=data_args.use_auth_token,
549
+ **tokenizer_kwargs,
550
+ )
551
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
552
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
553
+ )
554
+
555
+ # adapt config
556
+ config.update(
557
+ {
558
+ "feat_proj_dropout": model_args.feat_proj_dropout,
559
+ "attention_dropout": model_args.attention_dropout,
560
+ "hidden_dropout": model_args.hidden_dropout,
561
+ "final_dropout": model_args.final_dropout,
562
+ "mask_time_prob": model_args.mask_time_prob,
563
+ "mask_time_length": model_args.mask_time_length,
564
+ "mask_feature_prob": model_args.mask_feature_prob,
565
+ "mask_feature_length": model_args.mask_feature_length,
566
+ "gradient_checkpointing": training_args.gradient_checkpointing,
567
+ "layerdrop": model_args.layerdrop,
568
+ "ctc_loss_reduction": model_args.ctc_loss_reduction,
569
+ "pad_token_id": tokenizer.pad_token_id,
570
+ "vocab_size": len(tokenizer),
571
+ "activation_dropout": model_args.activation_dropout,
572
+ }
573
+ )
574
+
575
+ # create model
576
+ model = AutoModelForCTC.from_pretrained(
577
+ model_args.model_name_or_path,
578
+ cache_dir=model_args.cache_dir,
579
+ config=config,
580
+ use_auth_token=data_args.use_auth_token,
581
+ )
582
+
583
+ # freeze encoder
584
+ if model_args.freeze_feature_encoder:
585
+ model.freeze_feature_encoder()
586
+
587
+ # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
588
+ # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
589
+ # so that we just need to set the correct target sampling rate and normalize the input
590
+ # via the `feature_extractor`
591
+
592
+ # make sure that dataset decodes audio with correct sampling rate
593
+ dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
594
+ if dataset_sampling_rate != feature_extractor.sampling_rate:
595
+ raw_datasets = raw_datasets.cast_column(
596
+ data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
597
+ )
598
+
599
+ # derive max & min input length for sample rate & max duration
600
+ max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
601
+ min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
602
+ audio_column_name = data_args.audio_column_name
603
+ num_workers = data_args.preprocessing_num_workers
604
+
605
+ # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
606
+ phoneme_language = data_args.phoneme_language
607
+
608
+ # Preprocessing the datasets.
609
+ # We need to read the audio files as arrays and tokenize the targets.
610
+ def prepare_dataset(batch):
611
+ # load audio
612
+ sample = batch[audio_column_name]
613
+
614
+ inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
615
+ batch["input_values"] = inputs.input_values[0]
616
+ batch["input_length"] = len(batch["input_values"])
617
+
618
+ # encode targets
619
+ additional_kwargs = {}
620
+ if phoneme_language is not None:
621
+ additional_kwargs["phonemizer_lang"] = phoneme_language
622
+
623
+ batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
624
+ return batch
625
+
626
+ with training_args.main_process_first(desc="dataset map preprocessing"):
627
+ vectorized_datasets = raw_datasets.map(
628
+ prepare_dataset,
629
+ remove_columns=next(iter(raw_datasets.values())).column_names,
630
+ num_proc=num_workers,
631
+ desc="preprocess datasets",
632
+ )
633
+
634
+ def is_audio_in_length_range(length):
635
+ return length > min_input_length and length < max_input_length
636
+
637
+ # filter data that is shorter than min_input_length
638
+ vectorized_datasets = vectorized_datasets.filter(
639
+ is_audio_in_length_range,
640
+ num_proc=num_workers,
641
+ input_columns=["input_length"],
642
+ )
643
+
644
+ # 7. Next, we can prepare the training.
645
+ # Let's use word error rate (WER) as our evaluation metric,
646
+ # instantiate a data collator and the trainer
647
+
648
+ # Define evaluation metrics during training, *i.e.* word error rate, character error rate
649
+ eval_metrics = {metric: evaluate.load(metric) for metric in data_args.eval_metrics}
650
+
651
+ # for large datasets it is advised to run the preprocessing on a
652
+ # single machine first with ``args.preprocessing_only`` since there will mostly likely
653
+ # be a timeout when running the script in distributed mode.
654
+ # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
655
+ # cached dataset
656
+ if data_args.preprocessing_only:
657
+ logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
658
+ return
659
+
660
+ def compute_metrics(pred):
661
+ pred_logits = pred.predictions
662
+ pred_ids = np.argmax(pred_logits, axis=-1)
663
+
664
+ pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
665
+
666
+ pred_str = tokenizer.batch_decode(pred_ids)
667
+ # we do not want to group tokens when computing the metrics
668
+ label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
669
+
670
+ metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
671
+
672
+ return metrics
673
+
674
+ # Now save everything to be able to create a single processor later
675
+ if is_main_process(training_args.local_rank):
676
+ # save feature extractor, tokenizer and config
677
+ feature_extractor.save_pretrained(training_args.output_dir)
678
+ tokenizer.save_pretrained(training_args.output_dir)
679
+ config.save_pretrained(training_args.output_dir)
680
+
681
+ try:
682
+ processor = AutoProcessor.from_pretrained(training_args.output_dir)
683
+ except (OSError, KeyError):
684
+ warnings.warn(
685
+ "Loading a processor from a feature extractor config that does not"
686
+ " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
687
+ " attribute to your `preprocessor_config.json` file to suppress this warning: "
688
+ " `'processor_class': 'Wav2Vec2Processor'`",
689
+ FutureWarning,
690
+ )
691
+ processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
692
+
693
+ # Instantiate custom data collator
694
+ data_collator = DataCollatorCTCWithPadding(processor=processor)
695
+
696
+ # Initialize Trainer
697
+ trainer = Trainer(
698
+ model=model,
699
+ data_collator=data_collator,
700
+ args=training_args,
701
+ compute_metrics=compute_metrics,
702
+ train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
703
+ eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
704
+ tokenizer=feature_extractor,
705
+ )
706
+
707
+ # 8. Finally, we can start training
708
+
709
+ # Training
710
+ if training_args.do_train:
711
+
712
+ # use last checkpoint if exist
713
+ if last_checkpoint is not None:
714
+ checkpoint = last_checkpoint
715
+ elif os.path.isdir(model_args.model_name_or_path):
716
+ checkpoint = model_args.model_name_or_path
717
+ else:
718
+ checkpoint = None
719
+
720
+ train_result = trainer.train(resume_from_checkpoint=checkpoint)
721
+ trainer.save_model()
722
+
723
+ metrics = train_result.metrics
724
+ max_train_samples = (
725
+ data_args.max_train_samples
726
+ if data_args.max_train_samples is not None
727
+ else len(vectorized_datasets["train"])
728
+ )
729
+ metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
730
+
731
+ trainer.log_metrics("train", metrics)
732
+ trainer.save_metrics("train", metrics)
733
+ trainer.save_state()
734
+
735
+ # Evaluation
736
+ results = {}
737
+ if training_args.do_eval:
738
+ logger.info("*** Evaluate ***")
739
+ metrics = trainer.evaluate()
740
+ max_eval_samples = (
741
+ data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
742
+ )
743
+ metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
744
+
745
+ trainer.log_metrics("eval", metrics)
746
+ trainer.save_metrics("eval", metrics)
747
+
748
+ # Write model card and (optionally) push to hub
749
+ config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
750
+ kwargs = {
751
+ "finetuned_from": model_args.model_name_or_path,
752
+ "tasks": "automatic-speech-recognition",
753
+ "tags": ["automatic-speech-recognition", data_args.dataset_name],
754
+ "dataset_args": (
755
+ f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split:"
756
+ f" {data_args.eval_split_name}"
757
+ ),
758
+ "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
759
+ }
760
+ if "common_voice" in data_args.dataset_name:
761
+ kwargs["language"] = config_name
762
+
763
+ if training_args.push_to_hub:
764
+ trainer.push_to_hub(**kwargs)
765
+ else:
766
+ trainer.create_model_card(**kwargs)
767
+
768
+ return results
769
+
770
+
771
+ if __name__ == "__main__":
772
+ main()
runs/Dec14_13-29-56_129-213-22-31/1671025286.7583845/events.out.tfevents.1671025286.129-213-22-31.83694.1 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:84fe7f9eb9bdee3f0ce972e8ae5cd67b151f661955efd2d0d8e279a1d976c2b6
3
+ size 5629
runs/Dec14_13-29-56_129-213-22-31/events.out.tfevents.1671025286.129-213-22-31.83694.0 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6a77e764bccdcdc7651b56a1d6603a74c655cda4a0d8f991fb83461ec03d6517
3
+ size 13343
special_tokens_map.json ADDED
@@ -0,0 +1,22 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "additional_special_tokens": [
3
+ {
4
+ "content": "<s>",
5
+ "lstrip": false,
6
+ "normalized": true,
7
+ "rstrip": false,
8
+ "single_word": false
9
+ },
10
+ {
11
+ "content": "</s>",
12
+ "lstrip": false,
13
+ "normalized": true,
14
+ "rstrip": false,
15
+ "single_word": false
16
+ }
17
+ ],
18
+ "bos_token": "<s>",
19
+ "eos_token": "</s>",
20
+ "pad_token": "[PAD]",
21
+ "unk_token": "[UNK]"
22
+ }
tokenizer_config.json ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "bos_token": "<s>",
3
+ "do_lower_case": false,
4
+ "eos_token": "</s>",
5
+ "model_max_length": 1000000000000000019884624838656,
6
+ "name_or_path": "./",
7
+ "pad_token": "[PAD]",
8
+ "replace_word_delimiter_char": " ",
9
+ "special_tokens_map_file": null,
10
+ "tokenizer_class": "Wav2Vec2CTCTokenizer",
11
+ "unk_token": "[UNK]",
12
+ "word_delimiter_token": "|"
13
+ }
training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:52dd59ad82b06ca181351e49d7e245e5b6920cc4a1d35f95f33ca3f843e75b0c
3
+ size 3451
vocab.json ADDED
@@ -0,0 +1,117 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "[PAD]": 114,
3
+ "[UNK]": 113,
4
+ "_": 1,
5
+ "`": 2,
6
+ "a": 3,
7
+ "b": 4,
8
+ "c": 5,
9
+ "d": 6,
10
+ "e": 7,
11
+ "f": 8,
12
+ "g": 9,
13
+ "h": 10,
14
+ "i": 11,
15
+ "j": 12,
16
+ "k": 13,
17
+ "l": 14,
18
+ "m": 15,
19
+ "n": 16,
20
+ "o": 17,
21
+ "p": 18,
22
+ "q": 19,
23
+ "r": 20,
24
+ "s": 21,
25
+ "t": 22,
26
+ "u": 23,
27
+ "v": 24,
28
+ "w": 25,
29
+ "x": 26,
30
+ "y": 27,
31
+ "z": 28,
32
+ "|": 0,
33
+ "¥": 29,
34
+ "°": 30,
35
+ "½": 31,
36
+ "¾": 32,
37
+ "é": 33,
38
+ "í": 34,
39
+ "،": 35,
40
+ "؛": 36,
41
+ "؟": 37,
42
+ "ء": 38,
43
+ "آ": 39,
44
+ "أ": 40,
45
+ "ؤ": 41,
46
+ "ئ": 42,
47
+ "ا": 43,
48
+ "ب": 44,
49
+ "ت": 45,
50
+ "ث": 46,
51
+ "ج": 47,
52
+ "ح": 48,
53
+ "خ": 49,
54
+ "د": 50,
55
+ "ذ": 51,
56
+ "ر": 52,
57
+ "ز": 53,
58
+ "س": 54,
59
+ "ش": 55,
60
+ "ص": 56,
61
+ "ض": 57,
62
+ "ط": 58,
63
+ "ظ": 59,
64
+ "ع": 60,
65
+ "غ": 61,
66
+ "ـ": 62,
67
+ "ف": 63,
68
+ "ق": 64,
69
+ "ك": 65,
70
+ "ل": 66,
71
+ "م": 67,
72
+ "ن": 68,
73
+ "ه": 69,
74
+ "و": 70,
75
+ "ى": 71,
76
+ "ي": 72,
77
+ "ً": 73,
78
+ "ٌ": 74,
79
+ "َ": 75,
80
+ "ُ": 76,
81
+ "ّ": 77,
82
+ "٪": 78,
83
+ "ټ": 79,
84
+ "پ": 80,
85
+ "ځ": 81,
86
+ "څ": 82,
87
+ "چ": 83,
88
+ "ډ": 84,
89
+ "ړ": 85,
90
+ "ږ": 86,
91
+ "ژ": 87,
92
+ "ښ": 88,
93
+ "ک": 89,
94
+ "ګ": 90,
95
+ "گ": 91,
96
+ "ڼ": 92,
97
+ "ھ": 93,
98
+ "ی": 94,
99
+ "ۍ": 95,
100
+ "ې": 96,
101
+ "ے": 97,
102
+ "۔": 98,
103
+ "۰": 99,
104
+ "۱": 100,
105
+ "۲": 101,
106
+ "۳": 102,
107
+ "۴": 103,
108
+ "۵": 104,
109
+ "۶": 105,
110
+ "۷": 106,
111
+ "۸": 107,
112
+ "۹": 108,
113
+ "‌": 109,
114
+ "‍": 110,
115
+ "–": 111,
116
+ "—": 112
117
+ }