tomascufaro commited on
Commit
36107b3
1 Parent(s): 28d0b9a

copy repo from tomascufaro

Browse files
Files changed (50) hide show
  1. README.md +212 -0
  2. added_tokens.json +1 -0
  3. all_results.json +14 -0
  4. checkpoint-70500/config.json +115 -0
  5. checkpoint-70500/optimizer.pt +3 -0
  6. checkpoint-70500/preprocessor_config.json +9 -0
  7. checkpoint-70500/pytorch_model.bin +3 -0
  8. checkpoint-70500/rng_state.pth +3 -0
  9. checkpoint-70500/scaler.pt +3 -0
  10. checkpoint-70500/scheduler.pt +3 -0
  11. checkpoint-70500/trainer_state.json +0 -0
  12. checkpoint-70500/training_args.bin +3 -0
  13. checkpoint-71000/config.json +115 -0
  14. checkpoint-71000/optimizer.pt +3 -0
  15. checkpoint-71000/preprocessor_config.json +9 -0
  16. checkpoint-71000/pytorch_model.bin +3 -0
  17. checkpoint-71000/rng_state.pth +3 -0
  18. checkpoint-71000/scaler.pt +3 -0
  19. checkpoint-71000/scheduler.pt +3 -0
  20. checkpoint-71000/trainer_state.json +0 -0
  21. checkpoint-71000/training_args.bin +3 -0
  22. checkpoint-71500/config.json +115 -0
  23. checkpoint-71500/optimizer.pt +3 -0
  24. checkpoint-71500/preprocessor_config.json +9 -0
  25. checkpoint-71500/pytorch_model.bin +3 -0
  26. checkpoint-71500/rng_state.pth +3 -0
  27. checkpoint-71500/scaler.pt +3 -0
  28. checkpoint-71500/scheduler.pt +3 -0
  29. checkpoint-71500/trainer_state.json +0 -0
  30. checkpoint-71500/training_args.bin +3 -0
  31. config.json +115 -0
  32. eval.py +157 -0
  33. eval_results.json +9 -0
  34. log_mozilla-foundation_common_voice_8_0_es_test_predictions.txt +0 -0
  35. log_mozilla-foundation_common_voice_8_0_es_test_targets.txt +0 -0
  36. log_mozilla-foundation_common_voice_8_0_es_validation_predictions.txt +0 -0
  37. log_mozilla-foundation_common_voice_8_0_es_validation_targets.txt +0 -0
  38. mozilla-foundation_common_voice_8_0_es_test_eval_results.txt +2 -0
  39. mozilla-foundation_common_voice_8_0_es_validation_eval_results.txt +2 -0
  40. preprocessor_config.json +9 -0
  41. pytorch_model.bin +3 -0
  42. run.sh +35 -0
  43. run_speech_recognition_ctc.py +737 -0
  44. run_speech_recognition_ctc_bnb.py +804 -0
  45. special_tokens_map.json +1 -0
  46. tokenizer_config.json +1 -0
  47. train_results.json +8 -0
  48. trainer_state.json +0 -0
  49. training_args.bin +3 -0
  50. vocab.json +1 -0
README.md ADDED
@@ -0,0 +1,212 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - es
4
+ license: apache-2.0
5
+ tags:
6
+ - automatic-speech-recognition
7
+ - mozilla-foundation/common_voice_8_0
8
+ - generated_from_trainer
9
+ - "es"
10
+ - "robust-speech-event"
11
+ datasets:
12
+ - common_voice
13
+ model-index:
14
+ - name: xls-r-es-test
15
+ results: []
16
+ ---
17
+
18
+ <!-- This model card has been generated automatically according to the information the Trainer had access to. You
19
+ should probably proofread and complete it, then remove this comment. -->
20
+
21
+ # xls-r-es-test
22
+
23
+ This model is a fine-tuned version of [facebook/wav2vec2-large-xlsr-53](https://huggingface.co/facebook/wav2vec2-large-xlsr-53) on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - ES dataset.
24
+ It achieves the following results on the evaluation set:
25
+ - Loss: 0.1304
26
+ - WER: 0.1261
27
+ - CER: 0.035
28
+
29
+ ## Model description
30
+
31
+ More information needed
32
+
33
+ ## Intended uses & limitations
34
+
35
+ More information needed
36
+
37
+ ## Training and evaluation data
38
+
39
+ More information needed
40
+
41
+ ## Training procedure
42
+
43
+ ### Training hyperparameters
44
+
45
+ The following hyperparameters were used during training:
46
+ - learning_rate: 7.5e-05
47
+ - train_batch_size: 8
48
+ - eval_batch_size: 8
49
+ - seed: 42
50
+ - gradient_accumulation_steps: 4
51
+ - total_train_batch_size: 32
52
+ - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
53
+ - lr_scheduler_type: linear
54
+ - lr_scheduler_warmup_steps: 2000
55
+ - num_epochs: 10.0
56
+ - mixed_precision_training: Native AMP
57
+
58
+ ### Training results
59
+
60
+ | Training Loss | Epoch | Step | Validation Loss | Wer |
61
+ |:-------------:|:-----:|:-----:|:---------------:|:------:|
62
+ | 2.9613 | 0.07 | 500 | 2.9647 | 1.0 |
63
+ | 2.604 | 0.14 | 1000 | 1.8300 | 0.9562 |
64
+ | 1.177 | 0.21 | 1500 | 0.3652 | 0.3077 |
65
+ | 1.0745 | 0.28 | 2000 | 0.2707 | 0.2504 |
66
+ | 1.0103 | 0.35 | 2500 | 0.2338 | 0.2157 |
67
+ | 0.9858 | 0.42 | 3000 | 0.2321 | 0.2129 |
68
+ | 0.974 | 0.49 | 3500 | 0.2164 | 0.2031 |
69
+ | 0.9699 | 0.56 | 4000 | 0.2078 | 0.1970 |
70
+ | 0.9513 | 0.63 | 4500 | 0.2173 | 0.2139 |
71
+ | 0.9657 | 0.7 | 5000 | 0.2050 | 0.1979 |
72
+ | 0.9484 | 0.77 | 5500 | 0.2008 | 0.1919 |
73
+ | 0.9317 | 0.84 | 6000 | 0.2012 | 0.1911 |
74
+ | 0.9366 | 0.91 | 6500 | 0.2024 | 0.1976 |
75
+ | 0.9242 | 0.98 | 7000 | 0.2062 | 0.2028 |
76
+ | 0.9138 | 1.05 | 7500 | 0.1924 | 0.1863 |
77
+ | 0.921 | 1.12 | 8000 | 0.1935 | 0.1836 |
78
+ | 0.9117 | 1.19 | 8500 | 0.1887 | 0.1815 |
79
+ | 0.9064 | 1.26 | 9000 | 0.1909 | 0.1839 |
80
+ | 0.9118 | 1.32 | 9500 | 0.1869 | 0.1830 |
81
+ | 0.9121 | 1.39 | 10000 | 0.1863 | 0.1802 |
82
+ | 0.9048 | 1.46 | 10500 | 0.1845 | 0.1791 |
83
+ | 0.8955 | 1.53 | 11000 | 0.1863 | 0.1774 |
84
+ | 0.8947 | 1.6 | 11500 | 0.1907 | 0.1814 |
85
+ | 0.9073 | 1.67 | 12000 | 0.1892 | 0.1853 |
86
+ | 0.8927 | 1.74 | 12500 | 0.1821 | 0.1750 |
87
+ | 0.8732 | 1.81 | 13000 | 0.1815 | 0.1768 |
88
+ | 0.8761 | 1.88 | 13500 | 0.1822 | 0.1749 |
89
+ | 0.8751 | 1.95 | 14000 | 0.1789 | 0.1715 |
90
+ | 0.8889 | 2.02 | 14500 | 0.1819 | 0.1791 |
91
+ | 0.8864 | 2.09 | 15000 | 0.1826 | 0.1794 |
92
+ | 0.886 | 2.16 | 15500 | 0.1788 | 0.1776 |
93
+ | 0.8915 | 2.23 | 16000 | 0.1756 | 0.1719 |
94
+ | 0.8689 | 2.3 | 16500 | 0.1769 | 0.1711 |
95
+ | 0.879 | 2.37 | 17000 | 0.1777 | 0.1739 |
96
+ | 0.8692 | 2.44 | 17500 | 0.1765 | 0.1705 |
97
+ | 0.8504 | 2.51 | 18000 | 0.1699 | 0.1652 |
98
+ | 0.8728 | 2.58 | 18500 | 0.1705 | 0.1694 |
99
+ | 0.8523 | 2.65 | 19000 | 0.1674 | 0.1645 |
100
+ | 0.8513 | 2.72 | 19500 | 0.1661 | 0.1611 |
101
+ | 0.8498 | 2.79 | 20000 | 0.1660 | 0.1631 |
102
+ | 0.8432 | 2.86 | 20500 | 0.1636 | 0.1610 |
103
+ | 0.8492 | 2.93 | 21000 | 0.1708 | 0.1688 |
104
+ | 0.8561 | 3.0 | 21500 | 0.1663 | 0.1604 |
105
+ | 0.842 | 3.07 | 22000 | 0.1690 | 0.1625 |
106
+ | 0.857 | 3.14 | 22500 | 0.1642 | 0.1605 |
107
+ | 0.8518 | 3.21 | 23000 | 0.1626 | 0.1585 |
108
+ | 0.8506 | 3.28 | 23500 | 0.1651 | 0.1605 |
109
+ | 0.8394 | 3.35 | 24000 | 0.1647 | 0.1585 |
110
+ | 0.8431 | 3.42 | 24500 | 0.1632 | 0.1573 |
111
+ | 0.8566 | 3.49 | 25000 | 0.1614 | 0.1550 |
112
+ | 0.8534 | 3.56 | 25500 | 0.1645 | 0.1589 |
113
+ | 0.8386 | 3.63 | 26000 | 0.1632 | 0.1582 |
114
+ | 0.8357 | 3.7 | 26500 | 0.1631 | 0.1556 |
115
+ | 0.8299 | 3.77 | 27000 | 0.1612 | 0.1550 |
116
+ | 0.8421 | 3.84 | 27500 | 0.1602 | 0.1552 |
117
+ | 0.8375 | 3.91 | 28000 | 0.1592 | 0.1537 |
118
+ | 0.8328 | 3.97 | 28500 | 0.1587 | 0.1537 |
119
+ | 0.8155 | 4.04 | 29000 | 0.1587 | 0.1520 |
120
+ | 0.8335 | 4.11 | 29500 | 0.1624 | 0.1556 |
121
+ | 0.8138 | 4.18 | 30000 | 0.1581 | 0.1547 |
122
+ | 0.8195 | 4.25 | 30500 | 0.1560 | 0.1507 |
123
+ | 0.8092 | 4.32 | 31000 | 0.1561 | 0.1534 |
124
+ | 0.8191 | 4.39 | 31500 | 0.1549 | 0.1493 |
125
+ | 0.8008 | 4.46 | 32000 | 0.1540 | 0.1493 |
126
+ | 0.8138 | 4.53 | 32500 | 0.1544 | 0.1493 |
127
+ | 0.8173 | 4.6 | 33000 | 0.1553 | 0.1511 |
128
+ | 0.8081 | 4.67 | 33500 | 0.1541 | 0.1484 |
129
+ | 0.8192 | 4.74 | 34000 | 0.1560 | 0.1506 |
130
+ | 0.8068 | 4.81 | 34500 | 0.1540 | 0.1503 |
131
+ | 0.8105 | 4.88 | 35000 | 0.1529 | 0.1483 |
132
+ | 0.7976 | 4.95 | 35500 | 0.1507 | 0.1451 |
133
+ | 0.8143 | 5.02 | 36000 | 0.1505 | 0.1462 |
134
+ | 0.8053 | 5.09 | 36500 | 0.1517 | 0.1476 |
135
+ | 0.785 | 5.16 | 37000 | 0.1526 | 0.1478 |
136
+ | 0.7936 | 5.23 | 37500 | 0.1489 | 0.1421 |
137
+ | 0.807 | 5.3 | 38000 | 0.1483 | 0.1420 |
138
+ | 0.8092 | 5.37 | 38500 | 0.1481 | 0.1435 |
139
+ | 0.793 | 5.44 | 39000 | 0.1503 | 0.1438 |
140
+ | 0.814 | 5.51 | 39500 | 0.1495 | 0.1480 |
141
+ | 0.807 | 5.58 | 40000 | 0.1472 | 0.1424 |
142
+ | 0.7913 | 5.65 | 40500 | 0.1471 | 0.1422 |
143
+ | 0.7844 | 5.72 | 41000 | 0.1473 | 0.1422 |
144
+ | 0.7888 | 5.79 | 41500 | 0.1445 | 0.1385 |
145
+ | 0.7806 | 5.86 | 42000 | 0.1435 | 0.1394 |
146
+ | 0.7773 | 5.93 | 42500 | 0.1461 | 0.1424 |
147
+ | 0.786 | 6.0 | 43000 | 0.1450 | 0.1413 |
148
+ | 0.7784 | 6.07 | 43500 | 0.1463 | 0.1424 |
149
+ | 0.7937 | 6.14 | 44000 | 0.1438 | 0.1386 |
150
+ | 0.7738 | 6.21 | 44500 | 0.1437 | 0.1383 |
151
+ | 0.7728 | 6.28 | 45000 | 0.1424 | 0.1371 |
152
+ | 0.7681 | 6.35 | 45500 | 0.1416 | 0.1376 |
153
+ | 0.776 | 6.42 | 46000 | 0.1415 | 0.1380 |
154
+ | 0.7773 | 6.49 | 46500 | 0.1416 | 0.1371 |
155
+ | 0.7692 | 6.56 | 47000 | 0.1398 | 0.1345 |
156
+ | 0.7642 | 6.62 | 47500 | 0.1381 | 0.1341 |
157
+ | 0.7692 | 6.69 | 48000 | 0.1392 | 0.1334 |
158
+ | 0.7667 | 6.76 | 48500 | 0.1392 | 0.1348 |
159
+ | 0.7712 | 6.83 | 49000 | 0.1398 | 0.1333 |
160
+ | 0.7628 | 6.9 | 49500 | 0.1392 | 0.1344 |
161
+ | 0.7622 | 6.97 | 50000 | 0.1377 | 0.1329 |
162
+ | 0.7639 | 7.04 | 50500 | 0.1361 | 0.1316 |
163
+ | 0.742 | 7.11 | 51000 | 0.1376 | 0.1327 |
164
+ | 0.7526 | 7.18 | 51500 | 0.1387 | 0.1342 |
165
+ | 0.7606 | 7.25 | 52000 | 0.1363 | 0.1316 |
166
+ | 0.7626 | 7.32 | 52500 | 0.1365 | 0.1313 |
167
+ | 0.752 | 7.39 | 53000 | 0.1354 | 0.1309 |
168
+ | 0.7562 | 7.46 | 53500 | 0.1362 | 0.1312 |
169
+ | 0.7557 | 7.53 | 54000 | 0.1358 | 0.1325 |
170
+ | 0.7588 | 7.6 | 54500 | 0.1343 | 0.1311 |
171
+ | 0.7485 | 7.67 | 55000 | 0.1346 | 0.1301 |
172
+ | 0.7466 | 7.74 | 55500 | 0.1354 | 0.1314 |
173
+ | 0.7558 | 7.81 | 56000 | 0.1359 | 0.1325 |
174
+ | 0.7578 | 7.88 | 56500 | 0.1363 | 0.1334 |
175
+ | 0.7411 | 7.95 | 57000 | 0.1346 | 0.1301 |
176
+ | 0.7478 | 8.02 | 57500 | 0.1355 | 0.1305 |
177
+ | 0.7451 | 8.09 | 58000 | 0.1349 | 0.1302 |
178
+ | 0.7383 | 8.16 | 58500 | 0.1349 | 0.1294 |
179
+ | 0.7482 | 8.23 | 59000 | 0.1341 | 0.1293 |
180
+ | 0.742 | 8.3 | 59500 | 0.1338 | 0.1296 |
181
+ | 0.7343 | 8.37 | 60000 | 0.1348 | 0.1307 |
182
+ | 0.7385 | 8.44 | 60500 | 0.1324 | 0.1282 |
183
+ | 0.7567 | 8.51 | 61000 | 0.1334 | 0.1281 |
184
+ | 0.7342 | 8.58 | 61500 | 0.1338 | 0.1289 |
185
+ | 0.7401 | 8.65 | 62000 | 0.1331 | 0.1285 |
186
+ | 0.7362 | 8.72 | 62500 | 0.1329 | 0.1283 |
187
+ | 0.7241 | 8.79 | 63000 | 0.1323 | 0.1277 |
188
+ | 0.7244 | 8.86 | 63500 | 0.1317 | 0.1269 |
189
+ | 0.7274 | 8.93 | 64000 | 0.1308 | 0.1260 |
190
+ | 0.7411 | 9.0 | 64500 | 0.1309 | 0.1256 |
191
+ | 0.7255 | 9.07 | 65000 | 0.1316 | 0.1265 |
192
+ | 0.7406 | 9.14 | 65500 | 0.1315 | 0.1270 |
193
+ | 0.7418 | 9.21 | 66000 | 0.1315 | 0.1269 |
194
+ | 0.7301 | 9.27 | 66500 | 0.1315 | 0.1273 |
195
+ | 0.7248 | 9.34 | 67000 | 0.1323 | 0.1274 |
196
+ | 0.7423 | 9.41 | 67500 | 0.1309 | 0.1267 |
197
+ | 0.7152 | 9.48 | 68000 | 0.1312 | 0.1271 |
198
+ | 0.7295 | 9.55 | 68500 | 0.1306 | 0.1262 |
199
+ | 0.7231 | 9.62 | 69000 | 0.1308 | 0.1263 |
200
+ | 0.7344 | 9.69 | 69500 | 0.1313 | 0.1267 |
201
+ | 0.7264 | 9.76 | 70000 | 0.1305 | 0.1263 |
202
+ | 0.7309 | 9.83 | 70500 | 0.1303 | 0.1262 |
203
+ | 0.73 | 9.9 | 71000 | 0.1303 | 0.1261 |
204
+ | 0.7353 | 9.97 | 71500 | 0.1304 | 0.1260 |
205
+
206
+
207
+ ### Framework versions
208
+
209
+ - Transformers 4.17.0.dev0
210
+ - Pytorch 1.10.2+cu102
211
+ - Datasets 1.18.3
212
+ - Tokenizers 0.11.0
added_tokens.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"<s>": 34, "</s>": 35}
all_results.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "epoch": 10.0,
3
+ "eval_loss": 0.13038970530033112,
4
+ "eval_runtime": 829.1884,
5
+ "eval_samples": 15440,
6
+ "eval_samples_per_second": 18.621,
7
+ "eval_steps_per_second": 2.328,
8
+ "eval_wer": 0.12614800858667752,
9
+ "train_loss": 0.8738376914307662,
10
+ "train_runtime": 268775.0566,
11
+ "train_samples": 229440,
12
+ "train_samples_per_second": 8.537,
13
+ "train_steps_per_second": 0.267
14
+ }
checkpoint-70500/config.json ADDED
@@ -0,0 +1,115 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "facebook/wav2vec2-large-xlsr-53",
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_channel_length": 10,
63
+ "mask_channel_min_space": 1,
64
+ "mask_channel_other": 0.0,
65
+ "mask_channel_prob": 0.0,
66
+ "mask_channel_selection": "static",
67
+ "mask_feature_length": 64,
68
+ "mask_feature_min_masks": 0,
69
+ "mask_feature_prob": 0.25,
70
+ "mask_time_length": 10,
71
+ "mask_time_min_masks": 2,
72
+ "mask_time_min_space": 1,
73
+ "mask_time_other": 0.0,
74
+ "mask_time_prob": 0.75,
75
+ "mask_time_selection": "static",
76
+ "model_type": "wav2vec2",
77
+ "num_adapter_layers": 3,
78
+ "num_attention_heads": 16,
79
+ "num_codevector_groups": 2,
80
+ "num_codevectors_per_group": 320,
81
+ "num_conv_pos_embedding_groups": 16,
82
+ "num_conv_pos_embeddings": 128,
83
+ "num_feat_extract_layers": 7,
84
+ "num_hidden_layers": 24,
85
+ "num_negatives": 100,
86
+ "output_hidden_size": 1024,
87
+ "pad_token_id": 33,
88
+ "proj_codevector_dim": 768,
89
+ "tdnn_dilation": [
90
+ 1,
91
+ 2,
92
+ 3,
93
+ 1,
94
+ 1
95
+ ],
96
+ "tdnn_dim": [
97
+ 512,
98
+ 512,
99
+ 512,
100
+ 512,
101
+ 1500
102
+ ],
103
+ "tdnn_kernel": [
104
+ 5,
105
+ 3,
106
+ 3,
107
+ 1,
108
+ 1
109
+ ],
110
+ "torch_dtype": "float32",
111
+ "transformers_version": "4.17.0.dev0",
112
+ "use_weighted_layer_sum": false,
113
+ "vocab_size": 36,
114
+ "xvector_output_dim": 512
115
+ }
checkpoint-70500/optimizer.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:47a02c311c6d4acd312ad5496b5163fa4de0cef279fded334076f045a08fef71
3
+ size 625441705
checkpoint-70500/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
+ }
checkpoint-70500/pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6288a1bc25c7f7b2037d71bf0e0c1e5f8b3810fcf35a0cc84e9b281bc73949ba
3
+ size 1262071281
checkpoint-70500/rng_state.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:16195a6451d36ddc8b5fcfaa9eb0734ba3cb4e5d72dcf40491718947afcfc21a
3
+ size 14567
checkpoint-70500/scaler.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d8aa1005e9cffde08c96316b6177ad9bfcf3e109a09807974dca4303ffc79176
3
+ size 559
checkpoint-70500/scheduler.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4878d36d0a3143d31e8d70b00241f38f521125c72becf0f1f46c3a1f86ce8352
3
+ size 623
checkpoint-70500/trainer_state.json ADDED
The diff for this file is too large to render. See raw diff
 
checkpoint-70500/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:61af16ecf3f3e38206ca59d2abf9c3872cfd4e4f38971380d8bed24a74daad7d
3
+ size 3055
checkpoint-71000/config.json ADDED
@@ -0,0 +1,115 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "facebook/wav2vec2-large-xlsr-53",
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_channel_length": 10,
63
+ "mask_channel_min_space": 1,
64
+ "mask_channel_other": 0.0,
65
+ "mask_channel_prob": 0.0,
66
+ "mask_channel_selection": "static",
67
+ "mask_feature_length": 64,
68
+ "mask_feature_min_masks": 0,
69
+ "mask_feature_prob": 0.25,
70
+ "mask_time_length": 10,
71
+ "mask_time_min_masks": 2,
72
+ "mask_time_min_space": 1,
73
+ "mask_time_other": 0.0,
74
+ "mask_time_prob": 0.75,
75
+ "mask_time_selection": "static",
76
+ "model_type": "wav2vec2",
77
+ "num_adapter_layers": 3,
78
+ "num_attention_heads": 16,
79
+ "num_codevector_groups": 2,
80
+ "num_codevectors_per_group": 320,
81
+ "num_conv_pos_embedding_groups": 16,
82
+ "num_conv_pos_embeddings": 128,
83
+ "num_feat_extract_layers": 7,
84
+ "num_hidden_layers": 24,
85
+ "num_negatives": 100,
86
+ "output_hidden_size": 1024,
87
+ "pad_token_id": 33,
88
+ "proj_codevector_dim": 768,
89
+ "tdnn_dilation": [
90
+ 1,
91
+ 2,
92
+ 3,
93
+ 1,
94
+ 1
95
+ ],
96
+ "tdnn_dim": [
97
+ 512,
98
+ 512,
99
+ 512,
100
+ 512,
101
+ 1500
102
+ ],
103
+ "tdnn_kernel": [
104
+ 5,
105
+ 3,
106
+ 3,
107
+ 1,
108
+ 1
109
+ ],
110
+ "torch_dtype": "float32",
111
+ "transformers_version": "4.17.0.dev0",
112
+ "use_weighted_layer_sum": false,
113
+ "vocab_size": 36,
114
+ "xvector_output_dim": 512
115
+ }
checkpoint-71000/optimizer.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2bff2547a0d95e50b19fde82898cb9b16efd386633a5b8850b9165395329e057
3
+ size 625441705
checkpoint-71000/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
+ }
checkpoint-71000/pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e99d50f1fa2d84154fde43dac8860f190bad6ed3e1924bfcacb14150d1de5966
3
+ size 1262071281
checkpoint-71000/rng_state.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:97f1ee927a6f18a282804da2ad73c504b6f37ec114596d451e4826ae48a72697
3
+ size 14567
checkpoint-71000/scaler.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a0fa447ce75544c22980a80903eb85ec2a42fdeb1abc8a1b5a5bd6e10c09b83e
3
+ size 559
checkpoint-71000/scheduler.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ba620a5a4ef7914840abb2769d3561ca0e741367894940b97d888d0c6b4f2aa3
3
+ size 623
checkpoint-71000/trainer_state.json ADDED
The diff for this file is too large to render. See raw diff
 
checkpoint-71000/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:61af16ecf3f3e38206ca59d2abf9c3872cfd4e4f38971380d8bed24a74daad7d
3
+ size 3055
checkpoint-71500/config.json ADDED
@@ -0,0 +1,115 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "facebook/wav2vec2-large-xlsr-53",
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_channel_length": 10,
63
+ "mask_channel_min_space": 1,
64
+ "mask_channel_other": 0.0,
65
+ "mask_channel_prob": 0.0,
66
+ "mask_channel_selection": "static",
67
+ "mask_feature_length": 64,
68
+ "mask_feature_min_masks": 0,
69
+ "mask_feature_prob": 0.25,
70
+ "mask_time_length": 10,
71
+ "mask_time_min_masks": 2,
72
+ "mask_time_min_space": 1,
73
+ "mask_time_other": 0.0,
74
+ "mask_time_prob": 0.75,
75
+ "mask_time_selection": "static",
76
+ "model_type": "wav2vec2",
77
+ "num_adapter_layers": 3,
78
+ "num_attention_heads": 16,
79
+ "num_codevector_groups": 2,
80
+ "num_codevectors_per_group": 320,
81
+ "num_conv_pos_embedding_groups": 16,
82
+ "num_conv_pos_embeddings": 128,
83
+ "num_feat_extract_layers": 7,
84
+ "num_hidden_layers": 24,
85
+ "num_negatives": 100,
86
+ "output_hidden_size": 1024,
87
+ "pad_token_id": 33,
88
+ "proj_codevector_dim": 768,
89
+ "tdnn_dilation": [
90
+ 1,
91
+ 2,
92
+ 3,
93
+ 1,
94
+ 1
95
+ ],
96
+ "tdnn_dim": [
97
+ 512,
98
+ 512,
99
+ 512,
100
+ 512,
101
+ 1500
102
+ ],
103
+ "tdnn_kernel": [
104
+ 5,
105
+ 3,
106
+ 3,
107
+ 1,
108
+ 1
109
+ ],
110
+ "torch_dtype": "float32",
111
+ "transformers_version": "4.17.0.dev0",
112
+ "use_weighted_layer_sum": false,
113
+ "vocab_size": 36,
114
+ "xvector_output_dim": 512
115
+ }
checkpoint-71500/optimizer.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e9a1e38dad1e2483e267f3026a959de17baa64ef7c5a1cb1e8255dcac08ebf5d
3
+ size 625441705
checkpoint-71500/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
+ }
checkpoint-71500/pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:74552bd7b8157ca706aa3e198982daab293d03bb6cce08c9062f08b41f0f7b48
3
+ size 1262071281
checkpoint-71500/rng_state.pth ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8f15e452de37bdd2c2324cf1ed19c5fb9d48fddca22bdf522d58abb6a1edd30f
3
+ size 14567
checkpoint-71500/scaler.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:bae09b9c1fcdd03cf16eeecd7db55e390678977b40359da069ac343cf9f745a6
3
+ size 559
checkpoint-71500/scheduler.pt ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c0b408326a7d0881f3a7694ebe5e2383eedd9eb758558a68e1de5d9597629fd6
3
+ size 623
checkpoint-71500/trainer_state.json ADDED
The diff for this file is too large to render. See raw diff
 
checkpoint-71500/training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:61af16ecf3f3e38206ca59d2abf9c3872cfd4e4f38971380d8bed24a74daad7d
3
+ size 3055
config.json ADDED
@@ -0,0 +1,115 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "facebook/wav2vec2-large-xlsr-53",
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_channel_length": 10,
63
+ "mask_channel_min_space": 1,
64
+ "mask_channel_other": 0.0,
65
+ "mask_channel_prob": 0.0,
66
+ "mask_channel_selection": "static",
67
+ "mask_feature_length": 64,
68
+ "mask_feature_min_masks": 0,
69
+ "mask_feature_prob": 0.25,
70
+ "mask_time_length": 10,
71
+ "mask_time_min_masks": 2,
72
+ "mask_time_min_space": 1,
73
+ "mask_time_other": 0.0,
74
+ "mask_time_prob": 0.75,
75
+ "mask_time_selection": "static",
76
+ "model_type": "wav2vec2",
77
+ "num_adapter_layers": 3,
78
+ "num_attention_heads": 16,
79
+ "num_codevector_groups": 2,
80
+ "num_codevectors_per_group": 320,
81
+ "num_conv_pos_embedding_groups": 16,
82
+ "num_conv_pos_embeddings": 128,
83
+ "num_feat_extract_layers": 7,
84
+ "num_hidden_layers": 24,
85
+ "num_negatives": 100,
86
+ "output_hidden_size": 1024,
87
+ "pad_token_id": 33,
88
+ "proj_codevector_dim": 768,
89
+ "tdnn_dilation": [
90
+ 1,
91
+ 2,
92
+ 3,
93
+ 1,
94
+ 1
95
+ ],
96
+ "tdnn_dim": [
97
+ 512,
98
+ 512,
99
+ 512,
100
+ 512,
101
+ 1500
102
+ ],
103
+ "tdnn_kernel": [
104
+ 5,
105
+ 3,
106
+ 3,
107
+ 1,
108
+ 1
109
+ ],
110
+ "torch_dtype": "float32",
111
+ "transformers_version": "4.17.0.dev0",
112
+ "use_weighted_layer_sum": false,
113
+ "vocab_size": 36,
114
+ "xvector_output_dim": 512
115
+ }
eval.py ADDED
@@ -0,0 +1,157 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ import argparse
3
+ import re
4
+ from typing import Dict
5
+
6
+ import torch
7
+ from datasets import Audio, Dataset, load_dataset, load_metric
8
+
9
+ from transformers import AutoFeatureExtractor, pipeline
10
+
11
+
12
+ def log_results(result: Dataset, args: Dict[str, str]):
13
+ """DO NOT CHANGE. This function computes and logs the result metrics."""
14
+
15
+ log_outputs = args.log_outputs
16
+ dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
17
+
18
+ # load metric
19
+ wer = load_metric("wer")
20
+ cer = load_metric("cer")
21
+
22
+ # compute metrics
23
+ wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
24
+ cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
25
+
26
+ # print & log results
27
+ result_str = f"WER: {wer_result}\n" f"CER: {cer_result}"
28
+ print(result_str)
29
+
30
+ with open(f"{dataset_id}_eval_results.txt", "w") as f:
31
+ f.write(result_str)
32
+
33
+ # log all results in text file. Possibly interesting for analysis
34
+ if log_outputs is not None:
35
+ pred_file = f"log_{dataset_id}_predictions.txt"
36
+ target_file = f"log_{dataset_id}_targets.txt"
37
+
38
+ with open(pred_file, "w") as p, open(target_file, "w") as t:
39
+
40
+ # mapping function to write output
41
+ def write_to_file(batch, i):
42
+ p.write(f"{i}" + "\n")
43
+ p.write(batch["prediction"] + "\n")
44
+ t.write(f"{i}" + "\n")
45
+ t.write(batch["target"] + "\n")
46
+
47
+ result.map(write_to_file, with_indices=True)
48
+
49
+ def clean_batch(text):
50
+ text = re.sub("([^A-Za-zÀ-ú ])", '', text).lower()
51
+ text = re.sub("([ß|þ|ð|æ])",'',text)
52
+ return text
53
+
54
+ def homologate_accents(text):
55
+ text=re.sub("([â|ã|ä|å|à])","a",text)
56
+ text=re.sub("([é|ê|ë])","e",text)
57
+ text=re.sub("([ì|î|ï])","i",text)
58
+ text=re.sub("([ö|õ|ô|ò|ø])","o",text)
59
+ text=re.sub("ù","u",text)
60
+ text=re.sub("ç","c",text)
61
+ return text
62
+
63
+ def normalize_text(text: str) -> str:
64
+ """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
65
+
66
+ chars_to_ignore_regex = '[,?.!\-\;\:"“%‘”�—’…–]' # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
67
+ text = text.lower()
68
+ text = re.sub(chars_to_ignore_regex, "", text.lower())
69
+
70
+ # In addition, we can normalize the target text, e.g. removing new lines characters etc...
71
+ # note that order is important here!
72
+ token_sequences_to_ignore = ["\n\n", "\n", " ", " "]
73
+
74
+ for t in token_sequences_to_ignore:
75
+ text = " ".join(text.split(t))
76
+
77
+ #added functions
78
+ text = homologate_accents(text)
79
+ text = clean_batch(text)
80
+ return text
81
+
82
+
83
+ def main(args):
84
+ # load dataset
85
+ dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
86
+
87
+ # for testing: only process the first two examples as a test
88
+ #dataset = dataset.select(range(15))
89
+ # vocab = [character for character in "aábcdeéfghiíjklmnñoópqrstuúüvwxyz·-."]
90
+
91
+ # dataset = dataset.filter(
92
+ # lambda example: not any((c not in vocab) for c in example),
93
+ # input_columns='sentence',
94
+ # desc="remove examples with weird characters"
95
+ # )
96
+ # load processor
97
+ feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
98
+ sampling_rate = feature_extractor.sampling_rate
99
+
100
+ # resample audio
101
+ dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
102
+
103
+ # load eval pipeline
104
+ if args.device is None:
105
+ args.device = 0 if torch.cuda.is_available() else -1
106
+ asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
107
+
108
+ # map function to decode audio
109
+ def map_to_pred(batch):
110
+ prediction = asr(
111
+ batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
112
+ )
113
+ batch["prediction"] = prediction["text"]
114
+ batch["target"] = normalize_text(batch["sentence"])
115
+ return batch
116
+
117
+ # run inference on all examples
118
+ result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
119
+ # compute and log_results
120
+ # do not change function below
121
+ log_results(result, args)
122
+
123
+
124
+ if __name__ == "__main__":
125
+ parser = argparse.ArgumentParser()
126
+
127
+ parser.add_argument(
128
+ "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
129
+ )
130
+ parser.add_argument(
131
+ "--dataset",
132
+ type=str,
133
+ required=True,
134
+ help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
135
+ )
136
+ parser.add_argument(
137
+ "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'` for Common Voice"
138
+ )
139
+ parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
140
+ parser.add_argument(
141
+ "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
142
+ )
143
+ parser.add_argument(
144
+ "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
145
+ )
146
+ parser.add_argument(
147
+ "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
148
+ )
149
+ parser.add_argument(
150
+ "--device",
151
+ type=int,
152
+ default=None,
153
+ help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
154
+ )
155
+ args = parser.parse_args()
156
+
157
+ main(args)
eval_results.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "epoch": 10.0,
3
+ "eval_loss": 0.13038970530033112,
4
+ "eval_runtime": 829.1884,
5
+ "eval_samples": 15440,
6
+ "eval_samples_per_second": 18.621,
7
+ "eval_steps_per_second": 2.328,
8
+ "eval_wer": 0.12614800858667752
9
+ }
log_mozilla-foundation_common_voice_8_0_es_test_predictions.txt ADDED
The diff for this file is too large to render. See raw diff
 
log_mozilla-foundation_common_voice_8_0_es_test_targets.txt ADDED
The diff for this file is too large to render. See raw diff
 
log_mozilla-foundation_common_voice_8_0_es_validation_predictions.txt ADDED
The diff for this file is too large to render. See raw diff
 
log_mozilla-foundation_common_voice_8_0_es_validation_targets.txt ADDED
The diff for this file is too large to render. See raw diff
 
mozilla-foundation_common_voice_8_0_es_test_eval_results.txt ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ WER: 0.12618083227750462
2
+ CER: 0.035028395923434555
mozilla-foundation_common_voice_8_0_es_validation_eval_results.txt ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ WER: 0.10670647680293982
2
+ CER: 0.0284079393233586
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 sha256:6aec613a008767c6bbd9ffa13109e9c965582aedbf6728b3f9ab578ab6371fb9
3
+ size 1262071281
run.sh ADDED
@@ -0,0 +1,35 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ python run_speech_recognition_ctc_bnb.py \
2
+ --dataset_name="mozilla-foundation/common_voice_8_0" \
3
+ --model_name_or_path="facebook/wav2vec2-large-xlsr-53" \
4
+ --dataset_config_name="es" \
5
+ --output_dir="./" \
6
+ --overwrite_output_dir \
7
+ --num_train_epochs="10" \
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
+ --length_column_name="input_length" \
14
+ --evaluation_strategy="steps" \
15
+ --text_column_name="sentence" \
16
+ --chars_to_ignore , ? . ! \- \; \: \" “ % ‘ ” � — ’ … – \
17
+ --save_steps="500" \
18
+ --eval_steps="500" \
19
+ --logging_steps="100" \
20
+ --layerdrop="0.0" \
21
+ --activation_dropout="0.1" \
22
+ --save_total_limit="3" \
23
+ --freeze_feature_encoder \
24
+ --feat_proj_dropout="0.0" \
25
+ --mask_time_prob="0.75" \
26
+ --mask_time_length="10" \
27
+ --mask_feature_prob="0.25" \
28
+ --mask_feature_length="64" \
29
+ --hub_model_id "tomascufaro/xls-r-es-test" \
30
+ --gradient_checkpointing \
31
+ --use_auth_token \
32
+ --fp16 \
33
+ --group_by_length \
34
+ --do_train --do_eval \
35
+ --push_to_hub
run_speech_recognition_ctc.py ADDED
@@ -0,0 +1,737 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python
2
+ # coding=utf-8
3
+ # Copyright 2021 The HuggingFace Inc. team. All rights reserved.
4
+ #
5
+ # Licensed under the Apache License, Version 2.0 (the "License");
6
+ # you may not use this file except in compliance with the License.
7
+ # You may obtain a copy of the License at
8
+ #
9
+ # http://www.apache.org/licenses/LICENSE-2.0
10
+ #
11
+ # Unless required by applicable law or agreed to in writing, software
12
+ # distributed under the License is distributed on an "AS IS" BASIS,
13
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14
+ # See the License for the specific language governing permissions and
15
+
16
+ """ Fine-tuning a 🤗 Transformers CTC model for automatic speech recognition"""
17
+
18
+ import functools
19
+ import json
20
+ import logging
21
+ import os
22
+ import re
23
+ import sys
24
+ import warnings
25
+ from dataclasses import dataclass, field
26
+ from typing import Dict, List, Optional, Union
27
+
28
+ import datasets
29
+ import numpy as np
30
+ import torch
31
+ from datasets import DatasetDict, load_dataset, load_metric
32
+
33
+ import transformers
34
+ from transformers import (
35
+ AutoConfig,
36
+ AutoFeatureExtractor,
37
+ AutoModelForCTC,
38
+ AutoProcessor,
39
+ AutoTokenizer,
40
+ HfArgumentParser,
41
+ Trainer,
42
+ TrainingArguments,
43
+ Wav2Vec2Processor,
44
+ set_seed,
45
+ )
46
+ from transformers.trainer_utils import get_last_checkpoint, is_main_process
47
+ from transformers.utils import check_min_version
48
+ from transformers.utils.versions import require_version
49
+
50
+
51
+ # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
52
+ check_min_version("4.17.0.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+validation'"
150
+ },
151
+ )
152
+ eval_split_name: str = field(
153
+ default="test",
154
+ metadata={
155
+ "help": "The name of the evaluation 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 = (
438
+ f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
439
+ )
440
+ text_column_name = data_args.text_column_name
441
+
442
+ def remove_special_characters(batch):
443
+ if chars_to_ignore_regex is not None:
444
+ batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
445
+ else:
446
+ batch["target_text"] = batch[text_column_name].lower() + " "
447
+ return batch
448
+
449
+ with training_args.main_process_first(desc="dataset map special characters removal"):
450
+ raw_datasets = raw_datasets.map(
451
+ remove_special_characters,
452
+ remove_columns=[text_column_name],
453
+ desc="remove special characters from datasets",
454
+ )
455
+
456
+ # save special tokens for tokenizer
457
+ word_delimiter_token = data_args.word_delimiter_token
458
+ unk_token = data_args.unk_token
459
+ pad_token = data_args.pad_token
460
+
461
+ # 3. Next, let's load the config as we might need it to create
462
+ # the tokenizer
463
+ # load config
464
+ config = AutoConfig.from_pretrained(
465
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
466
+ )
467
+
468
+ # 4. Next, if no tokenizer file is defined,
469
+ # we create the vocabulary of the model by extracting all unique characters from
470
+ # the training and evaluation datasets
471
+ # We need to make sure that only first rank saves vocabulary
472
+ # make sure all processes wait until vocab is created
473
+ tokenizer_name_or_path = model_args.tokenizer_name_or_path
474
+ tokenizer_kwargs = {}
475
+ if tokenizer_name_or_path is None:
476
+ # save vocab in training output dir
477
+ tokenizer_name_or_path = training_args.output_dir
478
+
479
+ vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
480
+
481
+ with training_args.main_process_first():
482
+ if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
483
+ os.remove(vocab_file)
484
+
485
+ with training_args.main_process_first(desc="dataset map vocabulary creation"):
486
+ if not os.path.isfile(vocab_file):
487
+ os.makedirs(tokenizer_name_or_path, exist_ok=True)
488
+ vocab_dict = create_vocabulary_from_data(
489
+ raw_datasets,
490
+ word_delimiter_token=word_delimiter_token,
491
+ unk_token=unk_token,
492
+ pad_token=pad_token,
493
+ )
494
+
495
+ # save vocab dict to be loaded into tokenizer
496
+ with open(vocab_file, "w") as file:
497
+ json.dump(vocab_dict, file)
498
+
499
+ # if tokenizer has just been created
500
+ # it is defined by `tokenizer_class` if present in config else by `model_type`
501
+ tokenizer_kwargs = {
502
+ "config": config if config.tokenizer_class is not None else None,
503
+ "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
504
+ "unk_token": unk_token,
505
+ "pad_token": pad_token,
506
+ "word_delimiter_token": word_delimiter_token,
507
+ }
508
+
509
+ # 5. Now we can instantiate the feature extractor, tokenizer and model
510
+ # Note for distributed training, the .from_pretrained methods guarantee that only
511
+ # one local process can concurrently download model & vocab.
512
+
513
+ # load feature_extractor and tokenizer
514
+ tokenizer = AutoTokenizer.from_pretrained(
515
+ tokenizer_name_or_path,
516
+ use_auth_token=data_args.use_auth_token,
517
+ **tokenizer_kwargs,
518
+ )
519
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
520
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
521
+ )
522
+
523
+ # adapt config
524
+ config.update(
525
+ {
526
+ "feat_proj_dropout": model_args.feat_proj_dropout,
527
+ "attention_dropout": model_args.attention_dropout,
528
+ "hidden_dropout": model_args.hidden_dropout,
529
+ "final_dropout": model_args.final_dropout,
530
+ "mask_time_prob": model_args.mask_time_prob,
531
+ "mask_time_length": model_args.mask_time_length,
532
+ "mask_feature_prob": model_args.mask_feature_prob,
533
+ "mask_feature_length": model_args.mask_feature_length,
534
+ "gradient_checkpointing": training_args.gradient_checkpointing,
535
+ "layerdrop": model_args.layerdrop,
536
+ "ctc_loss_reduction": model_args.ctc_loss_reduction,
537
+ "pad_token_id": tokenizer.pad_token_id,
538
+ "vocab_size": len(tokenizer),
539
+ "activation_dropout": model_args.activation_dropout,
540
+ }
541
+ )
542
+
543
+ # create model
544
+ model = AutoModelForCTC.from_pretrained(
545
+ model_args.model_name_or_path,
546
+ cache_dir=model_args.cache_dir,
547
+ config=config,
548
+ use_auth_token=data_args.use_auth_token,
549
+ )
550
+
551
+ # freeze encoder
552
+ if model_args.freeze_feature_encoder:
553
+ model.freeze_feature_encoder()
554
+
555
+ # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
556
+ # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
557
+ # so that we just need to set the correct target sampling rate and normalize the input
558
+ # via the `feature_extractor`
559
+
560
+ # make sure that dataset decodes audio with correct sampling rate
561
+ dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
562
+ if dataset_sampling_rate != feature_extractor.sampling_rate:
563
+ raw_datasets = raw_datasets.cast_column(
564
+ data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
565
+ )
566
+
567
+ # derive max & min input length for sample rate & max duration
568
+ max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
569
+ min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
570
+ audio_column_name = data_args.audio_column_name
571
+ num_workers = data_args.preprocessing_num_workers
572
+
573
+ # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
574
+ phoneme_language = data_args.phoneme_language
575
+
576
+ # Preprocessing the datasets.
577
+ # We need to read the audio files as arrays and tokenize the targets.
578
+ def prepare_dataset(batch):
579
+ # load audio
580
+ sample = batch[audio_column_name]
581
+
582
+ inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
583
+ batch["input_values"] = inputs.input_values[0]
584
+ batch["input_length"] = len(batch["input_values"])
585
+
586
+ # encode targets
587
+ additional_kwargs = {}
588
+ if phoneme_language is not None:
589
+ additional_kwargs["phonemizer_lang"] = phoneme_language
590
+
591
+ batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
592
+ return batch
593
+
594
+ with training_args.main_process_first(desc="dataset map preprocessing"):
595
+ vectorized_datasets = raw_datasets.map(
596
+ prepare_dataset,
597
+ remove_columns=next(iter(raw_datasets.values())).column_names,
598
+ num_proc=num_workers,
599
+ desc="preprocess datasets",
600
+ )
601
+
602
+ def is_audio_in_length_range(length):
603
+ return length > min_input_length and length < max_input_length
604
+
605
+ # filter data that is shorter than min_input_length
606
+ vectorized_datasets = vectorized_datasets.filter(
607
+ is_audio_in_length_range,
608
+ num_proc=num_workers,
609
+ input_columns=["input_length"],
610
+ )
611
+
612
+ # 7. Next, we can prepare the training.
613
+ # Let's use word error rate (WER) as our evaluation metric,
614
+ # instantiate a data collator and the trainer
615
+
616
+ # Define evaluation metrics during training, *i.e.* word error rate, character error rate
617
+ eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
618
+
619
+ # for large datasets it is advised to run the preprocessing on a
620
+ # single machine first with ``args.preprocessing_only`` since there will mostly likely
621
+ # be a timeout when running the script in distributed mode.
622
+ # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
623
+ # cached dataset
624
+ if data_args.preprocessing_only:
625
+ logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
626
+ return
627
+
628
+ def compute_metrics(pred):
629
+ pred_logits = pred.predictions
630
+ pred_ids = np.argmax(pred_logits, axis=-1)
631
+
632
+ pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
633
+
634
+ pred_str = tokenizer.batch_decode(pred_ids)
635
+ # we do not want to group tokens when computing the metrics
636
+ label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
637
+
638
+ metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
639
+
640
+ return metrics
641
+
642
+ # Now save everything to be able to create a single processor later
643
+ if is_main_process(training_args.local_rank):
644
+ # save feature extractor, tokenizer and config
645
+ feature_extractor.save_pretrained(training_args.output_dir)
646
+ tokenizer.save_pretrained(training_args.output_dir)
647
+ config.save_pretrained(training_args.output_dir)
648
+
649
+ try:
650
+ processor = AutoProcessor.from_pretrained(training_args.output_dir)
651
+ except (OSError, KeyError):
652
+ warnings.warn(
653
+ "Loading a processor from a feature extractor config that does not"
654
+ " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
655
+ " attribute to your `preprocessor_config.json` file to suppress this warning: "
656
+ " `'processor_class': 'Wav2Vec2Processor'`",
657
+ FutureWarning,
658
+ )
659
+ processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
660
+
661
+ # Instantiate custom data collator
662
+ data_collator = DataCollatorCTCWithPadding(processor=processor)
663
+
664
+ # Initialize Trainer
665
+ trainer = Trainer(
666
+ model=model,
667
+ data_collator=data_collator,
668
+ args=training_args,
669
+ compute_metrics=compute_metrics,
670
+ train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
671
+ eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
672
+ tokenizer=feature_extractor,
673
+ )
674
+
675
+ # 8. Finally, we can start training
676
+
677
+ # Training
678
+ if training_args.do_train:
679
+
680
+ # use last checkpoint if exist
681
+ if last_checkpoint is not None:
682
+ checkpoint = last_checkpoint
683
+ elif os.path.isdir(model_args.model_name_or_path):
684
+ checkpoint = model_args.model_name_or_path
685
+ else:
686
+ checkpoint = None
687
+
688
+ train_result = trainer.train(resume_from_checkpoint=checkpoint)
689
+ trainer.save_model()
690
+
691
+ metrics = train_result.metrics
692
+ max_train_samples = (
693
+ data_args.max_train_samples
694
+ if data_args.max_train_samples is not None
695
+ else len(vectorized_datasets["train"])
696
+ )
697
+ metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
698
+
699
+ trainer.log_metrics("train", metrics)
700
+ trainer.save_metrics("train", metrics)
701
+ trainer.save_state()
702
+
703
+ # Evaluation
704
+ results = {}
705
+ if training_args.do_eval:
706
+ logger.info("*** Evaluate ***")
707
+ metrics = trainer.evaluate()
708
+ max_eval_samples = (
709
+ data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
710
+ )
711
+ metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
712
+
713
+ trainer.log_metrics("eval", metrics)
714
+ trainer.save_metrics("eval", metrics)
715
+
716
+ # Write model card and (optionally) push to hub
717
+ config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
718
+ kwargs = {
719
+ "finetuned_from": model_args.model_name_or_path,
720
+ "tasks": "speech-recognition",
721
+ "tags": ["automatic-speech-recognition", data_args.dataset_name],
722
+ "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
723
+ "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
724
+ }
725
+ if "common_voice" in data_args.dataset_name:
726
+ kwargs["language"] = config_name
727
+
728
+ if training_args.push_to_hub:
729
+ trainer.push_to_hub(**kwargs)
730
+ else:
731
+ trainer.create_model_card(**kwargs)
732
+
733
+ return results
734
+
735
+
736
+ if __name__ == "__main__":
737
+ main()
run_speech_recognition_ctc_bnb.py ADDED
@@ -0,0 +1,804 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python
2
+ # coding=utf-8
3
+ # Copyright 2021 The HuggingFace Inc. team. All rights reserved.
4
+ #
5
+ # Licensed under the Apache License, Version 2.0 (the "License");
6
+ # you may not use this file except in compliance with the License.
7
+ # You may obtain a copy of the License at
8
+ #
9
+ # http://www.apache.org/licenses/LICENSE-2.0
10
+ #
11
+ # Unless required by applicable law or agreed to in writing, software
12
+ # distributed under the License is distributed on an "AS IS" BASIS,
13
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14
+ # See the License for the specific language governing permissions and
15
+
16
+ """ Fine-tuning a 🤗 Transformers CTC model for automatic speech recognition"""
17
+
18
+ import functools
19
+ import json
20
+ import logging
21
+ import os
22
+ import re
23
+ import sys
24
+ import warnings
25
+ from dataclasses import dataclass, field
26
+ from typing import Dict, List, Optional, Union
27
+
28
+ import datasets
29
+ import numpy as np
30
+ import torch
31
+ from datasets import DatasetDict, load_dataset, load_metric
32
+
33
+ import bitsandbytes as bnb
34
+ import transformers
35
+ from transformers import (
36
+ AutoConfig,
37
+ AutoFeatureExtractor,
38
+ AutoModelForCTC,
39
+ AutoProcessor,
40
+ AutoTokenizer,
41
+ HfArgumentParser,
42
+ Trainer,
43
+ TrainingArguments,
44
+ Wav2Vec2Processor,
45
+ set_seed,
46
+ )
47
+ from transformers.trainer_pt_utils import get_parameter_names
48
+ from transformers.trainer_utils import get_last_checkpoint, is_main_process
49
+ from transformers.utils import check_min_version
50
+ from transformers.utils.versions import require_version
51
+
52
+
53
+ # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
54
+ check_min_version("4.16.0.dev0")
55
+
56
+ require_version("datasets>=1.13.3", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
57
+
58
+
59
+ logger = logging.getLogger(__name__)
60
+
61
+
62
+ def list_field(default=None, metadata=None):
63
+ return field(default_factory=lambda: default, metadata=metadata)
64
+
65
+
66
+ @dataclass
67
+ class ModelArguments:
68
+ """
69
+ Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
70
+ """
71
+
72
+ model_name_or_path: str = field(
73
+ metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
74
+ )
75
+ tokenizer_name_or_path: Optional[str] = field(
76
+ default=None,
77
+ metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
78
+ )
79
+ cache_dir: Optional[str] = field(
80
+ default=None,
81
+ metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
82
+ )
83
+ freeze_feature_encoder: bool = field(
84
+ default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
85
+ )
86
+ attention_dropout: float = field(
87
+ default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
88
+ )
89
+ activation_dropout: float = field(
90
+ default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
91
+ )
92
+ feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
93
+ hidden_dropout: float = field(
94
+ default=0.0,
95
+ metadata={
96
+ "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
97
+ },
98
+ )
99
+ final_dropout: float = field(
100
+ default=0.0,
101
+ metadata={"help": "The dropout probability for the final projection layer."},
102
+ )
103
+ mask_time_prob: float = field(
104
+ default=0.05,
105
+ metadata={
106
+ "help": "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
+ mask_time_length: int = field(
112
+ default=10,
113
+ metadata={"help": "Length of vector span to mask along the time axis."},
114
+ )
115
+ mask_feature_prob: float = field(
116
+ default=0.0,
117
+ metadata={
118
+ "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
119
+ "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
120
+ },
121
+ )
122
+ mask_feature_length: int = field(
123
+ default=10,
124
+ metadata={"help": "Length of vector span to mask along the feature axis."},
125
+ )
126
+ layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
127
+ ctc_loss_reduction: Optional[str] = field(
128
+ default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
129
+ )
130
+
131
+
132
+ @dataclass
133
+ class DataTrainingArguments:
134
+ """
135
+ Arguments pertaining to what data we are going to input our model for training and eval.
136
+
137
+ Using `HfArgumentParser` we can turn this class
138
+ into argparse arguments to be able to specify them on
139
+ the command line.
140
+ """
141
+
142
+ dataset_name: str = field(
143
+ metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
144
+ )
145
+ dataset_config_name: str = field(
146
+ default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
147
+ )
148
+ train_split_name: str = field(
149
+ default="train+validation",
150
+ metadata={
151
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
152
+ },
153
+ )
154
+ eval_split_name: str = field(
155
+ default="test",
156
+ metadata={
157
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
158
+ },
159
+ )
160
+ audio_column_name: str = field(
161
+ default="audio",
162
+ metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
163
+ )
164
+ text_column_name: str = field(
165
+ default="text",
166
+ metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
167
+ )
168
+ overwrite_cache: bool = field(
169
+ default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
170
+ )
171
+ preprocessing_num_workers: Optional[int] = field(
172
+ default=None,
173
+ metadata={"help": "The number of processes to use for the preprocessing."},
174
+ )
175
+ max_train_samples: Optional[int] = field(
176
+ default=None,
177
+ metadata={
178
+ "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
179
+ "value if set."
180
+ },
181
+ )
182
+ max_eval_samples: Optional[int] = field(
183
+ default=None,
184
+ metadata={
185
+ "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
186
+ "value if set."
187
+ },
188
+ )
189
+ chars_to_ignore: Optional[List[str]] = list_field(
190
+ default=None,
191
+ metadata={"help": "A list of characters to remove from the transcripts."},
192
+ )
193
+ eval_metrics: List[str] = list_field(
194
+ default=["wer"],
195
+ metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
196
+ )
197
+ max_duration_in_seconds: float = field(
198
+ default=20.0,
199
+ metadata={
200
+ "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
201
+ },
202
+ )
203
+ min_duration_in_seconds: float = field(
204
+ default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
205
+ )
206
+ preprocessing_only: bool = field(
207
+ default=False,
208
+ metadata={
209
+ "help": "Whether to only do data preprocessing and skip training. "
210
+ "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
211
+ "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
212
+ "so that the cached datasets can consequently be loaded in distributed training"
213
+ },
214
+ )
215
+ use_auth_token: bool = field(
216
+ default=False,
217
+ metadata={
218
+ "help": "If :obj:`True`, will use the token generated when running"
219
+ ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
220
+ },
221
+ )
222
+ unk_token: str = field(
223
+ default="[UNK]",
224
+ metadata={"help": "The unk token for the tokenizer"},
225
+ )
226
+ pad_token: str = field(
227
+ default="[PAD]",
228
+ metadata={"help": "The padding token for the tokenizer"},
229
+ )
230
+ word_delimiter_token: str = field(
231
+ default="|",
232
+ metadata={"help": "The word delimiter token for the tokenizer"},
233
+ )
234
+ phoneme_language: Optional[str] = field(
235
+ default=None,
236
+ metadata={
237
+ "help": "The target language that should be used be"
238
+ " passed to the tokenizer for tokenization. Note that"
239
+ " this is only relevant if the model classifies the"
240
+ " input audio to a sequence of phoneme sequences."
241
+ },
242
+ )
243
+
244
+
245
+ @dataclass
246
+ class DataCollatorCTCWithPadding:
247
+ """
248
+ Data collator that will dynamically pad the inputs received.
249
+ Args:
250
+ processor (:class:`~transformers.AutoProcessor`)
251
+ The processor used for proccessing the data.
252
+ padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
253
+ Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
254
+ among:
255
+ * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
256
+ sequence if provided).
257
+ * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
258
+ maximum acceptable input length for the model if that argument is not provided.
259
+ * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
260
+ different lengths).
261
+ max_length (:obj:`int`, `optional`):
262
+ Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
263
+ max_length_labels (:obj:`int`, `optional`):
264
+ Maximum length of the ``labels`` returned list and optionally padding length (see above).
265
+ pad_to_multiple_of (:obj:`int`, `optional`):
266
+ If set will pad the sequence to a multiple of the provided value.
267
+ This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
268
+ 7.5 (Volta).
269
+ """
270
+
271
+ processor: AutoProcessor
272
+ padding: Union[bool, str] = "longest"
273
+ pad_to_multiple_of: Optional[int] = None
274
+ pad_to_multiple_of_labels: Optional[int] = None
275
+
276
+ def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
277
+ # split inputs and labels since they have to be of different lenghts and need
278
+ # different padding methods
279
+ input_features = [{"input_values": feature["input_values"]} for feature in features]
280
+ label_features = [{"input_ids": feature["labels"]} for feature in features]
281
+
282
+ batch = self.processor.pad(
283
+ input_features,
284
+ padding=self.padding,
285
+ pad_to_multiple_of=self.pad_to_multiple_of,
286
+ return_tensors="pt",
287
+ )
288
+
289
+ with self.processor.as_target_processor():
290
+ labels_batch = self.processor.pad(
291
+ label_features,
292
+ padding=self.padding,
293
+ pad_to_multiple_of=self.pad_to_multiple_of_labels,
294
+ return_tensors="pt",
295
+ )
296
+
297
+ # replace padding with -100 to ignore loss correctly
298
+ labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
299
+
300
+ batch["labels"] = labels
301
+
302
+ return batch
303
+
304
+
305
+ def create_vocabulary_from_data(
306
+ datasets: DatasetDict,
307
+ word_delimiter_token: Optional[str] = None,
308
+ unk_token: Optional[str] = None,
309
+ pad_token: Optional[str] = None,
310
+ ):
311
+ # Given training and test labels create vocabulary
312
+ def extract_all_chars(batch):
313
+ all_text = " ".join(batch["target_text"])
314
+ vocab = list(set(all_text))
315
+ return {"vocab": [vocab], "all_text": [all_text]}
316
+
317
+ vocabs = datasets.map(
318
+ extract_all_chars,
319
+ batched=True,
320
+ batch_size=-1,
321
+ keep_in_memory=True,
322
+ remove_columns=datasets["train"].column_names,
323
+ )
324
+
325
+ # take union of all unique characters in each dataset
326
+ vocab_set = functools.reduce(
327
+ lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
328
+ )
329
+
330
+ vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
331
+
332
+ # replace white space with delimiter token
333
+ if word_delimiter_token is not None:
334
+ vocab_dict[word_delimiter_token] = vocab_dict[" "]
335
+ del vocab_dict[" "]
336
+
337
+ # add unk and pad token
338
+ if unk_token is not None:
339
+ vocab_dict[unk_token] = len(vocab_dict)
340
+
341
+ if pad_token is not None:
342
+ vocab_dict[pad_token] = len(vocab_dict)
343
+
344
+ return vocab_dict
345
+
346
+
347
+ def main():
348
+ # See all possible arguments in src/transformers/training_args.py
349
+ # or by passing the --help flag to this script.
350
+ # We now keep distinct sets of args, for a cleaner separation of concerns.
351
+
352
+ parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
353
+ if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
354
+ # If we pass only one argument to the script and it's the path to a json file,
355
+ # let's parse it to get our arguments.
356
+ model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
357
+ else:
358
+ model_args, data_args, training_args = parser.parse_args_into_dataclasses()
359
+
360
+ # Detecting last checkpoint.
361
+ last_checkpoint = None
362
+ if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
363
+ last_checkpoint = get_last_checkpoint(training_args.output_dir)
364
+ if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
365
+ raise ValueError(
366
+ f"Output directory ({training_args.output_dir}) already exists and is not empty. "
367
+ "Use --overwrite_output_dir to overcome."
368
+ )
369
+ elif last_checkpoint is not None:
370
+ logger.info(
371
+ f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
372
+ "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
373
+ )
374
+
375
+ # Setup logging
376
+ logging.basicConfig(
377
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
378
+ datefmt="%m/%d/%Y %H:%M:%S",
379
+ handlers=[logging.StreamHandler(sys.stdout)],
380
+ )
381
+ logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
382
+
383
+ # Log on each process the small summary:
384
+ logger.warning(
385
+ f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
386
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
387
+ )
388
+ # Set the verbosity to info of the Transformers logger (on main process only):
389
+ if is_main_process(training_args.local_rank):
390
+ transformers.utils.logging.set_verbosity_info()
391
+ logger.info("Training/evaluation parameters %s", training_args)
392
+
393
+ # Set seed before initializing model.
394
+ set_seed(training_args.seed)
395
+
396
+ # 1. First, let's load the dataset
397
+ raw_datasets = DatasetDict()
398
+
399
+ if training_args.do_train:
400
+ raw_datasets["train"] = load_dataset(
401
+ data_args.dataset_name,
402
+ data_args.dataset_config_name,
403
+ split=data_args.train_split_name,
404
+ use_auth_token=True,
405
+ )
406
+
407
+ if data_args.audio_column_name not in raw_datasets["train"].column_names:
408
+ raise ValueError(
409
+ f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
410
+ "Make sure to set `--audio_column_name` to the correct audio column - one of "
411
+ f"{', '.join(raw_datasets['train'].column_names)}."
412
+ )
413
+
414
+ if data_args.text_column_name not in raw_datasets["train"].column_names:
415
+ raise ValueError(
416
+ f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
417
+ "Make sure to set `--text_column_name` to the correct text column - one of "
418
+ f"{', '.join(raw_datasets['train'].column_names)}."
419
+ )
420
+
421
+ if data_args.max_train_samples is not None:
422
+ raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
423
+
424
+ if training_args.do_eval:
425
+ raw_datasets["eval"] = load_dataset(
426
+ data_args.dataset_name,
427
+ data_args.dataset_config_name,
428
+ split=data_args.eval_split_name,
429
+ use_auth_token=True,
430
+ )
431
+
432
+ if data_args.max_eval_samples is not None:
433
+ raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
434
+
435
+ # 2. We remove some special characters from the datasets
436
+ # that make training complicated and do not help in transcribing the speech
437
+ # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
438
+ # that could be easily picked up by the model
439
+ chars_to_ignore_regex = (
440
+ f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
441
+ )
442
+ text_column_name = data_args.text_column_name
443
+
444
+ def remove_special_characters(batch):
445
+ if chars_to_ignore_regex is not None:
446
+ batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
447
+ else:
448
+ batch["target_text"] = batch[text_column_name].lower() + " "
449
+ return batch
450
+
451
+ with training_args.main_process_first(desc="dataset map special characters removal"):
452
+ raw_datasets = raw_datasets.map(
453
+ remove_special_characters,
454
+ remove_columns=[text_column_name],
455
+ desc="remove special characters from datasets",
456
+ )
457
+
458
+ def clean_batch(batch):
459
+ batch["target_text"] = re.sub("([^A-Za-zÀ-ú ])", '', batch["target_text"]).lower()
460
+ batch["target_text"]= re.sub("([ß|þ|ð|æ])",'',batch['target_text'])
461
+ return batch
462
+
463
+ with training_args.main_process_first(desc="dataset map clean batch removal"):
464
+ raw_datasets = raw_datasets.map(
465
+ clean_batch,
466
+ #remove_columns=[text_column_name],
467
+ desc="remove rare characters from datasets",
468
+ )
469
+
470
+ def homologate_accents(batch):
471
+ batch["target_text"]=re.sub("([â|ã|ä|å|à])","a",batch["target_text"])
472
+ batch["target_text"]=re.sub("([é|ê|ë])","e",batch["target_text"])
473
+ batch["target_text"]=re.sub("([ì|î|ï])","i",batch["target_text"])
474
+ batch["target_text"]=re.sub("([ö|õ|ô|ò|ø])","o",batch["target_text"])
475
+ batch["target_text"]=re.sub("ù","u",batch["target_text"])
476
+ batch["target_text"]=re.sub("ç","c",batch["target_text"])
477
+ return batch
478
+
479
+ with training_args.main_process_first(desc="dataset map homologate batch removal"):
480
+ raw_datasets = raw_datasets.map(
481
+ homologate_accents,
482
+ #remove_columns=[text_column_name],
483
+ desc="homologate accents characters from datasets",
484
+ )
485
+
486
+ set_characters = set()
487
+ for string in raw_datasets["train"]["target_text"]:
488
+ set_characters.update(string.lower())
489
+
490
+ vocab = [character for character in "aábcdeéfghiíjklmnñoópqrstuúüvwxyz·-"]
491
+
492
+ unwanted_chars = set_characters-set(vocab)-set([' '])
493
+
494
+ with training_args.main_process_first(desc="dataset filter non vocab chars"):
495
+ raw_datasets = raw_datasets.filter(
496
+ lambda example: not any((c in unwanted_chars) for c in example),
497
+ input_columns="target_text",
498
+ desc="remove examples with weird characters"
499
+ )
500
+
501
+ # save special tokens for tokenizer
502
+ word_delimiter_token = data_args.word_delimiter_token
503
+ unk_token = data_args.unk_token
504
+ pad_token = data_args.pad_token
505
+
506
+ # 3. Next, let's load the config as we might need it to create
507
+ # the tokenizer
508
+ # load config
509
+ config = AutoConfig.from_pretrained(
510
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
511
+ )
512
+
513
+ # 4. Next, if no tokenizer file is defined,
514
+ # we create the vocabulary of the model by extracting all unique characters from
515
+ # the training and evaluation datasets
516
+ # We need to make sure that only first rank saves vocabulary
517
+ # make sure all processes wait until vocab is created
518
+ tokenizer_name_or_path = model_args.tokenizer_name_or_path
519
+ tokenizer_kwargs = {}
520
+ if tokenizer_name_or_path is None:
521
+ # save vocab in training output dir
522
+ tokenizer_name_or_path = training_args.output_dir
523
+
524
+ vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
525
+
526
+ with training_args.main_process_first():
527
+ if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
528
+ os.remove(vocab_file)
529
+
530
+ with training_args.main_process_first(desc="dataset map vocabulary creation"):
531
+ if not os.path.isfile(vocab_file):
532
+ os.makedirs(tokenizer_name_or_path, exist_ok=True)
533
+ vocab_dict = create_vocabulary_from_data(
534
+ raw_datasets,
535
+ word_delimiter_token=word_delimiter_token,
536
+ unk_token=unk_token,
537
+ pad_token=pad_token,
538
+ )
539
+
540
+ # save vocab dict to be loaded into tokenizer
541
+ with open(vocab_file, "w") as file:
542
+ json.dump(vocab_dict, file)
543
+
544
+ # if tokenizer has just been created
545
+ # it is defined by `tokenizer_class` if present in config else by `model_type`
546
+ tokenizer_kwargs = {
547
+ "config": config if config.tokenizer_class is not None else None,
548
+ "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
549
+ "unk_token": unk_token,
550
+ "pad_token": pad_token,
551
+ "word_delimiter_token": word_delimiter_token,
552
+ }
553
+
554
+ # 5. Now we can instantiate the feature extractor, tokenizer and model
555
+ # Note for distributed training, the .from_pretrained methods guarantee that only
556
+ # one local process can concurrently download model & vocab.
557
+
558
+ # load feature_extractor and tokenizer
559
+ tokenizer = AutoTokenizer.from_pretrained(
560
+ tokenizer_name_or_path,
561
+ use_auth_token=data_args.use_auth_token,
562
+ **tokenizer_kwargs,
563
+ )
564
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
565
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
566
+ )
567
+
568
+ # adapt config
569
+ config.update(
570
+ {
571
+ "feat_proj_dropout": model_args.feat_proj_dropout,
572
+ "attention_dropout": model_args.attention_dropout,
573
+ "hidden_dropout": model_args.hidden_dropout,
574
+ "final_dropout": model_args.final_dropout,
575
+ "mask_time_prob": model_args.mask_time_prob,
576
+ "mask_time_length": model_args.mask_time_length,
577
+ "mask_feature_prob": model_args.mask_feature_prob,
578
+ "mask_feature_length": model_args.mask_feature_length,
579
+ "gradient_checkpointing": training_args.gradient_checkpointing,
580
+ "layerdrop": model_args.layerdrop,
581
+ "ctc_loss_reduction": model_args.ctc_loss_reduction,
582
+ "pad_token_id": tokenizer.pad_token_id,
583
+ "vocab_size": len(tokenizer),
584
+ "activation_dropout": model_args.activation_dropout,
585
+ }
586
+ )
587
+
588
+ # create model
589
+ model = AutoModelForCTC.from_pretrained(
590
+ model_args.model_name_or_path,
591
+ cache_dir=model_args.cache_dir,
592
+ config=config,
593
+ use_auth_token=data_args.use_auth_token,
594
+ )
595
+
596
+ # freeze encoder
597
+ if model_args.freeze_feature_encoder:
598
+ model.freeze_feature_encoder()
599
+
600
+ # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
601
+ # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
602
+ # so that we just need to set the correct target sampling rate and normalize the input
603
+ # via the `feature_extractor`
604
+
605
+ # make sure that dataset decodes audio with correct sampling rate
606
+ dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
607
+ if dataset_sampling_rate != feature_extractor.sampling_rate:
608
+ raw_datasets = raw_datasets.cast_column(
609
+ data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
610
+ )
611
+
612
+ # derive max & min input length for sample rate & max duration
613
+ max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
614
+ min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
615
+ audio_column_name = data_args.audio_column_name
616
+ num_workers = data_args.preprocessing_num_workers
617
+
618
+ # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
619
+ phoneme_language = data_args.phoneme_language
620
+
621
+ # Preprocessing the datasets.
622
+ # We need to read the audio files as arrays and tokenize the targets.
623
+ def prepare_dataset(batch):
624
+ # load audio
625
+ sample = batch[audio_column_name]
626
+
627
+ inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
628
+ batch["input_values"] = inputs.input_values[0]
629
+ batch["input_length"] = len(batch["input_values"])
630
+
631
+ # encode targets
632
+ additional_kwargs = {}
633
+ if phoneme_language is not None:
634
+ additional_kwargs["phonemizer_lang"] = phoneme_language
635
+
636
+ batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
637
+ return batch
638
+
639
+ with training_args.main_process_first(desc="dataset map preprocessing"):
640
+ vectorized_datasets = raw_datasets.map(
641
+ prepare_dataset,
642
+ remove_columns=next(iter(raw_datasets.values())).column_names,
643
+ num_proc=num_workers,
644
+ desc="preprocess datasets",
645
+ )
646
+
647
+ def is_audio_in_length_range(length):
648
+ return length > min_input_length and length < max_input_length
649
+
650
+ # filter data that is shorter than min_input_length
651
+ vectorized_datasets = vectorized_datasets.filter(
652
+ is_audio_in_length_range,
653
+ num_proc=num_workers,
654
+ input_columns=["input_length"],
655
+ )
656
+
657
+ # 7. Next, we can prepare the training.
658
+ # Let's use word error rate (WER) as our evaluation metric,
659
+ # instantiate a data collator and the trainer
660
+
661
+ # Define evaluation metrics during training, *i.e.* word error rate, character error rate
662
+ eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
663
+
664
+ # for large datasets it is advised to run the preprocessing on a
665
+ # single machine first with ``args.preprocessing_only`` since there will mostly likely
666
+ # be a timeout when running the script in distributed mode.
667
+ # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
668
+ # cached dataset
669
+ if data_args.preprocessing_only:
670
+ logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
671
+ return
672
+
673
+ def compute_metrics(pred):
674
+ pred_logits = pred.predictions
675
+ pred_ids = np.argmax(pred_logits, axis=-1)
676
+
677
+ pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
678
+
679
+ pred_str = tokenizer.batch_decode(pred_ids)
680
+ # we do not want to group tokens when computing the metrics
681
+ label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
682
+
683
+ metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
684
+
685
+ return metrics
686
+
687
+ # Now save everything to be able to create a single processor later
688
+ if is_main_process(training_args.local_rank):
689
+ # save feature extractor, tokenizer and config
690
+ feature_extractor.save_pretrained(training_args.output_dir)
691
+ tokenizer.save_pretrained(training_args.output_dir)
692
+ config.save_pretrained(training_args.output_dir)
693
+
694
+ try:
695
+ processor = AutoProcessor.from_pretrained(training_args.output_dir)
696
+ except (OSError, KeyError):
697
+ warnings.warn(
698
+ "Loading a processor from a feature extractor config that does not"
699
+ " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
700
+ " attribute to your `preprocessor_config.json` file to suppress this warning: "
701
+ " `'processor_class': 'Wav2Vec2Processor'`",
702
+ FutureWarning,
703
+ )
704
+ processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
705
+
706
+ # Instantiate custom data collator
707
+ data_collator = DataCollatorCTCWithPadding(processor=processor)
708
+
709
+ decay_parameters = get_parameter_names(model, [torch.nn.LayerNorm])
710
+ decay_parameters = [name for name in decay_parameters if "bias" not in name]
711
+ optimizer_grouped_parameters = [
712
+ {
713
+ "params": [p for n, p in model.named_parameters() if n in decay_parameters],
714
+ "weight_decay": training_args.weight_decay,
715
+ },
716
+ {
717
+ "params": [p for n, p in model.named_parameters() if n not in decay_parameters],
718
+ "weight_decay": 0.0,
719
+ },
720
+ ]
721
+ optimizer = bnb.optim.Adam8bit(
722
+ params=optimizer_grouped_parameters,
723
+ lr=training_args.learning_rate,
724
+ betas=(training_args.adam_beta1, training_args.adam_beta2),
725
+ eps=training_args.adam_epsilon,
726
+ )
727
+
728
+ optimizers = (optimizer, None)
729
+
730
+ # Initialize Trainer
731
+ trainer = Trainer(
732
+ model=model,
733
+ data_collator=data_collator,
734
+ args=training_args,
735
+ compute_metrics=compute_metrics,
736
+ train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
737
+ eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
738
+ tokenizer=feature_extractor,
739
+ optimizers=optimizers,
740
+ )
741
+
742
+ # 8. Finally, we can start training
743
+
744
+ # Training
745
+ if training_args.do_train:
746
+
747
+ # use last checkpoint if exist
748
+ if last_checkpoint is not None:
749
+ checkpoint = last_checkpoint
750
+ elif os.path.isdir(model_args.model_name_or_path):
751
+ checkpoint = model_args.model_name_or_path
752
+ else:
753
+ checkpoint = None
754
+
755
+ train_result = trainer.train(resume_from_checkpoint=checkpoint)
756
+ trainer.save_model()
757
+
758
+ metrics = train_result.metrics
759
+ max_train_samples = (
760
+ data_args.max_train_samples
761
+ if data_args.max_train_samples is not None
762
+ else len(vectorized_datasets["train"])
763
+ )
764
+ metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
765
+
766
+ trainer.log_metrics("train", metrics)
767
+ trainer.save_metrics("train", metrics)
768
+ trainer.save_state()
769
+
770
+ # Evaluation
771
+ results = {}
772
+ if training_args.do_eval:
773
+ logger.info("*** Evaluate ***")
774
+ metrics = trainer.evaluate()
775
+ max_eval_samples = (
776
+ data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
777
+ )
778
+ metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
779
+
780
+ trainer.log_metrics("eval", metrics)
781
+ trainer.save_metrics("eval", metrics)
782
+
783
+ # Write model card and (optionally) push to hub
784
+ config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
785
+ kwargs = {
786
+ "finetuned_from": model_args.model_name_or_path,
787
+ "tasks": "speech-recognition",
788
+ "tags": ["automatic-speech-recognition", data_args.dataset_name],
789
+ "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
790
+ "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
791
+ }
792
+ if "common_voice" in data_args.dataset_name:
793
+ kwargs["language"] = config_name
794
+
795
+ if training_args.push_to_hub:
796
+ trainer.push_to_hub(**kwargs)
797
+ else:
798
+ trainer.create_model_card(**kwargs)
799
+
800
+ return results
801
+
802
+
803
+ if __name__ == "__main__":
804
+ main()
special_tokens_map.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"bos_token": "<s>", "eos_token": "</s>", "unk_token": "[UNK]", "pad_token": "[PAD]", "additional_special_tokens": [{"content": "<s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}, {"content": "</s>", "single_word": false, "lstrip": false, "rstrip": false, "normalized": true}]}
tokenizer_config.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"unk_token": "[UNK]", "bos_token": "<s>", "eos_token": "</s>", "pad_token": "[PAD]", "do_lower_case": false, "word_delimiter_token": "|", "special_tokens_map_file": null, "tokenizer_file": null, "name_or_path": "./", "tokenizer_class": "Wav2Vec2CTCTokenizer"}
train_results.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "epoch": 10.0,
3
+ "train_loss": 0.8738376914307662,
4
+ "train_runtime": 268775.0566,
5
+ "train_samples": 229440,
6
+ "train_samples_per_second": 8.537,
7
+ "train_steps_per_second": 0.267
8
+ }
trainer_state.json ADDED
The diff for this file is too large to render. See raw diff
 
training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:61af16ecf3f3e38206ca59d2abf9c3872cfd4e4f38971380d8bed24a74daad7d
3
+ size 3055
vocab.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"a": 1, "b": 2, "c": 3, "d": 4, "e": 5, "f": 6, "g": 7, "h": 8, "i": 9, "j": 10, "k": 11, "l": 12, "m": 13, "n": 14, "o": 15, "p": 16, "q": 17, "r": 18, "s": 19, "t": 20, "u": 21, "v": 22, "w": 23, "x": 24, "y": 25, "z": 26, "á": 27, "í": 28, "ñ": 29, "ó": 30, "ú": 31, "|": 0, "[UNK]": 32, "[PAD]": 33}