cahya commited on
Commit
83453d2
1 Parent(s): 88eeb6f

add model files

Browse files
.gitignore ADDED
@@ -0,0 +1,2 @@
 
 
 
1
+ checkpoint-*/
2
+ wandb/
README.md CHANGED
@@ -1,3 +1,130 @@
1
- ---
2
- license: cc
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ language:
3
+ - tr
4
+ tags:
5
+ - automatic-speech-recognition
6
+ - mozilla-foundation/common_voice_8_0
7
+ - generated_from_trainer
8
+ datasets:
9
+ - common_voice
10
+ model-index:
11
+ - name: ''
12
+ results: []
13
+ ---
14
+
15
+ <!-- This model card has been generated automatically according to the information the Trainer had access to. You
16
+ should probably proofread and complete it, then remove this comment. -->
17
+
18
+ #
19
+
20
+ This model is a fine-tuned version of [./checkpoint-1000](https://huggingface.co/./checkpoint-1000) on the MOZILLA-FOUNDATION/COMMON_VOICE_8_0 - TR dataset.
21
+ It achieves the following results on the evaluation set:
22
+ - Loss: 0.3282
23
+ - Wer: 0.2836
24
+
25
+ ## Model description
26
+
27
+ More information needed
28
+
29
+ ## Intended uses & limitations
30
+
31
+ More information needed
32
+
33
+ ## Training and evaluation data
34
+
35
+ More information needed
36
+
37
+ ## Training procedure
38
+
39
+ ### Training hyperparameters
40
+
41
+ The following hyperparameters were used during training:
42
+ - learning_rate: 0.0003
43
+ - train_batch_size: 96
44
+ - eval_batch_size: 8
45
+ - seed: 42
46
+ - gradient_accumulation_steps: 2
47
+ - total_train_batch_size: 192
48
+ - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
49
+ - lr_scheduler_type: linear
50
+ - lr_scheduler_warmup_steps: 100
51
+ - num_epochs: 100.0
52
+ - mixed_precision_training: Native AMP
53
+
54
+ ### Training results
55
+
56
+ | Training Loss | Epoch | Step | Validation Loss | Wer |
57
+ |:-------------:|:-----:|:-----:|:---------------:|:------:|
58
+ | 1.0671 | 2.04 | 200 | 0.3079 | 0.2752 |
59
+ | 0.6433 | 4.08 | 400 | 0.2728 | 0.2848 |
60
+ | 0.5687 | 6.12 | 600 | 0.2882 | 0.3036 |
61
+ | 0.5355 | 8.16 | 800 | 0.2778 | 0.2920 |
62
+ | 0.5116 | 10.2 | 1000 | 0.2906 | 0.3014 |
63
+ | 0.5313 | 9.16 | 1200 | 0.2984 | 0.3273 |
64
+ | 0.4996 | 10.69 | 1400 | 0.3170 | 0.3344 |
65
+ | 0.4845 | 12.21 | 1600 | 0.3202 | 0.3634 |
66
+ | 0.5092 | 13.74 | 1800 | 0.3167 | 0.3373 |
67
+ | 0.4777 | 15.27 | 2000 | 0.3292 | 0.3386 |
68
+ | 0.4651 | 16.79 | 2200 | 0.3070 | 0.3427 |
69
+ | 0.461 | 18.32 | 2400 | 0.3149 | 0.3561 |
70
+ | 0.4481 | 19.85 | 2600 | 0.3292 | 0.3441 |
71
+ | 0.4479 | 21.37 | 2800 | 0.3142 | 0.3209 |
72
+ | 0.4305 | 22.9 | 3000 | 0.3525 | 0.3547 |
73
+ | 0.4254 | 24.43 | 3200 | 0.3414 | 0.3400 |
74
+ | 0.4066 | 25.95 | 3400 | 0.3118 | 0.3207 |
75
+ | 0.4043 | 27.48 | 3600 | 0.3418 | 0.3483 |
76
+ | 0.3985 | 29.01 | 3800 | 0.3254 | 0.3166 |
77
+ | 0.3982 | 30.53 | 4000 | 0.3306 | 0.3453 |
78
+ | 0.3929 | 32.06 | 4200 | 0.3262 | 0.3229 |
79
+ | 0.378 | 33.59 | 4400 | 0.3546 | 0.3336 |
80
+ | 0.4062 | 35.11 | 4600 | 0.3174 | 0.3457 |
81
+ | 0.3648 | 36.64 | 4800 | 0.3377 | 0.3357 |
82
+ | 0.3609 | 38.17 | 5000 | 0.3346 | 0.3520 |
83
+ | 0.3483 | 39.69 | 5200 | 0.3350 | 0.3526 |
84
+ | 0.3548 | 41.22 | 5400 | 0.3330 | 0.3406 |
85
+ | 0.3446 | 42.75 | 5600 | 0.3398 | 0.3372 |
86
+ | 0.3346 | 44.27 | 5800 | 0.3449 | 0.3288 |
87
+ | 0.3309 | 45.8 | 6000 | 0.3320 | 0.3144 |
88
+ | 0.326 | 47.33 | 6200 | 0.3400 | 0.3279 |
89
+ | 0.3189 | 48.85 | 6400 | 0.3400 | 0.3150 |
90
+ | 0.3165 | 50.38 | 6600 | 0.3359 | 0.2995 |
91
+ | 0.3132 | 51.91 | 6800 | 0.3343 | 0.3096 |
92
+ | 0.3092 | 53.44 | 7000 | 0.3224 | 0.3029 |
93
+ | 0.2995 | 54.96 | 7200 | 0.3205 | 0.2985 |
94
+ | 0.304 | 56.49 | 7400 | 0.3523 | 0.3034 |
95
+ | 0.2952 | 58.02 | 7600 | 0.3289 | 0.2934 |
96
+ | 0.2875 | 59.54 | 7800 | 0.3350 | 0.3008 |
97
+ | 0.2868 | 61.07 | 8000 | 0.3537 | 0.3227 |
98
+ | 0.2875 | 62.6 | 8200 | 0.3389 | 0.2970 |
99
+ | 0.2778 | 64.12 | 8400 | 0.3370 | 0.2960 |
100
+ | 0.2706 | 65.65 | 8600 | 0.3250 | 0.2802 |
101
+ | 0.2669 | 67.18 | 8800 | 0.3351 | 0.2903 |
102
+ | 0.2615 | 68.7 | 9000 | 0.3382 | 0.2989 |
103
+ | 0.2563 | 70.23 | 9200 | 0.3312 | 0.2975 |
104
+ | 0.2546 | 71.76 | 9400 | 0.3212 | 0.3003 |
105
+ | 0.2482 | 73.28 | 9600 | 0.3337 | 0.3091 |
106
+ | 0.2504 | 74.81 | 9800 | 0.3308 | 0.3110 |
107
+ | 0.2456 | 76.34 | 10000 | 0.3157 | 0.3118 |
108
+ | 0.2363 | 77.86 | 10200 | 0.3251 | 0.3144 |
109
+ | 0.2319 | 79.39 | 10400 | 0.3253 | 0.3038 |
110
+ | 0.2266 | 80.92 | 10600 | 0.3374 | 0.3038 |
111
+ | 0.2279 | 82.44 | 10800 | 0.3268 | 0.2964 |
112
+ | 0.2231 | 83.97 | 11000 | 0.3278 | 0.2950 |
113
+ | 0.2185 | 85.5 | 11200 | 0.3462 | 0.2981 |
114
+ | 0.2245 | 87.02 | 11400 | 0.3311 | 0.2895 |
115
+ | 0.223 | 88.55 | 11600 | 0.3325 | 0.2877 |
116
+ | 0.2121 | 90.08 | 11800 | 0.3337 | 0.2828 |
117
+ | 0.2126 | 91.6 | 12000 | 0.3325 | 0.2808 |
118
+ | 0.2027 | 93.13 | 12200 | 0.3277 | 0.2820 |
119
+ | 0.2058 | 94.66 | 12400 | 0.3308 | 0.2827 |
120
+ | 0.1991 | 96.18 | 12600 | 0.3279 | 0.2820 |
121
+ | 0.1991 | 97.71 | 12800 | 0.3300 | 0.2822 |
122
+ | 0.1986 | 99.24 | 13000 | 0.3285 | 0.2835 |
123
+
124
+
125
+ ### Framework versions
126
+
127
+ - Transformers 4.17.0.dev0
128
+ - Pytorch 1.10.2+cu102
129
+ - Datasets 1.18.3
130
+ - Tokenizers 0.11.0
all_results.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "epoch": 100.0,
3
+ "eval_loss": 0.32822880148887634,
4
+ "eval_runtime": 237.0221,
5
+ "eval_samples": 8339,
6
+ "eval_samples_per_second": 35.182,
7
+ "eval_steps_per_second": 4.4,
8
+ "eval_wer": 0.2835930339138405,
9
+ "train_loss": 0.29656382378731067,
10
+ "train_runtime": 73649.3567,
11
+ "train_samples": 25058,
12
+ "train_samples_per_second": 34.023,
13
+ "train_steps_per_second": 0.178
14
+ }
alphabet.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"labels": [" ", "-", "a", "b", "c", "d", "e", "f", "g", "h", "i", "j", "k", "l", "m", "n", "o", "p", "q", "r", "s", "t", "u", "v", "w", "x", "y", "z", "\u00e2", "\u00e7", "\u00eb", "\u00ee", "\u00f6", "\u00fc", "\u011f", "\u0131", "\u015f", "\u0307", "\u2047", "" ], "is_bpe": false}
config.json ADDED
@@ -0,0 +1,106 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "./checkpoint-1000",
3
+ "activation_dropout": 0.055,
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": 0.1,
12
+ "bos_token_id": 1,
13
+ "classifier_proj_size": 256,
14
+ "codevector_dim": 256,
15
+ "contrastive_logits_temperature": 0.1,
16
+ "conv_bias": false,
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": true,
46
+ "diversity_loss_weight": 0.1,
47
+ "do_stable_layer_norm": false,
48
+ "eos_token_id": 2,
49
+ "feat_extract_activation": "gelu",
50
+ "feat_extract_norm": "group",
51
+ "feat_proj_dropout": 0.04,
52
+ "feat_quantizer_dropout": 0.0,
53
+ "final_dropout": 0.0,
54
+ "hidden_act": "gelu",
55
+ "hidden_dropout": 0.047,
56
+ "hidden_size": 768,
57
+ "initializer_range": 0.02,
58
+ "intermediate_size": 3072,
59
+ "layer_norm_eps": 1e-05,
60
+ "layerdrop": 0.041,
61
+ "mask_feature_length": 64,
62
+ "mask_feature_min_masks": 0,
63
+ "mask_feature_prob": 0.25,
64
+ "mask_time_length": 10,
65
+ "mask_time_min_masks": 2,
66
+ "mask_time_prob": 0.4,
67
+ "model_type": "wav2vec2",
68
+ "num_adapter_layers": 3,
69
+ "num_attention_heads": 12,
70
+ "num_codevector_groups": 2,
71
+ "num_codevectors_per_group": 320,
72
+ "num_conv_pos_embedding_groups": 16,
73
+ "num_conv_pos_embeddings": 128,
74
+ "num_feat_extract_layers": 7,
75
+ "num_hidden_layers": 12,
76
+ "num_negatives": 100,
77
+ "output_hidden_size": 768,
78
+ "pad_token_id": 39,
79
+ "proj_codevector_dim": 256,
80
+ "tdnn_dilation": [
81
+ 1,
82
+ 2,
83
+ 3,
84
+ 1,
85
+ 1
86
+ ],
87
+ "tdnn_dim": [
88
+ 512,
89
+ 512,
90
+ 512,
91
+ 512,
92
+ 1500
93
+ ],
94
+ "tdnn_kernel": [
95
+ 5,
96
+ 3,
97
+ 3,
98
+ 1,
99
+ 1
100
+ ],
101
+ "torch_dtype": "float32",
102
+ "transformers_version": "4.17.0.dev0",
103
+ "use_weighted_layer_sum": false,
104
+ "vocab_size": 40,
105
+ "xvector_output_dim": 512
106
+ }
eval.py ADDED
@@ -0,0 +1,138 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ import argparse
3
+ import re
4
+ from typing import Dict
5
+
6
+ import torch
7
+ from datasets import Audio, Dataset, load_dataset, load_metric
8
+
9
+ from transformers import AutoFeatureExtractor, pipeline
10
+
11
+
12
+ def log_results(result: Dataset, args: Dict[str, str]):
13
+ """DO NOT CHANGE. This function computes and logs the result metrics."""
14
+
15
+ log_outputs = args.log_outputs
16
+ dataset_id = "_".join(args.dataset.split("/") + [args.config, args.split])
17
+
18
+ # load metric
19
+ wer = load_metric("wer")
20
+ cer = load_metric("cer")
21
+
22
+ # compute metrics
23
+ wer_result = wer.compute(references=result["target"], predictions=result["prediction"])
24
+ cer_result = cer.compute(references=result["target"], predictions=result["prediction"])
25
+
26
+ # print & log results
27
+ result_str = f"WER: {wer_result}\n" f"CER: {cer_result}"
28
+ print(result_str)
29
+
30
+ with open(f"{dataset_id}_eval_results.txt", "w") as f:
31
+ f.write(result_str)
32
+
33
+ # log all results in text file. Possibly interesting for analysis
34
+ if log_outputs is not None:
35
+ pred_file = f"log_{dataset_id}_predictions.txt"
36
+ target_file = f"log_{dataset_id}_targets.txt"
37
+
38
+ with open(pred_file, "w") as p, open(target_file, "w") as t:
39
+
40
+ # mapping function to write output
41
+ def write_to_file(batch, i):
42
+ p.write(f"{i}" + "\n")
43
+ p.write(batch["prediction"] + "\n")
44
+ t.write(f"{i}" + "\n")
45
+ t.write(batch["target"] + "\n")
46
+
47
+ result.map(write_to_file, with_indices=True)
48
+
49
+
50
+ def normalize_text(text: str) -> str:
51
+ """DO ADAPT FOR YOUR USE CASE. this function normalizes the target text."""
52
+ chars_to_ignore_regex = '[,?.!-;:""%\'"\'\'`…’»«‘“”�éû]' # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
53
+
54
+ text = re.sub(chars_to_ignore_regex, "", text.lower())
55
+
56
+ # In addition, we can normalize the target text, e.g. removing new lines characters etc...
57
+ # note that order is important here!
58
+ token_sequences_to_ignore = ["\n\n", "\n", " ", " "]
59
+
60
+ for t in token_sequences_to_ignore:
61
+ text = " ".join(text.split(t))
62
+
63
+ return text
64
+
65
+
66
+ def main(args):
67
+ # load dataset
68
+ dataset = load_dataset(args.dataset, args.config, data_dir=args.data_dir, split=args.split, use_auth_token=True)
69
+
70
+ # for testing: only process the first two examples as a test
71
+ # dataset = dataset.select(range(10))
72
+
73
+ # load processor
74
+ feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
75
+ sampling_rate = feature_extractor.sampling_rate
76
+
77
+ # resample audio
78
+ dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
79
+
80
+ # load eval pipeline
81
+ if args.device is None:
82
+ args.device = 0 if torch.cuda.is_available() else -1
83
+ asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
84
+
85
+ # map function to decode audio
86
+ def map_to_pred(batch):
87
+ prediction = asr(
88
+ batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
89
+ )
90
+
91
+ batch["prediction"] = prediction["text"]
92
+ batch["target"] = normalize_text(batch["sentence"])
93
+ return batch
94
+
95
+ # run inference on all examples
96
+ result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
97
+
98
+ # compute and log_results
99
+ # do not change function below
100
+ log_results(result, args)
101
+
102
+
103
+ if __name__ == "__main__":
104
+ parser = argparse.ArgumentParser()
105
+
106
+ parser.add_argument(
107
+ "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
108
+ )
109
+ parser.add_argument(
110
+ "--dataset",
111
+ type=str,
112
+ required=True,
113
+ help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
114
+ )
115
+ parser.add_argument(
116
+ "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'` for Common Voice"
117
+ )
118
+ parser.add_argument("--data_dir", type=str, required=False, default=None,
119
+ help="The directory contains the dataset")
120
+ parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
121
+ parser.add_argument(
122
+ "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
123
+ )
124
+ parser.add_argument(
125
+ "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
126
+ )
127
+ parser.add_argument(
128
+ "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
129
+ )
130
+ parser.add_argument(
131
+ "--device",
132
+ type=int,
133
+ default=None,
134
+ help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
135
+ )
136
+ args = parser.parse_args()
137
+
138
+ main(args)
eval_results.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "epoch": 100.0,
3
+ "eval_loss": 0.32822880148887634,
4
+ "eval_runtime": 237.0221,
5
+ "eval_samples": 8339,
6
+ "eval_samples_per_second": 35.182,
7
+ "eval_steps_per_second": 4.4,
8
+ "eval_wer": 0.2835930339138405
9
+ }
language_model/5gram.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b750a609666264c2071c7fc4dec044155244557fab76815b229ffb6ee9172a0c
3
+ size 1040566788
language_model/attrs.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"alpha": 0.5, "beta": 1.5, "unk_score_offset": -10.0, "score_boundary": true}
language_model/unigrams.txt ADDED
The diff for this file is too large to render. See raw diff
 
preprocessor_config.json ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "do_normalize": true,
3
+ "feature_extractor_type": "Wav2Vec2FeatureExtractor",
4
+ "feature_size": 1,
5
+ "padding_side": "right",
6
+ "padding_value": 0.0,
7
+ "processor_class": "Wav2Vec2ProcessorWithLM",
8
+ "return_attention_mask": true,
9
+ "sampling_rate": 16000
10
+ }
pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5717124172f9f489cd469aa7632b2c999b8e577db0e9c8d585434d13d424400e
3
+ size 377694615
run.sh ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ export WANDB_ENTITY=cahya
2
+ export WANDB_LOG_MODEL=true
3
+ export WANDB_PROJECT=xlsr-turkish
4
+
5
+ python run_speech_recognition_ctc.py \
6
+ --dataset_name="mozilla-foundation/common_voice_8_0" \
7
+ --model_name_or_path="./checkpoint-1000" \
8
+ --dataset_config_name="tr" \
9
+ --output_dir="./" \
10
+ --overwrite_output_dir \
11
+ --num_train_epochs="100" \
12
+ --per_device_train_batch_size="96" \
13
+ --per_device_eval_batch_size="8" \
14
+ --gradient_accumulation_steps="2" \
15
+ --learning_rate="3e-4" \
16
+ --warmup_steps="100" \
17
+ --length_column_name="input_length" \
18
+ --evaluation_strategy="steps" \
19
+ --text_column_name="sentence" \
20
+ --save_steps="200" \
21
+ --eval_steps="200" \
22
+ --logging_steps="200" \
23
+ --layerdrop="0.041" \
24
+ --activation_dropout="0.055" \
25
+ --attention_dropout="0.1" \
26
+ --hidden_dropout="0.047" \
27
+ --save_total_limit="3" \
28
+ --freeze_feature_encoder \
29
+ --feat_proj_dropout="0.04" \
30
+ --mask_time_prob="0.4" \
31
+ --mask_time_length="10" \
32
+ --mask_feature_prob="0.25" \
33
+ --mask_feature_length="64" \
34
+ --gradient_checkpointing \
35
+ --use_auth_token \
36
+ --fp16=true \
37
+ --group_by_length \
38
+ --do_train=true \
39
+ --do_eval=true \
40
+ --push_to_hub=false \
41
+ --chars_to_ignore , ? . ! \; \: \"\" \% \' \" \' \' \` … \’ » « \‘ '“' '”' � é û \( \)
run_speech_recognition_ctc.py ADDED
@@ -0,0 +1,748 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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'"
150
+ },
151
+ )
152
+ eval_split_name: str = field(
153
+ default="test",
154
+ metadata={
155
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'test'"
156
+ },
157
+ )
158
+ audio_column_name: str = field(
159
+ default="audio",
160
+ metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
161
+ )
162
+ text_column_name: str = field(
163
+ default="text",
164
+ metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
165
+ )
166
+ overwrite_cache: bool = field(
167
+ default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
168
+ )
169
+ preprocessing_num_workers: Optional[int] = field(
170
+ default=None,
171
+ metadata={"help": "The number of processes to use for the preprocessing."},
172
+ )
173
+ max_train_samples: Optional[int] = field(
174
+ default=None,
175
+ metadata={
176
+ "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
177
+ "value if set."
178
+ },
179
+ )
180
+ max_eval_samples: Optional[int] = field(
181
+ default=None,
182
+ metadata={
183
+ "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
184
+ "value if set."
185
+ },
186
+ )
187
+ chars_to_ignore: Optional[List[str]] = list_field(
188
+ default=None,
189
+ metadata={"help": "A list of characters to remove from the transcripts."},
190
+ )
191
+ eval_metrics: List[str] = list_field(
192
+ default=["wer"],
193
+ metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
194
+ )
195
+ max_duration_in_seconds: float = field(
196
+ default=20.0,
197
+ metadata={
198
+ "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
199
+ },
200
+ )
201
+ min_duration_in_seconds: float = field(
202
+ default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
203
+ )
204
+ preprocessing_only: bool = field(
205
+ default=False,
206
+ metadata={
207
+ "help": "Whether to only do data preprocessing and skip training. "
208
+ "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
209
+ "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
210
+ "so that the cached datasets can consequently be loaded in distributed training"
211
+ },
212
+ )
213
+ use_auth_token: bool = field(
214
+ default=False,
215
+ metadata={
216
+ "help": "If :obj:`True`, will use the token generated when running"
217
+ ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
218
+ },
219
+ )
220
+ unk_token: str = field(
221
+ default="[UNK]",
222
+ metadata={"help": "The unk token for the tokenizer"},
223
+ )
224
+ pad_token: str = field(
225
+ default="[PAD]",
226
+ metadata={"help": "The padding token for the tokenizer"},
227
+ )
228
+ word_delimiter_token: str = field(
229
+ default="|",
230
+ metadata={"help": "The word delimiter token for the tokenizer"},
231
+ )
232
+ phoneme_language: Optional[str] = field(
233
+ default=None,
234
+ metadata={
235
+ "help": "The target language that should be used be"
236
+ " passed to the tokenizer for tokenization. Note that"
237
+ " this is only relevant if the model classifies the"
238
+ " input audio to a sequence of phoneme sequences."
239
+ },
240
+ )
241
+
242
+
243
+ @dataclass
244
+ class DataCollatorCTCWithPadding:
245
+ """
246
+ Data collator that will dynamically pad the inputs received.
247
+ Args:
248
+ processor (:class:`~transformers.AutoProcessor`)
249
+ The processor used for proccessing the data.
250
+ padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
251
+ Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
252
+ among:
253
+ * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
254
+ sequence if provided).
255
+ * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
256
+ maximum acceptable input length for the model if that argument is not provided.
257
+ * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
258
+ different lengths).
259
+ max_length (:obj:`int`, `optional`):
260
+ Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
261
+ max_length_labels (:obj:`int`, `optional`):
262
+ Maximum length of the ``labels`` returned list and optionally padding length (see above).
263
+ pad_to_multiple_of (:obj:`int`, `optional`):
264
+ If set will pad the sequence to a multiple of the provided value.
265
+ This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
266
+ 7.5 (Volta).
267
+ """
268
+
269
+ processor: AutoProcessor
270
+ padding: Union[bool, str] = "longest"
271
+ pad_to_multiple_of: Optional[int] = None
272
+ pad_to_multiple_of_labels: Optional[int] = None
273
+
274
+ def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
275
+ # split inputs and labels since they have to be of different lenghts and need
276
+ # different padding methods
277
+ input_features = [{"input_values": feature["input_values"]} for feature in features]
278
+ label_features = [{"input_ids": feature["labels"]} for feature in features]
279
+
280
+ batch = self.processor.pad(
281
+ input_features,
282
+ padding=self.padding,
283
+ pad_to_multiple_of=self.pad_to_multiple_of,
284
+ return_tensors="pt",
285
+ )
286
+
287
+ with self.processor.as_target_processor():
288
+ labels_batch = self.processor.pad(
289
+ label_features,
290
+ padding=self.padding,
291
+ pad_to_multiple_of=self.pad_to_multiple_of_labels,
292
+ return_tensors="pt",
293
+ )
294
+
295
+ # replace padding with -100 to ignore loss correctly
296
+ labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
297
+
298
+ batch["labels"] = labels
299
+
300
+ return batch
301
+
302
+
303
+ def create_vocabulary_from_data(
304
+ datasets: DatasetDict,
305
+ word_delimiter_token: Optional[str] = None,
306
+ unk_token: Optional[str] = None,
307
+ pad_token: Optional[str] = None,
308
+ ):
309
+ # Given training and test labels create vocabulary
310
+ def extract_all_chars(batch):
311
+ all_text = " ".join(batch["target_text"])
312
+ vocab = list(set(all_text))
313
+ return {"vocab": [vocab], "all_text": [all_text]}
314
+
315
+ print(f"dataset: {datasets}")
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
+ print("training_args.do_train:", training_args.do_train)
362
+ last_checkpoint = None
363
+ if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
364
+ last_checkpoint = get_last_checkpoint(training_args.output_dir)
365
+ if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
366
+ raise ValueError(
367
+ f"Output directory ({training_args.output_dir}) already exists and is not empty. "
368
+ "Use --overwrite_output_dir to overcome."
369
+ )
370
+ elif last_checkpoint is not None:
371
+ logger.info(
372
+ f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
373
+ "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
374
+ )
375
+
376
+ # Setup logging
377
+ logging.basicConfig(
378
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
379
+ datefmt="%m/%d/%Y %H:%M:%S",
380
+ handlers=[logging.StreamHandler(sys.stdout)],
381
+ )
382
+ logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
383
+
384
+ # Log on each process the small summary:
385
+ logger.warning(
386
+ f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
387
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
388
+ )
389
+ # Set the verbosity to info of the Transformers logger (on main process only):
390
+ if is_main_process(training_args.local_rank):
391
+ transformers.utils.logging.set_verbosity_info()
392
+ logger.info("Training/evaluation parameters %s", training_args)
393
+
394
+ # Set seed before initializing model.
395
+ set_seed(training_args.seed)
396
+
397
+ # 1. First, let's load the dataset
398
+ raw_datasets = DatasetDict()
399
+
400
+ print("do_train:", training_args.do_train)
401
+
402
+ if training_args.do_train:
403
+ print("load train")
404
+ raw_datasets["train"] = load_dataset(
405
+ data_args.dataset_name,
406
+ data_args.dataset_config_name,
407
+ split=data_args.train_split_name,
408
+ use_auth_token=data_args.use_auth_token,
409
+ )
410
+
411
+ if data_args.audio_column_name not in raw_datasets["train"].column_names:
412
+ raise ValueError(
413
+ f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
414
+ "Make sure to set `--audio_column_name` to the correct audio column - one of "
415
+ f"{', '.join(raw_datasets['train'].column_names)}."
416
+ )
417
+
418
+ if data_args.text_column_name not in raw_datasets["train"].column_names:
419
+ raise ValueError(
420
+ f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
421
+ "Make sure to set `--text_column_name` to the correct text column - one of "
422
+ f"{', '.join(raw_datasets['train'].column_names)}."
423
+ )
424
+
425
+ if data_args.max_train_samples is not None:
426
+ raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
427
+
428
+ if training_args.do_eval:
429
+ raw_datasets["eval"] = load_dataset(
430
+ data_args.dataset_name,
431
+ data_args.dataset_config_name,
432
+ split=data_args.eval_split_name,
433
+ use_auth_token=data_args.use_auth_token,
434
+ )
435
+
436
+ if data_args.max_eval_samples is not None:
437
+ raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
438
+
439
+ # 2. We remove some special characters from the datasets
440
+ # that make training complicated and do not help in transcribing the speech
441
+ # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
442
+ # that could be easily picked up by the model
443
+ chars_to_ignore_regex = (
444
+ f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
445
+ )
446
+ print(f"char ignored: {data_args.chars_to_ignore} {chars_to_ignore_regex}")
447
+ text_column_name = data_args.text_column_name
448
+
449
+ def remove_special_characters(batch):
450
+ if chars_to_ignore_regex is not None:
451
+ batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
452
+ else:
453
+ batch["target_text"] = batch[text_column_name].lower() + " "
454
+ return batch
455
+
456
+ with training_args.main_process_first(desc="dataset map special characters removal"):
457
+ raw_datasets = raw_datasets.map(
458
+ remove_special_characters,
459
+ remove_columns=[text_column_name],
460
+ desc="remove special characters from datasets",
461
+ )
462
+
463
+ # save special tokens for tokenizer
464
+ word_delimiter_token = data_args.word_delimiter_token
465
+ unk_token = data_args.unk_token
466
+ pad_token = data_args.pad_token
467
+
468
+ # 3. Next, let's load the config as we might need it to create
469
+ # the tokenizer
470
+ # load config
471
+ config = AutoConfig.from_pretrained(
472
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
473
+ )
474
+
475
+ print(f"config: {config}")
476
+
477
+ # 4. Next, if no tokenizer file is defined,
478
+ # we create the vocabulary of the model by extracting all unique characters from
479
+ # the training and evaluation datasets
480
+ # We need to make sure that only first rank saves vocabulary
481
+ # make sure all processes wait until vocab is created
482
+ tokenizer_name_or_path = model_args.tokenizer_name_or_path
483
+ tokenizer_kwargs = {}
484
+ if tokenizer_name_or_path is None:
485
+ # save vocab in training output dir
486
+ tokenizer_name_or_path = training_args.output_dir
487
+
488
+ vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
489
+
490
+ with training_args.main_process_first():
491
+ if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
492
+ os.remove(vocab_file)
493
+
494
+ with training_args.main_process_first(desc="dataset map vocabulary creation"):
495
+ if not os.path.isfile(vocab_file):
496
+ os.makedirs(tokenizer_name_or_path, exist_ok=True)
497
+ vocab_dict = create_vocabulary_from_data(
498
+ raw_datasets,
499
+ word_delimiter_token=word_delimiter_token,
500
+ unk_token=unk_token,
501
+ pad_token=pad_token,
502
+ )
503
+ print(f"vocab: {vocab_dict}")
504
+
505
+ # save vocab dict to be loaded into tokenizer
506
+ with open(vocab_file, "w") as file:
507
+ json.dump(vocab_dict, file)
508
+
509
+ # if tokenizer has just been created
510
+ # it is defined by `tokenizer_class` if present in config else by `model_type`
511
+ tokenizer_kwargs = {
512
+ "config": config if config.tokenizer_class is not None else None,
513
+ "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
514
+ "unk_token": unk_token,
515
+ "pad_token": pad_token,
516
+ "word_delimiter_token": word_delimiter_token,
517
+ }
518
+
519
+ # 5. Now we can instantiate the feature extractor, tokenizer and model
520
+ # Note for distributed training, the .from_pretrained methods guarantee that only
521
+ # one local process can concurrently download model & vocab.
522
+
523
+ # load feature_extractor and tokenizer
524
+ tokenizer = AutoTokenizer.from_pretrained(
525
+ tokenizer_name_or_path,
526
+ eos_token=None, bos_token=None,
527
+ use_auth_token=data_args.use_auth_token,
528
+ **tokenizer_kwargs,
529
+ )
530
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
531
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
532
+ )
533
+
534
+ # adapt config
535
+ config.update(
536
+ {
537
+ "feat_proj_dropout": model_args.feat_proj_dropout,
538
+ "attention_dropout": model_args.attention_dropout,
539
+ "hidden_dropout": model_args.hidden_dropout,
540
+ "final_dropout": model_args.final_dropout,
541
+ "mask_time_prob": model_args.mask_time_prob,
542
+ "mask_time_length": model_args.mask_time_length,
543
+ "mask_feature_prob": model_args.mask_feature_prob,
544
+ "mask_feature_length": model_args.mask_feature_length,
545
+ "gradient_checkpointing": training_args.gradient_checkpointing,
546
+ "layerdrop": model_args.layerdrop,
547
+ "ctc_loss_reduction": model_args.ctc_loss_reduction,
548
+ "pad_token_id": tokenizer.pad_token_id,
549
+ "vocab_size": len(tokenizer),
550
+ "activation_dropout": model_args.activation_dropout,
551
+ }
552
+ )
553
+
554
+ # create model
555
+ model = AutoModelForCTC.from_pretrained(
556
+ model_args.model_name_or_path,
557
+ cache_dir=model_args.cache_dir,
558
+ config=config,
559
+ use_auth_token=data_args.use_auth_token,
560
+ )
561
+
562
+ # freeze encoder
563
+ if model_args.freeze_feature_encoder:
564
+ model.freeze_feature_encoder()
565
+
566
+ # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
567
+ # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
568
+ # so that we just need to set the correct target sampling rate and normalize the input
569
+ # via the `feature_extractor`
570
+
571
+ # make sure that dataset decodes audio with correct sampling rate
572
+ dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
573
+ if dataset_sampling_rate != feature_extractor.sampling_rate:
574
+ raw_datasets = raw_datasets.cast_column(
575
+ data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
576
+ )
577
+
578
+ # derive max & min input length for sample rate & max duration
579
+ max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
580
+ min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
581
+ audio_column_name = data_args.audio_column_name
582
+ num_workers = data_args.preprocessing_num_workers
583
+
584
+ # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
585
+ phoneme_language = data_args.phoneme_language
586
+
587
+ # Preprocessing the datasets.
588
+ # We need to read the audio files as arrays and tokenize the targets.
589
+ def prepare_dataset(batch):
590
+ # load audio
591
+ sample = batch[audio_column_name]
592
+
593
+ inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
594
+ batch["input_values"] = inputs.input_values[0]
595
+ batch["input_length"] = len(batch["input_values"])
596
+
597
+ # encode targets
598
+ additional_kwargs = {}
599
+ if phoneme_language is not None:
600
+ additional_kwargs["phonemizer_lang"] = phoneme_language
601
+
602
+ batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
603
+ return batch
604
+
605
+ with training_args.main_process_first(desc="dataset map preprocessing"):
606
+ vectorized_datasets = raw_datasets.map(
607
+ prepare_dataset,
608
+ remove_columns=next(iter(raw_datasets.values())).column_names,
609
+ num_proc=num_workers,
610
+ desc="preprocess datasets",
611
+ )
612
+
613
+ def is_audio_in_length_range(length):
614
+ return length > min_input_length and length < max_input_length
615
+
616
+ # filter data that is shorter than min_input_length
617
+ vectorized_datasets = vectorized_datasets.filter(
618
+ is_audio_in_length_range,
619
+ num_proc=num_workers,
620
+ input_columns=["input_length"],
621
+ )
622
+
623
+ # 7. Next, we can prepare the training.
624
+ # Let's use word error rate (WER) as our evaluation metric,
625
+ # instantiate a data collator and the trainer
626
+
627
+ # Define evaluation metrics during training, *i.e.* word error rate, character error rate
628
+ eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
629
+
630
+ # for large datasets it is advised to run the preprocessing on a
631
+ # single machine first with ``args.preprocessing_only`` since there will mostly likely
632
+ # be a timeout when running the script in distributed mode.
633
+ # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
634
+ # cached dataset
635
+ if data_args.preprocessing_only:
636
+ logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
637
+ return
638
+
639
+ def compute_metrics(pred):
640
+ pred_logits = pred.predictions
641
+ pred_ids = np.argmax(pred_logits, axis=-1)
642
+
643
+ pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
644
+
645
+ pred_str = tokenizer.batch_decode(pred_ids)
646
+ # we do not want to group tokens when computing the metrics
647
+ label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
648
+
649
+ metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
650
+
651
+ return metrics
652
+
653
+ # Now save everything to be able to create a single processor later
654
+ if is_main_process(training_args.local_rank):
655
+ # save feature extractor, tokenizer and config
656
+ feature_extractor.save_pretrained(training_args.output_dir)
657
+ tokenizer.save_pretrained(training_args.output_dir)
658
+ config.save_pretrained(training_args.output_dir)
659
+
660
+ try:
661
+ processor = AutoProcessor.from_pretrained(training_args.output_dir)
662
+ except (OSError, KeyError):
663
+ warnings.warn(
664
+ "Loading a processor from a feature extractor config that does not"
665
+ " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
666
+ " attribute to your `preprocessor_config.json` file to suppress this warning: "
667
+ " `'processor_class': 'Wav2Vec2Processor'`",
668
+ FutureWarning,
669
+ )
670
+ processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
671
+
672
+ # Instantiate custom data collator
673
+ data_collator = DataCollatorCTCWithPadding(processor=processor)
674
+
675
+ # Initialize Trainer
676
+ trainer = Trainer(
677
+ model=model,
678
+ data_collator=data_collator,
679
+ args=training_args,
680
+ compute_metrics=compute_metrics,
681
+ train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
682
+ eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
683
+ tokenizer=feature_extractor,
684
+ )
685
+
686
+ # 8. Finally, we can start training
687
+
688
+ # Training
689
+ if training_args.do_train:
690
+
691
+ # use last checkpoint if exist
692
+ if last_checkpoint is not None:
693
+ checkpoint = last_checkpoint
694
+ elif os.path.isdir(model_args.model_name_or_path):
695
+ checkpoint = model_args.model_name_or_path
696
+ else:
697
+ checkpoint = None
698
+
699
+ train_result = trainer.train(resume_from_checkpoint=checkpoint)
700
+ trainer.save_model()
701
+
702
+ metrics = train_result.metrics
703
+ max_train_samples = (
704
+ data_args.max_train_samples
705
+ if data_args.max_train_samples is not None
706
+ else len(vectorized_datasets["train"])
707
+ )
708
+ metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
709
+
710
+ trainer.log_metrics("train", metrics)
711
+ trainer.save_metrics("train", metrics)
712
+ trainer.save_state()
713
+
714
+ # Evaluation
715
+ results = {}
716
+ if training_args.do_eval:
717
+ logger.info("*** Evaluate ***")
718
+ metrics = trainer.evaluate()
719
+ max_eval_samples = (
720
+ data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
721
+ )
722
+ metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
723
+
724
+ trainer.log_metrics("eval", metrics)
725
+ trainer.save_metrics("eval", metrics)
726
+
727
+ # Write model card and (optionally) push to hub
728
+ config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
729
+ kwargs = {
730
+ "finetuned_from": model_args.model_name_or_path,
731
+ "tasks": "speech-recognition",
732
+ "tags": ["automatic-speech-recognition", data_args.dataset_name],
733
+ "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
734
+ "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
735
+ }
736
+ if "common_voice" in data_args.dataset_name:
737
+ kwargs["language"] = config_name
738
+
739
+ if training_args.push_to_hub:
740
+ trainer.push_to_hub(**kwargs)
741
+ else:
742
+ trainer.create_model_card(**kwargs)
743
+
744
+ return results
745
+
746
+
747
+ if __name__ == "__main__":
748
+ main()
special_tokens_map.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"unk_token": "[UNK]", "pad_token": "[PAD]"}
tokenizer_config.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"unk_token": "[UNK]", "bos_token": null, "eos_token": null, "pad_token": "[PAD]", "do_lower_case": false, "word_delimiter_token": "|", "special_tokens_map_file": null, "tokenizer_file": null, "name_or_path": "./", "tokenizer_class": "Wav2Vec2CTCTokenizer"}
train_results.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "epoch": 100.0,
3
+ "train_loss": 0.29656382378731067,
4
+ "train_runtime": 73649.3567,
5
+ "train_samples": 25058,
6
+ "train_samples_per_second": 34.023,
7
+ "train_steps_per_second": 0.178
8
+ }
trainer_state.json ADDED
@@ -0,0 +1,1000 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_metric": null,
3
+ "best_model_checkpoint": null,
4
+ "epoch": 100.0,
5
+ "global_step": 13100,
6
+ "is_hyper_param_search": false,
7
+ "is_local_process_zero": true,
8
+ "is_world_process_zero": true,
9
+ "log_history": [
10
+ {
11
+ "epoch": 2.04,
12
+ "learning_rate": 0.0002969690721649484,
13
+ "loss": 1.0671,
14
+ "step": 200
15
+ },
16
+ {
17
+ "epoch": 2.04,
18
+ "eval_loss": 0.3079470694065094,
19
+ "eval_runtime": 237.7627,
20
+ "eval_samples_per_second": 35.073,
21
+ "eval_steps_per_second": 4.387,
22
+ "eval_wer": 0.2752062328139322,
23
+ "step": 200
24
+ },
25
+ {
26
+ "epoch": 4.08,
27
+ "learning_rate": 0.00029078350515463917,
28
+ "loss": 0.6433,
29
+ "step": 400
30
+ },
31
+ {
32
+ "epoch": 4.08,
33
+ "eval_loss": 0.27281925082206726,
34
+ "eval_runtime": 238.3892,
35
+ "eval_samples_per_second": 34.981,
36
+ "eval_steps_per_second": 4.375,
37
+ "eval_wer": 0.2847616865261228,
38
+ "step": 400
39
+ },
40
+ {
41
+ "epoch": 6.12,
42
+ "learning_rate": 0.0002845979381443299,
43
+ "loss": 0.5687,
44
+ "step": 600
45
+ },
46
+ {
47
+ "epoch": 6.12,
48
+ "eval_loss": 0.28817421197891235,
49
+ "eval_runtime": 238.0657,
50
+ "eval_samples_per_second": 35.028,
51
+ "eval_steps_per_second": 4.381,
52
+ "eval_wer": 0.3036205316223648,
53
+ "step": 600
54
+ },
55
+ {
56
+ "epoch": 8.16,
57
+ "learning_rate": 0.00027841237113402056,
58
+ "loss": 0.5355,
59
+ "step": 800
60
+ },
61
+ {
62
+ "epoch": 8.16,
63
+ "eval_loss": 0.27776163816452026,
64
+ "eval_runtime": 239.8353,
65
+ "eval_samples_per_second": 34.77,
66
+ "eval_steps_per_second": 4.349,
67
+ "eval_wer": 0.29200274977085244,
68
+ "step": 800
69
+ },
70
+ {
71
+ "epoch": 10.2,
72
+ "learning_rate": 0.00027222680412371134,
73
+ "loss": 0.5116,
74
+ "step": 1000
75
+ },
76
+ {
77
+ "epoch": 10.2,
78
+ "eval_loss": 0.2905969023704529,
79
+ "eval_runtime": 240.1935,
80
+ "eval_samples_per_second": 34.718,
81
+ "eval_steps_per_second": 4.342,
82
+ "eval_wer": 0.3013978001833181,
83
+ "step": 1000
84
+ },
85
+ {
86
+ "epoch": 9.16,
87
+ "learning_rate": 0.00027468461538461536,
88
+ "loss": 0.5313,
89
+ "step": 1200
90
+ },
91
+ {
92
+ "epoch": 9.16,
93
+ "eval_loss": 0.2984345555305481,
94
+ "eval_runtime": 234.3486,
95
+ "eval_samples_per_second": 35.584,
96
+ "eval_steps_per_second": 4.451,
97
+ "eval_wer": 0.327314390467461,
98
+ "step": 1200
99
+ },
100
+ {
101
+ "epoch": 10.69,
102
+ "learning_rate": 0.00027006923076923077,
103
+ "loss": 0.4996,
104
+ "step": 1400
105
+ },
106
+ {
107
+ "epoch": 10.69,
108
+ "eval_loss": 0.3169882297515869,
109
+ "eval_runtime": 237.8663,
110
+ "eval_samples_per_second": 35.058,
111
+ "eval_steps_per_second": 4.385,
112
+ "eval_wer": 0.3344179651695692,
113
+ "step": 1400
114
+ },
115
+ {
116
+ "epoch": 12.21,
117
+ "learning_rate": 0.0002654538461538461,
118
+ "loss": 0.4845,
119
+ "step": 1600
120
+ },
121
+ {
122
+ "epoch": 12.21,
123
+ "eval_loss": 0.32016345858573914,
124
+ "eval_runtime": 236.9291,
125
+ "eval_samples_per_second": 35.196,
126
+ "eval_steps_per_second": 4.402,
127
+ "eval_wer": 0.36338221814848765,
128
+ "step": 1600
129
+ },
130
+ {
131
+ "epoch": 13.74,
132
+ "learning_rate": 0.00026086153846153847,
133
+ "loss": 0.5092,
134
+ "step": 1800
135
+ },
136
+ {
137
+ "epoch": 13.74,
138
+ "eval_loss": 0.3166552186012268,
139
+ "eval_runtime": 236.2482,
140
+ "eval_samples_per_second": 35.298,
141
+ "eval_steps_per_second": 4.415,
142
+ "eval_wer": 0.3373052245646196,
143
+ "step": 1800
144
+ },
145
+ {
146
+ "epoch": 15.27,
147
+ "learning_rate": 0.0002562461538461538,
148
+ "loss": 0.4777,
149
+ "step": 2000
150
+ },
151
+ {
152
+ "epoch": 15.27,
153
+ "eval_loss": 0.32921522855758667,
154
+ "eval_runtime": 235.517,
155
+ "eval_samples_per_second": 35.407,
156
+ "eval_steps_per_second": 4.429,
157
+ "eval_wer": 0.3385655362053162,
158
+ "step": 2000
159
+ },
160
+ {
161
+ "epoch": 16.79,
162
+ "learning_rate": 0.0002516307692307692,
163
+ "loss": 0.4651,
164
+ "step": 2200
165
+ },
166
+ {
167
+ "epoch": 16.79,
168
+ "eval_loss": 0.3070097863674164,
169
+ "eval_runtime": 238.1366,
170
+ "eval_samples_per_second": 35.018,
171
+ "eval_steps_per_second": 4.38,
172
+ "eval_wer": 0.34271310724106324,
173
+ "step": 2200
174
+ },
175
+ {
176
+ "epoch": 18.32,
177
+ "learning_rate": 0.0002470153846153846,
178
+ "loss": 0.461,
179
+ "step": 2400
180
+ },
181
+ {
182
+ "epoch": 18.32,
183
+ "eval_loss": 0.3148922026157379,
184
+ "eval_runtime": 237.1783,
185
+ "eval_samples_per_second": 35.159,
186
+ "eval_steps_per_second": 4.398,
187
+ "eval_wer": 0.35609532538955085,
188
+ "step": 2400
189
+ },
190
+ {
191
+ "epoch": 19.85,
192
+ "learning_rate": 0.00024239999999999998,
193
+ "loss": 0.4481,
194
+ "step": 2600
195
+ },
196
+ {
197
+ "epoch": 19.85,
198
+ "eval_loss": 0.32919421792030334,
199
+ "eval_runtime": 236.0044,
200
+ "eval_samples_per_second": 35.334,
201
+ "eval_steps_per_second": 4.419,
202
+ "eval_wer": 0.34411090742438133,
203
+ "step": 2600
204
+ },
205
+ {
206
+ "epoch": 21.37,
207
+ "learning_rate": 0.00023778461538461536,
208
+ "loss": 0.4479,
209
+ "step": 2800
210
+ },
211
+ {
212
+ "epoch": 21.37,
213
+ "eval_loss": 0.3142247200012207,
214
+ "eval_runtime": 234.6363,
215
+ "eval_samples_per_second": 35.54,
216
+ "eval_steps_per_second": 4.445,
217
+ "eval_wer": 0.32089825847846015,
218
+ "step": 2800
219
+ },
220
+ {
221
+ "epoch": 22.9,
222
+ "learning_rate": 0.00023316923076923077,
223
+ "loss": 0.4305,
224
+ "step": 3000
225
+ },
226
+ {
227
+ "epoch": 22.9,
228
+ "eval_loss": 0.3525453507900238,
229
+ "eval_runtime": 236.821,
230
+ "eval_samples_per_second": 35.212,
231
+ "eval_steps_per_second": 4.404,
232
+ "eval_wer": 0.35467461044912924,
233
+ "step": 3000
234
+ },
235
+ {
236
+ "epoch": 24.43,
237
+ "learning_rate": 0.00022855384615384612,
238
+ "loss": 0.4254,
239
+ "step": 3200
240
+ },
241
+ {
242
+ "epoch": 24.43,
243
+ "eval_loss": 0.34136688709259033,
244
+ "eval_runtime": 235.4909,
245
+ "eval_samples_per_second": 35.411,
246
+ "eval_steps_per_second": 4.429,
247
+ "eval_wer": 0.3400091659028414,
248
+ "step": 3200
249
+ },
250
+ {
251
+ "epoch": 25.95,
252
+ "learning_rate": 0.00022393846153846153,
253
+ "loss": 0.4066,
254
+ "step": 3400
255
+ },
256
+ {
257
+ "epoch": 25.95,
258
+ "eval_loss": 0.3118491470813751,
259
+ "eval_runtime": 236.1268,
260
+ "eval_samples_per_second": 35.316,
261
+ "eval_steps_per_second": 4.417,
262
+ "eval_wer": 0.3207378551787351,
263
+ "step": 3400
264
+ },
265
+ {
266
+ "epoch": 27.48,
267
+ "learning_rate": 0.0002193230769230769,
268
+ "loss": 0.4043,
269
+ "step": 3600
270
+ },
271
+ {
272
+ "epoch": 27.48,
273
+ "eval_loss": 0.34181562066078186,
274
+ "eval_runtime": 235.2934,
275
+ "eval_samples_per_second": 35.441,
276
+ "eval_steps_per_second": 4.433,
277
+ "eval_wer": 0.3482584784601283,
278
+ "step": 3600
279
+ },
280
+ {
281
+ "epoch": 29.01,
282
+ "learning_rate": 0.0002147076923076923,
283
+ "loss": 0.3985,
284
+ "step": 3800
285
+ },
286
+ {
287
+ "epoch": 29.01,
288
+ "eval_loss": 0.32544735074043274,
289
+ "eval_runtime": 236.4374,
290
+ "eval_samples_per_second": 35.269,
291
+ "eval_steps_per_second": 4.411,
292
+ "eval_wer": 0.31663611365719524,
293
+ "step": 3800
294
+ },
295
+ {
296
+ "epoch": 30.53,
297
+ "learning_rate": 0.00021009230769230766,
298
+ "loss": 0.3982,
299
+ "step": 4000
300
+ },
301
+ {
302
+ "epoch": 30.53,
303
+ "eval_loss": 0.33056947588920593,
304
+ "eval_runtime": 239.8848,
305
+ "eval_samples_per_second": 34.763,
306
+ "eval_steps_per_second": 4.348,
307
+ "eval_wer": 0.3452795600366636,
308
+ "step": 4000
309
+ },
310
+ {
311
+ "epoch": 32.06,
312
+ "learning_rate": 0.00020547692307692307,
313
+ "loss": 0.3929,
314
+ "step": 4200
315
+ },
316
+ {
317
+ "epoch": 32.06,
318
+ "eval_loss": 0.3262433409690857,
319
+ "eval_runtime": 238.1544,
320
+ "eval_samples_per_second": 35.015,
321
+ "eval_steps_per_second": 4.38,
322
+ "eval_wer": 0.3228689275893675,
323
+ "step": 4200
324
+ },
325
+ {
326
+ "epoch": 33.59,
327
+ "learning_rate": 0.00020086153846153845,
328
+ "loss": 0.378,
329
+ "step": 4400
330
+ },
331
+ {
332
+ "epoch": 33.59,
333
+ "eval_loss": 0.3545904755592346,
334
+ "eval_runtime": 234.4177,
335
+ "eval_samples_per_second": 35.573,
336
+ "eval_steps_per_second": 4.449,
337
+ "eval_wer": 0.33357011915673696,
338
+ "step": 4400
339
+ },
340
+ {
341
+ "epoch": 35.11,
342
+ "learning_rate": 0.00019624615384615385,
343
+ "loss": 0.4062,
344
+ "step": 4600
345
+ },
346
+ {
347
+ "epoch": 35.11,
348
+ "eval_loss": 0.3174082636833191,
349
+ "eval_runtime": 235.3658,
350
+ "eval_samples_per_second": 35.43,
351
+ "eval_steps_per_second": 4.431,
352
+ "eval_wer": 0.34566911090742436,
353
+ "step": 4600
354
+ },
355
+ {
356
+ "epoch": 36.64,
357
+ "learning_rate": 0.0001916307692307692,
358
+ "loss": 0.3648,
359
+ "step": 4800
360
+ },
361
+ {
362
+ "epoch": 36.64,
363
+ "eval_loss": 0.3376729190349579,
364
+ "eval_runtime": 239.2202,
365
+ "eval_samples_per_second": 34.859,
366
+ "eval_steps_per_second": 4.36,
367
+ "eval_wer": 0.33572410632447297,
368
+ "step": 4800
369
+ },
370
+ {
371
+ "epoch": 38.17,
372
+ "learning_rate": 0.0001870153846153846,
373
+ "loss": 0.3609,
374
+ "step": 5000
375
+ },
376
+ {
377
+ "epoch": 38.17,
378
+ "eval_loss": 0.33455467224121094,
379
+ "eval_runtime": 236.9279,
380
+ "eval_samples_per_second": 35.196,
381
+ "eval_steps_per_second": 4.402,
382
+ "eval_wer": 0.351993583868011,
383
+ "step": 5000
384
+ },
385
+ {
386
+ "epoch": 39.69,
387
+ "learning_rate": 0.0001824,
388
+ "loss": 0.3483,
389
+ "step": 5200
390
+ },
391
+ {
392
+ "epoch": 39.69,
393
+ "eval_loss": 0.3349843919277191,
394
+ "eval_runtime": 235.432,
395
+ "eval_samples_per_second": 35.42,
396
+ "eval_steps_per_second": 4.43,
397
+ "eval_wer": 0.3525893675527039,
398
+ "step": 5200
399
+ },
400
+ {
401
+ "epoch": 41.22,
402
+ "learning_rate": 0.0001777846153846154,
403
+ "loss": 0.3548,
404
+ "step": 5400
405
+ },
406
+ {
407
+ "epoch": 41.22,
408
+ "eval_loss": 0.33302220702171326,
409
+ "eval_runtime": 237.3922,
410
+ "eval_samples_per_second": 35.128,
411
+ "eval_steps_per_second": 4.394,
412
+ "eval_wer": 0.34062786434463793,
413
+ "step": 5400
414
+ },
415
+ {
416
+ "epoch": 42.75,
417
+ "learning_rate": 0.00017316923076923075,
418
+ "loss": 0.3446,
419
+ "step": 5600
420
+ },
421
+ {
422
+ "epoch": 42.75,
423
+ "eval_loss": 0.3398281931877136,
424
+ "eval_runtime": 236.2326,
425
+ "eval_samples_per_second": 35.3,
426
+ "eval_steps_per_second": 4.415,
427
+ "eval_wer": 0.3372135655362053,
428
+ "step": 5600
429
+ },
430
+ {
431
+ "epoch": 44.27,
432
+ "learning_rate": 0.00016855384615384615,
433
+ "loss": 0.3346,
434
+ "step": 5800
435
+ },
436
+ {
437
+ "epoch": 44.27,
438
+ "eval_loss": 0.34486979246139526,
439
+ "eval_runtime": 236.4721,
440
+ "eval_samples_per_second": 35.264,
441
+ "eval_steps_per_second": 4.411,
442
+ "eval_wer": 0.3287809349220898,
443
+ "step": 5800
444
+ },
445
+ {
446
+ "epoch": 45.8,
447
+ "learning_rate": 0.00016393846153846153,
448
+ "loss": 0.3309,
449
+ "step": 6000
450
+ },
451
+ {
452
+ "epoch": 45.8,
453
+ "eval_loss": 0.3319507837295532,
454
+ "eval_runtime": 236.6263,
455
+ "eval_samples_per_second": 35.241,
456
+ "eval_steps_per_second": 4.408,
457
+ "eval_wer": 0.31439046746104493,
458
+ "step": 6000
459
+ },
460
+ {
461
+ "epoch": 47.33,
462
+ "learning_rate": 0.0001593230769230769,
463
+ "loss": 0.326,
464
+ "step": 6200
465
+ },
466
+ {
467
+ "epoch": 47.33,
468
+ "eval_loss": 0.3399747312068939,
469
+ "eval_runtime": 236.0765,
470
+ "eval_samples_per_second": 35.323,
471
+ "eval_steps_per_second": 4.418,
472
+ "eval_wer": 0.32786434463794684,
473
+ "step": 6200
474
+ },
475
+ {
476
+ "epoch": 48.85,
477
+ "learning_rate": 0.0001547076923076923,
478
+ "loss": 0.3189,
479
+ "step": 6400
480
+ },
481
+ {
482
+ "epoch": 48.85,
483
+ "eval_loss": 0.3399554491043091,
484
+ "eval_runtime": 236.2028,
485
+ "eval_samples_per_second": 35.304,
486
+ "eval_steps_per_second": 4.416,
487
+ "eval_wer": 0.31500916590284145,
488
+ "step": 6400
489
+ },
490
+ {
491
+ "epoch": 50.38,
492
+ "learning_rate": 0.0001500923076923077,
493
+ "loss": 0.3165,
494
+ "step": 6600
495
+ },
496
+ {
497
+ "epoch": 50.38,
498
+ "eval_loss": 0.3359447121620178,
499
+ "eval_runtime": 235.0525,
500
+ "eval_samples_per_second": 35.477,
501
+ "eval_steps_per_second": 4.437,
502
+ "eval_wer": 0.29945004582951423,
503
+ "step": 6600
504
+ },
505
+ {
506
+ "epoch": 51.91,
507
+ "learning_rate": 0.00014547692307692305,
508
+ "loss": 0.3132,
509
+ "step": 6800
510
+ },
511
+ {
512
+ "epoch": 51.91,
513
+ "eval_loss": 0.3342697024345398,
514
+ "eval_runtime": 235.4774,
515
+ "eval_samples_per_second": 35.413,
516
+ "eval_steps_per_second": 4.429,
517
+ "eval_wer": 0.3095554537121907,
518
+ "step": 6800
519
+ },
520
+ {
521
+ "epoch": 53.44,
522
+ "learning_rate": 0.00014086153846153845,
523
+ "loss": 0.3092,
524
+ "step": 7000
525
+ },
526
+ {
527
+ "epoch": 53.44,
528
+ "eval_loss": 0.3224042057991028,
529
+ "eval_runtime": 237.4291,
530
+ "eval_samples_per_second": 35.122,
531
+ "eval_steps_per_second": 4.393,
532
+ "eval_wer": 0.302910174152154,
533
+ "step": 7000
534
+ },
535
+ {
536
+ "epoch": 54.96,
537
+ "learning_rate": 0.00013624615384615383,
538
+ "loss": 0.2995,
539
+ "step": 7200
540
+ },
541
+ {
542
+ "epoch": 54.96,
543
+ "eval_loss": 0.3204595744609833,
544
+ "eval_runtime": 236.2646,
545
+ "eval_samples_per_second": 35.295,
546
+ "eval_steps_per_second": 4.415,
547
+ "eval_wer": 0.29851054078826766,
548
+ "step": 7200
549
+ },
550
+ {
551
+ "epoch": 56.49,
552
+ "learning_rate": 0.0001316307692307692,
553
+ "loss": 0.304,
554
+ "step": 7400
555
+ },
556
+ {
557
+ "epoch": 56.49,
558
+ "eval_loss": 0.35227909684181213,
559
+ "eval_runtime": 236.3256,
560
+ "eval_samples_per_second": 35.286,
561
+ "eval_steps_per_second": 4.413,
562
+ "eval_wer": 0.30339138405132904,
563
+ "step": 7400
564
+ },
565
+ {
566
+ "epoch": 58.02,
567
+ "learning_rate": 0.0001270153846153846,
568
+ "loss": 0.2952,
569
+ "step": 7600
570
+ },
571
+ {
572
+ "epoch": 58.02,
573
+ "eval_loss": 0.3288583755493164,
574
+ "eval_runtime": 238.857,
575
+ "eval_samples_per_second": 34.912,
576
+ "eval_steps_per_second": 4.367,
577
+ "eval_wer": 0.2934005499541705,
578
+ "step": 7600
579
+ },
580
+ {
581
+ "epoch": 59.54,
582
+ "learning_rate": 0.0001224,
583
+ "loss": 0.2875,
584
+ "step": 7800
585
+ },
586
+ {
587
+ "epoch": 59.54,
588
+ "eval_loss": 0.335005521774292,
589
+ "eval_runtime": 236.8029,
590
+ "eval_samples_per_second": 35.215,
591
+ "eval_steps_per_second": 4.405,
592
+ "eval_wer": 0.3008020164986251,
593
+ "step": 7800
594
+ },
595
+ {
596
+ "epoch": 61.07,
597
+ "learning_rate": 0.00011778461538461537,
598
+ "loss": 0.2868,
599
+ "step": 8000
600
+ },
601
+ {
602
+ "epoch": 61.07,
603
+ "eval_loss": 0.3537150025367737,
604
+ "eval_runtime": 236.5754,
605
+ "eval_samples_per_second": 35.249,
606
+ "eval_steps_per_second": 4.409,
607
+ "eval_wer": 0.3227314390467461,
608
+ "step": 8000
609
+ },
610
+ {
611
+ "epoch": 62.6,
612
+ "learning_rate": 0.00011316923076923076,
613
+ "loss": 0.2875,
614
+ "step": 8200
615
+ },
616
+ {
617
+ "epoch": 62.6,
618
+ "eval_loss": 0.3389277756214142,
619
+ "eval_runtime": 234.9228,
620
+ "eval_samples_per_second": 35.497,
621
+ "eval_steps_per_second": 4.44,
622
+ "eval_wer": 0.29704399633363887,
623
+ "step": 8200
624
+ },
625
+ {
626
+ "epoch": 64.12,
627
+ "learning_rate": 0.00010855384615384616,
628
+ "loss": 0.2778,
629
+ "step": 8400
630
+ },
631
+ {
632
+ "epoch": 64.12,
633
+ "eval_loss": 0.33703726530075073,
634
+ "eval_runtime": 234.6165,
635
+ "eval_samples_per_second": 35.543,
636
+ "eval_steps_per_second": 4.446,
637
+ "eval_wer": 0.29596700274977084,
638
+ "step": 8400
639
+ },
640
+ {
641
+ "epoch": 65.65,
642
+ "learning_rate": 0.00010393846153846154,
643
+ "loss": 0.2706,
644
+ "step": 8600
645
+ },
646
+ {
647
+ "epoch": 65.65,
648
+ "eval_loss": 0.32503727078437805,
649
+ "eval_runtime": 234.7831,
650
+ "eval_samples_per_second": 35.518,
651
+ "eval_steps_per_second": 4.442,
652
+ "eval_wer": 0.28015582034830433,
653
+ "step": 8600
654
+ },
655
+ {
656
+ "epoch": 67.18,
657
+ "learning_rate": 9.934615384615383e-05,
658
+ "loss": 0.2669,
659
+ "step": 8800
660
+ },
661
+ {
662
+ "epoch": 67.18,
663
+ "eval_loss": 0.335059255361557,
664
+ "eval_runtime": 234.788,
665
+ "eval_samples_per_second": 35.517,
666
+ "eval_steps_per_second": 4.442,
667
+ "eval_wer": 0.2902841429880843,
668
+ "step": 8800
669
+ },
670
+ {
671
+ "epoch": 68.7,
672
+ "learning_rate": 9.473076923076922e-05,
673
+ "loss": 0.2615,
674
+ "step": 9000
675
+ },
676
+ {
677
+ "epoch": 68.7,
678
+ "eval_loss": 0.3381515145301819,
679
+ "eval_runtime": 236.5391,
680
+ "eval_samples_per_second": 35.254,
681
+ "eval_steps_per_second": 4.409,
682
+ "eval_wer": 0.29885426214482125,
683
+ "step": 9000
684
+ },
685
+ {
686
+ "epoch": 70.23,
687
+ "learning_rate": 9.01153846153846e-05,
688
+ "loss": 0.2563,
689
+ "step": 9200
690
+ },
691
+ {
692
+ "epoch": 70.23,
693
+ "eval_loss": 0.3312215507030487,
694
+ "eval_runtime": 236.7043,
695
+ "eval_samples_per_second": 35.23,
696
+ "eval_steps_per_second": 4.406,
697
+ "eval_wer": 0.2974793767186068,
698
+ "step": 9200
699
+ },
700
+ {
701
+ "epoch": 71.76,
702
+ "learning_rate": 8.549999999999999e-05,
703
+ "loss": 0.2546,
704
+ "step": 9400
705
+ },
706
+ {
707
+ "epoch": 71.76,
708
+ "eval_loss": 0.3212486505508423,
709
+ "eval_runtime": 238.4241,
710
+ "eval_samples_per_second": 34.975,
711
+ "eval_steps_per_second": 4.375,
712
+ "eval_wer": 0.30034372135655363,
713
+ "step": 9400
714
+ },
715
+ {
716
+ "epoch": 73.28,
717
+ "learning_rate": 8.088461538461537e-05,
718
+ "loss": 0.2482,
719
+ "step": 9600
720
+ },
721
+ {
722
+ "epoch": 73.28,
723
+ "eval_loss": 0.3337170481681824,
724
+ "eval_runtime": 237.0835,
725
+ "eval_samples_per_second": 35.173,
726
+ "eval_steps_per_second": 4.399,
727
+ "eval_wer": 0.30907424381301557,
728
+ "step": 9600
729
+ },
730
+ {
731
+ "epoch": 74.81,
732
+ "learning_rate": 7.626923076923075e-05,
733
+ "loss": 0.2504,
734
+ "step": 9800
735
+ },
736
+ {
737
+ "epoch": 74.81,
738
+ "eval_loss": 0.33080631494522095,
739
+ "eval_runtime": 236.549,
740
+ "eval_samples_per_second": 35.253,
741
+ "eval_steps_per_second": 4.409,
742
+ "eval_wer": 0.3109761686526123,
743
+ "step": 9800
744
+ },
745
+ {
746
+ "epoch": 76.34,
747
+ "learning_rate": 7.165384615384615e-05,
748
+ "loss": 0.2456,
749
+ "step": 10000
750
+ },
751
+ {
752
+ "epoch": 76.34,
753
+ "eval_loss": 0.31574100255966187,
754
+ "eval_runtime": 235.6572,
755
+ "eval_samples_per_second": 35.386,
756
+ "eval_steps_per_second": 4.426,
757
+ "eval_wer": 0.3117781851512374,
758
+ "step": 10000
759
+ },
760
+ {
761
+ "epoch": 77.86,
762
+ "learning_rate": 6.703846153846153e-05,
763
+ "loss": 0.2363,
764
+ "step": 10200
765
+ },
766
+ {
767
+ "epoch": 77.86,
768
+ "eval_loss": 0.3251018524169922,
769
+ "eval_runtime": 236.3835,
770
+ "eval_samples_per_second": 35.277,
771
+ "eval_steps_per_second": 4.412,
772
+ "eval_wer": 0.31439046746104493,
773
+ "step": 10200
774
+ },
775
+ {
776
+ "epoch": 79.39,
777
+ "learning_rate": 6.242307692307691e-05,
778
+ "loss": 0.2319,
779
+ "step": 10400
780
+ },
781
+ {
782
+ "epoch": 79.39,
783
+ "eval_loss": 0.32527828216552734,
784
+ "eval_runtime": 236.1364,
785
+ "eval_samples_per_second": 35.314,
786
+ "eval_steps_per_second": 4.417,
787
+ "eval_wer": 0.30382676443629697,
788
+ "step": 10400
789
+ },
790
+ {
791
+ "epoch": 80.92,
792
+ "learning_rate": 5.7807692307692304e-05,
793
+ "loss": 0.2266,
794
+ "step": 10600
795
+ },
796
+ {
797
+ "epoch": 80.92,
798
+ "eval_loss": 0.3374153673648834,
799
+ "eval_runtime": 236.6995,
800
+ "eval_samples_per_second": 35.23,
801
+ "eval_steps_per_second": 4.406,
802
+ "eval_wer": 0.30382676443629697,
803
+ "step": 10600
804
+ },
805
+ {
806
+ "epoch": 82.44,
807
+ "learning_rate": 5.321538461538461e-05,
808
+ "loss": 0.2279,
809
+ "step": 10800
810
+ },
811
+ {
812
+ "epoch": 82.44,
813
+ "eval_loss": 0.32676786184310913,
814
+ "eval_runtime": 235.2934,
815
+ "eval_samples_per_second": 35.441,
816
+ "eval_steps_per_second": 4.433,
817
+ "eval_wer": 0.29640238313473877,
818
+ "step": 10800
819
+ },
820
+ {
821
+ "epoch": 83.97,
822
+ "learning_rate": 4.8599999999999995e-05,
823
+ "loss": 0.2231,
824
+ "step": 11000
825
+ },
826
+ {
827
+ "epoch": 83.97,
828
+ "eval_loss": 0.32775917649269104,
829
+ "eval_runtime": 237.2828,
830
+ "eval_samples_per_second": 35.144,
831
+ "eval_steps_per_second": 4.396,
832
+ "eval_wer": 0.29502749770852427,
833
+ "step": 11000
834
+ },
835
+ {
836
+ "epoch": 85.5,
837
+ "learning_rate": 4.398461538461538e-05,
838
+ "loss": 0.2185,
839
+ "step": 11200
840
+ },
841
+ {
842
+ "epoch": 85.5,
843
+ "eval_loss": 0.3461829721927643,
844
+ "eval_runtime": 234.8207,
845
+ "eval_samples_per_second": 35.512,
846
+ "eval_steps_per_second": 4.442,
847
+ "eval_wer": 0.29814390467461044,
848
+ "step": 11200
849
+ },
850
+ {
851
+ "epoch": 87.02,
852
+ "learning_rate": 3.9369230769230767e-05,
853
+ "loss": 0.2245,
854
+ "step": 11400
855
+ },
856
+ {
857
+ "epoch": 87.02,
858
+ "eval_loss": 0.3311368525028229,
859
+ "eval_runtime": 234.6868,
860
+ "eval_samples_per_second": 35.532,
861
+ "eval_steps_per_second": 4.444,
862
+ "eval_wer": 0.2894821264894592,
863
+ "step": 11400
864
+ },
865
+ {
866
+ "epoch": 88.55,
867
+ "learning_rate": 3.475384615384615e-05,
868
+ "loss": 0.223,
869
+ "step": 11600
870
+ },
871
+ {
872
+ "epoch": 88.55,
873
+ "eval_loss": 0.3325417935848236,
874
+ "eval_runtime": 234.9788,
875
+ "eval_samples_per_second": 35.488,
876
+ "eval_steps_per_second": 4.439,
877
+ "eval_wer": 0.2876947754353804,
878
+ "step": 11600
879
+ },
880
+ {
881
+ "epoch": 90.08,
882
+ "learning_rate": 3.0138461538461538e-05,
883
+ "loss": 0.2121,
884
+ "step": 11800
885
+ },
886
+ {
887
+ "epoch": 90.08,
888
+ "eval_loss": 0.333699494600296,
889
+ "eval_runtime": 237.3406,
890
+ "eval_samples_per_second": 35.135,
891
+ "eval_steps_per_second": 4.395,
892
+ "eval_wer": 0.282813932172319,
893
+ "step": 11800
894
+ },
895
+ {
896
+ "epoch": 91.6,
897
+ "learning_rate": 2.552307692307692e-05,
898
+ "loss": 0.2126,
899
+ "step": 12000
900
+ },
901
+ {
902
+ "epoch": 91.6,
903
+ "eval_loss": 0.3324645757675171,
904
+ "eval_runtime": 234.7938,
905
+ "eval_samples_per_second": 35.516,
906
+ "eval_steps_per_second": 4.442,
907
+ "eval_wer": 0.28077451879010085,
908
+ "step": 12000
909
+ },
910
+ {
911
+ "epoch": 93.13,
912
+ "learning_rate": 2.0907692307692305e-05,
913
+ "loss": 0.2027,
914
+ "step": 12200
915
+ },
916
+ {
917
+ "epoch": 93.13,
918
+ "eval_loss": 0.3276507258415222,
919
+ "eval_runtime": 235.1021,
920
+ "eval_samples_per_second": 35.47,
921
+ "eval_steps_per_second": 4.436,
922
+ "eval_wer": 0.2819660861594867,
923
+ "step": 12200
924
+ },
925
+ {
926
+ "epoch": 94.66,
927
+ "learning_rate": 1.629230769230769e-05,
928
+ "loss": 0.2058,
929
+ "step": 12400
930
+ },
931
+ {
932
+ "epoch": 94.66,
933
+ "eval_loss": 0.33078595995903015,
934
+ "eval_runtime": 236.8707,
935
+ "eval_samples_per_second": 35.205,
936
+ "eval_steps_per_second": 4.403,
937
+ "eval_wer": 0.2827451879010083,
938
+ "step": 12400
939
+ },
940
+ {
941
+ "epoch": 96.18,
942
+ "learning_rate": 1.1676923076923075e-05,
943
+ "loss": 0.1991,
944
+ "step": 12600
945
+ },
946
+ {
947
+ "epoch": 96.18,
948
+ "eval_loss": 0.3278521001338959,
949
+ "eval_runtime": 235.6753,
950
+ "eval_samples_per_second": 35.383,
951
+ "eval_steps_per_second": 4.426,
952
+ "eval_wer": 0.2820119156736939,
953
+ "step": 12600
954
+ },
955
+ {
956
+ "epoch": 97.71,
957
+ "learning_rate": 7.061538461538461e-06,
958
+ "loss": 0.1991,
959
+ "step": 12800
960
+ },
961
+ {
962
+ "epoch": 97.71,
963
+ "eval_loss": 0.3299693167209625,
964
+ "eval_runtime": 236.8407,
965
+ "eval_samples_per_second": 35.209,
966
+ "eval_steps_per_second": 4.404,
967
+ "eval_wer": 0.28221814848762605,
968
+ "step": 12800
969
+ },
970
+ {
971
+ "epoch": 99.24,
972
+ "learning_rate": 2.446153846153846e-06,
973
+ "loss": 0.1986,
974
+ "step": 13000
975
+ },
976
+ {
977
+ "epoch": 99.24,
978
+ "eval_loss": 0.3284846246242523,
979
+ "eval_runtime": 236.4429,
980
+ "eval_samples_per_second": 35.269,
981
+ "eval_steps_per_second": 4.411,
982
+ "eval_wer": 0.2834555453712191,
983
+ "step": 13000
984
+ },
985
+ {
986
+ "epoch": 100.0,
987
+ "step": 13100,
988
+ "total_flos": 8.36308166572502e+19,
989
+ "train_loss": 0.29656382378731067,
990
+ "train_runtime": 73649.3567,
991
+ "train_samples_per_second": 34.023,
992
+ "train_steps_per_second": 0.178
993
+ }
994
+ ],
995
+ "max_steps": 13100,
996
+ "num_train_epochs": 100,
997
+ "total_flos": 8.36308166572502e+19,
998
+ "trial_name": null,
999
+ "trial_params": null
1000
+ }
training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d1c646fe11ac72715f880fbe76c8724ae3a051b5d9a16b634b10373429a53a18
3
+ size 2991
vocab.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"-": 1, "a": 2, "b": 3, "c": 4, "d": 5, "e": 6, "f": 7, "g": 8, "h": 9, "i": 10, "j": 11, "k": 12, "l": 13, "m": 14, "n": 15, "o": 16, "p": 17, "q": 18, "r": 19, "s": 20, "t": 21, "u": 22, "v": 23, "w": 24, "x": 25, "y": 26, "z": 27, "â": 28, "ç": 29, "ë": 30, "î": 31, "ö": 32, "ü": 33, "ğ": 34, "ı": 35, "ş": 36, "̇": 37, "|": 0, "[UNK]": 38, "[PAD]": 39}