arampacha commited on
Commit
740882f
1 Parent(s): 8ee0ab8
.gitattributes CHANGED
@@ -25,3 +25,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
25
  *.zip filter=lfs diff=lfs merge=lfs -text
26
  *.zstandard filter=lfs diff=lfs merge=lfs -text
27
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
 
25
  *.zip filter=lfs diff=lfs merge=lfs -text
26
  *.zstandard filter=lfs diff=lfs merge=lfs -text
27
  *tfevents* filter=lfs diff=lfs merge=lfs -text
28
+ language_model/5gram.bin filter=lfs diff=lfs merge=lfs -text
29
+ pytorch_model.bin filter=lfs diff=lfs merge=lfs -text
README.md ADDED
@@ -0,0 +1,75 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ tags:
4
+ - automatic-speech-recognition
5
+ - /workspace/data/hy/noizy_student_1/
6
+ - generated_from_trainer
7
+ model-index:
8
+ - name: ''
9
+ results: []
10
+ ---
11
+
12
+ <!-- This model card has been generated automatically according to the information the Trainer had access to. You
13
+ should probably proofread and complete it, then remove this comment. -->
14
+
15
+ #
16
+
17
+ This model is a fine-tuned version of [facebook/wav2vec2-xls-r-1b](https://huggingface.co/facebook/wav2vec2-xls-r-1b) on the /WORKSPACE/DATA/HY/NOIZY_STUDENT_1/ - NA dataset.
18
+ It achieves the following results on the evaluation set:
19
+ - Loss: 0.3061
20
+ - Wer: 0.3899
21
+ - Cer: 0.0747
22
+
23
+ ## Model description
24
+
25
+ More information needed
26
+
27
+ ## Intended uses & limitations
28
+
29
+ More information needed
30
+
31
+ ## Training and evaluation data
32
+
33
+ More information needed
34
+
35
+ ## Training procedure
36
+
37
+ ### Training hyperparameters
38
+
39
+ The following hyperparameters were used during training:
40
+ - learning_rate: 8e-05
41
+ - train_batch_size: 16
42
+ - eval_batch_size: 64
43
+ - seed: 42
44
+ - gradient_accumulation_steps: 8
45
+ - total_train_batch_size: 128
46
+ - optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-08
47
+ - lr_scheduler_type: cosine
48
+ - lr_scheduler_warmup_ratio: 0.1
49
+ - training_steps: 1200
50
+ - mixed_precision_training: Native AMP
51
+
52
+ ### Training results
53
+
54
+ | Training Loss | Epoch | Step | Validation Loss | Wer | Cer |
55
+ |:-------------:|:------:|:----:|:---------------:|:------:|:------:|
56
+ | 5.195 | 9.09 | 100 | 3.1338 | 1.0 | 1.0 |
57
+ | 2.3769 | 18.18 | 200 | 0.4825 | 0.6616 | 0.1360 |
58
+ | 1.345 | 27.26 | 300 | 0.3781 | 0.5113 | 0.1057 |
59
+ | 1.2001 | 36.35 | 400 | 0.3571 | 0.4602 | 0.0931 |
60
+ | 1.0484 | 45.44 | 500 | 0.3121 | 0.4094 | 0.0776 |
61
+ | 0.926 | 54.53 | 600 | 0.3227 | 0.4094 | 0.0801 |
62
+ | 0.8854 | 63.62 | 700 | 0.3061 | 0.3899 | 0.0747 |
63
+ | 0.8054 | 72.7 | 800 | 0.3159 | 0.3891 | 0.0745 |
64
+ | 0.7442 | 81.79 | 900 | 0.3136 | 0.3802 | 0.0731 |
65
+ | 0.714 | 90.88 | 1000 | 0.3230 | 0.3716 | 0.0717 |
66
+ | 0.6641 | 99.97 | 1100 | 0.3193 | 0.3610 | 0.0693 |
67
+ | 0.6367 | 109.09 | 1200 | 0.3254 | 0.3587 | 0.0692 |
68
+
69
+
70
+ ### Framework versions
71
+
72
+ - Transformers 4.17.0.dev0
73
+ - Pytorch 1.10.2+cu102
74
+ - Datasets 1.18.2.dev0
75
+ - Tokenizers 0.11.0
added_tokens.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"<s>": 42, "</s>": 43}
all_results.json ADDED
@@ -0,0 +1,15 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "epoch": 109.09,
3
+ "eval_cer": 0.07465857359635812,
4
+ "eval_loss": 0.3061184883117676,
5
+ "eval_runtime": 15.8663,
6
+ "eval_samples": 335,
7
+ "eval_samples_per_second": 21.114,
8
+ "eval_steps_per_second": 0.378,
9
+ "eval_wer": 0.38992974238875877,
10
+ "train_loss": 1.3784224351247152,
11
+ "train_runtime": 12041.2604,
12
+ "train_samples": 1456,
13
+ "train_samples_per_second": 12.756,
14
+ "train_steps_per_second": 0.1
15
+ }
config.json ADDED
@@ -0,0 +1,107 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "facebook/wav2vec2-xls-r-1b",
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": 1024,
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": 1280,
58
+ "initializer_range": 0.02,
59
+ "intermediate_size": 5120,
60
+ "layer_norm_eps": 1e-05,
61
+ "layerdrop": 0.1,
62
+ "mask_feature_length": 64,
63
+ "mask_feature_min_masks": 0,
64
+ "mask_feature_prob": 0.25,
65
+ "mask_time_length": 10,
66
+ "mask_time_min_masks": 2,
67
+ "mask_time_prob": 0.75,
68
+ "model_type": "wav2vec2",
69
+ "num_adapter_layers": 3,
70
+ "num_attention_heads": 16,
71
+ "num_codevector_groups": 2,
72
+ "num_codevectors_per_group": 320,
73
+ "num_conv_pos_embedding_groups": 16,
74
+ "num_conv_pos_embeddings": 128,
75
+ "num_feat_extract_layers": 7,
76
+ "num_hidden_layers": 48,
77
+ "num_negatives": 100,
78
+ "output_hidden_size": 1280,
79
+ "pad_token_id": 41,
80
+ "proj_codevector_dim": 1024,
81
+ "tdnn_dilation": [
82
+ 1,
83
+ 2,
84
+ 3,
85
+ 1,
86
+ 1
87
+ ],
88
+ "tdnn_dim": [
89
+ 512,
90
+ 512,
91
+ 512,
92
+ 512,
93
+ 1500
94
+ ],
95
+ "tdnn_kernel": [
96
+ 5,
97
+ 3,
98
+ 3,
99
+ 1,
100
+ 1
101
+ ],
102
+ "torch_dtype": "float32",
103
+ "transformers_version": "4.17.0.dev0",
104
+ "use_weighted_layer_sum": false,
105
+ "vocab_size": 44,
106
+ "xvector_output_dim": 512
107
+ }
eval.py ADDED
@@ -0,0 +1,134 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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, Wav2Vec2ProcessorWithLM
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
+ """This function normalizes the target text."""
52
+
53
+ chars_to_ignore_regex = re.compile("[^\sաբգդեզէըթժիլխծկհձղճմյնշոչպջռսվտրցւփքօֆև]")
54
+ text = re.sub(chars_to_ignore_regex, "", text.lower())
55
+ text = " ".join(text.split())
56
+
57
+ return text
58
+
59
+
60
+ def main(args):
61
+ # load dataset
62
+ dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
63
+
64
+ # for testing: only process the first two examples as a test
65
+ # dataset = dataset.select(range(10))
66
+
67
+ # load processor
68
+ # feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
69
+ # sampling_rate = feature_extractor.sampling_rate
70
+ processor = Wav2Vec2ProcessorWithLM.from_pretrained(args.model_id)
71
+
72
+ # resample audio
73
+ dataset = dataset.cast_column("audio", Audio(sampling_rate=processor.feature_extractor.sampling_rate))
74
+
75
+ # load eval pipeline
76
+ if args.device is None:
77
+ args.device = 0 if torch.cuda.is_available() else -1
78
+ asr = pipeline(
79
+ "automatic-speech-recognition", model=args.model_id, device=args.device,
80
+ feature_extractor=processor.feature_extractor, decoder=processor.decoder
81
+ )
82
+
83
+ # map function to decode audio
84
+ def map_to_pred(batch):
85
+ prediction = asr(
86
+ batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
87
+ )
88
+
89
+ batch["prediction"] = prediction["text"]
90
+ batch["target"] = normalize_text(batch["sentence"])
91
+ return batch
92
+
93
+ # run inference on all examples
94
+ result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
95
+
96
+ # compute and log_results
97
+ # do not change function below
98
+ log_results(result, args)
99
+
100
+
101
+ if __name__ == "__main__":
102
+ parser = argparse.ArgumentParser()
103
+
104
+ parser.add_argument(
105
+ "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
106
+ )
107
+ parser.add_argument(
108
+ "--dataset",
109
+ type=str,
110
+ required=True,
111
+ help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
112
+ )
113
+ parser.add_argument(
114
+ "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'` for Common Voice"
115
+ )
116
+ parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
117
+ parser.add_argument(
118
+ "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
119
+ )
120
+ parser.add_argument(
121
+ "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
122
+ )
123
+ parser.add_argument(
124
+ "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
125
+ )
126
+ parser.add_argument(
127
+ "--device",
128
+ type=int,
129
+ default=None,
130
+ help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
131
+ )
132
+ args = parser.parse_args()
133
+
134
+ main(args)
eval_results.json ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "epoch": 109.09,
3
+ "eval_cer": 0.07465857359635812,
4
+ "eval_loss": 0.3061184883117676,
5
+ "eval_runtime": 15.8663,
6
+ "eval_samples": 335,
7
+ "eval_samples_per_second": 21.114,
8
+ "eval_steps_per_second": 0.378,
9
+ "eval_wer": 0.38992974238875877
10
+ }
language_model/5gram.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:39bcfd7d6afa5cc9c6286594ac121b9900d9e2f0934bd3da54b21373a303f064
3
+ size 879915787
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 +1,9 @@
 
 
 
 
 
 
 
 
 
 
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 sha256:32516ce95b5883aab22adf9ea26a65e31f5fecfa4cb664058ac379897433753e
3
+ size 3850538161
run.sh ADDED
@@ -0,0 +1,41 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ python run_speech_recognition_ctc.py \
2
+ --dataset_name="/workspace/data/hy/noizy_student_1/" \
3
+ --train_split_name train \
4
+ --model_name_or_path="facebook/wav2vec2-xls-r-1b" \
5
+ --output_dir="./" \
6
+ --overwrite_output_dir \
7
+ --max_steps 1200 \
8
+ --per_device_train_batch_size="16" \
9
+ --per_device_eval_batch_size="64" \
10
+ --gradient_accumulation_steps="8" \
11
+ --dataloader_num_workers 8 \
12
+ --learning_rate="8e-5" \
13
+ --lr_scheduler_type cosine \
14
+ --adam_beta2 0.98 \
15
+ --warmup_ratio 0.1 \
16
+ --evaluation_strategy="steps" \
17
+ --text_column_name="sentence" \
18
+ --chars_to_ignore \, \? \. \! \- \; \: \" \“ \% \‘ \” \� \' « » \( \) ։ ՝ ՞ ՛ ՚ \
19
+ --save_steps="100" \
20
+ --eval_steps="100" \
21
+ --logging_steps="100" \
22
+ --save_total_limit="2" \
23
+ --freeze_feature_encoder \
24
+ --layerdrop="0.1" \
25
+ --activation_dropout="0.1" \
26
+ --feat_proj_dropout="0.0" \
27
+ --mask_time_prob="0.75" \
28
+ --mask_time_length="10" \
29
+ --mask_feature_prob="0.25" \
30
+ --mask_feature_length="64" \
31
+ --gradient_checkpointing \
32
+ --use_auth_token \
33
+ --fp16 \
34
+ --group_by_length \
35
+ --do_train --do_eval \
36
+ --load_best_model_at_end \
37
+ --report_to all \
38
+ --run_name="xlsr-hy-ns-1b-1" \
39
+ --wandb_project="xlsr-hy" \
40
+ --seed 842 \
41
+ --bnb --tristage_sched
run_speech_recognition_ctc.py ADDED
@@ -0,0 +1,819 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 torch.optim.lr_scheduler import LambdaLR
32
+ from datasets import DatasetDict, load_dataset, load_metric, load_from_disk
33
+
34
+ import bitsandbytes as bnb
35
+ import transformers
36
+ from transformers import (
37
+ AutoConfig,
38
+ AutoFeatureExtractor,
39
+ AutoModelForCTC,
40
+ AutoProcessor,
41
+ AutoTokenizer,
42
+ HfArgumentParser,
43
+ Trainer,
44
+ TrainingArguments,
45
+ Wav2Vec2Processor,
46
+ set_seed,
47
+ )
48
+ from transformers.trainer_pt_utils import get_parameter_names
49
+ from transformers.trainer_utils import get_last_checkpoint, is_main_process
50
+ from transformers.utils import check_min_version
51
+ from transformers.utils.versions import require_version
52
+
53
+
54
+ # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
55
+ check_min_version("4.16.0.dev0")
56
+
57
+ require_version("datasets>=1.13.3", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
58
+
59
+
60
+ logger = logging.getLogger(__name__)
61
+
62
+
63
+ def list_field(default=None, metadata=None):
64
+ return field(default_factory=lambda: default, metadata=metadata)
65
+
66
+
67
+ @dataclass
68
+ class ModelArguments:
69
+ """
70
+ Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
71
+ """
72
+
73
+ model_name_or_path: str = field(
74
+ metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
75
+ )
76
+ tokenizer_name_or_path: Optional[str] = field(
77
+ default=None,
78
+ metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
79
+ )
80
+ cache_dir: Optional[str] = field(
81
+ default=None,
82
+ metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
83
+ )
84
+ freeze_feature_encoder: bool = field(
85
+ default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
86
+ )
87
+ attention_dropout: float = field(
88
+ default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
89
+ )
90
+ activation_dropout: float = field(
91
+ default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
92
+ )
93
+ feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
94
+ hidden_dropout: float = field(
95
+ default=0.0,
96
+ metadata={
97
+ "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
98
+ },
99
+ )
100
+ final_dropout: float = field(
101
+ default=0.0,
102
+ metadata={"help": "The dropout probability for the final projection layer."},
103
+ )
104
+ mask_time_prob: float = field(
105
+ default=0.05,
106
+ metadata={
107
+ "help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
108
+ "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
109
+ "vectors will be masked along the time axis."
110
+ },
111
+ )
112
+ mask_time_length: int = field(
113
+ default=10,
114
+ metadata={"help": "Length of vector span to mask along the time axis."},
115
+ )
116
+ mask_feature_prob: float = field(
117
+ default=0.0,
118
+ metadata={
119
+ "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
120
+ "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
121
+ },
122
+ )
123
+ mask_feature_length: int = field(
124
+ default=10,
125
+ metadata={"help": "Length of vector span to mask along the feature axis."},
126
+ )
127
+ layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
128
+ ctc_loss_reduction: Optional[str] = field(
129
+ default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
130
+ )
131
+
132
+
133
+ @dataclass
134
+ class DataTrainingArguments:
135
+ """
136
+ Arguments pertaining to what data we are going to input our model for training and eval.
137
+
138
+ Using `HfArgumentParser` we can turn this class
139
+ into argparse arguments to be able to specify them on
140
+ the command line.
141
+ """
142
+
143
+ dataset_name: str = field(
144
+ metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
145
+ )
146
+ dataset_config_name: str = field(
147
+ default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
148
+ )
149
+ train_split_name: str = field(
150
+ default="train+validation",
151
+ metadata={
152
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
153
+ },
154
+ )
155
+ eval_split_name: str = field(
156
+ default="test",
157
+ metadata={
158
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
159
+ },
160
+ )
161
+ audio_column_name: str = field(
162
+ default="audio",
163
+ metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
164
+ )
165
+ text_column_name: str = field(
166
+ default="text",
167
+ metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
168
+ )
169
+ overwrite_cache: bool = field(
170
+ default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
171
+ )
172
+ preprocessing_num_workers: Optional[int] = field(
173
+ default=None,
174
+ metadata={"help": "The number of processes to use for the preprocessing."},
175
+ )
176
+ max_train_samples: Optional[int] = field(
177
+ default=None,
178
+ metadata={
179
+ "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
180
+ "value if set."
181
+ },
182
+ )
183
+ max_eval_samples: Optional[int] = field(
184
+ default=None,
185
+ metadata={
186
+ "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
187
+ "value if set."
188
+ },
189
+ )
190
+ chars_to_ignore: Optional[List[str]] = list_field(
191
+ default=None,
192
+ metadata={"help": "A list of characters to remove from the transcripts."},
193
+ )
194
+ eval_metrics: List[str] = list_field(
195
+ default=["wer", "cer"],
196
+ metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
197
+ )
198
+ max_duration_in_seconds: float = field(
199
+ default=20.0,
200
+ metadata={
201
+ "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
202
+ },
203
+ )
204
+ min_duration_in_seconds: float = field(
205
+ default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
206
+ )
207
+ preprocessing_only: bool = field(
208
+ default=False,
209
+ metadata={
210
+ "help": "Whether to only do data preprocessing and skip training. "
211
+ "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
212
+ "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
213
+ "so that the cached datasets can consequently be loaded in distributed training"
214
+ },
215
+ )
216
+ use_auth_token: bool = field(
217
+ default=False,
218
+ metadata={
219
+ "help": "If :obj:`True`, will use the token generated when running"
220
+ ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
221
+ },
222
+ )
223
+ unk_token: str = field(
224
+ default="[UNK]",
225
+ metadata={"help": "The unk token for the tokenizer"},
226
+ )
227
+ pad_token: str = field(
228
+ default="[PAD]",
229
+ metadata={"help": "The padding token for the tokenizer"},
230
+ )
231
+ word_delimiter_token: str = field(
232
+ default="|",
233
+ metadata={"help": "The word delimiter token for the tokenizer"},
234
+ )
235
+ phoneme_language: Optional[str] = field(
236
+ default=None,
237
+ metadata={
238
+ "help": "The target language that should be used be"
239
+ " passed to the tokenizer for tokenization. Note that"
240
+ " this is only relevant if the model classifies the"
241
+ " input audio to a sequence of phoneme sequences."
242
+ },
243
+ )
244
+
245
+ @dataclass
246
+ class ExtraArguments:
247
+ "Additional training arguments"
248
+ bnb: bool = field(
249
+ default=False,
250
+ metadata = {"help":"If true uses 8bit Adam"}
251
+ )
252
+ tristage_sched: bool = field(
253
+ default=False,
254
+ metadata = {"help":"If true uses tristage LR scheduler (refer to XLS-R paper)"}
255
+ )
256
+ wandb_project: str = field(
257
+ default=None,
258
+ metadata = {"help":"Name of wandb project to log into"}
259
+ )
260
+
261
+
262
+ @dataclass
263
+ class DataCollatorCTCWithPadding:
264
+ """
265
+ Data collator that will dynamically pad the inputs received.
266
+ Args:
267
+ processor (:class:`~transformers.AutoProcessor`)
268
+ The processor used for proccessing the data.
269
+ padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
270
+ Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
271
+ among:
272
+ * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
273
+ sequence if provided).
274
+ * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
275
+ maximum acceptable input length for the model if that argument is not provided.
276
+ * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
277
+ different lengths).
278
+ max_length (:obj:`int`, `optional`):
279
+ Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
280
+ max_length_labels (:obj:`int`, `optional`):
281
+ Maximum length of the ``labels`` returned list and optionally padding length (see above).
282
+ pad_to_multiple_of (:obj:`int`, `optional`):
283
+ If set will pad the sequence to a multiple of the provided value.
284
+ This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
285
+ 7.5 (Volta).
286
+ """
287
+
288
+ processor: AutoProcessor
289
+ padding: Union[bool, str] = "longest"
290
+ pad_to_multiple_of: Optional[int] = None
291
+ pad_to_multiple_of_labels: Optional[int] = None
292
+
293
+ def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
294
+ # split inputs and labels since they have to be of different lenghts and need
295
+ # different padding methods
296
+ input_features = [{"input_values": feature["input_values"]} for feature in features]
297
+ label_features = [{"input_ids": feature["labels"]} for feature in features]
298
+
299
+ batch = self.processor.pad(
300
+ input_features,
301
+ padding=self.padding,
302
+ pad_to_multiple_of=self.pad_to_multiple_of,
303
+ return_tensors="pt",
304
+ )
305
+
306
+ with self.processor.as_target_processor():
307
+ labels_batch = self.processor.pad(
308
+ label_features,
309
+ padding=self.padding,
310
+ pad_to_multiple_of=self.pad_to_multiple_of_labels,
311
+ return_tensors="pt",
312
+ )
313
+
314
+ # replace padding with -100 to ignore loss correctly
315
+ labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
316
+
317
+ batch["labels"] = labels
318
+
319
+ return batch
320
+
321
+
322
+ def get_tri_stage_schedule(
323
+ optimizer, num_training_steps, ratios=[0.1, 0.4, 0.5], num_warmup_steps=None, num_hold_steps=None, start_ratio=0.01, end_ratio=0.05
324
+ ):
325
+ assert (num_warmup_steps is None) == (num_hold_steps is None)
326
+ if num_warmup_steps is None:
327
+ num_warmup_steps = int(ratios[0]*num_training_steps)
328
+ num_hold_steps = int(ratios[1]*num_training_steps)
329
+ start_decay_step = num_warmup_steps + num_hold_steps
330
+ a_w, b_w = (1-start_ratio)/num_warmup_steps, start_ratio
331
+ num_decay_steps = num_training_steps - start_decay_step
332
+ a_d, b_d = (end_ratio-1)/num_decay_steps, 1.
333
+
334
+ def lr_lambda(current_step):
335
+ if current_step < num_warmup_steps:
336
+ return a_w * float(current_step) + b_w
337
+ if current_step < start_decay_step:
338
+ return 1.
339
+ return max(end_ratio, a_d * float(current_step - start_decay_step) + b_d )
340
+
341
+ return LambdaLR(optimizer, lr_lambda)
342
+
343
+ def create_vocabulary_from_data(
344
+ datasets: DatasetDict,
345
+ word_delimiter_token: Optional[str] = None,
346
+ unk_token: Optional[str] = None,
347
+ pad_token: Optional[str] = None,
348
+ ):
349
+ # Given training and test labels create vocabulary
350
+ def extract_all_chars(batch):
351
+ all_text = " ".join(batch["target_text"])
352
+ vocab = list(set(all_text))
353
+ return {"vocab": [vocab], "all_text": [all_text]}
354
+
355
+ vocabs = datasets.map(
356
+ extract_all_chars,
357
+ batched=True,
358
+ batch_size=-1,
359
+ keep_in_memory=True,
360
+ remove_columns=datasets["train"].column_names,
361
+ )
362
+
363
+ # take union of all unique characters in each dataset
364
+ vocab_set = functools.reduce(
365
+ lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
366
+ )
367
+
368
+ vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
369
+
370
+ # replace white space with delimiter token
371
+ if word_delimiter_token is not None:
372
+ vocab_dict[word_delimiter_token] = vocab_dict[" "]
373
+ del vocab_dict[" "]
374
+
375
+ # add unk and pad token
376
+ if unk_token is not None:
377
+ vocab_dict[unk_token] = len(vocab_dict)
378
+
379
+ if pad_token is not None:
380
+ vocab_dict[pad_token] = len(vocab_dict)
381
+
382
+ return vocab_dict
383
+
384
+
385
+ def main():
386
+ # See all possible arguments in src/transformers/training_args.py
387
+ # or by passing the --help flag to this script.
388
+ # We now keep distinct sets of args, for a cleaner separation of concerns.
389
+
390
+ parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments, ExtraArguments))
391
+ if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
392
+ # If we pass only one argument to the script and it's the path to a json file,
393
+ # let's parse it to get our arguments.
394
+ model_args, data_args, training_args, extra_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
395
+ else:
396
+ model_args, data_args, training_args, extra_args = parser.parse_args_into_dataclasses()
397
+
398
+ # Detecting last checkpoint.
399
+ last_checkpoint = None
400
+ if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
401
+ last_checkpoint = get_last_checkpoint(training_args.output_dir)
402
+ if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
403
+ raise ValueError(
404
+ f"Output directory ({training_args.output_dir}) already exists and is not empty. "
405
+ "Use --overwrite_output_dir to overcome."
406
+ )
407
+ elif last_checkpoint is not None:
408
+ logger.info(
409
+ f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
410
+ "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
411
+ )
412
+
413
+ # Setup logging
414
+ logging.basicConfig(
415
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
416
+ datefmt="%m/%d/%Y %H:%M:%S",
417
+ handlers=[logging.StreamHandler(sys.stdout)],
418
+ )
419
+ logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
420
+
421
+ # Log on each process the small summary:
422
+ logger.warning(
423
+ f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
424
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
425
+ )
426
+ # Set the verbosity to info of the Transformers logger (on main process only):
427
+ if is_main_process(training_args.local_rank):
428
+ transformers.utils.logging.set_verbosity_info()
429
+ logger.info("Training/evaluation parameters %s", training_args)
430
+
431
+ # Set seed before initializing model.
432
+ set_seed(training_args.seed)
433
+
434
+ # configure wandb run
435
+ os.environ["WANDB_PROJECT"] = extra_args.wandb_project
436
+
437
+ # 1. First, let's load the dataset
438
+ raw_datasets = DatasetDict()
439
+
440
+ if training_args.do_train:
441
+ if data_args.dataset_name.endswith("/"):
442
+ raw_datasets["train"] = load_from_disk(f"{data_args.dataset_name}/{data_args.train_split_name}")
443
+ else:
444
+ raw_datasets["train"] = load_dataset(
445
+ data_args.dataset_name,
446
+ data_args.dataset_config_name,
447
+ split=data_args.train_split_name,
448
+ use_auth_token=data_args.use_auth_token,
449
+ )
450
+
451
+ if data_args.audio_column_name not in raw_datasets["train"].column_names:
452
+ raise ValueError(
453
+ f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
454
+ "Make sure to set `--audio_column_name` to the correct audio column - one of "
455
+ f"{', '.join(raw_datasets['train'].column_names)}."
456
+ )
457
+
458
+ if data_args.text_column_name not in raw_datasets["train"].column_names:
459
+ raise ValueError(
460
+ f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
461
+ "Make sure to set `--text_column_name` to the correct text column - one of "
462
+ f"{', '.join(raw_datasets['train'].column_names)}."
463
+ )
464
+
465
+ if data_args.max_train_samples is not None:
466
+ raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
467
+
468
+ if training_args.do_eval:
469
+ if data_args.dataset_name.endswith("/"):
470
+ raw_datasets["eval"] = load_from_disk(f"{data_args.dataset_name}/{data_args.eval_split_name}")
471
+ else:
472
+ raw_datasets["eval"] = load_dataset(
473
+ data_args.dataset_name,
474
+ data_args.dataset_config_name,
475
+ split=data_args.eval_split_name,
476
+ use_auth_token=data_args.use_auth_token,
477
+ )
478
+
479
+ if data_args.max_eval_samples is not None:
480
+ raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
481
+
482
+ # 2. We remove some special characters from the datasets
483
+ # that make training complicated and do not help in transcribing the speech
484
+ # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
485
+ # that could be easily picked up by the model
486
+ chars_to_ignore_regex = (
487
+ f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
488
+ )
489
+ text_column_name = data_args.text_column_name
490
+
491
+ def remove_special_characters(batch):
492
+ if chars_to_ignore_regex is not None:
493
+ batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
494
+ else:
495
+ batch["target_text"] = batch[text_column_name].lower() + " "
496
+ return batch
497
+
498
+ with training_args.main_process_first(desc="dataset map special characters removal"):
499
+ raw_datasets = raw_datasets.map(
500
+ remove_special_characters,
501
+ remove_columns=[text_column_name],
502
+ desc="remove special characters from datasets",
503
+ )
504
+
505
+ # save special tokens for tokenizer
506
+ word_delimiter_token = data_args.word_delimiter_token
507
+ unk_token = data_args.unk_token
508
+ pad_token = data_args.pad_token
509
+
510
+ # 3. Next, let's load the config as we might need it to create
511
+ # the tokenizer
512
+ # load config
513
+ config = AutoConfig.from_pretrained(
514
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
515
+ )
516
+
517
+ # 4. Next, if no tokenizer file is defined,
518
+ # we create the vocabulary of the model by extracting all unique characters from
519
+ # the training and evaluation datasets
520
+ # We need to make sure that only first rank saves vocabulary
521
+ # make sure all processes wait until vocab is created
522
+ tokenizer_name_or_path = model_args.tokenizer_name_or_path
523
+ tokenizer_kwargs = {}
524
+ if tokenizer_name_or_path is None:
525
+ # save vocab in training output dir
526
+ tokenizer_name_or_path = training_args.output_dir
527
+
528
+ vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
529
+
530
+ with training_args.main_process_first():
531
+ if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
532
+ os.remove(vocab_file)
533
+
534
+ with training_args.main_process_first(desc="dataset map vocabulary creation"):
535
+ if not os.path.isfile(vocab_file):
536
+ os.makedirs(tokenizer_name_or_path, exist_ok=True)
537
+ vocab_dict = create_vocabulary_from_data(
538
+ raw_datasets,
539
+ word_delimiter_token=word_delimiter_token,
540
+ unk_token=unk_token,
541
+ pad_token=pad_token,
542
+ )
543
+
544
+ # save vocab dict to be loaded into tokenizer
545
+ with open(vocab_file, "w") as file:
546
+ json.dump(vocab_dict, file)
547
+
548
+ # if tokenizer has just been created
549
+ # it is defined by `tokenizer_class` if present in config else by `model_type`
550
+ tokenizer_kwargs = {
551
+ "config": config if config.tokenizer_class is not None else None,
552
+ "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
553
+ "unk_token": unk_token,
554
+ "pad_token": pad_token,
555
+ "word_delimiter_token": word_delimiter_token,
556
+ }
557
+
558
+ # 5. Now we can instantiate the feature extractor, tokenizer and model
559
+ # Note for distributed training, the .from_pretrained methods guarantee that only
560
+ # one local process can concurrently download model & vocab.
561
+
562
+ # load feature_extractor and tokenizer
563
+ tokenizer = AutoTokenizer.from_pretrained(
564
+ tokenizer_name_or_path,
565
+ use_auth_token=data_args.use_auth_token,
566
+ **tokenizer_kwargs,
567
+ )
568
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
569
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
570
+ )
571
+
572
+ # adapt config
573
+ config.update(
574
+ {
575
+ "feat_proj_dropout": model_args.feat_proj_dropout,
576
+ "attention_dropout": model_args.attention_dropout,
577
+ "hidden_dropout": model_args.hidden_dropout,
578
+ "final_dropout": model_args.final_dropout,
579
+ "mask_time_prob": model_args.mask_time_prob,
580
+ "mask_time_length": model_args.mask_time_length,
581
+ "mask_feature_prob": model_args.mask_feature_prob,
582
+ "mask_feature_length": model_args.mask_feature_length,
583
+ "gradient_checkpointing": training_args.gradient_checkpointing,
584
+ "layerdrop": model_args.layerdrop,
585
+ "ctc_loss_reduction": model_args.ctc_loss_reduction,
586
+ "pad_token_id": tokenizer.pad_token_id,
587
+ "vocab_size": len(tokenizer),
588
+ "activation_dropout": model_args.activation_dropout,
589
+ }
590
+ )
591
+
592
+ # create model
593
+ model = AutoModelForCTC.from_pretrained(
594
+ model_args.model_name_or_path,
595
+ cache_dir=model_args.cache_dir,
596
+ config=config,
597
+ use_auth_token=data_args.use_auth_token,
598
+ )
599
+
600
+ # freeze encoder
601
+ if model_args.freeze_feature_encoder:
602
+ model.freeze_feature_encoder()
603
+
604
+ # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
605
+ # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
606
+ # so that we just need to set the correct target sampling rate and normalize the input
607
+ # via the `feature_extractor`
608
+
609
+ # make sure that dataset decodes audio with correct sampling rate
610
+ dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
611
+ if dataset_sampling_rate != feature_extractor.sampling_rate:
612
+ raw_datasets = raw_datasets.cast_column(
613
+ data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
614
+ )
615
+
616
+ # derive max & min input length for sample rate & max duration
617
+ max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
618
+ min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
619
+ audio_column_name = data_args.audio_column_name
620
+ num_workers = data_args.preprocessing_num_workers
621
+
622
+ # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
623
+ phoneme_language = data_args.phoneme_language
624
+
625
+ # Preprocessing the datasets.
626
+ # We need to read the audio files as arrays and tokenize the targets.
627
+ def prepare_dataset(batch):
628
+ # load audio
629
+ sample = batch[audio_column_name]
630
+
631
+ inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
632
+ batch["input_values"] = inputs.input_values[0]
633
+ batch["length"] = len(batch["input_values"])
634
+
635
+ # encode targets
636
+ additional_kwargs = {}
637
+ if phoneme_language is not None:
638
+ additional_kwargs["phonemizer_lang"] = phoneme_language
639
+
640
+ batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
641
+ return batch
642
+
643
+ with training_args.main_process_first(desc="dataset map preprocessing"):
644
+ vectorized_datasets = raw_datasets.map(
645
+ prepare_dataset,
646
+ remove_columns=next(iter(raw_datasets.values())).column_names,
647
+ num_proc=num_workers,
648
+ desc="preprocess datasets",
649
+ )
650
+
651
+ def is_audio_in_length_range(length):
652
+ return length > min_input_length and length < max_input_length
653
+
654
+ # filter data that is shorter than min_input_length
655
+ vectorized_datasets = vectorized_datasets.filter(
656
+ is_audio_in_length_range,
657
+ num_proc=num_workers,
658
+ input_columns=["length"],
659
+ )
660
+
661
+ # 7. Next, we can prepare the training.
662
+ # Let's use word error rate (WER) as our evaluation metric,
663
+ # instantiate a data collator and the trainer
664
+
665
+ # Define evaluation metrics during training, *i.e.* word error rate, character error rate
666
+ eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
667
+
668
+ # for large datasets it is advised to run the preprocessing on a
669
+ # single machine first with ``args.preprocessing_only`` since there will mostly likely
670
+ # be a timeout when running the script in distributed mode.
671
+ # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
672
+ # cached dataset
673
+ if data_args.preprocessing_only:
674
+ logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
675
+ return
676
+
677
+ def compute_metrics(pred):
678
+ pred_logits = pred.predictions
679
+ pred_ids = np.argmax(pred_logits, axis=-1)
680
+
681
+ pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
682
+
683
+ pred_str = tokenizer.batch_decode(pred_ids)
684
+ # we do not want to group tokens when computing the metrics
685
+ label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
686
+
687
+ metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
688
+
689
+ return metrics
690
+
691
+ # Now save everything to be able to create a single processor later
692
+ if is_main_process(training_args.local_rank):
693
+ # save feature extractor, tokenizer and config
694
+ feature_extractor.save_pretrained(training_args.output_dir)
695
+ tokenizer.save_pretrained(training_args.output_dir)
696
+ config.save_pretrained(training_args.output_dir)
697
+
698
+ try:
699
+ processor = AutoProcessor.from_pretrained(training_args.output_dir)
700
+ except (OSError, KeyError):
701
+ warnings.warn(
702
+ "Loading a processor from a feature extractor config that does not"
703
+ " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
704
+ " attribute to your `preprocessor_config.json` file to suppress this warning: "
705
+ " `'processor_class': 'Wav2Vec2Processor'`",
706
+ FutureWarning,
707
+ )
708
+ processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
709
+
710
+ # Instantiate custom data collator
711
+ data_collator = DataCollatorCTCWithPadding(processor=processor)
712
+
713
+ decay_parameters = get_parameter_names(model, [torch.nn.LayerNorm])
714
+ decay_parameters = [name for name in decay_parameters if "bias" not in name]
715
+ optimizer_grouped_parameters = [
716
+ {
717
+ "params": [p for n, p in model.named_parameters() if n in decay_parameters],
718
+ "weight_decay": training_args.weight_decay,
719
+ },
720
+ {
721
+ "params": [p for n, p in model.named_parameters() if n not in decay_parameters],
722
+ "weight_decay": 0.0,
723
+ },
724
+ ]
725
+ if extra_args.bnb:
726
+ optimizer = bnb.optim.Adam8bit(
727
+ params=optimizer_grouped_parameters,
728
+ lr=training_args.learning_rate,
729
+ betas=(training_args.adam_beta1, training_args.adam_beta2),
730
+ eps=training_args.adam_epsilon,
731
+ )
732
+ else:
733
+ optimizer = torch.optim.AdamW(
734
+ params=optimizer_grouped_parameters,
735
+ lr=training_args.learning_rate,
736
+ betas=(training_args.adam_beta1, training_args.adam_beta2),
737
+ eps=training_args.adam_epsilon,
738
+ )
739
+ if extra_args.tristage_sched:
740
+ scheduler = get_tri_stage_schedule(optimizer, training_args.max_steps)
741
+ else:
742
+ scheduler = None
743
+ optimizers = (optimizer, scheduler)
744
+
745
+ # Initialize Trainer
746
+ trainer = Trainer(
747
+ model=model,
748
+ data_collator=data_collator,
749
+ args=training_args,
750
+ compute_metrics=compute_metrics,
751
+ train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
752
+ eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
753
+ tokenizer=feature_extractor,
754
+ optimizers=optimizers,
755
+ )
756
+
757
+ # 8. Finally, we can start training
758
+
759
+ # Training
760
+ if training_args.do_train:
761
+
762
+ # use last checkpoint if exist
763
+ if last_checkpoint is not None:
764
+ checkpoint = last_checkpoint
765
+ elif os.path.isdir(model_args.model_name_or_path):
766
+ checkpoint = model_args.model_name_or_path
767
+ else:
768
+ checkpoint = None
769
+
770
+ train_result = trainer.train(resume_from_checkpoint=checkpoint)
771
+ trainer.save_model()
772
+
773
+ metrics = train_result.metrics
774
+ max_train_samples = (
775
+ data_args.max_train_samples
776
+ if data_args.max_train_samples is not None
777
+ else len(vectorized_datasets["train"])
778
+ )
779
+ metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
780
+
781
+ trainer.log_metrics("train", metrics)
782
+ trainer.save_metrics("train", metrics)
783
+ trainer.save_state()
784
+
785
+ # Evaluation
786
+ results = {}
787
+ if training_args.do_eval:
788
+ logger.info("*** Evaluate ***")
789
+ metrics = trainer.evaluate()
790
+ max_eval_samples = (
791
+ data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
792
+ )
793
+ metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
794
+
795
+ trainer.log_metrics("eval", metrics)
796
+ trainer.save_metrics("eval", metrics)
797
+
798
+ # Write model card and (optionally) push to hub
799
+ config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
800
+ kwargs = {
801
+ "finetuned_from": model_args.model_name_or_path,
802
+ "tasks": "speech-recognition",
803
+ "tags": ["automatic-speech-recognition", data_args.dataset_name],
804
+ "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
805
+ "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
806
+ }
807
+ if "common_voice" in data_args.dataset_name:
808
+ kwargs["language"] = config_name
809
+
810
+ if training_args.push_to_hub:
811
+ trainer.push_to_hub(**kwargs)
812
+ else:
813
+ trainer.create_model_card(**kwargs)
814
+
815
+ return results
816
+
817
+
818
+ if __name__ == "__main__":
819
+ main()
runs/Jan30_22-51-36_job-b1f4681b-d20d-47f2-af64-0c1734f4ff64/1643583136.495153/events.out.tfevents.1643583136.job-b1f4681b-d20d-47f2-af64-0c1734f4ff64.28099.1 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c822bc82b95a41418c740d3df0b4c87e020c9fc8acecf736978ffeb2f19a0e87
3
+ size 4772
runs/Jan30_22-51-36_job-b1f4681b-d20d-47f2-af64-0c1734f4ff64/events.out.tfevents.1643583136.job-b1f4681b-d20d-47f2-af64-0c1734f4ff64.28099.0 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4887a01c9b94b9a5c62bf28f5268fd5b1f73c89e0eb36532a103dea26da2c0a4
3
+ size 11310
runs/Jan30_22-51-36_job-b1f4681b-d20d-47f2-af64-0c1734f4ff64/events.out.tfevents.1643595198.job-b1f4681b-d20d-47f2-af64-0c1734f4ff64.28099.2 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:71a4c1fd0a3d8971ae0dfb139fb4339857f9f66f823a3c8d61b9ff59d3b2a462
3
+ size 405
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": 109.09,
3
+ "train_loss": 1.3784224351247152,
4
+ "train_runtime": 12041.2604,
5
+ "train_samples": 1456,
6
+ "train_samples_per_second": 12.756,
7
+ "train_steps_per_second": 0.1
8
+ }
trainer_state.json ADDED
@@ -0,0 +1,217 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "best_metric": 0.3061184883117676,
3
+ "best_model_checkpoint": "./checkpoint-700",
4
+ "epoch": 109.08791208791209,
5
+ "global_step": 1200,
6
+ "is_hyper_param_search": false,
7
+ "is_local_process_zero": true,
8
+ "is_world_process_zero": true,
9
+ "log_history": [
10
+ {
11
+ "epoch": 9.09,
12
+ "learning_rate": 6.548e-05,
13
+ "loss": 5.195,
14
+ "step": 100
15
+ },
16
+ {
17
+ "epoch": 9.09,
18
+ "eval_cer": 1.0,
19
+ "eval_loss": 3.133816719055176,
20
+ "eval_runtime": 15.8007,
21
+ "eval_samples_per_second": 21.202,
22
+ "eval_steps_per_second": 0.38,
23
+ "eval_wer": 1.0,
24
+ "step": 100
25
+ },
26
+ {
27
+ "epoch": 18.18,
28
+ "learning_rate": 8e-05,
29
+ "loss": 2.3769,
30
+ "step": 200
31
+ },
32
+ {
33
+ "epoch": 18.18,
34
+ "eval_cer": 0.13601416287303997,
35
+ "eval_loss": 0.48250746726989746,
36
+ "eval_runtime": 15.2805,
37
+ "eval_samples_per_second": 21.923,
38
+ "eval_steps_per_second": 0.393,
39
+ "eval_wer": 0.6615925058548009,
40
+ "step": 200
41
+ },
42
+ {
43
+ "epoch": 27.26,
44
+ "learning_rate": 8e-05,
45
+ "loss": 1.345,
46
+ "step": 300
47
+ },
48
+ {
49
+ "epoch": 27.26,
50
+ "eval_cer": 0.10571573090541224,
51
+ "eval_loss": 0.3780955374240875,
52
+ "eval_runtime": 15.0829,
53
+ "eval_samples_per_second": 22.211,
54
+ "eval_steps_per_second": 0.398,
55
+ "eval_wer": 0.5113192818110851,
56
+ "step": 300
57
+ },
58
+ {
59
+ "epoch": 36.35,
60
+ "learning_rate": 8e-05,
61
+ "loss": 1.2001,
62
+ "step": 400
63
+ },
64
+ {
65
+ "epoch": 36.35,
66
+ "eval_cer": 0.09312089023773394,
67
+ "eval_loss": 0.3571384847164154,
68
+ "eval_runtime": 15.2304,
69
+ "eval_samples_per_second": 21.996,
70
+ "eval_steps_per_second": 0.394,
71
+ "eval_wer": 0.4601873536299766,
72
+ "step": 400
73
+ },
74
+ {
75
+ "epoch": 45.44,
76
+ "learning_rate": 8e-05,
77
+ "loss": 1.0484,
78
+ "step": 500
79
+ },
80
+ {
81
+ "epoch": 45.44,
82
+ "eval_cer": 0.07764289327263531,
83
+ "eval_loss": 0.3121073842048645,
84
+ "eval_runtime": 15.2202,
85
+ "eval_samples_per_second": 22.01,
86
+ "eval_steps_per_second": 0.394,
87
+ "eval_wer": 0.4094457455113193,
88
+ "step": 500
89
+ },
90
+ {
91
+ "epoch": 54.53,
92
+ "learning_rate": 8e-05,
93
+ "loss": 0.926,
94
+ "step": 600
95
+ },
96
+ {
97
+ "epoch": 54.53,
98
+ "eval_cer": 0.08012139605462823,
99
+ "eval_loss": 0.3227134943008423,
100
+ "eval_runtime": 15.3003,
101
+ "eval_samples_per_second": 21.895,
102
+ "eval_steps_per_second": 0.392,
103
+ "eval_wer": 0.4094457455113193,
104
+ "step": 600
105
+ },
106
+ {
107
+ "epoch": 63.62,
108
+ "learning_rate": 6.758666666666667e-05,
109
+ "loss": 0.8854,
110
+ "step": 700
111
+ },
112
+ {
113
+ "epoch": 63.62,
114
+ "eval_cer": 0.07465857359635812,
115
+ "eval_loss": 0.3061184883117676,
116
+ "eval_runtime": 17.2128,
117
+ "eval_samples_per_second": 19.462,
118
+ "eval_steps_per_second": 0.349,
119
+ "eval_wer": 0.38992974238875877,
120
+ "step": 700
121
+ },
122
+ {
123
+ "epoch": 72.7,
124
+ "learning_rate": 5.492e-05,
125
+ "loss": 0.8054,
126
+ "step": 800
127
+ },
128
+ {
129
+ "epoch": 72.7,
130
+ "eval_cer": 0.07445624683864441,
131
+ "eval_loss": 0.315933495759964,
132
+ "eval_runtime": 17.3337,
133
+ "eval_samples_per_second": 19.327,
134
+ "eval_steps_per_second": 0.346,
135
+ "eval_wer": 0.38914910226385635,
136
+ "step": 800
137
+ },
138
+ {
139
+ "epoch": 81.79,
140
+ "learning_rate": 4.225333333333334e-05,
141
+ "loss": 0.7442,
142
+ "step": 900
143
+ },
144
+ {
145
+ "epoch": 81.79,
146
+ "eval_cer": 0.07309054122407689,
147
+ "eval_loss": 0.3135768473148346,
148
+ "eval_runtime": 15.2189,
149
+ "eval_samples_per_second": 22.012,
150
+ "eval_steps_per_second": 0.394,
151
+ "eval_wer": 0.3801717408274785,
152
+ "step": 900
153
+ },
154
+ {
155
+ "epoch": 90.88,
156
+ "learning_rate": 2.958666666666667e-05,
157
+ "loss": 0.714,
158
+ "step": 1000
159
+ },
160
+ {
161
+ "epoch": 90.88,
162
+ "eval_cer": 0.07172483560950936,
163
+ "eval_loss": 0.32300877571105957,
164
+ "eval_runtime": 15.24,
165
+ "eval_samples_per_second": 21.982,
166
+ "eval_steps_per_second": 0.394,
167
+ "eval_wer": 0.37158469945355194,
168
+ "step": 1000
169
+ },
170
+ {
171
+ "epoch": 99.97,
172
+ "learning_rate": 1.6920000000000004e-05,
173
+ "loss": 0.6641,
174
+ "step": 1100
175
+ },
176
+ {
177
+ "epoch": 99.97,
178
+ "eval_cer": 0.06934749620637329,
179
+ "eval_loss": 0.31931421160697937,
180
+ "eval_runtime": 15.1822,
181
+ "eval_samples_per_second": 22.065,
182
+ "eval_steps_per_second": 0.395,
183
+ "eval_wer": 0.36104605776736926,
184
+ "step": 1100
185
+ },
186
+ {
187
+ "epoch": 109.09,
188
+ "learning_rate": 4.253333333333336e-06,
189
+ "loss": 0.6367,
190
+ "step": 1200
191
+ },
192
+ {
193
+ "epoch": 109.09,
194
+ "eval_cer": 0.06924633282751644,
195
+ "eval_loss": 0.32542118430137634,
196
+ "eval_runtime": 15.2075,
197
+ "eval_samples_per_second": 22.029,
198
+ "eval_steps_per_second": 0.395,
199
+ "eval_wer": 0.358704137392662,
200
+ "step": 1200
201
+ },
202
+ {
203
+ "epoch": 109.09,
204
+ "step": 1200,
205
+ "total_flos": 9.090968694813691e+19,
206
+ "train_loss": 1.3784224351247152,
207
+ "train_runtime": 12041.2604,
208
+ "train_samples_per_second": 12.756,
209
+ "train_steps_per_second": 0.1
210
+ }
211
+ ],
212
+ "max_steps": 1200,
213
+ "num_train_epochs": 110,
214
+ "total_flos": 9.090968694813691e+19,
215
+ "trial_name": null,
216
+ "trial_params": null
217
+ }
training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9886d1fbbad75a820f26ae7488ec03668e4ac81f646bc9c64621c5a4caedeba1
3
+ size 3055
vocab.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"ա": 1, "բ": 2, "գ": 3, "դ": 4, "ե": 5, "զ": 6, "է": 7, "ը": 8, "թ": 9, "ժ": 10, "ի": 11, "լ": 12, "խ": 13, "ծ": 14, "կ": 15, "հ": 16, "ձ": 17, "ղ": 18, "ճ": 19, "մ": 20, "յ": 21, "ն": 22, "շ": 23, "ո": 24, "չ": 25, "պ": 26, "ջ": 27, "ռ": 28, "ս": 29, "վ": 30, "տ": 31, "ր": 32, "ց": 33, "ւ": 34, "փ": 35, "ք": 36, "օ": 37, "ֆ": 38, "և": 39, "|": 0, "[UNK]": 40, "[PAD]": 41}