infinitejoy commited on
Commit
657058f
1 Parent(s): 1026e4f

Training in progress, step 4000

Browse files
.gitignore ADDED
@@ -0,0 +1 @@
 
 
1
+ checkpoint-*/
config.json ADDED
@@ -0,0 +1,107 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_name_or_path": "facebook/wav2vec2-xls-r-300m",
3
+ "activation_dropout": 0.1,
4
+ "adapter_kernel_size": 3,
5
+ "adapter_stride": 2,
6
+ "add_adapter": false,
7
+ "apply_spec_augment": true,
8
+ "architectures": [
9
+ "Wav2Vec2ForCTC"
10
+ ],
11
+ "attention_dropout": 0.0,
12
+ "bos_token_id": 1,
13
+ "classifier_proj_size": 256,
14
+ "codevector_dim": 768,
15
+ "contrastive_logits_temperature": 0.1,
16
+ "conv_bias": true,
17
+ "conv_dim": [
18
+ 512,
19
+ 512,
20
+ 512,
21
+ 512,
22
+ 512,
23
+ 512,
24
+ 512
25
+ ],
26
+ "conv_kernel": [
27
+ 10,
28
+ 3,
29
+ 3,
30
+ 3,
31
+ 3,
32
+ 2,
33
+ 2
34
+ ],
35
+ "conv_stride": [
36
+ 5,
37
+ 2,
38
+ 2,
39
+ 2,
40
+ 2,
41
+ 2,
42
+ 2
43
+ ],
44
+ "ctc_loss_reduction": "mean",
45
+ "ctc_zero_infinity": false,
46
+ "diversity_loss_weight": 0.1,
47
+ "do_stable_layer_norm": true,
48
+ "eos_token_id": 2,
49
+ "feat_extract_activation": "gelu",
50
+ "feat_extract_dropout": 0.0,
51
+ "feat_extract_norm": "layer",
52
+ "feat_proj_dropout": 0.0,
53
+ "feat_quantizer_dropout": 0.0,
54
+ "final_dropout": 0.0,
55
+ "hidden_act": "gelu",
56
+ "hidden_dropout": 0.0,
57
+ "hidden_size": 1024,
58
+ "initializer_range": 0.02,
59
+ "intermediate_size": 4096,
60
+ "layer_norm_eps": 1e-05,
61
+ "layerdrop": 0.0,
62
+ "mask_feature_length": 64,
63
+ "mask_feature_min_masks": 0,
64
+ "mask_feature_prob": 0.25,
65
+ "mask_time_length": 10,
66
+ "mask_time_min_masks": 2,
67
+ "mask_time_prob": 0.75,
68
+ "model_type": "wav2vec2",
69
+ "num_adapter_layers": 3,
70
+ "num_attention_heads": 16,
71
+ "num_codevector_groups": 2,
72
+ "num_codevectors_per_group": 320,
73
+ "num_conv_pos_embedding_groups": 16,
74
+ "num_conv_pos_embeddings": 128,
75
+ "num_feat_extract_layers": 7,
76
+ "num_hidden_layers": 24,
77
+ "num_negatives": 100,
78
+ "output_hidden_size": 1024,
79
+ "pad_token_id": 43,
80
+ "proj_codevector_dim": 768,
81
+ "tdnn_dilation": [
82
+ 1,
83
+ 2,
84
+ 3,
85
+ 1,
86
+ 1
87
+ ],
88
+ "tdnn_dim": [
89
+ 512,
90
+ 512,
91
+ 512,
92
+ 512,
93
+ 1500
94
+ ],
95
+ "tdnn_kernel": [
96
+ 5,
97
+ 3,
98
+ 3,
99
+ 1,
100
+ 1
101
+ ],
102
+ "torch_dtype": "float32",
103
+ "transformers_version": "4.16.0.dev0",
104
+ "use_weighted_layer_sum": false,
105
+ "vocab_size": 46,
106
+ "xvector_output_dim": 512
107
+ }
eval.py ADDED
@@ -0,0 +1,137 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
+
53
+ chars_to_ignore_regex = '[,?.!\-\;\:"“%‘”�—’…–]' # noqa: W605 IMPORTANT: this should correspond to the chars that were ignored during training
54
+
55
+ text = re.sub(chars_to_ignore_regex, "", text.lower())
56
+
57
+ # In addition, we can normalize the target text, e.g. removing new lines characters etc...
58
+ # note that order is important here!
59
+ token_sequences_to_ignore = ["\n\n", "\n", " ", " "]
60
+
61
+ for t in token_sequences_to_ignore:
62
+ text = " ".join(text.split(t))
63
+
64
+ return text
65
+
66
+
67
+ def main(args):
68
+ # load dataset
69
+ dataset = load_dataset(args.dataset, args.config, split=args.split, use_auth_token=True)
70
+
71
+ # for testing: only process the first two examples as a test
72
+ # dataset = dataset.select(range(10))
73
+
74
+ # load processor
75
+ feature_extractor = AutoFeatureExtractor.from_pretrained(args.model_id)
76
+ sampling_rate = feature_extractor.sampling_rate
77
+
78
+ # resample audio
79
+ dataset = dataset.cast_column("audio", Audio(sampling_rate=sampling_rate))
80
+
81
+ # load eval pipeline
82
+ if args.device is None:
83
+ args.device = 0 if torch.cuda.is_available() else -1
84
+ asr = pipeline("automatic-speech-recognition", model=args.model_id, device=args.device)
85
+
86
+ # map function to decode audio
87
+ def map_to_pred(batch):
88
+ prediction = asr(
89
+ batch["audio"]["array"], chunk_length_s=args.chunk_length_s, stride_length_s=args.stride_length_s
90
+ )
91
+
92
+ batch["prediction"] = prediction["text"]
93
+ batch["target"] = normalize_text(batch["sentence"])
94
+ return batch
95
+
96
+ # run inference on all examples
97
+ result = dataset.map(map_to_pred, remove_columns=dataset.column_names)
98
+
99
+ # compute and log_results
100
+ # do not change function below
101
+ log_results(result, args)
102
+
103
+
104
+ if __name__ == "__main__":
105
+ parser = argparse.ArgumentParser()
106
+
107
+ parser.add_argument(
108
+ "--model_id", type=str, required=True, help="Model identifier. Should be loadable with 🤗 Transformers"
109
+ )
110
+ parser.add_argument(
111
+ "--dataset",
112
+ type=str,
113
+ required=True,
114
+ help="Dataset name to evaluate the `model_id`. Should be loadable with 🤗 Datasets",
115
+ )
116
+ parser.add_argument(
117
+ "--config", type=str, required=True, help="Config of the dataset. *E.g.* `'en'` for Common Voice"
118
+ )
119
+ parser.add_argument("--split", type=str, required=True, help="Split of the dataset. *E.g.* `'test'`")
120
+ parser.add_argument(
121
+ "--chunk_length_s", type=float, default=None, help="Chunk length in seconds. Defaults to 5 seconds."
122
+ )
123
+ parser.add_argument(
124
+ "--stride_length_s", type=float, default=None, help="Stride of the audio chunks. Defaults to 1 second."
125
+ )
126
+ parser.add_argument(
127
+ "--log_outputs", action="store_true", help="If defined, write outputs to log file for analysis."
128
+ )
129
+ parser.add_argument(
130
+ "--device",
131
+ type=int,
132
+ default=None,
133
+ help="The device to run the pipeline on. -1 for CPU (default), 0 for the first GPU and so on.",
134
+ )
135
+ args = parser.parse_args()
136
+
137
+ main(args)
out.log ADDED
@@ -0,0 +1,91 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 02/04/2022 03:04:30 - WARNING - __main__ - Process rank: -1, device: cuda:0, n_gpu: 1distributed training: False, 16-bits training: True
2
+ 02/04/2022 03:04:30 - INFO - __main__ - Training/evaluation parameters TrainingArguments(
3
+ _n_gpu=1,
4
+ adafactor=False,
5
+ adam_beta1=0.9,
6
+ adam_beta2=0.999,
7
+ adam_epsilon=1e-08,
8
+ bf16=False,
9
+ bf16_full_eval=False,
10
+ dataloader_drop_last=False,
11
+ dataloader_num_workers=0,
12
+ dataloader_pin_memory=True,
13
+ ddp_bucket_cap_mb=None,
14
+ ddp_find_unused_parameters=None,
15
+ debug=[],
16
+ deepspeed=None,
17
+ disable_tqdm=False,
18
+ do_eval=True,
19
+ do_predict=False,
20
+ do_train=True,
21
+ eval_accumulation_steps=None,
22
+ eval_steps=4000,
23
+ evaluation_strategy=IntervalStrategy.STEPS,
24
+ fp16=True,
25
+ fp16_backend=auto,
26
+ fp16_full_eval=False,
27
+ fp16_opt_level=O1,
28
+ gradient_accumulation_steps=1,
29
+ gradient_checkpointing=True,
30
+ greater_is_better=None,
31
+ group_by_length=True,
32
+ half_precision_backend=auto,
33
+ hub_model_id=None,
34
+ hub_strategy=HubStrategy.EVERY_SAVE,
35
+ hub_token=<HUB_TOKEN>,
36
+ ignore_data_skip=False,
37
+ label_names=None,
38
+ label_smoothing_factor=0.0,
39
+ learning_rate=7e-05,
40
+ length_column_name=input_length,
41
+ load_best_model_at_end=False,
42
+ local_rank=-1,
43
+ log_level=-1,
44
+ log_level_replica=-1,
45
+ log_on_each_node=True,
46
+ logging_dir=./wav2vec2-large-xls-r-300m-abkhaz-cv8/runs/Feb04_03-04-30_job-8be8b741-e32e-4579-bbec-1e00d9824b4f,
47
+ logging_first_step=False,
48
+ logging_nan_inf_filter=True,
49
+ logging_steps=1000,
50
+ logging_strategy=IntervalStrategy.STEPS,
51
+ lr_scheduler_type=SchedulerType.LINEAR,
52
+ max_grad_norm=1.0,
53
+ max_steps=-1,
54
+ metric_for_best_model=None,
55
+ mp_parameters=,
56
+ no_cuda=False,
57
+ num_train_epochs=50.0,
58
+ optim=OptimizerNames.ADAMW_HF,
59
+ output_dir=./wav2vec2-large-xls-r-300m-abkhaz-cv8,
60
+ overwrite_output_dir=True,
61
+ past_index=-1,
62
+ per_device_eval_batch_size=16,
63
+ per_device_train_batch_size=32,
64
+ prediction_loss_only=False,
65
+ push_to_hub=True,
66
+ push_to_hub_model_id=None,
67
+ push_to_hub_organization=None,
68
+ push_to_hub_token=<PUSH_TO_HUB_TOKEN>,
69
+ remove_unused_columns=True,
70
+ report_to=[],
71
+ resume_from_checkpoint=None,
72
+ run_name=./wav2vec2-large-xls-r-300m-abkhaz-cv8,
73
+ save_on_each_node=False,
74
+ save_steps=4000,
75
+ save_strategy=IntervalStrategy.STEPS,
76
+ save_total_limit=2,
77
+ seed=42,
78
+ sharded_ddp=[],
79
+ skip_memory_metrics=True,
80
+ tf32=None,
81
+ tpu_metrics_debug=False,
82
+ tpu_num_cores=None,
83
+ use_legacy_prediction_loop=False,
84
+ warmup_ratio=0.0,
85
+ warmup_steps=4000,
86
+ weight_decay=0.0,
87
+ xpu_backend=None,
88
+ )
89
+ 02/04/2022 03:04:33 - WARNING - datasets.builder - Reusing dataset common_voice (/workspace/.cache/huggingface/datasets/mozilla-foundation___common_voice/ab/7.0.0/fe20cac47c166e25b1f096ab661832e3da7cf298ed4a91dcaa1343ad972d175b)
90
+ 02/04/2022 03:04:35 - WARNING - datasets.builder - Reusing dataset common_voice (/workspace/.cache/huggingface/datasets/mozilla-foundation___common_voice/ab/7.0.0/fe20cac47c166e25b1f096ab661832e3da7cf298ed4a91dcaa1343ad972d175b)
91
+ 02/04/2022 03:05:07 - WARNING - huggingface_hub.repository - Cloning https://huggingface.co/infinitejoy/wav2vec2-large-xls-r-300m-abkhaz-cv8 into local empty directory.
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 sha256:5fd9ffe0c663fb7498e57833981968585a34c752d58c1e3b4ecf45b22a012795
3
+ size 1262112241
run_speech_recognition_ctc.py ADDED
@@ -0,0 +1,740 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python
2
+ # coding=utf-8
3
+ # Copyright 2021 The HuggingFace Inc. team. All rights reserved.
4
+ #
5
+ # Licensed under the Apache License, Version 2.0 (the "License");
6
+ # you may not use this file except in compliance with the License.
7
+ # You may obtain a copy of the License at
8
+ #
9
+ # http://www.apache.org/licenses/LICENSE-2.0
10
+ #
11
+ # Unless required by applicable law or agreed to in writing, software
12
+ # distributed under the License is distributed on an "AS IS" BASIS,
13
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14
+ # See the License for the specific language governing permissions and
15
+
16
+ """ Fine-tuning a 🤗 Transformers CTC model for automatic speech recognition"""
17
+
18
+ import functools
19
+ import json
20
+ import logging
21
+ import os
22
+ import re
23
+ import sys
24
+ import warnings
25
+ from dataclasses import dataclass, field
26
+ from typing import Dict, List, Optional, Union
27
+
28
+ import datasets
29
+ import numpy as np
30
+ import torch
31
+ from datasets import DatasetDict, load_dataset, load_metric
32
+
33
+ import transformers
34
+ from transformers import (
35
+ AutoConfig,
36
+ AutoFeatureExtractor,
37
+ AutoModelForCTC,
38
+ AutoProcessor,
39
+ AutoTokenizer,
40
+ HfArgumentParser,
41
+ Trainer,
42
+ TrainingArguments,
43
+ Wav2Vec2Processor,
44
+ set_seed,
45
+ )
46
+ from transformers.trainer_utils import get_last_checkpoint, is_main_process
47
+ from transformers.utils import check_min_version
48
+ from transformers.utils.versions import require_version
49
+
50
+
51
+ # Will error if the minimal version of Transformers is not installed. Remove at your own risks.
52
+ check_min_version("4.16.0.dev0")
53
+
54
+ require_version("datasets>=1.13.3", "To fix: pip install -r examples/pytorch/text-classification/requirements.txt")
55
+
56
+
57
+ logger = logging.getLogger(__name__)
58
+
59
+
60
+ def list_field(default=None, metadata=None):
61
+ return field(default_factory=lambda: default, metadata=metadata)
62
+
63
+
64
+ @dataclass
65
+ class ModelArguments:
66
+ """
67
+ Arguments pertaining to which model/config/tokenizer we are going to fine-tune from.
68
+ """
69
+
70
+ model_name_or_path: str = field(
71
+ metadata={"help": "Path to pretrained model or model identifier from huggingface.co/models"}
72
+ )
73
+ tokenizer_name_or_path: Optional[str] = field(
74
+ default=None,
75
+ metadata={"help": "Path to pretrained tokenizer or tokenizer identifier from huggingface.co/models"},
76
+ )
77
+ cache_dir: Optional[str] = field(
78
+ default=None,
79
+ metadata={"help": "Where do you want to store the pretrained models downloaded from huggingface.co"},
80
+ )
81
+ freeze_feature_encoder: bool = field(
82
+ default=True, metadata={"help": "Whether to freeze the feature encoder layers of the model."}
83
+ )
84
+ attention_dropout: float = field(
85
+ default=0.0, metadata={"help": "The dropout ratio for the attention probabilities."}
86
+ )
87
+ activation_dropout: float = field(
88
+ default=0.0, metadata={"help": "The dropout ratio for activations inside the fully connected layer."}
89
+ )
90
+ feat_proj_dropout: float = field(default=0.0, metadata={"help": "The dropout ratio for the projected features."})
91
+ hidden_dropout: float = field(
92
+ default=0.0,
93
+ metadata={
94
+ "help": "The dropout probability for all fully connected layers in the embeddings, encoder, and pooler."
95
+ },
96
+ )
97
+ final_dropout: float = field(
98
+ default=0.0,
99
+ metadata={"help": "The dropout probability for the final projection layer."},
100
+ )
101
+ mask_time_prob: float = field(
102
+ default=0.05,
103
+ metadata={
104
+ "help": "Probability of each feature vector along the time axis to be chosen as the start of the vector"
105
+ "span to be masked. Approximately ``mask_time_prob * sequence_length // mask_time_length`` feature"
106
+ "vectors will be masked along the time axis."
107
+ },
108
+ )
109
+ mask_time_length: int = field(
110
+ default=10,
111
+ metadata={"help": "Length of vector span to mask along the time axis."},
112
+ )
113
+ mask_feature_prob: float = field(
114
+ default=0.0,
115
+ metadata={
116
+ "help": "Probability of each feature vector along the feature axis to be chosen as the start of the vector"
117
+ "span to be masked. Approximately ``mask_feature_prob * sequence_length // mask_feature_length`` feature bins will be masked along the time axis."
118
+ },
119
+ )
120
+ mask_feature_length: int = field(
121
+ default=10,
122
+ metadata={"help": "Length of vector span to mask along the feature axis."},
123
+ )
124
+ layerdrop: float = field(default=0.0, metadata={"help": "The LayerDrop probability."})
125
+ ctc_loss_reduction: Optional[str] = field(
126
+ default="mean", metadata={"help": "The way the ctc loss should be reduced. Should be one of 'mean' or 'sum'."}
127
+ )
128
+
129
+
130
+ @dataclass
131
+ class DataTrainingArguments:
132
+ """
133
+ Arguments pertaining to what data we are going to input our model for training and eval.
134
+
135
+ Using `HfArgumentParser` we can turn this class
136
+ into argparse arguments to be able to specify them on
137
+ the command line.
138
+ """
139
+
140
+ dataset_name: str = field(
141
+ metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
142
+ )
143
+ dataset_config_name: str = field(
144
+ default=None, metadata={"help": "The configuration name of the dataset to use (via the datasets library)."}
145
+ )
146
+ train_split_name: str = field(
147
+ default="train+validation",
148
+ metadata={
149
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
150
+ },
151
+ )
152
+ eval_split_name: str = field(
153
+ default="test",
154
+ metadata={
155
+ "help": "The name of the training data set split to use (via the datasets library). Defaults to 'train'"
156
+ },
157
+ )
158
+ audio_column_name: str = field(
159
+ default="audio",
160
+ metadata={"help": "The name of the dataset column containing the audio data. Defaults to 'audio'"},
161
+ )
162
+ text_column_name: str = field(
163
+ default="text",
164
+ metadata={"help": "The name of the dataset column containing the text data. Defaults to 'text'"},
165
+ )
166
+ overwrite_cache: bool = field(
167
+ default=False, metadata={"help": "Overwrite the cached preprocessed datasets or not."}
168
+ )
169
+ preprocessing_num_workers: Optional[int] = field(
170
+ default=None,
171
+ metadata={"help": "The number of processes to use for the preprocessing."},
172
+ )
173
+ max_train_samples: Optional[int] = field(
174
+ default=None,
175
+ metadata={
176
+ "help": "For debugging purposes or quicker training, truncate the number of training examples to this "
177
+ "value if set."
178
+ },
179
+ )
180
+ max_eval_samples: Optional[int] = field(
181
+ default=None,
182
+ metadata={
183
+ "help": "For debugging purposes or quicker training, truncate the number of validation examples to this "
184
+ "value if set."
185
+ },
186
+ )
187
+ chars_to_ignore: Optional[List[str]] = list_field(
188
+ default=None,
189
+ metadata={"help": "A list of characters to remove from the transcripts."},
190
+ )
191
+ eval_metrics: List[str] = list_field(
192
+ default=["wer"],
193
+ metadata={"help": "A list of metrics the model should be evaluated on. E.g. `'wer cer'`"},
194
+ )
195
+ max_duration_in_seconds: float = field(
196
+ default=20.0,
197
+ metadata={
198
+ "help": "Filter audio files that are longer than `max_duration_in_seconds` seconds to 'max_duration_in_seconds`"
199
+ },
200
+ )
201
+ min_duration_in_seconds: float = field(
202
+ default=0.0, metadata={"help": "Filter audio files that are shorter than `min_duration_in_seconds` seconds"}
203
+ )
204
+ preprocessing_only: bool = field(
205
+ default=False,
206
+ metadata={
207
+ "help": "Whether to only do data preprocessing and skip training. "
208
+ "This is especially useful when data preprocessing errors out in distributed training due to timeout. "
209
+ "In this case, one should run the preprocessing in a non-distributed setup with `preprocessing_only=True` "
210
+ "so that the cached datasets can consequently be loaded in distributed training"
211
+ },
212
+ )
213
+ use_auth_token: bool = field(
214
+ default=False,
215
+ metadata={
216
+ "help": "If :obj:`True`, will use the token generated when running"
217
+ ":obj:`transformers-cli login` as HTTP bearer authorization for remote files."
218
+ },
219
+ )
220
+ unk_token: str = field(
221
+ default="[UNK]",
222
+ metadata={"help": "The unk token for the tokenizer"},
223
+ )
224
+ pad_token: str = field(
225
+ default="[PAD]",
226
+ metadata={"help": "The padding token for the tokenizer"},
227
+ )
228
+ word_delimiter_token: str = field(
229
+ default="|",
230
+ metadata={"help": "The word delimiter token for the tokenizer"},
231
+ )
232
+ phoneme_language: Optional[str] = field(
233
+ default=None,
234
+ metadata={
235
+ "help": "The target language that should be used be"
236
+ " passed to the tokenizer for tokenization. Note that"
237
+ " this is only relevant if the model classifies the"
238
+ " input audio to a sequence of phoneme sequences."
239
+ },
240
+ )
241
+
242
+
243
+ @dataclass
244
+ class DataCollatorCTCWithPadding:
245
+ """
246
+ Data collator that will dynamically pad the inputs received.
247
+ Args:
248
+ processor (:class:`~transformers.AutoProcessor`)
249
+ The processor used for proccessing the data.
250
+ padding (:obj:`bool`, :obj:`str` or :class:`~transformers.tokenization_utils_base.PaddingStrategy`, `optional`, defaults to :obj:`True`):
251
+ Select a strategy to pad the returned sequences (according to the model's padding side and padding index)
252
+ among:
253
+ * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
254
+ sequence if provided).
255
+ * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the
256
+ maximum acceptable input length for the model if that argument is not provided.
257
+ * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of
258
+ different lengths).
259
+ max_length (:obj:`int`, `optional`):
260
+ Maximum length of the ``input_values`` of the returned list and optionally padding length (see above).
261
+ max_length_labels (:obj:`int`, `optional`):
262
+ Maximum length of the ``labels`` returned list and optionally padding length (see above).
263
+ pad_to_multiple_of (:obj:`int`, `optional`):
264
+ If set will pad the sequence to a multiple of the provided value.
265
+ This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >=
266
+ 7.5 (Volta).
267
+ """
268
+
269
+ processor: AutoProcessor
270
+ padding: Union[bool, str] = "longest"
271
+ pad_to_multiple_of: Optional[int] = None
272
+ pad_to_multiple_of_labels: Optional[int] = None
273
+
274
+ def __call__(self, features: List[Dict[str, Union[List[int], torch.Tensor]]]) -> Dict[str, torch.Tensor]:
275
+ # split inputs and labels since they have to be of different lenghts and need
276
+ # different padding methods
277
+ input_features = [{"input_values": feature["input_values"]} for feature in features]
278
+ label_features = [{"input_ids": feature["labels"]} for feature in features]
279
+
280
+ batch = self.processor.pad(
281
+ input_features,
282
+ padding=self.padding,
283
+ pad_to_multiple_of=self.pad_to_multiple_of,
284
+ return_tensors="pt",
285
+ )
286
+
287
+ with self.processor.as_target_processor():
288
+ labels_batch = self.processor.pad(
289
+ label_features,
290
+ padding=self.padding,
291
+ pad_to_multiple_of=self.pad_to_multiple_of_labels,
292
+ return_tensors="pt",
293
+ )
294
+
295
+ # replace padding with -100 to ignore loss correctly
296
+ labels = labels_batch["input_ids"].masked_fill(labels_batch.attention_mask.ne(1), -100)
297
+
298
+ batch["labels"] = labels
299
+
300
+ return batch
301
+
302
+
303
+ def create_vocabulary_from_data(
304
+ datasets: DatasetDict,
305
+ word_delimiter_token: Optional[str] = None,
306
+ unk_token: Optional[str] = None,
307
+ pad_token: Optional[str] = None,
308
+ ):
309
+ # Given training and test labels create vocabulary
310
+ def extract_all_chars(batch):
311
+ all_text = " ".join(batch["target_text"])
312
+ vocab = list(set(all_text))
313
+ return {"vocab": [vocab], "all_text": [all_text]}
314
+
315
+ vocabs = datasets.map(
316
+ extract_all_chars,
317
+ batched=True,
318
+ batch_size=-1,
319
+ keep_in_memory=True,
320
+ remove_columns=datasets["train"].column_names,
321
+ )
322
+
323
+ # take union of all unique characters in each dataset
324
+ vocab_set = functools.reduce(
325
+ lambda vocab_1, vocab_2: set(vocab_1["vocab"][0]) | set(vocab_2["vocab"][0]), vocabs.values()
326
+ )
327
+
328
+ vocab_dict = {v: k for k, v in enumerate(sorted(list(vocab_set)))}
329
+
330
+ # replace white space with delimiter token
331
+ if word_delimiter_token is not None:
332
+ vocab_dict[word_delimiter_token] = vocab_dict[" "]
333
+ del vocab_dict[" "]
334
+
335
+ # add unk and pad token
336
+ if unk_token is not None:
337
+ vocab_dict[unk_token] = len(vocab_dict)
338
+
339
+ if pad_token is not None:
340
+ vocab_dict[pad_token] = len(vocab_dict)
341
+
342
+ return vocab_dict
343
+
344
+
345
+ def main():
346
+ # See all possible arguments in src/transformers/training_args.py
347
+ # or by passing the --help flag to this script.
348
+ # We now keep distinct sets of args, for a cleaner separation of concerns.
349
+
350
+ parser = HfArgumentParser((ModelArguments, DataTrainingArguments, TrainingArguments))
351
+ if len(sys.argv) == 2 and sys.argv[1].endswith(".json"):
352
+ # If we pass only one argument to the script and it's the path to a json file,
353
+ # let's parse it to get our arguments.
354
+ model_args, data_args, training_args = parser.parse_json_file(json_file=os.path.abspath(sys.argv[1]))
355
+ else:
356
+ model_args, data_args, training_args = parser.parse_args_into_dataclasses()
357
+
358
+ # Detecting last checkpoint.
359
+ last_checkpoint = None
360
+ if os.path.isdir(training_args.output_dir) and training_args.do_train and not training_args.overwrite_output_dir:
361
+ last_checkpoint = get_last_checkpoint(training_args.output_dir)
362
+ if last_checkpoint is None and len(os.listdir(training_args.output_dir)) > 0:
363
+ raise ValueError(
364
+ f"Output directory ({training_args.output_dir}) already exists and is not empty. "
365
+ "Use --overwrite_output_dir to overcome."
366
+ )
367
+ elif last_checkpoint is not None:
368
+ logger.info(
369
+ f"Checkpoint detected, resuming training at {last_checkpoint}. To avoid this behavior, change "
370
+ "the `--output_dir` or add `--overwrite_output_dir` to train from scratch."
371
+ )
372
+
373
+ # Setup logging
374
+ logging.basicConfig(
375
+ format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
376
+ datefmt="%m/%d/%Y %H:%M:%S",
377
+ handlers=[logging.StreamHandler(sys.stdout)],
378
+ )
379
+ logger.setLevel(logging.INFO if is_main_process(training_args.local_rank) else logging.WARN)
380
+
381
+ # Log on each process the small summary:
382
+ logger.warning(
383
+ f"Process rank: {training_args.local_rank}, device: {training_args.device}, n_gpu: {training_args.n_gpu}"
384
+ f"distributed training: {bool(training_args.local_rank != -1)}, 16-bits training: {training_args.fp16}"
385
+ )
386
+ # Set the verbosity to info of the Transformers logger (on main process only):
387
+ if is_main_process(training_args.local_rank):
388
+ transformers.utils.logging.set_verbosity_info()
389
+ logger.info("Training/evaluation parameters %s", training_args)
390
+
391
+ # Set seed before initializing model.
392
+ set_seed(training_args.seed)
393
+
394
+ # 1. First, let's load the dataset
395
+ raw_datasets = DatasetDict()
396
+
397
+ if training_args.do_train:
398
+ raw_datasets["train"] = load_dataset(
399
+ data_args.dataset_name,
400
+ data_args.dataset_config_name,
401
+ split=data_args.train_split_name,
402
+ use_auth_token=data_args.use_auth_token,
403
+ )
404
+
405
+ if data_args.audio_column_name not in raw_datasets["train"].column_names:
406
+ raise ValueError(
407
+ f"--audio_column_name '{data_args.audio_column_name}' not found in dataset '{data_args.dataset_name}'. "
408
+ "Make sure to set `--audio_column_name` to the correct audio column - one of "
409
+ f"{', '.join(raw_datasets['train'].column_names)}."
410
+ )
411
+
412
+ if data_args.text_column_name not in raw_datasets["train"].column_names:
413
+ raise ValueError(
414
+ f"--text_column_name {data_args.text_column_name} not found in dataset '{data_args.dataset_name}'. "
415
+ "Make sure to set `--text_column_name` to the correct text column - one of "
416
+ f"{', '.join(raw_datasets['train'].column_names)}."
417
+ )
418
+
419
+ if data_args.max_train_samples is not None:
420
+ raw_datasets["train"] = raw_datasets["train"].select(range(data_args.max_train_samples))
421
+
422
+ if training_args.do_eval:
423
+ raw_datasets["eval"] = load_dataset(
424
+ data_args.dataset_name,
425
+ data_args.dataset_config_name,
426
+ split=data_args.eval_split_name,
427
+ use_auth_token=data_args.use_auth_token,
428
+ )
429
+
430
+ if data_args.max_eval_samples is not None:
431
+ raw_datasets["eval"] = raw_datasets["eval"].select(range(data_args.max_eval_samples))
432
+
433
+ # 2. We remove some special characters from the datasets
434
+ # that make training complicated and do not help in transcribing the speech
435
+ # E.g. characters, such as `,` and `.` do not really have an acoustic characteristic
436
+ # that could be easily picked up by the model
437
+ chars_to_ignore_regex = (
438
+ f'[{"".join(data_args.chars_to_ignore)}]' if data_args.chars_to_ignore is not None else None
439
+ )
440
+ text_column_name = data_args.text_column_name
441
+
442
+ def remove_special_characters(batch):
443
+ if chars_to_ignore_regex is not None:
444
+ batch["target_text"] = re.sub(chars_to_ignore_regex, "", batch[text_column_name]).lower() + " "
445
+ else:
446
+ batch["target_text"] = batch[text_column_name].lower() + " "
447
+
448
+ batch["target_text"] = batch[text_column_name].lower().replace('c', '')
449
+
450
+ return batch
451
+
452
+ with training_args.main_process_first(desc="dataset map special characters removal"):
453
+ raw_datasets = raw_datasets.map(
454
+ remove_special_characters,
455
+ remove_columns=[text_column_name],
456
+ desc="remove special characters from datasets",
457
+ )
458
+
459
+ # save special tokens for tokenizer
460
+ word_delimiter_token = data_args.word_delimiter_token
461
+ unk_token = data_args.unk_token
462
+ pad_token = data_args.pad_token
463
+
464
+ # 3. Next, let's load the config as we might need it to create
465
+ # the tokenizer
466
+ # load config
467
+ config = AutoConfig.from_pretrained(
468
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
469
+ )
470
+
471
+ # 4. Next, if no tokenizer file is defined,
472
+ # we create the vocabulary of the model by extracting all unique characters from
473
+ # the training and evaluation datasets
474
+ # We need to make sure that only first rank saves vocabulary
475
+ # make sure all processes wait until vocab is created
476
+ tokenizer_name_or_path = model_args.tokenizer_name_or_path
477
+ tokenizer_kwargs = {}
478
+ if tokenizer_name_or_path is None:
479
+ # save vocab in training output dir
480
+ tokenizer_name_or_path = training_args.output_dir
481
+
482
+ vocab_file = os.path.join(tokenizer_name_or_path, "vocab.json")
483
+
484
+ with training_args.main_process_first():
485
+ if training_args.overwrite_output_dir and os.path.isfile(vocab_file):
486
+ os.remove(vocab_file)
487
+
488
+ with training_args.main_process_first(desc="dataset map vocabulary creation"):
489
+ if not os.path.isfile(vocab_file):
490
+ os.makedirs(tokenizer_name_or_path, exist_ok=True)
491
+ vocab_dict = create_vocabulary_from_data(
492
+ raw_datasets,
493
+ word_delimiter_token=word_delimiter_token,
494
+ unk_token=unk_token,
495
+ pad_token=pad_token,
496
+ )
497
+
498
+ # save vocab dict to be loaded into tokenizer
499
+ with open(vocab_file, "w") as file:
500
+ json.dump(vocab_dict, file)
501
+
502
+ # if tokenizer has just been created
503
+ # it is defined by `tokenizer_class` if present in config else by `model_type`
504
+ tokenizer_kwargs = {
505
+ "config": config if config.tokenizer_class is not None else None,
506
+ "tokenizer_type": config.model_type if config.tokenizer_class is None else None,
507
+ "unk_token": unk_token,
508
+ "pad_token": pad_token,
509
+ "word_delimiter_token": word_delimiter_token,
510
+ }
511
+
512
+ # 5. Now we can instantiate the feature extractor, tokenizer and model
513
+ # Note for distributed training, the .from_pretrained methods guarantee that only
514
+ # one local process can concurrently download model & vocab.
515
+
516
+ # load feature_extractor and tokenizer
517
+ tokenizer = AutoTokenizer.from_pretrained(
518
+ tokenizer_name_or_path,
519
+ use_auth_token=data_args.use_auth_token,
520
+ **tokenizer_kwargs,
521
+ )
522
+ feature_extractor = AutoFeatureExtractor.from_pretrained(
523
+ model_args.model_name_or_path, cache_dir=model_args.cache_dir, use_auth_token=data_args.use_auth_token
524
+ )
525
+
526
+ # adapt config
527
+ config.update(
528
+ {
529
+ "feat_proj_dropout": model_args.feat_proj_dropout,
530
+ "attention_dropout": model_args.attention_dropout,
531
+ "hidden_dropout": model_args.hidden_dropout,
532
+ "final_dropout": model_args.final_dropout,
533
+ "mask_time_prob": model_args.mask_time_prob,
534
+ "mask_time_length": model_args.mask_time_length,
535
+ "mask_feature_prob": model_args.mask_feature_prob,
536
+ "mask_feature_length": model_args.mask_feature_length,
537
+ "gradient_checkpointing": training_args.gradient_checkpointing,
538
+ "layerdrop": model_args.layerdrop,
539
+ "ctc_loss_reduction": model_args.ctc_loss_reduction,
540
+ "pad_token_id": tokenizer.pad_token_id,
541
+ "vocab_size": len(tokenizer),
542
+ "activation_dropout": model_args.activation_dropout,
543
+ }
544
+ )
545
+
546
+ # create model
547
+ model = AutoModelForCTC.from_pretrained(
548
+ model_args.model_name_or_path,
549
+ cache_dir=model_args.cache_dir,
550
+ config=config,
551
+ use_auth_token=data_args.use_auth_token,
552
+ )
553
+
554
+ # freeze encoder
555
+ if model_args.freeze_feature_encoder:
556
+ model.freeze_feature_encoder()
557
+
558
+ # 6. Now we preprocess the datasets including loading the audio, resampling and normalization
559
+ # Thankfully, `datasets` takes care of automatically loading and resampling the audio,
560
+ # so that we just need to set the correct target sampling rate and normalize the input
561
+ # via the `feature_extractor`
562
+
563
+ # make sure that dataset decodes audio with correct sampling rate
564
+ dataset_sampling_rate = next(iter(raw_datasets.values())).features[data_args.audio_column_name].sampling_rate
565
+ if dataset_sampling_rate != feature_extractor.sampling_rate:
566
+ raw_datasets = raw_datasets.cast_column(
567
+ data_args.audio_column_name, datasets.features.Audio(sampling_rate=feature_extractor.sampling_rate)
568
+ )
569
+
570
+ # derive max & min input length for sample rate & max duration
571
+ max_input_length = data_args.max_duration_in_seconds * feature_extractor.sampling_rate
572
+ min_input_length = data_args.min_duration_in_seconds * feature_extractor.sampling_rate
573
+ audio_column_name = data_args.audio_column_name
574
+ num_workers = data_args.preprocessing_num_workers
575
+
576
+ # `phoneme_language` is only relevant if the model is fine-tuned on phoneme classification
577
+ phoneme_language = data_args.phoneme_language
578
+
579
+ # Preprocessing the datasets.
580
+ # We need to read the audio files as arrays and tokenize the targets.
581
+ def prepare_dataset(batch):
582
+ # load audio
583
+ sample = batch[audio_column_name]
584
+
585
+ inputs = feature_extractor(sample["array"], sampling_rate=sample["sampling_rate"])
586
+ batch["input_values"] = inputs.input_values[0]
587
+ batch["input_length"] = len(batch["input_values"])
588
+
589
+ # encode targets
590
+ additional_kwargs = {}
591
+ if phoneme_language is not None:
592
+ additional_kwargs["phonemizer_lang"] = phoneme_language
593
+
594
+ batch["labels"] = tokenizer(batch["target_text"], **additional_kwargs).input_ids
595
+ return batch
596
+
597
+ with training_args.main_process_first(desc="dataset map preprocessing"):
598
+ vectorized_datasets = raw_datasets.map(
599
+ prepare_dataset,
600
+ remove_columns=next(iter(raw_datasets.values())).column_names,
601
+ num_proc=num_workers,
602
+ desc="preprocess datasets",
603
+ )
604
+
605
+ def is_audio_in_length_range(length):
606
+ return length > min_input_length and length < max_input_length
607
+
608
+ # filter data that is shorter than min_input_length
609
+ vectorized_datasets = vectorized_datasets.filter(
610
+ is_audio_in_length_range,
611
+ num_proc=num_workers,
612
+ input_columns=["input_length"],
613
+ )
614
+
615
+ # 7. Next, we can prepare the training.
616
+ # Let's use word error rate (WER) as our evaluation metric,
617
+ # instantiate a data collator and the trainer
618
+
619
+ # Define evaluation metrics during training, *i.e.* word error rate, character error rate
620
+ eval_metrics = {metric: load_metric(metric) for metric in data_args.eval_metrics}
621
+
622
+ # for large datasets it is advised to run the preprocessing on a
623
+ # single machine first with ``args.preprocessing_only`` since there will mostly likely
624
+ # be a timeout when running the script in distributed mode.
625
+ # In a second step ``args.preprocessing_only`` can then be set to `False` to load the
626
+ # cached dataset
627
+ if data_args.preprocessing_only:
628
+ logger.info(f"Data preprocessing finished. Files cached at {vectorized_datasets.cache_files}")
629
+ return
630
+
631
+ def compute_metrics(pred):
632
+ pred_logits = pred.predictions
633
+ pred_ids = np.argmax(pred_logits, axis=-1)
634
+
635
+ pred.label_ids[pred.label_ids == -100] = tokenizer.pad_token_id
636
+
637
+ pred_str = tokenizer.batch_decode(pred_ids)
638
+ # we do not want to group tokens when computing the metrics
639
+ label_str = tokenizer.batch_decode(pred.label_ids, group_tokens=False)
640
+
641
+ metrics = {k: v.compute(predictions=pred_str, references=label_str) for k, v in eval_metrics.items()}
642
+
643
+ return metrics
644
+
645
+ # Now save everything to be able to create a single processor later
646
+ if is_main_process(training_args.local_rank):
647
+ # save feature extractor, tokenizer and config
648
+ feature_extractor.save_pretrained(training_args.output_dir)
649
+ tokenizer.save_pretrained(training_args.output_dir)
650
+ config.save_pretrained(training_args.output_dir)
651
+
652
+ try:
653
+ processor = AutoProcessor.from_pretrained(training_args.output_dir)
654
+ except (OSError, KeyError):
655
+ warnings.warn(
656
+ "Loading a processor from a feature extractor config that does not"
657
+ " include a `processor_class` attribute is deprecated and will be removed in v5. Please add the following "
658
+ " attribute to your `preprocessor_config.json` file to suppress this warning: "
659
+ " `'processor_class': 'Wav2Vec2Processor'`",
660
+ FutureWarning,
661
+ )
662
+ processor = Wav2Vec2Processor.from_pretrained(training_args.output_dir)
663
+
664
+ # Instantiate custom data collator
665
+ data_collator = DataCollatorCTCWithPadding(processor=processor)
666
+
667
+ # Initialize Trainer
668
+ trainer = Trainer(
669
+ model=model,
670
+ data_collator=data_collator,
671
+ args=training_args,
672
+ compute_metrics=compute_metrics,
673
+ train_dataset=vectorized_datasets["train"] if training_args.do_train else None,
674
+ eval_dataset=vectorized_datasets["eval"] if training_args.do_eval else None,
675
+ tokenizer=feature_extractor,
676
+ )
677
+
678
+ # 8. Finally, we can start training
679
+
680
+ # Training
681
+ if training_args.do_train:
682
+
683
+ # use last checkpoint if exist
684
+ if last_checkpoint is not None:
685
+ checkpoint = last_checkpoint
686
+ elif os.path.isdir(model_args.model_name_or_path):
687
+ checkpoint = model_args.model_name_or_path
688
+ else:
689
+ checkpoint = None
690
+
691
+ train_result = trainer.train(resume_from_checkpoint=checkpoint)
692
+ trainer.save_model()
693
+
694
+ metrics = train_result.metrics
695
+ max_train_samples = (
696
+ data_args.max_train_samples
697
+ if data_args.max_train_samples is not None
698
+ else len(vectorized_datasets["train"])
699
+ )
700
+ metrics["train_samples"] = min(max_train_samples, len(vectorized_datasets["train"]))
701
+
702
+ trainer.log_metrics("train", metrics)
703
+ trainer.save_metrics("train", metrics)
704
+ trainer.save_state()
705
+
706
+ # Evaluation
707
+ results = {}
708
+ if training_args.do_eval:
709
+ logger.info("*** Evaluate ***")
710
+ metrics = trainer.evaluate()
711
+ max_eval_samples = (
712
+ data_args.max_eval_samples if data_args.max_eval_samples is not None else len(vectorized_datasets["eval"])
713
+ )
714
+ metrics["eval_samples"] = min(max_eval_samples, len(vectorized_datasets["eval"]))
715
+
716
+ trainer.log_metrics("eval", metrics)
717
+ trainer.save_metrics("eval", metrics)
718
+
719
+ # Write model card and (optionally) push to hub
720
+ config_name = data_args.dataset_config_name if data_args.dataset_config_name is not None else "na"
721
+ kwargs = {
722
+ "finetuned_from": model_args.model_name_or_path,
723
+ "tasks": "speech-recognition",
724
+ "tags": ["automatic-speech-recognition", data_args.dataset_name],
725
+ "dataset_args": f"Config: {config_name}, Training split: {data_args.train_split_name}, Eval split: {data_args.eval_split_name}",
726
+ "dataset": f"{data_args.dataset_name.upper()} - {config_name.upper()}",
727
+ }
728
+ if "common_voice" in data_args.dataset_name:
729
+ kwargs["language"] = config_name
730
+
731
+ if training_args.push_to_hub:
732
+ trainer.push_to_hub(**kwargs)
733
+ else:
734
+ trainer.create_model_card(**kwargs)
735
+
736
+ return results
737
+
738
+
739
+ if __name__ == "__main__":
740
+ main()
special_tokens_map.json CHANGED
@@ -1 +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}]}
 
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}, {"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 CHANGED
@@ -1 +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"}
 
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": "./wav2vec2-large-xls-r-300m-abkhaz-cv8", "tokenizer_class": "Wav2Vec2CTCTokenizer"}
training_args.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:767816e61ffa2269a45193b6a452ef9d82108a84b35d0b23e850bef890a77022
3
+ size 3055
wav2vec2-large-xls-r-300m-abkhaz-cv8 ADDED
@@ -0,0 +1 @@
 
 
1
+ Subproject commit 1026e4f5d2b106c5fe8fc30f818d70a305adde90