-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
69 lines (57 loc) · 2.4 KB
/
train.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
import warnings
import numpy as np
import yaml
from transformers import Trainer, TrainingArguments, Wav2Vec2ForCTC
from dataset import dataset
from utils import DataCollatorCTCWithPadding, compute_metrics
def train():
warnings.filterwarnings("ignore", category=np.VisibleDeprecationWarning)
warnings.filterwarnings("ignore", category=UserWarning)
with open("config_train.yml") as f:
args = yaml.load(f, Loader=yaml.FullLoader)
dataset_train, dataset_test, processor = dataset(args)
data_collator = DataCollatorCTCWithPadding(processor=processor, padding=True)
model = Wav2Vec2ForCTC.from_pretrained(
args["pretrained_checkpoint_dir"],
attention_dropout=args["attention_dropout"],
hidden_dropout=args["hidden_dropout"],
feat_proj_dropout=args["feat_proj_dropout"],
mask_time_prob=args["mask_time_prob"],
layerdrop=args["layerdrop"],
# gradient_checkpointing=args["gradient_checkpointing"],
ctc_loss_reduction=args["ctc_loss_reduction"],
pad_token_id=processor.tokenizer.pad_token_id,
vocab_size=len(processor.tokenizer),
)
model.freeze_feature_extractor()
print("-------load_pretrained_model_done----------")
training_args = TrainingArguments(
output_dir=args["checkpoint_dir"],
group_by_length=args["group_by_length"],
per_device_train_batch_size=args["batch_size"],
per_device_eval_batch_size=args["batch_size"],
gradient_accumulation_steps=args["gradient_accumulation_steps"],
evaluation_strategy=args["evaluation_strategy"],
num_train_epochs=args["num_train_epochs"],
fp16=args["fp16"],
save_steps=args["save_steps"],
eval_steps=args["eval_steps"],
logging_steps=args["logging_steps"],
weight_decay=args["weight_decay"],
learning_rate=args["learning_rate"],
warmup_steps=args["warmup_steps"],
save_total_limit=args["save_total_limit"],
dataloader_num_workers=args["dataloader_num_workers"],
)
print("-------train_ready_done---------")
trainer = Trainer(
model=model,
data_collator=data_collator,
args=training_args,
compute_metrics=compute_metrics,
train_dataset=dataset_train,
eval_dataset=dataset_test,
tokenizer=processor.feature_extractor,
)
print("-------training_start!---------")
trainer.train()