-
Notifications
You must be signed in to change notification settings - Fork 0
/
pre_train.py
137 lines (108 loc) · 4.07 KB
/
pre_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
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
import numpy as np
from datasets import Dataset, load_dataset
from dataclasses import dataclass, field
import torch
from transformers import (
DataCollatorForLanguageModeling,
Trainer,
TrainerCallback,
TrainingArguments,
Qwen2Tokenizer,
Qwen2Config,
Qwen2ForCausalLM,
)
from transformers.trainer_callback import TrainerControl, TrainerState
TRAIN_FILES = [
'./datasets/wiki-zh.parquet',
]
EVAL_FILE = "./datasets/pretrain_eval_512_1w.parquet"
# %%
@dataclass
class PretrainArguments:
tokenizer_dir: str = "./model_save/"
model_save_dir: str = "./model_save/pre/"
logs_dir: str = "./logs/"
train_files: list = field(default_factory=lambda: TRAIN_FILES)
eval_file: str = EVAL_FILE
max_seq_len: int = 1024
def get_mapped_dataset(tokenizer, files, dt) -> Dataset:
dataset = load_dataset(path="parquet", data_files=files, split="train", cache_dir=".cache", keep_in_memory=False)
def token_to_id(samples: dict) -> dict:
batch_txt = samples["text"]
outputs = tokenizer(
batch_txt,
padding=False,
return_attention_mask=False,
truncation=False
)
input_ids = [np.array(item, dtype=dt) for item in outputs["input_ids"]]
return {"input_ids": input_ids}
mapped_dataset = dataset.map(token_to_id, batched=True, batch_size=10000, remove_columns=dataset.column_names, num_proc=8, keep_in_memory=False)
return mapped_dataset
class MyTrainerCallback(TrainerCallback):
log_cnt = 0
def on_log(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, **kwargs):
self.log_cnt += 1
if self.log_cnt % 2 == 0:
torch.cuda.empty_cache()
def on_epoch_end(self, args: TrainingArguments, state: TrainerState, control: TrainerControl, **kwargs):
control.should_save = True
return control
def main():
pretrain_args = PretrainArguments()
tokenizer = Qwen2Tokenizer.from_pretrained(pretrain_args.tokenizer_dir)
vocab_size = len(tokenizer)
if vocab_size % 64 != 0:
vocab_size = (vocab_size // 64 + 1) * 64
print(f"final vocab size: {vocab_size}")
dt = np.uint16 if vocab_size < 65535 else np.uint32
train_dataset = get_mapped_dataset(tokenizer, pretrain_args.train_files, dt)
eval_dataset = get_mapped_dataset(tokenizer, pretrain_args.eval_file, dt)
# `mlm=False`表示要训练CLM模型,`mlm=True`表示要训练MLM模型
data_collator = DataCollatorForLanguageModeling(tokenizer, mlm=False)
config = Qwen2Config.from_pretrained(pretrain_args.tokenizer_dir)
model = Qwen2ForCausalLM(config)
model_size = sum(t.numel() for t in model.parameters())
print(f"QWen size: {model_size / 1000 ** 2:.1f}M parameters")
trainer_callback = MyTrainerCallback()
args = TrainingArguments(
output_dir=pretrain_args.model_save_dir,
per_device_train_batch_size=8,
per_device_eval_batch_size=8,
gradient_accumulation_steps=8,
num_train_epochs=1,
weight_decay=0.1,
ddp_find_unused_parameters=False,
warmup_steps=0,
learning_rate=1e-4,
evaluation_strategy="steps",
eval_steps=100,
save_steps=100,
save_strategy="steps",
save_total_limit=3,
report_to="tensorboard",
optim="adamw_torch",
lr_scheduler_type="cosine",
bf16=True,
logging_steps=10,
log_level="info",
logging_first_step=True,
)
trainer = Trainer(
model=model,
tokenizer=tokenizer,
args=args,
data_collator=data_collator,
train_dataset=train_dataset,
eval_dataset=eval_dataset,
callbacks=[trainer_callback],
)
# `resume_from_checkpoint=True`参数可以从上次保存的检查点继续训练
trainer.train( #'model_save/pre/checkpoint-3400'
# resume_from_checkpoint=True
)
eval_results = trainer.evaluate()
print(f"Perplexity: {np.exp(eval_results['eval_loss']):.2f}")
trainer.save_model(pretrain_args.model_save_dir)
if __name__ == "__main__":
main()