-
Notifications
You must be signed in to change notification settings - Fork 1
/
train.py
122 lines (110 loc) · 4.62 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
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
import math
from argparse import ArgumentParser
from functools import partial
from os.path import join
import torch
from clip import ClipAdaptionPromptV2ForMultiModalConditionalGeneration
from data import build_dataset, collate_data
from peft import AdaptionPromptV2Config, TaskType, PeftType
from transformers import AutoProcessor, AutoTokenizer, Trainer, TrainingArguments
def train():
parser = ArgumentParser()
parser.add_argument("--pretrained_language_model_name_or_path", type=str)
parser.add_argument("--pretrained_vision_model_name_or_path", type=str)
args = parser.parse_args()
pretrained_language_model_name_or_path = args.pretrained_language_model_name_or_path
pretrained_vision_model_name_or_path = args.pretrained_vision_model_name_or_path
train_args = TrainingArguments(
output_dir=join(pretrained_language_model_name_or_path, "mm_adaption_prompt_v2"),
per_device_train_batch_size=1,
per_device_eval_batch_size=1,
learning_rate=3e-3,
num_train_epochs=3,
deepspeed=None,
gradient_checkpointing=False,
gradient_accumulation_steps=2,
logging_strategy="steps",
logging_steps=10,
evaluation_strategy="epoch",
save_strategy="no",
save_total_limit=1,
load_best_model_at_end=False,
disable_tqdm=False,
remove_unused_columns=False,
local_rank=-1,
do_train=True,
do_eval=True,
seed=1024,
data_seed=1024,
fp16=True,
fp16_full_eval=True,
bf16=False,
bf16_full_eval=False
)
train_args.gradient_accumulation_steps = max(1, train_args.gradient_accumulation_steps // train_args.world_size)
print("preparing tokenizer and processor...")
tokenizer = AutoTokenizer.from_pretrained(pretrained_language_model_name_or_path, use_fast=False)
if not tokenizer.pad_token_id:
tokenizer.pad_token_id = tokenizer.eos_token_id
processor = AutoProcessor.from_pretrained(pretrained_vision_model_name_or_path)
print("preparing datasets...")
train_ds, eval_ds = build_dataset(
pretrained_language_model_name_or_path=pretrained_language_model_name_or_path,
pretrained_vision_model_name_or_path=pretrained_vision_model_name_or_path,
tokenizer=tokenizer,
processor=processor,
chat_utterance_max_num=20,
image_caption_sample_max_len=128,
image_caption_block_max_len=192,
instruction_following_sample_max_len=1024,
instruction_following_block_max_len=1024,
chat_sample_max_len=1024,
chat_block_max_len=1024,
num_image_caption_train_samples=100000,
num_image_caption_eval_samples=5000,
num_instruction_following_train_blocks=50000,
num_instruction_following_eval_blocks=5000,
num_chat_train_blocks=50000,
num_chat_eval_blocks=5000,
)
print("preparing model...")
model = ClipAdaptionPromptV2ForMultiModalConditionalGeneration.build_model_for_train(
pretrained_language_model_name_or_path=pretrained_language_model_name_or_path,
pretrained_vision_model_name_or_path=pretrained_vision_model_name_or_path,
hf_train_args=train_args,
adaption_prompt_v2_config=AdaptionPromptV2Config(
peft_type=PeftType.ADAPTION_PROMPT_V2,
task_type=TaskType.CAUSAL_LM,
adapter_len=10,
adapter_layers=30,
add_bias=True,
add_scale=True,
multi_modal=True,
supported_modals=["vision"]
),
language_model_loading_kwargs={"load_in_8bit": False, "low_cpu_mem_usage": True, "device_map": "auto", "torch_dtype": torch.float16},
vision_model_loading_kwargs={"load_in_8bit": False, "low_cpu_mem_usage": True, "torch_dtype": torch.float16},
)
print("preparing trainer...")
trainer = Trainer(
model=model,
args=train_args,
train_dataset=train_ds,
eval_dataset=eval_ds,
data_collator=partial(collate_data, pad_token_id=tokenizer.pad_token_id)
)
print(f"{train_args.parallel_mode=}")
print(f"{train_args.per_device_train_batch_size=}")
print(f"{train_args.gradient_accumulation_steps=}")
print(f"epoch_steps={len(trainer.get_train_dataloader())}")
print("training...")
train_res = trainer.train()
train_metrics = train_res.metrics
train_ppl = math.exp(train_metrics["train_loss"])
print("evaluating...")
eval_metrics = trainer.evaluate()
eval_ppl = math.exp(eval_metrics["eval_loss"])
print(f"{train_ppl=}\t{eval_ppl=}")
model.save_pretrained(train_args.output_dir)
if __name__ == "__main__":
train()