-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathtrain_alignment.py
177 lines (152 loc) · 8.32 KB
/
train_alignment.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
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
import os
import math
import pathlib
from typing import Optional, Dict, List
from dataclasses import dataclass, field
import json
import torch
from lightning.pytorch import LightningModule, Trainer, seed_everything
from lightning.pytorch.callbacks import ModelCheckpoint
from lightning.pytorch.loggers import WandbLogger
from torch.utils.data import DataLoader, random_split
import transformers
from transformers.training_args import TrainingArguments
from torchvision import transforms
from models.sd_condition import SD_Condition
from dataset import AlignmentDataset
from dataset import AligmentDataCollator
def default_gpus():
return [0,1,2,3]
@dataclass
class ModelArguments:
llm_model_name_or_path: Optional[str] = field(default="baichuan-inc/Baichuan2-7B-Chat")
sd_model_name_or_path: Optional[str] = field(default="stabilityai/stable-diffusion-xl-base-1.0")
model_save_name: str = field(default='model_{epoch}-{step}')
@dataclass
class DataArguments:
en_train_data_path: str = field(
default=None, metadata={"help": "Path to the english training data."}
)
ch_train_data_path: str = field(
default=None, metadata={"help": "Path to the chinese training data."}
)
val_data_path: str = field(
default=None, metadata={"help": "Path to the validation data."}
)
test_data_path: str = field(default=None, metadata={"help": "Path to the test data."})
training_length: int = field(default=None)
test_length: int = field(default=10000)
@dataclass
class TrainingArguments:
seed: int = field(default=42)
cache_dir: Optional[str] = field(default=None)
output_dir: str = field(default='results')
num_train_epochs: int = field(default=2)
per_device_train_batch_size:int = field(default=2)
per_device_eval_batch_size:int = field(default=2)
real_batch_size:int = field(default=48)
learning_rate:float = field(default=2e-5)
warmup_steps:int = field(default=1000)
adam_epsilon:float = field(default=1e-8)
gpus: List[int] = field(default_factory=default_gpus)
num_nodes: int = field(default=1)
num_workers: int = field(default=16)
strategy: str = field(default='ddp')
val_check_interval: float = field(default=0.1)
stage1_weight: Optional[str] = field(default=None)
caption_loss: str = field(default='mse')
resume: Optional[str] = field(default=None)
wandb_project_name: str = field(default="LLM_SD")
wandb_api_key: str = field(default=None)
def make_supervised_data_module(data_args, training_args, data_collator, tokenizer) -> Dict:
"""Make dataset and collator for multilingual textual aligment."""
test_length = data_args.test_length
en_data_paths = [item.strip() for item in data_args.en_train_data_path.split(',')]
ch_data_paths = [item.strip() for item in data_args.ch_train_data_path.split(',')]
en_datasets, ch_datasets = [], []
for data_path in en_data_paths:
en_dataset = AlignmentDataset(data_path=data_path, tokenizer=tokenizer, language='en')
en_datasets.append(en_dataset)
for data_path in ch_data_paths:
ch_dataset = AlignmentDataset(data_path=data_path, tokenizer=tokenizer, language='ch')
ch_datasets.append(ch_dataset)
en_ch_probs = [0.5, 0.5]
train_dataloaders, val_datasets = [], []
if en_ch_probs[1] > 1e-3: # multilingual
for dataset in en_datasets + ch_datasets:
train_bs = int(training_args.per_device_train_batch_size * en_ch_probs[0] / len(en_dataset)) if dataset.language == 'en' \
else int(training_args.per_device_train_batch_size * en_ch_probs[1] / len(ch_datasets))
if train_bs < 1: train_bs = 1
dataset_test_len = int(test_length * en_ch_probs[0] / len(en_datasets)) if dataset.language == 'en' \
else int(test_length * en_ch_probs[1] / len(ch_datasets))
train_dataset, eval_dataset = random_split(dataset, lengths=[len(dataset)-dataset_test_len, dataset_test_len], generator=torch.Generator().manual_seed(3407))
val_datasets.append(eval_dataset)
train_dataloader = DataLoader(train_dataset,
batch_size=train_bs,
num_workers=training_args.num_workers,
collate_fn=data_collator,
prefetch_factor=4,
pin_memory=False)
train_dataloaders.append(train_dataloader)
else: # eng only
for dataset in en_datasets:
train_bs = int(training_args.per_device_train_batch_size * en_ch_probs[0] / len(en_dataset))
if train_bs < 1: train_bs = 1
dataset_test_len = int(test_length * en_ch_probs[0] / len(en_datasets)) if dataset.language == 'en' \
else int(test_length * en_ch_probs[1] / len(ch_datasets))
train_dataset, eval_dataset = random_split(dataset, lengths=[len(dataset)-dataset_test_len, dataset_test_len], generator=torch.Generator().manual_seed(3407))
val_datasets.append(eval_dataset)
train_dataloader = DataLoader(train_dataset,
batch_size=train_bs,
num_workers=training_args.num_workers,
collate_fn=data_collator,
prefetch_factor=4,
pin_memory=False)
train_dataloaders.append(train_dataloader)
combined_val_dataset = torch.utils.data.ConcatDataset(val_datasets)
val_dataloader = DataLoader(combined_val_dataset,
batch_size=training_args.per_device_eval_batch_size,
num_workers=training_args.num_workers,
collate_fn=data_collator,
prefetch_factor=4,
pin_memory=True)
print("Data Loading Finished")
return train_dataloaders, val_dataloader
if __name__ == "__main__":
parser = transformers.HfArgumentParser(
(ModelArguments, DataArguments, TrainingArguments)
)
model_args, data_args, training_args = parser.parse_args_into_dataclasses()
seed_everything(training_args.seed)
torch.backends.cuda.matmul.allow_tf32 = True
os.environ["WANDB_API_KEY"] = training_args.wandb_api_key
batch_size = training_args.real_batch_size
devices = training_args.gpus
num_devices = len(devices) * training_args.num_nodes
gradient_accumulation_steps = max(1, batch_size // (training_args.per_device_train_batch_size*num_devices))
if training_args.stage1_weight: # load a pretrained ckpt and finetune on it
assert training_args.stage1_weight is not None
model = SD_Condition.load_from_checkpoint(training_args.stage1_weight, strict=False, map_location="cpu", model_args=model_args, **vars(training_args))
else: # train from scratch
model = SD_Condition(model_args, **vars(training_args))
tokenizer = model.llm_tokenizer
data_collator = AligmentDataCollator(tokenizer)
train_dataloaders, val_dataloaders = make_supervised_data_module(data_args, training_args, data_collator, tokenizer)
checkpoint_callback = ModelCheckpoint(
dirpath=training_args.output_dir,
filename=model_args.model_save_name,
monitor="val_loss",
save_top_k=-1,
)
wandb_logger = WandbLogger(save_dir=training_args.output_dir, project=training_args.wandb_project_name, offline=True, name=model_args.model_save_name)
trainer = Trainer(default_root_dir=training_args.output_dir, max_epochs=training_args.num_train_epochs,
accumulate_grad_batches=gradient_accumulation_steps,
accelerator="gpu", devices=devices,
num_nodes=training_args.num_nodes,
strategy=training_args.strategy,
logger=wandb_logger,
precision='bf16-mixed',
val_check_interval=training_args.val_check_interval,
num_sanity_val_steps=0,
callbacks=[checkpoint_callback])
trainer.fit(model, train_dataloaders, val_dataloaders, ckpt_path=training_args.resume)