-
Notifications
You must be signed in to change notification settings - Fork 3
/
train_evcap.py
176 lines (162 loc) · 6.59 KB
/
train_evcap.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
import os
import torch
from torch.utils.data import DataLoader, DistributedSampler
import random
import sys
import argparse
import numpy as np
import utils
from optims import LinearWarmupCosineLRScheduler, set_optimizer
from dataset.coco_dataset import COCODataset
from models.evcap import EVCap
from common.dist_utils import (
get_rank,
init_distributed_mode,
get_world_size,
)
def set_seed(seed: int) -> None:
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
def save_checkpoint(model,optimizer, cur_epoch, output_dir):
"""
Save the checkpoint at the current epoch.
"""
model_no_ddp = model
param_grad_dic = {
k: v.requires_grad for (k, v) in model_no_ddp.named_parameters()
}
state_dict = model_no_ddp.state_dict()
for k in list(state_dict.keys()):
if k in param_grad_dic.keys() and not param_grad_dic[k]:
del state_dict[k]
save_obj = {
"model": state_dict,
"optimizer": optimizer.state_dict(),
"epoch": cur_epoch,
}
print("Saving checkpoint at epoch {} to {}.".format(cur_epoch, output_dir))
torch.save(save_obj, output_dir)
def train(dataset, model, args):
device = torch.device(f"cuda:{get_rank()}")
batch_size = args.bs
epochs = args.epochs
accum_grad_iters = 1
output_dir = args.out_dir
if not os.path.exists(output_dir):
os.makedirs(output_dir, exist_ok=True)
if args.distributed:
sampler = DistributedSampler(
dataset,
shuffle=True,
num_replicas=get_world_size(),
rank=get_rank(),
)
model = model.to(device)
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[get_rank()])
else:
sampler = None
train_dataloader = DataLoader(dataset, batch_size=batch_size, pin_memory=True, sampler=sampler,shuffle=False, drop_last=True)
model.train()
optimizer = set_optimizer(model, init_lr=1e-4, weight_decay=0.05)
scheduler = LinearWarmupCosineLRScheduler(optimizer= optimizer,
max_epoch=epochs,
iters_per_epoch=len(train_dataloader),
min_lr=8e-5,
init_lr=1e-4,
decay_rate=None,
warmup_start_lr=1e-6,
warmup_steps=5000,)
if args.amp:
scaler = torch.cuda.amp.GradScaler()
use_amp = scaler is not None
print('use_amp', use_amp)
for epoch in range(epochs):
print(f">>> Training epoch {epoch}")
sys.stdout.flush()
metric_logger = utils.MetricLogger(delimiter=" ")
metric_logger.add_meter('lr', utils.SmoothedValue(window_size=50, fmt='{value:.6f}'))
metric_logger.add_meter('loss', utils.SmoothedValue(window_size=50, fmt='{value:.6f}'))
metric_logger.update(loss=1000.0)
metric_logger.update(lr=optimizer.param_groups[0]["lr"])
print_freq = 50
header = 'Train Epoch: [{}]'.format(epoch)
for idx, samples in enumerate(metric_logger.log_every(train_dataloader, print_freq, header)):
samples['image'] = samples['image'].to(device)
scheduler.step(cur_epoch=epoch, cur_step=idx)
with torch.cuda.amp.autocast(enabled=use_amp):
loss = model(samples)["loss"]
if use_amp:
scaler.scale(loss).backward()
else:
loss.backward()
if (idx + 1) % accum_grad_iters == 0:
if use_amp:
scaler.step(optimizer)
scaler.update()
else:
optimizer.step()
optimizer.zero_grad()
metric_logger.update(loss=loss.item())
metric_logger.update(lr = optimizer.param_groups[0]["lr"])
metric_logger.synchronize_between_processes()
print("Averaged stats:", metric_logger.global_avg())
if epoch == epochs - 1:
output_dir_model = os.path.join(output_dir, f"{epoch:03d}.pt")
save_checkpoint(model, optimizer, epoch, output_dir_model)
return model
def main():
os.environ["TOKENIZERS_PARALLELISM"] = "false"
print('Starts ...')
print(" # PID :", os.getpid())
parser = argparse.ArgumentParser()
parser.add_argument('--out_dir', default='./checkpoints')
parser.add_argument('--epochs', type=int, default=1)
parser.add_argument('--bs', type=int, default=6)
parser.add_argument('--is_rn', dest='is_rn', action='store_true')
parser.add_argument('--device', default = 'cuda', help = 'gpu for training')
parser.add_argument('--distributed', default = True)
parser.add_argument('--amp', default = True)
parser.add_argument('--dist_url', default = "env://")
parser.add_argument('--world_size', type = int, default = 1)
parser.add_argument('--num_query_token_txt', type = int, default = 8)
parser.add_argument('--topn', type = int, default = 9)
parser.add_argument('--disable_random_seed', action = 'store_true', default = False, help = 'set random seed for reproducing')
parser.add_argument('--random_seed', type = int, default = 42, help = 'set random seed for reproducing')
args = parser.parse_args()
print(f'args: {vars(args)}')
if not args.disable_random_seed:
set_seed(args.random_seed)
init_distributed_mode(args)
print(f'args: {vars(args)}')
data_root = 'data/coco/coco2014'
dataset = COCODataset(data_root=data_root)
model_type = "vicuna-13b-v1.3"
model = EVCap(
ext_path = 'ext_data/ext_memory_lvis.pkl',
vit_model="eva_clip_g",
q_former_model="https://storage.googleapis.com/sfr-vision-language-research/LAVIS/models/BLIP2/blip2_pretrained_flant5xxl.pth",
img_size=224,
drop_path_rate=0,
use_grad_checkpoint=False,
vit_precision="fp16",
freeze_vit=True,
freeze_qformer=True,
num_query_token=32,
num_query_token_txt=args.num_query_token_txt,
topn = args.topn,
llama_model=model_type,
prompt_path="prompts/prompt_evcap.txt",
prompt_template='###Human: {} ###Assistant: ',
max_txt_len=128,
end_sym='\n',
low_resource=False, # use 8 bit and put vit in cpu
device_8bit=0, # the device of 8bit model should be set when loading and cannot be changed anymore.
)
train(dataset, model, args)
if __name__ == '__main__':
main()