-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_onecycle.py
159 lines (141 loc) · 6.41 KB
/
train_onecycle.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
"""
Unofficial code for VPT(Visual Prompt Tuning) paper of arxiv 2203.12119
A toy Tuning process that demostrates the code
the code is based on timm
"""
import torch
import torch.nn as nn
import torch.optim as optim
import numpy as np
from tqdm import tqdm
from PromptModels.GetPromptModel import build_promptmodel
import argparse
from torch.optim.lr_scheduler import OneCycleLR
import torchvision
import torchvision.transforms as transforms
from torch.utils.data.dataloader import DataLoader
from timm.scheduler import CosineLRScheduler
from datasets import create_datasets
from utils import *
parser = argparse.ArgumentParser()
parser.add_argument('--dataset', type=str, default='cifar100')
parser.add_argument('--split', type=str, default='1000')
parser.add_argument('--data_path', type=str, default='data/')
parser.add_argument('--batch_size', type=int, default=128)
parser.add_argument('--base_lr', type=float, default=0.01)
parser.add_argument('--lr', type=float, default=0.005)
parser.add_argument('--momentum', type=float, default=0.9)
parser.add_argument('--weight_decay', type=float, default=0.01)
parser.add_argument('--image_size', type=int, default=224)
parser.add_argument('--epochs', type=int, default=100)
parser.add_argument('--warmup_epochs', type=int, default=10)
parser.add_argument('--optimizer', type=str, default='sgd')
parser.add_argument('--scheduler', type=str, default='cosine')
parser.add_argument('--prompt_length', type=int, default=10)
parser.add_argument('--name', type=str, default='vpt_deep')
parser.add_argument('--base_model', type=str, default='vit_base_patch16_224_in21k')
def setup_seed(seed): # setting up the random seed
import random
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
random.seed(seed)
torch.backends.cudnn.deterministic = True
def main(args):
setup_seed(42)
save_path = os.path.join('./save', args.name)
ensure_path(save_path)
set_log_path(save_path)
device = 'cuda' if torch.cuda.is_available() else 'cpu'
args.lr = args.base_lr * args.batch_size / 256
# labels = torch.ones(batch_size).long() # long ones
norm_params = {'mean': [0.485, 0.456, 0.406],
'std': [0.229, 0.224, 0.225]}
normalize = transforms.Normalize(**norm_params)
train_transforms = transforms.Compose([
transforms.RandomResizedCrop(args.image_size),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
normalize,
])
val_transforms = transforms.Compose([
transforms.Resize((args.image_size * 8 // 7, args.image_size * 8 // 7)),
transforms.CenterCrop((args.image_size, args.image_size)),
transforms.ToTensor(),
normalize,
])
train_dataset, val_dataset, num_classes = create_datasets(args.data_path, train_transforms, val_transforms, args.dataset, args.split)
log(f"train dataset: {len(train_dataset)} samples")
log(f"val dataset: {len(val_dataset)} samples")
train_loader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=8, pin_memory=True)
val_loader = DataLoader(val_dataset, batch_size=int(args.batch_size), shuffle=False)
model = build_promptmodel(num_classes=num_classes, img_size=args.image_size, base_model=args.base_model, model_idx='ViT', patch_size=16,
Prompt_Token_num=args.prompt_length, VPT_type="Deep") # VPT_type = "Shallow"
# test for updating
prompt_state_dict = model.obtain_prompt()
model.load_prompt(prompt_state_dict)
model = model.to(device)
optimizer = optim.SGD(model.parameters(), lr=args.lr, weight_decay=args.weight_decay, momentum=args.momentum)
steps_per_epoch = len(train_loader)
scheduler = OneCycleLR(optimizer, max_lr=args.lr, steps_per_epoch=steps_per_epoch, epochs=args.epochs, pct_start=0.2)
criterion = nn.CrossEntropyLoss()
# preds = model(data) # (1, class_number)
# print('before Tuning model output：', preds)
# check backwarding tokens
for param in model.parameters():
if param.requires_grad:
print(param.shape)
max_va = -1
for epoch in range(args.epochs):
print('epoch:',epoch)
aves_keys = ['tl', 'ta', 'vl', 'va']
aves = {k: Averager() for k in aves_keys}
iter_num = 0
model.Freeze()
for imgs, targets in tqdm(train_loader, desc='train', leave=False):
imgs = imgs.to(device)
targets = targets.to(device)
optimizer.zero_grad()
outputs = model(imgs)
loss = criterion(outputs, targets)
loss.backward()
optimizer.step()
scheduler.step()
acc = compute_acc(outputs, targets)
aves['tl'].add(loss.item())
aves['ta'].add(acc)
iter_num += 1
# print()
model.eval()
for imgs, targets in tqdm(val_loader, desc='val', leave=False):
imgs = imgs.to(device)
targets = targets.to(device)
outputs = model(imgs)
loss = criterion(outputs, targets)
acc = compute_acc(outputs, targets)
aves['vl'].add(loss.item())
aves['va'].add(acc)
log_str = 'epoch {}, lr: {:.4f}, train loss: {:.4f}|acc: {:.4f}'.format(
epoch, scheduler.get_last_lr()[0], aves['tl'].v, aves['ta'].v)
log_str += ', val loss: {:.4f}|acc: {:.4f}'.format(aves['vl'].v, aves['va'].v)
log(log_str)
# preds = model(data) # (1, class_number)
print('After Tuning model output: ', aves['va'].v)
save_obj = {
'config': vars(args),
'state_dict': model.state_dict(),
'val_acc': aves['va'].v,
}
if epoch <= args.epochs:
torch.save(save_obj, os.path.join(save_path, 'epoch-last.pth'))
torch.save(save_obj, os.path.join(
save_path, 'epoch-{}.pth'.format(epoch)))
if aves['va'].v > max_va:
max_va = aves['va'].v
torch.save(save_obj, os.path.join(save_path, 'max-va.pth'))
else:
torch.save(save_obj, os.path.join(save_path, 'epoch-ex.pth'))
# scheduler.step(epoch)
if __name__ == "__main__":
args = parser.parse_args()
main(args)