-
Notifications
You must be signed in to change notification settings - Fork 0
/
pretrain.py
200 lines (166 loc) · 7.53 KB
/
pretrain.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
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
import argparse
import time
import numpy as np
from PIL import ImageFile
ImageFile.LOAD_TRUNCATED_IMAGES = True
import torch.multiprocessing
import torchvision.transforms as transforms
from torch.nn.parallel import DataParallel
torch.multiprocessing.set_sharing_strategy('file_system')
from utils.utils import *
from utils.datasets import *
from utils.yaml_config_hook import *
from utils.save_model import *
from sync_batchnorm import convert_model
from modules import resnet, network, contrastive_loss
from modules.Transform import Transform
from utils.common import *
from utils.model_utils import *
from modules.transforms import build_transform#,ImageFolder
def train(args, scaler):
loss_epoch = 0
for step, (x, _) in enumerate(data_loader):
x0 = x[0].to(args.default_device)
x1 = x[1].to(args.default_device)
optimizer.zero_grad()
with torch.cuda.amp.autocast(True):
z_i, z_i_groups = model(x0)
z_j, z_j_groups = model(x1)
loss = criterion_instance([z_i, z_j])
if args.do_ca:
ca_loss = CAM(z_i_groups, z_j_groups, args.ca_t, args)
loss_ins_ca = loss + args.Lambda * (ca_loss)
if args.do_ca:
scaler.scale(loss_ins_ca).backward(retain_graph=True)
else:
if args.do_adv:
scaler.scale(loss).backward(retain_graph=True)
else:
scaler.scale(loss).backward()
if args.do_adv:
advm.attack()
with torch.cuda.amp.autocast(True):
z_i_adv, z_i_groups_adv = model(x0)
loss_instance_adv = criterion_instance([z_j, z_i_adv])
advm.restore()
scaler.scale(loss_instance_adv).backward()
scaler.step(optimizer)
scaler.update()
model.zero_grad()
if args.do_adv:
if step % 10 == 0:
if args.do_ca:
print(
f"Step [{step}/{len(data_loader)}]\t loss_instance: {loss.item()}",
f"Step [{step}/{len(data_loader)}]\t loss_adv: {loss_instance_adv.item()}",
f"Step [{step}/{len(data_loader)}]\t ca_loss: {loss_ins_ca.item()}")
else:
print(
f"Step [{step}/{len(data_loader)}]\t loss_instance: {loss.item()}",
f"Step [{step}/{len(data_loader)}]\t loss_adv: {loss_instance_adv.item()}")
else:
if step % 50 == 0:
print(
f"Step [{step}/{len(data_loader)}]\t loss_instance: {loss.item()}")
if args.do_adv:
if args.do_ca:
loss_epoch += loss_ins_ca.item() + loss_instance_adv.item()
else:
loss_epoch += loss.item() + loss_instance_adv.item()
else:
loss_epoch = loss.item()
return loss_epoch
if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument('--use_redis', action='store_true')
parser.add_argument('--default_device', type=str, default='cuda', help='default_device')
parser.add_argument('--arch', type=str, default='resnet18', help='model architecture')
parser.add_argument('--comet_name', type=str, default='pretrain',
help='对本次实验的命名, 用于comet.ml')
config = yaml_config_hook("config/config.yaml")
for k, v in config.items():
parser.add_argument(f"--{k}", default=v, type=type(v))
args = parser.parse_args()
args.model_path = f"save/{args.dataset}/pretrain/{args.resnet}_bs_{args.batch_size}_img_size_{args.image_size}_lr_{args.learning_rate}_\
wd_{args.weight_decay}_temperature_{args.instance_temperature}_adv_eps_{args.epsilon}_ca_{args.Lambda}_strong_weak"
if not os.path.exists(args.model_path):
os.makedirs(args.model_path)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
torch.cuda.manual_seed(args.seed)
np.random.seed(args.seed)
transform = Transform(init_size=224)
dataset = CRC_Dataset(
dataset_path=args.dataset_path,
transform=transform
)
print('dataset:', len(dataset))
data_loader = torch.utils.data.DataLoader(
dataset,
batch_size=args.batch_size,
shuffle=True,
drop_last=True,
num_workers=args.workers,
pin_memory=True
)
# initialize model
res = resnet.get_resnet(args.resnet)
model = network.Network(res, args.feature_dim)
load_imageweights = True
if load_imageweights:
if args.resnet == 'ResNet18' and not os.path.isfile('./save/resnet18-f37072fd.pth'):
print('Downloading imagenet weights...')
os.system('wget https://download.pytorch.org/models/resnet18-f37072fd.pth')
if args.resnet == 'ResNet34' and not os.path.isfile('./save/resnet34-333f7ec4.pth'):
print('Downloading imagenet weights...')
os.system('wget https://download.pytorch.org/models/resnet34-333f7ec4.pth')
if args.resnet == 'ResNet50' and not os.path.isfile('./save/resnet50-19c8e357.pth'):
print('Downloading imagenet weights...')
os.system('wget https://download.pytorch.org/models/resnet50-19c8e357.pth')
if args.resnet == 'ResNet18':
print('Loading imagenet ResNet18 weights...')
checkpoint = torch.load('./save/resnet18-f37072fd.pth', map_location='cpu')
elif args.resnet == 'ResNet34':
print('Loading imagenet ResNet34 weights...')
checkpoint = torch.load('./save/resnet34-333f7ec4.pth', map_location='cpu')
elif args.resnet == 'ResNet50':
print('Loading imagenet ResNet50 weights...')
checkpoint = torch.load('./save/resnet50-19c8e357.pth', map_location='cpu')
else:
raise NotImplementedError
model_dict = model.state_dict()
def rename_key(key):
if not 'resnet' in key:
return 'resnet.' + key
return key
checkpoint_dict = {}
for key, val in checkpoint.items():
checkpoint_dict[rename_key(key)] = val
model_dict.update(checkpoint_dict)
missing_keys, unexpected_keys = model.load_state_dict(model_dict, strict=False)
print(f'missing keys:{missing_keys}, unexpected_keys:{unexpected_keys}')
dataparallel = 1
if dataparallel:
model = convert_model(model)
model = DataParallel(model)
model = model.to(args.default_device)
print('load the model successfully')
# optimizer / loss
scaler = torch.cuda.amp.GradScaler()
optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate, weight_decay=args.weight_decay)
# loss_device = torch.device("cuda")
criterion_instance = contrastive_loss.InstanceLoss(args.batch_size, args.instance_temperature, args.default_device).to(
args.default_device)
print('load the criterion_instance successfully')
if args.do_adv:
advm = AdvModule(model, dataparallel, epsilon=args.epsilon, emb_name='resnet.conv1.weight')
print('start training')
for epoch in range(args.start_epoch, args.epochs):
start_ts = time.time()
lr = optimizer.param_groups[0]["lr"]
adjust_learning_rate(optimizer, epoch, args)
loss_epoch = train(args, scaler)
if epoch % 10 == 0:
save_model(args, model, optimizer, epoch)
print(f"Epoch [{epoch}/{args.epochs}]\t Loss: {loss_epoch / len(data_loader)}, cost:{time.time() - start_ts}")
save_model(args, model, optimizer, args.epochs)