-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_Strace.py
215 lines (158 loc) · 8.33 KB
/
train_Strace.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
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
import os, sys
import os.path as osp
os.environ['CUDA_VISIBLE_DEVICES'] = '1'
root_dir = os.path.abspath(os.getcwd())
sys.path.append(root_dir)
import model
import utils
from dataset import StegaData, StegaData_pos
from FSNet import FSNet256
import yaml, shutil
import random
import numpy as np
from glob import glob
from easydict import EasyDict
from PIL import Image, ImageOps
from torch import optim
import torch, cv2
from torch.utils.data import DataLoader
import time, dlib
random.seed(20)
np.random.seed(20)
torch.set_num_threads(1)
with open('{}/00_setting.yaml'.format(root_dir), 'r') as f:
args = EasyDict(yaml.load(f, Loader=yaml.SafeLoader))
def main():
dataset_train = StegaData_pos(args.train_path, args.secret_size, size=(256, 256), args=args, is_train=True, is_bio=True, no_mask=True)
dataset = dataset_train
dataloader_train = DataLoader(dataset_train, batch_size=args.batch_size, shuffle=True, num_workers=4, pin_memory=True)
dataset_test = StegaData_pos(args.test_path, args.secret_size, size=(256, 256), args=args, is_train=True, is_bio=True, no_mask=True)
dataloader_test = DataLoader(dataset_test, batch_size=args.batch_size, shuffle=False, num_workers=4, pin_memory=True)
encoder = model.StegaStampEncoder(secret_size=args.secret_size, args=args)
decoder = model.StegaStampDecoderPrior(secret_size=args.secret_size, binaryy=args.binaryy, args=args)
if args.cuda:
encoder = encoder.cuda()
decoder = decoder.cuda()
de_vars = decoder.parameters()
en_vars = encoder.parameters()
ende_vars = [{'params': encoder.parameters()},
{'params': decoder.parameters()}]
optimize_en = optim.Adam(en_vars, lr=args.lr)
optimize_ende = optim.Adam(ende_vars, lr=args.lr)
optimize_de = optim.Adam(de_vars, lr=args.lr)
global_step = 0
total_steps = len(dataset) // args.batch_size + 1
best_acc, best_iter, best_auc = 0, 0, 0
vae = FSNet256().cuda()
vae.load_ckpt(args.vae_ckpt_path)
vae.eval()
if args.setimg_mode:
if args.set_img_path is not None:
img_cv = cv2.imread(args.set_img_path)
img_torch = torch.from_numpy(img_cv).float()
img_torch = img_torch.permute((2, 0, 1))
img_torch /= 255.
img_torch = img_torch.unsqueeze(0)
set_image_o = img_torch.repeat(args.batch_size, 1, 1, 1).cuda()
else:
print('===========no set img!')
else:
print('===========no set img!')
while global_step < args.num_steps:
for _ in range(min(total_steps, args.num_steps - global_step)):
encoder = encoder.train()
img_gau, _, secret_input = next(iter(dataloader_train))
img_gau = img_gau.cuda()
secret_input = secret_input.cuda()
if args.setimg_mode:
set_image = set_image_o.clone()
else:
set_image = None
global_step += 1
no_im_loss = global_step < args.no_im_loss_steps
l2_edge_gain = 0
if global_step > args.l2_edge_delay:
l2_edge_gain = min(args.l2_edge_gain * (global_step-args.l2_edge_delay) / args.l2_edge_ramp, args.l2_edge_gain)
scale_loss1_img_rec = min(args.loss1_img_rec * global_step / (args.loss1_img_rec_ramp+args.no_im_loss_steps), args.loss1_img_rec)
scale_loss2_lpips = min(args.loss2_lpips * global_step / (args.loss2_lpips_ramp+args.no_im_loss_steps), args.loss2_lpips)
scale_loss3_binary = min(args.loss3_binary * global_step / args.loss3_binary_ramp, args.loss3_binary)
loss_scales = [scale_loss1_img_rec, scale_loss2_lpips, scale_loss3_binary]
meta = model.self_1019_build_model(secret_input, encoder, decoder, img_gau, args, global_step, vae, l2_edge_gain, test_mode=False, set_img=set_image, no_add_noise=False, mismatch=False)
if no_im_loss:
optimize_ende.zero_grad()
meta['loss3_binary'].backward()
optimize_en.step()
else:
optimize_ende.zero_grad()
loss_all = scale_loss1_img_rec * meta['loss1_img_rec'] + scale_loss2_lpips * meta['loss2_lpips'] + scale_loss3_binary * meta['loss3_binary']
loss_all.backward()
optimize_en.step()
meta = model.self_1019_build_model(secret_input, encoder, decoder, img_gau, args, global_step, vae, l2_edge_gain, test_mode=False, set_img=set_image, no_add_noise=False, mismatch=True)
if no_im_loss:
optimize_ende.zero_grad()
meta['loss3_binary'].backward()
optimize_en.step()
else:
optimize_ende.zero_grad()
loss_all = scale_loss1_img_rec * meta['loss1_img_rec'] + scale_loss2_lpips * meta['loss2_lpips'] + scale_loss3_binary * meta['loss3_binary']
loss_all.backward()
optimize_en.step()
meta = model.self_1019_build_model(secret_input, encoder, decoder, img_gau, args, global_step, vae, l2_edge_gain, test_mode=False, set_img=set_image, no_add_noise=False, mismatch=False)
optimize_de.zero_grad()
optimize_en.zero_grad()
meta['loss3_binary'].backward()
optimize_de.step()
meta = model.self_1019_build_model(secret_input, encoder, decoder, img_gau, args, global_step, vae, l2_edge_gain, test_mode=False, set_img=set_image, no_add_noise=False, mismatch=True)
optimize_de.zero_grad()
optimize_en.zero_grad()
meta['loss3_binary'].backward()
optimize_de.step()
if global_step % 1000 == 1:
os.makedirs(args.saved_models, exist_ok=True)
encoder = encoder.eval()
decoder = decoder.cuda()
num_all = 0.0
acc_bit = 0.0
acc_bit_mismatch = 0.0
for img_gau, _, secret_input in dataloader_test:
if args.setimg_mode:
set_image = set_image_o.clone()
else:
set_image = None
img_gau = img_gau.cuda()
secret_input = secret_input.cuda()
meta = model.self_1019_build_model(secret_input, encoder, decoder, img_gau, args, global_step, vae, l2_edge_gain, test_mode=True, set_img=set_image, no_add_noise=False, mismatch=False)
acc_bit += meta['bit_acc']
meta = model.self_1019_build_model(secret_input, encoder, decoder, img_gau, args, global_step, vae, l2_edge_gain, test_mode=True, set_img=set_image, no_add_noise=False, mismatch=True)
acc_bit_mismatch += meta['bit_acc']
num_all += 1
acc_bit_all = acc_bit / num_all
acc_bit_all_mismatch = acc_bit_mismatch / num_all
oneline = 'iter:{}, acc_bit:{:0>3f}, acc_bit_mismatch:{:0>3f}, loss_scales:{}\n'.format(global_step, acc_bit_all, acc_bit_all_mismatch, loss_scales)
with open(args.saved_models+'/acc.txt','a') as f:
f.write(oneline)
print(oneline)
print(args.saved_models)
if best_acc < acc_bit_all and global_step > 10000:
best_acc = acc_bit_all
best_iter = global_step
if acc_bit_all < 0.89:
args.loss3_binary *= 1.25
else:
args.loss1_img_rec *= 1.25
args.loss2_lpips *= 1.25
gopt_path = args.saved_models + '/gopt_{:07d}.pth'.format(global_step)
en_path = args.saved_models + '/encoder_{:07d}.pth'.format(global_step)
de_path = args.saved_models + '/decoder_{:07d}.pth'.format(global_step)
torch.save({
'optimize_ende': optimize_ende.state_dict()},
gopt_path)
torch.save({
'state_dict': encoder.state_dict()},
en_path)
torch.save({
'state_dict': decoder.state_dict()},
de_path)
print('====>save at {} in {}'.format(global_step, args.saved_models))
if __name__ == '__main__':
main()