-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_Decomposition_angle.py
264 lines (227 loc) · 13.6 KB
/
train_Decomposition_angle.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
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
import argparse
import torchvision as tv
import torchvision.utils as utils
from torch.utils.data import DataLoader
from tensorboardX import SummaryWriter
from lib.model import *
from lib.dataset import MyData_cityscape
from lib.utils import *
from lib.loss_function import *
from matplotlib import pyplot as plt
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
device_ids = [0]
parser = argparse.ArgumentParser(description="UDGNet")
parser.add_argument("--batchSize", type=int, default=8, help="Training batch size")
parser.add_argument("--epochs", type=int, default=150, help="Number of training epochs")
parser.add_argument("--milestone", type=list, default=[30, 60, 80], help="When to decay learning rate;")
parser.add_argument("--lr", type=float, default=1e-4, help="Initial learning rate")
parser.add_argument("--outf", type=str, default="output", help='path of log files')
parser.add_argument("--rain_path",type=str,default="./dataset/test/rain",help='path of rain dataset')
parser.add_argument("--angle_path",type=str,default="./dataset/train/angle",help='path of angle path')
parser.add_argument("--clean_path",type=str,default="./dataset/train/clean",help='path of clean path')
parser.add_argument("--reset", type=int, default=1, help='path of dataset')
opt = parser.parse_args()
transform = tv.transforms.Compose(
[#tv.transforms.Resize(128),
tv.transforms.ToTensor()
# tv.transforms.Normalize(mean=[0.5,0.5,0.5],std=[0.5,0.5,0.5])
])
def set_requires_grad(nets, requires_grad=False):
"""Set requies_grad=Fasle for all the networks to avoid unnecessary computations
Parameters:
nets (network list) -- a list of networks
requires_grad (bool) -- whether the networks require gradients or not
"""
if not isinstance(nets, list):
nets = [nets]
for net in nets:
if net is not None:
for param in net.parameters():
param.requires_grad = requires_grad
def main():
# Load dataset
print('Loading dataset .../n')
repeat = 2e3
dataset_train = MyData_cityscape(opt.rain_path,opt.clean_path,opt.angle_path,transforms=transform,patch_size=128,batch_size=opt.batchSize,repeat=repeat,channel=3)
loader_train = DataLoader(dataset=dataset_train, num_workers=4, batch_size=opt.batchSize, shuffle=True)
print("# of training samples: %d/n" % int(len(dataset_train)))
# 定义参数
generator = UDGNet(inchannel=3)
angle_model = angle_estimation(in_channels=3)
discriminator = NLayerDiscriminator(input_nc=3, ndf=64, n_layers=3, norm_layer=nn.BatchNorm2d)
generator.apply(weights_init_kaiming)
angle_model.apply(weights_init_kaiming)
discriminator.apply(weights_init_kaiming)
generator = nn.DataParallel(generator, device_ids=device_ids).cuda()
angle_model = nn.DataParallel(angle_model, device_ids=device_ids).cuda()
discriminator = nn.DataParallel(discriminator, device_ids=device_ids).cuda()
if opt.reset == 1:
generator = init_model(generator)
angle_model = init_model(angle_model)
discriminator = init_model(discriminator)
else:
generator.load_state_dict(torch.load('/home/ubuntu/2TB/YCF/2021YCF_ACM/Decomposition_UTV/output/model2_single_streak/generator_backup.pth'))
angle_model.load_state_dict(torch.load('/home/ubuntu/2TB/YCF/2021YCF_ACM/UTV_Decomposition/output/train_Decomposition_angle_model/angle_model_backup.pth'))
discriminator.load_state_dict(torch.load('/home/ubuntu/2TB/YCF/2021YCF_ACM/Decomposition_UTV/output/model2_single_streak/discriminator_backup.pth'))
fake_pool = ImagePool(50)
class_loss = GANLoss('lsgan')
Prior_criterion1 = nn.L1Loss(reduction='sum')
Prior_criterion2 = TV_Loss()
data_criterion2 = nn.MSELoss(size_average=False)
Prior_criterion1.cuda()
Prior_criterion2.cuda()
data_criterion2.cuda()
class_loss.cuda()
# training
output_folder = opt.outf + '/train_Decomposition_angle_model_cityscape'
if not os.path.exists(output_folder):
os.makedirs(output_folder)
writer = SummaryWriter(output_folder)
step = 0
current_lr = opt.lr
d_optimizer = torch.optim.Adam(discriminator.parameters(), lr=opt.lr,eps=1e-4,amsgrad=True)
g_optimizer = torch.optim.Adam(generator.parameters(), lr=opt.lr, eps=1e-4, amsgrad=True)
angle_optimizer = torch.optim.Adam(angle_model.parameters(), lr=opt.lr, eps=1e-4, amsgrad=True)
d_schedulr = torch.optim.lr_scheduler.MultiStepLR(d_optimizer,milestones=[50,70,100],gamma=0.1,last_epoch = -1)
g_schedulr = torch.optim.lr_scheduler.MultiStepLR(g_optimizer, milestones=[50, 70, 100], gamma=0.1, last_epoch=-1)
angle_schedulr = torch.optim.lr_scheduler.MultiStepLR(angle_optimizer, milestones=[10, 40, 50], gamma=0.1, last_epoch=-1)
discriminator_total_loss = 0
for epoch in range(opt.epochs):
D_loss = 0
G_loss = 0
A_loss = 0
for i, data in enumerate(loader_train): # 可同时获得索引和值
#angle model train
rain2, real_image,angle_label = data['rain_data'].type(torch.FloatTensor).cuda(), \
data['clean_data'].type(torch.FloatTensor).cuda(),\
data['angle_label'].type(torch.FloatTensor).cuda()
out_angle = angle_model(rain2)
angle_loss = data_criterion2(out_angle,angle_label)
angle_optimizer.zero_grad()
angle_loss.backward()
angle_optimizer.step()
if epoch < 20:
if step % 30 == 0 and step != 0:
# print("[epoch %d][%d/%d] loss: %.4f loss_angle: %.4f" %
# (epoch + 1, i + 1, len(loader_train), loss.item(), loss_angle.item()))
# print("Iter: {}, D: {:.4}, G:{:.4}".format(iter_count, d_total_error.item(), g_error.item()))
print("[epoch %d][%d/%d], angle_loss: %f " %
(epoch + 1, i + 1, len(loader_train), angle_loss.item()))
writer.add_scalar('angle loss', angle_loss.item(), step)
if step % 300 == 0:
# # results
with torch.no_grad():
out_angle_transform = transform_crop(rain2,out_angle)
label_angle_transform = transform_crop(rain2,angle_label)
Img_out_angle = utils.make_grid(out_angle_transform.data, nrow=8, normalize=True, scale_each=True)
Img_label_angle = utils.make_grid(label_angle_transform.data, nrow=8, normalize=True, scale_each=True)
writer.add_image('label angle image', Img_label_angle, epoch * repeat + step)
writer.add_image('out angle image', Img_out_angle, epoch * repeat + step)
else:
"""generator train"""
generator.train()
set_requires_grad([discriminator],False)
transform_rain = transform_crop(rain2,out_angle)
out_put = generator(transform_rain)
fake_images, Out_R = out_put['out_clean'],\
out_put['rain_streak']
reconstruct_image = fake_images + Out_R
w_x = fake_images.size()[3]
h_x = fake_images.size()[2]
#reconstruct loss
w1 = 0.1
consistencey_loss = data_criterion2(reconstruct_image,transform_rain)
# rain UTV Loss
w2 = 1.5
w3 = 1
w4 = 0.1
data_zero = torch.zeros_like(Out_R)
Gradient_y_OutR = gradient_y(Out_R)
gradient_zero = torch.zeros_like(Gradient_y_OutR)
R_prior1 = Prior_criterion1(gradient_y(Out_R),gradient_zero) #雨条层竖直方向梯度尽可能小
R_prior2 = Prior_criterion1(gradient_x(transform_rain),gradient_x(Out_R)) #雨条层的垂直方向梯度与退化图像的垂直方向梯度尽可能相似
R_prior3 = Prior_criterion1(Out_R,data_zero)
streak_total_loss = w2*R_prior1 + w3*R_prior2 + w4*R_prior3
#Clean image loss
w5 = 0.01
w6 = 0.01
w7 = 4e3
Clean_prior = data_criterion2(fake_images,transform_rain)
Tv_loss = Prior_criterion2(fake_images)
#adv loss
gen_logits_fake = discriminator(fake_images)
GAN_loss = class_loss(gen_logits_fake, True)
Image_total_loss = w5*Clean_prior + w6*Tv_loss + w7*GAN_loss
generator_total_loss = w1*consistencey_loss + 0.1*streak_total_loss + 0.1*Image_total_loss
g_optimizer.zero_grad()
generator_total_loss.backward()
g_optimizer.step()
"""discriminator train"""
# real data
transform_real_image = transform_crop(real_image,out_angle)
set_requires_grad([discriminator], True)
real_data = Variable(transform_real_image)
# transform_real_data = transform(real_data)
logits_real = discriminator(real_data)
# fake data
# fake_images = generator(rain2)
fake = fake_pool.query(fake_images)
logits_fake = discriminator(fake.detach())
#discriminator loss
discriminator_total_loss = class_loss(logits_real, True) + class_loss(logits_fake, False)
d_optimizer.zero_grad()
discriminator_total_loss.backward()
d_optimizer.step()
if step % 30 == 0 and step != 0:
# print("[epoch %d][%d/%d] loss: %.4f loss_angle: %.4f" %
# (epoch + 1, i + 1, len(loader_train), loss.item(), loss_angle.item()))
# print("Iter: {}, D: {:.4}, G:{:.4}".format(iter_count, d_total_error.item(), g_error.item()))
print("[epoch %d][%d/%d], angle_loss:%f,,D_loss: %f, consistencey_loss:%f,streak_total_loss:%f, Image_total_loss:%f,generator_total_loss:%f " %
(epoch + 1, i + 1, len(loader_train),\
angle_loss.item(),\
discriminator_total_loss.item(),\
consistencey_loss.item(),\
streak_total_loss.item(),\
Image_total_loss.item(),\
generator_total_loss.item()))
writer.add_scalar('discriminator_total_loss', discriminator_total_loss.item(), step)
writer.add_scalar('angle loss', angle_loss.item(), step)
writer.add_scalar('R_prior1_loss', R_prior1.item(), step)
writer.add_scalar('R_prior2_loss', R_prior2.item(), step)
writer.add_scalar('R_prior3_loss', R_prior2.item(), step)
writer.add_scalar('Clean_prior_loss', Clean_prior.item(), step)
writer.add_scalar('Tv_loss', Tv_loss.item(), step)
writer.add_scalar('GAN_loss', GAN_loss.item(), step)
writer.add_scalar('consistencey_loss', consistencey_loss.item(), step)
writer.add_scalar('streak_total_loss', streak_total_loss.item(), step)
writer.add_scalar('Image_total_loss', Image_total_loss.item(), step)
writer.add_scalar('generator_total_loss', generator_total_loss.item(), step)
if step % 300 == 0:
# # results
with torch.no_grad():
out2_train = torch.clamp(fake_images, 0., 1.)
out_mask_all = torch.clamp(Out_R, 0., 1.)
Img_rain = utils.make_grid(rain2.data, nrow=8, normalize=True, scale_each=True)
Img_clean = utils.make_grid(real_data.data, nrow=8, normalize=True, scale_each=True)
Img_out_img_all = utils.make_grid(out2_train.data, nrow=8, normalize=True, scale_each=True)
Img_out_mask_all = utils.make_grid(out_mask_all.data, nrow=8, normalize=True, scale_each=True)
writer.add_image('noisy image', Img_rain, epoch * repeat + step)
writer.add_image('clean image', Img_clean, epoch*repeat+step)
writer.add_image('reconstructed image2', Img_out_img_all, epoch * repeat + step)
writer.add_image('reconstructed all mask', Img_out_mask_all, epoch * repeat + step)
# torch.cuda.empty_cache()
# if you are using older version of PyTorch, you may need to change loss.item() to loss.data[0]
step += 1
d_schedulr.step()
g_schedulr.step()
angle_schedulr.step()
if epoch % 5 == 0:
torch.save(generator.state_dict(), os.path.join(output_folder, 'generator_epoch_' + str(epoch) + '.pth'))
torch.save(discriminator.state_dict(), os.path.join(output_folder, 'discriminator_epoch_' + str(epoch) + '.pth'))
torch.save(angle_model.state_dict(),os.path.join(output_folder, 'angle_model_epoch_' + str(epoch) + '.pth'))
torch.save(generator.state_dict(), os.path.join(output_folder, 'generator_backup.pth'))
torch.save(discriminator.state_dict(), os.path.join(output_folder, 'discriminator_backup.pth'))
torch.save(angle_model.state_dict(), os.path.join(output_folder, 'angle_model_backup.pth'))
if __name__ == "__main__":
main()