forked from zoharli/SRGAN-pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
259 lines (232 loc) · 10.2 KB
/
train.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
import argparse
import os
import shutil
import time
import numpy as np
from PIL import Image
import torch
import torch.nn as nn
import torch.nn.parallel
import torch.backends.cudnn as cudnn
import torch.optim
import torch.utils.data
import torchvision.transforms as transforms
import torchvision.datasets as datasets
import torchvision.models as models
from torch.autograd import Variable
from models import *
from dataset import *
import numpy as np
from setproctitle import *
parser = argparse.ArgumentParser(description='Pretraining')
parser.add_argument('-j', '--workers', default=4, type=int,
help='number of data loading workers (default: 4)')
parser.add_argument('--epochs', default=40, type=int,
help='number of total epochs to run')
parser.add_argument('--start-epoch', default=0, type=int,
help='manual epoch number (useful on restarts)')
parser.add_argument('-b', '--batch-size', default=8, type=int,
help='mini-batch size (default: 16)')
parser.add_argument('--crop-size','-c',default=256,type=int,
help='crop size of the hr image')
parser.add_argument('--lr', '--learning-rate', default=1e-4, type=float,
help='initial learning rate')
parser.add_argument('--momentum','-m',default=0.9,type=float,
help='momentum if using sgd optimization')
parser.add_argument('--weight-decay', '--wd', default=1e-4, type=float,
help='weight decay (default: 1e-4)')
parser.add_argument('--print-freq', '-p', default=100, type=int,
help='print frequency (default: 20)')
parser.add_argument('--resume', default='', type=str,
help='path to latest checkpoint (default: none)')
parser.add_argument('--generator',default='',type=str,
help='path to the pretrained srResNet if use it')
parser.add_argument('--logdir','-s',default='save',type=str,
help='path to save checkpoint')
parser.add_argument('--optim','-o',default='Adam',
help='the optimization method to be employed')
parser.add_argument('--traindir',default='r375-400.bin',
help=' the global name of training set dir')
parser.add_argument('--valdir',default='b100',type=str,
help='the global name of validation set dir')
parser.add_argument('--weight',default=0.001,type=float,
help='the weight of adversarial loss,i.e. gen_loss=content_loss+weight*adv_loss')
parser.add_argument('--separate',action='store_true',
help='wheather to separate real and fake minibatch when training discriminator')
parser.add_argument('--fixG',action='store_true',
help='wheather to fix generator and only train discriminator')
parser.add_argument('--fixD',action='store_true',
help='wheather to fix discriminator and only train generator')
parser.add_argument('--clip',default=None,type=float,
help='gradient clip norm')
global_step=0
best_psnr = -100
best_gen_loss=10000
best_disc_loss= 10000
args = parser.parse_args()
args.__dict__['upscale_factor']=4
#args.traindir=globals()[args.traindir]
args.valdir=globals()[args.valdir]
args.__dict__['model_base_name']='SRGAN_v%g_w%g_%s'%(args.lr,args.weight,args.optim)+('_fixG' if args.fixG else '')+('_fixD' if args.fixD else '')+('_separate' if args.separate else '')
args.__dict__['model_name']=args.model_base_name+'.pth'
args.__dict__['snapshot']='snapshot_'+args.model_base_name
setproctitle(args.model_base_name)
if not os.path.exists(args.logdir):
os.makedirs(args.logdir)
if not os.path.exists(args.snapshot):
os.makedirs(args.snapshot)
cudnn.benchmark = True
train_loader = torch.utils.data.DataLoader(
get_train_set(args),
batch_size=args.batch_size, shuffle=True,
num_workers=args.workers, pin_memory=True)
gen=GenNet().cuda()
disc=DisNet().cuda()
vgg=vgg19_54().cuda()
if args.resume:
if os.path.isfile(args.resume):
print("=> loading checkpoint '{}'".format(args.resume))
checkpoint = torch.load(args.resume)
args.start_epoch = checkpoint['epoch']
global_step=checkpoint['global_step']
best_psnr = checkpoint['best_psnr']
gen.load_state_dict(checkpoint['gen_state_dict'])
disc.load_state_dict(checkpoint['disc_state_dict'])
print("=> loaded checkpoint (epoch {})"
.format( checkpoint['epoch']))
else:
print("=> no checkpoint found at '{}'".format(args.resume))
if args.generator:
if os.path.isfile(args.generator):
print("=> loading checkpoint '{}'".format(args.generator))
checkpoint = torch.load(args.generator)
gen.load_state_dict(checkpoint['state_dict'])
print("=> loaded checkpoint")
else:
print("=>no checkpoint found at '{}'".format(args.generator))
label=Variable(torch.FloatTensor(args.batch_size)).cuda()
cont_criterion = nn.MSELoss().cuda()
adv_criterion = nn.BCELoss().cuda()
gen_optimizer = torch.optim.Adam(gen.parameters(), args.lr)
disc_optimizer = torch.optim.Adam(disc.parameters(),args.lr)
def normalize(tensor):
r,g,b=torch.split(tensor,1,1)
r=(r-0.485)/0.229
g=(g-0.456)/0.224
b=(b-0.406)/0.225
return torch.cat((r,g,b),1)
def rgb2y_matlab(img): #convert a PIL rgb image to y-channel image in `matlab` way
img=(img/255.0)[4:-4,4:-4,:]
img=65.481*img[:,:,0]+128.553*img[:,:,1]+24.966*img[:,:,2]+16.0
return img
def validate(model,vgg,criterion,valdir,epoch,factor,optimizer):
cnt=0
sum=0
ysum=0
cont_loss=0
model.eval()
for x in [os.path.join(valdir,y) for y in os.listdir(valdir)]:
im=Image.open(x)
im=im.resize((im.size[0]-im.size[0]%factor,im.size[1]-im.size[1]%factor),Image.BICUBIC)
target=Variable(torch.stack([transforms.ToTensor()(im)],0),volatile=True).cuda()
f_target=vgg(normalize(target))
input=torch.stack([transforms.ToTensor()(im.resize((im.size[0]//args.upscale_factor,im.size[1]//args.upscale_factor),Image.BICUBIC))],0)
input_var=Variable(input,volatile=True).cuda()
output=model(input_var)
f_output=vgg(normalize(output))
cont_loss+=criterion(f_target,f_output).cpu().data[0]
loss=criterion(target,output).cpu().data[0]
img=torch.squeeze(output.data.cpu())
rgb=transforms.ToPILImage()(img)
rgb.save(os.path.join(args.snapshot,os.path.basename(x)))
yorigin=rgb2y_matlab(np.asarray(im))
youtput=rgb2y_matlab(np.asarray(rgb))
ymse=np.mean((yorigin-youtput)**2)
ypsnr=10*(np.log10(255.0**2/ymse))
psnr=10*np.log10(1.0/loss)
sum+=psnr
ysum+=ypsnr
cnt+=1
psnr=float(sum)/cnt
ypsnr=float(ysum)/cnt
cont_loss=float(cont_loss)/cnt
s=' | psnr=%.3f | ypsnr=%.3f | cont_mse=%g'%(psnr,ypsnr,cont_loss)
model.train()
return s
def save_checkpoint(state, is_best,logdir):
filename=os.path.join(logdir,args.model_name)
torch.save(state, filename)
if is_best:
shutil.copyfile(filename, os.path.join(logdir,'best_'+args.model_name))
def train(epoch):
for i, (input, target) in enumerate(train_loader):
global global_step
input_var = Variable(input.cuda())
target_var = Variable(target.cuda())
if not args.fixD:
if args.separate:
real_loss=((disc(target_var)-1)**2).mean()
disc_optimizer.zero_grad()
real_loss.backward()
#torch.nn.utils.clip_grad_norm(disc.parameters(),args.clip)
disc_optimizer.step()
fake_loss=torch.mean(disc(gen(input_var))**2)
disc_optimizer.zero_grad()
fake_loss.backward()
#torch.nn.utils.clip_grad_norm(disc.parameters(),args.clip)
disc_optimizer.step()
disc_loss=(fake_loss+real_loss)/2
else:
disc_optimizer.zero_grad()
label.data.fill_(1)
output=disc(target_var)
real_loss=adv_criterion(output,label)
label.data.fill_(0)
output=disc(gen(input_var).detach())
fake_loss=adv_criterion(output,label)
disc_loss=args.weight*(fake_loss+real_loss)*100
disc_loss.backward()
disc_optimizer.step()
if not args.fixG and (disc_loss.data[0]<27.2/20 or i%args.print_freq==0):
gen_optimizer.zero_grad()
G_z=gen(input_var)
fake_feature=vgg(normalize(G_z))
real_feature=vgg(normalize(target_var)).detach()
content_loss=cont_criterion(fake_feature,real_feature)
label.data.fill_(1)
output=disc(G_z)
adv_loss=adv_criterion(output,label)
gen_loss=args.weight*adv_loss+content_loss
gen_loss.backward()
if args.clip is not None:
torch.nn.utils.clip_grad_norm(gen.parameters(),args.clip)
gen_optimizer.step()
if i % args.print_freq == 0:
s=time.strftime('%dth-%H:%M:%S',time.localtime(time.time()))+' | epoch%d(%d) | lr=%g'%(epoch,global_step,args.lr)
if not args.fixG:
vs=validate(gen,vgg,cont_criterion,args.valdir,epoch,args.upscale_factor,gen_optimizer)
s+=vs+' | Loss(G):%g[Cont:%g/Adv:%g]'%(gen_loss.data[0],content_loss.data[0],adv_loss.data[0])
if not args.fixD:
s+=' | Loss(D):%g[Real:%g/Fake:%g]'%(disc_loss.data[0],real_loss.data[0],fake_loss.data[0])
print(s)
f=open('info.'+args.model_base_name,'a')
f.write(s+'\n')
f.close()
if not args.fixG:
global best_gen_loss
is_best = gen_loss.data[0] > best_gen_loss
best_gen_loss = max(gen_loss.data[0] , best_gen_loss)
else:
global best_disc_loss
is_best = disc_loss.data[0] > best_disc_loss
best_disc_loss = max(disc_loss.data[0],best_disc_loss)
save_checkpoint({
'epoch': epoch + 1,
'global_step':global_step,
'gen_state_dict': gen.state_dict(),
'disc_state_dict':disc.state_dict(),
'best_psnr': best_psnr,
}, is_best,args.logdir)
global_step+=1
for epoch in range(args.start_epoch, args.epochs):
train(epoch)