GAN-Face-Attribute-Manipulation-pytorch/train1.py at master · fluctlux/GAN-Face-Attribute-Manipulation-pytorch · GitHub

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
import net
import load_data as ld
import numpy as np
import torch
import pdb
from PIL import Image
from torch.autograd import Variable as Vb
import torch.nn as nn
import torch.nn.functional as F
import torch.nn.init as init
import torchvision.models
import torch.optim as optim
import logging
import torchvision.utils  as tov
from torch import Tensor as tensor
#pdb.set_trace()
gpuid=0
lr_rate=0.0002
alpha=0.0005
beta=0.01*alpha
num_iter=500000
optim_betas=(0.9,0.999)
bs=64
labelfake=Vb(torch.from_numpy(np.full((bs),2,dtype=int))).cuda(gpuid)
labelpo=Vb(torch.from_numpy(np.full((bs),1,dtype=int))).cuda(gpuid)
labelne=Vb(torch.from_numpy(np.full((bs),0,dtype=int))).cuda(gpuid)
logging.basicConfig(filename='log/residualgan_v10.log', level=logging.INFO)
G1=net.Generator1().cuda(gpuid)
G2=net.Generator2().cuda(gpuid)
D=net.Discriminator().cuda(gpuid)
d_optimizer=optim.Adam(D.parameters(),lr=lr_rate,betas=optim_betas)
g1_optimizer=optim.Adam(G1.parameters(),lr=lr_rate,betas=optim_betas)
g2_optimizer=optim.Adam(G2.parameters(),lr=lr_rate,betas=optim_betas)
l1_crit=nn.L1Loss(size_average=False)
datalistpo=ld.getlist('../Eyeglasses_Positive.txt')
datalistne=ld.getlist('../Eyeglasses_Negative.txt')
iternow1=0
iternow2=0
for iter1 in xrange(num_iter):
    D.zero_grad()
    datapo,iternow1=ld.load_data('../img_align_celeba_crop/','../Eyeglasses_Positive.txt',datalistpo,iternow1,bs,gpuid=gpuid)
    datane,iternow2=ld.load_data('../img_align_celeba_crop/','../Eyeglasses_Negative.txt',datalistne,iternow2,bs,gpuid=gpuid)
    datapo_,_=G1(datane)
    datane_,_=G2(datapo)
    dlabelpo,_=D(datapo)
    dlabelne,_=D(datane)
    dlabelpo_,_=D(datapo_)
    dlabelne_,_=D(datane_)
    trueloss1=F.cross_entropy(dlabelpo,labelpo)
    trueloss2=F.cross_entropy(dlabelne,labelne)
    fakeloss1=F.cross_entropy(dlabelpo_,labelfake)
    fakeloss2=F.cross_entropy(dlabelne_,labelfake)
    trueloss1.backward()
    trueloss2.backward()
    fakeloss1.backward()
    fakeloss2.backward()
    if iter1%100==0:
        outinfo=str(trueloss1)+' '+str(trueloss2)+' '+str(fakeloss1)+' '+str(fakeloss2)
        logging.info(outinfo)
        print 'dloss',outinfo
    d_optimizer.step()


    datapo,iternow1=ld.load_data('../img_align_celeba_crop/','../Eyeglasses_Positive.txt',datalistpo,iternow1,bs,gpuid=gpuid)
    datane,iternow2=ld.load_data('../img_align_celeba_crop/','../Eyeglasses_Negative.txt',datalistne,iternow2,bs,gpuid=gpuid)
    G1.zero_grad()
    G2.zero_grad()
    datapo_,resu1=G1(datane)
    dlabelpo_,out3_=D(datapo_)
    dlabelne,out3=D(datane)
    datane_2,_=G2(datapo_)
    dlabelne_2,_=D(datane_2)
    ganloss1=F.cross_entropy(dlabelpo_,labelpo)
    pixloss1=torch.sum(torch.abs(resu1))/bs
    perloss1=torch.sum(torch.abs(out3-out3_))/bs
    dualloss1=F.cross_entropy(dlabelne_2,labelne)
    g1loss=ganloss1+alpha*pixloss1+beta*perloss1+dualloss1
    #g1loss=ganloss1+dualloss1+alpha*pixloss1
    if iter1%100==0:
        outinfo=str(g1loss)
        logging.info(outinfo)
        print 'g1loss',outinfo
    g1loss.backward()
    #g1_optimizer.step()

    datane_,resu2=G2(datapo)
    dlabelne_,out3_=D(datane_)
    dlabelpo,out3=D(datapo)
    datapo_2,_=G2(datane_)
    dlabelpo_2,_=D(datapo_2)
    ganloss2=F.cross_entropy(dlabelne_,labelne)
    pixloss2=torch.sum(torch.abs(resu2))/bs
    perloss2=torch.sum(torch.abs(out3-out3_))/bs
    dualloss2=F.cross_entropy(dlabelpo_2,labelpo)
    g2loss=ganloss2+alpha*pixloss2+beta*perloss2+dualloss2
    #g2loss=ganloss2+dualloss2+alpha*pixloss2
    if iter1%100==0:
        outinfo=str(g2loss)
        logging.info(outinfo)
        print 'g2loss',outinfo
        logging.info(str(iter1))
    g2loss.backward()
    g1_optimizer.step()
    g2_optimizer.step()
    print iter1
    #logging.info(str(iter1))
    if iter1 % 200 == 0:
        saveim=datapo.cpu().data
        tov.save_image(saveim,'img/datapo'+str(iter1)+'.jpg')
        saveim=datane.cpu().data
        tov.save_image(saveim,'img/datane'+str(iter1)+'.jpg')
        saveim=datapo_.cpu().data
        tov.save_image(saveim,'img/datapo_'+str(iter1)+'.jpg')
        saveim=datane_.cpu().data
        tov.save_image(saveim,'img/datane_'+str(iter1)+'.jpg')
        save_name = 'model/{}_iter_{}.pth.tar'.format('residualgan', iter1)
        torch.save({'G1': G1.state_dict(), 'G2': G2.state_dict(),'D': D.state_dict()}, save_name)
        logging.info('save model to {}'.format(save_name))