-
Notifications
You must be signed in to change notification settings - Fork 46
/
main.py
171 lines (145 loc) · 6.13 KB
/
main.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
import torch
import torch.nn as nn
import torchvision.transforms as transforms
from dataloader import TextDataset
import models
import argparse
parser = argparse.ArgumentParser()
parser.add_argument('--image_size1', type=int, default=64,help='stage1 image size')
parser.add_argument('--image_size2', type=int, default=256,help='stage2 image size')
parser.add_argument('--z_dim', type=int, default=100,help='noise dimension')
parser.add_argument('--embed_dim', type=int, default=128,help='embedding compressed dim')
parser.add_argument('--D1_dim', type=int, default=512,)
parser.add_argument('--G2_dim', type=int, default=512)
parser.add_argument('--data_dir', type=str, required=True,help='path to embedding directory containing pickle files')
parser.add_argument('--image_dir', type=str, required=True,help='path to folder containing jpg folder which contains images')
parser.add_argument('--epochs1', type=int, default=100)
parser.add_argument('--epochs2', type=int, default=100)
parser.add_argument('--batch_size', type=int, default=64)
parser.add_argument('--lrG1', type=float, default=0.0002)
parser.add_argument('--lrG2', type=float, default=0.0002)
parser.add_argument('--lrD1', type=float, default=0.0002)
parser.add_argument('--lrD2', type=float, default=0.0002)
parser.add_argument('--beta1', type=float, default=0.5)
parser.add_argument('--beta2', type=float, default=0.999)
args = parser.parse_args()
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
image_transform1 = transforms.Compose([
transforms.RandomCrop(args.image_size1),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])
image_transform2 = transforms.Compose([
transforms.RandomCrop(args.image_size2),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])
dataset = TextDataset(args.image_dir,args.data_dir,image_transform1,image_transform2 , 'train', args.image_size1,args.image_size2)
dataloader = torch.utils.data.DataLoader(dataset, batch_size=args.batch_size, drop_last=True)
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
m.weight.data.normal_(0.0, 0.02)
elif classname.find('BatchNorm') != -1:
m.weight.data.normal_(1.0, 0.02)
m.bias.data.fill_(0)
elif classname.find('Linear') != -1:
m.weight.data.normal_(0.0, 0.02)
if m.bias is not None:
m.bias.data.fill_(0.0)
def show_and_save(file_name,img):
npimg = np.transpose(img.numpy(),(1,2,0))
f = "/content/drive/My Drive/%s.png" % file_name
fig = plt.figure(dpi=250)
fig.suptitle(file_name, fontsize=14, fontweight='bold')
plt.imshow(npimg)
plt.imsave(f,npimg)
'''
STAGE 1
'''
netG1=models.G1(args.embed_dim,args.z_dim,args.image_size1,(args.embed_dim+args.z_dim)*8).to(device)
netG1.apply(weights_init)
netD1=models.D1(args.image_size1,args.D1_dim,args.embed_dim).to(device)
netD1.apply(weights_init)
netG1_para = []
for p in netG1.parameters():
if p.requires_grad:
netG1_para.append(p)
optimG1=torch.optim.Adam(netG1_para,lr=args.lrG1,betas=(args.beta1, args.beta2))
optimD1=torch.optim.Adam(netD1.parameters(),lr=args.lrD1,betas=(args.beta1, args.beta2))
criterion=nn.BCELoss().to(device)
def KL(logvar,mu):
KLD_element = mu.pow(2).add_(logvar.exp()).mul_(-1).add_(1).add_(logvar)
KLD = torch.mean(KLD_element).mul_(-0.5)
return KLD
ones=torch.ones(args.batch_size).to(device)
zeros=torch.zeros(args.batch_size).to(device)
z_fixed=torch.randn(args.batch_size,args.z_dim).to(device)
for epoch in range(args.epochs1):
G1_loss=[]
D1_loss=[]
for i,data in enumerate(dataloader):
data[0]=data[0].to(device)
data[2]=data[2].to(device)
z=torch.randn(args.batch_size,args.z_dim).to(device)
gen_img,mean,logvar=netG1(data[2],z)
out_fake=netD1(gen_img,data[2])
out_real=netD1(data[0],data[2])
lossD=criterion(out_fake,zeros)+criterion(out_real,ones)
optimD1.zero_grad()
lossD.backward(retain_graph=True)
optimD1.step()
D1_loss.append(lossD.item())
gen_img,mean,logvar=netG1(data[2],z)
out_fake=netD1(gen_img,data[2])
kl=KL(logvar,mean).to(device)
lossG=criterion(out_fake,zeros)+KL(logvar,mean).to(device)
G1_loss.append(lossG.item())
optimG1.zero_grad()
lossG.backward()
optimG1.step()
if i%10==0:
print("Epoch [%d/%d] Step[%d/%d] G_loss:%f D_loss:%f KL: %f"%(epoch+1,args.epochs1,i+1,len(dataloader),lossG.item(),lossD.item(),kl.item()) )
gen,_,_=netG(data[2],z_fixed)
gen=gen.detach()
show_and_save("%d"%(i) ,make_grid((gen*0.5+0.5).cpu(),8))
'''
STAGE 2
'''
gen_img64=[]
for i,data in enumerate(dataloader):
data[2]=data[2].to(device)
z=torch.randn(args.batch_size,args.z_dim).to(device)
gen_img,mean,logvar=netG1(data[2],z)
gen_img=gen_img.detach()
gen_img64.append(gen_img)
netG2=models.G2(64,512,256,512+128).to(device)
netG2.apply(weights_init)
netD2=models.D2(256,512,128).to(device)
netD2.apply(weights_init)
optimG2=torch.optim.Adam(netG2.parameters(),lr=args.lrG2,betas=(args.beta1,args.beta2))
optimD2=torch.optim.Adam(netD2.parameters(),lr=args.lrD2,betas=(args.beta1,args.beta2))
criterion=nn.BCELoss().to(device)
for epoch in range(args.epochs2):
G2_loss=[]
D2_loss=[]
for i,data in enumerate(dataloader):
data[2]=data[2].to(device)
data[1]=data[1].to(device)
gen_img64[i]=gen_img64[i].to(device)
gen_img,mean,logvar=netG2(data[2],gen_img64[i])
out_fake=netD2(gen_img,data[2])
out_real=netD2(data[1],data[2])
lossD=criterion(out_fake,zeros)+criterion(out_real,ones)
optimD2.zero_grad()
lossD.backward(retain_graph=True)
optimD2.step()
D2_loss.append(lossD.item())
gen_img,mean,logvar=netG2(data[2],gen_img64[i])
out_fake=netD2(gen_img,data[2])
kl=KL(logvar,mean).to(device)
lossG=criterion(out_fake,zeros)+KL(logvar,mean).to(device)
G2_loss.append(lossG.item())
optimG2.zero_grad()
lossG.backward()
optimG2.step()
if i%10==0:
print("Epoch [%d/%d] Step[%d/%d] G_loss:%f D_loss:%f KL: %f"%(epoch+1,args.epochs2,i+1,len(dataloader),lossG.item(),lossD.item(),kl.item()) )