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
| """
使用 CycleGAN 实现莫奈风格图像转化为真实照片风格图片
在同一根目录下需要：
datasets/monet2photo/trainA 文件夹存放莫奈风格画像
datasets/monet2photo/trainB 文件夹存放真实风景图片
"""
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from torchvision import transforms
from torchvision.datasets import ImageFolder
import os
batch_size = 1
lr = 0.0002
epochs = 1000
input_channel_size = 3
output_channel_size = 3
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
os.makedirs("output_cyclegan", exist_ok=True)
transform = transforms.Compose([
transforms.Resize((256, 256)),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
dataset_A = ImageFolder(root="datasets/monet2photo/trainA", transform=transform)
dataset_B = ImageFolder(root="datasets/monet2photo/trainB", transform=transform)
dataloader_A = DataLoader(dataset_A, batch_size=batch_size, shuffle=True)
dataloader_B = DataLoader(dataset_B, batch_size=batch_size, shuffle=True)
class ResnetBlock(nn.Module):
def __init__(self, dim):
super().__init__()
self.conv_block = nn.Sequential(
nn.ReflectionPad2d(1),
nn.Conv2d(dim, dim, kernel_size=3, padding=0, bias=False),
nn.InstanceNorm2d(dim),
nn.ReLU(True),
nn.ReflectionPad2d(1),
nn.Conv2d(dim, dim, kernel_size=3, padding=0, bias=False),
nn.InstanceNorm2d(dim)
)
def forward(self, x):
return x + self.conv_block(x)
class Generator(nn.Module):
def __init__(self, input_channel_size, output_channel_size):
super().__init__()
self.main = nn.Sequential(
nn.ReflectionPad2d(3),
nn.Conv2d(input_channel_size, 64, kernel_size=7, padding=0, bias=False),
nn.InstanceNorm2d(64),
nn.ReLU(True),
nn.Conv2d(64, 128, kernel_size=3, stride=2, padding=1, bias=False),
nn.InstanceNorm2d(128),
nn.ReLU(True),
nn.Conv2d(128, 256, kernel_size=3, stride=2, padding=1, bias=False),
nn.InstanceNorm2d(256),
nn.ReLU(True),
ResnetBlock(256),
ResnetBlock(256),
ResnetBlock(256),
ResnetBlock(256),
nn.ConvTranspose2d(256, 128, kernel_size=3, stride=2, padding=1, output_padding=1, bias=False),
nn.InstanceNorm2d(128),
nn.ReLU(True),
nn.ConvTranspose2d(128, 64, kernel_size=3, stride=2, padding=1, output_padding=1, bias=False),
nn.InstanceNorm2d(64),
nn.ReLU(True),
nn.ReflectionPad2d(3),
nn.Conv2d(64, output_channel_size, kernel_size=7, padding=0),
nn.Tanh()
)
def forward(self, input):
return self.main(input)
class Discriminator(nn.Module):
def __init__(self, input_channel_size):
super().__init__()
self.model = nn.Sequential(
nn.Conv2d(input_channel_size, 64, kernel_size=4, stride=2, padding=1),
nn.LeakyReLU(0.2, True),
nn.Conv2d(64, 128, kernel_size=4, stride=2, padding=1, bias=False),
nn.InstanceNorm2d(128),
nn.LeakyReLU(0.2, True),
nn.Conv2d(128, 256, kernel_size=4, stride=2, padding=1, bias=False),
nn.InstanceNorm2d(256),
nn.LeakyReLU(0.2, True),
nn.Conv2d(256, 1, kernel_size=4, stride=2, padding=1)
)
def forward(self, input):
return self.model(input)
netG = Generator(input_channel_size, output_channel_size).to(device)
netF = Generator(output_channel_size, input_channel_size).to(device)
netD_A = Discriminator(input_channel_size).to(device)
netD_B = Discriminator(output_channel_size).to(device)
criterion_GAN = nn.MSELoss()
criterion_cycle = nn.L1Loss()
optimizer_GF = optim.Adam(list(netG.parameters()) + list(netF.parameters()), lr=lr, betas=(0.5, 0.999))
optimizer_D_A = optim.Adam(netD_A.parameters(), lr=lr, betas=(0.5, 0.999))
optimizer_D_B = optim.Adam(netD_B.parameters(), lr=lr, betas=(0.5, 0.999))
for epoch in range(epochs):
for i, ((real_A, _), (real_B, _)) in enumerate(zip(dataloader_A, dataloader_B)):
real_A = real_A.to(device)
real_B = real_B.to(device)
optimizer_GF.zero_grad()
fake_B = netG(real_A)
fake_A = netF(real_B)
valid_B = torch.ones_like(netD_B(fake_B))
valid_A = torch.ones_like(netD_A(fake_A))
loss_G_GAN = criterion_GAN(netD_B(fake_B), valid_B)
loss_F_GAN = criterion_GAN(netD_A(fake_A), valid_A)
recovered_A = netF(fake_B)
recovered_B = netG(fake_A)
loss_cycle_A = criterion_cycle(recovered_A, real_A)
loss_cycle_B = criterion_cycle(recovered_B, real_B)
total_cycle_loss = (loss_cycle_A + loss_cycle_B) * 10
total_G_loss = loss_G_GAN + loss_F_GAN + total_cycle_loss
total_G_loss.backward()
optimizer_GF.step()
real_output_A = netD_A(real_A)
fake_output_A = netD_A(fake_A.detach())
loss_D_real_A = criterion_GAN(real_output_A, torch.ones_like(real_output_A))
loss_D_fake_A = criterion_GAN(fake_output_A, torch.zeros_like(fake_output_A))
loss_D_A = 0.5 * (loss_D_real_A + loss_D_fake_A)
loss_D_A.backward()
optimizer_D_A.step()
real_output_B = netD_B(real_B)
fake_output_B = netD_B(fake_B.detach())
loss_D_real_B = criterion_GAN(real_output_B, torch.ones_like(real_output_B))
loss_D_fake_B = criterion_GAN(fake_output_B, torch.zeros_like(fake_output_B))
loss_D_B = 0.5 * (loss_D_real_B + loss_D_fake_B)
loss_D_B.backward()
optimizer_D_B.step()
if i % 100 == 0:
print(f"Epoch [{epoch + 1}/{epochs}] Batch {i} G_loss: {total_G_loss.item():.4f} D_A: {loss_D_A.item():.4f} D_B: {loss_D_B.item():.4f}")
torch.save(netG.state_dict(), "generator_monet2real.pth")
torch.save(netF.state_dict(), "generator_real2monet.pth")
|
