/
models.py
123 lines (99 loc) · 4.18 KB
/
models.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
import torch.nn as nn
class ResBlock(nn.Module):
def __init__(self, conv_in, conv_out):
super(ResBlock, self).__init__()
self.res_sequence = nn.Sequential(
nn.ReflectionPad2d(padding=1),
nn.Conv2d(conv_in, conv_out, kernel_size=(3, 3), stride=1, padding=0),
nn.InstanceNorm2d(conv_out),
nn.ReLU(inplace=True),
nn.ReflectionPad2d(padding=1),
nn.Conv2d(conv_in, conv_out, kernel_size=(3, 3), stride=1, padding=0),
nn.InstanceNorm2d(conv_out)
)
def forward(self, input):
x = self.res_sequence(input)
#element-wise sum of output with input
output = input + x
return output
class Generator(nn.Module):
def __init__(self, channels, width, height, feature_space=64, n_res=16):
super(Generator, self).__init__()
self.init_conv = nn.Sequential(
nn.ReflectionPad2d(channels),
nn.Conv2d(channels, feature_space, kernel_size=(7, 7), stride=1, padding=0),
nn.InstanceNorm2d(feature_space),
nn.ReLU(inplace=True)
)
#downsizing
downsize_layer = []
conv_in = feature_space
conv_out = feature_space*2
for _ in range(2):
downsize_layer.append(nn.Conv2d(conv_in, conv_out, kernel_size=(3, 3), stride=2, padding=1))
downsize_layer.append(nn.InstanceNorm2d(conv_out))
downsize_layer.append(nn.ReLU(inplace=True))
conv_in = conv_out
conv_out = conv_out*2
self.downsize_sequence = nn.Sequential(*downsize_layer)
#residual layer
residual_layer = []
for _ in range(n_res):
residual_layer.append(ResBlock(conv_in, conv_in))
self.resid_sequence = nn.Sequential(*residual_layer)
#upsizing
upsize_layer = []
conv_out = conv_in // 2
for _ in range(2):
upsize_layer.append(nn.Upsample(scale_factor=2))
upsize_layer.append(nn.Conv2d(conv_in, conv_out, kernel_size=(3, 3), stride=1, padding=1))
upsize_layer.append(nn.InstanceNorm2d(conv_out))
upsize_layer.append(nn.ReLU(inplace=True))
conv_in = conv_out
conv_out = conv_out // 2
self.upsize_sequence = nn.Sequential(*upsize_layer)
#output conv
self.output = nn.Sequential(
nn.ReflectionPad2d(channels),
nn.Conv2d(conv_in, channels, kernel_size=(7, 7), stride=1, padding=0),
nn.Tanh()
)
def forward(self, input):
x = self.init_conv(input)
x = self.downsize_sequence(x)
x = self.resid_sequence(x)
x = self.upsize_sequence(x)
output = self.output(x)
return output
class DiscriminatorBlock(nn.Module):
def __init__(self, conv_in, conv_out, normalize):
super(DiscriminatorBlock, self).__init__()
self.normalize = normalize
self.conv = nn.Conv2d(conv_in, conv_out, kernel_size=(4, 4), stride=2, padding=1)
self.inst_norm = nn.InstanceNorm2d(conv_out)
self.lr = nn.LeakyReLU(0.2, inplace=True)
def forward(self, input):
x = self.conv(input)
if self.normalize:
x = self.inst_norm(x)
output = self.lr(x)
return output
class Discriminator(nn.Module):
def __init__(self, channels, width, height, feature_space=64):
super(Discriminator, self).__init__()
dis_block_layer = []
#init discriminator block feed through
dis_block_layer.append(DiscriminatorBlock(channels, feature_space, False))
conv_in = feature_space
for _ in range(3):
dis_block_layer.append(DiscriminatorBlock(conv_in, conv_in*2, True))
conv_in = conv_in * 2
self.block_sequence = nn.Sequential(*dis_block_layer)
self.output = nn.Sequential(
nn.ZeroPad2d((1, 0, 1, 0)),
nn.Conv2d(conv_in, 1, kernel_size=(4, 4), stride=1, padding=1)
)
def forward(self, input):
x = self.block_sequence(input)
output = self.output(x)
return output