-
Notifications
You must be signed in to change notification settings - Fork 26
/
Copy pathmodel.py
142 lines (118 loc) · 4.95 KB
/
model.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
import torch.nn as nn
import torch.nn.functional as F
from torchvision.models import vgg19
def calc_mean_std(features):
"""
:param features: shape of features -> [batch_size, c, h, w]
:return: features_mean, feature_s: shape of mean/std ->[batch_size, c, 1, 1]
"""
batch_size, c = features.size()[:2]
features_mean = features.reshape(batch_size, c, -1).mean(dim=2).reshape(batch_size, c, 1, 1)
features_std = features.reshape(batch_size, c, -1).std(dim=2).reshape(batch_size, c, 1, 1) + 1e-6
return features_mean, features_std
def adain(content_features, style_features):
"""
Adaptive Instance Normalization
:param content_features: shape -> [batch_size, c, h, w]
:param style_features: shape -> [batch_size, c, h, w]
:return: normalized_features shape -> [batch_size, c, h, w]
"""
content_mean, content_std = calc_mean_std(content_features)
style_mean, style_std = calc_mean_std(style_features)
normalized_features = style_std * (content_features - content_mean) / content_std + style_mean
return normalized_features
class VGGEncoder(nn.Module):
def __init__(self):
super().__init__()
vgg = vgg19(pretrained=True).features
self.slice1 = vgg[: 2]
self.slice2 = vgg[2: 7]
self.slice3 = vgg[7: 12]
self.slice4 = vgg[12: 21]
for p in self.parameters():
p.requires_grad = False
def forward(self, images, output_last_feature=False):
h1 = self.slice1(images)
h2 = self.slice2(h1)
h3 = self.slice3(h2)
h4 = self.slice4(h3)
if output_last_feature:
return h4
else:
return h1, h2, h3, h4
class RC(nn.Module):
"""A wrapper of ReflectionPad2d and Conv2d"""
def __init__(self, in_channels, out_channels, kernel_size=3, pad_size=1, activated=True):
super().__init__()
self.pad = nn.ReflectionPad2d((pad_size, pad_size, pad_size, pad_size))
self.conv = nn.Conv2d(in_channels, out_channels, kernel_size)
self.activated = activated
def forward(self, x):
h = self.pad(x)
h = self.conv(h)
if self.activated:
return F.relu(h)
else:
return h
class Decoder(nn.Module):
def __init__(self):
super().__init__()
self.rc1 = RC(512, 256, 3, 1)
self.rc2 = RC(256, 256, 3, 1)
self.rc3 = RC(256, 256, 3, 1)
self.rc4 = RC(256, 256, 3, 1)
self.rc5 = RC(256, 128, 3, 1)
self.rc6 = RC(128, 128, 3, 1)
self.rc7 = RC(128, 64, 3, 1)
self.rc8 = RC(64, 64, 3, 1)
self.rc9 = RC(64, 3, 3, 1, False)
def forward(self, features):
h = self.rc1(features)
h = F.interpolate(h, scale_factor=2)
h = self.rc2(h)
h = self.rc3(h)
h = self.rc4(h)
h = self.rc5(h)
h = F.interpolate(h, scale_factor=2)
h = self.rc6(h)
h = self.rc7(h)
h = F.interpolate(h, scale_factor=2)
h = self.rc8(h)
h = self.rc9(h)
return h
class Model(nn.Module):
def __init__(self):
super().__init__()
self.vgg_encoder = VGGEncoder()
self.decoder = Decoder()
def generate(self, content_images, style_images, alpha=1.0):
content_features = self.vgg_encoder(content_images, output_last_feature=True)
style_features = self.vgg_encoder(style_images, output_last_feature=True)
t = adain(content_features, style_features)
t = alpha * t + (1 - alpha) * content_features
out = self.decoder(t)
return out
@staticmethod
def calc_content_loss(out_features, t):
return F.mse_loss(out_features, t)
@staticmethod
def calc_style_loss(content_middle_features, style_middle_features):
loss = 0
for c, s in zip(content_middle_features, style_middle_features):
c_mean, c_std = calc_mean_std(c)
s_mean, s_std = calc_mean_std(s)
loss += F.mse_loss(c_mean, s_mean) + F.mse_loss(c_std, s_std)
return loss
def forward(self, content_images, style_images, alpha=1.0, lam=10):
content_features = self.vgg_encoder(content_images, output_last_feature=True)
style_features = self.vgg_encoder(style_images, output_last_feature=True)
t = adain(content_features, style_features)
t = alpha * t + (1 - alpha) * content_features
out = self.decoder(t)
output_features = self.vgg_encoder(out, output_last_feature=True)
output_middle_features = self.vgg_encoder(out, output_last_feature=False)
style_middle_features = self.vgg_encoder(style_images, output_last_feature=False)
loss_c = self.calc_content_loss(output_features, t)
loss_s = self.calc_style_loss(output_middle_features, style_middle_features)
loss = loss_c + lam * loss_s
return loss