-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
156 lines (116 loc) · 5.53 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
143
144
145
146
147
148
149
150
151
152
153
154
155
156
import numpy as np
import math
import chainer
import chainer.functions as F
import chainer.links as L
from chainer import cuda, optimizers, Variable
class InstanceNormalization(links.Link):
def __init__(self, size, eps=2e-5, dtype=numpy.float32,):
super(InstanceNormalization, self).__init__()
self.add_param("gamma", size, dtype=dtype)
self.add_param("beta", size dtype=dtype)
self.eps = eps
def __call__(self, x):
xp = cuda.get_array_module(x.data)
mean = x.mean(axis=(2,3), keepdims=True)
var = x.var(axis=(2,3), keepdims=True)
normalized_x = (x - mean) / xp.sqrt(var + self.eps)
return self.gamma*normalized_x + self.beta
class Block(chainer.Chain):
def __init__(self, n_in, N):
super(Block,self).__init__(
c1 = L.Convolution2D(n_in, N, 3, stride=1, pad=1),
b1 = L.InstanceNormalization(N),
c2 = L.Convolution2D(N, N, 3, stride=1, pad=1),
b2 = L.InstanceNormalization(N),
c3 = L.Convolution2D(N, N, 1, stride=1, pad=0),
b3 = L.InstanceNormalization(N)
)
def __call__(self, x, test=False):
h = F.relu(self.b1(self.c1(x), test=test))
h = F.relu(self.b2(self.c2(h), test=test))
h = F.relu(self.b3(self.c3(h), test=test))
return h
class FaceSwapNet(chainer.Chain):
def __init__(self):
super(FaceSwapNet, self).__init__(
b1 = Block(3,32),
b2_1 = Block(3,32),
b2_2 = Block(64,64),
b3_1 = Block(3,32),
b3_2 = Block(96,96),
b4_1 = Block(3,32),
b4_2 = Block(128,128),
b5_1 = Block(3,32),
b5_2 = Block(160,160),
fin_conv = L.Convolution2D(160, 3, 1, stride=1, pad=0)
)
def __call__(self, x1, x2, x3, x4, x5, test=False):
h1 = self.b1(x1, test=test)
h1 = F.unpooling_2d(h1, ksize=2, stride=2, pad=0, cover_all=False)
h2 = self.b2_1(x2, test=test)
h2 = self.b2_2(F.concat([h1, h2]))
h2 = F.unpooling_2d(h2, ksize=2, stride=2, pad=0, cover_all=False)
del h1,x1
h3 = self.b3_1(x3, test=test)
h3 = self.b3_2(F.concat([h2, h3]))
h3 = F.unpooling_2d(h3, ksize=2, stride=2, pad=0, cover_all=False)
del h2,x2
h4 = self.b4_1(x4, test=test)
h4 = self.b4_2(F.concat([h3, h4]))
h4 = F.unpooling_2d(h4, ksize=2, stride=2, pad=0, cover_all=False)
del h3,x3
h5 = self.b5_1(x5, test=test)
h5 = self.b5_2(F.concat([h4, h5]))
del h4,x4
h5 = F.sigmoid(self.fin_conv(h5))
return h5*255
def local_patch(self, content, style_patch, style_patch_norm):
xp = cuda.get_array_module(content.data)
b,ch,h,w = content.data.shape
correlation = F.convolution_2d(Variable(content.data,volatile=True), W=style_patch_norm.data, stride=1, pad=0)
indices = xp.argmax(correlation.data, axis=1)
nearest_style_patch = style_patch.data.take(indices, axis=0).reshape(b,-1)
content = F.convolution_2d(content, W=Variable(xp.identity(ch*3*3,dtype=xp.float32).reshape((ch*3*3,ch,3,3))),stride=1,pad=0).transpose(0,2,3,1).reshape(b,-1)
style_loss = F.mean_squared_error(content, nearest_style_patch)
return style_loss
class VGG19(chainer.Chain):
def __init__(self):
super(VGG19, self).__init__(
conv1_1 = L.Convolution2D(3, 64, 3, stride=1, pad=1),
conv1_2 = L.Convolution2D(64, 64, 3, stride=1, pad=1),
conv2_1 = L.Convolution2D(64, 128, 3, stride=1, pad=1),
conv2_2 = L.Convolution2D(128, 128, 3, stride=1, pad=1),
conv3_1 = L.Convolution2D(128, 256, 3, stride=1, pad=1),
conv3_2 = L.Convolution2D(256, 256, 3, stride=1, pad=1),
conv3_3 = L.Convolution2D(256, 256, 3, stride=1, pad=1),
conv3_4 = L.Convolution2D(256, 256, 3, stride=1, pad=1),
conv4_1 = L.Convolution2D(256, 512, 3, stride=1, pad=1),
conv4_2 = L.Convolution2D(512, 512, 3, stride=1, pad=1),
#conv4_3 = L.Convolution2D(512, 512, 3, stride=1, pad=1),
#conv4_4 = L.Convolution2D(512, 512, 3, stride=1, pad=1),
#conv5_1 = L.Convolution2D(512, 512, 3, stride=1, pad=1),
#conv5_2 = L.Convolution2D(512, 512, 3, stride=1, pad=1),
#conv5_3 = L.Convolution2D(512, 512, 3, stride=1, pad=1),
#conv5_4 = L.Convolution2D(512, 512, 3, stride=1, pad=1),
)
self.mean = np.asarray([104, 117, 124], dtype=np.float32)
def vgg_preprocess(X, input_type="trans"):
if input_type=="trans":
X -= np.asarray([[[124]],[[117]],[[104]]], dtype=np.float32)
elif input_type=="RGB":
X = np.rollaxis(X[:,:,::-1]-np.asarray([104, 117, 124], dtype=np.float32),2)
return X
def __call__(self, x):
layer_names = ['1_1', '1_2', 'pool', '2_1', '2_2', 'pool', '3_1',
'3_2', '3_3', '3_4', 'pool', '4_1', '4_2']#, '4_3', '4_4',
#'pool', '5_1', '5_2', '5_3', '5_4']
layers = {}
h = x
for layer_name in layer_names:
if layer_name == 'pool':
h = F.max_pooling_2d(h, 2, stride=2)
else:
h = F.relu(self['conv' + layer_name](h))
layers[layer_name] = h
return layers