/
model_celeba.py
231 lines (180 loc) · 8.01 KB
/
model_celeba.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
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
import torch
import torch.nn as nn
import functools
import numpy as np
import torch.nn.functional as F
class residual_block(nn.Module):
def __init__(self,dim):
super(residual_block,self).__init__()
self.block= nn.Sequential(nn.ReflectionPad2d(1),#tf.pad(h, [[0, 0], [1, 1], [1, 1], [0, 0]], mode='REFLECT')
nn.Conv2d(dim,dim, kernel_size=3, padding=0, bias=False),#keras.layers.Conv2D(dim, 3, padding='valid', use_bias=False)(h)
nn.BatchNorm2d(dim),#Norm()(h)
nn.ReLU(),#tf.nn.relu(h)
nn.ReflectionPad2d(1),#tf.pad(h, [[0, 0], [1, 1], [1, 1], [0, 0]], mode='REFLECT')
nn.Conv2d(dim,dim, kernel_size=3, padding=0, bias=False),#keras.layers.Conv2D(dim, 3, padding='valid', use_bias=False)(h)
nn.BatchNorm2d(dim)#Norm()(h)
)
def forward(self,x):
identity=x
h=self.block(x)
h+=identity
return h
class equalized_transform(nn.Module):
def __init__(self,dim):
super(equalized_transform,self).__init__()
self.block= nn.Sequential(
nn.ReflectionPad2d(1),#tf.pad(h, [[0, 0], [1, 1], [1, 1], [0, 0]], mode='REFLECT')
nn.Conv2d(dim,dim, kernel_size=3, padding=0, bias=False),#keras.layers.Conv2D(dim, 3, padding='valid', use_bias=False)(h)
nn.BatchNorm2d(dim),#Norm()(h)
nn.ReLU(),#tf.nn.relu(h)
nn.ReflectionPad2d(1),#tf.pad(h, [[0, 0], [1, 1], [1, 1], [0, 0]], mode='REFLECT')
nn.Conv2d(dim,dim, kernel_size=3, padding=0, bias=False),#keras.layers.Conv2D(dim, 3, padding='valid', use_bias=False)(h)
nn.BatchNorm2d(dim)
)
def forward(self,x):
h=self.block(x)
return h
def from_RGB(output_nc):
h = nn.Sequential(nn.ReflectionPad2d(3),#tf.pad(h, [[0, 0], [3, 3], [3, 3], [0, 0]], mode='REFLECT')
nn.Conv2d(3, output_nc, kernel_size=7, padding=0, bias=False),
nn.BatchNorm2d(output_nc),#Norm()(h)
nn.ReLU()#tf.nn.relu(h)
)
return h
def to_RGB(input_nc):
h = nn.Sequential(nn.ReflectionPad2d(3),#tf.pad(h, [[0, 0], [3, 3], [3, 3], [0, 0]], mode='REFLECT')
nn.Conv2d(input_nc, 3, kernel_size=7, padding=0, bias=True),
nn.Tanh()#tf.nn.tanh(h)
)
return h
def down_sampling(input_nc, output_nc):
h = nn.Sequential(nn.Conv2d(input_nc, output_nc, kernel_size=3, stride=2,padding=1, bias=False),
nn.BatchNorm2d(output_nc),#Norm()(h)
nn.ReLU()
)
return h
def up_sampling(input_nc, output_nc):
h = nn.Sequential(nn.ConvTranspose2d(input_nc, output_nc, kernel_size=3, stride=2,bias=False, padding=1, output_padding=1),
nn.BatchNorm2d(output_nc),
nn.ReLU()
)
return h
def dis_from_RGB(output_nc):
h=nn.Sequential(nn.Conv2d(3, output_nc, kernel_size=4, stride=2, padding=1,bias=True),
nn.BatchNorm2d(output_nc),
nn.LeakyReLU(0.2)
)
return h
def dis_down_sampling(input_nc, output_nc,stride=2):
if stride==1:
h = nn.Sequential(nn.Conv2d(input_nc, output_nc, kernel_size=4, stride=stride, padding=1, bias=False),
nn.BatchNorm2d(output_nc),#Norm()(h)
nn.LeakyReLU(0.2)#tf.nn.relu(h)
)
else:
h = nn.Sequential(nn.Conv2d(input_nc, output_nc, kernel_size=4, stride=stride, padding=1,bias=False),
nn.BatchNorm2d(output_nc),#Norm()(h)
nn.LeakyReLU(0.2)#tf.nn.relu(h)
)
return h
def build_z(ngf):
h=[from_RGB(ngf)]
dim=ngf
for i in range(2):
h+=[down_sampling(dim,dim*2)]
dim=dim*2
for i in range(9):
h+= [residual_block(dim)]
h=nn.Sequential(*h)
return h #Nx3x32x32
def build_g(ngf):
dim=ngf*2**3
h=[]
for i in range(2):
h+=[up_sampling(dim,dim//2)]
dim=dim//2
h+=[to_RGB(dim)]
h=nn.Sequential(*h)
return h #Nx3x32x32
class build_d_backbone(nn.Module):
def __init__(self,ngf):
super(build_d_backbone,self).__init__()
dim=ngf
h=[dis_from_RGB(dim)]
for i in range(2):
h+=[dis_down_sampling(dim,dim*2,stride=2)]
dim*=2
h+=[dis_down_sampling(dim,dim*2,stride=1)]
self.net=nn.Sequential(*h)
def forward(self,x):
return self.net(x)
def build_score_map(dim):
h=nn.Conv2d(dim,1, kernel_size=4, stride=1, padding=1,bias=True)
return h
class build_classifier(nn.Module):
def __init__(self,dim,num_classes):
super(build_classifier,self).__init__()
h=[residual_block(dim)]
h+=[down_sampling(dim,2*dim)]
h+=[nn.AvgPool2d(8)]
h+=[nn.Conv2d(2*dim,num_classes, kernel_size=1,bias=True)]
self.net=nn.Sequential(*h)
def forward(self,x):
x=self.net(x).squeeze(3).squeeze(2)
return x
class build_encoder_classifier(nn.Module):
def __init__(self,dim,num_classes):
super(build_encoder_classifier,self).__init__()
h=[down_sampling(dim,2*dim)]
h+=[nn.AvgPool2d(16)]
h+=[nn.Conv2d(2*dim,num_classes, kernel_size=1,bias=True)]
self.net=nn.Sequential(*h)
def forward(self,x):
x=self.net(x).squeeze(3).squeeze(2)
return x
class Encoder(nn.Module):
def __init__(self,ngf=16,num_classes=[17,4]):
super(Encoder,self).__init__()
self.ngf=ngf
self.networkZ=build_z(ngf)
self.networkZ_body=build_z(ngf)
self.embeds=nn.Embedding(sum(num_classes),64,max_norm=100)
self.num_classes=num_classes
self.classifier=build_encoder_classifier(ngf*4,sum(num_classes))
self.filter=equalized_transform(ngf*4)
self.fc=nn.Linear(64,ngf*4*2)
def forward(self,imgs,attrs,mask=None):
#
z_body=self.networkZ_body((1-mask)*imgs)
z_before=self.networkZ(mask*imgs)
attr_logits_real=self.classifier(z_before)
z=self.filter(z_before)
attr_logits_fake=self.classifier(z)
cur_index=0
attr_feats=0
for i in range(len(self.num_classes)):
attr_feats+=self.embeds(cur_index+attrs[:,i])
cur_index+=self.num_classes[i]
attr_feats=self.fc(attr_feats).view(-1,self.ngf*4*2,1,1)
mean,std=torch.chunk(attr_feats, chunks=2, dim=1)
z=(1+std)*z+mean
z=torch.cat((z,z_body),dim=1)
return z,attr_logits_real,attr_logits_fake
class Generator(nn.Module):
def __init__(self,ngf=16):
super(Generator,self).__init__()
self.networkG= build_g(ngf)
def forward(self,z):
nimgs=self.networkG(z)
return nimgs
class Discriminator(nn.Module):
def __init__(self,ngf=16,num_classes=[17,4]):
super(Discriminator,self).__init__()
self.networkD_backbone=build_d_backbone(ngf)
self.build_score_map=build_score_map(ngf*8)
self.networkD_classifier=build_classifier(ngf*8,sum(num_classes))
def forward(self,imgs):
backbone=self.networkD_backbone(imgs)
score_map=self.build_score_map(backbone)
logits=self.networkD_classifier(backbone)
return score_map,logits