-
Notifications
You must be signed in to change notification settings - Fork 18
/
models.py
322 lines (267 loc) · 10.5 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
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
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
import torch
import torch.optim as optim
import torch.nn as nn
from torch import cat
import numpy as np
class GeneratorCNN_Pose_UAEAfterResidual_256(nn.Module):
def block(self, ch_in, ch_out, kernel, stride=1, padding=1):
return nn.Sequential(
nn.Conv2d(ch_in, ch_out, kernel, stride, padding),
nn.ReLU(),
nn.Conv2d(ch_out, ch_out, kernel, stride, padding),
nn.ReLU()
)
def block_one(self, ch_in, ch_out, kernel, stride=1, padding=1):
return nn.Sequential(
nn.Conv2d(ch_in, ch_out, kernel, stride, padding),
nn.ReLU()
)
def conv(self, ch_in, ch_out, kernel, stride=1, padding=1):
return nn.Conv2d(ch_in, ch_out, kernel, stride, padding)
def fc(self, ch_in, ch_out):
return nn.Linear(ch_in, ch_out)
def __init__(self, ch_in, z_num, repeat_num, hidden_num=128):
super(GeneratorCNN_Pose_UAEAfterResidual_256, self).__init__()
self.min_fea_map_H = 8
self.z_num = z_num
self.hidden_num = hidden_num
self.repeat_num = repeat_num
self.block_1 = self.block_one(ch_in, self.hidden_num, 3, 1)
self.block1 = self.block(self.hidden_num, 128, 3, 1)
self.block2 = self.block(256, 256, 3, 1)
self.block3 = self.block(384, 384, 3, 1)
self.block4 = self.block(512, 512, 3, 1)
self.block5 = self.block(640, 640, 3, 1)
self.block6 = self.block(768, 768, 3, 1)
self.block_one1 = self.block_one(128, 256, 3, 2)
self.block_one2 = self.block_one(256, 384, 3, 2)
self.block_one3 = self.block_one(384, 512, 3, 2)
self.block_one4 = self.block_one(512, 640, 3, 2)
self.block_one5 = self.block_one(640, 768, 3, 2)
self.fc1 = self.fc(self.min_fea_map_H * self.min_fea_map_H * 768, self.z_num)
self.fc2 = self.fc(self.z_num, self.min_fea_map_H * self.min_fea_map_H * self.hidden_num)
self.block7 = self.block(896, 896, 3, 1)
self.block8 = self.block(1280, 1280, 3, 1)
self.block9 = self.block(1024, 1024, 3, 1)
self.block10 = self.block(768, 768, 3, 1)
self.block11 = self.block(512, 512, 3, 1)
self.block12 = self.block(256, 256, 3, 1)
self.block_one6 = self.block_one(896, 640, 1, 1, padding=0)
self.block_one7 = self.block_one(1280, 512, 1, 1, padding=0)
self.block_one8 = self.block_one(1024, 384, 1, 1, padding=0)
self.block_one9 = self.block_one(768, 256, 1, 1, padding=0)
self.block_one10 = self.block_one(512, 128, 1, 1, padding=0)
self.conv_last = self.conv(256, 3, 3, 1)
self.upscale = nn.Upsample(scale_factor=2)
def forward(self, x):
encoder_layer_list = []
x = self.block_1(x)
# 1st encoding layer
res = x
x = self.block1(x)
x = x + res
encoder_layer_list.append(x)
x = self.block_one1(x)
# 2nd encoding layer
res = x
x = self.block2(x)
x = x + res
encoder_layer_list.append(x)
x = self.block_one2(x)
# 3rd encoding layer
res = x
x = self.block3(x)
x = x + res
encoder_layer_list.append(x)
x = self.block_one3(x)
# 4th encoding layer
res = x
x = self.block4(x)
x = x + res
encoder_layer_list.append(x)
x = self.block_one4(x)
# 5th encoding layer
res = x
x = self.block5(x)
x = x + res
encoder_layer_list.append(x)
x = self.block_one5(x)
# 6th encoding layer
res = x
x = self.block6(x)
x = x + res
encoder_layer_list.append(x)
x = x.view(-1, self.min_fea_map_H * self.min_fea_map_H * 768)
x = self.fc1(x)
z = x
x = self.fc2(z)
x = x.view(-1, self.hidden_num, self.min_fea_map_H, self.min_fea_map_H)
# 1st decoding layer
x = torch.cat([x, encoder_layer_list[5]], dim=1)
res = x
x = self.block7(x)
x = x + res
x = self.upscale(x)
x = self.block_one6(x)
# 2nd decoding layer
x = torch.cat([x, encoder_layer_list[4]], dim=1)
res = x
x = self.block8(x)
x = x + res
x = self.upscale(x)
x = self.block_one7(x)
# 3rd decoding layer
x = torch.cat([x, encoder_layer_list[3]], dim=1)
res = x
x = self.block9(x)
x = x + res
x = self.upscale(x)
x = self.block_one8(x)
# 4th decoding layer
x = torch.cat([x, encoder_layer_list[2]], dim=1)
res = x
x = self.block10(x)
x = x + res
x = self.upscale(x)
x = self.block_one9(x)
# 5th decoding layer
x = torch.cat([x, encoder_layer_list[1]], dim=1)
res = x
x = self.block11(x)
x = x + res
x = self.upscale(x)
x = self.block_one10(x)
# 6th decoding layer
x = torch.cat([x, encoder_layer_list[0]], dim=1)
res = x
x = self.block12(x)
x = x + res
output = self.conv_last(x)
return output
class UAE_noFC_AfterNoise(nn.Module):
def block(self, ch_in, ch_out, kernel, stride=1, padding=1):
return nn.Sequential(
nn.Conv2d(ch_in, ch_out, kernel, stride, padding),
nn.ReLU(),
nn.Conv2d(ch_out, ch_out, kernel, stride, padding),
nn.ReLU()
)
def block_one(self, ch_in, ch_out, kernel, stride=1, padding=1):
return nn.Sequential(
nn.Conv2d(ch_in, ch_out, kernel, stride, padding),
nn.ReLU()
)
def conv(self, ch_in, ch_out, kernel, stride=1, padding=1):
return nn.Conv2d(ch_in, ch_out, kernel, stride, padding)
def __init__(self, ch_in, repeat_num, hidden_num=128):
super(UAE_noFC_AfterNoise, self).__init__()
self.hidden_num = hidden_num
self.repeat_num = repeat_num
self.block_1 = self.block_one(ch_in, self.hidden_num, 3, 1)
self.block1 = self.block(self.hidden_num, 128, 3, 1)
self.block2 = self.block(128, 256, 3, 1)
self.block3 = self.block(256, 384, 3, 1)
self.block4 = self.block(384, 512, 3, 1)
self.block_one1 = self.block_one(128, 128, 3, 2)
self.block_one2 = self.block_one(256, 256, 3, 2)
self.block_one3 = self.block_one(384, 384, 3, 2)
self.block5 = self.block(1024, 128, 3, 1)
self.block6 = self.block(512, 128, 3, 1)
self.block7 = self.block(384, 128, 3, 1)
self.block8 = self.block(256, 128, 3, 1)
self.conv_last = self.conv(128, 3, 3, 1)
self.upscale = nn.Upsample(scale_factor=2)
def forward(self, x):
encoder_layer_list = []
x = self.block_1(x)
# 1st encoding layer
x = self.block1(x)
encoder_layer_list.append(x)
x = self.block_one1(x)
# 2nd encoding layer
x = self.block2(x)
encoder_layer_list.append(x)
x = self.block_one2(x)
# 3rd encoding layer
x = self.block3(x)
encoder_layer_list.append(x)
x = self.block_one3(x)
# 4th encoding layer
x = self.block4(x)
encoder_layer_list.append(x)
# 1st decoding layer
x = torch.cat([x, encoder_layer_list[-1]], dim=1)
x = self.block5(x)
x = self.upscale(x)
# 2nd decoding layer
x = torch.cat([x, encoder_layer_list[-2]], dim=1)
x = self.block6(x)
x = self.upscale(x)
# 3rd decoding layer
x = torch.cat([x, encoder_layer_list[-3]], dim=1)
x = self.block7(x)
x = self.upscale(x)
# 4th decoding layer
x = torch.cat([x, encoder_layer_list[-4]], dim=1)
x = self.block8(x)
output = self.conv_last(x)
return output
class DCGANDiscriminator_256(nn.Module):
def uniform(self, stdev, size):
return np.random.uniform(
low=-stdev * np.sqrt(3),
high=stdev * np.sqrt(3),
size=size
).astype('float32')
def LeakyReLU(self, x, alpha=0.2):
return torch.max(alpha*x, x)
def conv2d(self, x, input_dim, filter_size, output_dim, gain=1, stride=1, padding=2):
filter_values = self.uniform(
self._weights_stdev,
(output_dim, input_dim, filter_size, filter_size)
)
filter_values *= gain
filters = torch.from_numpy(filter_values)
biases = torch.from_numpy(np.zeros(output_dim, dtype='float32'))
if self.use_gpu:
filters = filters.cuda()
biases = biases.cuda()
result = nn.functional.conv2d(x, filters, biases, stride, padding)
return result
def LayerNorm(self, ch):
return nn.BatchNorm2d(ch)
def __init__(self, bn=True, input_dim=3, dim=64, _weights_stdev=0.02, use_gpu=True):
super(DCGANDiscriminator_256, self).__init__()
self.bn = bn
self.input_dim = input_dim
self.dim = dim
self._weights_stdev = _weights_stdev
self.use_gpu = use_gpu
self.bn1 = self.LayerNorm(2*self.dim)
self.bn2 = self.LayerNorm(4*self.dim)
self.bn3 = self.LayerNorm(8*self.dim)
self.bn4 = self.LayerNorm(8*self.dim)
self.fc1 = nn.Linear(8*8*8*self.dim, 1)
def forward(self, x):
output = x
output = self.conv2d(output, self.input_dim, 5, self.dim, stride=2)
output = self.LeakyReLU(output)
output = self.conv2d(output, self.dim, 5, 2*self.dim, stride=2)
if self.bn:
output = self.bn1(output)
output = self.LeakyReLU(output)
output = self.conv2d(output, 2*self.dim, 5, 4*self.dim, stride=2)
if self.bn:
output = self.bn2(output)
output = self.LeakyReLU(output)
output = self.conv2d(output, 4*self.dim, 5, 8*self.dim, stride=2)
if self.bn:
output = self.bn3(output)
output = self.LeakyReLU(output)
output = self.conv2d(output, 8*self.dim, 5, 8*self.dim, stride=2)
if self.bn:
output = self.bn4(output)
output = self.LeakyReLU(output)
output = output.view(-1, 8*8*8*self.dim)
output = self.fc1(output)
return output