This repository has been archived by the owner on Oct 10, 2022. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 6
/
model_builder.py
287 lines (258 loc) · 8.69 KB
/
model_builder.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
"""Model builder
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from keras.layers import Activation, Dense, Input
from keras.layers import Conv2D, Flatten
from keras.layers import Conv2DTranspose, Dropout
from keras.layers import LeakyReLU
from keras.models import Model
from keras.layers.merge import concatenate
from keras.utils import plot_model
from keras_contrib.layers.normalization.instancenormalization import InstanceNormalization
import numpy as np
import argparse
def encoder_layer(inputs,
filters=16,
kernel_size=3,
strides=2,
activation='relu',
instance_norm=True,
postfix=None):
conv = Conv2D(filters=filters,
kernel_size=kernel_size,
strides=strides,
padding='same',
name='conv_'+postfix)
x = inputs
if instance_norm:
x = InstanceNormalization(name="in_"+postfix)(x)
if activation == 'relu':
x = Activation(activation, name='relu_'+postfix)(x)
else:
x = LeakyReLU(alpha=0.2, name='leaky_'+postfix)(x)
x = conv(x)
return x
def decoder_layer(inputs,
paired_inputs,
filters=16,
kernel_size=3,
strides=2,
activation='relu',
instance_norm=True,
postfix=None):
conv = Conv2DTranspose(filters=filters,
kernel_size=kernel_size,
strides=strides,
padding='same',
name='tconv_'+postfix)
x = inputs
if instance_norm:
x = InstanceNormalization(name='in_'+postfix)(x)
if activation == 'relu':
x = Activation(activation, name='relu_'+postfix)(x)
else:
x = LeakyReLU(alpha=0.2, name='lrelu_'+postfix)(x)
x = conv(x)
x = concatenate([x, paired_inputs], name='concat_'+postfix)
return x
def build_model(input_shape,
output_shape=None,
kernel_size=3,
name='pspu_skelnet'):
channels = int(output_shape[-1])
inputs0 = Input(shape=input_shape)
inputs1 = Input(shape=input_shape)
inputs2 = Input(shape=input_shape)
inputs3 = Input(shape=input_shape)
e11 = encoder_layer(inputs1,
32,
strides=1,
kernel_size=kernel_size,
postfix='e11')
# 256x256 x 32
e12 = encoder_layer(e11,
64,
kernel_size=kernel_size,
postfix='e12')
# 128x128 x 64
e13 = encoder_layer(e12,
128,
kernel_size=kernel_size,
postfix='e13')
# 64x64 x 128
e14 = encoder_layer(e13,
256,
kernel_size=kernel_size,
postfix='e14')
# 32x32 x 256
e15 = encoder_layer(e14,
512,
kernel_size=kernel_size,
postfix='e15')
# 16x16 x 512
e16 = encoder_layer(e15,
1024,
kernel_size=kernel_size,
postfix='e16')
# 8x8 x 1024
e17 = encoder_layer(e16,
2048,
kernel_size=kernel_size,
postfix='e17')
# 4x4 x 2048
d11 = decoder_layer(e17,
e16,
1024,
kernel_size=kernel_size,
postfix='d11')
# 8x8 x 1024+1024
d12 = decoder_layer(d11,
e15,
512,
kernel_size=kernel_size,
postfix='d12')
# 16x16 x 512+512
d13 = decoder_layer(d12,
e14,
256,
kernel_size=kernel_size,
postfix='d13')
# 32x32 x 256+256
d14 = decoder_layer(d13,
e13,
128,
kernel_size=kernel_size,
postfix='d14')
# 64x64 x 128+128
d15 = decoder_layer(d14,
e12,
64,
kernel_size=kernel_size,
postfix='d15')
# 128x128 x 64+64
d16 = decoder_layer(d15,
e11,
32,
kernel_size=kernel_size,
postfix='d16')
# 256x256 x 32+32
o1 = Conv2DTranspose(channels,
kernel_size=1,
strides=1,
padding='same',
name='tconv_o1')(d16)
# 256x256 x channels
e21 = encoder_layer(inputs2,
32,
strides=1,
kernel_size=kernel_size,
postfix='e21')
# 256x256 x 32
e22 = encoder_layer(e21,
64,
strides=4,
kernel_size=kernel_size,
postfix='e22')
# 64x64 x 64
e23 = encoder_layer(e22,
128,
strides=4,
kernel_size=kernel_size,
postfix='e23')
# 16x16 x 128
e24 = encoder_layer(e23,
256,
strides=4,
kernel_size=kernel_size,
postfix='e24')
# 4x4 x 256
d21 = decoder_layer(e24,
e23,
128,
strides=4,
kernel_size=kernel_size,
postfix='d21')
# 16x16 x 128+128
d22 = decoder_layer(d21,
e22,
64,
strides=4,
kernel_size=kernel_size,
postfix='d22')
# 64x64 x 64+64
d23 = decoder_layer(d22,
e21,
32,
strides=4,
kernel_size=kernel_size,
postfix='d23')
# 256x256 x 32+328
o2 = Conv2DTranspose(channels,
kernel_size=1,
strides=1,
padding='same',
name='tconv_o2')(d23)
# 256x256 x 1
e31 = encoder_layer(inputs3,
32,
strides=1,
kernel_size=kernel_size,
postfix='e31')
# 256x256 x 32
e32 = encoder_layer(e31,
64,
strides=8,
kernel_size=kernel_size,
postfix='e32')
# 32x32 x 64
e33 = encoder_layer(e32,
128,
strides=8,
kernel_size=kernel_size,
postfix='e33')
# 4x4 x 128
d31 = decoder_layer(e33,
e32,
64,
strides=8,
kernel_size=kernel_size,
postfix='d31')
# 32x32 x 64+64
d32 = decoder_layer(d31,
e31,
128,
strides=8,
kernel_size=kernel_size,
postfix='d32')
# 256x256 x 32+32
o3 = Conv2DTranspose(channels,
kernel_size=1,
strides=1,
padding='same',
name='tconv_o3')(d32)
y = concatenate([o1, o2, o3], name="concat_os")
y = Conv2DTranspose(32,
kernel_size=1,
strides=1,
padding='same',
name='tconv_pre')(y)
y = Conv2DTranspose(channels,
kernel_size=kernel_size,
strides=1,
activation='sigmoid',
padding='same',
name='tconv_out')(y)
outputs = y
inputs = [inputs0, inputs1, inputs2, inputs3]
model = Model(inputs, outputs, name=name)
return model
if __name__ == '__main__':
parser = argparse.ArgumentParser()
args = parser.parse_args()
input_shape = (256, 256, 1)
output_shape = (256, 256, 1)
model = build_model(input_shape, output_shape)
model.summary()
plot_model(model, to_file='model.png', show_shapes=True)