-
Notifications
You must be signed in to change notification settings - Fork 3
/
generator.py
260 lines (193 loc) · 8.07 KB
/
generator.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
# coding: utf-8
from chainer import Chain
from chainer.backends import cuda
from chainer.initializers import HeNormal
import chainer.functions as F
import chainer.links as L
from architecture import ASPP, PixelShuffler
from spectral_norms import define_conv
from atrous_conv import define_atrous_conv
class ResNetDeepLab(Chain):
def __init__(self, opt):
super().__init__()
he_w = HeNormal()
with self.init_scope():
# This ResNet101 use a pre-trained caffemodel that can be downloaded at GitHub
# <https://github.com/KaimingHe/deep-residual-networks>.
self.resnet101 = L.ResNet101Layers()
self.use_layer = ('res3', 512)
nf = self.use_layer[1]
self.c1 = define_atrous_conv(opt)(nf, nf, ksize=3, rate=2, initialW=he_w)
self.norm1 = L.BatchNormalization(nf)
self.c2 = define_atrous_conv(opt)(nf, nf, ksize=3, rate=4, initialW=he_w)
self.norm2 = L.BatchNormalization(nf)
self.aspp = ASPP(opt, nf, input_resolution=32)
self.up1 = PixelShuffler(opt, nf, nf // 2, rate=2) # 32 -> 64
self.up2 = PixelShuffler(opt, nf // 2, nf // 4, rate=2) # 64 -> 128
self.up3 = PixelShuffler(opt, nf // 4, nf // 8, rate=2) # 128 -> 256
self.to_class = define_conv(opt)(nf // 8, opt.class_num, ksize=3, pad=1, initialW=he_w)
self.activation = F.leaky_relu
def prepare(self, variable_img):
# out = F.resize_images(variable_img, (224, 224))
out = variable_img
# out = (out + 1) * 0.5
out = out[:, ::-1, :, :]
out = F.transpose(out, (0, 2, 3, 1))
out *= 255
xp = cuda.get_array_module(variable_img.array)
out -= xp.array([103.063, 115.903, 123.152], dtype=variable_img.dtype)
out = F.transpose(out, (0, 3, 1, 2))
return out
def __call__(self, x):
x = self.prepare(x)
h = self.resnet101(x, [self.use_layer[0]])[self.use_layer[0]]
h = self.activation(h)
h = self.c1(h)
h = self.norm1(h)
h = self.activation(h)
h = self.c2(h)
h = self.norm2(h)
h = self.activation(h)
h = self.aspp(h)
h = self.up1(h)
h = self.activation(h)
h = self.up2(h)
h = self.activation(h)
h = self.up3(h)
h = self.activation(h)
out = self.to_class(h)
out = F.softmax(out, axis=1)
return out
class DilatedFCN(Chain):
def __init__(self, opt):
super().__init__()
he_w = HeNormal()
down_sampling_num = 3
ngf = opt.ngf
with self.init_scope():
# [input] 3 x 256 x 256
self.c1 = define_conv(opt)(opt.img_shape[0], ngf, ksize=4, stride=2, pad=1, initialW=he_w)
self.norm1 = L.BatchNormalization(ngf)
# [input] ngf x 128 x 128
self.c2 = define_conv(opt)(ngf, ngf * 2, ksize=4, stride=2, pad=1, initialW=he_w)
self.norm2 = L.BatchNormalization(ngf * 2)
# [input] ngf*2 x 64 x 64
self.c3 = define_conv(opt)(ngf * 2, ngf * 4, ksize=4, stride=2, pad=1, initialW=he_w)
self.norm3 = L.BatchNormalization(ngf * 4)
# [input] ngf*4 x 32 x 32
self.a1 = define_atrous_conv(opt)(ngf * 4, ngf * 4, ksize=3, rate=2, initialW=he_w)
self.norm4 = L.BatchNormalization(ngf * 4)
# [input] ngf*4 x 32 x 32
self.a2 = define_atrous_conv(opt)(ngf * 4, ngf * 4, ksize=3, rate=4, initialW=he_w)
self.norm5 = L.BatchNormalization(ngf * 4)
# [input] ngf*4 x 32 x 32
resolution = max(opt.img_shape[1], opt.img_shape[2]) // 2 ** down_sampling_num
self.aspp = ASPP(opt, ngf * 4, input_resolution=resolution)
# [input] ngf*4 x 32 x 32
self.up1 = PixelShuffler(opt, ngf * 4, ngf * 2, rate=2) # 64 -> 128
self.up2 = PixelShuffler(opt, ngf * 2, ngf, rate=2) # 128 -> 256
self.to_class = define_conv(opt)(ngf, opt.class_num, ksize=3, pad=1, initialW=he_w)
# [output] class_num x 256 x 256
self.activation = F.relu
def __call__(self, x):
h = self.c1(x)
h = self.norm1(h)
h = self.activation(h)
h = self.c2(h)
h = self.norm2(h)
h = self.activation(h)
h = self.c3(h)
h = self.norm3(h)
h = self.activation(h)
h = self.a1(h)
h = self.norm4(h)
h = self.activation(h)
h = self.a2(h)
h = self.norm5(h)
h = self.activation(h)
h = self.aspp(h)
h = self.up1(h)
h = self.activation(h)
h = self.up2(h)
h = self.activation(h)
out = self.to_class(h)
out = F.softmax(out, axis=1)
return out
class UNet(Chain):
def __init__(self, opt):
super().__init__()
he_w = HeNormal()
ngf = opt.ngf
with self.init_scope():
# Encoder
# [input] 3 x 256 x 256
self.e1 = define_conv(opt)(opt.input_ch, ngf, ksize=3, stride=1, pad=1, initialW=he_w)
self.e1_bn = L.BatchNormalization(ngf)
# [input] ngf x 256 x 256
self.e2 = define_conv(opt)(ngf, ngf * 2, ksize=4, stride=2, pad=1, initialW=he_w)
self.e2_bn = L.BatchNormalization(ngf * 2)
# [input] ngf*2 x 128 x 128
self.e3 = define_conv(opt)(ngf * 2, ngf * 4, ksize=4, stride=2, pad=1, initialW=he_w)
self.e3_bn = L.BatchNormalization(ngf * 4)
# [input] ngf*4 x 64 x 64
self.e4 = define_conv(opt)(ngf * 4, ngf * 8, ksize=4, stride=2, pad=1, initialW=he_w)
self.e4_bn = L.BatchNormalization(ngf * 8)
# [input] ngf*8 256 x 32 x 32
self.e5 = define_conv(opt)(ngf * 8, ngf * 16, ksize=4, stride=2, pad=1, initialW=he_w)
self.e5_bn = L.BatchNormalization(ngf * 16)
# Decoder
# [input] ngf*16 x 16 x 16
self.d1 = L.Deconvolution2D(ngf * 16, ngf * 8, ksize=4, stride=2, pad=1, initialW=he_w)
self.d1_bn = L.BatchNormalization(ngf * 8)
# [input] ngf*8*2 x 32 x 32 (concat)
self.d2 = L.Deconvolution2D(ngf * 8 * 2, ngf * 4, ksize=4, stride=2, pad=1, initialW=he_w)
self.d2_bn = L.BatchNormalization(ngf * 4)
# [input] ngf*4*2 x 64 x 64 (concat)
self.d3 = L.Deconvolution2D(ngf * 4 * 2, ngf * 2, ksize=4, stride=2, pad=1, initialW=he_w)
self.d3_bn = L.BatchNormalization(ngf * 2)
# [input] ngf*2*2 x 128 x 128 (concat)
self.d4 = L.Deconvolution2D(ngf * 2 * 2, ngf, ksize=4, stride=2, pad=1, initialW=he_w)
self.d4_bn = L.BatchNormalization(ngf)
# [input] ngf x 256 x 256
self.to_class = define_conv(opt)(ngf, opt.nclass, ksize=3, pad=1, initialW=he_w)
# [output] nclass x 256 x 256
self.activation = F.relu
def __call__(self, x):
# Encoder
eh1 = self.e1(x)
eh1 = self.e1_bn(eh1)
eh1 = self.activation(eh1)
eh2 = self.e2(eh1)
eh2 = self.e2_bn(eh2)
eh2 = self.activation(eh2)
eh3 = self.e3(eh2)
eh3 = self.e3_bn(eh3)
eh3 = self.activation(eh3)
eh4 = self.e4(eh3)
eh4 = self.e4_bn(eh4)
eh4 = self.activation(eh4)
eh5 = self.e5(eh4)
eh5 = self.e5_bn(eh5)
eh5 = self.activation(eh5)
# Decoder
dh1 = self.d1(eh5)
dh1 = self.d1_bn(dh1)
dh1 = F.dropout(dh1)
dh1 = self.activation(dh1)
dh2 = F.concat((eh4, dh1), axis=1)
dh2 = self.d2(dh2)
dh2 = self.d2_bn(dh2)
dh2 = F.dropout(dh2)
dh2 = self.activation(dh2)
dh3 = F.concat((eh3, dh2), axis=1)
dh3 = self.d3(dh3)
dh3 = self.d3_bn(dh3)
dh3 = F.dropout(dh3)
dh3 = self.activation(dh3)
dh4 = F.concat((eh2, dh3), axis=1)
dh4 = self.d4(dh4)
dh4 = self.d4_bn(dh4)
dh4 = self.activation(dh4)
out = self.to_class(dh4)
out = F.softmax(out, axis=1)
return out