-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmodel.py
271 lines (189 loc) · 10.7 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
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
import torch.nn as nn
from torch.nn import init
from torch.nn.modules import conv, Linear
import torch.nn.functional as F
import torch
from collections import OrderedDict
import math
import torch.nn.utils.spectral_norm as spectral_norm
def weights_init_normal(m):
classname = m.__class__.__name__
if classname.find("Conv")!=-1:
m.weight.data.normal_(0,0.01)
elif classname.find("Linear")!=-1:
m.weight.data.normal_(0,0.01)
elif classname.find("BatchNorm")!=-1:
m.weight.data.normal_(1.0,0.01)
#m.bias.data.constant_(0.0)
def weight_init_kaiming(m):
classname = m.__class__.__name__
if classname.find("Conv")!=-1:
init.kaiming_normal(m.weight.data,a=0,mode="fan_in")
elif classname.find("Linear")!=-1:
init.kaiming_normal(m.weight.data,a=0,mode="fan_in")
elif classname.find("BatchNorm")!=-1:
init.normal(m.weight.data,1.0,0.02)
init.constant(m.bias.data,0.0)
def BuildResidualBlock(channels,dropout,kernel,depth,bias):
layers = []
for i in range(int(depth)):
layers += [nn.Conv3d(channels,channels,kernel_size=kernel,stride=1,padding=kernel//2,bias=bias),
#nn.BatchNorm3d(channels),
nn.ReLU(True)]
if dropout:
layers += [nn.Dropout(0.5)]
layers += [nn.Conv3d(channels,channels,kernel_size=kernel,stride=1,padding=kernel//2,bias=bias),
#nn.BatchNorm3d(channels),
]
return nn.Sequential(*layers)
class ResidualBlock(nn.Module):
def __init__(self,channels,dropout,kernel,depth,bias):
super(ResidualBlock,self).__init__()
self.block = BuildResidualBlock(channels,dropout,kernel,depth,bias)
def forward(self,x):
out = x+self.block(x)
return out
class Block(nn.Module):
def __init__(self,inchannels,outchannels,dropout,kernel,bias,depth,mode,factor=2):
super(Block,self).__init__()
layers = []
for i in range(int(depth)):
layers += [
spectral_norm(nn.Conv3d(inchannels,inchannels,kernel_size=kernel,padding=kernel//2,bias=bias),eps=1e-4),
nn.InstanceNorm3d(inchannels),
nn.LeakyReLU(0.2,inplace=True)]
if dropout:
layers += [nn.Dropout(0.5)]
self.model = nn.Sequential(*layers)
if factor == 2:
self.stride = 2
self.padding = 1
elif factor == 4:
self.stride = 1
self.padding = 0
if mode == 'down':
self.conv1 = spectral_norm(nn.Conv3d(inchannels,outchannels,4,stride=self.stride,padding=self.padding,bias=bias),eps=1e-4)
self.conv2 = spectral_norm(nn.Conv3d(inchannels,outchannels,4,stride=self.stride,padding=self.padding,bias=bias),eps=1e-4)
elif mode == 'up':
self.conv1 = spectral_norm(nn.ConvTranspose3d(inchannels,outchannels,4,stride=self.stride,padding=self.padding,bias=bias),eps=1e-4)
self.conv2 = spectral_norm(nn.ConvTranspose3d(inchannels,outchannels,4,stride=self.stride,padding=self.padding,bias=bias),eps=1e-4)
elif mode == 'same':
self.conv1 = nn.Sequential(*[nn.ReplicationPad3d(kernel//2),
spectral_norm(nn.Conv3d(inchannels,outchannels,kernel_size=kernel,bias=bias),eps=1e-4)
])
self.conv2 = nn.Sequential(*[nn.ReplicationPad3d(kernel//2),
spectral_norm(nn.Conv3d(inchannels,outchannels,kernel_size=kernel,bias=bias),eps=1e-4)
])
self.conv3 = nn.Sequential(*[nn.ReplicationPad3d(kernel//2),
spectral_norm(nn.Conv3d(2*outchannels,outchannels,kernel_size=kernel,bias=bias),eps=1e-4)
])
def forward(self,x):
y = self.model(x)
y = self.conv1(y)
x = self.conv2(x)
#return self.conv3(torch.cat([x,y],dim=1))
return x+y
class kV2V(nn.Module):
def __init__(self,inc,outc,init_channels):
super(kV2V,self).__init__()
### first model
self.input_pool = nn.Conv3d(inc,init_channels,4,2,1)
self.conv1 = nn.Conv3d(init_channels,init_channels,4,2,1)
self.conv2 = nn.Conv3d(init_channels,2*init_channels,4,2,1)
self.conv3 = nn.Conv3d(2*init_channels,4*init_channels,4,2,1)
self.conv4 = nn.Conv3d(4*init_channels,8*init_channels,4,2,1)
self.rb1 = ResidualBlock(8*init_channels,False,3,2,False)
self.rb2 = ResidualBlock(8*init_channels,False,3,2,False)
self.rb3 = ResidualBlock(8*init_channels,False,3,2,False)
self.b3 = Block(inchannels=4*init_channels,outchannels=4*init_channels,dropout=False,kernel=3,bias=False,depth=2,mode='same',factor=2)
self.b2 = Block(inchannels=2*init_channels,outchannels=2*init_channels,dropout=False,kernel=3,bias=False,depth=2,mode='same',factor=2)
self.b1 = Block(inchannels=init_channels,outchannels=init_channels,dropout=False,kernel=3,bias=False,depth=2,mode='same',factor=2)
### decoder
self.deconv41 = nn.ConvTranspose3d(8*init_channels,(8*init_channels)//2,4,2,1)
self.conv_u41 = nn.Conv3d(11*init_channels,(8*init_channels)//2,3,1,1)
self.deconv31 = nn.ConvTranspose3d((8*init_channels)//2,(8*init_channels)//4,4,2,1)
self.conv_u31 = nn.Conv3d(9*init_channels,(8*init_channels)//4,3,1,1)
self.deconv21 = nn.ConvTranspose3d((8*init_channels)//4,(8*init_channels)//8,4,2,1)
self.conv_u21 = nn.Conv3d(8*init_channels,(8*init_channels)//8,3,1,1)
self.deconv11 = nn.ConvTranspose3d((8*init_channels)//8,(8*init_channels)//16,4,2,1)
self.conv_u11 = nn.Conv3d((8*init_channels)//16,(8*init_channels)//16,3,1,1)
self.pool = nn.MaxPool3d(2, stride=2)
######################## second model
self.conv1_main = nn.Conv3d(inc,init_channels,4,2,1)
self.conv2_main = nn.Conv3d(init_channels+(8*init_channels)//16,2*init_channels,4,2,1)
self.conv3_main = nn.Conv3d(2*init_channels+(8*init_channels)//8,4*init_channels,4,2,1)
self.conv4_main = nn.Conv3d(4*init_channels+(8*init_channels)//4,8*init_channels,4,2,1)
self.rb1_main = ResidualBlock(8*init_channels+(8*init_channels)//2,False,3,2,False)
self.rb2_main = ResidualBlock(8*init_channels+(8*init_channels)//2,False,3,2,False)
self.rb3_main = ResidualBlock(8*init_channels+(8*init_channels)//2,False,3,2,False)
self.b3_main = Block(inchannels=4*init_channels,outchannels=4*init_channels,dropout=False,kernel=3,bias=False,depth=2,mode='same',factor=2)
self.b2_main = Block(inchannels=2*init_channels,outchannels=2*init_channels,dropout=False,kernel=3,bias=False,depth=2,mode='same',factor=2)
self.b1_main = Block(inchannels=init_channels,outchannels=init_channels,dropout=False,kernel=3,bias=False,depth=2,mode='same',factor=2)
# decoder
self.deconv41_main = nn.ConvTranspose3d(8*init_channels+(8*init_channels)//2,(8*init_channels)//2,4,2,1)
self.conv_u41_main = nn.Conv3d(11*init_channels,(8*init_channels)//2,3,1,1)
self.deconv31_main = nn.ConvTranspose3d((8*init_channels)//2,(8*init_channels)//4,4,2,1)
self.conv_u31_main = nn.Conv3d(9*init_channels,(8*init_channels)//4,3,1,1)
self.deconv21_main = nn.ConvTranspose3d((8*init_channels)//4,(8*init_channels)//8,4,2,1)
self.conv_u21_main = nn.Conv3d(8*init_channels,(8*init_channels)//8,3,1,1)
self.deconv11_main = nn.ConvTranspose3d((8*init_channels)//8,(8*init_channels)//16,4,2,1)
self.conv_u11_main = nn.Conv3d((8*init_channels)//16,outc,3,1,1)
def forward(self,x):
# first model
x_pool = F.relu(self.input_pool(x))
x1 = F.relu(self.conv1(x_pool))
x2 = F.relu(self.conv2(x1))
x3 = F.relu(self.conv3(x2))
x4 = F.relu(self.conv4(x3))
x4 = self.rb1(x4)
x4 = self.rb2(x4)
x4 = self.rb3(x4)
x1 = self.b1(x1)
x2 = self.b2(x2)
x3 = self.b3(x3)
# decoder
u11 = F.relu(self.deconv41(x4))
x1_pool1 = self.pool(x1)
x1_pool2 = self.pool(x1_pool1)
x2_pool1 = self.pool(x2)
u11 = F.relu(self.conv_u41(torch.cat((u11,x3,x2_pool1,x1_pool2),dim=1)))
u21 = F.relu(self.deconv31(u11)) # 2*init_channels
x3_up1 = F.interpolate(x3, scale_factor=2.0, mode='trilinear', align_corners=True)
u21 = F.relu(self.conv_u31(torch.cat((u21,x2,x1_pool1,x3_up1),dim=1)))
u31 = F.relu(self.deconv21(u21)) # init_channels
x3_up2 = F.interpolate(x3, scale_factor=4.0, mode='trilinear', align_corners=True)
x2_up1 = F.interpolate(x2, scale_factor=2.0, mode='trilinear', align_corners=True)
u31 = F.relu(self.conv_u21(torch.cat((u31,x1,x2_up1,x3_up2),dim=1)))
u41 = F.relu(self.deconv11(u31))
out = self.conv_u11(u41)
#### second model
x1_main = F.relu(self.conv1_main(x))
fuse1 = torch.cat([out, x1_main], 1) #(8*init_channels)//16+
x2_main = F.relu(self.conv2_main(fuse1))
fuse2 = torch.cat([u31, x2_main], 1) #
x3_main = F.relu(self.conv3_main(fuse2))
fuse3 = torch.cat([u21, x3_main], 1)
x4_main = F.relu(self.conv4_main(fuse3))
fuse4 = torch.cat([u11, x4_main], 1)
x4_main = self.rb1_main(fuse4)
x4_main = self.rb2_main(x4_main)
x4_main = self.rb3_main(x4_main)
x1_main = self.b1_main(x1_main)
x2_main = self.b2_main(x2_main)
x3_main = self.b3_main(x3_main)
# decoder
u11_main = F.relu(self.deconv41_main(x4_main))
x1_pool1_main = self.pool(x1_main)
x1_pool2_main = self.pool(x1_pool1_main)
x2_pool1_main = self.pool(x2_main)
u11_main = F.relu(self.conv_u41_main(torch.cat((u11_main,x3_main,x2_pool1_main,x1_pool2_main),dim=1)))
u21_main = F.relu(self.deconv31_main(u11_main)) # 2*init_channels
x3_up1_main = F.interpolate(x3_main, scale_factor=2.0, mode='trilinear', align_corners=True)
u21_main = F.relu(self.conv_u31_main(torch.cat((u21_main,x2_main,x1_pool1_main,x3_up1_main),dim=1)))
u31_main = F.relu(self.deconv21_main(u21_main)) # init_channels
x3_up2_main = F.interpolate(x3_main, scale_factor=4.0, mode='trilinear', align_corners=True)
x2_up1_main = F.interpolate(x2_main, scale_factor=2.0, mode='trilinear', align_corners=True)
u31_main = F.relu(self.conv_u21_main(torch.cat((u31_main,x1_main,x2_up1_main,x3_up2_main),dim=1)))
u41_main = F.relu(self.deconv11_main(u31_main))
out_main = self.conv_u11_main(u41_main)
return out_main