-
Notifications
You must be signed in to change notification settings - Fork 13
/
IVD_Net.py
316 lines (236 loc) · 13.9 KB
/
IVD_Net.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
from Blocks import *
import torch.nn.init as init
import pdb
import math
import torch.nn.functional as F
import numpy as np
def croppCenter(tensorToCrop,finalShape):
org_shape = tensorToCrop.shape
diff = np.zeros(3)
diff[0] = org_shape[2] - finalShape[2]
diff[1] = org_shape[3] - finalShape[3]
croppBorders = np.zeros(2,dtype=int)
croppBorders[0] = int(diff[0]/2)
croppBorders[1] = int(diff[1]/2)
return tensorToCrop[:,
:,
croppBorders[0]:org_shape[2]-croppBorders[0],
croppBorders[1]:org_shape[3]-croppBorders[1]]
class Conv_residual_conv_Inception_Dilation(nn.Module):
def __init__(self,in_dim,out_dim,act_fn):
super(Conv_residual_conv_Inception_Dilation,self).__init__()
self.in_dim = in_dim
self.out_dim = out_dim
act_fn = act_fn
self.conv_1 = conv_block(self.in_dim,self.out_dim,act_fn)
self.conv_2_1 = conv_block(self.out_dim, self.out_dim, act_fn, kernel_size=1, stride=1, padding=0, dilation=1 )
self.conv_2_2 = conv_block(self.out_dim, self.out_dim, act_fn, kernel_size=3, stride=1, padding=1, dilation=1 )
self.conv_2_3 = conv_block(self.out_dim, self.out_dim, act_fn, kernel_size=5, stride=1, padding=2, dilation=1 )
self.conv_2_4 = conv_block(self.out_dim, self.out_dim, act_fn, kernel_size=3, stride=1, padding=2, dilation=2 )
self.conv_2_5 = conv_block(self.out_dim, self.out_dim, act_fn, kernel_size=3, stride=1, padding=4, dilation=4 )
self.conv_2_output = conv_block(self.out_dim*5, self.out_dim, act_fn, kernel_size=1, stride=1, padding=0, dilation=1 )
self.conv_3 = conv_block(self.out_dim,self.out_dim,act_fn)
def forward(self,input):
conv_1 = self.conv_1(input)
conv_2_1 = self.conv_2_1(conv_1)
conv_2_2 = self.conv_2_2(conv_1)
conv_2_3 = self.conv_2_3(conv_1)
conv_2_4 = self.conv_2_4(conv_1)
conv_2_5 = self.conv_2_5(conv_1)
out1 = torch.cat([conv_2_1, conv_2_2, conv_2_3, conv_2_4, conv_2_5], 1)
out1 = self.conv_2_output(out1)
conv_3 = self.conv_3(out1+conv_1)
return conv_3
class Conv_residual_conv_Inception_Dilation_asymmetric(nn.Module):
def __init__(self,in_dim,out_dim,act_fn):
super(Conv_residual_conv_Inception_Dilation_asymmetric,self).__init__()
self.in_dim = in_dim
self.out_dim = out_dim
act_fn = act_fn
self.conv_1 = conv_block(self.in_dim,self.out_dim,act_fn)
self.conv_2_1 = conv_block_Asym_Inception(self.out_dim, self.out_dim, act_fn, kernel_size=1, stride=1, padding=0, dilation=1 )
self.conv_2_2 = conv_block_Asym_Inception(self.out_dim, self.out_dim, act_fn, kernel_size=3, stride=1, padding=1, dilation=1 )
self.conv_2_3 = conv_block_Asym_Inception(self.out_dim, self.out_dim, act_fn, kernel_size=5, stride=1, padding=2, dilation=1 )
self.conv_2_4 = conv_block_Asym_Inception(self.out_dim, self.out_dim, act_fn, kernel_size=3, stride=1, padding=2, dilation=2 )
self.conv_2_5 = conv_block_Asym_Inception(self.out_dim, self.out_dim, act_fn, kernel_size=3, stride=1, padding=4, dilation=4 )
self.conv_2_output = conv_block(self.out_dim*5, self.out_dim, act_fn, kernel_size=1, stride=1, padding=0, dilation=1 )
self.conv_3 = conv_block(self.out_dim,self.out_dim,act_fn)
def forward(self,input):
conv_1 = self.conv_1(input)
conv_2_1 = self.conv_2_1(conv_1)
conv_2_2 = self.conv_2_2(conv_1)
conv_2_3 = self.conv_2_3(conv_1)
conv_2_4 = self.conv_2_4(conv_1)
conv_2_5 = self.conv_2_5(conv_1)
out1 = torch.cat([conv_2_1, conv_2_2, conv_2_3, conv_2_4, conv_2_5], 1)
out1 = self.conv_2_output(out1)
conv_3 = self.conv_3(out1+conv_1)
return conv_3
class IVD_Net_asym(nn.Module):
def __init__(self,input_nc, output_nc, ngf):
super(IVD_Net_asym,self).__init__()
print('~'*50)
print(' ----- Creating FUSION_NET HD (Assymetric) network...')
print('~'*50)
self.in_dim = input_nc
self.out_dim = ngf
self.final_out_dim = output_nc
#act_fn = nn.LeakyReLU(0.2, inplace=True)
act_fn = nn.ReLU()
act_fn_2 = nn.ReLU()
# ~~~ Encoding Paths ~~~~~~ #
# Encoder (Modality 1)
self.down_1_0 = Conv_residual_conv_Inception_Dilation_asymmetric(self.in_dim, self.out_dim, act_fn)
self.pool_1_0 = maxpool()
self.down_2_0 = Conv_residual_conv_Inception_Dilation_asymmetric(self.out_dim * 4, self.out_dim * 2, act_fn)
self.pool_2_0 = maxpool()
self.down_3_0 = Conv_residual_conv_Inception_Dilation_asymmetric(self.out_dim * 12, self.out_dim * 4, act_fn)
self.pool_3_0 = maxpool()
self.down_4_0 = Conv_residual_conv_Inception_Dilation_asymmetric(self.out_dim * 28, self.out_dim * 8, act_fn)
self.pool_4_0 = maxpool()
# Encoder (Modality 2)
self.down_1_1 = Conv_residual_conv_Inception_Dilation_asymmetric(self.in_dim, self.out_dim, act_fn)
self.pool_1_1 = maxpool()
self.down_2_1 = Conv_residual_conv_Inception_Dilation_asymmetric(self.out_dim * 4, self.out_dim * 2, act_fn)
self.pool_2_1 = maxpool()
self.down_3_1 = Conv_residual_conv_Inception_Dilation_asymmetric(self.out_dim * 12, self.out_dim * 4, act_fn)
self.pool_3_1 = maxpool()
self.down_4_1 = Conv_residual_conv_Inception_Dilation_asymmetric(self.out_dim * 28, self.out_dim * 8, act_fn)
self.pool_4_1 = maxpool()
# Encoder (Modality 3)
self.down_1_2 = Conv_residual_conv_Inception_Dilation_asymmetric(self.in_dim, self.out_dim, act_fn)
self.pool_1_2 = maxpool()
self.down_2_2 = Conv_residual_conv_Inception_Dilation_asymmetric(self.out_dim * 4, self.out_dim * 2, act_fn)
self.pool_2_2 = maxpool()
self.down_3_2 = Conv_residual_conv_Inception_Dilation_asymmetric(self.out_dim * 12, self.out_dim * 4, act_fn)
self.pool_3_2 = maxpool()
self.down_4_2 = Conv_residual_conv_Inception_Dilation_asymmetric(self.out_dim * 28, self.out_dim * 8, act_fn)
self.pool_4_2 = maxpool()
# Encoder (Modality 4)
self.down_1_3 = Conv_residual_conv_Inception_Dilation_asymmetric(self.in_dim, self.out_dim, act_fn)
self.pool_1_3 = maxpool()
self.down_2_3 = Conv_residual_conv_Inception_Dilation_asymmetric(self.out_dim * 4, self.out_dim * 2, act_fn)
self.pool_2_3 = maxpool()
self.down_3_3 = Conv_residual_conv_Inception_Dilation_asymmetric(self.out_dim * 12, self.out_dim * 4, act_fn)
self.pool_3_3 = maxpool()
self.down_4_3 = Conv_residual_conv_Inception_Dilation_asymmetric(self.out_dim * 28, self.out_dim * 8, act_fn)
self.pool_4_3 = maxpool()
# Bridge between Encoder-Decoder
self.bridge = Conv_residual_conv_Inception_Dilation_asymmetric(self.out_dim * 60, self.out_dim * 16, act_fn)
# ~~~ Decoding Path ~~~~~~ #
self.deconv_1 = conv_decod_block(self.out_dim * 16, self.out_dim * 8, act_fn_2)
self.up_1 = Conv_residual_conv_Inception_Dilation(self.out_dim * 8, self.out_dim * 8, act_fn_2)
self.deconv_2 = conv_decod_block(self.out_dim * 8, self.out_dim * 4, act_fn_2)
self.up_2 = Conv_residual_conv_Inception_Dilation(self.out_dim * 4, self.out_dim * 4, act_fn_2)
self.deconv_3 = conv_decod_block(self.out_dim * 4, self.out_dim * 2, act_fn_2)
self.up_3 = Conv_residual_conv_Inception_Dilation(self.out_dim * 2, self.out_dim * 2, act_fn_2)
self.deconv_4 = conv_decod_block(self.out_dim * 2, self.out_dim, act_fn_2)
self.up_4 = Conv_residual_conv_Inception_Dilation(self.out_dim, self.out_dim, act_fn_2)
self.out = nn.Conv2d(self.out_dim,self.final_out_dim, kernel_size=3, stride=1, padding=1)
# Params initialization
for m in self.modules():
if isinstance(m, nn.Conv2d):
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(0, math.sqrt(2. / n))
#init.xavier_uniform(m.weight.data)
#init.xavier_uniform(m.bias.data)
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
def forward(self,input):
# ############################# #
# ~~~~~~ Encoding path ~~~~~~~ #
i0 = input[:,0:1,:,:]
i1 = input[:,1:2,:,:]
i2 = input[:,2:3,:,:]
i3 = input[:,3:4,:,:]
# ----- First Level --------
down_1_0 = self.down_1_0(i0)
down_1_1 = self.down_1_1(i1)
down_1_2 = self.down_1_2(i2)
down_1_3 = self.down_1_3(i3)
# ----- Second Level --------
#input_2nd = torch.cat((down_1_0,down_1_1,down_1_2,down_1_3),dim=1)
input_2nd_0 = torch.cat((self.pool_1_0(down_1_0),
self.pool_1_1(down_1_1),
self.pool_1_2(down_1_2),
self.pool_1_3(down_1_3)),dim=1)
input_2nd_1 = torch.cat((self.pool_1_1(down_1_1),
self.pool_1_2(down_1_2),
self.pool_1_3(down_1_3),
self.pool_1_0(down_1_0)),dim=1)
input_2nd_2 = torch.cat((self.pool_1_2(down_1_2),
self.pool_1_3(down_1_3),
self.pool_1_0(down_1_0),
self.pool_1_1(down_1_1)),dim=1)
input_2nd_3 = torch.cat((self.pool_1_3(down_1_3),
self.pool_1_0(down_1_0),
self.pool_1_1(down_1_1),
self.pool_1_2(down_1_2)),dim=1)
down_2_0 = self.down_2_0(input_2nd_0)
down_2_1 = self.down_2_1(input_2nd_1)
down_2_2 = self.down_2_2(input_2nd_2)
down_2_3 = self.down_2_3(input_2nd_3)
# ----- Third Level --------
# Max-pool
down_2_0m = self.pool_2_0(down_2_0)
down_2_1m = self.pool_2_0(down_2_1)
down_2_2m = self.pool_2_0(down_2_2)
down_2_3m = self.pool_2_0(down_2_3)
input_3rd_0 = torch.cat((down_2_0m,down_2_1m,down_2_2m,down_2_3m),dim=1)
input_3rd_0 = torch.cat((input_3rd_0,croppCenter(input_2nd_0, input_3rd_0.shape)), dim=1)
input_3rd_1 = torch.cat((down_2_1m,down_2_2m,down_2_3m,down_2_0m),dim=1)
input_3rd_1 = torch.cat((input_3rd_1,croppCenter(input_2nd_1, input_3rd_1.shape)), dim=1)
input_3rd_2 = torch.cat((down_2_2m,down_2_3m,down_2_0m,down_2_1m),dim=1)
input_3rd_2 = torch.cat((input_3rd_2,croppCenter(input_2nd_2, input_3rd_2.shape)), dim=1)
input_3rd_3 = torch.cat((down_2_3m,down_2_0m,down_2_1m,down_2_2m),dim=1)
input_3rd_3 = torch.cat((input_3rd_3,croppCenter(input_2nd_3, input_3rd_3.shape)), dim=1)
down_3_0 = self.down_3_0(input_3rd_0)
down_3_1 = self.down_3_1(input_3rd_1)
down_3_2 = self.down_3_2(input_3rd_2)
down_3_3 = self.down_3_3(input_3rd_3)
# ----- Fourth Level --------
# Max-pool
down_3_0m = self.pool_3_0(down_3_0)
down_3_1m = self.pool_3_0(down_3_1)
down_3_2m = self.pool_3_0(down_3_2)
down_3_3m = self.pool_3_0(down_3_3)
input_4th_0 = torch.cat((down_3_0m,down_3_1m,down_3_2m,down_3_3m),dim=1)
input_4th_0 = torch.cat((input_4th_0,croppCenter(input_3rd_0, input_4th_0.shape)), dim=1)
input_4th_1 = torch.cat((down_3_1m,down_3_2m,down_3_3m,down_3_0m),dim=1)
input_4th_1 = torch.cat((input_4th_1,croppCenter(input_3rd_1, input_4th_1.shape)), dim=1)
input_4th_2 = torch.cat((down_3_2m,down_3_3m,down_3_0m,down_3_1m),dim=1)
input_4th_2 = torch.cat((input_4th_2,croppCenter(input_3rd_2, input_4th_2.shape)), dim=1)
input_4th_3 = torch.cat((down_3_3m,down_3_0m,down_3_1m,down_3_2m),dim=1)
input_4th_3 = torch.cat((input_4th_3,croppCenter(input_3rd_3, input_4th_3.shape)), dim=1)
down_4_0 = self.down_4_0(input_4th_0)
down_4_1 = self.down_4_1(input_4th_1)
down_4_2 = self.down_4_2(input_4th_2)
down_4_3 = self.down_4_3(input_4th_3)
#----- Bridge -----
# Max-pool
down_4_0m = self.pool_4_0(down_4_0)
down_4_1m = self.pool_4_0(down_4_1)
down_4_2m = self.pool_4_0(down_4_2)
down_4_3m = self.pool_4_0(down_4_3)
inputBridge = torch.cat((down_4_0m,down_4_1m,down_4_2m,down_4_3m),dim=1)
inputBridge = torch.cat((inputBridge,croppCenter(input_4th_0, inputBridge.shape)), dim=1)
bridge = self.bridge(inputBridge)
#
# ############################# #
# ~~~~~~ Decoding path ~~~~~~~ #
deconv_1 = self.deconv_1(bridge)
skip_1 = (deconv_1 + down_4_0 + down_4_1 + down_4_2 + down_4_3)/5 # Residual connection
up_1 = self.up_1(skip_1)
deconv_2 = self.deconv_2(up_1)
skip_2 = (deconv_2 + down_3_0 + down_3_1 + down_3_2 + down_3_3)/5 # Residual connection
up_2 = self.up_2(skip_2)
deconv_3 = self.deconv_3(up_2)
skip_3 = (deconv_3 + down_2_0 + down_2_1 + down_2_2 + down_2_3)/5 # Residual connection
up_3 = self.up_3(skip_3)
deconv_4 = self.deconv_4(up_3)
skip_4 = (deconv_4 + down_1_0 + down_1_1 + down_1_2 + down_1_3)/5 # Residual connection
up_4 = self.up_4(skip_4)
# Last output
#return F.softmax(self.out(up_4))
return self.out(up_4)