-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathOhModel.py
266 lines (219 loc) · 11.2 KB
/
OhModel.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
import torch
from torch import Tensor
import torch.nn as nn
import torch.nn.functional as F
import torchvision.transforms as vtransforms
from torch.autograd import Variable
from typing import Type, Any, Callable, Union, List, Optional
import blocks
from torch.nn import init
import functools
#####################
# Custom Generator and Discriminator for Oh et al.'s model
# Oh and Yun. 2018. Oh, D. Y., & Yun, I. D. (2018). Learning Bone Suppression from Dual Energy Chest X-rays using Adversarial Networks. https://arxiv.org/abs/1811.02628
#####################
# Residual Block:
class ConvResBlock(nn.Module):
def __init__(self, in_channels, out_channels, use_bias=True, reluType="normal"):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.use_bias = use_bias
self.reluType = reluType
self.conv1_1 = nn.Conv2d(self.in_channels, self.out_channels,
kernel_size=1, stride=2, padding=0, bias=self.use_bias)
self.conv3_1 = nn.Conv2d(self.in_channels, self.out_channels,
kernel_size=3, stride=2, padding = 1, bias=self.use_bias)
self.BN1 = nn.BatchNorm2d(self.out_channels)
if self.reluType == "normal":
self.relu = nn.ReLU()
if self.reluType == "leaky":
self.relu = nn.LeakyReLU(0.2)
self.conv3_2 = nn.Conv2d(self.out_channels, self.out_channels, kernel_size=3,
stride=1, padding=1, bias=self.use_bias)
self.BN2 = nn.BatchNorm2d(self.out_channels)
def forward(self, x):
out_skip = self.conv1_1(x)
out = self.conv3_1(x)
out = self.BN1(out)
out = self.relu(out)
out = self.conv3_2(out)
out = out + out_skip
out = self.BN2(out)
out = self.relu(out)
return out
class DeconvResBlock(nn.Module):
def __init__(self, in_channels, out_channels, use_bias=True, reluType="normal"):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.use_bias = use_bias
self.reluType = reluType
self.convTranspose2d_padding = 1
#self.conv1_1 = nn.ConvTranspose2d(self.in_channels, self.out_channels,
# kernel_size=1, stride=2, padding=0, output_padding=1, bias=self.use_bias)
self.conv1_1 = blocks.UpsampleConvolution(in_channels=self.in_channels, out_channels=self.out_channels, upsample_scale_factor=2, kernel_size=1, stride=1, padding=0, bias=self.use_bias)
#self.conv3_1 = nn.ConvTranspose2d(self.in_channels, self.out_channels,
# kernel_size=3, stride=2, padding = 1, output_padding=1, bias=self.use_bias)
self.conv3_1 = blocks.UpsampleConvolution(in_channels=self.in_channels, out_channels=self.out_channels, upsample_scale_factor=2, kernel_size=3, stride=1, padding=1, bias=self.use_bias)
self.BN1 = nn.BatchNorm2d(self.out_channels)
if self.reluType == "normal":
self.relu = nn.ReLU()
if self.reluType == "leaky":
self.relu = nn.LeakyReLU(0.2)
self.conv3_2 = nn.Conv2d(self.out_channels, self.out_channels, kernel_size=3,
stride=1, padding=1, bias=self.use_bias)
self.BN2 = nn.BatchNorm2d(self.out_channels)
def forward(self, x):
# Skip is correct
out_skip = self.conv1_1(x)
out = self.conv3_1(x)
out = self.BN1(out)
out = self.relu(out)
out = self.conv3_2(out)
out = out + out_skip
out = self.BN2(out)
out = self.relu(out)
return out
class SqueezeExcitationBlock(nn.Module):
def __init__(self, in_c, out_c, reduction_ratio=16, use_bias=True, reluType="normal"):
super().__init__()
self.in_channels = in_c
self.out_channels = out_c
self.reduction_ratio = reduction_ratio
self.use_bias = use_bias
self.reluType = reluType
# RESIDUAL BLOCK
if self.reluType == "normal":
self.relu = nn.ReLU()
if self.reluType == "leaky":
self.relu = nn.LeakyReLU(0.2)
self.conv_skip = nn.Conv2d(self.in_channels, self.out_channels, 1, 1, 0)
self.conv3_1 = nn.Conv2d(self.in_channels, self.out_channels, 3, 1, 1)
self.norm1 = nn.BatchNorm2d(self.out_channels, self.out_channels)
self.conv3_2 = nn.Conv2d(self.out_channels, self.out_channels, 3, 1, 1)
self.norm2 = nn.BatchNorm2d(self.out_channels, self.out_channels)
self.linear1 = nn.Linear(in_features=self.out_channels, out_features=self.out_channels//self.reduction_ratio)
self.linear2 = nn.Linear(in_features=self.out_channels//self.reduction_ratio, out_features=self.out_channels)
self.sigmoid = nn.Sigmoid()
def forward(self, x):
# Residual Block
out_skip = self.conv_skip(x)
out = self.conv3_1(x)
out = self.norm1(out)
out = self.relu(out)
out = self.conv3_2(out)
out = out + out_skip
out = self.norm2(out)
out_after_residual = self.relu(out)
# SQUEEZE + EXCITATION
out = torch.mean(out_after_residual,[2,3]) # global average pooling
out = self.linear1(out)
out = self.relu(out)
out = self.linear2(out)
out = self.sigmoid(out)
out = out.unsqueeze(-1).unsqueeze(-1)
out = out*out_after_residual
return out
class Generator(nn.Module):
def __init__(self, input_array_shape, reluType="normal", use_bias=True):
super().__init__()
self.input_array_shape = input_array_shape
self.num_ini_filters = 64
self.reluType = reluType
self.use_bias = use_bias
in_channels = input_array_shape[1]
self.encblk1 = ConvResBlock(in_channels, self.num_ini_filters, use_bias=self.use_bias, reluType=self.reluType)
self.encblk2 = ConvResBlock(self.num_ini_filters, self.num_ini_filters, use_bias=self.use_bias, reluType=self.reluType)
self.encblk3 = ConvResBlock(self.num_ini_filters, self.num_ini_filters*2, use_bias=self.use_bias, reluType=self.reluType)
self.encblk4 = ConvResBlock(self.num_ini_filters*2, self.num_ini_filters*2, use_bias=self.use_bias, reluType=self.reluType)
self.encblk5 = ConvResBlock(self.num_ini_filters*2, self.num_ini_filters*4, use_bias=self.use_bias, reluType=self.reluType)
self.encblk6 = ConvResBlock(self.num_ini_filters*4, 320, use_bias=self.use_bias, reluType=self.reluType)
self.SQblk = SqueezeExcitationBlock(320, 320,
reduction_ratio=16, use_bias=self.use_bias, reluType=self.reluType)
self.decblk6 = DeconvResBlock(320, self.num_ini_filters*4, use_bias=self.use_bias, reluType=self.reluType)
self.decblk5 = DeconvResBlock(self.num_ini_filters*4, self.num_ini_filters*2, use_bias=self.use_bias, reluType=self.reluType)
self.decblk4 = DeconvResBlock(self.num_ini_filters*2, self.num_ini_filters*2, use_bias=self.use_bias, reluType=self.reluType)
self.decblk3 = DeconvResBlock(self.num_ini_filters*2, self.num_ini_filters*1, use_bias=self.use_bias, reluType=self.reluType)
self.decblk2 = DeconvResBlock(self.num_ini_filters*1, self.num_ini_filters*1, use_bias=self.use_bias, reluType=self.reluType)
self.decblk1 = DeconvResBlock(self.num_ini_filters*1, in_channels, use_bias=self.use_bias, reluType=self.reluType)
print("Oh Model Generator thought to use summation skip connection.")
def forward(self, x):
out1 = self.encblk1(x)
out2 = self.encblk2(out1)
out3 = self.encblk3(out2)
out4 = self.encblk4(out3)
out5 = self.encblk5(out4)
out6 = self.encblk6(out5)
out = self.SQblk(out6)
#print(out.shape)
out = self.decblk6(out)
# Using Summation Skip Connection
out = out + out5
out = self.decblk5(out)
out = out + out4
out = self.decblk4(out)
out = out + out3
out = self.decblk3(out)
out = out + out2
out = self.decblk2(out)
out = out + out1
out = self.decblk1(out)
return out
class Discriminator_ConvBlock(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size, padding, use_bias=True, reluType="normal"):
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.stride = 2
self.padding = padding
self.use_bias = use_bias
self.reluType = reluType
self.conv = nn.Conv2d(self.in_channels, self.out_channels, self.kernel_size, self.stride, self.padding, self.use_bias)
self.norm = nn.BatchNorm2d(self.out_channels)
if self.reluType == "normal":
self.relu = nn.ReLU()
if self.reluType == "leaky":
self.relu = nn.LeakyReLU(0.2)
def forward(self, x):
out = self.conv(x)
out = self.norm(out)
out = self.relu(out)
return out
class Discriminator(nn.Module):
def __init__(self, input_array_shape, num_kernels, kernel_dims, use_bias = True, reluType="normal"):
super().__init__()
self.input_array_shape = input_array_shape
self.initial_out_channel = 32
self.use_bias = use_bias
self.reluType = reluType
self.num_kernels = num_kernels
self.kernel_dims = kernel_dims
in_c = self.input_array_shape[1]
out_c = self.initial_out_channel
self.encodeBlk1 = Discriminator_ConvBlock(in_c, out_c, 3, 1, self.use_bias, self.reluType)
self.encodeBlk2 = Discriminator_ConvBlock(out_c, out_c, 3, 1, self.use_bias, self.reluType)
self.encodeBlk3 = Discriminator_ConvBlock(out_c, out_c*2, 3, 1, self.use_bias, self.reluType)
self.encodeBlk4 = Discriminator_ConvBlock(out_c*2, out_c*2, 3, 1, self.use_bias, self.reluType)
self.encodeBlk5 = Discriminator_ConvBlock(out_c*2, out_c*4, 3, 1, self.use_bias, self.reluType)
self.encodeBlk6 = Discriminator_ConvBlock(out_c*4, out_c*4, 3, 1, self.use_bias, self.reluType)
self.encodeBlk7 = Discriminator_ConvBlock(out_c*4, out_c*8, 3, 1, self.use_bias, self.reluType)
# Flatten array
array_size = [self.input_array_shape[0], out_c*8, input_array_shape[2]//128, input_array_shape[3]//128]
self.flatten = nn.Flatten()
feature_num = array_size[1]*array_size[2]*array_size[3]
self.miniBatchDisc = blocks.MiniBatchDiscrimination(feature_num, self.num_kernels, self.kernel_dims, mean=False)
self.fc = nn.Linear(feature_num + self.num_kernels, 1)
def forward(self, x):
out = self.encodeBlk1(x)
out = self.encodeBlk2(out)
out = self.encodeBlk3(out)
out = self.encodeBlk4(out)
out = self.encodeBlk5(out)
out = self.encodeBlk6(out)
out = self.encodeBlk7(out)
out = self.flatten(out)
out = self.miniBatchDisc(out)
out = self.fc(out)
return out