-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathReconNet_model_max.py
186 lines (158 loc) · 11.7 KB
/
ReconNet_model_max.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
# This implementation is based on the DenseNet implementation by Geoff Pleiss
# https://github.com/gpleiss/efficient_densenet_pytorch/blob/master/models/densenet.py
import torch.nn as nn
import torch
def _bn_function_factory(norm, relu, conv):
def bn_function(*inputs):
concated_features = torch.cat(inputs, 1)
bottleneck_output = conv(relu(norm(concated_features)))
return bottleneck_output
return bn_function
class _DenseLayer(nn.Module):
def __init__(self, num_input_features, growth_rate, bn_size):
super(_DenseLayer, self).__init__()
self.add_module('norm1', nn.BatchNorm2d(num_input_features)),
self.add_module('relu1', nn.PReLU()),
self.add_module('conv1', nn.Conv2d(num_input_features, bn_size * growth_rate,
kernel_size=1, stride=1, bias=False)),
self.add_module('norm2', nn.BatchNorm2d(bn_size * growth_rate)),
self.add_module('relu2', nn.PReLU()),
self.add_module('conv2', nn.Conv2d(bn_size * growth_rate, growth_rate,
kernel_size=3, stride=1, padding=1, bias=False)),
def forward(self, *prev_features):
bn_function = _bn_function_factory(self.norm1, self.relu1, self.conv1)
bottleneck_output = bn_function(*prev_features)
new_features = self.conv2(self.relu2(self.norm2(bottleneck_output)))
return new_features
class _DenseBlock(nn.Module):
def __init__(self, num_layers, num_input_features, bn_size, growth_rate):
super(_DenseBlock, self).__init__()
for i in range(num_layers):
layer = _DenseLayer(
num_input_features + i * growth_rate,
growth_rate=growth_rate,
bn_size=bn_size,
)
self.add_module('denselayer%d' % (i + 1), layer)
def forward(self, init_features):
features = [init_features]
for name, layer in self.named_children():
new_features = layer(*features)
features.append(new_features)
return torch.cat(features, 1)
class _Transition(nn.Sequential):
def __init__(self, num_input_features, num_output_features):
super(_Transition, self).__init__()
self.add_module('norm1', nn.BatchNorm2d(num_input_features))
self.add_module('relu1', nn.PReLU())
self.add_module('conv1', nn.Conv2d(num_input_features, num_output_features,
kernel_size=1, stride=1, bias=False))
self.add_module('pool', nn.MaxPool2d(kernel_size=3, stride=2,padding=1))
# self.add_module('norm2', nn.BatchNorm2d(num_output_features))
# self.add_module('relu2', nn.LeakyReLU(0.1))
# self.add_module('conv2', nn.Conv2d(num_output_features, num_output_features,
# kernel_size=3, stride=2, padding=1,bias=False))
class _Transition_up(nn.Sequential):
def __init__(self, num_input_features, num_output_features,out_padding):
super(_Transition_up, self).__init__()
self.add_module('norm1', nn.BatchNorm2d(num_input_features))
self.add_module('relu1', nn.PReLU())
self.add_module('conv1', nn.Conv2d(num_input_features, num_output_features,
kernel_size=1, stride=1, bias=False))
self.add_module('norm2', nn.BatchNorm2d(num_output_features))
self.add_module('relu2', nn.PReLU())
self.add_module('conv2', nn.ConvTranspose2d(num_output_features, num_output_features,
kernel_size=3,stride=2,padding=1,output_padding=out_padding,bias=False))
class ReconNet(nn.Module):
def __init__(self, num_input_feature = 16, num_init_features=64,growth_rate=24,
bn_size=4,depth_of_block=5, compression_downsample=0.8,compression_upsample=0.2):
super(ReconNet, self).__init__()
assert 0 < compression_downsample <= 1, 'compression of densenet should be between 0 and 1'
assert 0 < compression_upsample <= 1, 'compression of densenet should be between 0 and 1'
self.conv0 = nn.Conv2d(num_input_feature,num_init_features, kernel_size=3, stride=1, padding=1, bias=False)
num_channel1 = num_init_features
self.dense_down1 = _DenseBlock(depth_of_block,num_channel1,bn_size=bn_size,growth_rate=growth_rate)
self.downsample1 = _Transition(num_channel1 + depth_of_block * growth_rate,
int((num_channel1 + depth_of_block * growth_rate) * compression_downsample))
num_channel2 = int((num_channel1 + depth_of_block * growth_rate) * compression_downsample)
self.dense_down2 = _DenseBlock(depth_of_block,num_channel2,bn_size=bn_size,growth_rate=growth_rate)
self.downsample2 = _Transition(num_channel2 + depth_of_block * growth_rate,
int((num_channel2 + depth_of_block * growth_rate) * compression_downsample))
num_channel3 = int((num_channel2 + depth_of_block * growth_rate) * compression_downsample)
self.dense_down3 = _DenseBlock(depth_of_block,num_channel3,bn_size=bn_size,growth_rate=growth_rate)
self.downsample3 = _Transition(num_channel3 + depth_of_block * growth_rate,
int((num_channel3 + depth_of_block * growth_rate) * compression_downsample))
num_channel4 = int((num_channel3 + depth_of_block * growth_rate) * compression_downsample)
self.dense_down4 = _DenseBlock(depth_of_block,num_channel4,bn_size=bn_size,growth_rate=growth_rate)
self.downsample4 = _Transition(num_channel4 + depth_of_block * growth_rate,
int((num_channel4 + depth_of_block * growth_rate) * compression_downsample))
num_channel5 = int((num_channel4 + depth_of_block * growth_rate) * compression_downsample)
self.dense_down5 = _DenseBlock(depth_of_block,num_channel5,bn_size=bn_size,growth_rate=growth_rate)
self.downsample5 = _Transition(num_channel5 + depth_of_block * growth_rate,
int((num_channel5 + depth_of_block * growth_rate) * compression_downsample))
num_channel6 = int((num_channel5 + depth_of_block * growth_rate) * compression_downsample)
self.dense_down6 = _DenseBlock(depth_of_block,num_channel6,bn_size=bn_size,growth_rate=growth_rate)
self.downsample6 = _Transition(num_channel6 + depth_of_block * growth_rate,
int((num_channel6 + depth_of_block * growth_rate) * compression_downsample))
num_channel7 = int((num_channel6 + depth_of_block * growth_rate) * compression_downsample)
self.dense_down7 = _DenseBlock(depth_of_block,num_channel7,bn_size=bn_size,growth_rate=growth_rate)
self.downsample7 = _Transition(num_channel7 + depth_of_block * growth_rate,
int((num_channel7 + depth_of_block * growth_rate) * compression_downsample))
num_channel8 = int((num_channel7 + depth_of_block * growth_rate) * compression_downsample)
self.dense_up1 = _DenseBlock(depth_of_block,num_channel8,bn_size=bn_size,growth_rate=growth_rate)
self.upsample1 = _Transition_up(num_channel8 + depth_of_block * growth_rate,
int((num_channel8 + depth_of_block * growth_rate) * compression_upsample),out_padding=(0,1))
num_channel9 = int((num_channel8 + depth_of_block * growth_rate) * compression_upsample) + num_channel7 + depth_of_block * growth_rate
self.dense_up2 = _DenseBlock(depth_of_block,num_channel9,bn_size=bn_size,growth_rate=growth_rate)
self.upsample2 = _Transition_up(num_channel9 + depth_of_block * growth_rate,
int((num_channel9 + depth_of_block * growth_rate) * compression_upsample),out_padding=1)
num_channel10 = int((num_channel9 + depth_of_block * growth_rate) * compression_upsample) + num_channel6 + depth_of_block * growth_rate
self.dense_up3 = _DenseBlock(depth_of_block,num_channel10,bn_size=bn_size,growth_rate=growth_rate)
self.upsample3 = _Transition_up(num_channel10 + depth_of_block * growth_rate,
int((num_channel10 + depth_of_block * growth_rate) * compression_upsample),out_padding=1)
num_channel11 = int((num_channel10 + depth_of_block * growth_rate) * compression_upsample) + num_channel5 + depth_of_block * growth_rate
self.dense_up4 = _DenseBlock(depth_of_block,num_channel11,bn_size=bn_size,growth_rate=growth_rate)
self.upsample4 = _Transition_up(num_channel11 + depth_of_block * growth_rate,
int((num_channel11 + depth_of_block * growth_rate) * compression_upsample),out_padding=1)
num_channel12 = int((num_channel11 + depth_of_block * growth_rate) * compression_upsample) + num_channel4 + depth_of_block * growth_rate
self.dense_up5 = _DenseBlock(depth_of_block,num_channel12,bn_size=bn_size,growth_rate=growth_rate)
self.upsample5 = _Transition_up(num_channel12 + depth_of_block * growth_rate,
int((num_channel12 + depth_of_block * growth_rate) * compression_upsample),out_padding=1)
num_channel13 = int((num_channel12 + depth_of_block * growth_rate) * compression_upsample) + num_channel3 + depth_of_block * growth_rate
self.dense_up6 = _DenseBlock(depth_of_block,num_channel13,bn_size=bn_size,growth_rate=growth_rate)
self.upsample6 = _Transition_up(num_channel13 + depth_of_block * growth_rate,
int((num_channel13 + depth_of_block * growth_rate) * compression_upsample),out_padding=1)
num_channel14 = int((num_channel13 + depth_of_block * growth_rate) * compression_upsample) + num_channel2 + depth_of_block * growth_rate
self.dense_up7 = _DenseBlock(depth_of_block,num_channel14,bn_size=bn_size,growth_rate=growth_rate)
self.upsample7 = _Transition_up(num_channel14 + depth_of_block * growth_rate,
int((num_channel14 + depth_of_block * growth_rate) * compression_upsample),out_padding=1)
num_channel15 = int((num_channel14 + depth_of_block * growth_rate) * compression_upsample) + num_channel1 + depth_of_block * growth_rate
self.final = nn.Sequential(_DenseBlock(depth_of_block,num_channel15,bn_size=bn_size,growth_rate=growth_rate),
nn.BatchNorm2d(num_channel15 + depth_of_block * growth_rate),
nn.LeakyReLU(0.1),
nn.Conv2d(num_channel15 + depth_of_block * growth_rate,1,kernel_size=1, stride=1),
nn.Sigmoid())
def forward(self, x):
x1 = self.dense_down1(self.conv0(x))
x2 = self.dense_down2(self.downsample1(x1))
x3 = self.dense_down3(self.downsample2(x2))
x4 = self.dense_down4(self.downsample3(x3))
x5 = self.dense_down5(self.downsample4(x4))
x6 = self.dense_down6(self.downsample5(x5))
x7 = self.dense_down7(self.downsample6(x6))
x8 = self.dense_up1(self.downsample7(x7))
x9 = torch.cat([x7,self.upsample1(x8)],1)
x9 = self.dense_up2(x9)
x10 = torch.cat([x6,self.upsample2(x9)],1)
x10 = self.dense_up3(x10)
x11 = torch.cat([x5,self.upsample3(x10)],1)
x11 = self.dense_up4(x11)
x12 = torch.cat([x4,self.upsample4(x11)],1)
x12 = self.dense_up5(x12)
x13 = torch.cat([x3,self.upsample5(x12)],1)
x13 = self.dense_up6(x13)
x14 = torch.cat([x2,self.upsample6(x13)],1)
x14 = self.dense_up7(x14)
x15 = torch.cat([x1,self.upsample7(x14)],1)
x = self.final(x15)
return x