-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathjonas_net.py
211 lines (165 loc) · 7.77 KB
/
jonas_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
from torch import nn
import torch
import torchvision.models as models
def conv2d_to_conv3d(layer):
in_channels = layer.in_channels
out_channels = layer.out_channels
new_weight = layer.weight.unsqueeze(2)
kernel_size = tuple([1] + list(layer.kernel_size))
stride = tuple([1] + list(layer.stride))
padding = tuple([0] + list(layer.padding))
new_layer = nn.Conv3d(in_channels, out_channels, kernel_size, stride, padding)
new_weight = layer.weight.unsqueeze(2)
new_layer.weight.data = new_weight.data
if layer.bias is not None:
new_layer.bias.data = layer.bias.data
return new_layer
def batch2d_to_batch3d(layer):
num_features = layer.num_features
eps = layer.eps
momentum = layer.momentum
affine = layer.affine
new_bn = nn.BatchNorm3d(num_features, eps, momentum, affine)
new_bn.load_state_dict(layer.state_dict())
return new_bn
def pool2d_to_pool3d(layer):
kernel_size = tuple([1, layer.kernel_size, layer.kernel_size])
stride = tuple([1, layer.stride, layer.stride])
padding = layer.padding
dilation = layer.dilation
return_indices = layer.return_indices
ceil_mode = layer.ceil_mode
return nn.MaxPool3d(kernel_size, stride, padding, dilation, return_indices, ceil_mode)
def transform_module_list(module_list):
transformed_list = []
for item in module_list:
if isinstance(item, nn.Conv2d):
transformed_list.append(conv2d_to_conv3d(item))
elif isinstance(item, nn.BatchNorm2d):
transformed_list.append(batch2d_to_batch3d(item))
elif isinstance(item, nn.MaxPool2d):
transformed_list.append(pool2d_to_pool3d(item))
elif isinstance(item, nn.ReLU):
transformed_list.append(item)
elif isinstance(item, nn.Sequential):
transformed_list += transform_module_list(item.children())
else:
transformed_list.append(item)
return transformed_list
def translate_block(block, downsample=False):
block.conv1 = conv2d_to_conv3d(block.conv1)
block.bn1 = batch2d_to_batch3d(block.bn1)
block.conv2 = conv2d_to_conv3d(block.conv2)
block.bn2 = batch2d_to_batch3d(block.bn2)
if block.downsample is not None:
conv = conv2d_to_conv3d(block.downsample[0])
bn = batch2d_to_batch3d(block.downsample[1])
block.downsample = nn.Sequential(*[conv, bn])
return block
def conv1x3x3(in_, out):
return nn.Conv3d(in_, out, kernel_size=(1,3,3), padding=(0,1,1))
def conv3x1x1(in_, out):
# padding=(0,0,0) when depth reduction, otherwise (1,0,0)
# return nn.Conv3d(in_, out, kernel_size=(3,1,1), padding=(0,0,0))
return nn.Conv3d(in_, out, kernel_size=(3,1,1), padding=(1,0,0))
class ConvRelu(nn.Module):
def __init__(self, in_, out, depth=False):
super().__init__()
if depth:
self.conv = conv3x1x1(in_, out)
else:
self.conv = conv1x3x3(in_, out)
self.bn = nn.BatchNorm3d(out)
self.activation = nn.ReLU(inplace=True)
def forward(self, x):
x = self.conv(x)
x = self.bn(x)
x = self.activation(x)
return x
class ConvReluFromEnc(nn.Module):
def __init__(self, conv, bn, relu):
super().__init__()
self.conv = conv
self.bn = bn
self.relu = relu
def forward(self, x):
x = self.conv(x)
x = self.bn(x)
x = self.relu(x)
return x
class Upsample2D(nn.Module):
def __init__(self, in_channels, middle_channels, out_channels):
super().__init__()
self.block = nn.Sequential(
ConvRelu(in_channels, middle_channels),
nn.ConvTranspose3d(middle_channels, out_channels, kernel_size=(1,4,4), stride=(1,2,2), padding=(0,1,1)),
nn.ReLU(inplace=True)
)
def forward(self, x):
return self.block(x)
class AlbuNet3D34(nn.Module):
def __init__(self, num_classes=1, num_filters=16, pretrained=False, is_deconv=False):
super(AlbuNet3D34, self).__init__()
# resnet34 has a lot of blocks
self.num_classes = num_classes
self.encoder = models.resnet34(pretrained=pretrained)
self.pool = nn.MaxPool3d(kernel_size=(1,2,2), stride=(1,2,2))
conv0_modules = [self.encoder.conv1, self.encoder.bn1, self.encoder.relu, self.pool]
self.conv0 = nn.Sequential(*transform_module_list(conv0_modules))
self.conv1_0 = translate_block(self.encoder.layer1[0])
self.conv1_1 = translate_block(self.encoder.layer1[1])
self.conv1_2 = translate_block(self.encoder.layer1[2])
# number output filters: 64
self.conv2_0 = translate_block(self.encoder.layer2[0])
self.conv2_1 = translate_block(self.encoder.layer2[1])
self.conv2_2 = translate_block(self.encoder.layer2[2])
self.conv2_3 = translate_block(self.encoder.layer2[3])
# number output filters: 128
self.conv3_0 = translate_block(self.encoder.layer3[0])
self.conv3_1 = translate_block(self.encoder.layer3[1])
self.conv3_2 = translate_block(self.encoder.layer3[2])
self.conv3_3 = translate_block(self.encoder.layer3[3])
self.conv3_4 = translate_block(self.encoder.layer3[4])
self.conv3_5 = translate_block(self.encoder.layer3[5])
# number output filters: 256
self.conv4_0 = translate_block(self.encoder.layer4[0])
self.conv4_1 = translate_block(self.encoder.layer4[1])
self.conv4_2 = translate_block(self.encoder.layer4[2])
# number output filters: 512
# self.center = Upsample2D(512, num_filters * 8 * 2, num_filters * 8)
self.center = ConvRelu(512, num_filters * 8) # 128
self.dec5 = Upsample2D(512 + num_filters * 8, num_filters * 8 * 2, num_filters * 8)
# self.dec5 = Upsample2D(512, num_filters * 8 * 2, num_filters * 8)
self.dec4 = Upsample2D(256 + num_filters * 8, num_filters * 8 * 2, num_filters * 8)
self.dec3 = Upsample2D(128 + num_filters * 8, num_filters * 4 * 2, num_filters * 2)
self.dec2 = Upsample2D(64 + num_filters * 2, num_filters * 2 * 2, num_filters * 2 * 2)
self.dec1 = Upsample2D(num_filters * 2 * 2, num_filters * 2 * 2, num_filters)
self.dec0 = ConvRelu(num_filters, num_filters)
self.final = nn.Conv3d(num_filters, num_classes, kernel_size=1)
self.depth1 = ConvRelu(512, 512, depth=True)
self.depth2 = ConvRelu(num_filters * 8, num_filters * 8, depth=True)
self.depth3 = ConvRelu(num_filters * 2 * 2, num_filters * 2 * 2, depth=True)
self.depth4 = ConvRelu(num_filters, num_filters, depth=True)
def forward(self, x):
conv0 = self.conv0(x)
conv1 = self.conv1_2(self.conv1_1(self.conv1_0(conv0)))
conv2 = self.conv2_3(self.conv2_2(self.conv2_1(self.conv2_0(conv1))))
conv3 = self.conv3_5(self.conv3_4(self.conv3_3(self.conv3_2(self.conv3_1(self.conv3_0(conv2))))))
conv4 = self.conv4_2(self.conv4_1(self.conv4_0(conv3)))
conv4 = self.depth1(conv4)
center = self.center(conv4)
dec5 = self.dec5(torch.cat([center, conv4], 1))
dec4 = self.dec4(torch.cat([dec5, conv3], 1))
dec4 = self.depth2(dec4)
dec3 = self.dec3(torch.cat([dec4, conv2], 1))
dec2 = self.dec2(torch.cat([dec3, conv1], 1))
dec2 = self.depth3(dec2)
dec1 = self.dec1(dec2)
dec0 = self.dec0(dec1)
dec0 = self.depth4(dec0)
if self.num_classes > 1:
# x_out = F.log_softmax(self.final(dec0), dim=1)
x_out = self.final(dec0) # softmax is moved to loss metric
else:
x_out = self.final(dec0)
return x_out