-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathunet_network.py
81 lines (62 loc) · 3.41 KB
/
unet_network.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
import torch
import numpy as np
import os
import torch.nn as nn
### 3D U-net implementation for segmenting organs at risk
class UNet_3D(nn.Module):
def __init__(self, input_channel = 1, output_channel = 1, num_features = 32, num_layers = 4):
super(UNet_3D, self).__init__()
self.num_features = num_features
# Identify each layers in the UNet
self.encoder_1 = UNet_3D._build_conv_block(input_channel, num_features)
self.maxpool_1 = nn.MaxPool3d(kernel_size = 2, stride = 2)
self.encoder_2 = UNet_3D._build_conv_block(num_features, num_features*2)
self.maxpool_2 = nn.MaxPool3d(kernel_size = 2, stride = 2)
self.encoder_3 = UNet_3D._build_conv_block(num_features*2, num_features*4)
self.maxpool_3 = nn.MaxPool3d(kernel_size = 2, stride = 2)
self.encoder_4 = UNet_3D._build_conv_block(num_features*4, num_features*8)
self.maxpool_4 = nn.MaxPool3d(kernel_size = 2, stride = 2)
self.bottle_neck = UNet_3D._build_conv_block(num_features *8, num_features * 16)
self.upconv_4 = nn.ConvTranspose3d(num_features * 16, num_features * 8, kernel_size = 2, stride = 2)
self.decoder_4 = UNet_3D._build_conv_block((num_features*8)*2, num_features *8)
self.upconv_3 = nn.ConvTranspose3d(num_features * 8, num_features * 4, kernel_size = 2, stride = 2)
self.decoder_3 = UNet_3D._build_conv_block((num_features*4)*2, num_features *4)
self.upconv_2 = nn.ConvTranspose3d(num_features * 4, num_features * 2, kernel_size = 2, stride = 2)
self.decoder_2 = UNet_3D._build_conv_block((num_features*2)*2, num_features *2)
self.upconv_1 = nn.ConvTranspose3d(num_features*2 , num_features, kernel_size = 2, stride = 2)
self.decoder_1 = UNet_3D._build_conv_block(num_features*2, num_features)
self.final_conv = torch.nn.Conv3d(num_features, output_channel, kernel_size = 1)
def forward(self, x):
enc1 = self.encoder_1(x)
enc2 = self.encoder_2(self.maxpool_1(enc1))
enc3 = self.encoder_3(self.maxpool_2(enc2))
enc4 = self.encoder_4(self.maxpool_3(enc3))
bottleneck = self.bottle_neck(self.maxpool_4(enc4))
dec4 = self.upconv_4(bottleneck)
dec4 = torch.cat((dec4, enc4), dim=1)
dec4 = self.decoder_4(dec4)
dec3 = self.upconv_3(dec4)
dec3 = torch.cat((dec3, enc3), dim=1)
dec3 = self.decoder_3(dec3)
dec2 = self.upconv_2(dec3)
dec2 = torch.cat((dec2, enc2), dim=1)
dec2 = self.decoder_2(dec2)
dec1 = self.upconv_1(dec2)
dec1 = torch.cat((dec1, enc1), dim=1)
dec1 = self.decoder_1(dec1)
return torch.sigmoid(self.final_conv(dec1))
@staticmethod
def _build_conv_block(input_channel, num_features):
conv_block = nn.Sequential(
nn.Conv3d(in_channels = input_channel, out_channels = num_features, kernel_size = 3, padding = 1, bias = False),
nn.BatchNorm3d(num_features = num_features),
nn.ReLU(inplace= True),
nn.Conv3d(in_channels = num_features, out_channels = num_features, kernel_size = 3, padding = 1, bias = False),
nn.BatchNorm3d(num_features = num_features),
nn.ReLU(inplace=True))
return conv_block
def _centre_crop(encoder_layer, decoder_layer):
"""
A function that centre crops encoder layer to match that of the encoder layer
"""
pass