RITnet/densenet.py at master · kbarkevich/RITnet · GitHub

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
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Mon Sep  2 11:20:33 2019

@author: Shusil Dangi

References:
    https://github.com/ShusilDangi/DenseUNet-K
It is a simplied version of DenseNet with U-NET architecture.
2D implementation
"""
import torch
import math
import torch.nn as nn
import torch.nn.functional as F

class DenseNet2D_down_block(nn.Module):
    def __init__(self,input_channels,output_channels,down_size,dropout=False,prob=0,use_bn=True):
        super(DenseNet2D_down_block, self).__init__()
        self.conv1 = nn.Conv2d(input_channels,output_channels,kernel_size=(3,3),padding=(1,1))
        self.conv21 = nn.Conv2d(input_channels+output_channels,output_channels,kernel_size=(1,1),padding=(0,0))
        self.conv22 = nn.Conv2d(output_channels,output_channels,kernel_size=(3,3),padding=(1,1))
        self.conv31 = nn.Conv2d(input_channels+2*output_channels,output_channels,kernel_size=(1,1),padding=(0,0))
        self.conv32 = nn.Conv2d(output_channels,output_channels,kernel_size=(3,3),padding=(1,1))
        self.max_pool = nn.AvgPool2d(kernel_size=down_size)

        self.relu = nn.LeakyReLU()
        self.down_size = down_size
        self.dropout = dropout
        self.dropout1 = nn.Dropout(p=prob)
        self.dropout2 = nn.Dropout(p=prob)
        self.dropout3 = nn.Dropout(p=prob)
        print(down_size)
        if use_bn:
            self.ins = None
            self.bn = torch.nn.BatchNorm2d(num_features=output_channels)  # for best_model.pkl
        else:
            self.bn = None
            self.ins= torch.nn.InstanceNorm2d(num_features=output_channels)  # for other

    def forward(self, x):
        if self.down_size != None:
            x = self.max_pool(x)

        if self.dropout:
            x1 = self.relu(self.dropout1(self.conv1(x)))
            x21 = torch.cat((x,x1),dim=1)
            x22 = self.relu(self.dropout2(self.conv22(self.conv21(x21))))
            x31 = torch.cat((x21,x22),dim=1)
            out = self.relu(self.dropout3(self.conv32(self.conv31(x31))))
        else:
            x1 = self.relu(self.conv1(x))
            x21 = torch.cat((x,x1),dim=1)
            x22 = self.relu(self.conv22(self.conv21(x21)))
            x31 = torch.cat((x21,x22),dim=1)
            out = self.relu(self.conv32(self.conv31(x31)))
        if self.bn is not None:
            return self.bn(out)  # for best_model.pkl
        else:
            return self.ins(out)  # for other


class DenseNet2D_up_block_concat(nn.Module):
    def __init__(self,skip_channels,input_channels,output_channels,up_stride,dropout=False,prob=0):
        super(DenseNet2D_up_block_concat, self).__init__()
        self.conv11 = nn.Conv2d(skip_channels+input_channels,output_channels,kernel_size=(1,1),padding=(0,0))
        self.conv12 = nn.Conv2d(output_channels,output_channels,kernel_size=(3,3),padding=(1,1))
        self.conv21 = nn.Conv2d(skip_channels+input_channels+output_channels,output_channels,
                                kernel_size=(1,1),padding=(0,0))
        self.conv22 = nn.Conv2d(output_channels,output_channels,kernel_size=(3,3),padding=(1,1))
        self.relu = nn.LeakyReLU()
        self.up_stride = up_stride
        self.dropout = dropout
        self.dropout1 = nn.Dropout(p=prob)
        self.dropout2 = nn.Dropout(p=prob)

    def forward(self,prev_feature_map,x):
        x = nn.functional.interpolate(x,scale_factor=self.up_stride,mode='bilinear')
        x = torch.cat((x,prev_feature_map),dim=1)
        if self.dropout:
            x1 = self.relu(self.dropout1(self.conv12(self.conv11(x))))
            x21 = torch.cat((x,x1),dim=1)
            out = self.relu(self.dropout2(self.conv22(self.conv21(x21))))
        else:
            x1 = self.relu(self.conv12(self.conv11(x)))
            x21 = torch.cat((x,x1),dim=1)
            out = self.relu(self.conv22(self.conv21(x21)))
        return out

class DenseNet2D(nn.Module):
    def __init__(self,in_channels=1,out_channels=4,channel_size=32,concat=True,dropout=False,prob=0):
        super(DenseNet2D, self).__init__()

        self.down_block1 = DenseNet2D_down_block(input_channels=in_channels,output_channels=channel_size,
                                                 down_size=None,dropout=dropout,prob=prob,use_bn=out_channels==4)
        self.down_block2 = DenseNet2D_down_block(input_channels=channel_size,output_channels=channel_size,
                                                 down_size=(2,2),dropout=dropout,prob=prob,use_bn=out_channels==4)
        self.down_block3 = DenseNet2D_down_block(input_channels=channel_size,output_channels=channel_size,
                                                 down_size=(2,2),dropout=dropout,prob=prob,use_bn=out_channels==4)
        self.down_block4 = DenseNet2D_down_block(input_channels=channel_size,output_channels=channel_size,
                                                 down_size=(2,2),dropout=dropout,prob=prob,use_bn=out_channels==4)
        self.down_block5 = DenseNet2D_down_block(input_channels=channel_size,output_channels=channel_size,
                                                 down_size=(2,2),dropout=dropout,prob=prob,use_bn=out_channels==4)

        self.up_block1 = DenseNet2D_up_block_concat(skip_channels=channel_size,input_channels=channel_size,
                                                    output_channels=channel_size,up_stride=(2,2),dropout=dropout,prob=prob)
        self.up_block2 = DenseNet2D_up_block_concat(skip_channels=channel_size,input_channels=channel_size,
                                                    output_channels=channel_size,up_stride=(2,2),dropout=dropout,prob=prob)
        self.up_block3 = DenseNet2D_up_block_concat(skip_channels=channel_size,input_channels=channel_size,
                                                    output_channels=channel_size,up_stride=(2,2),dropout=dropout,prob=prob)
        self.up_block4 = DenseNet2D_up_block_concat(skip_channels=channel_size,input_channels=channel_size,
                                                    output_channels=channel_size,up_stride=(2,2),dropout=dropout,prob=prob)

        self.out_conv1 = nn.Conv2d(in_channels=channel_size,out_channels=out_channels,kernel_size=1,padding=0)
        self.concat = concat
        self.dropout = dropout
        self.dropout1 = nn.Dropout(p=prob)

        self._initialize_weights()

    def _initialize_weights(self):
        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))
                if m.bias is not None:
                    m.bias.data.zero_()
            elif isinstance(m, nn.BatchNorm2d):
                m.weight.data.fill_(1)
                m.bias.data.zero_()
            elif isinstance(m, nn.Linear):
                n = m.weight.size(1)
                m.weight.data.normal_(0, 0.01)
                m.bias.data.zero_()

    def forward(self,x):
        self.x1 = self.down_block1(x)
        self.x2 = self.down_block2(self.x1)
        self.x3 = self.down_block3(self.x2)
        self.x4 = self.down_block4(self.x3)
        self.x5 = self.down_block5(self.x4)
        self.x6 = self.up_block1(self.x4,self.x5)
        self.x7 = self.up_block2(self.x3,self.x6)
        self.x8 = self.up_block3(self.x2,self.x7)
        self.x9 = self.up_block4(self.x1,self.x8)
        if self.dropout:
            out = self.out_conv1(self.dropout1(self.x9))
        else:
            out = self.out_conv1(self.x9)

        return out