TOPAL/model.py at master · Jzy2017/TOPAL · 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
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
import torch
import torch.nn as nn
from blocks import *
from functools import reduce
import numpy as np
import cv2
import time
import numpy as np
import os

class RDB(nn.Module):
    def __init__(self, inChannels, growRate, kSize=3):
        super(RDB, self).__init__()
        Cin = inChannels
        G = growRate

        self.conv1 = nn.Conv2d(Cin, G, kSize, padding=(kSize -1 )//2, stride=1)
        self.conv2 = nn.Conv2d(Cin + G, G, kSize, padding=(kSize -1 )//2, stride=1)
        self.conv3 = nn.Conv2d(Cin + 2 * G, G, kSize, padding=(kSize -1 )//2, stride=1)
        self.conv4 = nn.Conv2d(Cin + 3 * G, G, kSize, padding=(kSize - 1) // 2, stride=1)
        self.conv5 = nn.Conv2d(Cin + 4 * G, G, kSize, padding=(kSize - 1) // 2, stride=1)
        self.conv6 = nn.Conv2d(Cin + 5 * G, Cin, kSize, padding=(kSize - 1) // 2, stride=1)
        self.act = nn.LeakyReLU(negative_slope=0.2, inplace=True)

    def forward(self, x):
        x1 = self.act(self.conv1(x))
        x2 = self.act(self.conv2(torch.cat((x, x1), 1)))
        x3 = self.act(self.conv3(torch.cat((x, x1, x2), 1)))
        x4 = self.act(self.conv4(torch.cat((x, x1, x2, x3), 1)))
        x5 = self.act(self.conv5(torch.cat((x, x1, x2, x3, x4), 1)))
        x6 = self.conv6(torch.cat((x, x1, x2, x3, x4, x5), 1))

        return 0.1 * x6 + x

class WRDB(nn.Module):
    def __init__(self, num_features, growRate, kSize=3):
        super(WRDB, self).__init__()
        self.RDB1 = RDB(num_features, growRate, kSize)
        self.RDB2 = RDB(num_features, growRate, kSize)
        self.RDB3 = RDB(num_features, growRate, kSize)

        self.conv0= nn.Conv2d(num_features, num_features, kSize, padding=(kSize -1 )//2, stride=1)
        self.conv1_1 = nn.Conv2d(in_channels=num_features, out_channels=num_features,
                               kernel_size=1, padding=0, stride=1, groups=num_features)
        self.conv1_2 = nn.Conv2d(in_channels=num_features, out_channels=num_features,
                                 kernel_size=1, padding=0, stride=1, groups=num_features)
        self.conv1_3 = nn.Conv2d(in_channels=num_features, out_channels=num_features,
                                 kernel_size=1, padding=0, stride=1, groups=num_features)

        self.conv2_1 = nn.Conv2d(in_channels=num_features, out_channels=num_features,
                               kernel_size=1, padding=0, stride=1, groups=num_features)
        self.conv2_2 = nn.Conv2d(in_channels=num_features, out_channels=num_features,
                                 kernel_size=1, padding=0, stride=1, groups=num_features)

        self.conv3_1 = nn.Conv2d(in_channels=num_features, out_channels=num_features,
                               kernel_size=1, padding=0, stride=1, groups=num_features)

        self.act = nn.LeakyReLU(negative_slope=0.2, inplace=True)

    def forward(self, x):
        x1_0 = self.act(self.conv0(x))
        x1_1 = self.conv1_1(x1_0)
        x2_0 = self.RDB1(x1_0)
        x2 = x2_0 + x1_1
        x2_1 = self.conv2_1(x2)
        x1_2 = self.conv1_2(x1_0)
        x3_0 = self.RDB2(x2)
        x3 = x2_1 + x1_2 + x3_0
        x1_3 = self.conv1_3(x1_0)
        x2_2 = self.conv2_2(x2)
        x3_1 = self.conv3_1(x3)
        x4 = self.RDB3(x3)
        out = x1_3 + x2_2 + x3_1 + x4
        out = 0.1 * out + x
        return out

class Upsampler(nn.Module):
    def __init__(self, num_features, out_channels, kSize, scale):
        super(Upsampler, self).__init__()
        # Up-sampling net
        if scale == 2 or scale == 3:
            self.UPNet = nn.Sequential(*[
                nn.Conv2d(num_features, num_features * scale * scale, kSize, padding=(kSize - 1) // 2, stride=1),
                nn.PixelShuffle(scale),
                nn.Conv2d(num_features, out_channels, kSize, padding=(kSize - 1) // 2, stride=1)
            ])
        elif scale == 4:
            self.UPNet = nn.Sequential(*[
                nn.Conv2d(num_features, num_features * 4, kSize, padding=(kSize - 1) // 2, stride=1),
                nn.PixelShuffle(2),
                nn.Conv2d(num_features, num_features * 4, kSize, padding=(kSize - 1) // 2, stride=1),
                nn.PixelShuffle(2),
                nn.Conv2d(num_features, out_channels, kSize, padding=(kSize - 1) // 2, stride=1)
            ])
        elif scale == 8:
            self.UPNet = nn.Sequential(*[
            nn.Conv2d(num_features, num_features * 4, kSize, padding=(kSize - 1) // 2, stride=1),
            nn.PixelShuffle(2),
            nn.Conv2d(num_features, num_features * 4, kSize, padding=(kSize - 1) // 2, stride=1),
            nn.PixelShuffle(2),
            nn.Conv2d(num_features, num_features * 4, kSize, padding=(kSize - 1) // 2, stride=1),
            nn.PixelShuffle(2),
            nn.Conv2d(num_features, out_channels, kSize, padding=(kSize - 1) // 2, stride=1)
        ])
        else:
            raise ValueError("scale must be 2 or 3 or 4.")

    def forward(self, x):
        out = self.UPNet(x)

        return out

class AsyCA(nn.Module):
    def __init__(self, num_features, ratio):
        super(AsyCA, self).__init__()
        self.out_channels = num_features
        self.conv_init = nn.Conv2d(num_features * 2, num_features, kernel_size=1, padding=0, stride=1)
        self.conv_dc = nn.Conv2d(num_features, num_features // ratio, kernel_size=1, padding=0, stride=1)
        self.conv_ic = nn.Conv2d(num_features // ratio, num_features * 2, kernel_size=1, padding=0, stride=1)
        self.act = nn.ReLU(inplace=True)
        self.avg_pool = nn.AdaptiveAvgPool2d(1)
        self.softmax = nn.Softmax(dim=1)

    def forward(self, x1, x2):
        batch_size = x1.size(0)

        feat_init = torch.cat((x1, x2), 1)
        feat_init = self.conv_init(feat_init)
        fea_avg = self.avg_pool(feat_init)
        feat_ca = self.conv_dc(fea_avg)
        feat_ca = self.conv_ic(self.act(feat_ca))

        a_b = feat_ca.reshape(batch_size, 2, self.out_channels, -1)

        a_b = self.softmax(a_b)
        # print(a_b[0,0,0,0],)
        a_b = list(a_b.chunk(2, dim=1))  # split to a and b
        a_b = list(map(lambda x1: x1.reshape(batch_size, self.out_channels, 1, 1), a_b))
        self.V1 =V1= a_b[0] * x1
        self.V2 =V2= a_b[1] * x2
        V = V1 + V2
        return V

class WaterFusion(nn.Module):
    def __init__(self, in_channels, out_channels, num_features, growthrate):
        super(WaterFusion, self).__init__()
        kSize = 3
        ratio = 4

        self.feat_conv1 = nn.Conv2d(in_channels, num_features, kSize, padding=(kSize - 1) // 2, stride=1)
        self.feat_conv2 = nn.Conv2d(in_channels, num_features, kSize, padding=(kSize - 1) // 2, stride=1)
        self.RDB0 = RDB(num_features, growthrate, kSize)
        self.RDB1 = RDB(num_features, growthrate, kSize)
        self.RDB2 = RDB(num_features, growthrate, kSize)
        # self.RDB3 = RDB(num_features, growthrate, kSize)

        self.AsyCA1 = AsyCA(num_features, ratio)
        # self.AsyCA2 = AsyCA(num_features, ratio)
        # self.AsyCA3 = AsyCA(num_features, ratio)

        self.out_conv = nn.Conv2d(num_features, out_channels, 1, padding=0, stride=1)

    	# self.tanh = nn.Tanh()

    def forward(self, pre1, pre2, w=None):
        x1 = self.feat_conv1(pre1)
        self.x1 = x1 = self.RDB0(x1)

        x2 = self.feat_conv2(pre2)
        self.x2 = x2 = self.RDB1(x2)

        if w == None:
            self.fused = x = self.AsyCA1(x1, x2)
        else:
            self.fused = x = w*x1+(1-w)*x2
        #


        x = self.RDB2(x)
        x = self.out_conv(x)
        # x = self.tanh(x)
        return x