File size: 7,816 Bytes
40ed350
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
212
import math

import torch
import torch.nn as nn
import torch.nn.functional as F


class SqeezeExcite(nn.Module):
    def __init__(self, channel, reduction_ratio = 16):
        super(SqeezeExcite,self).__init__()
        self.GAP = nn.AdaptiveAvgPool2d(1)

        self.mlp = nn.Sequential(
            nn.Linear(channel, channel//reduction_ratio, bias = False),
            nn.ReLU(inplace=True),
            nn.Linear(channel//reduction_ratio,channel,bias = False),
            nn.Sigmoid()
        )

    def forward(self,x):
        b,c,_,_ = x.size()
        out = self.GAP(x).view(b,c)
        out = self.mlp(out).view(b,c,1,1)
        return x * out.expand_as(x)


class ECA(nn.Module):
    # https://wandb.ai/diganta/ECANet-sweep/reports/Efficient-Channel-Attention--VmlldzozNzgwOTE
    def __init__(self,channels, b = 1, gamma = 2):
        super(ECA, self).__init__()
        self.GAP = nn.AdaptiveAvgPool2d(1)
        self.channels = channels
        self.b = b
        self.gamma = gamma
        self.conv = nn.Conv1d(1, 1, kernel_size=self.adaptive_kernel(),padding = (self.adaptive_kernel()-1)//2, bias = False)
        self.sigmoid = nn.Sigmoid()

    def forward(self,x):
        attn = self.GAP(x)
        attn = self.conv(attn.squeeze(-1).transpose(-1,-2)).transpose(-1,-2).unsqueeze(-1)
        attn = self.sigmoid(attn)
        return x * attn.expand_as(x)


    def adaptive_kernel(self):
        k = int(abs(math.log2(self.channels)/self.gamma) + self.b)
        ksize = k if k%2 else k+1
        return ksize



class UNetConvBlock(nn.Module):
    def __init__(self, in_channel, out_channel, ca_layer):
        super(UNetConvBlock, self).__init__()
        block = []

        block.append(nn.Conv2d(in_channel, out_channel, kernel_size=3, stride=1, padding=1))
        block.append(nn.PReLU())

        block.append(nn.Conv2d(out_channel, out_channel, kernel_size=3, padding=1, stride=1))
        block.append(nn.PReLU())

        if ca_layer:
            block.append(ECA(out_channel))
        self.block = nn.Sequential(*block)

    def forward(self, x):
        out = self.block(x)
        return out

class AttentionGate(nn.Module):
    def __init__(self, F_g, F_l, dimensions):
        super(AttentionGate, self).__init__()
        self.W_gate = nn.Sequential(
            nn.Conv2d(F_g, dimensions, kernel_size=1, stride=1, padding=0, bias=True),
            nn.BatchNorm2d(dimensions)
        )

        self.W_x = nn.Sequential(
            nn.Conv2d(F_l, dimensions, kernel_size=1, stride=1, padding=0, bias=True),
            nn.BatchNorm2d(dimensions)
        )

        self.psi = nn.Sequential(
            nn.Conv2d(dimensions, 1, kernel_size=1, stride=1, padding=0, bias=True),
            nn.BatchNorm2d(1),
            nn.Sigmoid()
        )

        self.relu = nn.PReLU()

    def forward(self, g, x):
        g1 = self.W_gate(g)
        x1 = self.W_x(x)
        psi = self.relu(g1 + x1)
        psi = self.psi(psi)
        out = x * psi
        return out

class UNetUpConvBlock(nn.Module):
    def __init__(self, in_channel, out_channel, upmode, ca_layer, up_factor = 2, att_mode = "standard"):
        super(UNetUpConvBlock, self).__init__()
        self.att_mode = att_mode
        self.ca_layer = ca_layer
        if upmode == 'upsample':
            self.Upsize = nn.Sequential(
                                        nn.Upsample(scale_factor=up_factor, mode='bilinear', align_corners=False),
                                        nn.Conv2d(in_channel, out_channel, kernel_size=1, stride=1, padding=0),
                                      )
        elif upmode == 'upconv':
            self.Upsize = nn.ConvTranspose2d(in_channel,out_channel,kernel_size=2,stride = 2)
        elif upmode == 'shuffle':
            self.Upsize = nn.Sequential(
                                        nn.Conv2d(in_channel,out_channel*4,kernel_size=3,stride=1,padding=1),
                                        nn.PReLU(),
                                        nn.PixelShuffle(2),
                                        nn.Conv2d(out_channel,out_channel,kernel_size=3,stride = 1,padding=1)
                                        )


        # self.conv = UNetConvBlock(in_channel, out_channel)
        if self.att_mode == 'standard':
            self.attention_gate = AttentionGate(out_channel, out_channel, out_channel)
            self.conv = UNetConvBlock(in_channel, out_channel, ca_layer=self.ca_layer)
        elif self.att_mode == 'modified':
            self.attention_gate = AttentionGate(out_channel, out_channel, out_channel )
            self.conv = UNetConvBlock(3*out_channel, out_channel, ca_layer = self.ca_layer)
        elif self.att_mode == 'None':
            self.conv = UNetConvBlock(in_channel, out_channel, ca_layer=self.ca_layer)



    def forward(self, x, residue):
        x = self.Upsize(x)
        x = F.interpolate(x, size=residue.shape[2:], mode='bilinear')
        if self.att_mode == "standard":
            attn = self.attention_gate(g = x, x=residue)
            out = torch.cat([x, attn],dim = 1)
            out = self.conv(out)
        elif self.att_mode == 'modified':
            attn = self.attention_gate(g = x, x = residue)
            out = torch.cat([x,residue,attn],dim = 1)
            out = self.conv(out)
        elif self.att_mode == 'None':
            out = torch.cat([x,residue], dim = 1)
            out = self.conv(out)
        return out





class AUNet(nn.Module):
    def __init__(self,in_channels = 6,out_channels = 6,depth = 3,growth_factor = 6,
                 interp_mode = 'bicubic', up_mode = 'upconv',spatial_attention = "standard", ca_layer = True):
        super(AUNet,self).__init__()

        if not spatial_attention in ['None', 'modified', 'standard']:
            raise AssertionError("spatial_attention options : \'None\'- no spatial attention, \'standard\'-spatial attention as in attention unet paper, \'modified\'-modified attention unet")

        self.in_channels = in_channels
        self.out_channels = out_channels
        self.depth = depth
        self.growth_factor = growth_factor
        self.interp_mode = interp_mode
        prev_channels = self.in_channels
        self.up_mode = up_mode
        self.att_mode = spatial_attention
        self.ca_layer = ca_layer

        self.encoding_module = nn.ModuleList()
        for i in range(self.depth):
            self.encoding_module.append(UNetConvBlock(in_channel=prev_channels,out_channel=2**(self.growth_factor + i), ca_layer=self.ca_layer))
            prev_channels = 2**(self.growth_factor+i)

        self.decoding_module = nn.ModuleList()
        for i in reversed(range(self.depth-1)):
            self.decoding_module.append(UNetUpConvBlock(prev_channels,2**(self.growth_factor+i),upmode = self.up_mode, att_mode = self.att_mode, ca_layer = self.ca_layer))
            prev_channels = 2**(self.growth_factor+i)

        self.final = nn.Conv2d(prev_channels,out_channels,1,1,0)

    def forward(self,MS,PAN = None):
        if PAN == None:
            x = MS
        else:
            x = torch.cat([MS,PAN],dim = 1)
        blocks = []
        for i,down in enumerate(self.encoding_module):
            x = down(x)
            if i != len(self.encoding_module)-1:
                blocks.append(x)
                x = F.avg_pool2d(x,2)

        for i,up in enumerate(self.decoding_module):
            x = up(x,blocks[-i-1])

        x = self.final(x)
        return x


if __name__ == '__main__':
    x = torch.rand([9,7,256,256]).cuda()
    model = AUNet(in_channels=7, out_channels=6, depth=5, spatial_attention="modified", growth_factor=6,
                  interp_mode='bilinear', up_mode='upconv', ca_layer=True).cuda()
    x = model(x)
    # print(model)


    activation = {}
    for layer in model:
        print(layer)