File size: 4,221 Bytes
beb7843
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
from ..basic.conv import Conv2d


# Channel Self Attetion Module
class CSAM(nn.Module):
    """ Channel attention module """
    def __init__(self):
        super(CSAM, self).__init__()
        self.gamma = nn.Parameter(torch.zeros(1))
        self.softmax  = nn.Softmax(dim=-1)


    def forward(self, x):
        """
            inputs :
                x : input feature maps( B x C x H x W )
            returns :
                out : attention value + input feature
                attention: B x C x C
        """
        B, C, H, W = x.size()
        # query / key / value
        query = x.view(B, C, -1)
        key = x.view(B, C, -1).permute(0, 2, 1)
        value = x.view(B, C, -1)

        # attention matrix
        energy = torch.bmm(query, key)
        energy_new = torch.max(energy, -1, keepdim=True)[0].expand_as(energy) - energy
        attention = self.softmax(energy_new)

        # attention
        out = torch.bmm(attention, value)
        out = out.view(B, C, H, W)

        # output
        out = self.gamma * out + x

        return out


# Spatial Self Attetion Module
class SSAM(nn.Module):
    """ Channel attention module """
    def __init__(self):
        super(SSAM, self).__init__()
        self.gamma = nn.Parameter(torch.zeros(1))
        self.softmax  = nn.Softmax(dim=-1)


    def forward(self, x):
        """
            inputs :
                x : input feature maps( B x C x H x W )
            returns :
                out : attention value + input feature
                attention: B x C x C
        """
        B, C, H, W = x.size()
        # query / key / value
        query = x.view(B, C, -1).permute(0, 2, 1)   # [B, N, C]
        key = x.view(B, C, -1)                      # [B, C, N]
        value = x.view(B, C, -1).permute(0, 2, 1)   # [B, N, C]

        # attention matrix
        energy = torch.bmm(query, key)
        energy_new = torch.max(energy, -1, keepdim=True)[0].expand_as(energy) - energy
        attention = self.softmax(energy_new)

        # attention
        out = torch.bmm(attention, value)
        out = out.permute(0, 2, 1).contiguous().view(B, C, H, W)

        # output
        out = self.gamma * out + x

        return out


# Channel Encoder
class ChannelEncoder(nn.Module):
    def __init__(self, in_dim, out_dim, act_type='', norm_type=''):
        super().__init__()
        self.fuse_convs = nn.Sequential(
            Conv2d(in_dim, out_dim, k=1, act_type=act_type, norm_type=norm_type),
            Conv2d(out_dim, out_dim, k=3, p=1, act_type=act_type, norm_type=norm_type),
            CSAM(),
            Conv2d(out_dim, out_dim, k=3, p=1, act_type=act_type, norm_type=norm_type),
            nn.Dropout(0.1, inplace=False),
            nn.Conv2d(out_dim, out_dim, kernel_size=1)
        )

    def forward(self, x1, x2):
        """
            x: [B, C, H, W]
        """
        x = torch.cat([x1, x2], dim=1)
        # [B, CN, H, W] -> [B, C, H, W]
        x = self.fuse_convs(x)

        return x


# Spatial Encoder
class SpatialEncoder(nn.Module):
    def __init__(self, in_dim, out_dim, act_type='', norm_type=''):
        super().__init__()
        self.fuse_convs = nn.Sequential(
            Conv2d(in_dim, out_dim, k=1, act_type=act_type, norm_type=norm_type),
            Conv2d(out_dim, out_dim, k=3, p=1, act_type=act_type, norm_type=norm_type),
            SSAM(),
            Conv2d(out_dim, out_dim, k=3, p=1, act_type=act_type, norm_type=norm_type),
            nn.Dropout(0.1, inplace=False),
            nn.Conv2d(out_dim, out_dim, kernel_size=1)
        )

    def forward(self, x):
        """
            x: [B, C, H, W]
        """
        x = self.fuse_convs(x)

        return x


def build_channel_encoder(cfg, in_dim, out_dim):
    encoder = ChannelEncoder(
            in_dim=in_dim,
            out_dim=out_dim,
            act_type=cfg['head_act'],
            norm_type=cfg['head_norm']
        )

    return encoder


def build_spatial_encoder(cfg, in_dim, out_dim):
    encoder = SpatialEncoder(
            in_dim=in_dim,
            out_dim=out_dim,
            act_type=cfg['head_act'],
            norm_type=cfg['head_norm']
        )

    return encoder