File size: 5,997 Bytes
d419456
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torch.nn.functional as F
from torchvision.models import efficientnet_b0, EfficientNet_B0_Weights


# ---------------------- Attention Modules ----------------------
class PositionAttentionModule(nn.Module):
    def __init__(self, in_channels):
        super().__init__()
        mid = max(1, in_channels // 8)
        self.conv_b = nn.Conv2d(in_channels, mid, 1, bias=False)
        self.conv_c = nn.Conv2d(in_channels, mid, 1, bias=False)
        self.conv_d = nn.Conv2d(in_channels, in_channels, 1, bias=False)
        self.softmax = nn.Softmax(dim=-1)

    def forward(self, x):
        b, c, h, w = x.size()
        B = self.conv_b(x).view(b, -1, h*w)
        C = self.conv_c(x).view(b, -1, h*w).permute(0, 2, 1)
        D = self.conv_d(x).view(b, -1, h*w)
        A = self.softmax(torch.bmm(C, B))
        out = torch.bmm(D, A.permute(0, 2, 1)).view(b, c, h, w)
        return out + x


class ChannelAttentionModule(nn.Module):
    def __init__(self, in_channels):
        super().__init__()
        self.gamma = nn.Parameter(torch.zeros(1))

    def forward(self, x):
        b, c, h, w = x.size()
        xf = x.view(b, c, -1)
        A = torch.bmm(xf, xf.permute(0, 2, 1))
        A = F.softmax(A, dim=-1)
        out = torch.bmm(A, xf).view(b, c, h, w)
        return self.gamma * out + x


class DANet(nn.Module):
    def __init__(self, in_channels):
        super().__init__()
        self.pam = PositionAttentionModule(in_channels)
        self.cam = ChannelAttentionModule(in_channels)
        self.fuse = nn.Conv2d(in_channels, in_channels, 1, bias=False)

    def forward(self, x):
        return self.fuse(self.pam(x) + self.cam(x))


class MSFE(nn.Module):
    def __init__(self, in_channels, branch_out=None):
        super().__init__()
        if branch_out is None:
            branch_out = max(1, in_channels // 3)
        self.b1 = nn.Conv2d(in_channels, branch_out, 1, bias=False)
        self.b2 = nn.Conv2d(in_channels, branch_out, 3, padding=1, bias=False)
        self.b3 = nn.Conv2d(in_channels, branch_out, 5, padding=2, bias=False)
        self.bn = nn.BatchNorm2d(branch_out * 3)
        self.proj = nn.Conv2d(branch_out * 3, in_channels, 1, bias=False)
        self.act = nn.ReLU(inplace=True)

    def forward(self, x):
        y = torch.cat([self.b1(x), self.b2(x), self.b3(x)], dim=1)
        y = self.proj(self.act(self.bn(y)))
        return y


class PreNorm(nn.Module):
    def __init__(self, c):
        super().__init__()
        self.bn = nn.BatchNorm2d(c)
        self.act = nn.ReLU(inplace=True)

    def forward(self, x):
        return self.act(self.bn(x))


class MSFE_then_DANet_Block(nn.Module):
    def __init__(self, c, branch_out=None, drop_p=0.0):
        super().__init__()
        self.pre = PreNorm(c)
        self.msfe = MSFE(c, branch_out=branch_out)
        self.danet = DANet(c)
        self.dp = nn.Dropout2d(drop_p) if drop_p > 0 else nn.Identity()

    def forward(self, x):
        y = self.pre(x)
        y = self.msfe(y)
        y = self.danet(y)
        y = self.dp(y)
        return x + y


class DANet_Block(nn.Module):
    def __init__(self, c, drop_p=0.0):
        super().__init__()
        self.pre = PreNorm(c)
        self.danet = DANet(c)
        self.dp = nn.Dropout2d(drop_p) if drop_p > 0 else nn.Identity()

    def forward(self, x):
        y = self.dp(self.danet(self.pre(x)))
        return x + y


# ---------------------- Main Model ----------------------
class EfficientNetB0Hybrid(nn.Module):
    def __init__(self, num_classes=7, msfe_then_danet_indices=(6,), danet_only_indices=(4,),
                 branch_out_ratio=1/3, drop_p=0.0, use_pretrained=True):
        super().__init__()
        self.backbone = efficientnet_b0(weights=EfficientNet_B0_Weights.DEFAULT if use_pretrained else None)
        self.features = self.backbone.features
        self.pool = nn.AdaptiveAvgPool2d(1)
        self.classifier = nn.Linear(self.backbone.classifier[1].in_features, num_classes)

        # Capture feature shapes
        self._feat_shapes = {}
        self._register_shape_hooks(msfe_then_danet_indices, danet_only_indices)
        with torch.no_grad():
            dummy = torch.zeros(1, 3, 224, 224)
            _ = self._run_features_with_capture(dummy)
        for h in self._hooks: h.remove()

        self.msfe_then_danet_blocks = nn.ModuleDict()
        self.danet_only_blocks = nn.ModuleDict()

        for idx in msfe_then_danet_indices:
            C = self._feat_shapes[idx][1]
            br = max(1, int(C * branch_out_ratio))
            self.msfe_then_danet_blocks[str(idx)] = MSFE_then_DANet_Block(C, branch_out=br, drop_p=drop_p)

        for idx in danet_only_indices:
            C = self._feat_shapes[idx][1]
            self.danet_only_blocks[str(idx)] = DANet_Block(C, drop_p=drop_p)

        self.msfe_then_danet_indices = set(msfe_then_danet_indices)
        self.danet_only_indices = set(danet_only_indices)

    def _register_shape_hooks(self, msfe_then_danet_indices, danet_only_indices):
        self._hooks = []
        wanted = set(msfe_then_danet_indices) | set(danet_only_indices)

        def make_hook(i):
            def hook(_m, _inp, out):
                if isinstance(out, (list, tuple)):
                    out = out[0]
                self._feat_shapes[i] = out.shape
            return hook

        for i, m in enumerate(self.features):
            if i in wanted:
                self._hooks.append(m.register_forward_hook(make_hook(i)))

    def _run_features_with_capture(self, x):
        for m in self.features:
            x = m(x)
        return x

    def forward(self, x):
        for i, m in enumerate(self.features):
            x = m(x)
            if i in self.msfe_then_danet_indices:
                x = self.msfe_then_danet_blocks[str(i)](x)
            if i in self.danet_only_indices:
                x = self.danet_only_blocks[str(i)](x)
        x = self.pool(x).flatten(1)
        return self.classifier(x)