每日Attention学习13——Adaptive Feature Fusion

模块出处

[AAAI 23] [link] [code] High-Resolution Iterative Feedback Network for Camoufaged Object Detection

---

模块名称

Adaptive Feature Fusion

---

模块作用

多尺度特征融合

---

模块结构

---

模块代码

import torch
import torch.nn as nn

class AFF(nn.Module):
    def __init__(self, ch_in=32, reduction=16):
        super(AFF, self).__init__()
        self.avg_pool = nn.AdaptiveAvgPool2d(1)
        self.fc = nn.Sequential(
            nn.Linear(ch_in, ch_in // reduction, bias=False),
            nn.ReLU(inplace=True),
            nn.Linear(ch_in // reduction, ch_in, bias=False),
            nn.Sigmoid()
        )
        self.fc_wight = nn.Sequential(
            nn.Linear(ch_in, ch_in // reduction, bias=False),
            nn.ReLU(inplace=True),
            nn.Linear(ch_in // reduction, 1, bias=False),
            nn.Sigmoid()
        )
    
    def forward(self, x_h, x_l):
        # 增强high level特征, 在Squeeze and Excite上引入了一个额外的h_weight($\alpha_1$)
        b, c, _, _ = x_h.size()
        y_h = self.avg_pool(x_h).view(b, c)
        h_weight = self.fc_wight(y_h)
        y_h = self.fc(y_h).view(b, c, 1, 1)
        x_fusion_h = x_h * y_h.expand_as(x_h)
        x_fusion_h = torch.mul(x_fusion_h, h_weight.view(b, 1, 1, 1))

        # 增强low level特征, 在Squeeze and Excite上引入了一个额外的l_weight($\alpha_2$)
        b, c, _, _ = x_l.size()
        y_l = self.avg_pool(x_l).view(b, c) 
        l_weight = self.fc_wight(y_l)
        y_l = self.fc(y_l).view(b, c, 1, 1)
        x_fusion_l = x_l * y_l.expand_as(x_l)
        x_fusion_l = torch.mul(x_fusion_l, l_weight.view(b, 1, 1, 1))

        # 多级特征融合
        x_fusion = x_fusion_h + x_fusion_l
        return x_fusion
    
if __name__ == '__main__':
    aff = AFF()
    x_h = torch.randn([1, 32, 16, 16])
    x_l = torch.randn([1, 32, 16, 16])
    out = aff(x_h, x_l)
    print(out.shape)  # 1, 32, 16, 16