每日Attention学习25——Multi-Scale Attention Fusion

模块出处

[TCSVT 24] [link] [code] DSNet: A Novel Way to Use Atrous Convolutions in Semantic Segmentation

---

模块名称

Multi-Scale Attention Fusion (MSAF)

---

模块作用

双级特征融合

---

模块结构

---

模块思想

MSAF的主要思想是让网络根据损失学习特征权重，允许模型选择性地融合来自不同尺度的信息。

---

模块代码

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


class MSAF(nn.Module):
    def __init__(self, channels=64, r=4):
        super(MSAF, self).__init__()
        inter_channels = int(channels // r)

        self.local_att = nn.Sequential(
            nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
            nn.BatchNorm2d(inter_channels),
            nn.ReLU(inplace=True),
            nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
            nn.BatchNorm2d(channels),
        )

        self.context1 = nn.Sequential(
            nn.AdaptiveAvgPool2d((4, 4)),
            nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
            nn.BatchNorm2d(inter_channels),
            nn.ReLU(inplace=True),
            nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
            nn.BatchNorm2d(channels)
        )

        self.context2 = nn.Sequential(
            nn.AdaptiveAvgPool2d((8, 8)),
            nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
            nn.BatchNorm2d(inter_channels),
            nn.ReLU(inplace=True),
            nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
            nn.BatchNorm2d(channels)
        )

        self.global_att = nn.Sequential(
            nn.AdaptiveAvgPool2d(1),
            nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
            nn.BatchNorm2d(inter_channels),
            nn.ReLU(inplace=True),
            nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
            nn.BatchNorm2d(channels),
        )

        self.sigmoid = nn.Sigmoid()

    def forward(self, x, residual):
        h, w = x.shape[2], x.shape[3]
        xa = x + residual
        xl = self.local_att(xa)
        c1 = self.context1(xa)
        c2 = self.context2(xa)
        xg = self.global_att(xa)
        c1 = F.interpolate(c1, size=[h, w], mode='nearest')
        c2 = F.interpolate(c2, size=[h, w], mode='nearest')
        xlg = xl + xg + c1 + c2 
        wei = self.sigmoid(xlg)
        xo = 2 * x * wei + 2 * residual * (1 - wei)
        return xo
    

if __name__ == '__main__':
    msaf = MSAF()
    x1 = torch.randn([2, 64, 16, 16])
    x2 = torch.randn([2, 64, 16, 16])
    out = msaf(x1, x2)  
    print(out.shape)  # 2, 64, 16, 16

---