模块出处

[link] [code] [MM 21] Complementary Trilateral Decoder for Fast and Accurate Salient Object Detection

---

模块名称

Spatial Attention Module (SAM)

---

模块作用

空间注意力

---

模块结构

---

模块代码

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


def conv3x3(in_planes, out_planes, stride=1, padding=1, dilation=1, bias=False):
    "3x3 convolution with padding"
    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=padding, dilation=dilation, bias=bias)


class SAM(nn.Module):
    def __init__(self, in_chan, out_chan):
        super(SAM, self).__init__()
        self.conv_atten = conv3x3(2, 1)
        self.conv = conv3x3(in_chan, out_chan)
        self.bn = nn.BatchNorm2d(out_chan)

    def forward(self, x):
        avg_out = torch.mean(x, dim=1, keepdim=True)
        max_out, _ = torch.max(x, dim=1, keepdim=True)
        atten = torch.cat([avg_out, max_out], dim=1)
        atten = torch.sigmoid(self.conv_atten(atten))
        out = torch.mul(x, atten)
        out = F.relu(self.bn(self.conv(out)), inplace=True)
        return out

    
if __name__ == '__main__':
    x = torch.randn([1, 256, 16, 16])
    sam = SAM(in_chan=256, out_chan=64)
    out = sam(x)
    print(out.shape)  # 1, 64, 16, 16

---

原文表述

我们设计了空间注意力模块 (SAM)，以有效地完善特征（见图 3）。我们首先沿通道轴使用平均和最大运算，分别生成两个不同的单通道空间图$S_{avg}$和$S_{max}$。然后，我们将它们连接起来，通过3×3卷积和sigmoid函数计算出空间注意力图。空间注意力图$M_{sa}$可以通过元素级相乘从空间维度对特征重新加权。最后，细化后的特征被送入3×3卷积层，将通道压缩至64。