【CA】《Coordinate Attention for Efficient Mobile Network Design》

CVPR-2021

github：https://github.com/Andrew-Qibin/CoordAttention

---

---

1 Background and Motivation

SENet（【SENet】《Squeeze-and-Excitation Networks》） 注意力忽视了 position information

作者 embedding positional information into channel attention，采用 two 1D feature encoding processes，提出 coordinate attention，direction-aware and position-sensitive

2 Related Work

• Mobile Network Architecture
• Attention Mechanisms

3 Advantages / Contributions

• 设计提出轻量级 Coordinate Attention，兼顾 channel 和 spatial
• 在分类 / 目标检测 / 分割 benchmarks 上验证了其有效性

4 Method

SE / CBAM / CA 结构对比

CBAM 虽然有 spatial attention，但是 fail in modeling long-range dependencies that are essential for vision tasks

下面看看 CA 的公式表达

结合代码看看

import torch
import torch.nn as nn
import math
import torch.nn.functional as F
 
class h_sigmoid(nn.Module):
    def __init__(self, inplace=True):
        super(h_sigmoid, self).__init__()
        self.relu = nn.ReLU6(inplace=inplace)
 
    def forward(self, x):
        return self.relu(x + 3) / 6
 
class h_swish(nn.Module):
    def __init__(self, inplace=True):
        super(h_swish, self).__init__()
        self.sigmoid = h_sigmoid(inplace=inplace)
 
    def forward(self, x):
        return x * self.sigmoid(x)
 
class CoordAtt(nn.Module):
    def __init__(self, inp, oup, reduction=32):
        super(CoordAtt, self).__init__()
        self.pool_h = nn.AdaptiveAvgPool2d((None, 1))
        self.pool_w = nn.AdaptiveAvgPool2d((1, None))
 
        mip = max(8, inp // reduction)
 
        self.conv1 = nn.Conv2d(inp, mip, kernel_size=1, stride=1, padding=0)
        self.bn1 = nn.BatchNorm2d(mip)
        self.act = h_swish()
        
        self.conv_h = nn.Conv2d(mip, oup, kernel_size=1, stride=1, padding=0)
        self.conv_w = nn.Conv2d(mip, oup, kernel_size=1, stride=1, padding=0)
        
 
    def forward(self, x):
        identity = x
        
        n,c,h,w = x.size()
        x_h = self.pool_h(x) #（n, c, h, 1）
        x_w = self.pool_w(x).permute(0, 1, 3, 2)  #（n, c, 1, w）-> (n, c, w, 1) 注意这里转化成和 x_h 一样的排列形式了
        
        y = torch.cat([x_h, x_w], dim=2)  # 方便拼一起，conv->bn->act
        y = self.conv1(y)
        y = self.bn1(y)
        y = self.act(y) 
        
        x_h, x_w = torch.split(y, [h, w], dim=2)  # 再拆开
        x_w = x_w.permute(0, 1, 3, 2)  # 这里转化回去了
 
        a_h = self.conv_h(x_h).sigmoid()
        a_w = self.conv_w(x_w).sigmoid()
 
        out = identity * a_w * a_h
 
        return out

在网络中的插入位置，以 mobileNetV2（【MobileNet V2】《MobileNetV2：Inverted Residuals and Linear Bottlenecks》） 和 MobileNeXt 为例

5 Experiments

5.1 Datasets and Metrics

• ImageNet：Top-1 Acc
• COCO：mAP
• Cityscapes：mIoU

5.2 Ablation Studies

（1）Importance of coordinate attention

相比于仅水平或者竖直方向的 spatial attention，合起来最猛

（2）Different weight multipliers


0.25，0.75，1.0 均有一致性的提升

比 NAS 出来的 MobileNeXt + SE 都猛

AutoML 时代，提点的方式之一，加新的原料，加 search space 中的素材

（3）The impact of reduction ratio r

猛，均有提升

5.3 Comparison with Other Methods

（1）Attention for Mobile Networks

capture long-range dependencies among spatial locations that are essential for vision tasks

（2）Stronger Baseline

5.4 Applications

（1） Object Detection

COCO 和 VOC 都有提升

（2） Semantic Segmentation

比 CBAM 猛

6 Conclusion

来自：注意力的理解心得