注意力机制（一）SE模块（Squeeze-and-Excitation Networks）论文总结和代码实现

Squeeze-and-Excitation Networks（压缩和激励网络）

论文地址：Squeeze-and-Excitation Networks

论文中文版：Squeeze-and-Excitation Networks_中文版

代码地址：GitHub - hujie-frank/SENet: Squeeze-and-Excitation Networks

 

一、论文出发点

为了提高网络的表示能力，许多现有的工作已经显示出增强空间编码的好处。而作者专注于通道，希望能够提出了一种新的架构单元，通过显式地建模出卷积特征通道之间的相互依赖性来提高网络的表示能力。

这里引用“博文：Squeeze-and-Excitation Networks解读”中的总结：核心思想是不同通道的权重应该自适应分配，由网络自己学习出来的，而不是像Inception net一样留下过多人工干预的痕迹。

 

 

二、论文的主要工作

1.提出了一种新的架构单元Squeeze-and-Excitation模块，该模块可以显式地建模卷积特征通道之间的相互依赖性来提高网络的表示能力。

2.提出了一种机制，使网络能够执行特征重新校准，通过这种机制可以学习使用全局信息来选择性地强调信息特征并抑制不太有用的特征。

 

 

三、Squeeze-and-Excitation模块

3.1 Transformation（Ftr）（转型）

3.2 Squeeze（全局信息嵌入）

 

 

四、模型：SE-Inception和SE-ResNet

五、实验

六、结论

本文提出的SE模块，这是一种新颖的架构单元，旨在通过使网络能够执行动态通道特征重新校准来提高网络的表示能力。大量实验证明了SENets的有效性，其在多个数据集上取得了最先进的性能。

七、源码分析

将SEblock嵌入ResNet的残差模块中。

7.1 SE模块

'''-------------一、SE模块-----------------------------'''
#全局平均池化+1*1卷积核+ReLu+1*1卷积核+Sigmoid
class SE_Block(nn.Module):
    def __init__(self, inchannel, ratio=16):
        super(SE_Block, self).__init__()
        # 全局平均池化(Fsq操作)
        self.gap = nn.AdaptiveAvgPool2d((1, 1))
        # 两个全连接层(Fex操作)
        self.fc = nn.Sequential(
            nn.Linear(inchannel, inchannel // ratio, bias=False),  # 从 c -> c/r
            nn.ReLU(),
            nn.Linear(inchannel // ratio, inchannel, bias=False),  # 从 c/r -> c
            nn.Sigmoid()
        )
 
    def forward(self, x):
            # 读取批数据图片数量及通道数
            b, c, h, w = x.size()
            # Fsq操作：经池化后输出b*c的矩阵
            y = self.gap(x).view(b, c)
            # Fex操作：经全连接层输出（b，c，1，1）矩阵
            y = self.fc(y).view(b, c, 1, 1)
            # Fscale操作：将得到的权重乘以原来的特征图x
            return x * y.expand_as(x)

7.2 SE-ResNet完整代码

不同版本的ResNet各层主要是由BasicBlock模块（18-layer、34-layer）或Bottleneck模块（50-layer、101-layer、152-layer）构成的。

添加SE模块位置：在BasicBlock模块或Bottleneck模块尾部添加，但是要注意放在shortcut之前。

import torch
import torch.nn as nn
import torch.nn.functional as F
from torchsummary import summary
 
'''-------------SE模块-----------------------------'''
#全局平均池化+1*1卷积核+ReLu+1*1卷积核+Sigmoid
class SE_Block(nn.Module):
    def __init__(self, inchannel, ratio=16):
        super(SE_Block, self).__init__()
        # 全局平均池化(Fsq操作)
        self.gap = nn.AdaptiveAvgPool2d((1, 1))
        # 两个全连接层(Fex操作)
        self.fc = nn.Sequential(
            nn.Linear(inchannel, inchannel // ratio, bias=False),  # 从 c -> c/r
            nn.ReLU(),
            nn.Linear(inchannel // ratio, inchannel, bias=False),  # 从 c/r -> c
            nn.Sigmoid()
        )
 
    def forward(self, x):
            # 读取批数据图片数量及通道数
            b, c, h, w = x.size()
            # Fsq操作：经池化后输出b*c的矩阵
            y = self.gap(x).view(b, c)
            # Fex操作：经全连接层输出（b，c，1，1）矩阵
            y = self.fc(y).view(b, c, 1, 1)
            # Fscale操作：将得到的权重乘以原来的特征图x
            return x * y.expand_as(x)
 
'''-------------（18-layer、34-layer）BasicBlock模块-----------------------------'''
# residual block 结构
class BasicBlock(nn.Module):
    expansion = 1
 
    def __init__(self, inchannel, outchannel, stride=1):
        super(BasicBlock, self).__init__()
        self.conv1 = nn.Conv2d(inchannel, outchannel, kernel_size=3,
                               stride=stride, padding=1, bias=False)
        self.bn1 = nn.BatchNorm2d(outchannel)
        self.conv2 = nn.Conv2d(outchannel, outchannel, kernel_size=3,
                               stride=1, padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(outchannel)
        # SE_Block放在BN之后，shortcut之前
        self.SE = SE_Block(outchannel)
 
        self.shortcut = nn.Sequential()
        if stride != 1 or inchannel != self.expansion*outchannel:
            self.shortcut = nn.Sequential(
                nn.Conv2d(inchannel, self.expansion*outchannel,
                          kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(self.expansion*outchannel)
            )
 
    def forward(self, x):
        out = F.relu(self.bn1(self.conv1(x)))
        out = self.bn2(self.conv2(out))
        SE_out = self.SE(out)
        out = out * SE_out
        out += self.shortcut(x)
        out = F.relu(out)
        return out
 
'''-------------（50-layer、101-layer、152-layer）Bottleneck模块-----------------------------'''
# residual block 结构
class Bottleneck(nn.Module):
    expansion = 4
 
    def __init__(self, inchannel, outchannel, stride=1):
        super(Bottleneck, self).__init__()
        self.conv1 = nn.Conv2d(inchannel, outchannel, kernel_size=1, bias=False)
        self.bn1 = nn.BatchNorm2d(outchannel)
        self.conv2 = nn.Conv2d(outchannel, outchannel, kernel_size=3,
                               stride=stride, padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(outchannel)
        self.conv3 = nn.Conv2d(outchannel, self.expansion*outchannel,
                               kernel_size=1, bias=False)
        self.bn3 = nn.BatchNorm2d(self.expansion*outchannel)
        # SE_Block放在BN之后，shortcut之前
        self.SE = SE_Block(self.expansion*outchannel)
 
        self.shortcut = nn.Sequential()
        if stride != 1 or inchannel != self.expansion*outchannel:
            self.shortcut = nn.Sequential(
                nn.Conv2d(inchannel, self.expansion*outchannel,
                          kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(self.expansion*outchannel)
            )
 
    def forward(self, x):
        out = F.relu(self.bn1(self.conv1(x)))
        out = F.relu(self.bn2(self.conv2(out)))
        out = self.bn3(self.conv3(out))
        SE_out = self.SE(out)
        out = out * SE_out
        out += self.shortcut(x)
        out = F.relu(out)
        return out
 
'''-------------搭建SE_ResNet结构-----------------------------'''
class SE_ResNet(nn.Module):
    def __init__(self, block, num_blocks, num_classes=10):
        super(SE_ResNet, self).__init__()
        self.in_planes = 64
 
        self.conv1 = nn.Conv2d(3, 64, kernel_size=3,
                               stride=1, padding=1, bias=False)                  # conv1
        self.bn1 = nn.BatchNorm2d(64)
        self.layer1 = self._make_layer(block, 64, num_blocks[0], stride=1)       # conv2_x
        self.layer2 = self._make_layer(block, 128, num_blocks[1], stride=2)      # conv3_x
        self.layer3 = self._make_layer(block, 256, num_blocks[2], stride=2)      # conv4_x
        self.layer4 = self._make_layer(block, 512, num_blocks[3], stride=2)      # conv5_x
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        self.linear = nn.Linear(512 * block.expansion, num_classes)
 
    def _make_layer(self, block, planes, num_blocks, stride):
        strides = [stride] + [1]*(num_blocks-1)
        layers = []
        for stride in strides:
            layers.append(block(self.in_planes, planes, stride))
            self.in_planes = planes * block.expansion
        return nn.Sequential(*layers)
 
    def forward(self, x):
        x = F.relu(self.bn1(self.conv1(x)))
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)
        x = self.avgpool(x)
        x = torch.flatten(x, 1)
        out = self.linear(x)
        return out
 
 
def SE_ResNet18():
    return SE_ResNet(BasicBlock, [2, 2, 2, 2])
 
 
def SE_ResNet34():
    return SE_ResNet(BasicBlock, [3, 4, 6, 3])
 
 
def SE_ResNet50():
    return SE_ResNet(Bottleneck, [3, 4, 6, 3])
 
 
def SE_ResNet101():
    return SE_ResNet(Bottleneck, [3, 4, 23, 3])
 
 
def SE_ResNet152():
    return SE_ResNet(Bottleneck, [3, 8, 36, 3])
 
 
'''
if __name__ == '__main__':
    model = SE_ResNet50()
    print(model)
    input = torch.randn(1, 3, 224, 224)
    out = model(input)
    print(out.shape)
# test()
'''
if __name__ == '__main__':
    device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
    net = SE_ResNet50().to(device)
    # 打印网络结构和参数
    summary(net, (3, 224, 224))