【学习】SE注意力，SEresnet代码实现。

SE是一个通道注意力机制。
参考：SE、CBAM、ECA注意力机制

se模块代码：

import torch.nn as nn
 
class SE_block(nn.Module):
    def __init__(self, channel, scaling=16): #scaling为缩放比例，
                                           # 用来控制两个全连接层中间神经网络神经元的个数，一般设置为16，具体可以根据需要微调
        super(SE_block, self).__init__()
        self.avg_pool = nn.AdaptiveAvgPool2d(1)
        self.fc = nn.Sequential(
                nn.Linear(channel, channel // scaling, bias=False),
                nn.ReLU(inplace=True),
                nn.Linear(channel // scaling, channel, bias=False),
                nn.Sigmoid()
        )
 
    def forward(self, x):
        b, c, _, _ = x.size()
        y = self.avg_pool(x).view(b, c)
        y = self.fc(y).view(b, c, 1, 1)
        return x * y

SEresnet18：

import torch
import torch.nn as nn
import torch.nn.functional as F
from torchsummary import summary
 
'''-------------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)
 
'''-------------二、BasicBlock模块-----------------------------'''
# 左侧的 residual block 结构（18-layer、34-layer）
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
 
'''-------------四、搭建SE_ResNet结构-----------------------------'''
class SE_ResNet(nn.Module):
    def __init__(self ,num_blocks=[2, 2, 2, 2], num_classes=10):
        super(SE_ResNet, self).__init__()
        self.in_planes = 64
 
        self.conv1 = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=7, stride=2, bias=False),
            nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True),
            nn.ReLU(inplace=True),
        )

        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
        self.layer1 = self._make_layer(BasicBlock, 64, num_blocks[0], stride=1)       # conv2_x
        self.layer2 = self._make_layer(BasicBlock, 128, num_blocks[1], stride=2)      # conv3_x
        self.layer3 = self._make_layer(BasicBlock, 256, num_blocks[2], stride=2)      # conv4_x
        self.layer4 = self._make_layer(BasicBlock, 512, num_blocks[3], stride=2)      # conv5_x
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        self.linear = nn.Linear(512 * BasicBlock.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 = self.conv1(x)
        x = self.maxpool(x)
        x1 = self.layer1(x)
        x2 = self.layer2(x1)
        x3 = self.layer3(x2)
        x4 = self.layer4(x3)

        return x2,x3,x4
 

 

这里的代码是只提取出了resnet的后三层。

注意：

在Resnet中添加SE注意力，如果是添加到残差模块中的话，只能自己手写Resnet代码，不能使用torch预训练的resnet。因为torch预训练的resnet的权重是按照原来resnet的结构分配的，在残差块中加入SE注意力机制改变了其原来的结构，所以会出现权重分配不匹配而报错的情况。

如果是加在预训练Resnet的第一个Conv 或者 是输出的最后一个Conv，便可以使用预训练的权重。