pytorch实现SeNet

论文链接：https://arxiv.org/abs/1709.01507
senet的全称为Squeeze-and-Excitation Networks，压缩-激活网络。

网络结构

$F_{tr}$可以看做是对feature-map的transform，将$X(H^{{}^{\prime }}\times W^{{}^{\prime }}\times C^{{}^{\prime }})$压缩到$U(H\times W\times C)$,这个transform过程可以是一个卷积。之后，$U$后面接一个se模块。se模块分为两个分支，上面的分支先是一个global average pooling，作为一个squeeze过程$F_{sq}(.)$,此时特征变为了$1\times 1\times C$的特征，之后经过一个特征激活过程$F_{ex}(.,W)$,这里接的是全连接层和激活函数，处理之后的特征$S$维度依然是$1\times 1\times C$。通过$F_{scale}(.,.)$过程将$U$和$S$沿着channel相乘，得到输出的特征图$H\times W\times C$。
作者在Inception和ResNet模块上使用了se模块：

上述的se模块的顺序是global average pooling-FC-ReLU-FC-Sigmoid过程。$r$是中间隐藏状态特征的维度，作者在后面也就是了调参实验。作者最后用了Sigmoid函数确保了权重大小在(0,1)之间。

参数量分析

作者分析了加入se模块到ResNet的参数的增加量。SE-ResNet-50在原来25M参数量的基础上增加了2.5M参数，增加了10%。作者分析，参数量的增加主要在于final stage的全连接层，此时具有最大数量的channel。此时删掉final stage的se模块，参数量只增加了4%，精度下降了<0.1%。

代码链接：https://github.com/kuangliu/pytorch-cifar/blob/master/models/senet.py

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


class BasicBlock(nn.Module):
    def __init__(self, in_planes, planes, stride=1):
        super(BasicBlock, self).__init__()
        self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
        self.bn1 = nn.BatchNorm2d(planes)
        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(planes)

        self.shortcut = nn.Sequential()
        if stride != 1 or in_planes != planes:
            self.shortcut = nn.Sequential(
                nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride, bias=False),
                nn.BatchNorm2d(planes)
            )

        # SE layers
        self.fc1 = nn.Conv2d(planes, planes//16, kernel_size=1)  # Use nn.Conv2d instead of nn.Linear
        self.fc2 = nn.Conv2d(planes//16, planes, kernel_size=1)

    def forward(self, x):
        out = F.relu(self.bn1(self.conv1(x)))
        out = self.bn2(self.conv2(out))

        # Squeeze
        w = F.avg_pool2d(out, out.size(2))
        w = F.relu(self.fc1(w))
        w = F.sigmoid(self.fc2(w))
        # Excitation
        out = out * w  # New broadcasting feature from v0.2!

        out += self.shortcut(x)
        out = F.relu(out)
        return out


class PreActBlock(nn.Module):
    def __init__(self, in_planes, planes, stride=1):
        super(PreActBlock, self).__init__()
        self.bn1 = nn.BatchNorm2d(in_planes)
        self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
        self.bn2 = nn.BatchNorm2d(planes)
        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, padding=1, bias=False)

        if stride != 1 or in_planes != planes:
            self.shortcut = nn.Sequential(
                nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride, bias=False)
            )

        # SE layers
        self.fc1 = nn.Conv2d(planes, planes//16, kernel_size=1)
        self.fc2 = nn.Conv2d(planes//16, planes, kernel_size=1)

    def forward(self, x):
        out = F.relu(self.bn1(x))
        shortcut = self.shortcut(out) if hasattr(self, 'shortcut') else x
        out = self.conv1(out)
        out = self.conv2(F.relu(self.bn2(out)))

        # Squeeze
        w = F.avg_pool2d(out, out.size(2))
        w = F.relu(self.fc1(w))
        w = F.sigmoid(self.fc2(w))
        # Excitation
        out = out * w

        out += shortcut
        return out


class SENet(nn.Module):
    def __init__(self, block, num_blocks, num_classes=10):
        super(SENet, self).__init__()
        self.in_planes = 64

        self.conv1 = nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False)
        self.bn1 = nn.BatchNorm2d(64)
        self.layer1 = self._make_layer(block,  64, num_blocks[0], stride=1)
        self.layer2 = self._make_layer(block, 128, num_blocks[1], stride=2)
        self.layer3 = self._make_layer(block, 256, num_blocks[2], stride=2)
        self.layer4 = self._make_layer(block, 512, num_blocks[3], stride=2)
        self.linear = nn.Linear(512, 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
        return nn.Sequential(*layers)

    def forward(self, x):
        out = F.relu(self.bn1(self.conv1(x)))
        out = self.layer1(out)
        out = self.layer2(out)
        out = self.layer3(out)
        out = self.layer4(out)
        out = F.avg_pool2d(out, 4)
        out = out.view(out.size(0), -1)
        out = self.linear(out)
        return out


def SENet18():
    return SENet(PreActBlock, [2,2,2,2])


def test():
    net = SENet18()
    y = net(torch.randn(1,3,32,32))
    print(y.size())

# test()