2017年,Hu等人发表了《Squeeze-and-Excitation Networks》,在文中创新性地提出了一种基于特征通道关系的轻量化模块,并将其称之为SE(Squeeze-and-Excitation) block,即SE注意力机制。这一模块为当时已经达到SOTA(state-of-the-art)水平的网络模型带来了进一步的效果提升,凭借其出色的效果和即插即用的特性,SE注意力机制成为了软注意力机制中的经典。
从文章中看
作者在文章中提出,SE模块可以利用图像特征通道之间的相互依赖性,使网络对于全局信息有选择性地强调某些重要的特征,抑制不太有用的特征,以此来增强模型的表达能力,提高网络的性能。这一功能即为一种注意力机制。文中对SE模块的各阶段步骤和原理进行了详尽介绍,并将SE模块与ResNet和Inception两种经典结构进行结合,通过大量实验进一步证明了SE模块的有效性。
浅谈原理
SE模块的结构示意如上图所示。可以看到对于输入信息X,先对其进行基本的卷积操作 F t r F_{tr} Ftr,得到大小为 W ∗ H ∗ C W*H*C W∗H∗C的特征U,其中W和H分别表示宽和高,C表示特征通道数。
接下来进入到了SE模块的核心操作,这部分操作分为两个阶段,即:Squeeze和Excitation,我们暂且称之为压缩和激活,两个阶段的具体步骤如下图所示。
压缩(Squeeze)
这部分是对输入的 W ∗ H ∗ C W*H*C W∗H∗C的特征U执行 F s q F_{sq} Fsq操作,将其输出为 1 ∗ 1 ∗ C 1*1*C 1∗1∗C的向量,可以理解为一个全局平均池化层,
激活(Excitation)
这一部分是对压缩后得到的输出向量执行 F e x F_{ex} Fex操作,操作由全连接+ReLU激活+全连接+Sigmoid复合构成。其中设置缩放参数r减少特征通道个数进而控制计算量。
在这一部分输出得到 1 ∗ 1 ∗ C 1*1*C 1∗1∗C的向量,其中便包含了特征中各通道权重值,将其与先前的特征U相乘,即可得到根据各通道之间相互依赖关系所激活的特征 X ^ \hat{X} X^。
以上便完成了SE模块的主要工作,可以看到,SE模块的引入对于现有网络模型结构的改动是极小的,而两个阶段所增加的计算量相较于性能的提升而言,基本可以忽略不计。
ResNet-18与SE注意力机制相结合的简易实现(PyTorch)
在残差网络中引入SE模块的设计结构如上图所示,作者在文中的实验里对ResNet-50进行了改进,而考虑到现有GPU条件有限,在这里我们尝试实现改进最简单的ResNet-18,更深层网络的改进步骤与ResNet-18类似,并选择MNIST数据集完成手写数字识别的任务,就实验实际效果而言,增加SE注意力机制后的模型的确得到了很大程度的提升。
import torch
from torch import nn
from torchvision.datasets import MNIST
from torch.nn import functional as F
from torch import optim
import torchvision.transforms as transforms
from torch.utils.data import DataLoader
class SEBlock(nn.Module):
def __init__(self, in_channels, reduction=16):
super(SEBlock, self).__init__()
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.fc = nn.Sequential(
nn.Linear(in_channels, in_channels // reduction, bias=False),
nn.ReLU(inplace=True),
nn.Linear(in_channels // reduction, in_channels, 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.expand_as(x)
class SE_ResidualBlock(nn.Module):
def __init__(self, in_channels, out_channels, strides=1):
super(SE_ResidualBlock, self).__init__()
self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=strides, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(out_channels)
self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(out_channels)
self.se = SEBlock(out_channels)
self.shortcut = None
if strides != 1 or in_channels != out_channels:
self.shortcut = nn.Sequential(
nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=strides, bias=False),
nn.BatchNorm2d(out_channels)
)
def forward(self, x):
out = F.relu(self.bn1(self.conv1(x)))
out = self.bn2(self.conv2(out))
out = self.se(out)
if self.shortcut:
x = self.shortcut(x)
out += x
out = F.relu(out)
return out
class SE_ResNet(nn.Module):
def __init__(self):
super(SE_ResNet, self).__init__()
self.conv1 = nn.Conv2d(1, 64, kernel_size=7, stride=2, padding=3)
self.bn1 = nn.BatchNorm2d(64)
self.pool1 = nn.MaxPool2d(kernel_size=3, stride=2)
self.layer1 = self.make_layer(64, 64, 2, strides=2)
self.layer2 = self.make_layer(64, 128, 2, strides=2)
self.layer3 = self.make_layer(128, 256, 2, strides=2)
self.layer4 = self.make_layer(256, 512, 2, strides=2)
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.fc = nn.Linear(512, 10)
def make_layer(self, in_channels, out_channels, num_blocks, strides=1):
layers = []
layers.append(SE_ResidualBlock(in_channels, out_channels, strides))
for _ in range(num_blocks - 1):
layers.append(SE_ResidualBlock(out_channels, out_channels, strides))
return nn.Sequential(*layers)
def forward(self, x):
out = self.conv1(x)
out = self.bn1(out)
out = nn.ReLU(inplace=True)(out)
out = self.pool1(out)
out = self.layer1(out)
out = self.layer2(out)
out = self.layer3(out)
out = self.layer4(out)
out = self.avgpool(out)
out = out.view(out.size(0), -1)
out = self.fc(out)
return out
# 定义数据预处理函数
transform = transforms.Compose([
transforms.RandomResizedCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
])
transform = transforms.Compose([transforms.ToTensor(),
transforms.Normalize((0.1307,),
(0.3081,))])
# 加载训练集和验证集
train_dataset = MNIST(root='../data/', train=True, transform=transform, download=True)
validation_dataset = MNIST(root='../data/', train=False, transform=transform)
batch_size = 128
train_loader = DataLoader(dataset=train_dataset, batch_size=batch_size, shuffle=True)
validation_loader = DataLoader(dataset=validation_dataset, batch_size=batch_size, shuffle=False)
model = SE_ResNet()
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print('Device:', device)
model = model.to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.9)
# 训练模型
num_epochs = 10
for epoch in range(num_epochs):
for i, (inputs, labels) in enumerate(train_loader):
inputs, labels = inputs.to(device), labels.to(device)
optimizer.zero_grad()
outputs = model(inputs)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
if (i + 1) % 100 == 0:
print('Epoch [{}/{}], Step [{}/{}], Loss: {:.4f}'.format(epoch + 1,
num_epochs, i + 1,
len(train_loader),
loss.item()))
# 在验证集上测试模型
with torch.no_grad():
correct = 0
total = 0
for inputs, labels in validation_loader:
inputs, labels = inputs.to(device), labels.to(device)
outputs = model(inputs)
_, predicted = torch.max(outputs.data, 1)
total += labels.size(0)
correct += (predicted == labels).sum().item()
accuracy = 100 * correct / total
print('Accuracy of the network on the validation images: %d %%' % (accuracy))