在 PyTorch 中实现 SENet（Squeeze-and-Excitation Networks），核心是实现它的 "通道注意力机制"------ 通过 SE 模块给每个特征通道分配重要性权重，增强有用特征、抑制无用特征。我们从最基础的 SE 模块开始，一步步嵌入到 ResNet 中（形成 SE-ResNet），确保你能理解每个环节的作用。
一、先明确 SENet 的核心结构
SENet 的本质是 "在现有 CNN 中嵌入 SE 模块"，以经典的 SE-ResNet 为例，结构可以概括为：
输入图像 →
初始卷积层 → 池化层 →
多个残差块（每个残差块后嵌入SE模块） →
全局平均池化 → 全连接层(输出类别)其中，SE 模块（Squeeze-and-Excitation Module）是核心组件，由 "压缩→激励→重标定" 三步组成。
二、PyTorch 实现 SENet 的步骤
步骤 1：导入必要的库
import torch # 核心库
import torch.nn as nn # 神经网络层
import torch.optim as optim # 优化器
from torch.utils.data import DataLoader # 数据加载器
from torchvision import datasets, transforms # 图像数据处理步骤 2：实现核心组件 ------SE 模块
SE 模块是 SENet 的灵魂，包含三个关键步骤：压缩（Squeeze）、激励（Excitation）、重标定（Scale）：
python
class SEBlock(nn.Module):
def __init__(self, channel, reduction=16):
super(SEBlock, self).__init__()
# 1. 压缩（Squeeze）：全局平均池化，将H×W×C压缩为1×1×C
self.avg_pool = nn.AdaptiveAvgPool2d(1)
# 2. 激励（Excitation）：两个全连接层，学习通道权重
self.fc = nn.Sequential(
# 降维：减少计算量（通道数从C→C/reduction）
nn.Linear(channel, channel // reduction, bias=False),
nn.ReLU(inplace=True),
# 升维：恢复通道数（从C/reduction→C）
nn.Linear(channel // reduction, channel, bias=False),
nn.Sigmoid() # 将权重压缩到0~1之间
)
def forward(self, x):
b, c, _, _ = x.size() # 获取批次大小b和通道数c
# 压缩：全局平均池化 → 输出形状：(b, c, 1, 1)
y = self.avg_pool(x)
# 拉平成向量：(b, c)
y = y.view(b, c)
# 激励：学习权重 → 输出形状：(b, c)，每个元素是对应通道的权重
y = self.fc(y)
# 调整形状：(b, c, 1, 1)，方便后续广播乘法
y = y.view(b, c, 1, 1)
# 3. 重标定：将权重与原特征图相乘（通道级加权）
return x * y # 广播机制：每个通道的所有像素都乘以该通道的权重参数解释：
channel：输入特征图的通道数；reduction：降维系数（论文中推荐 16），用于减少全连接层的计算量（如通道数 256→256/16=16）。
通俗理解：
- 压缩步骤：给每个通道打一个 "全局平均分"（比如 "猫眼睛" 通道得分高，"背景噪音" 通道得分低）；
- 激励步骤：根据平均分学习每个通道的 "重要性权重"（0~1 之间）；
- 重标定步骤：用权重调整原特征（重要通道放大，不重要通道缩小）。
步骤 3：实现 SE-ResNet 的残差块（嵌入 SE 模块）
SENet 通常基于 ResNet 改进，我们以 ResNet 的 "瓶颈残差块"（Bottleneck）为例，在其中嵌入 SE 模块：
python
class SEBottleneck(nn.Module):
expansion = 4 # 残差块输出通道数是输入的4倍（ResNet瓶颈块的特性）
def __init__(self, in_channels, out_channels, stride=1, downsample=None, reduction=16):
super(SEBottleneck, self).__init__()
# 1×1卷积：降维（减少计算量）
self.conv1 = nn.Conv2d(
in_channels, out_channels, kernel_size=1, stride=1, bias=False
)
self.bn1 = nn.BatchNorm2d(out_channels)
# 3×3卷积：提取特征（主卷积层）
self.conv2 = nn.Conv2d(
out_channels, out_channels, kernel_size=3, stride=stride,
padding=1, bias=False
)
self.bn2 = nn.BatchNorm2d(out_channels)
# 1×1卷积：升维（恢复通道数）
self.conv3 = nn.Conv2d(
out_channels, out_channels * self.expansion, kernel_size=1,
stride=1, bias=False
)
self.bn3 = nn.BatchNorm2d(out_channels * self.expansion)
self.relu = nn.ReLU(inplace=True)
# 嵌入SE模块（在特征提取后、残差连接前）
self.se = SEBlock(out_channels * self.expansion, reduction)
# 下采样（当输入输出尺寸/通道不同时使用，确保残差连接维度匹配）
self.downsample = downsample
self.stride = stride
def forward(self, x):
identity = x # 残差连接的捷径分支
# 主分支：特征提取
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
out = self.relu(out)
out = self.conv3(out)
out = self.bn3(out)
# 关键：通过SE模块对特征加权
out = self.se(out)
# 残差连接：主分支输出 + 捷径分支
if self.downsample is not None:
identity = self.downsample(x) # 下采样调整捷径分支维度
out += identity
out = self.relu(out)
return out结构解释：
- 这是 ResNet 的 "瓶颈残差块"（1×1 Conv→3×3 Conv→1×1 Conv），在最后添加了 SE 模块；
- SE 模块位于 "特征提取完成后、残差连接前"，确保加权后的特征与捷径分支融合；
expansion=4表示输出通道数是中间 3×3 卷积通道数的 4 倍（如中间 3×3 用 64 通道，输出则为 256 通道）。
步骤 4：搭建 SE-ResNet 完整网络
以 SE-ResNet50 为例（50 层），由 4 个残差块组组成，每组分别包含 3、4、6、3 个 SEBottleneck：
python
class SEResNet(nn.Module):
def __init__(self, block, layers, num_classes=1000, reduction=16):
super(SEResNet, self).__init__()
self.in_channels = 64 # 初始卷积后的通道数
# 初始卷积层
self.conv1 = nn.Conv2d(
3, self.in_channels, kernel_size=7, stride=2, padding=3, bias=False
)
self.bn1 = nn.BatchNorm2d(self.in_channels)
self.relu = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) # 尺寸减半
# 4个残差块组（每个组的输出通道数翻倍）
self.layer1 = self._make_layer(
block, 64, layers[0], reduction=reduction
) # 输出通道：64×4=256
self.layer2 = self._make_layer(
block, 128, layers[1], stride=2, reduction=reduction
) # 输出通道：128×4=512
self.layer3 = self._make_layer(
block, 256, layers[2], stride=2, reduction=reduction
) # 输出通道：256×4=1024
self.layer4 = self._make_layer(
block, 512, layers[3], stride=2, reduction=reduction
) # 输出通道：512×4=2048
# 分类部分
self.avgpool = nn.AdaptiveAvgPool2d((1, 1)) # 全局平均池化
self.fc = nn.Linear(512 * block.expansion, num_classes) # 全连接层
def _make_layer(self, block, out_channels, blocks, stride=1, reduction=16):
"""创建一个残差块组"""
downsample = None
# 当步长>1或输入输出通道不同时，需要下采样调整捷径分支
if stride != 1 or self.in_channels != out_channels * block.expansion:
downsample = nn.Sequential(
nn.Conv2d(
self.in_channels, out_channels * block.expansion,
kernel_size=1, stride=stride, bias=False
),
nn.BatchNorm2d(out_channels * block.expansion),
)
layers = []
# 添加第一个残差块（可能需要下采样）
layers.append(
block(self.in_channels, out_channels, stride, downsample, reduction)
)
self.in_channels = out_channels * block.expansion # 更新输入通道数
# 添加剩余的残差块（步长=1，不需要下采样）
for _ in range(1, blocks):
layers.append(block(self.in_channels, out_channels, reduction=reduction))
return nn.Sequential(*layers)
def forward(self, x):
# 初始卷积和池化
x = self.conv1(x)
x = self.bn1(x)
x = self.relu(x)
x = self.maxpool(x)
# 4个残差块组（每个块都包含SE模块）
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
# 分类
x = self.avgpool(x)
x = x.view(x.size(0), -1) # 拉平成向量
x = self.fc(x)
return xSE-ResNet50 配置 ：
通过layers=[3,4,6,3]定义 4 个残差块组的数量，总层数计算为：
3+4+6+3=16个残差块 × 3 层卷积 / 块 + 初始卷积层 + 全连接层 ≈ 50 层。
步骤 5：初始化 SE-ResNet50 模型
用上面定义的模块组装 SE-ResNet50：
python
def se_resnet50(num_classes=1000, reduction=16):
"""创建SE-ResNet50模型"""
return SEResNet(SEBottleneck, [3, 4, 6, 3], num_classes, reduction)步骤 6：准备数据（用 CIFAR-10 演示）
SE-ResNet 适合高精度分类任务，我们用 CIFAR-10（10 类）演示，输入尺寸调整为 224×224：
python
# 数据预处理：缩放+裁剪+翻转+标准化
transform = transforms.Compose([
transforms.Resize(256), # 缩放为256×256
transforms.RandomCrop(224), # 随机裁剪成224×224
transforms.RandomHorizontalFlip(), # 数据增强
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) # ImageNet标准化
])
# 加载CIFAR-10数据集
train_dataset = datasets.CIFAR10(
root='./data', train=True, download=True, transform=transform
)
test_dataset = datasets.CIFAR10(
root='./data', train=False, download=True, transform=transform
)
# 批量加载数据
train_loader = DataLoader(train_dataset, batch_size=32, shuffle=True, num_workers=4)
test_loader = DataLoader(test_dataset, batch_size=32, shuffle=False, num_workers=4)步骤 7：初始化模型、损失函数和优化器
python
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# 初始化SE-ResNet50，输出10类（CIFAR-10）
model = se_resnet50(num_classes=10, reduction=16).to(device)
criterion = nn.CrossEntropyLoss() # 交叉熵损失
# 优化器：推荐用SGD+动量
optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.9, weight_decay=5e-4)步骤 8：训练和测试函数
SE-ResNet 的训练逻辑与普通 ResNet 类似，由于 SE 模块计算量极小，训练速度几乎不受影响：
python
def train(model, train_loader, criterion, optimizer, epoch):
model.train()
for batch_idx, (data, target) in enumerate(train_loader):
data, target = data.to(device), target.to(device)
optimizer.zero_grad() # 清空梯度
output = model(data) # 模型预测
loss = criterion(output, target) # 计算损失
loss.backward() # 反向传播
optimizer.step() # 更新参数
# 打印进度
if batch_idx % 50 == 0:
print(f'Epoch {epoch}, Batch {batch_idx}, Loss: {loss.item():.6f}')
def test(model, test_loader):
model.eval()
correct = 0
total = 0
with torch.no_grad():
for data, target in test_loader:
data, target = data.to(device), target.to(device)
output = model(data)
_, predicted = torch.max(output.data, 1)
total += target.size(0)
correct += (predicted == target).sum().item()
print(f'Test Accuracy: {100 * correct / total:.2f}%')步骤 9：开始训练和测试
SE-ResNet50 收敛速度与普通 ResNet50 相近，建议训练 30-50 轮：
python
for epoch in range(1, 31):
train(model, train_loader, criterion, optimizer, epoch)
test(model, test_loader)在 CIFAR-10 上，SE-ResNet50 的准确率比普通 ResNet50 高 1-2%，体现了 SE 模块的特征增强效果。
三、完整代码总结
python
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from torchvision import datasets, transforms
# 1. 实现SE模块（Squeeze-and-Excitation）
class SEBlock(nn.Module):
def __init__(self, channel, reduction=16):
super(SEBlock, self).__init__()
# 压缩：全局平均池化
self.avg_pool = nn.AdaptiveAvgPool2d(1)
# 激励：两个全连接层学习通道权重
self.fc = nn.Sequential(
nn.Linear(channel, channel // reduction, bias=False),
nn.ReLU(inplace=True),
nn.Linear(channel // reduction, channel, bias=False),
nn.Sigmoid()
)
def forward(self, x):
b, c, _, _ = x.size()
# 压缩：(b, c, h, w) → (b, c, 1, 1) → (b, c)
y = self.avg_pool(x).view(b, c)
# 激励：学习权重 → (b, c) → (b, c, 1, 1)
y = self.fc(y).view(b, c, 1, 1)
# 重标定：通道级加权
return x * y
# 2. 实现SE-ResNet的残差块（嵌入SE模块）
class SEBottleneck(nn.Module):
expansion = 4 # 输出通道是中间通道的4倍
def __init__(self, in_channels, out_channels, stride=1, downsample=None, reduction=16):
super(SEBottleneck, self).__init__()
self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
self.bn1 = nn.BatchNorm2d(out_channels)
self.conv2 = nn.Conv2d(
out_channels, out_channels, kernel_size=3, stride=stride,
padding=1, bias=False
)
self.bn2 = nn.BatchNorm2d(out_channels)
self.conv3 = nn.Conv2d(
out_channels, out_channels * self.expansion, kernel_size=1, bias=False
)
self.bn3 = nn.BatchNorm2d(out_channels * self.expansion)
self.relu = nn.ReLU(inplace=True)
self.se = SEBlock(out_channels * self.expansion, reduction) # 嵌入SE模块
self.downsample = downsample
self.stride = stride
def forward(self, x):
identity = x
# 主分支特征提取
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
out = self.relu(out)
out = self.conv3(out)
out = self.bn3(out)
# SE模块加权
out = self.se(out)
# 残差连接
if self.downsample is not None:
identity = self.downsample(x)
out += identity
out = self.relu(out)
return out
# 3. 搭建SE-ResNet完整网络
class SEResNet(nn.Module):
def __init__(self, block, layers, num_classes=1000, reduction=16):
super(SEResNet, self).__init__()
self.in_channels = 64
# 初始卷积层
self.conv1 = nn.Conv2d(
3, self.in_channels, kernel_size=7, stride=2, padding=3, bias=False
)
self.bn1 = nn.BatchNorm2d(self.in_channels)
self.relu = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
# 4个残差块组
self.layer1 = self._make_layer(block, 64, layers[0], reduction=reduction)
self.layer2 = self._make_layer(block, 128, layers[1], stride=2, reduction=reduction)
self.layer3 = self._make_layer(block, 256, layers[2], stride=2, reduction=reduction)
self.layer4 = self._make_layer(block, 512, layers[3], stride=2, reduction=reduction)
# 分类部分
self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
self.fc = nn.Linear(512 * block.expansion, num_classes)
def _make_layer(self, block, out_channels, blocks, stride=1, reduction=16):
downsample = None
if stride != 1 or self.in_channels != out_channels * block.expansion:
downsample = nn.Sequential(
nn.Conv2d(
self.in_channels, out_channels * block.expansion,
kernel_size=1, stride=stride, bias=False
),
nn.BatchNorm2d(out_channels * block.expansion),
)
layers = []
layers.append(block(self.in_channels, out_channels, stride, downsample, reduction))
self.in_channels = out_channels * block.expansion
for _ in range(1, blocks):
layers.append(block(self.in_channels, out_channels, reduction=reduction))
return nn.Sequential(*layers)
def forward(self, x):
x = self.conv1(x)
x = self.bn1(x)
x = self.relu(x)
x = self.maxpool(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
x = self.avgpool(x)
x = x.view(x.size(0), -1)
x = self.fc(x)
return x
# 4. 初始化SE-ResNet50模型
def se_resnet50(num_classes=1000, reduction=16):
return SEResNet(SEBottleneck, [3, 4, 6, 3], num_classes, reduction)
# 5. 准备CIFAR-10数据
transform = transforms.Compose([
transforms.Resize(256),
transforms.RandomCrop(224),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])
train_dataset = datasets.CIFAR10(
root='./data', train=True, download=True, transform=transform
)
test_dataset = datasets.CIFAR10(
root='./data', train=False, download=True, transform=transform
)
train_loader = DataLoader(train_dataset, batch_size=32, shuffle=True, num_workers=4)
test_loader = DataLoader(test_dataset, batch_size=32, shuffle=False, num_workers=4)
# 6. 初始化模型、损失函数和优化器
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = se_resnet50(num_classes=10, reduction=16).to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.SGD(model.parameters(), lr=0.01, momentum=0.9, weight_decay=5e-4)
# 7. 训练函数
def train(model, train_loader, criterion, optimizer, epoch):
model.train()
for batch_idx, (data, target) in enumerate(train_loader):
data, target = data.to(device), target.to(device)
optimizer.zero_grad()
output = model(data)
loss = criterion(output, target)
loss.backward()
optimizer.step()
if batch_idx % 50 == 0:
print(f'Epoch {epoch}, Batch {batch_idx}, Loss: {loss.item():.6f}')
# 8. 测试函数
def test(model, test_loader):
model.eval()
correct = 0
total = 0
with torch.no_grad():
for data, target in test_loader:
data, target = data.to(device), target.to(device)
output = model(data)
_, predicted = torch.max(output.data, 1)
total += target.size(0)
correct += (predicted == target).sum().item()
print(f'Test Accuracy: {100 * correct / total:.2f}%')
# 9. 开始训练和测试
for epoch in range(1, 31):
train(model, train_loader, criterion, optimizer, epoch)
test(model, test_loader)
四、关键知识点回顾
- SE 模块核心逻辑：通过 "压缩（全局平均池化）→激励（全连接层学习权重）→重标定（通道级加权）" 三步，让模型自动关注重要特征通道；
- 嵌入方式：SE 模块通常放在残差块的 "特征提取后、残差连接前"，确保加权后的优质特征参与后续计算；
- 参数设置 ：
reduction=16：降维系数，平衡计算量和性能（值越大，计算量越小，但可能损失精度）；- SE-ResNet50 的
layers=[3,4,6,3]：定义 4 个残差块组的数量，总层数约 50 层；
- 优势：计算量极小（额外参数仅 0.03%），兼容性强（可嵌入任何 CNN），精度提升明显（比普通 ResNet 高 1-2%）。
通过这段代码，你能亲手实现这个 "智能特征管家"，感受注意力机制如何用极小代价提升模型性能！