当前位置：首页 > news >正文

python打卡day49

news 来源：原创 2025/6/10 6:45:31

CBAM注意力

知识点回顾：

通道注意力模块复习
空间注意力模块
CBAM的定义

所有的模块本质上只是对特征进一步提取，之前学习了SE通道注意力，今天进一步了解一下CBAM注意力，它的核心目标是通过学习的方式，自动获取特征图在通道和空间维度上的重要性，进而对特征图进行自适应调整，增强重要特征，抑制不重要特征。所以很明显，CBAM有两个主要部分，通道注意力模块和空间注意力模块，输入特征图 → 通道注意力模块 → 空间注意力模块 → 输出增强后的特征图

通道注意力：分析 “哪些通道的特征更关键”（如图像中的颜色、纹理通道）；空间注意力：定位 “关键特征在图像中的具体位置”（如物体所在区域），二者结合：让模型同时学会 “关注什么” 和 “关注哪里”，提升特征表达能力

CBAM的通道注意力相较于传统的通道注意力机制有了点变动，同时使用全局平均池化和全局最大池化，通过共享的全连接层学习通道权重，能更捕获更全面的通道统计信息

import torch
import torch.nn as nn# 定义通道注意力
class ChannelAttention(nn.Module):def __init__(self, in_channels, ratio=16):"""通道注意力机制初始化参数:in_channels: 输入特征图的通道数ratio: 降维比例，用于减少参数量，默认为16"""super().__init__()# 全局平均池化，将每个通道的特征图压缩为1x1，保留通道间的平均值信息self.avg_pool = nn.AdaptiveAvgPool2d(1)# 全局最大池化，将每个通道的特征图压缩为1x1，保留通道间的最显著特征self.max_pool = nn.AdaptiveMaxPool2d(1)# 共享全连接层，用于学习通道间的关系# 先降维（除以ratio），再通过ReLU激活，最后升维回原始通道数self.fc = nn.Sequential(nn.Linear(in_channels, in_channels // ratio, bias=False),  # 降维层nn.ReLU(),  # 非线性激活函数nn.Linear(in_channels // ratio, in_channels, bias=False)   # 升维层)# Sigmoid函数将输出映射到0-1之间，作为各通道的权重self.sigmoid = nn.Sigmoid()def forward(self, x):"""前向传播函数参数:x: 输入特征图，形状为 [batch_size, channels, height, width]返回:调整后的特征图，通道权重已应用"""# 获取输入特征图的维度信息，这是一种元组的解包写法b, c, h, w = x.shape# 对平均池化结果进行处理：展平后通过全连接网络avg_out = self.fc(self.avg_pool(x).view(b, c))# 对最大池化结果进行处理：展平后通过全连接网络max_out = self.fc(self.max_pool(x).view(b, c))# 将平均池化和最大池化的结果相加并通过sigmoid函数得到通道权重attention = self.sigmoid(avg_out + max_out).view(b, c, 1, 1)# 将注意力权重与原始特征相乘，增强重要通道，抑制不重要通道return x * attention #这个运算是pytorch的广播机制

空间注意力需要从空间维度（H×W）上计算权重，关键步骤：

通道压缩：将输入特征图（形状 [B, C, H, W]）沿通道维度进行池化，生成2个特征图（平均池化和最大池化结果，形状 [B, 1, H, W]）这里的池化和上面通道注意力不一样，是直接对通道取均值和最大值
空间特征融合：将两种池化结果拼接（形状 [B, 2, H, W]），通过一个卷积层学习空间权重
权重应用：用Sigmoid生成空间注意力权重（0到1之间的值），与原特征图相乘

## 空间注意力模块
class SpatialAttention(nn.Module):def __init__(self, kernel_size=7):super().__init__()self.conv = nn.Conv2d(2, 1, kernel_size, padding=kernel_size//2, bias=False)self.sigmoid = nn.Sigmoid()def forward(self, x):# 通道维度池化avg_out = torch.mean(x, dim=1, keepdim=True)  # 平均池化：(B,1,H,W)max_out, _ = torch.max(x, dim=1, keepdim=True)  # 最大池化：(B,1,H,W)pool_out = torch.cat([avg_out, max_out], dim=1)  # 拼接：(B,2,H,W)attention = self.conv(pool_out)  # 卷积提取空间特征return x * self.sigmoid(attention)  # 特征与空间权重相乘

CBAM就是通道注意力+空间注意力，二者的输出是串行的

# CBAM模块
class CBAM(nn.Module):def __init__(self, in_channels, ratio=16, kernel_size=7):super().__init__()self.channel_attn = ChannelAttention(in_channels, ratio)self.spatial_attn = SpatialAttention(kernel_size)def forward(self, x):x = self.channel_attn(x)x = self.spatial_attn(x)return x

把CBMA模块插入到CNN网络里面一起训练

import torch
import torch.nn as nn
import torch.optim as optim
from torchvision import datasets, transforms
from torch.utils.data import DataLoader
import matplotlib.pyplot as plt
import numpy as np# 设置中文字体支持
plt.rcParams["font.family"] = ["SimHei"]
plt.rcParams['axes.unicode_minus'] = False  # 解决负号显示问题# 检查GPU是否可用
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"使用设备: {device}")# 数据预处理（与原代码一致）
train_transform = transforms.Compose([transforms.RandomCrop(32, padding=4),transforms.RandomHorizontalFlip(),transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.1),transforms.RandomRotation(15),transforms.ToTensor(),transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))
])test_transform = transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))
])# 加载数据集（与原代码一致）
train_dataset = datasets.CIFAR10(root='./data', train=True, download=True, transform=train_transform)
test_dataset = datasets.CIFAR10(root='./data', train=False, transform=test_transform)
train_loader = DataLoader(train_dataset, batch_size=64, shuffle=True)
test_loader = DataLoader(test_dataset, batch_size=64, shuffle=False)# 定义带有CBAM的CNN模型
class CBAM_CNN(nn.Module):def __init__(self):super(CBAM_CNN, self).__init__()# ---------------------- 第一个卷积块（带CBAM） ----------------------self.conv1 = nn.Conv2d(3, 32, kernel_size=3, padding=1)self.bn1 = nn.BatchNorm2d(32) # 批归一化self.relu1 = nn.ReLU()self.pool1 = nn.MaxPool2d(kernel_size=2)self.cbam1 = CBAM(in_channels=32)  # 在第一个卷积块后添加CBAM# ---------------------- 第二个卷积块（带CBAM） ----------------------self.conv2 = nn.Conv2d(32, 64, kernel_size=3, padding=1)self.bn2 = nn.BatchNorm2d(64)self.relu2 = nn.ReLU()self.pool2 = nn.MaxPool2d(kernel_size=2)self.cbam2 = CBAM(in_channels=64)  # 在第二个卷积块后添加CBAM# ---------------------- 第三个卷积块（带CBAM） ----------------------self.conv3 = nn.Conv2d(64, 128, kernel_size=3, padding=1)self.bn3 = nn.BatchNorm2d(128)self.relu3 = nn.ReLU()self.pool3 = nn.MaxPool2d(kernel_size=2)self.cbam3 = CBAM(in_channels=128)  # 在第三个卷积块后添加CBAM# ---------------------- 全连接层 ----------------------self.fc1 = nn.Linear(128 * 4 * 4, 512)self.dropout = nn.Dropout(p=0.5)self.fc2 = nn.Linear(512, 10)def forward(self, x):# 第一个卷积块x = self.conv1(x)x = self.bn1(x)x = self.relu1(x)x = self.pool1(x)x = self.cbam1(x)  # 应用CBAM# 第二个卷积块x = self.conv2(x)x = self.bn2(x)x = self.relu2(x)x = self.pool2(x)x = self.cbam2(x)  # 应用CBAM# 第三个卷积块x = self.conv3(x)x = self.bn3(x)x = self.relu3(x)x = self.pool3(x)x = self.cbam3(x)  # 应用CBAM# 全连接层x = x.view(-1, 128 * 4 * 4)x = self.fc1(x)x = self.relu3(x)x = self.dropout(x)x = self.fc2(x)return x# 初始化模型并移至设备
model = CBAM_CNN().to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=0.001)
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', patience=3, factor=0.5)# 训练函数
def train(model, train_loader, test_loader, criterion, optimizer, scheduler, device, epochs):model.train()all_iter_losses = []iter_indices = []train_acc_history = []test_acc_history = []train_loss_history = []test_loss_history = []for epoch in range(epochs):running_loss = 0.0correct = 0total = 0for 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()iter_loss = loss.item()all_iter_losses.append(iter_loss)iter_indices.append(epoch * len(train_loader) + batch_idx + 1)running_loss += iter_loss_, predicted = output.max(1)total += target.size(0)correct += predicted.eq(target).sum().item()if (batch_idx + 1) % 100 == 0:print(f'Epoch: {epoch+1}/{epochs} | Batch: {batch_idx+1}/{len(train_loader)} 'f'| 单Batch损失: {iter_loss:.4f} | 累计平均损失: {running_loss/(batch_idx+1):.4f}')epoch_train_loss = running_loss / len(train_loader)epoch_train_acc = 100. * correct / totaltrain_acc_history.append(epoch_train_acc)train_loss_history.append(epoch_train_loss)# 测试阶段model.eval()test_loss = 0correct_test = 0total_test = 0with torch.no_grad():for data, target in test_loader:data, target = data.to(device), target.to(device)output = model(data)test_loss += criterion(output, target).item()_, predicted = output.max(1)total_test += target.size(0)correct_test += predicted.eq(target).sum().item()epoch_test_loss = test_loss / len(test_loader)epoch_test_acc = 100. * correct_test / total_testtest_acc_history.append(epoch_test_acc)test_loss_history.append(epoch_test_loss)scheduler.step(epoch_test_loss)print(f'Epoch {epoch+1}/{epochs} 完成 | 训练准确率: {epoch_train_acc:.2f}% | 测试准确率: {epoch_test_acc:.2f}%')plot_iter_losses(all_iter_losses, iter_indices)plot_epoch_metrics(train_acc_history, test_acc_history, train_loss_history, test_loss_history)return epoch_test_acc# 绘图函数
def plot_iter_losses(losses, indices):plt.figure(figsize=(10, 4))plt.plot(indices, losses, 'b-', alpha=0.7, label='Iteration Loss')plt.xlabel('Iteration（Batch序号）')plt.ylabel('损失值')plt.title('每个 Iteration 的训练损失')plt.legend()plt.grid(True)plt.tight_layout()plt.show()def plot_epoch_metrics(train_acc, test_acc, train_loss, test_loss):epochs = range(1, len(train_acc) + 1)plt.figure(figsize=(12, 4))plt.subplot(1, 2, 1)plt.plot(epochs, train_acc, 'b-', label='训练准确率')plt.plot(epochs, test_acc, 'r-', label='测试准确率')plt.xlabel('Epoch')plt.ylabel('准确率 (%)')plt.title('训练和测试准确率')plt.legend()plt.grid(True)plt.subplot(1, 2, 2)plt.plot(epochs, train_loss, 'b-', label='训练损失')plt.plot(epochs, test_loss, 'r-', label='测试损失')plt.xlabel('Epoch')plt.ylabel('损失值')plt.title('训练和测试损失')plt.legend()plt.grid(True)plt.tight_layout()plt.show()# 执行训练
epochs = 50
print("开始使用带CBAM的CNN训练模型...")
final_accuracy = train(model, train_loader, test_loader, criterion, optimizer, scheduler, device, epochs)
print(f"训练完成！最终测试准确率: {final_accuracy:.2f}%")# # 保存模型
# torch.save(model.state_dict(), 'cifar10_cbam_cnn_model.pth')
# print("模型已保存为: cifar10_cbam_cnn_model.pth")

作业：尝试对今天的模型检查参数数目，并用tensorboard查看训练过程

检查参数好办，直接看就行

# 在模型初始化后查看参数量
def count_parameters(model):return sum(p.numel() for p in model.parameters() if p.requires_grad)print(f"模型总参数量: {count_parameters(model):,}")
print("各层参数分布:")
for name, param in model.named_parameters():if param.requires_grad:print(f"{name}: {param.numel():,}")

模型总参数量: 1,150,896
各层参数分布:
conv1.weight: 864
conv1.bias: 32
bn1.weight: 32
bn1.bias: 32
cbam1.channel_attn.fc.0.weight: 64
cbam1.channel_attn.fc.2.weight: 64
cbam1.spatial_attn.conv.weight: 98
conv2.weight: 18,432
conv2.bias: 64
bn2.weight: 64
bn2.bias: 64
cbam2.channel_attn.fc.0.weight: 256
cbam2.channel_attn.fc.2.weight: 256
cbam2.spatial_attn.conv.weight: 98
conv3.weight: 73,728
conv3.bias: 128
bn3.weight: 128
bn3.bias: 128
cbam3.channel_attn.fc.0.weight: 1,024
cbam3.channel_attn.fc.2.weight: 1,024
cbam3.spatial_attn.conv.weight: 98
fc1.weight: 1,048,576
fc1.bias: 512
fc2.weight: 5,120
fc2.bias: 10

用TensorBoard查看训练过程，先创建writer

# 创建TensorBoard的SummaryWriter，指定日志保存目录
log_dir = 'runs/CBAM_CNN_experiment'
# 如果目录已存在，添加后缀避免覆盖
if os.path.exists(log_dir):i = 1while os.path.exists(f"{log_dir}_{i}"):i += 1log_dir = f"{log_dir}_{i}"
writer = SummaryWriter(log_dir)

然后直接在训练函数里面修改就行，图方便只查看训练过程的损失、准确率（主要是改完代码忘了还有测试集的事儿，训练一次太费时间了不搞了）

# 训练函数
def train(model, train_loader, test_loader, criterion, optimizer, scheduler, device, epochs):model.train()# 记录训练开始时间，用于计算训练速度global_step = 0all_iter_losses = []iter_indices = []train_acc_history = []test_acc_history = []train_loss_history = []test_loss_history = []for epoch in range(epochs):running_loss = 0.0correct = 0total = 0for 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()iter_loss = loss.item()all_iter_losses.append(iter_loss)iter_indices.append(epoch * len(train_loader) + batch_idx + 1)running_loss += iter_loss_, predicted = output.max(1)total += target.size(0)correct += predicted.eq(target).sum().item()if (batch_idx + 1) % 100 == 0:batch_loss = loss.item()batch_acc = 100. * correct / total# 记录标量数据（损失、准确率）writer.add_scalar('Train/Batch_Loss', batch_loss, global_step)writer.add_scalar('Train/Batch_Accuracy', batch_acc, global_step)# 记录学习率writer.add_scalar('Train/Learning_Rate', optimizer.param_groups[0]['lr'], global_step)# 每500个批次记录一次直方图（权重和梯度）if (batch_idx + 1) % 500 == 0:for name, param in model.named_parameters():writer.add_histogram(f'weights/{name}', param, global_step)if param.grad is not None:writer.add_histogram(f'grads/{name}', param.grad, global_step)print(f'Epoch: {epoch+1}/{epochs} | Batch: {batch_idx+1}/{len(train_loader)} 'f'| 单Batch损失: {iter_loss:.4f} | 累计平均损失: {running_loss/(batch_idx+1):.4f}')global_step += 1# 计算当前epoch的平均训练损失和准确率epoch_train_loss = running_loss / len(train_loader)epoch_train_acc = 100. * correct / total# 记录每个epoch的训练损失和准确率writer.add_scalar('Train/Epoch_Loss', epoch_train_loss, epoch)writer.add_scalar('Train/Epoch_Accuracy', epoch_train_acc, epoch)# 测试阶段model.eval()test_loss = 0correct_test = 0total_test = 0with torch.no_grad():for data, target in test_loader:data, target = data.to(device), target.to(device)output = model(data)test_loss += criterion(output, target).item()_, predicted = output.max(1)total_test += target.size(0)correct_test += predicted.eq(target).sum().item()epoch_test_loss = test_loss / len(test_loader)epoch_test_acc = 100. * correct_test / total_testtest_acc_history.append(epoch_test_acc)test_loss_history.append(epoch_test_loss)scheduler.step(epoch_test_loss)print(f'Epoch {epoch+1}/{epochs} 完成 | 训练准确率: {epoch_train_acc:.2f}% | 测试准确率: {epoch_test_acc:.2f}%')plot_iter_losses(all_iter_losses, iter_indices)plot_epoch_metrics(train_acc_history, test_acc_history, train_loss_history, test_loss_history)# 关闭TensorBoard写入器writer.close()return epoch_test_acc

最后直接执行命令 tensorboard --logdir=runs 查看可视化结果就行

@浙大疏锦行