DAY 52 神经网络调参指南

 

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# 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

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}")# 1. 数据预处理
# 训练集：使用多种数据增强方法提高模型泛化能力
train_transform = transforms.Compose([# 随机裁剪图像，从原图中随机截取32x32大小的区域transforms.RandomCrop(32, padding=4),# 随机水平翻转图像（概率0.5）transforms.RandomHorizontalFlip(),# 随机颜色抖动：亮度、对比度、饱和度和色调随机变化transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.1),# 随机旋转图像（最大角度15度）transforms.RandomRotation(15),# 将PIL图像或numpy数组转换为张量transforms.ToTensor(),# 标准化处理：每个通道的均值和标准差，使数据分布更合理transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))
])# 测试集：仅进行必要的标准化，保持数据原始特性，标准化不损失数据信息，可还原
test_transform = transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010))
])# 2. 加载CIFAR-10数据集
train_dataset = datasets.CIFAR10(root='./data',train=True,download=True,transform=train_transform  # 使用增强后的预处理
)test_dataset = datasets.CIFAR10(root='./data',train=False,transform=test_transform  # 测试集不使用增强
)# 3. 创建数据加载器
batch_size = 80
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
test_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False)

# 4. 定义CNN模型的定义（替代原MLP）
class CNN(nn.Module):def __init__(self):super(CNN, self).__init__()# 初始卷积层self.conv_init = nn.Sequential(nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False),nn.BatchNorm2d(64),nn.ReLU())# 第一卷积块（含CBAM）self.block1 = nn.Sequential(nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=1, bias=False),nn.BatchNorm2d(64),nn.ReLU(),nn.Conv2d(64, 64, kernel_size=3, stride=1, padding=1, bias=False),nn.BatchNorm2d(64),CBAM(64)  # 在卷积块后添加CBAM)self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2)self.drop1 = nn.Dropout2d(0.1)# 第二卷积块（含CBAM）self.block2 = nn.Sequential(nn.Conv2d(64, 128, kernel_size=3, stride=2, padding=1, bias=False),  # stride=2降维nn.BatchNorm2d(128),nn.ReLU(),nn.Conv2d(128, 128, kernel_size=3, stride=1, padding=1, bias=False),nn.BatchNorm2d(128),CBAM(128)  # 在卷积块后添加CBAM)self.pool2 = nn.AvgPool2d(kernel_size=2, stride=2)self.drop2 = nn.Dropout2d(0.2)# 第三卷积块（含CBAM）self.block3 = nn.Sequential(nn.Conv2d(128, 256, kernel_size=3, stride=2, padding=1, bias=False),nn.BatchNorm2d(256),nn.ReLU(),nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1, bias=False),nn.BatchNorm2d(256),CBAM(256)  # 在卷积块后添加CBAM)self.pool3 = nn.AdaptiveAvgPool2d(4)self.drop3 = nn.Dropout2d(0.3)# 全连接层self.fc = nn.Sequential(nn.Linear(256 * 4 * 4, 512),nn.BatchNorm1d(512),nn.ReLU(),nn.Dropout(0.5),nn.Linear(512, 128),nn.BatchNorm1d(128),nn.ReLU(),nn.Dropout(0.3),nn.Linear(128, 10))def forward(self, x):x = self.conv_init(x)x = self.block1(x)x = self.pool1(x)x = self.drop1(x)x = self.block2(x)x = self.pool2(x)x = self.drop2(x)x = self.block3(x)x = self.pool3(x)x = self.drop3(x)x = x.view(-1, 256 * 4 * 4)x = self.fc(x)return x# 初始化模型
model = CNN()
model = model.to(device)  # 将模型移至GPU（如果可用）

criterion = nn.CrossEntropyLoss()  # 交叉熵损失函数
optimizer = optim.Adam(model.parameters(), lr=0.001)  # Adam优化器# 引入学习率调度器，在训练过程中动态调整学习率--训练初期使用较大的 LR 快速降低损失，训练后期使用较小的 LR 更精细地逼近全局最优解。
# 在每个 epoch 结束后，需要手动调用调度器来更新学习率，可以在训练过程中调用 scheduler.step()
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer,        # 指定要控制的优化器（这里是Adam）mode='min',       # 监测的指标是"最小化"（如损失函数）patience=3,       # 如果连续3个epoch指标没有改善，才降低LRfactor=0.5        # 降低LR的比例（新LR = 旧LR × 0.5）
)

# 5. 训练模型（记录每个 iteration 的损失）
def train(model, train_loader, test_loader, criterion, optimizer, scheduler, device, epochs):model.train()  # 设置为训练模式# 记录每个 iteration 的损失all_iter_losses = []  # 存储所有 batch 的损失iter_indices = []     # 存储 iteration 序号# 记录每个 epoch 的准确率和损失train_acc_history = []test_acc_history = []train_loss_history = []test_loss_history = []# 早停相关参数best_test_acc = 0.0patience = 5  # 早停耐心值，5个epochcounter = 0   # 计数器，记录连续未改进的epoch数early_stop = False  # 早停标志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)  # 移至GPUoptimizer.zero_grad()  # 梯度清零output = model(data)  # 前向传播loss = criterion(output, target)  # 计算损失loss.backward()  # 反向传播optimizer.step()  # 更新参数# 记录当前 iteration 的损失iter_loss = loss.item()all_iter_losses.append(iter_loss)iter_indices.append(epoch * len(train_loader) + batch_idx + 1)# 统计准确率和损失running_loss += iter_loss_, predicted = output.max(1)total += target.size(0)correct += predicted.eq(target).sum().item()# 每100个批次打印一次训练信息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}')# 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 / totaltrain_acc_history.append(epoch_train_acc)train_loss_history.append(epoch_train_loss)# 测试阶段model.eval()  # 设置为评估模式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)# 更新学习率调度器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}%')# 早停检查if epoch_test_acc > best_test_acc:best_test_acc = epoch_test_acccounter = 0# 保存最佳模型（可选）torch.save(model.state_dict(), 'best_model.pth')print(f"找到更好的模型，准确率: {best_test_acc:.2f}%，已保存")else:counter += 1print(f"早停计数器: {counter}/{patience}")if counter >= patience:print(f"早停触发！连续 {patience} 个epoch测试准确率未提高")early_stop = True# 如果触发早停，跳出训练循环if early_stop:print(f"训练在第 {epoch+1} 个epoch提前结束")break# 绘制所有 iteration 的损失曲线plot_iter_losses(all_iter_losses, iter_indices)# 绘制每个 epoch 的准确率和损失曲线plot_epoch_metrics(train_acc_history, test_acc_history, train_loss_history, test_loss_history)return epoch_test_acc  # 返回最终测试准确率# 6. 绘制每个 iteration 的损失曲线
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()# 7. 绘制每个 epoch 的准确率和损失曲线
def plot_epoch_metrics(train_acc, test_acc, train_loss, test_loss):epochs = range(1, len(train_acc) + 1)plt.figure(figsize=(12, 4))# 绘制准确率曲线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)# 绘制损失曲线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()# 8. 执行训练和测试
epochs = 40  # 增加训练轮次以获得更好效果
print("开始使用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_cnn_model.pth')
# print("模型已保存为: cifar10_cnn_model.pth")

训练完成！最终测试准确率: 87.04%

@浙大疏精行