Python训练营---DAY52
DAY 52 神经网络调参指南
知识点回顾：
- 随机种子
- 内参的初始化
- 神经网络调参指南
- 参数的分类
- 调参的顺序
- 各部分参数的调整心得
作业：对于day'41的简单cnn，看看是否可以借助调参指南进一步提高精度。
对于day41的CNN进行调参：
早停类：
# 早停类
class EarlyStopping:"""如果验证损失在给定轮数内没有改善，则提前停止训练"""def __init__(self, patience=7, verbose=False, delta=0):"""参数:patience (int): 在最后一次验证损失改善后等待的轮数默认: 7verbose (bool): 如果为True，每次验证损失改善时打印消息默认: Falsedelta (float): 监控指标的最小变化，以视为改善默认: 0"""self.patience = patienceself.verbose = verboseself.counter = 0 # 记录验证损失未改善的连续轮数self.best_score = None # 记录历史最佳分数（通常是负的验证损失，因为我们希望最大化这个值）self.early_stop = False # 触发早停的标志self.val_loss_min = np.Inf # 记录历史最低验证损失self.delta = delta # 改善阈值def __call__(self, val_loss, model):score = val_lossif self.best_score is None:self.best_score = scoreself.save_checkpoint(val_loss, model)elif score < self.best_score + self.delta:self.counter += 1print(f'早停计数器: {self.counter} / {self.patience}')if self.counter >= self.patience:self.early_stop = Trueelse:self.best_score = scoreself.save_checkpoint(val_loss, model)self.counter = 0def save_checkpoint(self, val_loss, model):'''当验证损失降低时保存模型'''if self.verbose:print(f'测试集损失降低 ({self.val_loss_min:.6f} --> {val_loss:.6f})')# torch.save(model.state_dict(), 'checkpoint.pth')self.val_loss_min = val_loss测试最大允许的 batch size 函数
def find_max_batch_size(train_dataset, device, model):max_batch_size = 512try:train_loader = DataLoader(train_dataset, batch_size=max_batch_size, shuffle=True)for data, target in train_loader:data, target = data.to(device), target.to(device)output = model(data)breakexcept RuntimeError as e:print(f"最大 batch size 导致显存不足，尝试减小")while True:max_batch_size //= 2try:train_loader = DataLoader(train_dataset, batch_size=max_batch_size, shuffle=True)for data, target in train_loader:data, target = data.to(device), target.to(device)output = model(data)print(f"最大允许的 batch size 为 {max_batch_size}")breakexcept RuntimeError:continuereturn max_batch_size配置优化器和学习率调度器函数
# 配置优化器和学习率调度器函数
def configure_optimizer_scheduler(model):criterion = nn.CrossEntropyLoss()# 设置 weight_decay 参数添加 L2 正则化optimizer = optim.Adam(model.parameters(), lr=0.001, weight_decay=0.0001) scheduler = ReduceLROnPlateau(optimizer, 'min', patience=3, factor=0.5)return criterion, optimizer, scheduler完整代码：
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
from torch.optim.lr_scheduler import ReduceLROnPlateau # 添加这行# 设置替代中文字体（适用于Linux）
plt.rcParams["font.family"] = ["WenQuanYi Micro Hei", "sans-serif"]
plt.rcParams['axes.unicode_minus'] = False# 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='./cifar_data/cifar_data',train=True,download=True,transform=train_transform # 使用增强后的预处理
)test_dataset = datasets.CIFAR10(root='./cifar_data/cifar_data',train=False,transform=test_transform # 测试集不使用增强
)# 4. 定义CNN模型的定义（替代原MLP）
class CNN(nn.Module):def __init__(self):super(CNN, self).__init__() # 继承父类初始化# ---------------------- 第一个卷积块 ----------------------# 卷积层1：输入3通道（RGB），输出32个特征图，卷积核3x3，边缘填充1像素self.conv1 = nn.Conv2d(in_channels=3, # 输入通道数（图像的RGB通道）out_channels=32, # 输出通道数（生成32个新特征图）kernel_size=3, # 卷积核尺寸（3x3像素）padding=1 # 边缘填充1像素，保持输出尺寸与输入相同)# 批量归一化层：对32个输出通道进行归一化，加速训练self.bn1 = nn.BatchNorm2d(num_features=32)# ReLU激活函数：引入非线性，公式：max(0, x)self.relu1 = nn.ReLU()# 最大池化层：窗口2x2，步长2，特征图尺寸减半（32x32→16x16）self.pool1 = nn.MaxPool2d(kernel_size=2, stride=2) # stride默认等于kernel_size# ---------------------- 第二个卷积块 ----------------------# 卷积层2：输入32通道（来自conv1的输出），输出64通道self.conv2 = nn.Conv2d(in_channels=32, # 输入通道数（前一层的输出通道数）out_channels=64, # 输出通道数（特征图数量翻倍）kernel_size=3, # 卷积核尺寸不变padding=1 # 保持尺寸：16x16→16x16（卷积后）→8x8（池化后）)self.bn2 = nn.BatchNorm2d(num_features=64)self.relu2 = nn.ReLU()self.pool2 = nn.MaxPool2d(kernel_size=2) # 尺寸减半：16x16→8x8# ---------------------- 第三个卷积块 ----------------------# 卷积层3：输入64通道，输出128通道self.conv3 = nn.Conv2d(in_channels=64, # 输入通道数（前一层的输出通道数）out_channels=128, # 输出通道数（特征图数量再次翻倍）kernel_size=3,padding=1 # 保持尺寸：8x8→8x8（卷积后）→4x4（池化后）)self.bn3 = nn.BatchNorm2d(num_features=128)self.relu3 = nn.ReLU() # 复用激活函数对象（节省内存）self.pool3 = nn.MaxPool2d(kernel_size=2) # 尺寸减半：8x8→4x4# ---------------------- 全连接层（分类器） ----------------------# 计算展平后的特征维度：128通道 × 4x4尺寸 = 128×16=2048维self.fc1 = nn.Linear(in_features=128 * 4 * 4, # 输入维度（卷积层输出的特征数）out_features=512 # 输出维度（隐藏层神经元数）)# Dropout层：训练时随机丢弃50%神经元，防止过拟合self.dropout = nn.Dropout(p=0.5)# 输出层：将512维特征映射到10个类别（CIFAR-10的类别数）self.fc2 = nn.Linear(in_features=512, out_features=10)def forward(self, x):# 输入尺寸：[batch_size, 3, 32, 32]（batch_size=批量大小，3=通道数，32x32=图像尺寸）# ---------- 卷积块1处理 ----------x = self.conv1(x) # 卷积后尺寸：[batch_size, 32, 32, 32]（padding=1保持尺寸）x = self.bn1(x) # 批量归一化，不改变尺寸x = self.relu1(x) # 激活函数，不改变尺寸x = self.pool1(x) # 池化后尺寸：[batch_size, 32, 16, 16]（32→16是因为池化窗口2x2）# ---------- 卷积块2处理 ----------x = self.conv2(x) # 卷积后尺寸：[batch_size, 64, 16, 16]（padding=1保持尺寸）x = self.bn2(x)x = self.relu2(x)x = self.pool2(x) # 池化后尺寸：[batch_size, 64, 8, 8]# ---------- 卷积块3处理 ----------x = self.conv3(x) # 卷积后尺寸：[batch_size, 128, 8, 8]（padding=1保持尺寸）x = self.bn3(x)x = self.relu3(x)x = self.pool3(x) # 池化后尺寸：[batch_size, 128, 4, 4]# ---------- 展平与全连接层 ----------# 将多维特征图展平为一维向量：[batch_size, 128*4*4] = [batch_size, 2048]x = x.view(-1, 128 * 4 * 4) # -1自动计算批量维度，保持批量大小不变x = self.fc1(x) # 全连接层：2048→512，尺寸变为[batch_size, 512]x = self.relu3(x) # 激活函数（复用relu3，与卷积块3共用）x = self.dropout(x) # Dropout随机丢弃神经元，不改变尺寸x = self.fc2(x) # 全连接层：512→10，尺寸变为[batch_size, 10]（未激活，直接输出logits）return x # 输出未经过Softmax的logits，适用于交叉熵损失函数# 早停类
class EarlyStopping:"""如果验证损失在给定轮数内没有改善，则提前停止训练"""def __init__(self, patience=7, verbose=False, delta=0):"""参数:patience (int): 在最后一次验证损失改善后等待的轮数默认: 7verbose (bool): 如果为True，每次验证损失改善时打印消息默认: Falsedelta (float): 监控指标的最小变化，以视为改善默认: 0"""self.patience = patienceself.verbose = verboseself.counter = 0 # 记录验证损失未改善的连续轮数self.best_score = None # 记录历史最佳分数（通常是负的验证损失，因为我们希望最大化这个值）self.early_stop = False # 触发早停的标志self.val_loss_min = np.Inf # 记录历史最低验证损失self.delta = delta # 改善阈值def __call__(self, val_loss, model):score = val_lossif self.best_score is None:self.best_score = scoreself.save_checkpoint(val_loss, model)elif score < self.best_score + self.delta:self.counter += 1print(f'早停计数器: {self.counter} / {self.patience}')if self.counter >= self.patience:self.early_stop = Trueelse:self.best_score = scoreself.save_checkpoint(val_loss, model)self.counter = 0def save_checkpoint(self, val_loss, model):'''当验证损失降低时保存模型'''if self.verbose:print(f'测试集损失降低 ({self.val_loss_min:.6f} --> {val_loss:.6f})')# torch.save(model.state_dict(), 'checkpoint.pth')self.val_loss_min = val_loss# 2. 测试最大允许的 batch size 函数
def find_max_batch_size(train_dataset, device, model):max_batch_size = 512try:train_loader = DataLoader(train_dataset, batch_size=max_batch_size, shuffle=True)for data, target in train_loader:data, target = data.to(device), target.to(device)output = model(data)breakexcept RuntimeError as e:print(f"最大 batch size 导致显存不足，尝试减小")while True:max_batch_size //= 2try:train_loader = DataLoader(train_dataset, batch_size=max_batch_size, shuffle=True)for data, target in train_loader:data, target = data.to(device), target.to(device)output = model(data)print(f"最大允许的 batch size 为 {max_batch_size}")breakexcept RuntimeError:continuereturn max_batch_size# 配置优化器和学习率调度器函数
def configure_optimizer_scheduler(model):criterion = nn.CrossEntropyLoss()# 设置 weight_decay 参数添加 L2 正则化optimizer = optim.Adam(model.parameters(), lr=0.001, weight_decay=0.0001) scheduler = ReduceLROnPlateau(optimizer, 'min', patience=3, factor=0.5)return criterion, optimizer, scheduler# 初始化模型
model = CNN()
model = model.to(device) # 将模型移至GPU（如果可用）# 5. 训练模型（记录每个 iteration 的损失）
def train(model, train_dataset, test_dataset, device, epochs):model.train() # 设置为训练模式max_batch_size = find_max_batch_size(train_dataset, device, model)criterion, optimizer, scheduler = configure_optimizer_scheduler(model)early_stopping = EarlyStopping(patience=5, verbose=True)train_loader = DataLoader(train_dataset, batch_size=max_batch_size, shuffle=True)test_loader = DataLoader(test_dataset, batch_size=max_batch_size, shuffle=False)# 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 = []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)# 早停检查early_stopping(epoch_test_acc, model)if early_stopping.early_stop:print("早停触发，停止训练。")breakprint(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}%')# 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 = 30 # 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_dataset, test_dataset, device, epochs)
print(f"训练完成！最终测试准确率: {final_accuracy:.2f}%")
测试集准确率最好为80.44%，跟day41相差不是很大