增加交叉驗證和超參數調優

前文中，只是給了基礎模型：?

PyTorch 實現 CIFAR-10 圖像分類：從數據預處理到模型訓練與評估-CSDN博客

今天我們增加交叉驗證和超參數調優，

先看運行結果：
===== 在測試集上評估最終模型 =====
最終模型在測試集上的準確率：60.14%
最優模型已保存為 'cifar10_best_model.pth'（超參數：{'batch_size': 32, 'epochs': 5, 'lr': 0.01, 'momentum': 0.85}）

Process finished with exit code 0
比基礎模型準確率高了一點，

?完整代碼如下：

import torch
import torch.nn as nn
import torch.optim as optim
from torchvision import datasets, transforms
from torch.utils.data import DataLoader, Subset
import matplotlib.pyplot as plt
import numpy as np
import torchvision
from sklearn.model_selection import KFold, ParameterGrid  # 用于交叉驗證和超參數網格搜索# --------------------------
# 1. 數據準備（與原代碼一致，但后續會在訓練集內部做交叉驗證）
# --------------------------
# 數據預處理：標準化（與原代碼相同）
transform = transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])# 數據集路徑（請替換為你的實際路徑）
data_path = r'D:\workspace_py\deeplean\data'# 加載完整訓練集和測試集（測試集始終不變，用于最終評估）
full_trainset = datasets.CIFAR10(root=data_path, train=True, download=False, transform=transform)
testset = datasets.CIFAR10(root=data_path, train=False, download=False, transform=transform)
classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck')# --------------------------
# 2. 定義CNN模型（與原代碼一致）
# --------------------------
class Net(nn.Module):def __init__(self):super(Net, self).__init__()self.conv1 = nn.Conv2d(3, 6, 5)self.pool = nn.MaxPool2d(2, 2)self.conv2 = nn.Conv2d(6, 16, 5)self.fc1 = nn.Linear(16 * 5 * 5, 120)self.fc2 = nn.Linear(120, 84)self.fc3 = nn.Linear(84, 10)def forward(self, x):x = self.pool(torch.relu(self.conv1(x)))x = self.pool(torch.relu(self.conv2(x)))x = x.view(-1, 16 * 5 * 5)x = torch.relu(self.fc1(x))x = torch.relu(self.fc2(x))x = self.fc3(x)return x# --------------------------
# 3. 交叉驗證函數（核心新增）
# --------------------------
def cross_validate(model, train_dataset, k_folds=5, epochs=5, lr=0.001, batch_size=32, momentum=0.9):"""5折交叉驗證：將訓練集分成5份，每次用4份訓練，1份驗證，返回平均準確率"""kfold = KFold(n_splits=k_folds, shuffle=True, random_state=42)  # 固定隨機種子，結果可復現fold_results = []  # 存儲每折的驗證準確率for fold, (train_ids, val_ids) in enumerate(kfold.split(train_dataset)):print(f'\n===== 第 {fold + 1}/{k_folds} 折交叉驗證 =====')# 1. 劃分當前折的訓練集和驗證集train_subset = Subset(train_dataset, train_ids)  # 本次訓練用的數據val_subset = Subset(train_dataset, val_ids)  # 本次驗證用的數據# 2. 創建數據加載器train_loader = DataLoader(train_subset, batch_size=batch_size, shuffle=True)val_loader = DataLoader(val_subset, batch_size=batch_size, shuffle=False)# 3. 初始化模型和優化器（每折都重新訓練新模型，避免干擾）model_instance = Net()  # 重新實例化模型criterion = nn.CrossEntropyLoss()optimizer = optim.SGD(model_instance.parameters(), lr=lr, momentum=momentum)# 4. 訓練當前折的模型for epoch in range(epochs):model_instance.train()  # 訓練模式running_loss = 0.0for i, data in enumerate(train_loader, 0):inputs, labels = dataoptimizer.zero_grad()outputs = model_instance(inputs)loss = criterion(outputs, labels)loss.backward()optimizer.step()running_loss += loss.item()# 每200步打印一次損失（簡化輸出）if i % 200 == 199:print(f'折 {fold + 1}，輪次 {epoch + 1}，第 {i + 1} 步：平均損失 {running_loss / 200:.3f}')running_loss = 0.0# 5. 在驗證集上評估當前折的模型model_instance.eval()  # 驗證模式correct = 0total = 0with torch.no_grad():for data in val_loader:images, labels = dataoutputs = model_instance(images)_, predicted = torch.max(outputs.data, 1)total += labels.size(0)correct += (predicted == labels).sum().item()val_acc = 100 * correct / totalprint(f'第 {fold + 1} 折驗證準確率：{val_acc:.2f}%')fold_results.append(val_acc)# 計算所有折的平均準確率（該超參數組合的最終得分）avg_acc = sum(fold_results) / len(fold_results)print(f'\n===== 該超參數組合的平均驗證準確率：{avg_acc:.2f}% =====')return avg_acc# --------------------------
# 4. 超參數調優（核心新增）
# --------------------------
def hyperparameter_tuning(train_dataset):"""超參數網格搜索：嘗試不同的超參數組合，用交叉驗證選最優"""# 定義要測試的超參數組合（可根據需要增減）param_grid = {'lr': [0.001, 0.01],  # 學習率：嘗試兩個值'batch_size': [32, 64],  # 批大小：嘗試兩個值'momentum': [0.9, 0.85],  # 動量：嘗試兩個值'epochs': [5]  # 訓練輪次（固定為5，減少計算量）}best_acc = 0.0best_params = None  # 存儲最優超參數# 遍歷所有超參數組合（共 2×2×2=8 種組合）for params in ParameterGrid(param_grid):print(f'\n---------- 測試超參數組合：{params} ----------')# 用交叉驗證評估當前組合的性能current_acc = cross_validate(model=Net(),train_dataset=train_dataset,k_folds=5,epochs=params['epochs'],lr=params['lr'],batch_size=params['batch_size'],momentum=params['momentum'])# 記錄最優組合if current_acc > best_acc:best_acc = current_accbest_params = paramsprint(f'★ 發現更優組合！當前最優準確率：{best_acc:.2f}%')print(f'\n===== 超參數調優完成 =====')print(f'最優超參數：{best_params}')print(f'最優平均驗證準確率：{best_acc:.2f}%')return best_params# --------------------------
# 5. 主函數：執行超參數調優 + 最終訓練 + 測試集評估
# --------------------------
if __name__ == '__main__':# 步驟1：超參數調優（用交叉驗證選最優參數）print('===== 開始超參數調優（這一步比較慢，需要耐心等待）=====')best_params = hyperparameter_tuning(full_trainset)# 步驟2：用最優超參數在完整訓練集上訓練最終模型print('\n===== 用最優超參數訓練最終模型 =====')final_model = Net()criterion = nn.CrossEntropyLoss()optimizer = optim.SGD(final_model.parameters(),lr=best_params['lr'],momentum=best_params['momentum'])train_loader = DataLoader(full_trainset,batch_size=best_params['batch_size'],shuffle=True)# 訓練最終模型（輪次與調優時一致）for epoch in range(best_params['epochs']):final_model.train()running_loss = 0.0for i, data in enumerate(train_loader, 0):inputs, labels = dataoptimizer.zero_grad()outputs = final_model(inputs)loss = criterion(outputs, labels)loss.backward()optimizer.step()running_loss += loss.item()if i % 200 == 199:print(f'最終模型訓練 - 輪次 {epoch + 1}，第 {i + 1} 步：平均損失 {running_loss / 200:.3f}')running_loss = 0.0# 步驟3：在測試集上評估最終模型（用從未見過的測試數據）print('\n===== 在測試集上評估最終模型 =====')final_model.eval()test_loader = DataLoader(testset, batch_size=32, shuffle=False)correct = 0total = 0with torch.no_grad():for data in test_loader:images, labels = dataoutputs = final_model(images)_, predicted = torch.max(outputs.data, 1)total += labels.size(0)correct += (predicted == labels).sum().item()test_acc = 100 * correct / totalprint(f'最終模型在測試集上的準確率：{test_acc:.2f}%')# 步驟4：保存最優模型torch.save(final_model.state_dict(), 'cifar10_best_model.pth')print(f"最優模型已保存為 'cifar10_best_model.pth'（超參數：{best_params}）")