DAY 44 預訓練模型

知識點回顧：

1. 預訓練的概念
2. 常見的分類預訓練模型
3. 圖像預訓練模型的發展史
4. 預訓練的策略
5. 預訓練代碼實戰：resnet18

作業：

1. 嘗試在cifar10對比如下其他的預訓練模型，觀察差異，盡可能和他人選擇的不同
2. 嘗試通過ctrl進入resnet的內部，觀察殘差究竟是什么

選用?DenseNet121預訓練模型，注意DenseNet121 模型的最后分類層名為classifier，而不是 ResNet 中的fc

import torch
import torch.nn as nn
import torch.optim as optim
from torchvision import datasets, transforms, models
from torch.utils.data import DataLoader
import matplotlib.pyplot as plt
import os
from torchvision.models import resnet18, densenet121, vgg16# 設置中文字體支持
plt.rcParams["font.family"] = ["SimHei"]
plt.rcParams['axes.unicode_minus'] = False  # 解決負號顯示問題# 檢查GPU是否可用
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"使用設備: {device}")# 1. 數據預處理（訓練集增強，測試集標準化）
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))
])# 2. 加載CIFAR-10數據集
train_dataset = datasets.CIFAR10(root='./cifar_data',train=True,download=True,transform=train_transform
)test_dataset = datasets.CIFAR10(root='./cifar_data',train=False,transform=test_transform
)# 3. 創建數據加載器
batch_size = 64
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
test_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False)# 4. 定義DenseNet121模型
def create_densenet121(pretrained=True, num_classes=10):model = models.densenet121(pretrained=pretrained)# 修改最后一層全連接層in_features = model.classifier.in_featuresmodel.classifier = nn.Linear(in_features, num_classes) # DenseNet121 的最后一層分類器名稱是classifierreturn model.to(device)# 5. 凍結/解凍模型層的函數
# 這種設計允許我們在遷移學習中保留預訓練模型的特征提取部分（卷積層），只訓練新添加的分類層（全連接層）。
def freeze_model(model, freeze=True):"""凍結或解凍模型的卷積層參數"""# 凍結/解凍除fc層外的所有參數for name, param in model.named_parameters():if 'classifier' not in name:    #排除名稱中包含 "fc" 的參數，這些通常是全連接層的參數param.requires_grad = not freeze    #param.requires_grad是 PyTorch 中控制參數是否參與反向傳播和梯度更新的標志# 打印凍結狀態frozen_params = sum(p.numel() for p in model.parameters() if not p.requires_grad)   #統計所有requires_grad=False的參數數量total_params = sum(p.numel() for p in model.parameters())if freeze:print(f"已凍結模型卷積層參數 ({frozen_params}/{total_params} 參數)")else:print(f"已解凍模型所有參數 ({total_params}/{total_params} 參數可訓練)")return model# 6. 訓練函數（支持階段式訓練）
def train_with_freeze_schedule(model, train_loader, test_loader, criterion, optimizer, scheduler, device, epochs, freeze_epochs=5):"""前freeze_epochs輪凍結卷積層，之后解凍所有層進行訓練"""train_loss_history = []test_loss_history = []train_acc_history = []test_acc_history = []all_iter_losses = []iter_indices = []# 初始凍結卷積層if freeze_epochs > 0:model = freeze_model(model, freeze=True)for epoch in range(epochs):# 解凍控制：在指定輪次后解凍所有層if epoch == freeze_epochs:model = freeze_model(model, freeze=False)# 解凍后調整優化器（可選）optimizer.param_groups[0]['lr'] = 1e-4  # 降低學習率防止過擬合model.train()  # 設置為訓練模式running_loss = 0.0correct_train = 0total_train = 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()# 記錄Iteration損失iter_loss = loss.item()all_iter_losses.append(iter_loss)iter_indices.append(epoch * len(train_loader) + batch_idx + 1)# 統計訓練指標running_loss += iter_loss_, predicted = output.max(1)total_train += target.size(0)correct_train += predicted.eq(target).sum().item()# 每100批次打印進度if (batch_idx + 1) % 100 == 0:print(f"Epoch {epoch+1}/{epochs} | Batch {batch_idx+1}/{len(train_loader)} "f"| 單Batch損失: {iter_loss:.4f}")# 計算 epoch 級指標epoch_train_loss = running_loss / len(train_loader)epoch_train_acc = 100. * correct_train / total_train# 測試階段model.eval()correct_test = 0total_test = 0test_loss = 0.0with 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_test# 記錄歷史數據train_loss_history.append(epoch_train_loss)test_loss_history.append(epoch_test_loss)train_acc_history.append(epoch_train_acc)test_acc_history.append(epoch_test_acc)# 更新學習率調度器if scheduler is not None:scheduler.step(epoch_test_loss)# 打印 epoch 結果print(f"Epoch {epoch+1} 完成 | 訓練損失: {epoch_train_loss:.4f} "f"| 訓練準確率: {epoch_train_acc:.2f}% | 測試準確率: {epoch_test_acc:.2f}%")# 繪制損失和準確率曲線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  # 返回最終測試準確率# 7. 繪制Iteration損失曲線
def plot_iter_losses(losses, indices):plt.figure(figsize=(10, 4))plt.plot(indices, losses, 'b-', alpha=0.7)plt.xlabel('Iteration（Batch序號）')plt.ylabel('損失值')plt.title('訓練過程中的Iteration損失變化')plt.grid(True)plt.show()# 8. 繪制Epoch級指標曲線
def plot_epoch_metrics(train_acc, test_acc, train_loss, test_loss):epochs = range(1, len(train_acc) + 1)plt.figure(figsize=(12, 5))# 準確率曲線plt.subplot(1, 2, 1)plt.plot(epochs, train_acc, 'b-', label='訓練準確率')plt.plot(epochs, test_acc, 'r-', label='測試準確率')plt.xlabel('Epoch')plt.ylabel('準確率 (%)')plt.title('準確率隨Epoch變化')plt.legend()plt.grid(True)# 損失曲線plt.subplot(1, 2, 2)plt.plot(epochs, train_loss, 'b-', label='訓練損失')plt.plot(epochs, test_loss, 'r-', label='測試損失')plt.xlabel('Epoch')plt.ylabel('損失值')plt.title('損失值隨Epoch變化')plt.legend()plt.grid(True)plt.tight_layout()plt.show()# 主函數：訓練模型
def main():# 參數設置epochs = 40  # 總訓練輪次freeze_epochs = 5  # 凍結卷積層的輪次learning_rate = 1e-3  # 初始學習率weight_decay = 1e-4  # 權重衰減# 創建DenseNet121模型（加載預訓練權重）model = create_densenet121(pretrained=True, num_classes=10)# 定義優化器和損失函數optimizer = optim.Adam(model.parameters(), lr=learning_rate, weight_decay=weight_decay)criterion = nn.CrossEntropyLoss()# 定義學習率調度器scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=0.5, patience=2, verbose=True)# 開始訓練（前5輪凍結卷積層，之后解凍）final_accuracy = train_with_freeze_schedule(model=model,train_loader=train_loader,test_loader=test_loader,criterion=criterion,optimizer=optimizer,scheduler=scheduler,device=device,epochs=epochs,freeze_epochs=freeze_epochs)print(f"訓練完成！最終測試準確率: {final_accuracy:.2f}%")# # 保存模型# torch.save(model.state_dict(), 'resnet18_cifar10_finetuned.pth')# print("模型已保存至: resnet18_cifar10_finetuned.pth")if __name__ == "__main__":main()