DAY 41 簡單CNN
知識回顧
- 數據增強
- 卷積神經網絡定義的寫法
- batch歸一化：調整一個批次的分布，常用與圖像數據
- 特征圖：只有卷積操作輸出的才叫特征圖
- 調度器：直接修改基礎學習率
卷積操作常見流程如下：
1. 輸入 → 卷積層 → Batch歸一化層（可選） → 池化層 → 激活函數 → 下一層
- Flatten -> Dense (with Dropout，可選) -> Dense (Output)
這里相關的概念比較多，如果之前沒有學習過復試班強化班中的計算機視覺部分，請自行上網檢索視頻了解下基礎概念，也可以對照我提供的之前的講義學習下。
計算機視覺入門
作業：嘗試手動修改下不同的調度器和CNN的結構，觀察訓練的差異。
"""
DAY 41 實驗：比較不同的調度器和CNN結構本文件提供了多種CNN結構和學習率調度器的實現，
用于比較不同配置下的訓練效果。
"""import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torchvision
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
import numpy as np
from torch.utils.data import DataLoader
from torch.optim.lr_scheduler import StepLR, ReduceLROnPlateau, CosineAnnealingLR, OneCycleLR# 設置中文字體
plt.rcParams['font.sans-serif'] = ['SimHei']
plt.rcParams['axes.unicode_minus'] = False# 設置隨機種子
torch.manual_seed(42)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")# CIFAR-10數據集的類別
classes = ('飛機', '汽車', '鳥', '貓', '鹿', '狗', '青蛙', '馬', '船', '卡車')#====================== 1. 數據加載與增強 ======================def load_data(batch_size=64, is_train=True):"""加載CIFAR-10數據集，并應用數據增強"""if is_train:# 訓練集使用數據增強transform = transforms.Compose([transforms.RandomHorizontalFlip(),transforms.RandomRotation(10),transforms.RandomAffine(0, shear=10, scale=(0.8, 1.2)),transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2),transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])else:# 測試集只需要標準化transform = transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])dataset = torchvision.datasets.CIFAR10(root='./data', train=is_train,download=True,transform=transform)dataloader = DataLoader(dataset,batch_size=batch_size,shuffle=is_train,num_workers=2)return dataloader#====================== 2. 不同的CNN模型結構 ======================class BasicCNN(nn.Module):"""基礎CNN模型：3個卷積層"""def __init__(self, use_bn=True, dropout_rate=0.5):super(BasicCNN, self).__init__()self.use_bn = use_bn# 第一個卷積塊self.conv1 = nn.Conv2d(3, 32, kernel_size=3, padding=1)self.bn1 = nn.BatchNorm2d(32) if use_bn else nn.Identity()# 第二個卷積塊self.conv2 = nn.Conv2d(32, 64, kernel_size=3, padding=1)self.bn2 = nn.BatchNorm2d(64) if use_bn else nn.Identity()# 第三個卷積塊self.conv3 = nn.Conv2d(64, 128, kernel_size=3, padding=1)self.bn3 = nn.BatchNorm2d(128) if use_bn else nn.Identity()# 全連接層self.fc1 = nn.Linear(128 * 4 * 4, 512)self.dropout = nn.Dropout(dropout_rate)self.fc2 = nn.Linear(512, 10)def forward(self, x):# 第一個卷積塊x = self.conv1(x)x = self.bn1(x)x = F.relu(x)x = F.max_pool2d(x, 2)# 第二個卷積塊x = self.conv2(x)x = self.bn2(x)x = F.relu(x)x = F.max_pool2d(x, 2)# 第三個卷積塊x = self.conv3(x)x = self.bn3(x)x = F.relu(x)x = F.max_pool2d(x, 2)# 全連接層x = torch.flatten(x, 1)x = self.fc1(x)x = F.relu(x)x = self.dropout(x)x = self.fc2(x)return F.log_softmax(x, dim=1)class DeepCNN(nn.Module):"""更深的CNN模型：5個卷積層"""def __init__(self, use_bn=True, dropout_rate=0.5):super(DeepCNN, self).__init__()self.use_bn = use_bn# 第一個卷積塊self.conv1 = nn.Conv2d(3, 32, kernel_size=3, padding=1)self.bn1 = nn.BatchNorm2d(32) if use_bn else nn.Identity()# 第二個卷積塊self.conv2 = nn.Conv2d(32, 64, kernel_size=3, padding=1)self.bn2 = nn.BatchNorm2d(64) if use_bn else nn.Identity()# 第三個卷積塊self.conv3 = nn.Conv2d(64, 128, kernel_size=3, padding=1)self.bn3 = nn.BatchNorm2d(128) if use_bn else nn.Identity()# 第四個卷積塊self.conv4 = nn.Conv2d(128, 256, kernel_size=3, padding=1)self.bn4 = nn.BatchNorm2d(256) if use_bn else nn.Identity()# 第五個卷積塊self.conv5 = nn.Conv2d(256, 512, kernel_size=3, padding=1)self.bn5 = nn.BatchNorm2d(512) if use_bn else nn.Identity()# 全連接層self.fc1 = nn.Linear(512, 512)self.dropout = nn.Dropout(dropout_rate)self.fc2 = nn.Linear(512, 10)def forward(self, x):# 第一個卷積塊x = self.conv1(x)x = self.bn1(x)x = F.relu(x)x = F.max_pool2d(x, 2)# 第二個卷積塊x = self.conv2(x)x = self.bn2(x)x = F.relu(x)x = F.max_pool2d(x, 2)# 第三個卷積塊x = self.conv3(x)x = self.bn3(x)x = F.relu(x)# 第四個卷積塊x = self.conv4(x)x = self.bn4(x)x = F.relu(x)x = F.max_pool2d(x, 2)# 第五個卷積塊x = self.conv5(x)x = self.bn5(x)x = F.relu(x)x = F.adaptive_avg_pool2d(x, (1, 1))# 全連接層x = torch.flatten(x, 1)x = self.fc1(x)x = F.relu(x)x = self.dropout(x)x = self.fc2(x)return F.log_softmax(x, dim=1)class ResidualBlock(nn.Module):"""殘差塊"""def __init__(self, in_channels, out_channels, stride=1):super(ResidualBlock, self).__init__()self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=stride, padding=1, bias=False)self.bn1 = nn.BatchNorm2d(out_channels)self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False)self.bn2 = nn.BatchNorm2d(out_channels)self.shortcut = nn.Sequential()if stride != 1 or in_channels != out_channels:self.shortcut = nn.Sequential(nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=stride, bias=False),nn.BatchNorm2d(out_channels))def forward(self, x):out = F.relu(self.bn1(self.conv1(x)))out = self.bn2(self.conv2(out))out += self.shortcut(x)out = F.relu(out)return outclass ResNetCNN(nn.Module):"""帶有殘差連接的CNN模型"""def __init__(self, dropout_rate=0.5):super(ResNetCNN, self).__init__()self.conv1 = nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False)self.bn1 = nn.BatchNorm2d(64)# 殘差塊self.layer1 = self._make_layer(64, 64, 2, stride=1)self.layer2 = self._make_layer(64, 128, 2, stride=2)self.layer3 = self._make_layer(128, 256, 2, stride=2)# 全連接層（修正輸入維度：256 * 8 * 8）self.fc = nn.Linear(256 * 8 * 8, 10)self.dropout = nn.Dropout(dropout_rate)def _make_layer(self, in_channels, out_channels, num_blocks, stride):layers = []layers.append(ResidualBlock(in_channels, out_channels, stride))for _ in range(1, num_blocks):layers.append(ResidualBlock(out_channels, out_channels, 1))return nn.Sequential(*layers)def forward(self, x):out = F.relu(self.bn1(self.conv1(x))) # 32x32out = self.layer1(out) # 32x32out = self.layer2(out) # 16x16out = self.layer3(out) # 8x8out = torch.flatten(out, 1) # 256*8*8out = self.dropout(out)out = self.fc(out)return F.log_softmax(out, dim=1)#====================== 3. 訓練函數 ======================def train(model, train_loader, optimizer, scheduler, epoch, history):"""訓練一個epoch"""model.train()train_loss = 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 = F.nll_loss(output, target)loss.backward()optimizer.step()train_loss += loss.item()pred = output.max(1, keepdim=True)[1]correct += pred.eq(target.view_as(pred)).sum().item()total += target.size(0)if batch_idx % 100 == 0:print(f'Train Epoch: {epoch} [{batch_idx * len(data)}/{len(train_loader.dataset)} 'f'({100. * batch_idx / len(train_loader):.0f}%)]\t'f'Loss: {loss.item():.6f}\t'f'Accuracy: {100. * correct / total:.2f}%')# 如果使用ReduceLROnPlateau，需要在每個epoch結束時更新if isinstance(scheduler, optim.lr_scheduler.ReduceLROnPlateau):scheduler.step(train_loss)elif isinstance(scheduler, optim.lr_scheduler.OneCycleLR):pass # OneCycleLR在每個batch后更新，不在epoch結束時更新else:scheduler.step()epoch_loss = train_loss / len(train_loader)epoch_acc = 100. * correct / totalhistory['train_loss'].append(epoch_loss)history['train_acc'].append(epoch_acc)return epoch_loss, epoch_acc#====================== 4. 測試函數 ======================def test(model, test_loader, history):"""在測試集上評估模型"""model.eval()test_loss = 0correct = 0with torch.no_grad():for data, target in test_loader:data, target = data.to(device), target.to(device)output = model(data)test_loss += F.nll_loss(output, target, reduction='sum').item()pred = output.max(1, keepdim=True)[1]correct += pred.eq(target.view_as(pred)).sum().item()test_loss /= len(test_loader.dataset)accuracy = 100. * correct / len(test_loader.dataset)history['test_loss'].append(test_loss)history['test_acc'].append(accuracy)print(f'\nTest set: Average loss: {test_loss:.4f}, 'f'Accuracy: {correct}/{len(test_loader.dataset)} 'f'({accuracy:.2f}%)\n')return test_loss, accuracy#====================== 5. 可視化函數 ======================def plot_training_history(history, title):"""繪制訓練歷史曲線"""epochs = range(1, len(history['train_loss']) + 1)plt.figure(figsize=(12, 4))plt.subplot(1, 2, 1)plt.plot(epochs, history['train_loss'], 'b-', label='訓練損失')plt.plot(epochs, history['test_loss'], 'r-', label='測試損失')plt.title(f'{title} - 訓練和測試損失')plt.xlabel('Epoch')plt.ylabel('損失')plt.legend()plt.grid(True)plt.subplot(1, 2, 2)plt.plot(epochs, history['train_acc'], 'b-', label='訓練準確率')plt.plot(epochs, history['test_acc'], 'r-', label='測試準確率')plt.title(f'{title} - 訓練和測試準確率')plt.xlabel('Epoch')plt.ylabel('準確率 (%)')plt.legend()plt.grid(True)plt.tight_layout()plt.show()def compare_models(histories, titles):"""比較不同模型的訓練歷史"""epochs = range(1, len(histories[0]['train_loss']) + 1)plt.figure(figsize=(12, 8))# 比較訓練損失plt.subplot(2, 2, 1)for i, history in enumerate(histories):plt.plot(epochs, history['train_loss'], label=titles[i])plt.title('訓練損失比較')plt.xlabel('Epoch')plt.ylabel('損失')plt.legend()plt.grid(True)# 比較測試損失plt.subplot(2, 2, 2)for i, history in enumerate(histories):plt.plot(epochs, history['test_loss'], label=titles[i])plt.title('測試損失比較')plt.xlabel('Epoch')plt.ylabel('損失')plt.legend()plt.grid(True)# 比較訓練準確率plt.subplot(2, 2, 3)for i, history in enumerate(histories):plt.plot(epochs, history['train_acc'], label=titles[i])plt.title('訓練準確率比較')plt.xlabel('Epoch')plt.ylabel('準確率 (%)')plt.legend()plt.grid(True)# 比較測試準確率plt.subplot(2, 2, 4)for i, history in enumerate(histories):plt.plot(epochs, history['test_acc'], label=titles[i])plt.title('測試準確率比較')plt.xlabel('Epoch')plt.ylabel('準確率 (%)')plt.legend()plt.grid(True)plt.tight_layout()plt.show()#====================== 6. 實驗函數 ======================def run_experiment(model_type, scheduler_type, epochs=10, batch_size=64, lr=0.01):"""運行一個實驗：訓練指定的模型和調度器"""# 初始化訓練歷史記錄history = {'train_loss': [],'train_acc': [],'test_loss': [],'test_acc': []}# 加載數據print("正在加載訓練集...")train_loader = load_data(batch_size, is_train=True)print("正在加載測試集...")test_loader = load_data(batch_size, is_train=False)# 創建模型if model_type == 'basic':model = BasicCNN(use_bn=True).to(device)model_name = "基礎CNN"elif model_type == 'deep':model = DeepCNN(use_bn=True).to(device)model_name = "深層CNN"elif model_type == 'resnet':model = ResNetCNN().to(device)model_name = "殘差CNN"else:raise ValueError(f"未知的模型類型: {model_type}")# 創建優化器optimizer = optim.SGD(model.parameters(), lr=lr, momentum=0.9, weight_decay=5e-4)# 創建學習率調度器if scheduler_type == 'step':scheduler = StepLR(optimizer, step_size=5, gamma=0.1)scheduler_name = "StepLR"elif scheduler_type == 'plateau':scheduler = ReduceLROnPlateau(optimizer, mode='min', factor=0.1, patience=3)scheduler_name = "ReduceLROnPlateau"elif scheduler_type == 'cosine':scheduler = CosineAnnealingLR(optimizer, T_max=epochs)scheduler_name = "CosineAnnealingLR"elif scheduler_type == 'onecycle':scheduler = OneCycleLR(optimizer, max_lr=lr*10, epochs=epochs, steps_per_epoch=len(train_loader))scheduler_name = "OneCycleLR"else:raise ValueError(f"未知的調度器類型: {scheduler_type}")# 訓練和測試print(f"開始訓練 {model_name} 使用 {scheduler_name}，設備: {device}")for epoch in range(1, epochs + 1):train_loss, train_acc = train(model, train_loader, optimizer, scheduler, epoch, history)test_loss, test_acc = test(model, test_loader, history)# 返回訓練歷史和實驗標題return history, f"{model_name} + {scheduler_name}"#====================== 7. 主函數 ======================def main():# 超參數設置epochs = 10batch_size = 64lr = 0.01# 運行不同的實驗experiments = [# 比較不同的CNN結構（使用相同的調度器）('basic', 'cosine'),('deep', 'cosine'),('resnet', 'cosine'),# 比較不同的調度器（使用相同的CNN結構）# ('basic', 'step'),# ('basic', 'plateau'),# ('basic', 'cosine'),# ('basic', 'onecycle'),]histories = []titles = []for model_type, scheduler_type in experiments:history, title = run_experiment(model_type, scheduler_type, epochs, batch_size, lr)histories.append(history)titles.append(title)# 比較不同實驗的結果compare_models(histories, titles)if __name__ == '__main__':main()"""
實驗說明：1. 模型結構比較
- BasicCNN: 3個卷積層的基礎模型
- DeepCNN: 5個卷積層的深層模型
- ResNetCNN: 帶有殘差連接的模型2. 學習率調度器比較
- StepLR: 按步長降低學習率
- ReduceLROnPlateau: 當指標不再改善時降低學習率
- CosineAnnealingLR: 余弦退火調整學習率
- OneCycleLR: 一個周期的學習率策略3. 如何使用
- 默認比較不同的CNN結構（使用相同的余弦退火調度器）
- 取消注釋相應的代碼可以比較不同的調度器（使用相同的基礎CNN）
- 可以修改epochs、batch_size和lr參數來調整訓練過程4. 觀察重點
- 不同模型的收斂速度
- 最終的測試準確率
- 是否出現過擬合
- 學習率調度器對訓練過程的影響
"""正在加載訓練集...
Files already downloaded and verified
正在加載測試集...
Files already downloaded and verified
開始訓練，使用設備: cpu
Train Epoch: 1 [0/50000 (0%)] Loss: 2.293068 Accuracy: 12.50%
Train Epoch: 1 [6400/50000 (13%)] Loss: 1.712008 Accuracy: 27.32%
Train Epoch: 1 [12800/50000 (26%)] Loss: 1.600459 Accuracy: 32.20%
Train Epoch: 1 [19200/50000 (38%)] Loss: 1.775305 Accuracy: 34.62%
Train Epoch: 1 [25600/50000 (51%)] Loss: 1.768592 Accuracy: 36.12%
Train Epoch: 1 [32000/50000 (64%)] Loss: 1.276903 Accuracy: 37.53%
Train Epoch: 1 [38400/50000 (77%)] Loss: 1.339285 Accuracy: 38.52%
Train Epoch: 1 [44800/50000 (90%)] Loss: 1.675705 Accuracy: 39.45%Test set: Average loss: 1.2459, Accuracy: 5433/10000 (54.33%)Train Epoch: 2 [0/50000 (0%)] Loss: 1.580209 Accuracy: 40.62%
Train Epoch: 2 [6400/50000 (13%)] Loss: 1.495638 Accuracy: 47.74%
Train Epoch: 2 [12800/50000 (26%)] Loss: 1.260183 Accuracy: 48.11%
Train Epoch: 2 [19200/50000 (38%)] Loss: 1.548963 Accuracy: 48.80%
Train Epoch: 2 [25600/50000 (51%)] Loss: 1.085625 Accuracy: 49.33%
Train Epoch: 2 [32000/50000 (64%)] Loss: 1.324473 Accuracy: 49.83%
Train Epoch: 2 [38400/50000 (77%)] Loss: 1.382297 Accuracy: 50.29%
Train Epoch: 2 [44800/50000 (90%)] Loss: 1.364151 Accuracy: 50.72%Test set: Average loss: 1.1879, Accuracy: 5817/10000 (58.17%)Train Epoch: 3 [0/50000 (0%)] Loss: 1.418277 Accuracy: 43.75%
Train Epoch: 3 [6400/50000 (13%)] Loss: 0.966939 Accuracy: 58.01%
Train Epoch: 3 [12800/50000 (26%)] Loss: 1.144377 Accuracy: 58.63%
Train Epoch: 3 [19200/50000 (38%)] Loss: 0.795806 Accuracy: 59.14%
Train Epoch: 3 [25600/50000 (51%)] Loss: 0.949723 Accuracy: 59.52%
Train Epoch: 3 [32000/50000 (64%)] Loss: 1.286763 Accuracy: 59.81%
Train Epoch: 3 [38400/50000 (77%)] Loss: 1.195282 Accuracy: 59.94%
Train Epoch: 3 [44800/50000 (90%)] Loss: 0.957277 Accuracy: 60.14%Test set: Average loss: 0.8982, Accuracy: 6821/10000 (68.21%)Train Epoch: 4 [0/50000 (0%)] Loss: 0.887834 Accuracy: 71.88%
Train Epoch: 4 [6400/50000 (13%)] Loss: 1.282257 Accuracy: 62.07%
Train Epoch: 4 [12800/50000 (26%)] Loss: 0.960115 Accuracy: 62.15%
Train Epoch: 4 [19200/50000 (38%)] Loss: 1.008237 Accuracy: 62.35%
Train Epoch: 4 [25600/50000 (51%)] Loss: 1.156668 Accuracy: 62.07%
Train Epoch: 4 [32000/50000 (64%)] Loss: 0.733899 Accuracy: 62.38%
Train Epoch: 4 [38400/50000 (77%)] Loss: 1.111758 Accuracy: 62.18%
Train Epoch: 4 [44800/50000 (90%)] Loss: 1.084628 Accuracy: 62.32%Test set: Average loss: 0.8728, Accuracy: 6886/10000 (68.86%)訓練完成，繪制訓練歷史...
可視化特征圖...
PS C:\Users\I.Love.I\Desktop\Python_code> & C:/Users/I.Love.I/.conda/envs/DL/python.exe c:/Users/I.Love.I/Desktop/Python_code/python60-days-challenge/1_python-learning-library/Day41_simple.py
加載數據...
Files already downloaded and verified
Files already downloaded and verified實驗1：訓練基礎CNN（3層）...
Train Epoch: 1 [0/50000 (0%)] Loss: 2.314068 Accuracy: 14.06% LR: 0.010000
Train Epoch: 1 [6400/50000 (13%)] Loss: 1.568440 Accuracy: 30.23% LR: 0.010000
Train Epoch: 1 [12800/50000 (26%)] Loss: 1.659113 Accuracy: 35.59% LR: 0.010000
Train Epoch: 1 [19200/50000 (38%)] Loss: 1.617478 Accuracy: 38.03% LR: 0.010000
Train Epoch: 1 [25600/50000 (51%)] Loss: 1.440859 Accuracy: 39.80% LR: 0.010000
Train Epoch: 1 [32000/50000 (64%)] Loss: 1.178089 Accuracy: 41.20% LR: 0.010000
Train Epoch: 1 [38400/50000 (77%)] Loss: 1.137583 Accuracy: 42.30% LR: 0.010000 rain Epoch: 1 [44800/50000 (90%)] Loss: 1.342647 Accuracy: 43.36% LR: 0.010000
Train Epoch: 1 [44800/50000 (90%)] Loss: 1.342647 Accuracy: 43.36% LR: 0.010000 rain Epoch: 2 [0/50000 (0%)] Loss: 1.309829 Accuracy: 64.06% LR: 0.010000
Train Epoch: 1 [38400/50000 (77%)] Loss: 1.137583 Accuracy: 42.30% LR: 0.010000
Train Epoch: 1 [44800/50000 (90%)] Loss: 1.342647 Accuracy: 43.36% LR: 0.010000
Epoch 1: Test loss: 1.2134, Accuracy: 56.47%
Train Epoch: 2 [0/50000 (0%)] Loss: 1.309829 Accuracy: 64.06% LR: 0.010000
Train Epoch: 2 [6400/50000 (13%)] Loss: 1.275745 Accuracy: 52.75% LR: 0.010000
Train Epoch: 2 [12800/50000 (26%)] Loss: 1.147640 Accuracy: 53.40% LR: 0.010000
Train Epoch: 2 [19200/50000 (38%)] Loss: 1.347642 Accuracy: 54.10% LR: 0.010000
Train Epoch: 2 [25600/50000 (51%)] Loss: 1.003166 Accuracy: 54.91% LR: 0.010000
Train Epoch: 2 [32000/50000 (64%)] Loss: 1.219489 Accuracy: 55.54% LR: 0.010000
Train Epoch: 2 [38400/50000 (77%)] Loss: 1.321691 Accuracy: 56.07% LR: 0.010000
Train Epoch: 2 [44800/50000 (90%)] Loss: 1.249267 Accuracy: 56.47% LR: 0.010000
Epoch 2: Test loss: 1.0821, Accuracy: 62.85%
Train Epoch: 3 [0/50000 (0%)] Loss: 1.189534 Accuracy: 46.88% LR: 0.001000
Train Epoch: 3 [6400/50000 (13%)] Loss: 0.800227 Accuracy: 63.40% LR: 0.001000
Train Epoch: 3 [12800/50000 (26%)] Loss: 1.102200 Accuracy: 64.44% LR: 0.001000
Train Epoch: 3 [19200/50000 (38%)] Loss: 0.698571 Accuracy: 64.95% LR: 0.001000
Train Epoch: 3 [25600/50000 (51%)] Loss: 1.008356 Accuracy: 65.14% LR: 0.001000
Train Epoch: 3 [32000/50000 (64%)] Loss: 1.045420 Accuracy: 65.29% LR: 0.001000
Train Epoch: 3 [38400/50000 (77%)] Loss: 0.906162 Accuracy: 65.34% LR: 0.001000
Train Epoch: 3 [44800/50000 (90%)] Loss: 0.756010 Accuracy: 65.63% LR: 0.001000
Epoch 3: Test loss: 0.8147, Accuracy: 71.67%
Train Epoch: 4 [0/50000 (0%)] Loss: 0.931888 Accuracy: 70.31% LR: 0.001000
Train Epoch: 4 [6400/50000 (13%)] Loss: 1.229634 Accuracy: 67.79% LR: 0.001000
Train Epoch: 4 [12800/50000 (26%)] Loss: 0.853696 Accuracy: 67.86% LR: 0.001000
Train Epoch: 4 [19200/50000 (38%)] Loss: 0.732714 Accuracy: 68.29% LR: 0.001000
Train Epoch: 4 [25600/50000 (51%)] Loss: 0.950875 Accuracy: 67.96% LR: 0.001000
Train Epoch: 4 [32000/50000 (64%)] Loss: 0.639869 Accuracy: 68.21% LR: 0.001000
Train Epoch: 4 [38400/50000 (77%)] Loss: 0.861224 Accuracy: 68.15% LR: 0.001000
Train Epoch: 4 [44800/50000 (90%)] Loss: 1.007439 Accuracy: 68.26% LR: 0.001000
Epoch 4: Test loss: 0.7860, Accuracy: 72.78%
Train Epoch: 5 [0/50000 (0%)] Loss: 0.875544 Accuracy: 64.06% LR: 0.000100
Train Epoch: 5 [6400/50000 (13%)] Loss: 0.798205 Accuracy: 69.00% LR: 0.000100
Train Epoch: 5 [12800/50000 (26%)] Loss: 0.822474 Accuracy: 69.38% LR: 0.000100
Train Epoch: 5 [19200/50000 (38%)] Loss: 1.100589 Accuracy: 69.60% LR: 0.000100
Train Epoch: 5 [25600/50000 (51%)] Loss: 0.659224 Accuracy: 69.65% LR: 0.000100
Train Epoch: 5 [32000/50000 (64%)] Loss: 0.748625 Accuracy: 69.64% LR: 0.000100
Train Epoch: 5 [38400/50000 (77%)] Loss: 0.924533 Accuracy: 69.63% LR: 0.000100
Train Epoch: 5 [44800/50000 (90%)] Loss: 0.868795 Accuracy: 69.70% LR: 0.000100
Epoch 5: Test loss: 0.7704, Accuracy: 73.29%實驗2：訓練深層CNN（4層）...
Train Epoch: 1 [0/50000 (0%)] Loss: 2.361547 Accuracy: 10.94% LR: 0.010000
Train Epoch: 1 [6400/50000 (13%)] Loss: 1.508603 Accuracy: 31.33% LR: 0.010000
Train Epoch: 1 [12800/50000 (26%)] Loss: 1.625223 Accuracy: 36.70% LR: 0.010000
Train Epoch: 1 [19200/50000 (38%)] Loss: 1.560336 Accuracy: 39.16% LR: 0.010000
Train Epoch: 1 [25600/50000 (51%)] Loss: 1.314837 Accuracy: 41.68% LR: 0.010000
Train Epoch: 1 [32000/50000 (64%)] Loss: 1.158690 Accuracy: 43.56% LR: 0.010000
Train Epoch: 1 [38400/50000 (77%)] Loss: 1.332538 Accuracy: 45.04% LR: 0.010000
Train Epoch: 1 [44800/50000 (90%)] Loss: 1.487419 Accuracy: 46.50% LR: 0.010000
Epoch 1: Test loss: 1.0876, Accuracy: 61.16%
Train Epoch: 2 [0/50000 (0%)] Loss: 0.995175 Accuracy: 68.75% LR: 0.009045
Train Epoch: 2 [6400/50000 (13%)] Loss: 1.211926 Accuracy: 58.34% LR: 0.009045
Train Epoch: 2 [12800/50000 (26%)] Loss: 1.182561 Accuracy: 58.05% LR: 0.009045
Train Epoch: 2 [19200/50000 (38%)] Loss: 0.999875 Accuracy: 58.66% LR: 0.009045
Train Epoch: 2 [25600/50000 (51%)] Loss: 1.049008 Accuracy: 58.99% LR: 0.009045
Train Epoch: 2 [32000/50000 (64%)] Loss: 0.869797 Accuracy: 59.48% LR: 0.009045
Train Epoch: 2 [38400/50000 (77%)] Loss: 1.135825 Accuracy: 59.95% LR: 0.009045
Train Epoch: 2 [44800/50000 (90%)] Loss: 1.115167 Accuracy: 60.38% LR: 0.009045
Epoch 2: Test loss: 0.9288, Accuracy: 67.52%
Train Epoch: 3 [0/50000 (0%)] Loss: 0.859160 Accuracy: 79.69% LR: 0.006545
Train Epoch: 3 [6400/50000 (13%)] Loss: 0.984113 Accuracy: 67.08% LR: 0.006545
Train Epoch: 3 [12800/50000 (26%)] Loss: 0.934199 Accuracy: 66.88% LR: 0.006545
Train Epoch: 3 [19200/50000 (38%)] Loss: 1.200163 Accuracy: 67.24% LR: 0.006545
Train Epoch: 3 [25600/50000 (51%)] Loss: 0.886985 Accuracy: 67.39% LR: 0.006545
Train Epoch: 3 [32000/50000 (64%)] Loss: 0.588489 Accuracy: 67.56% LR: 0.006545
Train Epoch: 3 [38400/50000 (77%)] Loss: 0.802626 Accuracy: 67.73% LR: 0.006545
Train Epoch: 3 [44800/50000 (90%)] Loss: 0.769826 Accuracy: 68.06% LR: 0.006545
Epoch 3: Test loss: 0.8234, Accuracy: 71.23%
Train Epoch: 4 [0/50000 (0%)] Loss: 0.797326 Accuracy: 71.88% LR: 0.003455
Train Epoch: 4 [6400/50000 (13%)] Loss: 0.657178 Accuracy: 72.43% LR: 0.003455
Train Epoch: 4 [12800/50000 (26%)] Loss: 0.851532 Accuracy: 72.68% LR: 0.003455
Train Epoch: 4 [19200/50000 (38%)] Loss: 0.643618 Accuracy: 72.88% LR: 0.003455
Train Epoch: 4 [25600/50000 (51%)] Loss: 0.614715 Accuracy: 73.11% LR: 0.003455
Train Epoch: 4 [32000/50000 (64%)] Loss: 0.621094 Accuracy: 73.54% LR: 0.003455
Train Epoch: 4 [38400/50000 (77%)] Loss: 0.869759 Accuracy: 73.46% LR: 0.003455
Train Epoch: 4 [44800/50000 (90%)] Loss: 0.878335 Accuracy: 73.54% LR: 0.003455
Epoch 4: Test loss: 0.6981, Accuracy: 75.48%
Train Epoch: 5 [0/50000 (0%)] Loss: 0.766669 Accuracy: 68.75% LR: 0.000955
Train Epoch: 5 [6400/50000 (13%)] Loss: 0.553567 Accuracy: 75.59% LR: 0.000955
Train Epoch: 5 [12800/50000 (26%)] Loss: 0.499649 Accuracy: 76.55% LR: 0.000955
Train Epoch: 5 [19200/50000 (38%)] Loss: 0.672797 Accuracy: 76.39% LR: 0.000955
Train Epoch: 5 [25600/50000 (51%)] Loss: 0.780928 Accuracy: 76.45% LR: 0.000955
Train Epoch: 5 [32000/50000 (64%)] Loss: 0.580088 Accuracy: 76.59% LR: 0.000955
Train Epoch: 5 [38400/50000 (77%)] Loss: 0.698451 Accuracy: 76.64% LR: 0.000955
Train Epoch: 5 [44800/50000 (90%)] Loss: 0.682562 Accuracy: 76.69% LR: 0.000955
Epoch 5: Test loss: 0.6332, Accuracy: 78.08%
PS C:\Users\I.Love.I\Desktop\Python_code> & C:/Users/I.Love.I/.conda/envs/DL/python.exe c:/Users/I.Love.I/Desktop/Python_code/python60-days-challenge/1_python-learning-library/Day41_experiments.py
正在加載訓練集...
Files already downloaded and verified
正在加載測試集...
Files already downloaded and verified
開始訓練 基礎CNN 使用 CosineAnnealingLR，設備: cpu
Train Epoch: 1 [0/50000 (0%)] Loss: 2.293068 Accuracy: 12.50%
Train Epoch: 1 [6400/50000 (13%)] Loss: 1.722358 Accuracy: 27.48%
Train Epoch: 1 [12800/50000 (26%)] Loss: 1.556882 Accuracy: 31.74%
Train Epoch: 1 [19200/50000 (38%)] Loss: 1.758241 Accuracy: 34.25%
Train Epoch: 1 [25600/50000 (51%)] Loss: 1.674700 Accuracy: 35.67%
Train Epoch: 1 [32000/50000 (64%)] Loss: 1.315244 Accuracy: 37.07%
Train Epoch: 1 [38400/50000 (77%)] Loss: 1.367913 Accuracy: 38.25%
Train Epoch: 1 [44800/50000 (90%)] Loss: 1.559623 Accuracy: 39.41%Test set: Average loss: 1.2118, Accuracy: 5600/10000 (56.00%)Train Epoch: 2 [0/50000 (0%)] Loss: 1.484513 Accuracy: 50.00%
Train Epoch: 2 [6400/50000 (13%)] Loss: 1.664975 Accuracy: 47.22%
Train Epoch: 2 [12800/50000 (26%)] Loss: 1.220539 Accuracy: 48.27%
Train Epoch: 2 [19200/50000 (38%)] Loss: 1.407580 Accuracy: 49.03%
Train Epoch: 2 [25600/50000 (51%)] Loss: 1.092925 Accuracy: 49.81%
Train Epoch: 2 [32000/50000 (64%)] Loss: 1.206660 Accuracy: 50.29%
Train Epoch: 2 [38400/50000 (77%)] Loss: 1.379342 Accuracy: 50.75%
Train Epoch: 2 [44800/50000 (90%)] Loss: 1.501670 Accuracy: 51.13%Test set: Average loss: 1.1633, Accuracy: 6020/10000 (60.20%)Train Epoch: 3 [0/50000 (0%)] Loss: 1.454061 Accuracy: 50.00%
Train Epoch: 3 [6400/50000 (13%)] Loss: 0.939104 Accuracy: 57.24%
Train Epoch: 3 [12800/50000 (26%)] Loss: 1.240231 Accuracy: 56.58%
Train Epoch: 3 [19200/50000 (38%)] Loss: 0.877654 Accuracy: 57.00%
Train Epoch: 3 [25600/50000 (51%)] Loss: 1.166832 Accuracy: 57.22%
Train Epoch: 3 [32000/50000 (64%)] Loss: 1.057176 Accuracy: 57.53%
Train Epoch: 3 [38400/50000 (77%)] Loss: 1.339103 Accuracy: 57.71%
Train Epoch: 3 [44800/50000 (90%)] Loss: 1.041523 Accuracy: 58.04%Test set: Average loss: 0.9137, Accuracy: 6791/10000 (67.91%)Train Epoch: 4 [0/50000 (0%)] Loss: 0.974381 Accuracy: 68.75%
Train Epoch: 4 [6400/50000 (13%)] Loss: 1.366313 Accuracy: 62.33%
Train Epoch: 4 [12800/50000 (26%)] Loss: 0.929462 Accuracy: 62.22%
Train Epoch: 4 [19200/50000 (38%)] Loss: 0.884589 Accuracy: 62.59%
Train Epoch: 4 [25600/50000 (51%)] Loss: 1.171518 Accuracy: 62.30%
Train Epoch: 4 [32000/50000 (64%)] Loss: 0.842188 Accuracy: 62.71%
Train Epoch: 4 [38400/50000 (77%)] Loss: 1.012197 Accuracy: 62.51%
Train Epoch: 4 [44800/50000 (90%)] Loss: 1.134873 Accuracy: 62.71%Test set: Average loss: 0.8598, Accuracy: 6938/10000 (69.38%)Train Epoch: 5 [0/50000 (0%)] Loss: 0.878952 Accuracy: 67.19%
Train Epoch: 5 [6400/50000 (13%)] Loss: 1.031704 Accuracy: 64.60%
Train Epoch: 5 [12800/50000 (26%)] Loss: 1.064272 Accuracy: 65.48%
Train Epoch: 5 [19200/50000 (38%)] Loss: 1.377383 Accuracy: 65.85%
Train Epoch: 5 [25600/50000 (51%)] Loss: 0.908284 Accuracy: 66.02%
Train Epoch: 5 [32000/50000 (64%)] Loss: 0.973955 Accuracy: 66.14%
Train Epoch: 5 [38400/50000 (77%)] Loss: 1.087591 Accuracy: 66.22%
Train Epoch: 5 [44800/50000 (90%)] Loss: 0.889676 Accuracy: 66.24%Test set: Average loss: 0.7889, Accuracy: 7208/10000 (72.08%)Train Epoch: 6 [0/50000 (0%)] Loss: 0.876616 Accuracy: 70.31%
Train Epoch: 6 [6400/50000 (13%)] Loss: 0.897449 Accuracy: 67.44%
Train Epoch: 6 [12800/50000 (26%)] Loss: 1.097278 Accuracy: 67.63%
Train Epoch: 6 [19200/50000 (38%)] Loss: 0.918221 Accuracy: 68.12%
Train Epoch: 6 [25600/50000 (51%)] Loss: 0.754201 Accuracy: 68.52%
Train Epoch: 6 [32000/50000 (64%)] Loss: 0.915608 Accuracy: 68.48%
Train Epoch: 6 [38400/50000 (77%)] Loss: 1.108561 Accuracy: 68.63%
Train Epoch: 6 [44800/50000 (90%)] Loss: 0.889088 Accuracy: 68.67%Test set: Average loss: 0.6993, Accuracy: 7569/10000 (75.69%)Train Epoch: 7 [0/50000 (0%)] Loss: 0.958089 Accuracy: 65.62%
Train Epoch: 7 [6400/50000 (13%)] Loss: 1.091723 Accuracy: 70.81%
Train Epoch: 7 [12800/50000 (26%)] Loss: 0.777307 Accuracy: 70.35%
Train Epoch: 7 [19200/50000 (38%)] Loss: 1.069645 Accuracy: 70.53%
Train Epoch: 7 [25600/50000 (51%)] Loss: 0.896357 Accuracy: 70.77%
Train Epoch: 7 [32000/50000 (64%)] Loss: 0.776956 Accuracy: 70.76%
Train Epoch: 7 [38400/50000 (77%)] Loss: 0.742222 Accuracy: 70.88%
Train Epoch: 7 [44800/50000 (90%)] Loss: 0.773929 Accuracy: 70.91%Test set: Average loss: 0.6808, Accuracy: 7606/10000 (76.06%)Train Epoch: 8 [0/50000 (0%)] Loss: 0.693062 Accuracy: 73.44%
Train Epoch: 8 [6400/50000 (13%)] Loss: 0.597719 Accuracy: 71.81%
Train Epoch: 8 [12800/50000 (26%)] Loss: 0.577101 Accuracy: 72.22%
Train Epoch: 8 [19200/50000 (38%)] Loss: 0.715675 Accuracy: 72.38%
Train Epoch: 8 [25600/50000 (51%)] Loss: 0.894164 Accuracy: 72.39%
Train Epoch: 8 [32000/50000 (64%)] Loss: 0.889195 Accuracy: 72.42%
Train Epoch: 8 [38400/50000 (77%)] Loss: 0.777797 Accuracy: 72.43%
Train Epoch: 8 [44800/50000 (90%)] Loss: 0.707688 Accuracy: 72.41%Test set: Average loss: 0.6419, Accuracy: 7769/10000 (77.69%)Train Epoch: 9 [0/50000 (0%)] Loss: 0.801439 Accuracy: 73.44%
Train Epoch: 9 [6400/50000 (13%)] Loss: 0.908550 Accuracy: 72.94%
Train Epoch: 9 [12800/50000 (26%)] Loss: 0.692507 Accuracy: 73.35%
Train Epoch: 9 [19200/50000 (38%)] Loss: 0.560888 Accuracy: 73.52%
Train Epoch: 9 [25600/50000 (51%)] Loss: 0.740002 Accuracy: 73.79%
Train Epoch: 9 [32000/50000 (64%)] Loss: 0.751597 Accuracy: 73.90%
Train Epoch: 9 [38400/50000 (77%)] Loss: 0.994493 Accuracy: 73.81%
Train Epoch: 9 [44800/50000 (90%)] Loss: 0.777613 Accuracy: 73.88%Test set: Average loss: 0.6190, Accuracy: 7830/10000 (78.30%)Train Epoch: 10 [0/50000 (0%)] Loss: 0.687152 Accuracy: 79.69%
Train Epoch: 10 [6400/50000 (13%)] Loss: 0.801341 Accuracy: 73.86%
Train Epoch: 10 [12800/50000 (26%)] Loss: 0.684478 Accuracy: 73.98%
Train Epoch: 10 [19200/50000 (38%)] Loss: 0.705056 Accuracy: 74.19%
Train Epoch: 10 [25600/50000 (51%)] Loss: 0.853381 Accuracy: 74.05%
Train Epoch: 10 [32000/50000 (64%)] Loss: 0.580054 Accuracy: 74.07%
Train Epoch: 10 [38400/50000 (77%)] Loss: 0.817881 Accuracy: 74.00%
Train Epoch: 10 [44800/50000 (90%)] Loss: 0.817918 Accuracy: 74.10%Test set: Average loss: 0.6143, Accuracy: 7862/10000 (78.62%)正在加載訓練集...
Files already downloaded and verified
正在加載測試集...
Files already downloaded and verified
開始訓練 深層CNN 使用 CosineAnnealingLR，設備: cpu
Train Epoch: 1 [0/50000 (0%)] Loss: 2.342756 Accuracy: 14.06%
Train Epoch: 1 [6400/50000 (13%)] Loss: 1.850860 Accuracy: 25.28%
Train Epoch: 1 [12800/50000 (26%)] Loss: 1.762461 Accuracy: 30.99%
Train Epoch: 1 [19200/50000 (38%)] Loss: 1.590530 Accuracy: 34.26%
Train Epoch: 1 [25600/50000 (51%)] Loss: 1.340634 Accuracy: 36.95%
Train Epoch: 1 [32000/50000 (64%)] Loss: 1.501504 Accuracy: 39.23%
Train Epoch: 1 [38400/50000 (77%)] Loss: 1.322111 Accuracy: 41.06%
Train Epoch: 1 [44800/50000 (90%)] Loss: 1.156101 Accuracy: 42.66%Test set: Average loss: 1.2036, Accuracy: 5570/10000 (55.70%)Train Epoch: 2 [0/50000 (0%)] Loss: 1.286792 Accuracy: 54.69%
Train Epoch: 2 [6400/50000 (13%)] Loss: 1.453677 Accuracy: 55.48%
Train Epoch: 2 [12800/50000 (26%)] Loss: 0.941933 Accuracy: 56.44%
Train Epoch: 2 [19200/50000 (38%)] Loss: 1.107067 Accuracy: 56.65%
Train Epoch: 2 [25600/50000 (51%)] Loss: 1.146062 Accuracy: 56.92%
Train Epoch: 2 [32000/50000 (64%)] Loss: 1.112035 Accuracy: 57.31%
Train Epoch: 2 [38400/50000 (77%)] Loss: 1.183208 Accuracy: 57.59%
Train Epoch: 2 [44800/50000 (90%)] Loss: 1.217674 Accuracy: 58.32%Test set: Average loss: 1.0009, Accuracy: 6483/10000 (64.83%)Train Epoch: 3 [0/50000 (0%)] Loss: 1.032111 Accuracy: 62.50%
Train Epoch: 3 [6400/50000 (13%)] Loss: 0.904841 Accuracy: 64.26%
Train Epoch: 3 [12800/50000 (26%)] Loss: 1.111103 Accuracy: 64.02%
Train Epoch: 3 [19200/50000 (38%)] Loss: 1.111288 Accuracy: 64.01%
Train Epoch: 3 [25600/50000 (51%)] Loss: 0.887255 Accuracy: 64.14%
Train Epoch: 3 [32000/50000 (64%)] Loss: 0.821008 Accuracy: 64.42%
Train Epoch: 3 [38400/50000 (77%)] Loss: 0.922995 Accuracy: 64.60%
Train Epoch: 3 [44800/50000 (90%)] Loss: 0.852235 Accuracy: 64.90%Test set: Average loss: 0.7871, Accuracy: 7225/10000 (72.25%)Train Epoch: 4 [0/50000 (0%)] Loss: 1.008779 Accuracy: 68.75%
Train Epoch: 4 [6400/50000 (13%)] Loss: 0.877065 Accuracy: 68.30%
Train Epoch: 4 [12800/50000 (26%)] Loss: 0.890234 Accuracy: 68.38%
Train Epoch: 4 [19200/50000 (38%)] Loss: 0.822239 Accuracy: 68.46%
Train Epoch: 4 [25600/50000 (51%)] Loss: 1.126034 Accuracy: 68.56%
Train Epoch: 4 [32000/50000 (64%)] Loss: 0.859953 Accuracy: 68.46%
Train Epoch: 4 [38400/50000 (77%)] Loss: 0.792203 Accuracy: 68.64%
Train Epoch: 4 [44800/50000 (90%)] Loss: 0.921908 Accuracy: 68.79%Test set: Average loss: 0.7564, Accuracy: 7332/10000 (73.32%)Train Epoch: 5 [0/50000 (0%)] Loss: 0.739065 Accuracy: 68.75%
Train Epoch: 5 [6400/50000 (13%)] Loss: 0.773354 Accuracy: 71.61%
Train Epoch: 5 [12800/50000 (26%)] Loss: 0.605493 Accuracy: 72.07%
Train Epoch: 5 [19200/50000 (38%)] Loss: 0.617809 Accuracy: 72.11%
Train Epoch: 5 [25600/50000 (51%)] Loss: 0.883427 Accuracy: 72.07%
Train Epoch: 5 [32000/50000 (64%)] Loss: 0.842254 Accuracy: 71.99%
Train Epoch: 5 [38400/50000 (77%)] Loss: 0.533248 Accuracy: 72.05%
Train Epoch: 5 [44800/50000 (90%)] Loss: 1.000729 Accuracy: 72.02%Test set: Average loss: 0.6809, Accuracy: 7611/10000 (76.11%)Train Epoch: 6 [0/50000 (0%)] Loss: 0.810092 Accuracy: 73.44%
Train Epoch: 6 [6400/50000 (13%)] Loss: 0.616662 Accuracy: 73.48%
Train Epoch: 6 [12800/50000 (26%)] Loss: 0.562376 Accuracy: 74.10%
Train Epoch: 6 [19200/50000 (38%)] Loss: 0.832556 Accuracy: 73.90%
Train Epoch: 6 [25600/50000 (51%)] Loss: 0.878011 Accuracy: 74.10%
Train Epoch: 6 [32000/50000 (64%)] Loss: 0.509199 Accuracy: 74.05%
Train Epoch: 6 [38400/50000 (77%)] Loss: 0.673968 Accuracy: 74.19%
Train Epoch: 6 [44800/50000 (90%)] Loss: 0.744006 Accuracy: 74.30%Test set: Average loss: 0.6272, Accuracy: 7805/10000 (78.05%)Train Epoch: 7 [0/50000 (0%)] Loss: 0.555623 Accuracy: 81.25%
Train Epoch: 7 [6400/50000 (13%)] Loss: 0.578787 Accuracy: 76.49%
Train Epoch: 7 [12800/50000 (26%)] Loss: 0.823545 Accuracy: 76.82%
Train Epoch: 7 [19200/50000 (38%)] Loss: 0.716045 Accuracy: 76.79%
Train Epoch: 7 [25600/50000 (51%)] Loss: 0.815958 Accuracy: 76.64%
Train Epoch: 7 [32000/50000 (64%)] Loss: 0.755133 Accuracy: 76.55%
Train Epoch: 7 [38400/50000 (77%)] Loss: 0.602492 Accuracy: 76.47%
Train Epoch: 7 [44800/50000 (90%)] Loss: 0.758558 Accuracy: 76.46%Test set: Average loss: 0.5867, Accuracy: 7979/10000 (79.79%)Train Epoch: 8 [0/50000 (0%)] Loss: 0.834396 Accuracy: 73.44%
Train Epoch: 8 [6400/50000 (13%)] Loss: 0.533800 Accuracy: 78.57%
Train Epoch: 8 [12800/50000 (26%)] Loss: 0.626373 Accuracy: 78.33%
Train Epoch: 8 [19200/50000 (38%)] Loss: 0.643600 Accuracy: 78.60%
Train Epoch: 8 [25600/50000 (51%)] Loss: 0.740194 Accuracy: 78.72%
Train Epoch: 8 [32000/50000 (64%)] Loss: 0.602970 Accuracy: 78.66%
Train Epoch: 8 [38400/50000 (77%)] Loss: 0.432281 Accuracy: 78.71%
Train Epoch: 8 [44800/50000 (90%)] Loss: 0.451608 Accuracy: 78.71%Test set: Average loss: 0.5375, Accuracy: 8155/10000 (81.55%)Train Epoch: 9 [0/50000 (0%)] Loss: 0.863743 Accuracy: 68.75%
Train Epoch: 9 [6400/50000 (13%)] Loss: 0.412280 Accuracy: 79.07%
Train Epoch: 9 [12800/50000 (26%)] Loss: 0.553952 Accuracy: 79.86%
Train Epoch: 9 [19200/50000 (38%)] Loss: 0.724988 Accuracy: 79.91%
Train Epoch: 9 [25600/50000 (51%)] Loss: 0.498772 Accuracy: 79.94%
Train Epoch: 9 [32000/50000 (64%)] Loss: 0.791269 Accuracy: 80.01%
Train Epoch: 9 [38400/50000 (77%)] Loss: 0.518802 Accuracy: 79.95%
Train Epoch: 9 [44800/50000 (90%)] Loss: 0.408082 Accuracy: 79.84%Test set: Average loss: 0.5060, Accuracy: 8271/10000 (82.71%)Train Epoch: 10 [0/50000 (0%)] Loss: 0.534627 Accuracy: 84.38%
Train Epoch: 10 [6400/50000 (13%)] Loss: 0.785889 Accuracy: 80.35%
Train Epoch: 10 [12800/50000 (26%)] Loss: 0.448817 Accuracy: 80.48%
Train Epoch: 10 [19200/50000 (38%)] Loss: 0.501668 Accuracy: 80.71%
Train Epoch: 10 [25600/50000 (51%)] Loss: 0.587260 Accuracy: 80.95%
Train Epoch: 10 [32000/50000 (64%)] Loss: 0.423947 Accuracy: 81.13%
Train Epoch: 10 [38400/50000 (77%)] Loss: 0.846690 Accuracy: 81.14%
Train Epoch: 10 [44800/50000 (90%)] Loss: 0.598511 Accuracy: 81.15%Test set: Average loss: 0.4925, Accuracy: 8322/10000 (83.22%)正在加載訓練集...
Files already downloaded and verified
正在加載測試集...
Files already downloaded and verified
開始訓練 殘差CNN 使用 CosineAnnealingLR，設備: cpu
Traceback (most recent call last):File "c:\Users\I.Love.I\Desktop\Python_code\python60-days-challenge\1_python-learning-library\Day41_experiments.py", line 506, in <module>main()File "c:\Users\I.Love.I\Desktop\Python_code\python60-days-challenge\1_python-learning-library\Day41_experiments.py", line 498, in mainhistory, title = run_experiment(model_type, scheduler_type, epochs, batch_size, lr)File "c:\Users\I.Love.I\Desktop\Python_code\python60-days-challenge\1_python-learning-library\Day41_experiments.py", line 466, in run_experimenttrain_loss, train_acc = train(model, train_loader, optimizer, scheduler, epoch, history)File "c:\Users\I.Love.I\Desktop\Python_code\python60-days-challenge\1_python-learning-library\Day41_experiments.py", line 267, in trainoutput = model(data)File "C:\Users\I.Love.I\.conda\envs\DL\lib\site-packages\torch\nn\modules\module.py", line 1553, in _wrapped_call_implreturn self._call_impl(*args, **kwargs)File "C:\Users\I.Love.I\.conda\envs\DL\lib\site-packages\torch\nn\modules\module.py", line 1562, in _call_implreturn forward_call(*args, **kwargs)File "c:\Users\I.Love.I\Desktop\Python_code\python60-days-challenge\1_python-learning-library\Day41_experiments.py", line 249, in forwardout = self.fc(out)File "C:\Users\I.Love.I\.conda\envs\DL\lib\site-packages\torch\nn\modules\module.py", line 1553, in _wrapped_call_implreturn self._call_impl(*args, **kwargs)File "C:\Users\I.Love.I\.conda\envs\DL\lib\site-packages\torch\nn\modules\module.py", line 1562, in _call_implreturn forward_call(*args, **kwargs)File "C:\Users\I.Love.I\.conda\envs\DL\lib\site-packages\torch\nn\modules\linear.py", line 117, in forwardreturn F.linear(input, self.weight, self.bias)
RuntimeError: mat1 and mat2 shapes cannot be multiplied (64x1024 and 16384x10)
PS C:\Users\I.Love.I\Desktop\Python_code> & C:/Users/I.Love.I/.conda/envs/DL/python.exe c:/Users/I.Love.I/Desktop/Python_code/python60-days-challenge/1_python-learning-library/Day41_experiments.py
正在加載訓練集...
Files already downloaded and verified
正在加載測試集...
Files already downloaded and verified
開始訓練 基礎CNN 使用 CosineAnnealingLR，設備: cpu
Train Epoch: 1 [0/50000 (0%)] Loss: 2.293068 Accuracy: 12.50%
Train Epoch: 1 [6400/50000 (13%)] Loss: 1.722358 Accuracy: 27.48%
Train Epoch: 1 [12800/50000 (26%)] Loss: 1.556882 Accuracy: 31.74%
Train Epoch: 1 [19200/50000 (38%)] Loss: 1.758241 Accuracy: 34.25%
Train Epoch: 1 [25600/50000 (51%)] Loss: 1.674700 Accuracy: 35.67%
Train Epoch: 1 [32000/50000 (64%)] Loss: 1.315244 Accuracy: 37.07%
Train Epoch: 1 [38400/50000 (77%)] Loss: 1.367913 Accuracy: 38.25%
Train Epoch: 1 [44800/50000 (90%)] Loss: 1.559623 Accuracy: 39.41%Test set: Average loss: 1.2118, Accuracy: 5600/10000 (56.00%)Train Epoch: 2 [0/50000 (0%)] Loss: 1.484513 Accuracy: 50.00%
Train Epoch: 2 [6400/50000 (13%)] Loss: 1.664975 Accuracy: 47.22%
Train Epoch: 2 [12800/50000 (26%)] Loss: 1.220539 Accuracy: 48.27%
Train Epoch: 2 [19200/50000 (38%)] Loss: 1.407580 Accuracy: 49.03%
Train Epoch: 2 [25600/50000 (51%)] Loss: 1.092925 Accuracy: 49.81%
Train Epoch: 2 [32000/50000 (64%)] Loss: 1.206660 Accuracy: 50.29%
Train Epoch: 2 [38400/50000 (77%)] Loss: 1.379342 Accuracy: 50.75%
Train Epoch: 2 [44800/50000 (90%)] Loss: 1.501670 Accuracy: 51.13%Test set: Average loss: 1.1633, Accuracy: 6020/10000 (60.20%)Train Epoch: 3 [0/50000 (0%)] Loss: 1.454061 Accuracy: 50.00%
Train Epoch: 3 [6400/50000 (13%)] Loss: 0.939104 Accuracy: 57.24%
Train Epoch: 3 [12800/50000 (26%)] Loss: 1.240231 Accuracy: 56.58%
Train Epoch: 3 [19200/50000 (38%)] Loss: 0.877654 Accuracy: 57.00%
Train Epoch: 3 [25600/50000 (51%)] Loss: 1.166832 Accuracy: 57.22%
Train Epoch: 3 [32000/50000 (64%)] Loss: 1.057176 Accuracy: 57.53%
Train Epoch: 3 [38400/50000 (77%)] Loss: 1.339103 Accuracy: 57.71%
Train Epoch: 3 [44800/50000 (90%)] Loss: 1.041523 Accuracy: 58.04%Test set: Average loss: 0.9137, Accuracy: 6791/10000 (67.91%)Train Epoch: 4 [0/50000 (0%)] Loss: 0.974381 Accuracy: 68.75%
Train Epoch: 4 [6400/50000 (13%)] Loss: 1.366313 Accuracy: 62.33%
Train Epoch: 4 [12800/50000 (26%)] Loss: 0.929462 Accuracy: 62.22%
Train Epoch: 4 [19200/50000 (38%)] Loss: 0.884589 Accuracy: 62.59%
Train Epoch: 4 [25600/50000 (51%)] Loss: 1.171518 Accuracy: 62.30%
Train Epoch: 4 [32000/50000 (64%)] Loss: 0.842188 Accuracy: 62.71%
Train Epoch: 4 [38400/50000 (77%)] Loss: 1.012197 Accuracy: 62.51%
Train Epoch: 4 [44800/50000 (90%)] Loss: 1.134873 Accuracy: 62.71%Test set: Average loss: 0.8598, Accuracy: 6938/10000 (69.38%)Train Epoch: 5 [0/50000 (0%)] Loss: 0.878952 Accuracy: 67.19%
Train Epoch: 5 [6400/50000 (13%)] Loss: 1.031704 Accuracy: 64.60%
Train Epoch: 5 [12800/50000 (26%)] Loss: 1.064272 Accuracy: 65.48%
Train Epoch: 5 [19200/50000 (38%)] Loss: 1.377383 Accuracy: 65.85%
Train Epoch: 5 [25600/50000 (51%)] Loss: 0.908284 Accuracy: 66.02%
Train Epoch: 5 [32000/50000 (64%)] Loss: 0.973955 Accuracy: 66.14%
Train Epoch: 5 [38400/50000 (77%)] Loss: 1.087591 Accuracy: 66.22%
Train Epoch: 5 [44800/50000 (90%)] Loss: 0.889676 Accuracy: 66.24%Test set: Average loss: 0.7889, Accuracy: 7208/10000 (72.08%)Train Epoch: 6 [0/50000 (0%)] Loss: 0.876616 Accuracy: 70.31%
Train Epoch: 6 [6400/50000 (13%)] Loss: 0.897449 Accuracy: 67.44%
Train Epoch: 6 [12800/50000 (26%)] Loss: 1.097278 Accuracy: 67.63%
Train Epoch: 6 [19200/50000 (38%)] Loss: 0.918221 Accuracy: 68.12%
Train Epoch: 6 [25600/50000 (51%)] Loss: 0.754201 Accuracy: 68.52%
Train Epoch: 6 [32000/50000 (64%)] Loss: 0.915608 Accuracy: 68.48%
Train Epoch: 6 [38400/50000 (77%)] Loss: 1.108561 Accuracy: 68.63%
Train Epoch: 6 [44800/50000 (90%)] Loss: 0.889088 Accuracy: 68.67%Test set: Average loss: 0.6993, Accuracy: 7569/10000 (75.69%)Train Epoch: 7 [0/50000 (0%)] Loss: 0.958089 Accuracy: 65.62%
Train Epoch: 7 [6400/50000 (13%)] Loss: 1.091723 Accuracy: 70.81%
Train Epoch: 7 [12800/50000 (26%)] Loss: 0.777307 Accuracy: 70.35%
Train Epoch: 7 [19200/50000 (38%)] Loss: 1.069645 Accuracy: 70.53%
Train Epoch: 7 [25600/50000 (51%)] Loss: 0.896357 Accuracy: 70.77%
Train Epoch: 7 [32000/50000 (64%)] Loss: 0.776956 Accuracy: 70.76%
Train Epoch: 7 [38400/50000 (77%)] Loss: 0.742222 Accuracy: 70.88%
Train Epoch: 7 [44800/50000 (90%)] Loss: 0.773929 Accuracy: 70.91%Test set: Average loss: 0.6808, Accuracy: 7606/10000 (76.06%)Train Epoch: 8 [0/50000 (0%)] Loss: 0.693062 Accuracy: 73.44%
Train Epoch: 8 [6400/50000 (13%)] Loss: 0.597719 Accuracy: 71.81%
Train Epoch: 8 [12800/50000 (26%)] Loss: 0.577101 Accuracy: 72.22%
Train Epoch: 8 [19200/50000 (38%)] Loss: 0.715675 Accuracy: 72.38%
Train Epoch: 8 [25600/50000 (51%)] Loss: 0.894164 Accuracy: 72.39%
Train Epoch: 8 [32000/50000 (64%)] Loss: 0.889195 Accuracy: 72.42%
Train Epoch: 8 [38400/50000 (77%)] Loss: 0.777797 Accuracy: 72.43%
Train Epoch: 8 [44800/50000 (90%)] Loss: 0.707688 Accuracy: 72.41%Test set: Average loss: 0.6419, Accuracy: 7769/10000 (77.69%)Train Epoch: 9 [0/50000 (0%)] Loss: 0.801439 Accuracy: 73.44%
Train Epoch: 9 [6400/50000 (13%)] Loss: 0.908550 Accuracy: 72.94%
Train Epoch: 9 [12800/50000 (26%)] Loss: 0.692507 Accuracy: 73.35%
Train Epoch: 9 [19200/50000 (38%)] Loss: 0.560888 Accuracy: 73.52%
Train Epoch: 9 [25600/50000 (51%)] Loss: 0.740002 Accuracy: 73.79%
Train Epoch: 9 [32000/50000 (64%)] Loss: 0.751597 Accuracy: 73.90%
Train Epoch: 9 [38400/50000 (77%)] Loss: 0.994493 Accuracy: 73.81%
Train Epoch: 9 [44800/50000 (90%)] Loss: 0.777613 Accuracy: 73.88%Test set: Average loss: 0.6190, Accuracy: 7830/10000 (78.30%)Train Epoch: 10 [0/50000 (0%)] Loss: 0.687152 Accuracy: 79.69%
Train Epoch: 10 [6400/50000 (13%)] Loss: 0.801341 Accuracy: 73.86%
Train Epoch: 10 [12800/50000 (26%)] Loss: 0.684478 Accuracy: 73.98%
Train Epoch: 10 [19200/50000 (38%)] Loss: 0.705056 Accuracy: 74.19%
Train Epoch: 10 [25600/50000 (51%)] Loss: 0.853381 Accuracy: 74.05%
Train Epoch: 10 [32000/50000 (64%)] Loss: 0.580054 Accuracy: 74.07%
Train Epoch: 10 [38400/50000 (77%)] Loss: 0.817881 Accuracy: 74.00%
Train Epoch: 10 [44800/50000 (90%)] Loss: 0.817918 Accuracy: 74.10%Test set: Average loss: 0.6143, Accuracy: 7862/10000 (78.62%)正在加載訓練集...
Files already downloaded and verified
正在加載測試集...
Files already downloaded and verified
開始訓練 深層CNN 使用 CosineAnnealingLR，設備: cpu
Train Epoch: 1 [0/50000 (0%)] Loss: 2.342756 Accuracy: 14.06%
Train Epoch: 1 [6400/50000 (13%)] Loss: 1.850860 Accuracy: 25.28%"""
DAY 41 簡單CNN本節重點：
1. 數據增強
2. CNN結構定義
3. Batch Normalization
4. 特征圖可視化
5. 學習率調度器
"""import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torchvision
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
import numpy as np
from torch.utils.data import DataLoader# 設置中文字體
plt.rcParams['font.sans-serif'] = ['SimHei']
plt.rcParams['axes.unicode_minus'] = False# 設置隨機種子
torch.manual_seed(42)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")# CIFAR-10數據集的類別
classes = ('飛機', '汽車', '鳥', '貓', '鹿', '狗', '青蛙', '馬', '船', '卡車')#====================== 1. 數據加載與增強 ======================def load_data(batch_size=64, is_train=True):"""加載CIFAR-10數據集，并應用數據增強"""if is_train:# 訓練集使用數據增強transform = transforms.Compose([transforms.RandomHorizontalFlip(), # 隨機水平翻轉transforms.RandomRotation(10), # 隨機旋轉transforms.RandomAffine(0, shear=10, scale=(0.8, 1.2)), # 隨機仿射變換transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2), # 顏色抖動transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])else:# 測試集只需要標準化transform = transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])dataset = torchvision.datasets.CIFAR10(root='./data', train=is_train,download=True,transform=transform)dataloader = DataLoader(dataset,batch_size=batch_size,shuffle=is_train,num_workers=2)return dataloader#====================== 2. CNN模型定義 ======================class SimpleNet(nn.Module):def __init__(self, dropout_rate=0.5):super(SimpleNet, self).__init__()# 第一個卷積塊self.conv1 = nn.Conv2d(3, 32, kernel_size=3, padding=1)self.bn1 = nn.BatchNorm2d(32)# 第二個卷積塊self.conv2 = nn.Conv2d(32, 64, kernel_size=3, padding=1)self.bn2 = nn.BatchNorm2d(64)# 第三個卷積塊self.conv3 = nn.Conv2d(64, 128, kernel_size=3, padding=1)self.bn3 = nn.BatchNorm2d(128)# 全連接層self.fc1 = nn.Linear(128 * 4 * 4, 512)self.dropout = nn.Dropout(dropout_rate)self.fc2 = nn.Linear(512, 10)def forward(self, x):# 保存特征圖用于可視化self.feature_maps = []# 第一個卷積塊x = self.conv1(x)x = self.bn1(x)x = F.relu(x)x = F.max_pool2d(x, 2)self.feature_maps.append(x)# 第二個卷積塊x = self.conv2(x)x = self.bn2(x)x = F.relu(x)x = F.max_pool2d(x, 2)self.feature_maps.append(x)# 第三個卷積塊x = self.conv3(x)x = self.bn3(x)x = F.relu(x)x = F.max_pool2d(x, 2)self.feature_maps.append(x)# 全連接層x = torch.flatten(x, 1)x = self.fc1(x)x = F.relu(x)x = self.dropout(x)x = self.fc2(x)return F.log_softmax(x, dim=1)#====================== 3. 訓練函數 ======================def train(model, train_loader, optimizer, epoch, history):"""訓練一個epoch"""model.train()train_loss = 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 = F.nll_loss(output, target)loss.backward()optimizer.step()train_loss += loss.item()pred = output.max(1, keepdim=True)[1]correct += pred.eq(target.view_as(pred)).sum().item()total += target.size(0)if batch_idx % 100 == 0:print(f'Train Epoch: {epoch} [{batch_idx * len(data)}/{len(train_loader.dataset)} 'f'({100. * batch_idx / len(train_loader):.0f}%)]\t'f'Loss: {loss.item():.6f}\t'f'Accuracy: {100. * correct / total:.2f}%')epoch_loss = train_loss / len(train_loader)epoch_acc = 100. * correct / totalhistory['train_loss'].append(epoch_loss)history['train_acc'].append(epoch_acc)return epoch_loss, epoch_acc#====================== 4. 測試函數 ======================def test(model, test_loader, history):"""在測試集上評估模型"""model.eval()test_loss = 0correct = 0with torch.no_grad():for data, target in test_loader:data, target = data.to(device), target.to(device)output = model(data)test_loss += F.nll_loss(output, target, reduction='sum').item()pred = output.max(1, keepdim=True)[1]correct += pred.eq(target.view_as(pred)).sum().item()test_loss /= len(test_loader.dataset)accuracy = 100. * correct / len(test_loader.dataset)history['test_loss'].append(test_loss)history['test_acc'].append(accuracy)print(f'\nTest set: Average loss: {test_loss:.4f}, 'f'Accuracy: {correct}/{len(test_loader.dataset)} 'f'({accuracy:.2f}%)\n')return test_loss, accuracy#====================== 5. 可視化函數 ======================def plot_training_history(history):"""繪制訓練歷史曲線"""epochs = range(1, len(history['train_loss']) + 1)plt.figure(figsize=(12, 4))plt.subplot(1, 2, 1)plt.plot(epochs, history['train_loss'], 'b-', label='訓練損失')plt.plot(epochs, history['test_loss'], 'r-', label='測試損失')plt.title('訓練和測試損失')plt.xlabel('Epoch')plt.ylabel('損失')plt.legend()plt.grid(True)plt.subplot(1, 2, 2)plt.plot(epochs, history['train_acc'], 'b-', label='訓練準確率')plt.plot(epochs, history['test_acc'], 'r-', label='測試準確率')plt.title('訓練和測試準確率')plt.xlabel('Epoch')plt.ylabel('準確率 (%)')plt.legend()plt.grid(True)plt.tight_layout()plt.show()def visualize_feature_maps(model, test_loader):"""可視化特征圖"""# 獲取一個批次的數據dataiter = iter(test_loader)images, _ = next(dataiter)# 獲取特征圖with torch.no_grad():_ = model(images[0:1].to(device))# 顯示原始圖像和每層的特征圖plt.figure(figsize=(15, 5))# 顯示原始圖像plt.subplot(1, 4, 1)img = images[0] / 2 + 0.5npimg = img.numpy()plt.imshow(np.transpose(npimg, (1, 2, 0)))plt.title('原始圖像')plt.axis('off')# 顯示每層的特征圖for i, feature_map in enumerate(model.feature_maps, 2):plt.subplot(1, 4, i)# 選擇第一個樣本的第一個特征圖plt.imshow(feature_map[0, 0].cpu(), cmap='viridis')plt.title(f'層 {i-1} 特征圖')plt.axis('off')plt.tight_layout()plt.show()#====================== 6. 主函數 ======================def main():# 超參數設置batch_size = 64epochs = 4lr = 0.01dropout_rate = 0.5# 初始化訓練歷史記錄history = {'train_loss': [],'train_acc': [],'test_loss': [],'test_acc': []}# 加載數據print("正在加載訓練集...")train_loader = load_data(batch_size, is_train=True)print("正在加載測試集...")test_loader = load_data(batch_size, is_train=False)# 創建模型model = SimpleNet(dropout_rate=dropout_rate).to(device)optimizer = optim.SGD(model.parameters(), lr=lr, momentum=0.9)scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=2, gamma=0.1)# 訓練和測試print(f"開始訓練，使用設備: {device}")for epoch in range(1, epochs + 1):train_loss, train_acc = train(model, train_loader, optimizer, epoch, history)test_loss, test_acc = test(model, test_loader, history)scheduler.step()# 可視化訓練過程print("訓練完成，繪制訓練歷史...")plot_training_history(history)# 可視化特征圖print("可視化特征圖...")visualize_feature_maps(model, test_loader)if __name__ == '__main__':main()"""
學習要點：1. 數據增強
- transforms.RandomHorizontalFlip(): 隨機水平翻轉
- transforms.RandomRotation(): 隨機旋轉
- transforms.RandomAffine(): 隨機仿射變換
- transforms.ColorJitter(): 顏色抖動2. CNN結構
- 常見流程：輸入 → 卷積層 → BN → 激活函數 → 池化層
- 特征提取：多個卷積塊串聯
- 分類器：Flatten后接全連接層3. Batch Normalization
- 在卷積層后添加
- 訓練時計算并更新均值和方差
- 測試時使用訓練階段的統計量4. 特征圖
- 保存每層的特征圖用于可視化
- 觀察模型學習到的特征
- 幫助理解模型的工作原理5. 學習率調度器
- 使用StepLR按步長降低學習率
- 幫助模型更好地收斂
- 避免學習率過大或過小
"""
"""
DAY 41 簡單CNN - 基礎示例這個文件提供了一個簡單的CNN示例，展示了如何：
1. 修改CNN結構
2. 使用不同的學習率調度器
3. 觀察訓練效果的變化
"""import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import torchvision
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
from torch.utils.data import DataLoader
from torch.optim.lr_scheduler import StepLR, CosineAnnealingLR# 設置中文字體
plt.rcParams['font.sans-serif'] = ['SimHei']
plt.rcParams['axes.unicode_minus'] = False# 設置隨機種子和設備
torch.manual_seed(42)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")#====================== 1. 數據加載 ======================def load_data(batch_size=64, use_augmentation=True):"""加載CIFAR-10數據集參數：batch_size: 批次大小use_augmentation: 是否使用數據增強"""if use_augmentation:# 使用數據增強train_transform = transforms.Compose([transforms.RandomHorizontalFlip(), # 隨機水平翻轉transforms.RandomRotation(10), # 隨機旋轉transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])else:# 不使用數據增強train_transform = transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])# 測試集轉換test_transform = transforms.Compose([transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])# 加載數據集train_set = torchvision.datasets.CIFAR10(root='./data', train=True, download=True, transform=train_transform)test_set = torchvision.datasets.CIFAR10(root='./data', train=False, download=True, transform=test_transform)# 創建數據加載器train_loader = DataLoader(train_set, batch_size=batch_size, shuffle=True, num_workers=2)test_loader = DataLoader(test_set, batch_size=batch_size, shuffle=False, num_workers=2)return train_loader, test_loader#====================== 2. 模型定義 ======================class SimpleCNN(nn.Module):"""簡單的CNN模型，可以通過參數調整結構參數：num_conv_layers: 卷積層數量channels: 每層的通道數列表use_bn: 是否使用Batch Normalizationdropout_rate: Dropout比率"""def __init__(self, num_conv_layers=2, channels=[16, 32, 64], use_bn=False, dropout_rate=0.5):super(SimpleCNN, self).__init__()self.use_bn = use_bn# 創建卷積層self.conv_layers = nn.ModuleList()in_channels = 3 # 輸入圖像是RGB三通道for i in range(num_conv_layers):# 添加卷積層self.conv_layers.append(nn.Conv2d(in_channels, channels[i], kernel_size=3, padding=1))# 添加BN層（如果使用）if use_bn:self.conv_layers.append(nn.BatchNorm2d(channels[i]))in_channels = channels[i]# 計算全連接層的輸入維度# 假設輸入圖像是32x32，每經過一次池化層大小減半final_size = 32 // (2 ** num_conv_layers)fc_input = channels[-1] * final_size * final_size# 全連接層self.fc1 = nn.Linear(fc_input, 512)self.dropout = nn.Dropout(dropout_rate)self.fc2 = nn.Linear(512, 10)def forward(self, x):# 保存特征圖用于可視化self.feature_maps = []# 通過卷積層for i in range(0, len(self.conv_layers), 2 if self.use_bn else 1):x = self.conv_layers[i](x) # 卷積if self.use_bn:x = self.conv_layers[i+1](x) # BNx = F.relu(x) # 激活函數x = F.max_pool2d(x, 2) # 池化self.feature_maps.append(x) # 保存特征圖# 展平x = torch.flatten(x, 1)# 全連接層x = F.relu(self.fc1(x))x = self.dropout(x)x = self.fc2(x)return F.log_softmax(x, dim=1)#====================== 3. 訓練函數 ======================def train_model(model, train_loader, test_loader, optimizer, scheduler, epochs=5):"""訓練模型并記錄歷史"""history = {'train_loss': [], 'train_acc': [], 'test_loss': [], 'test_acc': []}for epoch in range(1, epochs + 1):model.train()train_loss = 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 = F.nll_loss(output, target)loss.backward()optimizer.step()train_loss += loss.item()pred = output.max(1, keepdim=True)[1]correct += pred.eq(target.view_as(pred)).sum().item()total += target.size(0)if batch_idx % 100 == 0:print(f'Train Epoch: {epoch} [{batch_idx * len(data)}/{len(train_loader.dataset)} 'f'({100. * batch_idx / len(train_loader):.0f}%)]\t'f'Loss: {loss.item():.6f}\t'f'Accuracy: {100. * correct / total:.2f}%\t'f'LR: {scheduler.get_last_lr()[0]:.6f}')# 記錄訓練指標history['train_loss'].append(train_loss / len(train_loader))history['train_acc'].append(100. * correct / total)# 測試model.eval()test_loss = 0correct = 0with torch.no_grad():for data, target in test_loader:data, target = data.to(device), target.to(device)output = model(data)test_loss += F.nll_loss(output, target, reduction='sum').item()pred = output.max(1, keepdim=True)[1]correct += pred.eq(target.view_as(pred)).sum().item()test_loss /= len(test_loader.dataset)accuracy = 100. * correct / len(test_loader.dataset)# 記錄測試指標history['test_loss'].append(test_loss)history['test_acc'].append(accuracy)print(f'Epoch {epoch}: Test loss: {test_loss:.4f}, Accuracy: {accuracy:.2f}%')# 更新學習率scheduler.step()return history#====================== 4. 可視化函數 ======================def plot_history(history, title):"""繪制訓練歷史"""plt.figure(figsize=(12, 4))# 繪制損失plt.subplot(1, 2, 1)plt.plot(history['train_loss'], label='訓練損失')plt.plot(history['test_loss'], label='測試損失')plt.title(f'{title} - 損失曲線')plt.xlabel('Epoch')plt.ylabel('損失')plt.legend()plt.grid(True)# 繪制準確率plt.subplot(1, 2, 2)plt.plot(history['train_acc'], label='訓練準確率')plt.plot(history['test_acc'], label='測試準確率')plt.title(f'{title} - 準確率曲線')plt.xlabel('Epoch')plt.ylabel('準確率 (%)')plt.legend()plt.grid(True)plt.tight_layout()plt.show()def visualize_feature_maps(model, test_loader):"""可視化特征圖"""# 獲取一個批次的數據dataiter = iter(test_loader)images, _ = next(dataiter)# 獲取特征圖with torch.no_grad():_ = model(images[0:1].to(device))# 顯示原始圖像和特征圖plt.figure(figsize=(15, 5))# 顯示原始圖像plt.subplot(1, len(model.feature_maps) + 1, 1)img = images[0] / 2 + 0.5 # 反歸一化plt.imshow(img.permute(1, 2, 0))plt.title('原始圖像')plt.axis('off')# 顯示每層的特征圖for i, feature_map in enumerate(model.feature_maps, 1):plt.subplot(1, len(model.feature_maps) + 1, i + 1)# 顯示第一個特征圖plt.imshow(feature_map[0, 0].cpu(), cmap='viridis')plt.title(f'特征圖 {i}')plt.axis('off')plt.tight_layout()plt.show()#====================== 5. 主函數 ======================def main():# 基礎配置batch_size = 64epochs = 5lr = 0.01# 加載數據print("加載數據...")train_loader, test_loader = load_data(batch_size, use_augmentation=True)# 實驗1：基礎CNN（3層）print("\n實驗1：訓練基礎CNN（3層）...")model1 = SimpleCNN(num_conv_layers=3, channels=[32, 64, 128], use_bn=True).to(device)optimizer1 = optim.SGD(model1.parameters(), lr=lr, momentum=0.9)scheduler1 = StepLR(optimizer1, step_size=2, gamma=0.1)history1 = train_model(model1, train_loader, test_loader, optimizer1, scheduler1, epochs)plot_history(history1, "基礎CNN（3層）+ StepLR")visualize_feature_maps(model1, test_loader)# 實驗2：深層CNN（4層）print("\n實驗2：訓練深層CNN（4層）...")model2 = SimpleCNN(num_conv_layers=4, channels=[32, 64, 128, 256], use_bn=True).to(device)optimizer2 = optim.SGD(model2.parameters(), lr=lr, momentum=0.9)scheduler2 = CosineAnnealingLR(optimizer2, T_max=epochs)history2 = train_model(model2, train_loader, test_loader, optimizer2, scheduler2, epochs)plot_history(history2, "深層CNN（4層）+ CosineAnnealingLR")visualize_feature_maps(model2, test_loader)if __name__ == '__main__':main()"""
學習要點：1. CNN結構修改
- 可以通過修改num_conv_layers和channels參數來改變網絡深度和寬度
- use_bn參數控制是否使用Batch Normalization
- dropout_rate參數調整Dropout比率2. 學習率調度器選擇
- StepLR：按固定步長降低學習率
- CosineAnnealingLR：余弦周期調整學習率3. 觀察重點
- 不同深度的網絡收斂速度
- 是否出現過擬合（訓練準確率高但測試準確率低）
- 特征圖的變化4. 實驗建議
- 嘗試不同的網絡深度（修改num_conv_layers和channels）
- 對比有無Batch Normalization的效果（修改use_bn）
- 測試不同的學習率調度策略
- 觀察數據增強的影響（修改use_augmentation）
"""
激情開賽)
