当前位置：首页 > news >正文

Day 41 训练

news 2025/6/7 6:52:27

Day 41 简单 CNN：探索卷积神经网络的奥秘

在深度学习领域中，卷积神经网络（CNN）一直是计算机视觉任务的中流砥柱。今天，我将通过一系列实验来探索 CNN 的奥秘，并分享我的代码和详细解释。

实验准备：数据加载与增强

Python

复制

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# 设置中文字体
plt.rcParams['font.sans-serif'] = ['SimHei']
plt.rcParams['axes.unicode_minus'] = False# 设置随机种子和设备
torch.manual_seed(42)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")# CIFAR-10数据集的类别
classes = ('飞机', '汽车', '鸟', '猫', '鹿', '狗', '青蛙', '马', '船', '卡车')#====================== 1. 数据加载与增强 ======================def load_data(batch_size=64, is_train=True):"""加载CIFAR-10数据集，并应用数据增强"""if is_train:# 训练集使用数据增强transform = transforms.Compose([transforms.RandomHorizontalFlip(),  # 随机水平翻转transforms.RandomRotation(10),      # 随机旋转transforms.RandomAffine(0, shear=10, scale=(0.8, 1.2)),  # 随机仿射变换transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2),  # 颜色抖动transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])else:# 测试集只需要标准化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

我首先准备了 CIFAR-10 数据集，并在训练集上应用了数据增强技术，包括随机水平翻转、随机旋转、随机仿射变换和颜色抖动，以提高模型的泛化能力。

CNN 模型定义

Python

复制

#====================== 2. CNN模型定义 ======================class SimpleNet(nn.Module):def __init__(self, dropout_rate=0.5):super(SimpleNet, self).__init__()# 第一个卷积块self.conv1 = nn.Conv2d(3, 32, kernel_size=3, padding=1)self.bn1 = nn.BatchNorm2d(32)# 第二个卷积块self.conv2 = nn.Conv2d(32, 64, kernel_size=3, padding=1)self.bn2 = nn.BatchNorm2d(64)# 第三个卷积块self.conv3 = nn.Conv2d(64, 128, kernel_size=3, padding=1)self.bn3 = nn.BatchNorm2d(128)# 全连接层self.fc1 = nn.Linear(128 * 4 * 4, 512)self.dropout = nn.Dropout(dropout_rate)self.fc2 = nn.Linear(512, 10)def forward(self, x):# 保存特征图用于可视化self.feature_maps = []# 第一个卷积块x = self.conv1(x)x = self.bn1(x)x = F.relu(x)x = F.max_pool2d(x, 2)self.feature_maps.append(x)# 第二个卷积块x = self.conv2(x)x = self.bn2(x)x = F.relu(x)x = F.max_pool2d(x, 2)self.feature_maps.append(x)# 第三个卷积块x = self.conv3(x)x = self.bn3(x)x = F.relu(x)x = F.max_pool2d(x, 2)self.feature_maps.append(x)# 全连接层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)

我定义了一个简单的 CNN 模型 SimpleNet，它包含三个卷积块，每个卷积块后都跟着 Batch Normalization 层、ReLU 激活函数和最大池化层。最后是全连接层用于分类。

训练函数

Python

复制

#====================== 3. 训练函数 ======================def train(model, train_loader, optimizer, epoch, history):"""训练一个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

训练函数用于执行一个 epoch 的训练过程，计算损失并更新模型参数。同时，它记录了训练的损失和准确率。

测试函数

Python

复制

#====================== 4. 测试函数 ======================def test(model, test_loader, history):"""在测试集上评估模型"""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

测试函数用于在测试集上评估模型的性能，计算测试损失和准确率。

可视化函数

Python

复制

#====================== 5. 可视化函数 ======================def plot_training_history(history):"""绘制训练历史曲线"""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='训练损失')plt.plot(epochs, history['test_loss'], 'r-', label='测试损失')plt.title('训练和测试损失')plt.xlabel('Epoch')plt.ylabel('损失')plt.legend()plt.grid(True)plt.subplot(1, 2, 2)plt.plot(epochs, history['train_acc'], 'b-', label='训练准确率')plt.plot(epochs, history['test_acc'], 'r-', label='测试准确率')plt.title('训练和测试准确率')plt.xlabel('Epoch')plt.ylabel('准确率 (%)')plt.legend()plt.grid(True)plt.tight_layout()plt.show()def visualize_feature_maps(model, test_loader):"""可视化特征图"""# 获取一个批次的数据dataiter = iter(test_loader)images, _ = next(dataiter)# 获取特征图with torch.no_grad():_ = model(images[0:1].to(device))# 显示原始图像和每层的特征图plt.figure(figsize=(15, 5))# 显示原始图像plt.subplot(1, 4, 1)img = images[0] / 2 + 0.5npimg = img.numpy()plt.imshow(np.transpose(npimg, (1, 2, 0)))plt.title('原始图像')plt.axis('off')# 显示每层的特征图for i, feature_map in enumerate(model.feature_maps, 2):plt.subplot(1, 4, i)# 选择第一个样本的第一个特征图plt.imshow(feature_map[0, 0].cpu(), cmap='viridis')plt.title(f'层 {i-1} 特征图')plt.axis('off')plt.tight_layout()plt.show()

可视化函数用于绘制训练历史曲线和特征图，帮助我们直观地理解模型的训练过程和内部特征提取情况。

主函数

Python

复制

#====================== 6. 主函数 ======================def main():# 超参数设置batch_size = 64epochs = 4lr = 0.01dropout_rate = 0.5# 初始化训练历史记录history = {'train_loss': [],'train_acc': [],'test_loss': [],'test_acc': []}# 加载数据print("正在加载训练集...")train_loader = load_data(batch_size, is_train=True)print("正在加载测试集...")test_loader = load_data(batch_size, is_train=False)# 创建模型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)# 训练和测试print(f"开始训练，使用设备: {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()# 可视化训练过程print("训练完成，绘制训练历史...")plot_training_history(history)# 可视化特征图print("可视化特征图...")visualize_feature_maps(model, test_loader)if __name__ == '__main__':main()

主函数整合了整个实验流程，包括数据加载、模型创建、训练、测试和可视化等步骤。