目录
完整代码：
进阶:
📂 项目目录结构
1. models/cnn_model.py (模型定义)
2. utils/visualizer.py (绘图工具)
3. data_loader.py (数据准备)
4. train_engine.py (训练引擎)
5. main.py (主入口)
在kaggle 找到一个图像数据集，用 cnn 网络进行训练并且用 grad-cam 做可视化
以Dogs vs. Cats —— 经典的二分类问题为例
真实场景：图片的分辨率不一，背景复杂，更接近现实项目。
进阶必备：你会学到如何调整图片大小（Resizing）、数据增强（Data Augmentation）以防止过拟合。
迁移学习入门：这是练习使用预训练模型（如 VGG16, ResNet）进行迁移学习（Transfer Learning）的最佳战场。
适合练习：二分类交叉熵损失函数（Binary Crossentropy）、数据流加载（ImageDataGenerator）。
Kaggle 链接：Dogs vs. Cats
完整代码：
import os 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 # 设置中文字体支持 plt.rcParams["font.family"] = ["SimHei"] plt.rcParams['axes.unicode_minus'] = False # 4. 定义CNN模型 class CNN(nn.Module): def __init__(self): super(CNN, self).__init__() # 第一个卷积块: 128x128 -> 64x64 self.conv1 = nn.Conv2d(3, 32, kernel_size=3, padding=1) self.bn1 = nn.BatchNorm2d(32) self.relu1 = nn.ReLU() self.pool1 = nn.MaxPool2d(2, 2) # 第二个卷积块: 64x64 -> 32x32 self.conv2 = nn.Conv2d(32, 64, kernel_size=3, padding=1) self.bn2 = nn.BatchNorm2d(64) self.relu2 = nn.ReLU() self.pool2 = nn.MaxPool2d(2, 2) # 第三个卷积块: 32x32 -> 16x16 self.conv3 = nn.Conv2d(64, 128, kernel_size=3, padding=1) self.bn3 = nn.BatchNorm2d(128) self.relu3 = nn.ReLU() self.pool3 = nn.MaxPool2d(2, 2) # 全连接层 # 注意：128x128 经过 3 次池化变为 16x16 self.fc1 = nn.Linear(128 * 16 * 16, 512) self.dropout = nn.Dropout(p=0.5) self.fc2 = nn.Linear(512, 2) # 猫狗双分类 def forward(self, x): x = self.pool1(self.relu1(self.bn1(self.conv1(x)))) x = self.pool2(self.relu2(self.bn2(self.conv2(x)))) x = self.pool3(self.relu3(self.bn3(self.conv3(x)))) x = x.view(x.size(0), -1) # 动态展平 x = torch.relu(self.fc1(x)) x = self.dropout(x) x = self.fc2(x) return x # 绘图函数保留在外面 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() def plot_epoch_metrics(train_acc, test_acc, train_loss, test_loss): epochs = range(1, len(train_acc) + 1) plt.figure(figsize=(12, 4)) 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.title('准确率曲线'); plt.legend(); plt.grid(True) 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.title('损失曲线'); plt.legend(); plt.grid(True) plt.tight_layout(); plt.show() # 5. 训练函数 def train(model, train_loader, val_loader, test_loader, criterion, optimizer, scheduler, device, epochs): """ 完整的训练逻辑 :param val_loader: 验证集加载器，用于训练过程中调整超参数 :param test_loader: 测试集加载器，用于最后评估模型泛化能力 """ # 记录数据用于绘图 all_iter_losses, iter_indices = [], [] train_acc_history, val_acc_history = [], [] train_loss_history, val_loss_history = [], [] print(f"开始训练，共 {epochs} 个 Epoch...") for epoch in range(epochs): # ==================== 1. 训练阶段 (Training) ==================== model.train() running_loss, correct, total = 0.0, 0, 0 for batch_idx, (data, target) in enumerate(train_loader): data, target = data.to(device), target.to(device) optimizer.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 iteration 级别的数据 iter_loss = loss.item() all_iter_losses.append(iter_loss) iter_indices.append(epoch * len(train_loader) + batch_idx + 1) running_loss += iter_loss _, predicted = output.max(1) total += target.size(0) correct += predicted.eq(target).sum().item() # 每 50 个 batch 打印一次进度 if (batch_idx + 1) % 50 == 0: print(f'Epoch: {epoch+1}/{epochs} [{batch_idx+1}/{len(train_loader)}] ' f'Loss: {iter_loss:.4f} | Acc: {100.*correct/total:.2f}%') epoch_train_loss = running_loss / len(train_loader) epoch_train_acc = 100. * correct / total train_acc_history.append(epoch_train_acc) train_loss_history.append(epoch_train_loss) # ==================== 2. 验证阶段 (Validation) ==================== # 每个 epoch 跑完都要去验证集“考试”，根据考试成绩调整学习率 model.eval() val_loss, correct_val, total_val = 0, 0, 0 with torch.no_grad(): # 验证阶段不计算梯度，节省内存和显存 for data, target in val_loader: data, target = data.to(device), target.to(device) output = model(data) val_loss += criterion(output, target).item() _, predicted = output.max(1) total_val += target.size(0) correct_val += predicted.eq(target).sum().item() epoch_val_loss = val_loss / len(val_loader) epoch_val_acc = 100. * correct_val / total_val val_acc_history.append(epoch_val_acc) val_loss_history.append(epoch_val_loss) # 根据验证集的损失调整学习率 scheduler.step(epoch_val_loss) # 获取当前学习率（用于打印） current_lr = optimizer.param_groups[0]['lr'] print(f'--- Epoch {epoch+1} 结束 | Train Acc: {epoch_train_acc:.2f}% | Val Acc: {epoch_val_acc:.2f}% | LR: {current_lr} ---') # ==================== 3. 最终测试阶段 (Testing) ==================== # 所有的训练都结束后，用完全没见过的测试集做最后的评估 print("\n" + "="*30) print("训练完成！正在进行最终测试...") model.eval() test_correct, test_total = 0, 0 with torch.no_grad(): for data, target in test_loader: data, target = data.to(device), target.to(device) output = model(data) _, predicted = output.max(1) test_total += target.size(0) test_correct += predicted.eq(target).sum().item() final_test_acc = 100. * test_correct / test_total print(f'终极测试准确率: {final_test_acc:.2f}%') print("="*30) # 绘制图表 plot_iter_losses(all_iter_losses, iter_indices) plot_epoch_metrics(train_acc_history, val_acc_history, train_loss_history, val_loss_history) return final_test_acc # 6. 主执行入口 if __name__ == '__main__': device = torch.device("cuda" if torch.cuda.is_available() else "cpu") print(f"使用设备: {device}") # 数据路径处理 current_dir = os.path.dirname(os.path.abspath(__file__)) train_path = os.path.join(current_dir, 'dataset', 'train') val_path = os.path.join(current_dir, 'dataset', 'validation') test_path = os.path.join(current_dir, 'dataset', 'test') # 数据预处理 target_size = (128, 128) train_transform = transforms.Compose([ transforms.Resize(target_size), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ]) test_transform = transforms.Compose([ transforms.Resize(target_size), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ]) # 加载器 train_dataset = datasets.ImageFolder(train_path, transform=train_transform) test_dataset = datasets.ImageFolder(test_path, transform=test_transform) train_loader = DataLoader(train_dataset, batch_size=64, shuffle=True, num_workers=2) test_loader = DataLoader(test_dataset, batch_size=64, shuffle=False, num_workers=2) # 加载验证集 val_dataset = datasets.ImageFolder(val_path, transform=test_transform) val_loader = DataLoader(val_dataset, batch_size=64, shuffle=False, num_workers=2) # 模型初始化 model = CNN().to(device) criterion = nn.CrossEntropyLoss() optimizer = optim.Adam(model.parameters(), lr=0.001) scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', patience=3, factor=0.5) # 启动训练 final_acc = train(model, train_loader, val_loader, test_loader, criterion, optimizer, scheduler, device, epochs=20) print(f"最终准确率: {final_acc:.2f}%")进阶:
将代码拆分为多个文件（模块化）是开发深度学习项目的标准操作。这样做不仅让代码清晰，还方便你以后更换模型（比如换成 ResNet）或更换数据集而不需要大规模改动代码。
按照以下结构拆分：
📂 项目目录结构
day49/ ├── dataset/ # 数据集文件夹 ├── models/ │ └── cnn_model.py # 存放模型结构 (CNN类) ├── utils/ │ └── visualizer.py # 存放绘图函数 (plot_... 函数) ├── data_loader.py # 存放数据预处理和 DataLoader 逻辑 ├── train_engine.py # 存放 train 核心函数 └── main.py # 执行入口1.models/cnn_model.py(模型定义)
将模型单独拎出来，方便以后在其他项目复用。
import torch import torch.nn as nn class CNN(nn.Module): def __init__(self): super(CNN, self).__init__() self.conv1 = nn.Conv2d(3, 32, kernel_size=3, padding=1) self.bn1 = nn.BatchNorm2d(32) self.relu1 = nn.ReLU() self.pool1 = nn.MaxPool2d(2, 2) self.conv2 = nn.Conv2d(32, 64, kernel_size=3, padding=1) self.bn2 = nn.BatchNorm2d(64) self.relu2 = nn.ReLU() self.pool2 = nn.MaxPool2d(2, 2) self.conv3 = nn.Conv2d(64, 128, kernel_size=3, padding=1) self.bn3 = nn.BatchNorm2d(128) self.relu3 = nn.ReLU() self.pool3 = nn.MaxPool2d(2, 2) self.fc1 = nn.Linear(128 * 16 * 16, 512) self.dropout = nn.Dropout(p=0.5) self.fc2 = nn.Linear(512, 2) def forward(self, x): x = self.pool1(self.relu1(self.bn1(self.conv1(x)))) x = self.pool2(self.relu2(self.bn2(self.conv2(x)))) x = self.pool3(self.relu3(self.bn3(self.conv3(x)))) x = x.view(x.size(0), -1) x = torch.relu(self.fc1(x)) x = self.dropout(x) x = self.fc2(x) return x2.utils/visualizer.py(绘图工具)
绘图逻辑通常比较占篇幅，且与训练逻辑无关。
import matplotlib.pyplot as plt # 设置中文字体支持 plt.rcParams["font.family"] = ["SimHei"] plt.rcParams['axes.unicode_minus'] = False 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') plt.ylabel('损失值') plt.title('训练损失') plt.legend(); plt.grid(True); plt.show() def plot_epoch_metrics(train_acc, val_acc, train_loss, val_loss): epochs = range(1, len(train_acc) + 1) plt.figure(figsize=(12, 4)) # ... 之前的绘图逻辑 ... plt.tight_layout(); plt.show() import cv2 import numpy as np import torch import torch.nn.functional as F import matplotlib.pyplot as plt def show_gradcam(model, img_tensor, original_img_path, target_layer): """ model: 训练好的模型 img_tensor: 经过 transform 处理后的图像张量 [1, 3, 128, 128] original_img_path: 原始图片的路径（用于叠加显示） target_layer: 想要可视化的卷积层（通常是最后一个卷积层） """ model.eval() # 1. 注册 Hook 获取梯度和特征图 gradients = [] activations = [] def backward_hook(module, grad_input, grad_output): gradients.append(grad_output[0]) def forward_hook(module, input, output): activations.append(output) # 绑定到目标层 handle_forward = target_layer.register_forward_hook(forward_hook) handle_backward = target_layer.register_full_backward_hook(backward_hook) # 2. 前向传播 output = model(img_tensor) category_index = output.argmax(dim=1).item() # 3. 反向传播获取梯度 model.zero_grad() loss = output[0, category_index] loss.backward() # 4. 计算 Grad-CAM grads = gradients[0].cpu().data.numpy()[0] # [C, H, W] f_maps = activations[0].cpu().data.numpy()[0] # [C, H, W] # 对通道维度取平均值作为权重 weights = np.mean(grads, axis=(1, 2)) cam = np.zeros(f_maps.shape[1:], dtype=np.float32) for i, w in enumerate(weights): cam += w * f_maps[i] # ReLU 激活并归一化 cam = np.maximum(cam, 0) cam = cv2.resize(cam, (128, 128)) cam = (cam - np.min(cam)) / (np.max(cam) - np.min(cam)) # 5. 叠加到原图 img = cv2.imdecode(np.fromfile(original_img_path, dtype=np.uint8), cv2.IMREAD_COLOR) img = cv2.resize(img, (128, 128)) heatmap = cv2.applyColorMap(np.uint8(255 * cam), cv2.COLORMAP_JET) result = heatmap * 0.4 + img * 0.6 # 0.4 是热力图透明度 # 移除 Hook handle_forward.remove() handle_backward.remove() # 展示结果 plt.figure(figsize=(8, 4)) plt.subplot(1, 2, 1) plt.imshow(cv2.cvtColor(img, cv2.COLOR_BGR2RGB)) plt.title("Original Image") plt.subplot(1, 2, 2) plt.imshow(cv2.cvtColor(np.uint8(result), cv2.COLOR_BGR2RGB)) plt.title(f"Grad-CAM (Class: {'Dog' if category_index==1 else 'Cat'})") plt.show()3.data_loader.py(数据准备)
这部分负责把原始图片变成模型能吃的DataLoader。
import os from torchvision import datasets, transforms from torch.utils.data import DataLoader def get_loaders(current_dir, batch_size=64): target_size = (128, 128) train_transform = transforms.Compose([ transforms.Resize(target_size), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ]) test_transform = transforms.Compose([ transforms.Resize(target_size), transforms.ToTensor(), transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]) ]) train_path = os.path.join(current_dir, 'dataset', 'train') val_path = os.path.join(current_dir, 'dataset', 'validation') test_path = os.path.join(current_dir, 'dataset', 'test') train_loader = DataLoader(datasets.ImageFolder(train_path, train_transform), batch_size=batch_size, shuffle=True, num_workers=2) val_loader = DataLoader(datasets.ImageFolder(val_path, test_transform), batch_size=batch_size, shuffle=False, num_workers=2) test_loader = DataLoader(datasets.ImageFolder(test_path, test_transform), batch_size=batch_size, shuffle=False, num_workers=2) return train_loader, val_loader, test_loader, test_transform4.train_engine.py(训练引擎)
把train函数放进来。注意要从其他模块导入绘图工具。
import torch from utils.visualizer import plot_iter_losses, plot_epoch_metrics def train(model, train_loader, val_loader, test_loader, criterion, optimizer, scheduler, device, epochs): """ 完整的训练逻辑 :param val_loader: 验证集加载器，用于训练过程中调整超参数 :param test_loader: 测试集加载器，用于最后评估模型泛化能力 """ # 记录数据用于绘图 all_iter_losses, iter_indices = [], [] train_acc_history, val_acc_history = [], [] train_loss_history, val_loss_history = [], [] print(f"开始训练，共 {epochs} 个 Epoch...") for epoch in range(epochs): # ==================== 1. 训练阶段 (Training) ==================== model.train() running_loss, correct, total = 0.0, 0, 0 for batch_idx, (data, target) in enumerate(train_loader): data, target = data.to(device), target.to(device) optimizer.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 iteration 级别的数据 iter_loss = loss.item() all_iter_losses.append(iter_loss) iter_indices.append(epoch * len(train_loader) + batch_idx + 1) running_loss += iter_loss _, predicted = output.max(1) total += target.size(0) correct += predicted.eq(target).sum().item() # 每 50 个 batch 打印一次进度 if (batch_idx + 1) % 50 == 0: print(f'Epoch: {epoch+1}/{epochs} [{batch_idx+1}/{len(train_loader)}] ' f'Loss: {iter_loss:.4f} | Acc: {100.*correct/total:.2f}%') epoch_train_loss = running_loss / len(train_loader) epoch_train_acc = 100. * correct / total train_acc_history.append(epoch_train_acc) train_loss_history.append(epoch_train_loss) # ==================== 2. 验证阶段 (Validation) ==================== # 每个 epoch 跑完都要去验证集“考试”，根据考试成绩调整学习率 model.eval() val_loss, correct_val, total_val = 0, 0, 0 with torch.no_grad(): # 验证阶段不计算梯度，节省内存和显存 for data, target in val_loader: data, target = data.to(device), target.to(device) output = model(data) val_loss += criterion(output, target).item() _, predicted = output.max(1) total_val += target.size(0) correct_val += predicted.eq(target).sum().item() epoch_val_loss = val_loss / len(val_loader) epoch_val_acc = 100. * correct_val / total_val val_acc_history.append(epoch_val_acc) val_loss_history.append(epoch_val_loss) # 根据验证集的损失调整学习率 scheduler.step(epoch_val_loss) # 获取当前学习率（用于打印） current_lr = optimizer.param_groups[0]['lr'] print(f'--- Epoch {epoch+1} 结束 | Train Acc: {epoch_train_acc:.2f}% | Val Acc: {epoch_val_acc:.2f}% | LR: {current_lr} ---') # ==================== 3. 最终测试阶段 (Testing) ==================== # 所有的训练都结束后，用完全没见过的测试集做最后的评估 print("\n" + "="*30) print("训练完成！正在进行最终测试...") model.eval() test_correct, test_total = 0, 0 with torch.no_grad(): for data, target in test_loader: data, target = data.to(device), target.to(device) output = model(data) _, predicted = output.max(1) test_total += target.size(0) test_correct += predicted.eq(target).sum().item() final_test_acc = 100. * test_correct / test_total print(f'终极测试准确率: {final_test_acc:.2f}%') print("="*30) # 绘制图表 plot_iter_losses(all_iter_losses, iter_indices) plot_epoch_metrics(train_acc_history, val_acc_history, train_loss_history, val_loss_history) return final_test_acc5.main.py(主入口)
主文件现在变得非常干净，只负责调度。
import os import torch import torch.nn as nn import torch.optim as optim # 导入你拆分的模块 from models.cnn_model import CNN from data_loader import get_loaders from train_engine import train if __name__ == '__main__': # 1. 配置 device = torch.device("cuda" if torch.cuda.is_available() else "cpu") current_dir = os.path.dirname(os.path.abspath(__file__)) # 2. 获取数据 train_loader, val_loader, test_loader,test_transform = get_loaders(current_dir) # 3. 初始化模型 model = CNN().to(device) criterion = nn.CrossEntropyLoss() optimizer = optim.Adam(model.parameters(), lr=0.001) scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', patience=3, factor=0.5) # 4. 运行 final_acc = train(model, train_loader, val_loader, test_loader, criterion, optimizer, scheduler, device, epochs=2) print(f"最终准确率: {final_acc:.2f}%") # --- Grad-CAM 可视化部分 --- from utils.visualizer import show_gradcam # 1. 挑一张测试图片（或者你本地找一张猫/狗的图） # 手动拼接完整路径 target_img_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'dataset', 'test', 'dogs', 'dog (1001).jpg') # 2. 对这张图做相同的预处理 from PIL import Image raw_img = Image.open(target_img_path).convert('RGB') input_tensor = test_transform(raw_img).unsqueeze(0).to(device) # 增加 batch 维度并移至 GPU # 3. 指定可视化最后一层卷积层 target_layer = model.conv3 # 4. 绘图 print("生成 Grad-CAM 可视化中...") show_gradcam(model, input_tensor, target_img_path, target_layer)
