学习来源：@浙大疏锦行
import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision
import torchvision.transforms as transforms
import numpy as np
import matplotlib.pyplot as plt
from PIL import Image
import os
# 设置随机种子确保结果可复现
torch.manual_seed(42)
np.random.seed(42)
# 将 'path/to/your_dataset' 替换为你的数据集所在的根目录
data_dir = './data/10 Big Cats of the Wild - Image Classification'
train_dir = os.path.join(data_dir, 'train')
test_dir = os.path.join(data_dir, 'test')
if not os.path.isdir(data_dir):
raise FileNotFoundError(
f"Dataset directory not found at '{data_dir}'. "
f"Please update the 'data_dir' variable to your dataset's path."
)
transform = transforms.Compose([
transforms.Resize((32, 32)),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
# 加载训练集和测试集
trainset = torchvision.datasets.ImageFolder(root=train_dir, transform=transform)
testset = torchvision.datasets.ImageFolder(root=test_dir, transform=transform)
# 从训练数据集中自动获取类别名称和数量
classes = trainset.classes
num_classes = len(classes)
print(f"从数据集中找到 {num_classes} 个类别: {classes}")
# 创建数据加载器
trainloader = torch.utils.data.DataLoader(trainset, batch_size=64, shuffle=True, num_workers=2)
testloader = torch.utils.data.DataLoader(testset, batch_size=64, shuffle=False, num_workers=2)
# --- MODIFICATION 3: 动态调整CNN模型以适应你的数据集 ---
class SimpleCNN(nn.Module):
def __init__(self, num_classes): # 将类别数量作为参数传入
super(SimpleCNN, self).__init__()
self.conv1 = nn.Conv2d(3, 32, kernel_size=3, padding=1)
self.conv2 = nn.Conv2d(32, 64, kernel_size=3, padding=1)
self.conv3 = nn.Conv2d(64, 128, kernel_size=3, padding=1)
self.pool = nn.MaxPool2d(2, 2)
# 输入特征数128 * 4 * 4取决于输入图像大小和网络结构。
# 由于我们将所有图像调整为32x32，经过3次2x2的池化后，尺寸变为 32 -> 16 -> 8 -> 4。所以这里是4*4。
self.fc1 = nn.Linear(128 * 4 * 4, 512)
# **重要**: 输出层的大小现在由num_classes决定
self.fc2 = nn.Linear(512, num_classes)
def forward(self, x):
x = self.pool(F.relu(self.conv1(x)))
x = self.pool(F.relu(self.conv2(x)))
x = self.pool(F.relu(self.conv3(x)))
x = x.view(-1, 128 * 4 * 4)
x = F.relu(self.fc1(x))
x = self.fc2(x)
return x
# 初始化模型，传入你的数据集的类别数量
model = SimpleCNN(num_classes=num_classes)
print("模型已创建")
# 如果有GPU则使用GPU
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = model.to(device)
# 训练模型函数 (现在使用传入的trainloader)
def train_model(model, trainloader, epochs=5): # 增加训练周期以获得更好效果
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
print("开始训练...")
for epoch in range(epochs):
running_loss = 0.0
for i, data in enumerate(trainloader, 0):
inputs, labels = data
inputs, labels = inputs.to(device), labels.to(device)
optimizer.zero_grad()
outputs = model(inputs)
loss = criterion(outputs, labels)
loss.backward()
optimizer.step()
running_loss += loss.item()
if i % 100 == 99:
print(f'[{epoch + 1}, {i + 1:5d}] 损失: {running_loss / 100:.3f}')
running_loss = 0.0
print("训练完成")
# 定义模型保存路径
model_save_path = 'my_custom_cnn.pth'
# 尝试加载预训练模型
try:
model.load_state_dict(torch.load(model_save_path))
print(f"已从 '{model_save_path}' 加载预训练模型")
except FileNotFoundError:
print("无法加载预训练模型，将开始训练新模型。")
train_model(model, trainloader, epochs=5) # 训练新模型
torch.save(model.state_dict(), model_save_path) # 保存训练好的模型
print(f"新模型已训练并保存至 '{model_save_path}'")
# 设置模型为评估模式
model.eval()
# Grad-CAM实现 (这部分无需修改)
class GradCAM:
def __init__(self, model, target_layer):
self.model = model
self.target_layer = target_layer
self.gradients = None
self.activations = None
self.register_hooks()
def register_hooks(self):
def forward_hook(module, input, output):
self.activations = output.detach()
def backward_hook(module, grad_input, grad_output):
self.gradients = grad_output[0].detach()
self.target_layer.register_forward_hook(forward_hook)
self.target_layer.register_backward_hook(backward_hook)
def generate_cam(self, input_image, target_class=None):
model_output = self.model(input_image)
if target_class is None:
target_class = torch.argmax(model_output, dim=1).item()
self.model.zero_grad()
one_hot = torch.zeros_like(model_output)
one_hot[0, target_class] = 1
model_output.backward(gradient=one_hot, retain_graph=True) # retain_graph=True可能需要
gradients = self.gradients
activations = self.activations
weights = torch.mean(gradients, dim=(2, 3), keepdim=True)
cam = torch.sum(weights * activations, dim=1, keepdim=True)
cam = F.relu(cam)
cam = F.interpolate(cam, size=(32, 32), mode='bilinear', align_corners=False)
cam = cam - cam.min()
cam = cam / cam.max() if cam.max() > 0 else cam
return cam.cpu().squeeze().numpy(), target_class
grad_cam = GradCAM(model, model.conv3)
# 从测试集中获取一张图片
img, label = testset[0]
img_tensor = img.unsqueeze(0).to(device)
# 生成CAM
cam, predicted_class_idx = grad_cam.generate_cam(img_tensor)
# 可视化结果
def visualize_cam(img, cam, predicted_class, true_class):
img = img.permute(1, 2, 0).numpy() # 转换回 (H, W, C)
# 反归一化以便显示
img = img * 0.5 + 0.5
img = np.clip(img, 0, 1)
heatmap = plt.cm.jet(cam)
heatmap = heatmap[:, :, :3] # 去掉alpha通道
overlay = heatmap * 0.4 + img * 0.6
plt.figure(figsize=(10, 5))
plt.subplot(1, 3, 1)
plt.imshow(img)
plt.title(f'Original Image\nTrue: {true_class}')
plt.axis('off')
plt.subplot(1, 3, 2)
plt.imshow(heatmap)
plt.title('Grad-CAM Heatmap')
plt.axis('off')
plt.subplot(1, 3, 3)
plt.imshow(overlay)
plt.title(f'Overlay\nPredicted: {predicted_class}')
plt.axis('off')
plt.show()
# 显示结果
predicted_class_name = classes[predicted_class_idx]
true_class_name = classes[label]
visualize_cam(img, cam, predicted_class_name, true_class_name)
