python打卡第47天

昨天代码中注意力热图的部分顺移至今天

知识点回顾：

热力图

作业：对比不同卷积层热图可视化的结果

def visualize_attention_map(model, test_loader, device, class_names, num_samples=3):"""可视化模型的注意力热力图，展示模型关注的图像区域参数:model: 训练好的PyTorch模型test_loader: 测试数据加载器device: 计算设备 (cuda/cpu)class_names: 类别名称列表num_samples: 要可视化的样本数量"""# 反标准化参数 (CIFAR10)mean = torch.tensor([0.4914, 0.4822, 0.4465]).view(1, 3, 1, 1)std = torch.tensor([0.2023, 0.1994, 0.2010]).view(1, 3, 1, 1)model.eval()  # 设置为评估模式fig, axes = plt.subplots(num_samples, 4, figsize=(16, 4*num_samples))if num_samples == 1:axes = axes.reshape(1, -1)  # 确保单样本时保持2D结构with torch.no_grad():for sample_idx, (images, labels) in enumerate(test_loader):if sample_idx >= num_samples:breakimages, labels = images.to(device), labels.to(device)# 创建钩子捕获特征图activation = []hook = model.conv3.register_forward_hook(lambda module, input, output: activation.append(output.detach()))# 前向传播outputs = model(images)hook.remove()  # 立即移除钩子# 获取预测结果_, preds = torch.max(outputs, 1)# 反标准化图像denorm_img = images.cpu() * std + meanimg_np = denorm_img[0].permute(1, 2, 0).numpy()img_np = np.clip(img_np, 0, 1)# 获取特征图feat_map = activation[0][0].cpu()  # 第一个样本# 计算通道重要性 (全局平均池化)channel_weights = torch.mean(feat_map, dim=(1, 2))topk_indices = torch.topk(channel_weights, k=3).indices# 显示原始图像ax = axes[sample_idx, 0]ax.imshow(img_np)ax.set_title(f"Original\nTrue: {class_names[labels[0]]}\nPred: {class_names[preds[0]]}")ax.axis('off')# 显示前3个重要通道的热力图for idx, channel_idx in enumerate(topk_indices, start=1):# 处理单通道特征图channel_heatmap = feat_map[channel_idx].numpy()# 归一化并缩放到原图尺寸channel_heatmap = (channel_heatmap - channel_heatmap.min()) / (channel_heatmap.max() - channel_heatmap.min() + 1e-8)resized_heatmap = cv2.resize(channel_heatmap, img_np.shape[:2][::-1])# 创建叠加可视化ax = axes[sample_idx, idx]ax.imshow(img_np)ax.imshow(resized_heatmap, alpha=0.5, cmap='jet')ax.set_title(f"Channel {channel_idx.item()} Attention")ax.axis('off')plt.tight_layout()plt.show()# 调用示例
visualize_attention_map(model, test_loader, device, class_names, num_samples=3)

 @浙大疏锦行