疏锦行Python打卡 DAY 47 注意力热图可视化

# 可视化空间注意力热力图（显示模型关注的图像区域）
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"""
    model.eval()  # 设置为评估模式
    
    with torch.no_grad():
        for i, (images, labels) in enumerate(test_loader):
            if i >= num_samples:  # 只可视化前几个样本
                break
                
            images, labels = images.to(device), labels.to(device)
            
            # 创建一个钩子，捕获中间特征图
            activation_maps = []
            
            def hook(module, input, output):
                activation_maps.append(output.cpu())
            
            # 为最后一个卷积层注册钩子（获取特征图）
            hook_handle = model.conv3.register_forward_hook(hook)
            
            # 前向传播，触发钩子
            outputs = model(images)
            
            # 移除钩子
            hook_handle.remove()
            
            # 获取预测结果
            _, predicted = torch.max(outputs, 1)
            
            # 获取原始图像
            img = images[0].cpu().permute(1, 2, 0).numpy()
            # 反标准化处理
            img = img * np.array([0.2023, 0.1994, 0.2010]).reshape(1, 1, 3) + np.array([0.4914, 0.4822, 0.4465]).reshape(1, 1, 3)
            img = np.clip(img, 0, 1)
            
            # 获取激活图（最后一个卷积层的输出）
            feature_map = activation_maps[0][0].cpu()  # 取第一个样本
            
            # 计算通道注意力权重（使用SE模块的全局平均池化）
            channel_weights = torch.mean(feature_map, dim=(1, 2))  # [C]
            
            # 按权重对通道排序
            sorted_indices = torch.argsort(channel_weights, descending=True)
            
            # 创建子图
            fig, axes = plt.subplots(1, 4, figsize=(16, 4))
            
            # 显示原始图像
            axes[0].imshow(img)
            axes[0].set_title(f'原始图像\n真实: {class_names[labels[0]]}\n预测: {class_names[predicted[0]]}')
            axes[0].axis('off')
            
            # 显示前3个最活跃通道的热力图
            for j in range(3):
                channel_idx = sorted_indices[j]
                # 获取对应通道的特征图
                channel_map = feature_map[channel_idx].numpy()
                # 归一化到[0,1]
                channel_map = (channel_map - channel_map.min()) / (channel_map.max() - channel_map.min() + 1e-8)
                
                # 调整热力图大小以匹配原始图像
                from scipy.ndimage import zoom
                heatmap = zoom(channel_map, (32/feature_map.shape[1], 32/feature_map.shape[2]))
                
                # 显示热力图
                axes[j+1].imshow(img)
                axes[j+1].imshow(heatmap, alpha=0.5, cmap='jet')
                axes[j+1].set_title(f'注意力热力图 - 通道 {channel_idx}')
                axes[j+1].axis('off')
            
            plt.tight_layout()
            plt.show()
 
# 调用可视化函数
visualize_attention_map(model, test_loader, device, class_names, num_samples=3)
import torch
import numpy as np
import matplotlib.pyplot as plt
from scipy.ndimage import zoom
 
def visualize_feature_maps(model, test_loader, device, class_names, num_samples=3, conv_layer_name='conv3'):
    model.eval()
    with torch.no_grad():
        for i, (images, labels) in enumerate(test_loader):
            if i >= num_samples:
                break
            images, labels = images.to(device), labels.to(device)
            activation_maps = []
 
            def hook(module, input, output):
                activation_maps.append(output.cpu())
 
            # 支持任意层名
            layer = dict([*model.named_modules()])[conv_layer_name]
            handle = layer.register_forward_hook(hook)
            outputs = model(images)
            handle.remove()
 
            _, predicted = torch.max(outputs, 1)
 
            img = images[0].cpu().permute(1, 2, 0).numpy()
            img = img * np.array([0.2023, 0.1994, 0.2010]).reshape(1, 1, 3) + np.array([0.4914, 0.4822, 0.4465]).reshape(1, 1, 3)
            img = np.clip(img, 0, 1)
 
            feature_map = activation_maps[0][0]  # shape: [C,H,W]
            channel_weights = torch.mean(feature_map, dim=(1, 2))
            sorted_indices = torch.argsort(channel_weights, descending=True)
 
            fig, axes = plt.subplots(1, 4, figsize=(16, 4))
            axes[0].imshow(img)
            axes[0].set_title(f'原始图像\n真实: {class_names[labels[0]]}\n预测: {class_names[predicted[0]]}')
            axes[0].axis('off')
 
            for j in range(3):
                channel_idx = sorted_indices[j]
                channel_map = feature_map[channel_idx].numpy()
                channel_map = (channel_map - channel_map.min()) / (channel_map.max() - channel_map.min() + 1e-8)
                heatmap = zoom(channel_map, (32/feature_map.shape[1], 32/feature_map.shape[2]))
                axes[j+1].imshow(img)
                axes[j+1].imshow(heatmap, alpha=0.5, cmap='jet')
                axes[j+1].set_title(f'特征图热力图 - 通道 {channel_idx}')
                axes[j+1].axis('off')
            plt.tight_layout()
            plt.show()
import torch
import numpy as np
import matplotlib.pyplot as plt
from scipy.ndimage import zoom
 
def visualize_gradcam(model, test_loader, device, class_names, num_samples=3, conv_layer_name='conv3'):
    model.eval()
    for i, (images, labels) in enumerate(test_loader):
        if i >= num_samples:
            break
        images, labels = images.to(device), labels.to(device)
        activations = []
        gradients = []
 
        def forward_hook(module, input, output):
            activations.append(output)
 
        def backward_hook(module, grad_in, grad_out):
            gradients.append(grad_out[0])
 
        # 支持任意层名
        layer = dict([*model.named_modules()])[conv_layer_name]
        fh = layer.register_forward_hook(forward_hook)
        bh = layer.register_full_backward_hook(backward_hook)
 
        outputs = model(images)
        _, predicted = torch.max(outputs, 1)
        score = outputs[0, predicted[0]]
        model.zero_grad()
        score.backward()
 
        fh.remove()
        bh.remove()
 
        # Grad-CAM权重：全局平均池化梯度
        grad = gradients[0][0].cpu().data.numpy()     # [C,H,W]
        act = activations[0][0].cpu().data.numpy()    # [C,H,W]
        weights = np.mean(grad, axis=(1, 2))          # [C]
 
        # 线性加权求和（Grad-CAM图）
        cam = np.zeros(act.shape[1:], dtype=np.float32)
        for k, w in enumerate(weights):
            cam += w * act[k, :, :]
        cam = np.maximum(cam, 0)   # 只保留正值（ReLU）
 
        # 归一化到[0,1]
        cam = (cam - cam.min()) / (cam.max() - cam.min() + 1e-8)
        cam = zoom(cam, (32/cam.shape[0], 32/cam.shape[1]))
 
        # 反标准化原图
        img = images[0].cpu().permute(1, 2, 0).numpy()
        img = img * np.array([0.2023, 0.1994, 0.2010]).reshape(1, 1, 3) + np.array([0.4914, 0.4822, 0.4465]).reshape(1, 1, 3)
        img = np.clip(img, 0, 1)
 
        plt.figure(figsize=(8,4))
        plt.subplot(1,2,1)
        plt.imshow(img)
        plt.title(f'原图\n真实: {class_names[labels[0]]}\n预测: {class_names[predicted[0]]}')
        plt.axis('off')
 
        plt.subplot(1,2,2)
        plt.imshow(img)
        plt.imshow(cam, alpha=0.5, cmap='jet')
        plt.title('Grad-CAM')
        plt.axis('off')
        plt.tight_layout()
        plt.show()

打卡：@浙大疏锦行