Grad-CAM，即梯度加权类激活映射 (Gradient-weighted Class Activation Mapping)，是一种用于解释卷积神经网络决策的方法。它通过可视化模型对于给定输入的关注区域来提供洞察。

原理:

Grad-CAM的关键思想是将输出类别的梯度（相对于特定卷积层的输出）与该层的输出相乘，然后取平均，得到一个“粗糙”的热力图。这个热力图可以被放大并叠加到原始图像上，以显示模型在分类时最关注的区域。

具体步骤如下：

1. 选择一个卷积层作为解释的来源。通常，我们会选择网络的最后一个卷积层，因为它既包含了高级特征，也保留了空间信息。
2. 前向传播图像到网络，得到你想解释的类别的得分。
3. 计算此得分 相对于我们选择的卷积层 输出的梯度。
4. 对于该卷积层的每个通道，使用上述梯度的全局平均值对该通道进行加权。
5. 结果是一个与卷积层的空间维度相同的加权热力图。

优势

Grad-CAM的优点是它可以用于任何卷积神经网络，无需进行结构修改或重新训练。它为我们提供了一个简单但直观的方式来理解模型对于特定输入的决策。

Code

import torch
import cv2
import torch.nn.functional as F
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
from PIL import Imageclass GradCAM:def __init__(self, model, target_layer):self.model = modelself.target_layer = target_layerself.feature_maps = Noneself.gradients = None# Hook layerstarget_layer.register_forward_hook(self.save_feature_maps)target_layer.register_backward_hook(self.save_gradients)def save_feature_maps(self, module, input, output):self.feature_maps = output.detach()def save_gradients(self, module, grad_input, grad_output):self.gradients = grad_output[0].detach()def generate_cam(self, image, class_idx=None):# Set model to evaluation modeself.model.eval()# Forward passoutput = self.model(image)if class_idx is None:class_idx = torch.argmax(output).item()# Zero out gradientsself.model.zero_grad()# Backward pass for target classone_hot = torch.zeros((1, output.size()[-1]), dtype=torch.float32)one_hot[0][class_idx] = 1output.backward(gradient=one_hot.cuda(), retain_graph=True)# Get pooled gradients and feature mapspooled_gradients = torch.mean(self.gradients, dim=[0, 2, 3])activation = self.feature_maps.squeeze(0)for i in range(activation.size(0)):activation[i, :, :] *= pooled_gradients[i]# Create heatmapheatmap = torch.mean(activation, dim=0).squeeze().cpu().numpy()heatmap = np.maximum(heatmap, 0)heatmap /= torch.max(heatmap)heatmap = cv2.resize(heatmap, (image.size(3), image.size(2)))heatmap = np.uint8(255 * heatmap)heatmap = cv2.applyColorMap(heatmap, cv2.COLORMAP_JET)# Superimpose heatmap on original imageoriginal_image = self.unprocess_image(image.squeeze().cpu().numpy())superimposed_img = heatmap * 0.4 + original_imagesuperimposed_img = np.clip(superimposed_img, 0, 255).astype(np.uint8)return heatmap, superimposed_imgdef unprocess_image(self, image):# Reverse the preprocessing stepmean = np.array([0.485, 0.456, 0.406])std = np.array([0.229, 0.224, 0.225])image = (((image.transpose(1, 2, 0) * std) + mean) * 255).astype(np.uint8)return imagedef visualize_gradcam(model, input_image_path, target_layer):# Load imageimg = Image.open(input_image_path)preprocess = transforms.Compose([transforms.Resize((224, 224)),transforms.ToTensor(),transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])input_tensor = preprocess(img).unsqueeze(0).cuda()# Create GradCAMgradcam = GradCAM(model, target_layer)heatmap, result = gradcam.generate_cam(input_tensor)plt.figure(figsize=(10,10))plt.subplot(1,2,1)plt.imshow(heatmap)plt.title('Heatmap')plt.axis('off')plt.subplot(1,2,2)plt.imshow(result)plt.title('Superimposed Image')plt.axis('off')plt.show()# Load your model (e.g., resnet20 in this case)
# model = resnet20()
# model.load_state_dict(torch.load("path_to_your_weights.pth"))
# model.to('cuda')# Visualize GradCAM
# visualize_gradcam(model, "path_to_your_input_image.jpg", model.layer3[-1])

中文注释详细版

import torch
import cv2
import torch.nn.functional as F
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
from PIL import Imageclass GradCAM:def __init__(self, model, target_layer):self.model = model  # 要进行Grad-CAM处理的模型self.target_layer = target_layer  # 要进行特征可视化的目标层self.feature_maps = None  # 存储特征图self.gradients = None  # 存储梯度# 为目标层添加钩子，以保存输出和梯度target_layer.register_forward_hook(self.save_feature_maps)target_layer.register_backward_hook(self.save_gradients)def save_feature_maps(self, module, input, output):"""保存特征图"""self.feature_maps = output.detach()def save_gradients(self, module, grad_input, grad_output):"""保存梯度"""self.gradients = grad_output[0].detach()def generate_cam(self, image, class_idx=None):"""生成CAM热力图"""# 将模型设置为评估模式self.model.eval()# 正向传播output = self.model(image)if class_idx is None:class_idx = torch.argmax(output).item()# 清空所有梯度self.model.zero_grad()# 对目标类进行反向传播one_hot = torch.zeros((1, output.size()[-1]), dtype=torch.float32)one_hot[0][class_idx] = 1output.backward(gradient=one_hot.cuda(), retain_graph=True)# 获取平均梯度和特征图pooled_gradients = torch.mean(self.gradients, dim=[0, 2, 3])activation = self.feature_maps.squeeze(0)for i in range(activation.size(0)):activation[i, :, :] *= pooled_gradients[i]# 创建热力图heatmap = torch.mean(activation, dim=0).squeeze().cpu().numpy()heatmap = np.maximum(heatmap, 0)heatmap /= torch.max(heatmap)heatmap = cv2.resize(heatmap, (image.size(3), image.size(2)))heatmap = np.uint8(255 * heatmap)heatmap = cv2.applyColorMap(heatmap, cv2.COLORMAP_JET)# 将热力图叠加到原始图像上original_image = self.unprocess_image(image.squeeze().cpu().numpy())superimposed_img = heatmap * 0.4 + original_imagesuperimposed_img = np.clip(superimposed_img, 0, 255).astype(np.uint8)return heatmap, superimposed_imgdef unprocess_image(self, image):"""反预处理图像，将其转回原始图像"""mean = np.array([0.485, 0.456, 0.406])std = np.array([0.229, 0.224, 0.225])image = (((image.transpose(1, 2, 0) * std) + mean) * 255).astype(np.uint8)return imagedef visualize_gradcam(model, input_image_path, target_layer):"""可视化Grad-CAM热力图"""# 加载图像img = Image.open(input_image_path)preprocess = transforms.Compose([transforms.Resize((224, 224)),transforms.ToTensor(),transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])input_tensor = preprocess(img).unsqueeze(0).cuda()# 创建GradCAMgradcam = GradCAM(model, target_layer)heatmap, result = gradcam.generate_cam(input_tensor)# 显示图像和热力图plt.figure(figsize=(10,10))plt.subplot(1,2,1)plt.imshow(heatmap)plt.title('热力图')plt.axis('off')plt.subplot(1,2,2)plt.imshow(result)plt.title('叠加后的图像')plt.axis('off')plt.show()# 以下是示例代码，显示如何使用上述代码。
# 首先，你需要加载你的模型和权重。
# model = resnet20()
# model.load_state_dict(torch.load("path_to_your_weights.pth"))
# model.to('cuda')# 然后，调用`visualize_gradcam`函数来查看结果。
# visualize_gradcam(model, "path_to_your_input_image.jpg", model.layer3[-1])

论文链接：https://openaccess.thecvf.com/content_ICCV_2017/papers/Selvaraju_Grad-CAM_Visual_Explanations_ICCV_2017_paper.pdf