Saliency Maps 简介——基础梯度

k^* = \arg \max_{i \in [1,K]} f^i(X)

G = \nabla_X f^{k^}(X) = \frac{\partial f^{k^}(X)}{\partial X}

1import torch
2import torchvision
3import torchvision.transforms as T
4from torchvision.models import mobilenet_v2
5from PIL import Image
6import json
7import matplotlib.pyplot as plt
8
9class_idx = json.load(open("imagenet_class_index.json", 'r')) # ImageNet ID 到标签的映射
10idx2label = [class_idx[str(k)][1] for k in range(len(class_idx))] 
11
12image = Image.open("dog.jpg").resize((224,224)) # 原始图像
13transform = T.Compose([
14    T.Resize((224,224)),
15    T.ToTensor(),
16    T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
17])
18
19image = transform(image) # 对图像进行变换和归一化
20image.requires_grad = True 
21torchvision.utils.save_image(image, "./output/image.png", normalize=True)
22
23model = mobilenet_v2() # 加载 MobileNetV2 模型
24ckpt = torch.load("./mobilenet_v2-b0353104.pth", map_location="cpu")
25model.load_state_dict(ckpt)
26model.eval()
27
28logits = model(image.unsqueeze(0))[0] # 输出 logits，前向传播
29prediction = torch.argmax(logits) # 网络预测结果
30print("模型预测: ", idx2label[prediction.item()])
31
32grad = torch.autograd.grad(logits[prediction.item()], [image])[0] # 反向传播
33grad, _ = torch.max(grad.relu(), 0) # 仅保留正值梯度
34
35plt.imshow(grad, cmap=plt.cm.hot) # 绘制显著性图
36plt.axis("off")
37plt.tight_layout()
38plt.savefig("./output/saliency.png", bbox_inches="tight")