真实标签一致性——特征匹配

FA_{k}(S, S_g) = \frac{ \vert \text{topK}(S, k) \cap \text{topK}(S_g, k) \vert }{k}

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
8import numpy as np
9from utils import *
10
11k = 100
12
13def get_feature_agreement(attrA, attrB, k):
14    attrA = attrA.detach().numpy().reshape(-1)
15    attrB = attrB.detach().numpy().reshape(-1)
16
17    topk_idA = np.argsort(-np.abs(attrA))[0:k]
18    topk_idB = np.argsort(-np.abs(attrB))[0:k]
19
20    topk_setsA = set(topk_idA)
21    topk_setsB = set(topk_idB)
22    
23    # 检查：相同 ID 的比例
24    metric_distr = len(topk_setsA.intersection(topk_setsB)) / k
25    return metric_distr
26
27image = Image.open("dog.jpg").resize((224, 224))  # 原始图像
28transform = T.Compose([
29    T.Resize((224, 224)),
30    T.ToTensor(),
31    T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
32])
33
34image = transform(image)  # 变换并归一化
35
36model = mobilenet_v2()  # 加载 MobileNetV2 模型
37ckpt = torch.load("./mobilenet_v2-b0353104.pth", map_location="cpu")
38model.load_state_dict(ckpt)
39model.eval()
40
41integrated_grad = get_integrated_gradients(model, image, num_steps=20)
42smooth_grad = get_smooth_gradients(model, image, num_samples=20, alpha=0.01)
43vanilla_grad = get_vanilla_gradients(model, image)
44
45fa_ig_sg = get_feature_agreement(smooth_grad, integrated_grad, k=k)
46print("Top-{} 特征一致性（IG 与 SG 之间）为：".format(k), fa_ig_sg)
47
48fa_ig_vg = get_feature_agreement(vanilla_grad, integrated_grad, k=k)
49print("Top-{} 特征一致性（IG 与 VG 之间）为：".format(k), fa_ig_vg)