反事实解释

X' = \argmin_{X'} \lambda \cdot (f(X') - y')^2 + d(X, X')

1import torch
2import torch.nn as nn
3import torchvision
4import torchvision.transforms as T
5from torchvision import datasets
6from PIL import Image
7import matplotlib.pyplot as plt
8import sklearn.linear_model as lm
9import numpy as np
10
11NUM_ITERATIONS = 1000 # 采样次数
12LAMBDA = 10
13
14class Network(nn.Module): # 全连接神经网络
15    def __init__(self):
16        super(Network, self).__init__()
17        self.fc1 = nn.Linear(784, 512)
18        self.relu =  nn.ReLU() # ReLU激活函数
19        self.fc2 = nn.Linear(512, 256)
20        self.fc3 = nn.Linear(256, 128,)
21        self.fc4 = nn.Linear(128, 64)
22        self.fc5 = nn.Linear(64, 10) # 输出长度为10的向量
23    def forward(self, x):
24        x =  self.relu(self.fc1(x))
25        x =  self.relu(self.fc2(x))
26        x =  self.relu(self.fc3(x))
27        x =  self.relu(self.fc4(x))
28        x =  self.relu(self.fc5(x))
29        return x
30
31# 加载MNIST测试数据集
32test_data = datasets.MNIST(root = './',  train = False, transform = T.ToTensor(), download=True)
33image = test_data.data[4].float() # 从测试集中选取一个输入样本
34pil_image = T.functional.to_pil_image(image*255) 
35pil_image.resize((224, 224)).save("./output/image.png") 
36
37model = Network() # 加载神经网络模型
38ckpt = torch.load("./mnist_fc.pth", map_location="cpu") # 加载模型权重
39model.load_state_dict(ckpt)
40model.eval()
41
42logits = model(image.view(-1).unsqueeze(0))[0] # 前向传播
43prediction = torch.argmax(logits) # 模型预测结果
44print("原始图像的模型预测结果 : ", prediction.item())
45target = torch.tensor(np.eye(1, 10, prediction.item()))
46
47cf = image.clone()
48cf.requires_grad = True
49
50for i in range(NUM_ITERATIONS):
51  logits = model(cf.view(-1).unsqueeze(0))[0] # 扰动输入的前向传播
52  prediction = torch.argmax(logits) 
53  regularization = torch.sum(torch.abs(cf - image))
54  top2_predictions = torch.topk(logits, 2)
55  loss = nn.functional.relu(top2_predictions[0] - top2_predictions[1])
56  
57  objective = torch.mean(LAMBDA*loss + regularization)
58  grad = torch.autograd.grad(objective, [cf])[0]
59 
60  cf = cf  - 0.01*grad
61  cf = torch.clip(cf, 0., 1.0)
62
63print("反事实样本的模型预测结果 : ", prediction.item())
64cf = torch.nn.functional.interpolate(cf[None, None, :, :], (224,224), mode='bilinear')
65torchvision.utils.save_image(cf.detach(), "./output/saliency.png", normalize=True)