• How to get gradients with respect to the inputs in pytorch


    This is one way to find adversarial examples of CNN.

    The boilerplate:

    import torch
    from torch.autograd import Variable
    import torch.nn as nn
    import torch.optim as optim
    import numpy as np
    

      Define a simple network:

    class lolnet(nn.Module):
        def __init__(self):
            super(lolnet,self).__init__()
            self.a=nn.Linear(in_features=1,out_features=1,bias=False)
            self.a.weight = nn.Parameter(torch.FloatTensor([[0.6]]))
            self.b=nn.Linear(in_features=1,out_features=1,bias=False)
            self.b.weight=nn.Parameter(torch.FloatTensor([[0.6]]))
            
        def forward(self, inputs):
            return self.b(
                self.a(inputs)
            )
    

      The inputs

    inputs=np.array([[5]])
    inputs=torch.from_numpy(inputs).float()
    inputs=Variable(inputs)
    inputs.requires_grad=True
    net=lolnet()
    

      The optimizer

    opx=optim.SGD(
        params=[
            {"params":inputs}
        ],lr=0.5
    )
    

      The optimization process

    for i in range(50):
        x=net(inputs)
        loss=(x-1)**2
        opx.zero_grad() 
        loss.backward()
        opx.step()
        print(net.a.weight.data.numpy()[0][0],inputs.data.numpy()[0][0],loss.data.numpy()[0][0])
    

      The results are as below:

    0.6 4.712 0.6400001
    0.6 4.4613247 0.4848616
    0.6 4.243137 0.36732942
    0.6 4.0532265 0.27828723
    0.6 3.8879282 0.2108294
    0.6 3.7440526 0.15972354
    0.6 3.6188233 0.1210059
    0.6 3.5098238 0.09167358
    0.6 3.4149506 0.069451585
    0.6 3.332373 0.052616227
    0.6 3.2604973 0.039861854
    0.6 3.1979368 0.030199187
    0.6 3.143484 0.022878764
    0.6 3.0960886 0.017332876
    0.6 3.0548356 0.013131317
    0.6 3.0189288 0.00994824
    0.6 2.9876754 0.0075367615
    0.6 2.9604726 0.005709796
    0.6 2.9367952 0.0043257284
    0.6 2.9161866 0.003277142
    0.6 2.8982487 0.0024827516
    0.6 2.8826356 0.0018809267
    0.6 2.869046 0.001424982
    0.6 2.8572176 0.0010795629
    0.6 2.8469222 0.0008178701
    0.6 2.837961 0.00061961624
    0.6 2.830161 0.00046941772
    0.6 2.8233721 0.000355627
    0.6 2.8174632 0.0002694209
    0.6 2.81232 0.00020411481
    0.6 2.8078432 0.0001546371
    0.6 2.8039467 0.00011715048
    0.6 2.8005552 8.875507e-05
    0.6 2.7976031 6.724081e-05
    0.6 2.7950337 5.093933e-05
    0.6 2.7927973 3.8591857e-05
    0.6 2.7908509 2.9236677e-05
    0.6 2.7891567 2.2150038e-05
    0.6 2.7876818 1.6781378e-05
    0.6 2.7863982 1.2713146e-05
    0.6 2.785281 9.631679e-06
    0.6 2.7843084 7.296927e-06
    0.6 2.783462 5.527976e-06
    0.6 2.7827253 4.1880226e-06
    0.6 2.782084 3.1727632e-06
    0.6 2.7815259 2.4034823e-06
    0.6 2.78104 1.821013e-06
    0.6 2.7806172 1.3793326e-06
    0.6 2.780249 1.044933e-06
    0.6 2.7799287 7.9170513e-07
    
    Process finished with exit code 0
    

      

  • 相关阅读:
    【转】Android Lint分类及常见错误
    备忘
    【转】QRCode二维码生成方案及其在带LOGO型二维码中的应用
    如何控制横向和纵向滚动条的显隐?
    网页屏保
    过度方式
    让背景图不滚动
    自定义指定区域的文字大小
    定义本网页关键字
    通过层来实现渐淡淡出
  • 原文地址:https://www.cnblogs.com/cxxszz/p/8974640.html
Copyright © 2020-2023  润新知