Skip to content

关于噪声处理的问题? #13

Open

Description

@Nonbiuld

训练生成的噪声的数值是有正有负的,下面这个操作把负的值都变成0了,这种做法不会有问题吗?不应该是噪声加图像再clamp到0-255?
self.perturb_tensor.mul(255).clamp_(0, 255).permute(0, 2, 3, 1).to('cpu').numpy()

Activity

  1. HanxunH commented on Nov 24, 2022

    @HanxunH
    Owner

    Yes, the noise should be added first. This could improve the protection performance. Keep me updated if you notice any difference in the results.

  2. Nonbiuld commented on Nov 24, 2022

    @Nonbiuld
    Author

    训练生成噪声的代码似乎也有一点小问题,images[i] = images[i] + sample_noise,比如这里这个image加了sample_noise之后,是不是需要再clamp到(0,1)会更加合理一点,如果更改后效果有所提升我会告诉你的。

  3. HanxunH commented on Nov 24, 2022

    @HanxunH
    Owner

    Good catch, thanks!

  4. pzs19 commented on Mar 23, 2023

    @pzs19

    @Nonbiuld 老哥,我也注意到这个问题了,请问更改之后效果如何?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions