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
HanxunH commentedon Nov 24, 2022
Yes, the noise should be added first. This could improve the protection performance. Keep me updated if you notice any difference in the results.
Nonbiuld commentedon Nov 24, 2022
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。
HanxunH commentedon Nov 24, 2022
Good catch, thanks!
pzs19 commentedon Mar 23, 2023
@Nonbiuld 老哥,我也注意到这个问题了,请问更改之后效果如何?