different results based on grad_mnist

[lee-wt]

Hi, ansuini!
Thanks for your work very much! I learned a lot through your paper and your codes. But I came across a problem in running the source code. As the image uploaded shows, I got a different result in grad_mnist where there is an initial increase in the first layers. While the ID of output layer is not equal to th original mnist. In my opinion, this result accords with your mention in the paper that the first layers prune the irrelevant features such as the luminance perturbation. So why may my result be different from yours?
Sincerely looking forward to your reply.