Closed baodingge closed 6 years ago
Hi @baodingge Yes, you're right - you'll need to modify the gradient for batch normalization. We have included a case for layer normalization but not for batch normalization.
However, we shuould have a similar case for batch normalization too - one that simply passes back the gradient that comes in from the succeeding layer to the preceding one. If you do make a change in tf_cnnvis, send us a PR and we'll merge or, we'll add the required case in a few days.
Also, another plan would be to try out reconstruction without any change to the gradients, which would be a bit grey, but my guess is that it should work fine too. Let us know how it works out.
Thanks!
Yes, I use the code without change the code, it also works well, and it looks a bit grey,but it is useful.
Alright. Thanks for the update. Applying Histogram equalization on the above image might get you an even better output.
my network was trained from scratch, so i use some BN layer.Should I modify the BN layer gradient?