shuyuezhuo
commented
5 years ago Sorry, I realised I messed up the mean varience normalization. Now its working fine with above settings.
Thank you again for your code, it is very helpful! Wish you all the best!
Hi, thank you for uploading the code! I am trying to reimplement the original design and I came across your code, it was big help for me.
By referring to your code, I rebuilded the code from Adain repo, I added the SANet module and tuned the parameters according to the original Adain paper and this issue . With regard to the loss weigth, I set it to sw = 3, cw = 1, l1 = 50, l2 = 1, because in that issue post dypark86 pointed out at last that this was the right set up. However, after I trained the network, the evaluation output was very strange, color and form are distored. But when I evaluate with the network model downloaded from your repo, it works fine. Therefore, I am wondering if this weight setting is correct, otherwise it might be some bugs in my code. Many thanks, cheers!