Hi maybe some bug in self-attention block

[Johnson-yue]

Hi, does your code has good performance with Self_attention??

I think your implement about self-attention is something wrong ,because basically self-attention have 4 convolutions( f, g, h, and v) you dont have v!!

and according official tensorflow version code , you have not max_pool followed by g()

[4m4n5]

Hey. You might be right I didn’t really look into it that much because the results were quite good regardless. I wrote the code when the SAGAN paper was pretty new. If you can spare some time to submit a PR, I’d totally merge it.

[Johnson-yue]

Hi, I check SAGAN paper again , and I think your implement is v1 code for v1 paper, but I read v2 paper and code, So ,it is different.

and did you get good performance ?Because I can't get it for my implement and cifar10 dataset