search

liruihui / SP-GAN

MIT License
111 stars 25 forks source link

Discriminator and per-point scores #3

Open yushiangw opened 3 years ago

yushiangw commented 3 years ago

Hi, thanks for sharing your implementation. It looks like the per point discriminator loss described in the main paper is disabled. In my understanding, the output will be (N+1)x1. I am wondering that this is intended or?

https://github.com/liruihui/SP-GAN/blob/main/Generation/Discriminator.py#L110

YanhaoZhang commented 3 years ago

Hi @yushiangw I got the same question here. May I know if you have figure it out? Many thanks.

ShahiraAbousamra commented 2 years ago

Hi @yushiangw and @YanhaoZhang, I have same question. From what I see both the max pool and average pool are over all the points. There are no point-wise scores in the discriminator. Were you able to figure it out? Thank you!

liruihui commented 2 years ago

Thanks for pointing out this issue.

After several checking, we found that the per-point discriminator loss tends to result in unstable training.
Without it, the performance and result are actually slightly better. So I suggest removing it.