wiseodd
Gibbsnet issue
I've been running the new gibbsnet implementation, but I can't find it to converge with the given settings.
Do you have any examples on what the results should look like?
My point on that implementation is to get the intuition of GibbsNet, and how to quickly implement it, esp. given ALI/BiGAN implementation. Given this intuition feel free to experiment with another architecture and hyperparams.
Cheers, Agustinus
-------- Original Message -------- Subject: [wiseodd/generative-models] Gibbsnet issue (#44) Local Time: December 25, 2017 12:32 PM UTC Time: December 25, 2017 11:32 AM From: [email protected] To: wiseodd/generative-models [email protected] Subscribed [email protected]
I've been running the new gibbsnet implementation, but I can't find it to converge with the given settings.
Do you have any examples on what the results should look like?
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.
Yes, I understand.
I just wanted to point out that it seems like this doesn't converge though.
After 2000 iterations:
The loss of the generator is very high (18) and discriminator loss goes to zero. Have you gotten it to work with your code? I am using the code as is, no modifications.
Yeah, it's also what I experienced. It won't generate good result and the loss is blowing up. I suspect we need to have stronger networks + much lower learning rate.
However, as I'd like to have everything in this repo with very simple architecture, I decided to push it as is :)
-------- Original Message -------- Subject: Re: [wiseodd/generative-models] Gibbsnet issue (#44) Local Time: December 25, 2017 5:27 PM UTC Time: December 25, 2017 4:27 PM From: [email protected] To: wiseodd/generative-models [email protected] Agustinus Kristiadi [email protected], Comment [email protected]
Yes, I understand. I just wanted to point out that it seems like this doesn't converge though. After 2000 iterations: 200
The loss of the generator is very high (18) and discriminator loss goes to zero. Have you gotten it to work with your code? I am using the code as is, no modifications.
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.
I tried an 8 layer MLP for each network, as suggested in the paper with lr of 1e-5. Still no convergence. Would be nice to get a version working. If I make any progress I'll update.