edward icon indicating copy to clipboard operation
edward copied to clipboard

Published 20 hours ago • verified publisher icon blei-lab

Implementation of SVAE - Johnson, Matthew, et al. NIPS 2016.

Open pauloabelha opened this issue 7 years ago • 9 comments

Hi all,

First, congratulations on Edward: a really nice project!

I wonder if anyone is already working on re-implementing SVAE - Johnson, Matthew, et al. NIPS 2016?

I've been trying to use the original author's code, but seems like a fresh example implementation (pin wheel data, like in the paper for instance) would be nice on Edward /examples.

I would do it, but I'm still learning about VAE - Edward tutorial :) If no one does it, I'll probably be the one giving it a try it in the future.

Best regards, Paulo

pauloabelha avatar Feb 15 '18 15:02 pauloabelha

@mattjj, @matthewdhoffman

SVAEs require a symbolic algebra system that properly accounts for integral identities (conjugacy relationships). We're working on a separate package doing this with Autograd/NumPy; but note it likely won't be ready for at least a few months.

dustinvtran avatar Feb 15 '18 18:02 dustinvtran

Ok. Thank you.

pauloabelha avatar Feb 15 '18 18:02 pauloabelha

Yeah and my public SVAE code kinda sucks. Sorry :)

I have an improved version that I've been threatening to release, but I haven't touched it or the corresponding preprint in many months. That code is an improvement and might serve as a stopgap, but the system Dustin mentioned is really the right way to build SVAEs.

Thanks for your interest! It's great motivation to keep working.

mattjj avatar Feb 17 '18 05:02 mattjj

Ok, Matthew. Thank you for your answer. :)

pauloabelha avatar Feb 19 '18 10:02 pauloabelha

Just out of interest is the symbolic algebra system that properly accounts for integral identities package open source somewhere?

JohnReid avatar Mar 07 '18 10:03 JohnReid

Also are you in interested in a particular structured model @pauloabelha? I would be keen to get a mixture model working. I'm happy to collaborate on this. Are are you waiting for the symbolic algebra system?

JohnReid avatar Mar 09 '18 11:03 JohnReid

Hi @JohnReid , I am most interested in an S-LDS for video clustering/inference. Yes, I am also waiting for the symbolic algebra system, but not sure if I will have enough time to wait on it or try a different approach other than SVAE for my research. Are you currently implementing something related to SVAEs? Best regards,

pauloabelha avatar May 18 '18 14:05 pauloabelha

Hi @pauloabelha, I have to say I found some other models that work for my data so I am not currently implementing anything related to SVAEs. I would be interested but I'm not sure I'll find the time anytime soon. Best regards, John.

JohnReid avatar May 18 '18 14:05 JohnReid

Hi there, Has anyone been working on this? I'm not sure I should try again to do it on my own or something official is about to come out in Edward. Thank you and kind regards,

pauloabelha avatar Jun 30 '18 16:06 pauloabelha