VaDE-pytorch icon indicating copy to clipboard operation
VaDE-pytorch copied to clipboard

Published 20 hours ago • verified publisher icon GuHongyang

multivariate Gaussian distribution

Open ThibaultGROUEIX opened this issue 5 years ago • 8 comments

Hi @GuHongyang , Thanks for the reimplementation! they mention in the paper that As for the generative process in Section 3.1, the multivariate Bernoulli distribution is used for MNIST dataset, and the multivariate Gaussian distribution is used for the others . In your code, you've implemented the multivariate Gaussian distribution right? Sincerely, Thibault

ThibaultGROUEIX avatar May 31 '20 09:05 ThibaultGROUEIX

My setting is the same as the original paper. As you can see in the Decoder of model.py, the output is the parameters of the bernoulli distribution.

GuHY777 avatar May 31 '20 09:05 GuHY777

Thanks ! Sorry i didn't dare dive in the details of the ELBO loss, i am not familiar with it and I would need comments to follow through. I just saw a lot of gaussians in there so I assumed you were using a multivariate Gaussian distribution. Thanks again

ThibaultGROUEIX avatar May 31 '20 10:05 ThibaultGROUEIX

Related question: I am trying the code on AFFNIST and it systematically crashes. It seems that z_sigma_2log becomes too large here which gives NaN after the exponential. Do you have an idea why that might happen? Thanks again.

ThibaultGROUEIX avatar May 31 '20 10:05 ThibaultGROUEIX

hi, I run your code in my computer, but I didn't get the good results as your figure (acc about 81%), I'm not sure whether I need to reset the parameters?

huyong1369 avatar Jun 09 '20 00:06 huyong1369

hi, I run your code in my computer, but I didn't get the good results as your figure (acc about 81%), I'm not sure whether I need to reset the parameters?

Hi, when I set the hid_dim as 50, the accuracy I got was is worse. I don't know what happened here? Any ideas?

goodman1204 avatar Jul 04 '20 02:07 goodman1204

@huyong1369 I'm running related experiments using this codes. The average accuracy is about 81% for mnist. Try to repeat the experiment to achieve better results.

wzm2256 avatar Jul 13 '20 02:07 wzm2256

Related question: I am trying the code on AFFNIST and it systematically crashes. It seems that z_sigma_2log becomes too large here which gives NaN after the exponential. Do you have an idea why that might happen? Thanks again.

Hello, this happened to me today. Were you able to solve this issue? Thank you.

ferb2015 avatar Nov 14 '20 08:11 ferb2015

No I did not solve it. As you can see from table 2 here, we reported the divergence on problem on Affnist, but the code ran on our other toy datasets.

ThibaultGROUEIX avatar Nov 14 '20 16:11 ThibaultGROUEIX