schlumberger / joint-vae Goto Github PK

Hi Emilien, it is really a nice work! However, I saw in your paper that you have done the disentanglement metric experiment proposed by FatorVAE paper. Could you please also share the codes for the experiment? I think that would be very helpful, since I did not find any codes on that experiement online. Thanks in advance!

Hi Emilien,

Thanks for open sourcing this really clean code. I'm trying to replicate your results and understand disentangling in VAE. Although it's a pleasure to work with this library, there's a minor detail that I cannot find information about?

In the dataloaders for chairs you say that you centre cropped the images, then resized them to 64*64. Do you chose the same centre cropping for all of the images, if so what cropping?

I tried looking for this information in many other papers, but no one seems to explain what they actually do, which makes me think that it's not that important.

Cheers 💯

From the code, I see that the analytic KL term is used in the ELBO, which does not yield a lower bound for the original likelihood objective. The authors of the CONCRETE paper suggest to replace these discrete mass functions with Concrete densities, which is theoretically more appealing.

I was just wondering if you did indeed try both, and came to the conclusion that the former is more stable, or was it a random decision?