search

ratschlab / SVGP-VAE

Tensorflow implementation for the SVGP-VAE model.
MIT License
21 stars 4 forks source link

A follow-up question for the issue 2 about the CE_term #3

Closed JohnD-I closed 1 year ago

JohnD-I commented 1 year ago

Hi， Is it going to minimize the cross-entropy between (latent) sparse GP posterior and (rearranged) VAE approximate posterior during the optimization? Because gauss_cross_entropy function seems to return the negative CE of the two posteriors. Another related question is the benefits for adding this ce_term in optimization, could we simply remove this term in the loss and will the results be very different? Best wishes John

JohnD-I commented 1 year ago

My question was solved with carefully reading this paper: The Gaussian Process Prior VAE for Interpretable Latent Dynamics from Pixels