Normalizing Flow vs Normalizing Flow VAE behavior about normalizing-flows HOT 2 CLOSED

Hi @maulberto3,

in this package, flows are defined as maps from the latent to the observation space. To compute the forward KL divergence, you have to map the observations to the latent space, i.e. apply the inverse.
On the other side, in a variational autoencoder you sample from the latent space (prior) and transform it with the flow layers, i.e. us the forward direction.

In any case, the error is not related to the fact whether you use the forward and the inverse map. Probably, it is due to a misspecification in the flow layer, e.g. that certain parameters that you use to initialize the flow layer do not have the correct shape.
I'll close the issue for now, but if this does not resolve your problem, feel free to open it and add more details about what you are doing and in which context the error occurs.

Best regards,
Vincent

from normalizing-flows.

The bug has been fix, see this issue.

from normalizing-flows.