VAE vs autoencoder

[axu-git]

If I run a RHVAE with encoder = decoder = none, it seems from the documentation that encoder = decoder = MLP. In this case, is it true that this is just an autoencoder (and not a VAE) because there is no distribution?

Thanks!

[clementchadebec]

No, it is still a VAE since the MLPs will provide in this case the parameters of the distributions. Mean and covariance of the posterior q(z|x) are given by the encoder while the mean of p(x|z) is given by the decoder.

[axu-git]

So we still have q(z | x) and p(x | z) are each Gaussians even if encoder = decoder = none?

[clementchadebec]

yes, this is correct

[axu-git]

Great thanks!