Hi everyone, I’ve implemented ADVI (i think) using ReparameterizationKLqp class, in order to approximate the posterior of a Gamma variable, Z.

To do so, I define the input latent variables dictionary as {Z : LogNormal} where the log Normal distribution is devised as a TransformedDistribuition instance.

It works perfectly fine!

I’ve used the LogNormal because if we apply the log transformation to Z and define the variational distribution of the transformed variable as a Normal, the resulting Z’s distribution is a LogNormal

However, both key and value have the same (constrainted) support as such no transformation is performed, so

- how is it working fine, if in ADVI the latent variables should be unconstrained?
- Am I implementing ADVI correctly?
- Or should I pass as input {log(Z): Normal}?