I’m using az.ess to find out the effective sample size of pymc3 variables. I noticed that some variables have been transformed during sampling and I’m interested in the effective sample size for variables in the transformed space. For instance, these variales returned by model.free_RVs:
However, it seems that az.ess by default only returns the effective sample size in the original space.
But in fact the trace returned by pymc3 containes both transformed and untranformed versions.
I’m not sure why the transformed versions are not contained in the arviz result.
So I’m wondering if there’s a way for arviz to return the effective sample size for the transformed variables? Thank you so much!