Obtaining log probability from ADVI


I’m curious if there is an out-of-the-box way of evaluating log probability from ADVI or other variational inference Approximation objects, or if there is appetite to make this available.

I’m asking because I’d like to implement some of the checks made in (Yes, but Did It Work?: Evaluating Variational Inference)[http://www.stat.columbia.edu/~gelman/research/published/Evaluating_Variational_Inference.pdf].


You can take a look at https://github.com/junpenglao/Planet_Sakaar_Data_Science/blob/master/Ports/Simulation%20Based%20Calibration.ipynb