Evaluating Variational Inference: Does rstan expose VI approx density?

Hi all,

I am exploring ADVI and in this regard I wanted to use the PSIS diagnostic introduced by @yuling, @avehtari et al. in Yes, but did it work?: Evaluating variational inference.

Using their notation of algorithm 1, I was wondering whether there is a method / object that allows me to access the fitted (log) VI approximation density for the samples, that is q(\theta_s)? Or did you guys manually write a function for the log of the, say, mean field VI density at fitted parameters?

Cheers!

@yuling has a branch for CmdStan


This adds log q to the csv output.

I hope we can get the diagnostics in CmdStan develop branch and in Rstan in next couple weeks (and before StanCon Helsinki).

2 Likes

Has this been added to CmdStan 2.18. in the meantime?

Another related question: Is the PSIS diagnostic applicable for variational approximations when working with right censored data and a corresponding likelihood?

2.18 has already been tagged, and it’s not in there. Maybe in 2.19 @yuling?

PSIS diagnostic for VI and ADVI do not know anything about likelihood, they have access only to the joint log density (target).

1 Like

No

It would be very handy to have code for this in R, too, given a stanfit-class object that gets returned by vb. If I wanted to code this, do I have all the ingredients available/exposed through rstan (e.g. the mean and standard deviation of each Gaussian factor)?

No. The (co)variance(s) of VB does not get returned in the stanfit object currently.

1 Like