Understand root cause of high memory utilization

[ricardoV94]

You can get draws from the VI fit and compute the log-likelihood from those. If the VI gives a good approximation to the posterior I don’t see why the same metrics wouldn’t be relevant, but I am not familiar with the theory either