bob-carpenter
commented
5 years ago +1 to unifying output across sampling, optimization and variational inference. Even better if we can include the log density for both the approximate and true models so that we could do something like importance sampling to compute expectations (I think that's already in the ADVI output).
Until v3 when we can finally standardize fitted model objects across algorithms (this has been a problem for years, see eg #41), I think we should provide a function in rstan to convert the list returned by optimizing() into a proper stanfit object consistent with vb() and sampling(). This would make a lot of people happy, or at least a bit less frustrated.