Closed Edenhofer closed 7 months ago
That is a nice and easy-to-understand demo! I would just tweak the optimization parameters a bit. I think something like 5 VI iterations should be enough and a sane default to copy to other regression problems. The 12 VI iterations make the demo a bit slow, given that it is only a linear regression.
Great to hear that you like it too! :)
Agreed! Fewer global iterations are probably enough. Note, the model is slightly nonlinear though because the priors are non-Gaussian. (The usual "what is the prior" vs "what is the likelihood" question...)
I think this simple demonstration of a nonlinear regression nicely shows how convenient it is to use NIFTy.re's modeling approach. @roth-jakob WDYT?