I wanted to use pymc.sample(…, init=“advi+adapt_diag”) to sample from a complicated posterior, but encountered convergence problems in the advi optimization similar to this post: https://discourse.pymc.io/t/convergence-problems-with-advi/1103
As in the linked post, changing the obj_optimizer to pymc.adadelta solves the convergence issue when calling advi alone with advi.fit(). But to get an accurate posterior, I need to use NUTS initialized by advi, not advi alone. Is there a way to specify the obj_optimizer when calling pymc.sample with init=“advi+adapt_diag”?
