You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Based on this forum post, I think a lot of new users of MCMC struggle with using plate handlers to correctly annotate their batch dimensions (which isn't strictly needed for doing inference).
While this isn't necessarily a requirement for MCMC, it becomes important if users would like to use the predictive utility to do predictions after running MCMC. This utility wraps the model in an outermost plate to vectorize predictions. I think we should support a parallel=False argument to this utility which by default will run the model forward on a per trace basis and only use plate if parallel is set to True. This will make it easier for new users to play around with MCMC who may not be well acquainted with our auto batching semantics.
The text was updated successfully, but these errors were encountered:
Based on this forum post, I think a lot of new users of MCMC struggle with using plate handlers to correctly annotate their batch dimensions (which isn't strictly needed for doing inference).
While this isn't necessarily a requirement for MCMC, it becomes important if users would like to use the
predictive
utility to do predictions after running MCMC. This utility wraps the model in an outermost plate to vectorize predictions. I think we should support aparallel=False
argument to this utility which by default will run the model forward on a per trace basis and only useplate
if parallel is set to True. This will make it easier for new users to play around with MCMC who may not be well acquainted with our auto batching semantics.The text was updated successfully, but these errors were encountered: