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
With the changes to AD handling, we are now in the position that a user could assign HMC to invalid nodes. For example, dcat and other discrete nodes or to nodes that have truncated or constraint dependencies.
At the moment we (as of AD-stoch-ind branching being merged into nimble`) simply have a run-time error when we run the HMC.
I think that in assigning NUTS samplers, we should check for invalid cases. I think these are:
discrete nodes
nodes assigned truncated distributions
nodes with dependencies of truncated distributions, dinterval, dconstraint
user-defined distribution without derivative support (I haven't thought through how difficult this would be)
@paciorek Yes, I think that makes sense about the cases to catch. In the HMC samplers we already were checking for discrete nodes, but not the other cases.
And @perrydv is there a way to determine if a user-defined distribution has AD support?
With the changes to AD handling, we are now in the position that a user could assign HMC to invalid nodes. For example,
dcat
and other discrete nodes or to nodes that have truncated or constraint dependencies.At the moment we (as of
AD-stoch-ind
branching being merged into nimble`) simply have a run-time error when we run the HMC.I think that in assigning NUTS samplers, we should check for invalid cases. I think these are:
dinterval
,dconstraint
@danielturek @perrydv Thoughts?
The text was updated successfully, but these errors were encountered: