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
I have a custom multi-dimensional distribution where the support may be truncated along some dimensions. In terms of constraints, some dimensions will either be real, greater_than, less_than, or interval. I naively was then implementing the support as, e.g.:
Right now, this is not really supported by the numpyro.distributions.constraints.Interval class because of how feasible_like() works, or how the scale is computed in the unconstrained transform. Would you be open to making these things inf-safe? So far I instead implemented a custom subclass InfSafeInterval(constraints._Interval) to support this, but thought I would check in on this. Thanks!
The text was updated successfully, but these errors were encountered:
Hi @adrn, for parameters with different domains, it is better to split them out, e.g. ivl0, ivl1, ivl2 in your case. We don't plan to support mixed-domain support.
Thanks for the response! That makes sense. But hm, it's possible what I want to do is not supported at the moment. For my custom distribution, some pairs of the parameters are not independent and so I can't split them out easily. And some may have real support, some interval, and some greater_than/less_than. This is probably a hack, but I ended up implementing a custom Constraint and Transform that can handle the joint and independent parameters and their supports separately...
Hello!
I have a custom multi-dimensional distribution where the support may be truncated along some dimensions. In terms of constraints, some dimensions will either be
real
,greater_than
,less_than
, orinterval
. I naively was then implementing thesupport
as, e.g.:Right now, this is not really supported by the
numpyro.distributions.constraints.Interval
class because of howfeasible_like()
works, or how thescale
is computed in the unconstrained transform. Would you be open to making these things inf-safe? So far I instead implemented a custom subclassInfSafeInterval(constraints._Interval)
to support this, but thought I would check in on this. Thanks!The text was updated successfully, but these errors were encountered: