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
It would be nice, in principle, to define chainrules for the chebpoly constructors, so that we support computation of sensitivities of e.g. polynomial evaluations to the data used to construct the polynomial.
In principle Zygote can do this for us, but it seems to be incredibly slow. (Before we implemented the rrule in #9, it was taking over an hour to backpropagate through a ChebPoly evaluation.)
There are a couple of pieces to implement here. First, chain rules for chebinterp and chebregression, which should be very straightforward since they are linear operations. Second, we need to fix the @notimplemented in the rrule for ChebPoly evaluation (#9) to optionally compute the tangent for sensitivity to polynomial coefficients and bounds — this will need a "thunk" tangent of some sort in order to defer the computation until/unless it is actually needed, because the coefficient tangent is both large and expensive to compute.
The text was updated successfully, but these errors were encountered:
It would be nice, in principle, to define chainrules for the chebpoly constructors, so that we support computation of sensitivities of e.g. polynomial evaluations to the data used to construct the polynomial.
In principle Zygote can do this for us, but it seems to be incredibly slow. (Before we implemented the
rrule
in #9, it was taking over an hour to backpropagate through aChebPoly
evaluation.)There are a couple of pieces to implement here. First, chain rules for
chebinterp
andchebregression
, which should be very straightforward since they are linear operations. Second, we need to fix the@notimplemented
in therrule
forChebPoly
evaluation (#9) to optionally compute the tangent for sensitivity to polynomial coefficients and bounds — this will need a "thunk" tangent of some sort in order to defer the computation until/unless it is actually needed, because the coefficient tangent is both large and expensive to compute.The text was updated successfully, but these errors were encountered: