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
Let's revisit the PROBE paper and double-check if the hinge-loss function is correctly defined. For consistency, I also think invalidation rate should not be a separate convergence criterium but handled through generator conditions.
double-check PROBE loss function (see first observation below)
Sort out nested AD problem
invalidation rate as generator condition
Observations
Currently gradients are simply not computed with respect to the Hinge loss. The desired invalidation rate is attained simply as a by-product of continuing the standard counterfactual search.
Let's revisit the PROBE paper and double-check if the hinge-loss function is correctly defined. For consistency, I also think invalidation rate should not be a separate convergence criterium but handled through generator conditions.
Observations
Enzyme
orDifferentiationInterface
for autodiff #300 (see here and here for relevant threads)The text was updated successfully, but these errors were encountered: