Optimize categorical crossentropy gradient update
NicolasHug opened this issue · 3 comments
NicolasHug commented
The gradient and hessian update of the categorical crossentropy loss computes p_k k times, but it only needs to compute it once (see scikit-learn/scikit-learn@9e68984 which led to serious improvement in fit time).
ogrisel commented
+1 as well.
aditya1702 commented
@NicolasHug Can I also work on this project?
ogrisel commented
@aditya1702 please feel free to submit a PR.