Custom loss function

[ogencoglu]

Ability to define arbitrary loss function just like in lightgbm would be great. Any such plans?

[deadsoul44]

It is part of the plan, but this might take a while. If you need any specific loss function, let me know.

[ogencoglu]

The ability to define different custom loss functions for different subsets of the training data would be my use case. I can do this with lightgbm (a bit hack-ish but gets the job done). My loss function is not constant as it is part of an optimization framework. Obviously gradients and Hessians should be available but there are frameworks for approximating those.