CustomObjective

Base class for writing your own regression loss. Subclass it, then pass an instance as the regressor's loss. See the User Guide: custom objectives and metrics.

Base class for user-defined regression objectives.

Subclass and implement grad_hess(y, raw) -> (gradient, hessian) and eval(y, raw, sample_weight=None) -> scalar (lower is better; drives early stopping). Optionally override init(y, sample_weight=None) (the starting raw score, default 0.0) and transform(raw) (raw scores -> predictions, default identity).

Pass an instance as the regressor's loss. Instances must be stateless across fits and picklable, so define the subclass at module level: bagged members fit in worker processes.

Read more in the User Guide.