lightgbm_with_custom_loss.py

"""Modified version of Driverless AI's internal LightGBM implementation with a custom objective function (used for tree split finding).
"""
from h2oaicore.models import BaseCustomModel, LightGBMModel
import numpy as np


# The custom objective function will be pickled along with the underlying LightGBM model for persistance purposes
# as a result it can't a lambda function or a method of the custom model object
# The only option is to make the function global in the following manner
def custom_asymmetric_objective(y_true, y_pred):
    """Asymetric MSE loss
    A custom loss has to return the gradient and the hessian of the loss or objective function
    This is not an evaluation but the loss LightGBM will optimize during training
    """
    residual = (y_true - y_pred).astype("float")
    grad = np.where(residual < 0, -2 * 10.0 * residual, -2 * residual)
    hess = np.where(residual < 0, 2 * 10.0, 2.0)
    return grad, hess


class MyLGBMAsymMSE(BaseCustomModel, LightGBMModel):
    """Custom model class that re-uses DAI LightGBMModel
    The class inherits :
      - BaseCustomModel that really is just a tag. It's there to make sure DAI knows it's a custom model and not
      its inner LightGBM Model
      - LightGBMModel object so that the custom model inherits all the properties and methods, especially for params
      mutation
    """
    # The loss is a regression loss
    _regression = True
    # The loss is not for classification
    _binary = False
    _multiclass = False  # WIP
    _mojo = True
    # Give the display name and description that will be shown in the UI
    _display_name = "MyLGBMAsymMSE"
    _description = "LightGBM with custom asymetric loss/objective"
    _is_reproducible = False  # not be exactly reproducible on GPUs
    _testing_can_skip_failure = False  # ensure tested as if shouldn't fail

    def set_default_params(self,
                           accuracy=None, time_tolerance=None, interpretability=None,
                           **kwargs):
        # Define the global loss
        # global custom_asymmetric_objective

        # First call the LightGBM set_default_params
        # This will input all model parameters just like DAI would do.
        LightGBMModel.set_default_params(
            self,
            accuracy=accuracy,
            time_tolerance=time_tolerance,
            interpretability=interpretability,
            **kwargs
        )
        # Now we just need to tell LightGBM that it has to optimize for our custom objective
        # And we are done
        self.params["objective"] = custom_asymmetric_objective

    def mutate_params(
            self, get_best=False, time_tolerance=10, accuracy=10, interpretability=1,
            imbalance_ratio=1.0,
            train_shape=(1, 1), ncol_effective=1,
            time_series=False, ensemble_level=0,
            score_f_name: str = None, **kwargs):
        # If we don't override the parent mutate_params method, DAI would have the opportunity
        # to modify the objective and select the winner
        # For demonstration purposes we purposely make sure that the objective
        # is the one we want
        # So first call the parent method to mutate parameters
        super().mutate_params(
            get_best=get_best, time_tolerance=time_tolerance, accuracy=accuracy,
            interpretability=interpretability,
            imbalance_ratio=imbalance_ratio, train_shape=train_shape, ncol_effective=ncol_effective,
            time_series=time_series, ensemble_level=ensemble_level,
            score_f_name=score_f_name, **kwargs)
        # Now set the objective
        self.params["objective"] = custom_asymmetric_objective