Add sample_weight to eval_metric

[BenEfrati]

Hi,
After #6751 provide a custom eval_metric via fit is deprecated.

In our eval_metric function we call to get_weight() from dmatrix.
When passing the metric function via constructor or set_params method we're unable to get sample_weight.

It might be useful for other sklearn metric function that accept sample_weight argument

[idanmoradarthas]

This is an important ability!

[trivialfis]

Thank you for the new feature! Let's not add an additional parameter and pass the weight when present.

[BenEfrati]

Thank you for the new feature! Let's not add an additional parameter and pass the weight when present.

Agree,
Are you ok with the try except approach?

[trivialfis]

I think we can patch the _metric_decorator as follow:

def _metric_decorator(func: Callable) -> Metric:
    """Decorate a metric function from sklearn.

    Converts a metric function that uses the typical sklearn metric signature so that
    it is compatible with :py:func:`train`

    """

    def inner(y_score: np.ndarray, dmatrix: DMatrix) -> Tuple[str, float]:
        y_true = dmatrix.get_label()
        weight = dmatrix.get_weight()
        if weight.size == 0:
            return func.__name__, func(y_true, y_score)
        else:
            return func.__name__, func(y_true, y_score, sample_weight=weight)

    return inner

and add a test.

[trivialfis]

Could you please help add a test in https://github.com/dmlc/xgboost/blob/master/tests/python/test_with_sklearn.py ?

[BenEfrati]

Could you please help add a test in https://github.com/dmlc/xgboost/blob/master/tests/python/test_with_sklearn.py ?

Sure

[trivialfis]

Thank you for the nice feature!