A wrapper for functions so that they can be parametrized with get_params and set_params in scikit-learn: proof of concept

paramfunc.py

from collections import defaultdict
import pandas as pd


class parametrized_function:
    def __init__(self, _func, **kwargs):
        self._func = _func
        self.__doc__ = self._func.__doc__
        self.__name__ = self._func.__name__
        # TODO use inspect to automatically find parameters with defaults
        self._params = kwargs

    def __call__(self, *args, **kwargs):
        kw = self._params
        kw.update(kwargs)
        return self._func(*args, **kw)

    def get_params(self, deep=False):
        out = self._params.copy()
        out['_func'] = self._func
        for key, value in out.items():
            if deep and hasattr(value, 'get_parms'):
                deep_items = value.get_params().items()
                out.update((key + '__' + k, val) for k, val in deep_items)
        return out

    def set_params(self, **params):
        if not params:
            # Simple optimization to gain speed (inspect is slow)
            return self
        valid_params = self.get_params(deep=True)

        nested_params = defaultdict(dict)  # grouped by prefix
        for key, value in params.items():
            key, delim, sub_key = key.partition('__')
            if key not in valid_params:
                raise ValueError('Invalid parameter %s for estimator %s. '
                                 'Check the list of available parameters '
                                 'with `estimator.get_params().keys()`.' %
                                 (key, self))

            if delim:
                nested_params[key][sub_key] = value
            else:
                self._params[key] = value
                valid_params[key] = value

        for key, sub_params in nested_params.items():
            valid_params[key].set_params(**sub_params)

        return self


if __name__ == '__main__':
    from sklearn.feature_selection import mutual_info_regression, SelectKBest
    from sklearn.model_selection import GridSearchCV
    from sklearn.pipeline import make_pipeline
    from sklearn.linear_model import LinearRegression
    from sklearn.datasets import make_regression

    mutual_info_regression = parametrized_function(mutual_info_regression,
                                                   n_neighbors=3)
    X, y = make_regression()

    gs = GridSearchCV(make_pipeline(SelectKBest(mutual_info_regression, k=1),
                                    LinearRegression()),
                      {'selectkbest__score_func__n_neighbors': [3, 4]},
                      cv=5, return_train_score=False).fit(X, y)
    print(pd.DataFrame(gs.cv_results_))

Note that if BaseEstimator's get_params and set_params were refactored to some BaseParametrized where _get_param was used instead of getattr and _set_param was used instead of setattr, we wouldn't need to duplicate so much code here

Reading through the comments, I do not find it too magically. I see this as an extension of how Pipeline or ColumnTransformer exposes their estimators with get_params and set_params.

As for refactoring into a BaseParametrized, if we make BaseParametrized easy to use for third parties, then they can easily extend any object to have the get_params and set_params interface.