I need to make a VotingRegressor ensemble with some estimators that accept sample weights during fitting and some that don't. Currently, mixed ensembles raise an exception:

from sklearn.datasets import make_regression
from sklearn.linear_model import LinearRegression
from sklearn.ensemble import RandomForestRegressor, VotingRegressor
from sklearn.neighbors import KNeighborsRegressor

X, y = make_regression()
weights = abs(y)

rgr = VotingRegressor(estimators=[('LR', LinearRegression()), 
                                  ('KNN', KNeighborsRegressor()),
                                  ('XGBR', RandomForestRegressor())])

rgr.fit(X, y, sample_weight=weights)

result:
TypeError: Underlying estimator KNeighborsRegressor does not support sample weights.

A possible solution would be to have the ensemble class (e.g. VotingRegressor, VotingClassifier, StackingRegressor, StackingClassifier) read the fit() signatures of the estimators in the ensemble and not pass sample_weight to estimators that don't accept sample_weight. Or more realistically, catch exceptions caused by calls to fit() with the sample_weight parameter and then default to calling fit() without this parameter. This behavior could be default, or activated by flag like "enable_mixed_sample_weight" in the ensemble class's __init__ method. If it's important to notify the user when an estimator doesn't accept the sample_weight parameter, notification and the exception currently in place could be enabled with a flag like "enforce_sample_weight".

As a workaround I'm using the Ensemble class from the pipecaster library (https://github.com/ajcallegari/pipecaster) which allows mixed ensembles by catching exceptions caused by fit() and then defaulting to a fit() call without the sample_weight parameter. This Ensemble class has the scikit-learn interface and supports classification, regression, voting, and model stacking.