We are looking to have n_features_out_ for transformers. This naming makes the difference explicit.

I would like to see some guidance on how an estimator implementation (e.g. in scikit-learn-contrib) is advised to maintain compatibility with Scikit-learn pre- and post- SLEP010.

That is, we want to encourage developers to take advantage of super()._validate_data(X, y), but we also don't want to force them to set a minimal Scikit-learn >= 0.23 dependency (or do we?). What's the recommended way to do implement fit and predict in such an implementation?

Is it to
(a) not use _validate_data until the minimal dependency is reached?
(b) implement a patched BaseEstimator in the library which inherits from Scikit-learn's BaseEstimator and adds _validate_data?
(c) something else?


Joel