check_estimator

[scikit-learn-contrib]

check_estimator

[vighneshbirodkar]

@RomainBrault You can use an if else statement based on the version number. I can't think of a clearner way to do it.

Thoughts @mblondel ?

[mblondel]

We typically try to maintain backward compatibility for two releases so this seems like a bug.

@amueller Would it be possible to support both classes and instances for two releases?

[amueller]

From the discussion, it seems the 0.17 version is a bug and didn't work. All the checks expect a class object, not an instance.

[RomainBrault]

@vighneshbirodkar Yes, it's what I did, but It looks a bit hackish to do it at user's level.

Anyway If I pass an instantiation of the object in 0.17, check_estimator only uses its class name. I think
it's actually useful if the check is done at the class level: one can do check_estimator(self) instead of check_estimator(self.__class__).

However if not applying on self, the code in master looks more natural to me (no need to (default) construct the estimator). I would be in favour of supporting both.

[azrdev]

In the meantime (I'm looking at current sklearn 0.19.1) check_estimator can work with an instance or class, so this can be closed.