Warn subclassed UserWarning when unpickling model using different verison of sklearn

Describe the workflow you want to enable

I want to terminate execution if I’m unpickling a model using a different version of sklearn.

Describe your proposed solution

Currently it warns with a UserWarning:

https://github.com/scikit-learn/scikit-learn/blob/4abd9f2d9e12451fa3e32bd77a97b5c4d8be8b19/sklearn/base.py#L320-L334

If, instead, it used a subclassed UserWarning you could wrap up loading a model in a catch_warnings context manager and filter that warning to raise instead, and then exit if it is raised.

Describe alternatives you’ve considered, if relevant

You could potentially do this by catching UserWarning and checking matching such as user_warning.message.startswith("Trying to unpickle estimator") but this is not very robust.

Additional context

No response

[thomasjpfan]

we’ve gone back and forth on the types of warnings and exceptions we raise.

I’ve seen the specializing issues/expcetions issue pop up from time to time, so there is a demand for them. I think it’s common for a Python library to have specialized warnings and exceptions.

The only documentation we have for exceptions are in the API docs. The last specialized warning we added was PositiveSpectrumWarning which has limited scope.

As for this issue, I think a unpickling warning also has limited scope, which makes it easier to ensure consistency. (Limited scope == less places can use it == less need to think about consistency). To move forward, I think we can move ahead with a PR and see what everyone thinks.

[adrinjalali]

I agree, happy to see the PR.