Make `NotInitializedError` a subclass of `NotFittedError`

[timokau]

Hi,

currently calling predict on an uninitialized estimator throws a NotInitializedError instead of scikit-learn's NotFittedError. This is problematic in cases where we want to be able to handle different kinds of estimators while providing some kind of fallback behavior. Currently we need

try:
    prediction = estimator.predict(X)
except (NotFittedError, NotInitializedError):
    # some fallback behavior

I think it would be better if NotInitializedError was a subclass of NotFittedError. In that case we could treat an skorch estimator like a regular scikit-learn estimator and simply catch NotFittedError.

Would a PR in that direction be welcome?

[BenjaminBossan]

Thanks for the suggestion, sounds indeed like a good idea. PR is welcome.

[timokau]

Great :) See #860.