Make all non-canonical modules private?

This is maybe for 1.0. Recently we started marking files like model_selection._split private, so that everything has a single canonical import:

from sklearn.model_selection import cross_val_score

For many (older?) models that’s not the case, we have

from sklearn.linear_model.logistic import LogisticRegression
from sklearn.linear_model import LogisticRegression

etc. I think it would be nice to make all the files that are not the canonical import (according to the API documentation) private (with deprecation obviously). That might be a bit annoying for existing users that used long import paths, but it makes auto-complete much more helpful and the module structure much less confusing for newcomers.

For example sklearn.linear_model.ridge is a module, while sklearn.linear_model.ridge_regression is a function that implements ridge regression and sklearn.linear_model.Ridge is a class that implements ridge regression. From the names this is totally unclear.

[glemaitre]

you can import from sklearn.datasets._base import ...

However, now you are aware that _base will be private and you might have to change your code at a new scikit-learn release since we might not provide backward-compatible code.

This said, be aware that we are going to define a developer API which will define some backward support for these types of utilities used in third-party.

[lucasdavid]

I had a ton of private datasets, such as names, text and geographical data. I used to use the following utils to download and manage these:

from sklearn.datasets.base import (get_data_home, RemoteFileMetadata, _fetch_remote)

What’s the recommended approach now?