Compatibility with Scikit-Learn Feature Selection

[sanjaradylov]

BaseGroupLasso is implemented as a scikit-learn transformer. As an intermediate step inside a pipeline it works well, but in combination with SelectFromModel (to further remove infinitesimal coefs), it throws an error:

>>> import sklearn, group_lasso
>>> sklearn.__version__, group_lasso.__version__
('1.0.2', '1.5.0')
>>> from group_lasso import GroupLasso
>>> from sklearn.datasets import make_regression
>>> from sklearn.feature_selection import SelectFromModel
>>> from sklearn.linear_model import Ridge
>>> from sklearn.pipeline import make_pipeline
>>> X, y = make_regression(n_features=5, n_informative=3, random_state=0)
>>> pipe = make_pipeline(SelectFromModel(GroupLasso(supress_warning=True)), Ridge())
>>> pipe.fit(X, y)
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
[<ipython-input-3-e63339588fc2>](https://localhost:8080/#) in <module>()
      3 
      4 X, y = make_regression(n_features=5, n_informative=3, random_state=0)
----> 5 pipe.fit(X, y)

6 frames
[/usr/local/lib/python3.7/dist-packages/sklearn/feature_selection/_base.py](https://localhost:8080/#) in _transform(self, X)
    101             return np.empty(0).reshape((X.shape[0], 0))
    102         if len(mask) != X.shape[1]:
--> 103             raise ValueError("X has a different shape than during fitting.")
    104         return X[:, safe_mask(X, mask)]
    105 

ValueError: X has a different shape than during fitting.

Possible solutions:

1. Add a new parameter min_coef and zero out all the coefficients s.t. np.abs(coef_) < min_coef. Optionally, reimplement TransformerMixin methods by inheriting from SelectorMixin to support get_support() and get_feature_names_out() methods.
   or
2. Remove fit_transform and transform methods to enable SelectFromModel(Grouplasso(), threshold=min_coef) inside a pipeline.