Lasso error when precompute='auto'

Describe the bug

The docs state that the precompute parameter can be “‘auto’, bool or array-like of shape (n_features, n_features).” However, attempting to use precompute='auto' leads to a ValueError.

Steps/Code to Reproduce

from sklearn.linear_model import Lasso

model = Lasso(precompute='auto')
model.fit([[1, 2], [3, 4]], [[1], [3]])

Expected Results

No error, should “just work.”

Actual Results

Traceback (most recent call last):
  File "/opt/project/vpf/elm/tmp.py", line 4, in <module>
    model.fit([[1, 2], [3, 4]], [[1], [3]])
  File "/usr/local/lib/python3.8/site-packages/sklearn/linear_model/_coordinate_descent.py", line 757, in fit
    raise ValueError('precompute should be one of True, False or'
ValueError: precompute should be one of True, False or array-like. Got 'auto'

Versions

Output from sklearn.show_versions():

System:
    python: 3.8.6 (default, Nov 18 2020, 13:49:49)  [GCC 8.3.0]
executable: /usr/local/bin/python
   machine: Linux-4.19.104-microsoft-standard-x86_64-with-glibc2.2.5
Python dependencies:
          pip: 20.3.3
   setuptools: 50.3.2
      sklearn: 0.24.0
        numpy: 1.19.5
        scipy: 1.6.0
       Cython: None
       pandas: None
   matplotlib: None
       joblib: 1.0.0
threadpoolctl: 2.1.0
Built with OpenMP: True

[jjerphan]

I can reproduce. This bug does not exist for the other models supporting precompute:

import numpy as np
from sklearn.linear_model import (
    Lasso,
    LassoCV,
    ElasticNetCV,
)

models = [
    Lasso,
    LassoCV,
    ElasticNetCV,
]

if __name__ == "__main__":
    X = np.random.rand(10, 2)
    y = np.random.rand(10)

    for Model in models:
        try:
            model = Model(precompute='auto')
            model.fit(X, y)
            print("✅", Model)
        except Exception as e:
            print("❌", Model)
            print(type(e), ":", e)
        print()

❌ <class 'sklearn.linear_model._coordinate_descent.Lasso'>
<class 'ValueError'> : precompute should be one of True, False or array-like. Got 'auto'

✅ <class 'sklearn.linear_model._coordinate_descent.LassoCV'>

✅ <class 'sklearn.linear_model._coordinate_descent.ElasticNetCV'>

[jjerphan]

AFAIK the precompute ‘auto’ in Lasso was removed as it was not clear when to toggle True or False. We decided to let the user decide and use True by default as it’s the best default. I would not put it back unless there is a thorough benchmark (but I think it’s hard).

Thanks for the details. It looks like precompute was set to False to handle sparsity: https://github.com/scikit-learn/scikit-learn/pull/19348 offers some clarification.

I would just remove the mentions of 'auto' in the docstring as its support is not present and not relevant.