IndexError thrown with LogisticRegressionCV and refit=False

[rowan-stein]

IndexError thrown with LogisticRegressionCV and refit=False

Description

The following error is thrown when trying to estimate a regularization parameter via cross-validation, without refitting.

Steps/Code to Reproduce

import sys
import sklearn
from sklearn.linear_model import LogisticRegressionCV
import numpy as np

np.random.seed(29)
X = np.random.normal(size=(1000, 3))
beta = np.random.normal(size=3)
intercept = np.random.normal(size=None)
y = np.sign(intercept + X @ beta)

LogisticRegressionCV(
    cv=5,
    solver='saga',  # same error with 'liblinear'
    tol=1e-2,
    refit=False,
).fit(X, y)

Expected Results

No error is thrown.

Actual Results

Traceback (most recent call last):
  File "<stdin>", line 11, in <module>
  File "/workspace/scikit-learn/sklearn/linear_model/logistic.py", line 2178, in fit
    for i in range(len(folds))], axis=0)
  File "/workspace/scikit-learn/sklearn/linear_model/logistic.py", line 2178, in <listcomp>
    for i in range(len(folds))], axis=0)
IndexError: too many indices for array: array is 3-dimensional, but 4 were indexed

Versions

... as in original issue ...

---

Researcher Specification

Root cause

• In sklearn/linear_model/logistic.py, LogisticRegressionCV.fit indexes coefs_paths assuming a fold dimension exists when refit=False. Under certain shape constructions, per-class entries can be 2D (missing the fold dimension), leading to IndexError: too many indices for array.

Proposed fix (high-level)

• Explicitly stack coefs_paths with np.stack(..., axis=0) prior to reshape to avoid object arrays and guarantee dimensions.
• Reshape to ensure per-class entries are 3D (n_folds, n_cs * n_l1_ratios, n_features [+1]).
• Use vectorized selection for per-fold best indices and add defensive shape assertions.

Tests

• Add refit=False tests for binary ovr with saga and liblinear, and multinomial with saga (and elasticnet with l1_ratios) to validate shapes and absence of IndexError.

We will implement the fix and reference this Issue in a single PR targeting branch scikit-learn__scikit-learn-14087.

[casey-brooks]

Reproduced locally using the current scikit-learn__scikit-learn-14087 base branch with the snippet in the issue. Running under Python 3.7.10 (Miniconda) with scikit-learn built from source (OpenMP disabled) produced the following stack trace:

Traceback (most recent call last):
  File "<stdin>", line 11, in <module>
  File "/workspace/scikit-learn/sklearn/linear_model/logistic.py", line 2178, in fit
    for i in range(len(folds))], axis=0)
  File "/workspace/scikit-learn/sklearn/linear_model/logistic.py", line 2178, in <listcomp>
    for i in range(len(folds))], axis=0)
IndexError: too many indices for array: array is 3-dimensional, but 4 were indexed

[casey-brooks]

Full stack trace reproduced locally is now documented under the PR section "Actual Results" for convenience.