Fix AttributeError use_fallback_lbfgs_solve for newton-cholesky when fitting with max_iter=0 (#26653)

Co-authored-by: Christian Lorentzen <lorentzen.ch@gmail.com>
Co-authored-by: Jérémie du Boisberranger <34657725+jeremiedbb@users.noreply.github.com>

Author: 3 people

doc/whats_new/v1.3.rst

@@ -455,6 +455,11 @@ Changelog
   on linearly separable problems.
   :pr:`25214` by `Tom Dupre la Tour`_.
 
+- |Fix| Fix a crash when calling `fit` on
+  :class:`linear_model.LogisticRegression(solver="newton-cholesky", max_iter=0)`
+  which failed to inspect the state of the model prior to the first parameter update.
+  :pr:`26653` by :user:`Olivier Grisel <ogrisel>`.
+
 - |API| Deprecates `n_iter` in favor of `max_iter` in
   :class:`linear_model.BayesianRidge` and :class:`linear_model.ARDRegression`.
   `n_iter` will be removed in scikit-learn 1.5. This change makes those

sklearn/linear_model/_glm/_newton_solver.py

@@ -375,6 +375,7 @@ def solve(self, X, y, sample_weight):
 
         self.iteration = 1
         self.converged = False
+        self.use_fallback_lbfgs_solve = False
 
         while self.iteration <= self.max_iter and not self.converged:
             if self.verbose:

sklearn/linear_model/tests/test_logistic.py

@@ -2063,3 +2063,29 @@ def test_liblinear_not_stuck():
     with warnings.catch_warnings():
         warnings.simplefilter("error", ConvergenceWarning)
         clf.fit(X_prep, y)
+
+
+@pytest.mark.parametrize("solver", SOLVERS)
+def test_zero_max_iter(solver):
+    # Make sure we can inspect the state of LogisticRegression right after
+    # initialization (before the first weight update).
+    X, y = load_iris(return_X_y=True)
+    y = y == 2
+    with ignore_warnings(category=ConvergenceWarning):
+        clf = LogisticRegression(solver=solver, max_iter=0).fit(X, y)
+    if solver not in ["saga", "sag"]:
+        # XXX: sag and saga have n_iter_ = [1]...
+        assert clf.n_iter_ == 0
+
+    if solver != "lbfgs":
+        # XXX: lbfgs has already started to update the coefficients...
+        assert_allclose(clf.coef_, np.zeros_like(clf.coef_))
+        assert_allclose(
+            clf.decision_function(X),
+            np.full(shape=X.shape[0], fill_value=clf.intercept_),
+        )
+        assert_allclose(
+            clf.predict_proba(X),
+            np.full(shape=(X.shape[0], 2), fill_value=0.5),
+        )
+    assert clf.score(X, y) < 0.7