__getattr__ can silently catch attribute error from filters_ property

Author: Genuster

mne/decoding/base.py

@@ -476,14 +476,16 @@ def fit(self, X, y, **fit_params):
 
     @property
     def filters_(self):
-        check_is_fitted(self)
+        check_is_fitted(self.model_)
         if hasattr(self.model_, "coef_"):
             # Standard Linear Model
             filters = self.model_.coef_
         elif hasattr(self.model_, "estimators_"):
             # Linear model with OneVsRestClassifier
             filters = np.vstack([est.coef_ for est in self.model_.estimators_])
-        elif hasattr(self.model_.best_estimator_, "coef_"):
+        elif hasattr(self.model_, "best_estimator_") and hasattr(
+            self.model_.best_estimator_, "coef_"
+        ):
             # Linear Model with GridSearchCV
             filters = self.model_.best_estimator_.coef_
         else:
@@ -511,6 +513,11 @@ def model(self):
     def model(self, value):
         self._orig_model = value
 
+    # XXX Remove this after 'model' warning cycle
+    def __repr__(self):
+        """Avoid FutureWarning from filter_ when printing the instance."""
+        return f"LinearModel(model={self._orig_model})"
+
 
 def _set_cv(cv, estimator=None, X=None, y=None):
     """Set the default CV depending on whether clf is classifier/regressor."""

mne/decoding/tests/test_base.py

@@ -385,6 +385,16 @@ def test_linearmodel():
     clf = LinearModel(LinearDiscriminantAnalysis())
     _ = clf.fit_transform(X, y)
 
+    # check that model has to have coef_, RBF-SVM doesn't
+    clf = LinearModel(svm.SVC(kernel="rbf"))
+    with pytest.raises(ValueError, match="does not have a `coef_`"):
+        clf.fit(X, y)
+
+    # check that model has to be a predictor
+    clf = LinearModel(StandardScaler())
+    with pytest.raises(ValueError, match="classifier or regressor"):
+        clf.fit(X, y)
+
     # check categorical target fit in standard linear model with GridSearchCV
     parameters = {"kernel": ["linear"], "C": [1, 10]}
     clf = LinearModel(
@@ -421,11 +431,6 @@ def test_linearmodel():
         wrong_y = rng.rand(n, n_features, 99)
         clf.fit(X, wrong_y)
 
-    # check that model has to be a predictor
-    clf = LinearModel(StandardScaler())
-    with pytest.raises(ValueError, match="classifier or regressor"):
-        clf.fit(X, Y)
-
 
 def test_cross_val_multiscore():
     """Test cross_val_multiscore for computing scores on decoding over time."""