Fix compatibility with scikit-learn 1.8 (#852)

* fixing more colab issues

* updating the utils

Author: bruAristimunha

moabb/datasets/preprocessing.py

@@ -7,7 +7,20 @@
 import numpy as np
 from sklearn.base import BaseEstimator, TransformerMixin
 from sklearn.pipeline import FunctionTransformer, Pipeline, _name_estimators
-from sklearn.utils._repr_html.estimator import _VisualBlock
+
+
+# Handle different scikit-learn versions for _VisualBlock import
+# sklearn >= 1.6 moved _VisualBlock to sklearn.utils._repr_html.estimator
+# sklearn < 1.6 has it in sklearn.utils._estimator_html_repr
+try:
+    from sklearn.utils._repr_html.estimator import _VisualBlock
+except (ImportError, ModuleNotFoundError):
+    try:
+        from sklearn.utils._estimator_html_repr import _VisualBlock
+    except (ImportError, ModuleNotFoundError):
+        # Fallback: create a dummy _VisualBlock for older sklearn versions
+        # that don't have HTML representation support
+        _VisualBlock = None
 
 
 log = logging.getLogger(__name__)
@@ -88,7 +101,9 @@ def __sklearn_is_fitted__(self):
         return True
 
     def _sk_visual_block_(self):
-        """Tell sklearn’s diagrammer to lay us out in parallel."""
+        """Tell sklearn's diagrammer to lay us out in parallel."""
+        if _VisualBlock is None:
+            return NotImplemented
         names, estimators = zip(*self.transformers)
         return _VisualBlock(
             kind="parallel",
@@ -114,7 +129,9 @@ def __sklearn_is_fitted__(self):
         return True
 
     def _sk_visual_block_(self):
-        """Tell sklearn’s diagrammer to lay us out in parallel."""
+        """Tell sklearn's diagrammer to lay us out in parallel."""
+        if _VisualBlock is None:
+            return NotImplemented
         return _VisualBlock(
             kind="parallel",
             name_caption=str(self.__class__.__name__),
@@ -405,6 +422,8 @@ def __repr__(self):
         return self._display_name
 
     def _sk_visual_block_(self):
+        if _VisualBlock is None:
+            return NotImplemented
         return _VisualBlock(
             kind="single",
             estimators=self,

moabb/evaluations/evaluations.py

@@ -25,6 +25,7 @@
 )
 from moabb.evaluations.utils import (
     _create_save_path,
+    _ensure_fitted,
     _save_model_cv,
 )
 
@@ -226,6 +227,7 @@ def _evaluate(
                         cvclf = clone(grid_clf)
 
                         cvclf.fit(X_[train], y_[train])
+                        _ensure_fitted(cvclf)
 
                         score = scorer(cvclf, X_[test], y_[test])
 
@@ -308,6 +310,7 @@ def score_explicit(self, clf, X_train, y_train, X_test, y_test):
         t_start = time()
         try:
             model = clf.fit(X_train, y_train)
+            _ensure_fitted(model)
             score = _score(
                 estimator=model,
                 X_test=X_test,
@@ -540,6 +543,7 @@ def evaluate(
                     cvclf = clone(grid_clf)
 
                     cvclf.fit(X[train], y[train])
+                    _ensure_fitted(cvclf)
 
                     model_list.append(cvclf)
                     score = scorer(cvclf, X[test], y[test])
@@ -724,6 +728,7 @@ def evaluate(
                     )
 
                 model = deepcopy(clf).fit(X[train], y[train])
+                _ensure_fitted(model)
 
                 if _carbonfootprint:
                     emissions = tracker.stop()

moabb/evaluations/utils.py

@@ -23,6 +23,67 @@
     optuna_available = False
 
 
+def _ensure_fitted(estimator):
+    """Ensure an estimator is properly marked as fitted for sklearn 1.8+.
+
+    In sklearn 1.8+, Pipeline.predict() calls check_is_fitted(self) which
+    may fail for some estimators (especially deep learning wrappers) that
+    don't properly set fitted attributes. This function adds the necessary
+    attributes to ensure the estimator passes sklearn's fitted check.
+
+    Parameters
+    ----------
+    estimator : sklearn-compatible estimator
+        The fitted estimator to mark as fitted. This should be called
+        after fit() has been called on the estimator.
+
+    Returns
+    -------
+    estimator : sklearn-compatible estimator
+        The same estimator with fitted attributes set.
+
+    Notes
+    -----
+    This function modifies the estimator in-place and returns it for
+    convenience. sklearn's check_is_fitted looks for:
+    1. __sklearn_is_fitted__() method returning True
+    2. Or any attribute ending with '_' (like classes_, coef_, etc.)
+
+    We add a __sklearn_is_fitted__ method that returns True.
+    """
+
+    # Define a method that returns True to indicate fitted state
+    def _sklearn_is_fitted_true(self):
+        return True
+
+    # Add __sklearn_is_fitted__ method if not present or if it returns False
+    if not hasattr(estimator, "__sklearn_is_fitted__"):
+        import types
+
+        estimator.__sklearn_is_fitted__ = types.MethodType(
+            _sklearn_is_fitted_true, estimator
+        )
+    else:
+        # Check if existing method returns False (unfitted)
+        try:
+            if not estimator.__sklearn_is_fitted__():
+                import types
+
+                estimator.__sklearn_is_fitted__ = types.MethodType(
+                    _sklearn_is_fitted_true, estimator
+                )
+        except Exception:
+            pass
+
+    # For Pipeline objects, also ensure all steps are marked
+    if isinstance(estimator, Pipeline):
+        for name, step in estimator.steps:
+            if step is not None:
+                _ensure_fitted(step)
+
+    return estimator
+
+
 def _check_if_is_pytorch_model(model):
     """Check if the model is a skorch model.