FIX: Fix KNN predictions on X=None (#2821)

* fix predictions on X=None

* don't fall back on predict_proba

* add comment about misleading log

* don't deselect valid conformance tests

* limit scope of deselections

Author: david-cortes-intel

deselected_tests.yaml

@@ -34,10 +34,6 @@ deselected_tests:
   - neighbors/tests/test_neighbors.py::test_nan_euclidean_support[KNeighborsClassifier-params0]
   - neighbors/tests/test_neighbors.py::test_nan_euclidean_support[KNeighborsRegressor-params1]
   - neighbors/tests/test_neighbors.py::test_nan_euclidean_support[LocalOutlierFactor-params6]
-  - neighbors/tests/test_neighbors.py::test_neighbor_classifiers_loocv[ball_tree-nn_model0]
-  - neighbors/tests/test_neighbors.py::test_neighbor_classifiers_loocv[brute-nn_model0]
-  - neighbors/tests/test_neighbors.py::test_neighbor_classifiers_loocv[kd_tree-nn_model0]
-  - neighbors/tests/test_neighbors.py::test_neighbor_classifiers_loocv[auto-nn_model0]
   # sklearn 1.7 unsupported features
   - tests/test_common.py::test_estimators[LinearRegression()-check_sample_weight_equivalence_on_dense_data]
 

onedal/neighbors/neighbors.py

@@ -472,7 +472,7 @@ def predict_proba(self, X, queue=None):
             _y = self._y.reshape((-1, 1))
             classes_ = [self.classes_]
 
-        n_queries = _num_samples(X)
+        n_queries = _num_samples(X if X is not None else self._fit_X)
 
         weights = self._get_weights(neigh_dist, self.weights)
         if weights is None:

sklearnex/neighbors/common.py

@@ -156,6 +156,19 @@ def _onedal_supported(self, device, method_name, *data):
         patching_status = PatchingConditionsChain(
             f"sklearn.neighbors.{class_name}.{method_name}"
         )
+        # TODO: with verbosity enabled, here it would emit a log saying that it fell
+        # back to sklearn, but internally, sklearn will end up calling 'kneighbors'
+        # which is overridden in the sklearnex classes, thus it will end up calling
+        # oneDAL in the end, but the log will say otherwise. Find a way to make the
+        # log consistent with what happens in practice.
+        patching_status.and_conditions(
+            [
+                (
+                    not (data[0] is None and method_name in ["predict", "score"]),
+                    "Predictions on 'None' data are handled by internal sklearn methods.",
+                )
+            ]
+        )
         if not patching_status.and_condition(
             "radius" not in method_name, "RadiusNeighbors not implemented in sklearnex"
         ):

sklearnex/neighbors/knn_classification.py

@@ -79,7 +79,8 @@ def fit(self, X, y):
     @wrap_output_data
     def predict(self, X):
         check_is_fitted(self)
-        check_feature_names(self, X, reset=False)
+        if X is not None:
+            check_feature_names(self, X, reset=False)
         return dispatch(
             self,
             "predict",
@@ -93,7 +94,8 @@ def predict(self, X):
     @wrap_output_data
     def predict_proba(self, X):
         check_is_fitted(self)
-        check_feature_names(self, X, reset=False)
+        if X is not None:
+            check_feature_names(self, X, reset=False)
         return dispatch(
             self,
             "predict_proba",
@@ -107,7 +109,8 @@ def predict_proba(self, X):
     @wrap_output_data
     def score(self, X, y, sample_weight=None):
         check_is_fitted(self)
-        check_feature_names(self, X, reset=False)
+        if X is not None:
+            check_feature_names(self, X, reset=False)
         return dispatch(
             self,
             "score",