Merge pull request #2505 from abigailgold/sklearn_nbclasses

Support sklearn models with multiple outputs

Author: beat-buesser

art/estimators/classification/classifier.py

@@ -116,7 +116,7 @@ def nb_classes(self, nb_classes: int):
         """
         Set the number of output classes.
         """
-        if nb_classes is None or nb_classes < 2:
+        if nb_classes is None or (isinstance(nb_classes, (int, np.integer)) and nb_classes < 2):
             raise ValueError("nb_classes must be greater than or equal to 2.")
 
         self._nb_classes = nb_classes

art/utils.py

@@ -799,15 +799,18 @@ def check_and_transform_label_format(
     labels: np.ndarray, nb_classes: int | None, return_one_hot: bool = True
 ) -> np.ndarray:
     """
-    Check label format and transform to one-hot-encoded labels if necessary
+    Check label format and transform to one-hot-encoded labels if necessary. Only supports single-output classification.
 
     :param labels: An array of integer labels of shape `(nb_samples,)`, `(nb_samples, 1)` or `(nb_samples, nb_classes)`.
-    :param nb_classes: The number of classes. If None the number of classes is determined automatically.
+    :param nb_classes: The number of classes, as an integer. If None the number of classes is determined automatically.
     :param return_one_hot: True if returning one-hot encoded labels, False if returning index labels.
     :return: Labels with shape `(nb_samples, nb_classes)` (one-hot) or `(nb_samples,)` (index).
     """
     labels_return = labels
 
+    if nb_classes is not None and not isinstance(nb_classes, (int, np.integer)):
+        raise TypeError("nb_classes that is not an integer is not supported")
+
     if len(labels.shape) == 2 and labels.shape[1] > 1:  # multi-class, one-hot encoded
         if not return_one_hot:
             labels_return = np.argmax(labels, axis=1)

tests/estimators/classification/test_scikitlearn.py

@@ -47,6 +47,7 @@
     ScikitlearnSVC,
 )
 from art.estimators.classification.scikitlearn import SklearnClassifier
+from art.utils import check_and_transform_label_format
 
 from tests.utils import TestBase, master_seed
 
@@ -80,6 +81,28 @@ def test_save(self):
     def test_clone_for_refitting(self):
         _ = self.classifier.clone_for_refitting()
 
+    def test_multi_label(self):
+        x_train = self.x_train_iris
+        y_train = self.y_train_iris
+        x_test = self.x_test_iris
+        y_test = self.y_test_iris
+
+        # make multi-label binary
+        y_train = np.column_stack((y_train, y_train, y_train))
+        y_train[y_train > 1] = 1
+        y_test = np.column_stack((y_test, y_test, y_test))
+        y_test[y_test > 1] = 1
+
+        underlying_model = DecisionTreeClassifier()
+        underlying_model.fit(x_train, y_train)
+        model = ScikitlearnDecisionTreeClassifier(model=underlying_model)
+
+        pred = model.predict(x_test)
+        assert pred[0].shape[0] == x_test.shape[0]
+        assert isinstance(model.nb_classes, np.ndarray)
+        with self.assertRaises(TypeError):
+            check_and_transform_label_format(y_train, nb_classes=model.nb_classes)
+
 
 class TestScikitlearnExtraTreeClassifier(TestBase):
     @classmethod