Merge pull request #1443 from Trusted-AI/development_issue_1334

Update check_and_transform_label_format for index labels

Author: beat-buesser

art/attacks/evasion/pixel_threshold.py

@@ -156,9 +156,11 @@ def generate(self, x: np.ndarray, y: Optional[np.ndarray] = None, **kwargs) -> n
                 raise ValueError(
                     "This attack has not yet been tested for binary classification with a single output classifier."
                 )
-            if len(y.shape) > 1:
+            if y.ndim > 1 and y.shape[1] > 1:
                 y = np.argmax(y, axis=1)
 
+        y = np.squeeze(y)
+
         if self.th is None:
             logger.info(
                 "Performing minimal perturbation Attack. \

art/estimators/classification/keras.py

@@ -562,13 +562,14 @@ def fit(self, x: np.ndarray, y: np.ndarray, batch_size: int = 128, nb_epochs: in
                `fit_generator` function in Keras and will be passed to this function as such. Including the number of
                epochs or the number of steps per epoch as part of this argument will result in as error.
         """
+        y_ndim = y.ndim
         y = check_and_transform_label_format(y, self.nb_classes)
 
         # Apply preprocessing
         x_preprocessed, y_preprocessed = self._apply_preprocessing(x, y, fit=True)
 
         # Adjust the shape of y for loss functions that do not take labels in one-hot encoding
-        if self._reduce_labels:
+        if self._reduce_labels or y_ndim == 1:
             y_preprocessed = np.argmax(y_preprocessed, axis=1)
 
         self._model.fit(x=x_preprocessed, y=y_preprocessed, batch_size=batch_size, epochs=nb_epochs, **kwargs)

art/utils.py

@@ -535,21 +535,27 @@ def check_and_transform_label_format(
     :return: Labels with shape `(nb_samples, nb_classes)` (one-hot) or `(nb_samples,)` (index).
     """
     if labels is not None:
-        if len(labels.shape) == 2 and labels.shape[1] > 1:
+        if len(labels.shape) == 2 and labels.shape[1] > 1:  # multi-class, one-hot encoded
             if not return_one_hot:
                 labels = np.argmax(labels, axis=1)
-        elif len(labels.shape) == 2 and labels.shape[1] == 1 and nb_classes is not None and nb_classes > 2:
-            labels = np.squeeze(labels)
+                labels = np.expand_dims(labels, axis=1)
+        elif (
+            len(labels.shape) == 2 and labels.shape[1] == 1 and nb_classes is not None and nb_classes > 2
+        ):  # multi-class, index labels
             if return_one_hot:
                 labels = to_categorical(labels, nb_classes)
-        elif len(labels.shape) == 2 and labels.shape[1] == 1 and nb_classes is not None and nb_classes == 2:
-            pass
-        elif len(labels.shape) == 1:
+            else:
+                labels = np.expand_dims(labels, axis=1)
+        elif (
+            len(labels.shape) == 2 and labels.shape[1] == 1 and nb_classes is not None and nb_classes == 2
+        ):  # binary, index labels
             if return_one_hot:
-                if nb_classes == 2:
-                    labels = np.expand_dims(labels, axis=1)
-                else:
-                    labels = to_categorical(labels, nb_classes)
+                labels = to_categorical(labels, nb_classes)
+        elif len(labels.shape) == 1:  # index labels
+            if return_one_hot:
+                labels = to_categorical(labels, nb_classes)
+            else:
+                labels = np.expand_dims(labels, axis=1)
         else:
             raise ValueError(
                 "Shape of labels not recognised."

tests/test_utils.py

@@ -171,26 +171,44 @@ def test_to_categorical(self):
 
     def test_check_and_transform_label_format(self):
         labels_expected = np.array([[0, 0, 0, 1, 0], [0, 1, 0, 0, 0], [0, 0, 0, 0, 1]])
+        labels_expected_binary = np.array([[0, 1], [1, 0], [0, 1]])
 
         # test input shape (nb_samples,)
         labels = np.array([3, 1, 4])
-        labels_transformed = check_and_transform_label_format(labels, 5)
+        labels_transformed = check_and_transform_label_format(labels, nb_classes=5, return_one_hot=True)
         np.testing.assert_array_equal(labels_transformed, labels_expected)
 
         # test input shape (nb_samples, 1)
         labels = np.array([[3], [1], [4]])
-        labels_transformed = check_and_transform_label_format(labels, 5)
+        labels_transformed = check_and_transform_label_format(labels, nb_classes=5, return_one_hot=True)
         np.testing.assert_array_equal(labels_transformed, labels_expected)
 
+        # test input shape (nb_samples, 1) - binary
+        labels = np.array([[1], [0], [1]])
+        labels_transformed = check_and_transform_label_format(labels, nb_classes=2, return_one_hot=True)
+        np.testing.assert_array_equal(labels_transformed, labels_expected_binary)
+
+        # test input shape (nb_samples, 1) - binary
+        labels = np.array([[0, 1], [1, 0], [0, 1]])
+        labels_transformed = check_and_transform_label_format(labels, nb_classes=2, return_one_hot=True)
+        np.testing.assert_array_equal(labels_transformed, labels_expected_binary)
+
         # test input shape (nb_samples, nb_classes)
         labels = np.array([[0, 0, 0, 1, 0], [0, 1, 0, 0, 0], [0, 0, 0, 0, 1]])
-        labels_transformed = check_and_transform_label_format(labels, 5)
+        labels_transformed = check_and_transform_label_format(labels, nb_classes=5, return_one_hot=True)
         np.testing.assert_array_equal(labels_transformed, labels_expected)
 
         # test input shape (nb_samples, nb_classes) with return_one_hot=False
         labels = np.array([[0, 0, 0, 1, 0], [0, 1, 0, 0, 0], [0, 0, 0, 0, 1]])
-        labels_transformed = check_and_transform_label_format(labels, 5, return_one_hot=False)
-        np.testing.assert_array_equal(labels_transformed, np.argmax(labels_expected, axis=1))
+        labels_transformed = check_and_transform_label_format(labels, nb_classes=5, return_one_hot=False)
+        np.testing.assert_array_equal(labels_transformed, np.expand_dims(np.argmax(labels_expected, axis=1), axis=1))
+
+        # test input shape (nb_samples, 1) - binary
+        labels = np.array([[1], [0], [1]])
+        labels_transformed = check_and_transform_label_format(labels, nb_classes=2, return_one_hot=False)
+        np.testing.assert_array_equal(
+            labels_transformed, np.expand_dims(np.argmax(labels_expected_binary, axis=1), axis=1)
+        )
 
         # ValueError for len(labels.shape) > 2
         labels = np.array([[[0, 0, 0, 1, 0], [0, 1, 0, 0, 0], [0, 0, 0, 0, 1]]])