amueller
diff --git a/‎sklearn/preprocessing.py
Lines changed: 12 additions & 1 deletion b/‎sklearn/preprocessing.py
Lines changed: 12 additions & 1 deletion
diff --git a/‎sklearn/tests/test_preprocessing.py
Lines changed: 10 additions & 1 deletion b/‎sklearn/tests/test_preprocessing.py
Lines changed: 10 additions & 1 deletion
@@ -932,6 +932,11 @@ class LabelBinarizer(BaseEstimator, TransformerMixin):
         Whether or not data will be multilabel.
         If None, it will be inferred during fitting.
 
+    indicator_matrix : bool or None (default)
+        Whether ``inverse_transform`` will produce an indicator
+        matrix encoding (if False it will return list of lists).
+        If None, it will be inferred during fitting.
+
     Attributes
     ----------
     `classes_` : array of shape [n_class]
@@ -940,6 +945,10 @@ class LabelBinarizer(BaseEstimator, TransformerMixin):
     `multilabel_` : bool
         Whether the estimator was fitted for multi-label data.
 
+    `indicator_matrix_` : bool
+        Whether the estimator was fitted with a label indicator matrix.
+        This will determine the result of ``inverse_transform``.
+
     Examples
     --------
     >>> from sklearn import preprocessing
@@ -960,21 +969,23 @@ class LabelBinarizer(BaseEstimator, TransformerMixin):
     """
 
     def __init__(self, neg_label=0, pos_label=1, classes=None,
-                 multilabel=None):
+                 multilabel=None, indicator_matrix=None):
         if neg_label >= pos_label:
             raise ValueError("neg_label must be strictly less than pos_label.")
 
         self.neg_label = neg_label
         self.pos_label = pos_label
         self.classes = classes
         self.multilabel = multilabel
+        self.indicator_matrix = indicator_matrix
 
     def _check_fitted(self):
         if not hasattr(self, "classes_"):
             if self.classes is not None:
                 self.classes_ = np.unique(self.classes)
                 # default to not doing multi-label things
                 self.multilabel_ = bool(self.multilabel)
+                self.indicator_matrix_ = bool(self.indicator_matrix)
             else:
                 raise ValueError("LabelBinarizer was not fitted yet.")
 
 
@@ -657,15 +657,24 @@ def test_label_binarizer_classes():
     transformed = lb.transform(['see', 'see'])
     assert_equal(transformed.shape, (2, 3))
     assert_array_equal(np.argmax(transformed, axis=1), [2, 2])
+    # test inverse transform
+    assert_array_equal(['see', 'see'], lb.inverse_transform(transformed))
 
     # also works with multilabel data if we say so:
     lb = LabelBinarizer(classes=np.arange(1, 3), multilabel=True)
-    y = [[1, 2], [1], []]
+    y = [(1, 2), (1,), ()]
     Y = np.array([[1, 1],
                   [1, 0],
                   [0, 0]])
     assert_array_equal(lb.transform(y), Y)
     assert_array_equal(lb.fit_transform(y), Y)
+    # inverse transform of label indicator matrix to label
+    assert_array_equal(lb.inverse_transform(Y), y)
+
+    # inverse transform  with indicator_matrix=True
+    lb = LabelBinarizer(classes=np.arange(1, 3), multilabel=True,
+                        indicator_matrix=True)
+    assert_array_equal(lb.inverse_transform(Y), Y)
 
     lb = LabelBinarizer(classes=np.arange(1, 3))
     assert_raise_message(ValueError, "not fitted with multilabel",