scikit-learn · TomDLT · Jan 15, 2020 · Jan 9, 2020 · Jan 9, 2020 · Jan 9, 2020
diff --git a/doc/whats_new/v0.23.rst b/doc/whats_new/v0.23.rst
@@ -118,6 +118,15 @@ Changelog
   when `y=None`.
   :pr: `15918` by :user: `Luca Kubin <lkubin>`.
 
+:mod:`sklearn.neighbors`
+..............................
+
+- |Fix| Fix a bug which converted a list of arrays into a 2-D object 
+  array instead of a 1-D array containing NumPy arrays. This bug
+  was affecting :meth:`neighbors.NearestNeighbors.radius_neighbors`.
+  :pr:`16076` by :user:`Guillaume Lemaitre <glemaitre>` and  
+  :user:`Alex Shacked <alexshacked>`.
+
 :mod:`sklearn.preprocessing`
 ............................
 

diff --git a/sklearn/neighbors/_base.py b/sklearn/neighbors/_base.py
@@ -24,6 +24,7 @@
 from ..metrics import pairwise_distances_chunked
 from ..metrics.pairwise import PAIRWISE_DISTANCE_FUNCTIONS
 from ..utils import check_X_y, check_array, gen_even_slices
+from ..utils import _to_object_array
 from ..utils.multiclass import check_classification_targets
 from ..utils.validation import check_is_fitted
 from ..utils.validation import check_non_negative
@@ -276,8 +277,8 @@ def _radius_neighbors_from_graph(graph, radius, return_distance):
     indices = indices.astype(np.intp, copy=no_filter_needed)
 
     if return_distance:
-        neigh_dist = np.array(np.split(data, indptr[1:-1]), dtype=object)
-    neigh_ind = np.array(np.split(indices, indptr[1:-1]), dtype=object)
+        neigh_dist = _to_object_array(np.split(data, indptr[1:-1]))
+    neigh_ind = _to_object_array(np.split(indices, indptr[1:-1]))
 
     if return_distance:
         return neigh_dist, neigh_ind
@@ -940,17 +941,12 @@ class from an array representing our data set and ask who's
                 neigh_dist_chunks, neigh_ind_chunks = zip(*chunked_results)
                 neigh_dist_list = sum(neigh_dist_chunks, [])
                 neigh_ind_list = sum(neigh_ind_chunks, [])
-                # See https://github.com/numpy/numpy/issues/5456
-                # to understand why this is initialized this way.
-                neigh_dist = np.empty(len(neigh_dist_list), dtype='object')
-                neigh_dist[:] = neigh_dist_list
-                neigh_ind = np.empty(len(neigh_ind_list), dtype='object')
-                neigh_ind[:] = neigh_ind_list
+                neigh_dist = _to_object_array(neigh_dist_list)
+                neigh_ind = _to_object_array(neigh_ind_list)
                 results = neigh_dist, neigh_ind
             else:
                 neigh_ind_list = sum(chunked_results, [])
-                results = np.empty(len(neigh_ind_list), dtype='object')
-                results[:] = neigh_ind_list
+                results = _to_object_array(neigh_ind_list)
 
         elif self._fit_method in ['ball_tree', 'kd_tree']:
             if issparse(X):

diff --git a/sklearn/neighbors/tests/test_neighbors.py b/sklearn/neighbors/tests/test_neighbors.py
@@ -649,6 +649,30 @@ def test_radius_neighbors_boundary_handling():
         assert_array_equal(results[0], [0, 1])
 
 
+def test_radius_neighbors_returns_array_of_objects():
+    # check that we can pass precomputed distances to
+    # NearestNeighbors.radius_neighbors()
+    # non-regression test for
+    # https://github.com/scikit-learn/scikit-learn/issues/16036
+    X = csr_matrix(np.ones((4, 4)))
+    X.setdiag([0, 0, 0, 0])
+
+    nbrs = neighbors.NearestNeighbors(radius=0.5, algorithm='auto',
+                                      leaf_size=30,
+                                      metric='precomputed').fit(X)
+    neigh_dist, neigh_ind = nbrs.radius_neighbors(X, return_distance=True)
+
+    expected_dist = np.empty(X.shape[0], dtype=object)
+    expected_dist[:] = [np.array([0]), np.array([0]), np.array([0]),
+                        np.array([0])]
+    expected_ind = np.empty(X.shape[0], dtype=object)
+    expected_ind[:] = [np.array([0]), np.array([1]), np.array([2]),
+                       np.array([3])]
+
+    assert_array_equal(neigh_dist, expected_dist)
+    assert_array_equal(neigh_ind, expected_ind)
+
+
 def test_RadiusNeighborsClassifier_multioutput():
     # Test k-NN classifier on multioutput data
     rng = check_random_state(0)

diff --git a/sklearn/preprocessing/tests/test_label.py b/sklearn/preprocessing/tests/test_label.py
@@ -14,6 +14,7 @@
 from sklearn.utils._testing import assert_array_equal
 from sklearn.utils._testing import assert_warns_message
 from sklearn.utils._testing import ignore_warnings
+from sklearn.utils import _to_object_array
 
 from sklearn.preprocessing._label import LabelBinarizer
 from sklearn.preprocessing._label import MultiLabelBinarizer
@@ -433,8 +434,7 @@ def test_multilabel_binarizer_same_length_sequence():
 
 
 def test_multilabel_binarizer_non_integer_labels():
-    tuple_classes = np.empty(3, dtype=object)
-    tuple_classes[:] = [(1,), (2,), (3,)]
+    tuple_classes = _to_object_array([(1,), (2,), (3,)])
     inputs = [
         ([('2', '3'), ('1',), ('1', '2')], ['1', '2', '3']),
         ([('b', 'c'), ('a',), ('a', 'b')], ['a', 'b', 'c']),

diff --git a/sklearn/utils/__init__.py b/sklearn/utils/__init__.py
@@ -819,6 +819,45 @@ def tosequence(x):
         return list(x)
 
 
+def _to_object_array(sequence):
+    """Convert sequence to a 1-D NumPy array of object dtype.
+
+    numpy.array constructor has a similar use but it's output
+    is ambiguous. It can be 1-D NumPy array of object dtype if
+    the input is a ragged array, but if the input is a list of
+    equal length arrays, then the output is a 2D numpy.array.
+    _to_object_array solves this ambiguity by guarantying that
+    the output is a 1-D NumPy array of objects for any input.
+
+    Parameters
+    ----------
+    sequence : array-like of shape (n_elements,)
+        The sequence to be converted.
+
+    Returns
+    -------
+    out : ndarray of shape (n_elements,), dtype=object
+        The converted sequence into a 1-D NumPy array of object dtype.
+
+    Examples
+    --------
+    >>> import numpy as np
+    >>> from sklearn.utils import _to_object_array
+    >>> _to_object_array([np.array([0]), np.array([1])])
+    array([array([0]), array([1])], dtype=object)
+    >>> _to_object_array([np.array([0]), np.array([1, 2])])
+    array([array([0]), array([1, 2])], dtype=object)
+    >>> np.array([np.array([0]), np.array([1])])
+    array([[0],
+       [1]])
+    >>> np.array([np.array([0]), np.array([1, 2])])
+    array([array([0]), array([1, 2])], dtype=object)
+    """
+    out = np.empty(len(sequence), dtype=object)
+    out[:] = sequence
+    return out
+
+
 def indices_to_mask(indices, mask_length):
     """Convert list of indices to boolean mask.
 

diff --git a/sklearn/utils/tests/test_utils.py b/sklearn/utils/tests/test_utils.py
@@ -28,6 +28,7 @@
 from sklearn.utils import _message_with_time, _print_elapsed_time
 from sklearn.utils import get_chunk_n_rows
 from sklearn.utils import is_scalar_nan
+from sklearn.utils import _to_object_array
 from sklearn.utils._mocking import MockDataFrame
 from sklearn import config_context
 
@@ -646,3 +647,14 @@ def test_deprecation_joblib_api(tmpdir):
 
     from sklearn.utils._joblib import joblib
     del joblib.parallel.BACKENDS['failing']
+
+
+@pytest.mark.parametrize(
+    "sequence",
+    [[np.array(1), np.array(2)], [[1, 2], [3, 4]]]
+)
+def test_to_object_array(sequence):
+    out = _to_object_array(sequence)
+    assert isinstance(out, np.ndarray)
+    assert out.dtype.kind == 'O'
+    assert out.ndim == 1