Merge pull request #17973 from seberg/require-sequence-array-coercion

DEP: Futurewarn on requiring __len__ on array-likes

3 files changed, 146 insertions, 3 deletions

diff --git a/numpy/core/src/multiarray/array_coercion.c b/numpy/core/src/multiarray/array_coercion.c
index 603e9d93b..1eac401bc 100644
--- a/numpy/core/src/multiarray/array_coercion.c
+++ b/numpy/core/src/multiarray/array_coercion.c
@@ -922,6 +922,53 @@ PyArray_DiscoverDTypeAndShape_Recursive(
             Py_DECREF(arr);
             arr = NULL;
         }
+        else if (curr_dims > 0 && curr_dims != max_dims) {
+            /*
+             * Deprecated 2020-12-09, NumPy 1.20
+             *
+             * See https://github.com/numpy/numpy/issues/17965
+             * Shapely had objects which are not sequences but did export
+             * the array-interface (and so are arguably array-like).
+             * Previously numpy would not use array-like information during
+             * shape discovery, so that it ended up acting as if this was
+             * an (unknown) scalar but with the specified dtype.
+             * Thus we ignore "scalars" here, as the value stored in the
+             * array should be acceptable.
+             */
+            if (PyArray_NDIM(arr) > 0 && NPY_UNLIKELY(!PySequence_Check(obj))) {
+                if (PyErr_WarnFormat(PyExc_FutureWarning, 1,
+                        "The input object of type '%s' is an array-like "
+                        "implementing one of the corresponding protocols "
+                        "(`__array__`, `__array_interface__` or "
+                        "`__array_struct__`); but not a sequence (or 0-D). "
+                        "In the future, this object will be coerced as if it "
+                        "was first converted using `np.array(obj)`. "
+                        "To retain the old behaviour, you have to either "
+                        "modify the type '%s', or assign to an empty array "
+                        "created with `np.empty(correct_shape, dtype=object)`.",
+                        Py_TYPE(obj)->tp_name, Py_TYPE(obj)->tp_name) < 0) {
+                    Py_DECREF(arr);
+                    return -1;
+                }
+                /*
+                 * Strangely enough, even though we threw away the result here,
+                 * we did use it during descriptor discovery, so promote it:
+                 */
+                if (update_shape(curr_dims, &max_dims, out_shape,
+                        0, NULL, NPY_FALSE, flags) < 0) {
+                    *flags |= FOUND_RAGGED_ARRAY;
+                    Py_DECREF(arr);
+                    return max_dims;
+                }
+                if (!(*flags & DESCRIPTOR_WAS_SET) && handle_promotion(
+                        out_descr, PyArray_DESCR(arr), fixed_DType, flags) < 0) {
+                    Py_DECREF(arr);
+                    return -1;
+                }
+                Py_DECREF(arr);
+                return max_dims;
+            }
+        }
     }
     if (arr != NULL) {
         /*
diff --git a/numpy/core/tests/test_array_coercion.py b/numpy/core/tests/test_array_coercion.py
index b966ee7b0..08b32dfcc 100644
--- a/numpy/core/tests/test_array_coercion.py
+++ b/numpy/core/tests/test_array_coercion.py
@@ -38,8 +38,18 @@ def arraylikes():
 
     yield subclass
 
+    class _SequenceLike():
+        # We are giving a warning that array-like's were also expected to be
+        # sequence-like in `np.array([array_like])`, this can be removed
+        # when the deprecation exired (started NumPy 1.20)
+        def __len__(self):
+            raise TypeError
+
+        def __getitem__(self):
+            raise TypeError
+
     # Array-interface
-    class ArrayDunder:
+    class ArrayDunder(_SequenceLike):
         def __init__(self, a):
             self.a = a
 
@@ -52,7 +62,7 @@ def arraylikes():
     yield param(memoryview, id="memoryview")
 
     # Array-interface
-    class ArrayInterface:
+    class ArrayInterface(_SequenceLike):
         def __init__(self, a):
             self.a = a  # need to hold on to keep interface valid
             self.__array_interface__ = a.__array_interface__
@@ -60,7 +70,7 @@ def arraylikes():
     yield param(ArrayInterface, id="__array_interface__")
 
     # Array-Struct
-    class ArrayStruct:
+    class ArrayStruct(_SequenceLike):
         def __init__(self, a):
             self.a = a  # need to hold on to keep struct valid
             self.__array_struct__ = a.__array_struct__
diff --git a/numpy/core/tests/test_deprecations.py b/numpy/core/tests/test_deprecations.py
index a67fe62c3..ed238da9f 100644
--- a/numpy/core/tests/test_deprecations.py
+++ b/numpy/core/tests/test_deprecations.py
@@ -773,6 +773,92 @@ class TestDeprecateSubarrayDTypeDuringArrayCoercion(_DeprecationTestCase):
         self.assert_deprecated(check)
 
 
+class TestFutureWarningArrayLikeNotIterable(_DeprecationTestCase):
+    # Deprecated 2020-12-09, NumPy 1.20
+    warning_cls = FutureWarning
+    message = "The input object of type.*but not a sequence"
+
+    @pytest.mark.parametrize("protocol",
+            ["__array__", "__array_interface__", "__array_struct__"])
+    def test_deprecated(self, protocol):
+        """Test that these objects give a warning since they are not 0-D,
+        not coerced at the top level `np.array(obj)`, but nested, and do
+        *not* define the sequence protocol.
+
+        NOTE: Tests for the versions including __len__ and __getitem__ exist
+              in `test_array_coercion.py` and they can be modified or ammended
+              when this deprecation expired.
+        """
+        blueprint = np.arange(10)
+        MyArr = type("MyArr", (), {protocol: getattr(blueprint, protocol)})
+        self.assert_deprecated(lambda: np.array([MyArr()], dtype=object))
+
+    @pytest.mark.parametrize("protocol",
+             ["__array__", "__array_interface__", "__array_struct__"])
+    def test_0d_not_deprecated(self, protocol):
+        # 0-D always worked (albeit it would use __float__ or similar for the
+        # conversion, which may not happen anymore)
+        blueprint = np.array(1.)
+        MyArr = type("MyArr", (), {protocol: getattr(blueprint, protocol)})
+        myarr = MyArr()
+
+        self.assert_not_deprecated(lambda: np.array([myarr], dtype=object))
+        res = np.array([myarr], dtype=object)
+        expected = np.empty(1, dtype=object)
+        expected[0] = myarr
+        assert_array_equal(res, expected)
+
+    @pytest.mark.parametrize("protocol",
+             ["__array__", "__array_interface__", "__array_struct__"])
+    def test_unnested_not_deprecated(self, protocol):
+        blueprint = np.arange(10)
+        MyArr = type("MyArr", (), {protocol: getattr(blueprint, protocol)})
+        myarr = MyArr()
+
+        self.assert_not_deprecated(lambda: np.array(myarr))
+        res = np.array(myarr)
+        assert_array_equal(res, blueprint)
+
+    @pytest.mark.parametrize("protocol",
+             ["__array__", "__array_interface__", "__array_struct__"])
+    def test_strange_dtype_handling(self, protocol):
+        """The old code would actually use the dtype from the array, but
+        then end up not using the array (for dimension discovery)
+        """
+        blueprint = np.arange(10).astype("f4")
+        MyArr = type("MyArr", (), {protocol: getattr(blueprint, protocol),
+                                   "__float__": lambda _: 0.5})
+        myarr = MyArr()
+
+        # Make sure we warn (and capture the FutureWarning)
+        with pytest.warns(FutureWarning, match=self.message):
+            res = np.array([[myarr]])
+
+        assert res.shape == (1, 1)
+        assert res.dtype == "f4"
+        assert res[0, 0] == 0.5
+
+    @pytest.mark.parametrize("protocol",
+             ["__array__", "__array_interface__", "__array_struct__"])
+    def test_assignment_not_deprecated(self, protocol):
+        # If the result is dtype=object we do not unpack a nested array or
+        # array-like, if it is nested at exactly the right depth.
+        # NOTE: We actually do still call __array__, etc. but ignore the result
+        #       in the end. For `dtype=object` we could optimize that away.
+        blueprint = np.arange(10).astype("f4")
+        MyArr = type("MyArr", (), {protocol: getattr(blueprint, protocol),
+                                   "__float__": lambda _: 0.5})
+        myarr = MyArr()
+
+        res = np.empty(3, dtype=object)
+        def set():
+            res[:] = [myarr, myarr, myarr]
+        self.assert_not_deprecated(set)
+        assert res[0] is myarr
+        assert res[1] is myarr
+        assert res[2] is myarr
+
+
 class TestDeprecatedUnpickleObjectScalar(_DeprecationTestCase):
     # Deprecated 2020-11-24, NumPy 1.20
     """