feat(types) Numpy.typing.NDArray (#5212)

* tests passing

* lint

* add comment

* remove empty tuple[()]

* test io_name

* style: pre-commit fixes

* remove accidental >

Signed-off-by: Michael Carlstrom <[email protected]>

* try T

* make both const_name

Signed-off-by: Michael Carlstrom <[email protected]>

* try and treat as string

* style: pre-commit fixes

* Update Numpy type hints

* style: pre-commit fixes

* re-run ci

Signed-off-by: Michael Carlstrom <[email protected]>

* re-run ci

Signed-off-by: Michael Carlstrom <[email protected]>

* remove escape characters

* Added tests for ArrayLike in signatures and fixed wrong types for Refs

---------

Signed-off-by: Michael Carlstrom <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Tim Ohliger <[email protected]>

Author: 3 people

include/pybind11/eigen/matrix.h

@@ -225,19 +225,22 @@ struct EigenProps {
         = !show_c_contiguous && show_order && requires_col_major;
 
     static constexpr auto descriptor
-        = const_name("numpy.ndarray[") + npy_format_descriptor<Scalar>::name + const_name("[")
+        = const_name("typing.Annotated[")
+          + io_name("numpy.typing.ArrayLike, ", "numpy.typing.NDArray[")
+          + npy_format_descriptor<Scalar>::name + io_name("", "]") + const_name(", \"[")
           + const_name<fixed_rows>(const_name<(size_t) rows>(), const_name("m")) + const_name(", ")
-          + const_name<fixed_cols>(const_name<(size_t) cols>(), const_name("n")) + const_name("]")
-          +
+          + const_name<fixed_cols>(const_name<(size_t) cols>(), const_name("n"))
+          + const_name("]\"")
           // For a reference type (e.g. Ref<MatrixXd>) we have other constraints that might need to
           // be satisfied: writeable=True (for a mutable reference), and, depending on the map's
           // stride options, possibly f_contiguous or c_contiguous.  We include them in the
           // descriptor output to provide some hint as to why a TypeError is occurring (otherwise
-          // it can be confusing to see that a function accepts a 'numpy.ndarray[float64[3,2]]' and
-          // an error message that you *gave* a numpy.ndarray of the right type and dimensions.
-          const_name<show_writeable>(", flags.writeable", "")
-          + const_name<show_c_contiguous>(", flags.c_contiguous", "")
-          + const_name<show_f_contiguous>(", flags.f_contiguous", "") + const_name("]");
+          // it can be confusing to see that a function accepts a
+          // 'typing.Annotated[numpy.typing.NDArray[numpy.float64], "[3,2]"]' and an error message
+          // that you *gave* a numpy.ndarray of the right type and dimensions.
+          + const_name<show_writeable>(", \"flags.writeable\"", "")
+          + const_name<show_c_contiguous>(", \"flags.c_contiguous\"", "")
+          + const_name<show_f_contiguous>(", \"flags.f_contiguous\"", "") + const_name("]");
 };
 
 // Casts an Eigen type to numpy array.  If given a base, the numpy array references the src data,
@@ -441,7 +444,9 @@ struct eigen_map_caster {
         }
     }
 
-    static constexpr auto name = props::descriptor;
+    // return_descr forces the use of NDArray instead of ArrayLike in args
+    // since Ref<...> args can only accept arrays.
+    static constexpr auto name = return_descr(props::descriptor);
 
     // Explicitly delete these: support python -> C++ conversion on these (i.e. these can be return
     // types but not bound arguments).  We still provide them (with an explicitly delete) so that

include/pybind11/eigen/tensor.h

@@ -124,13 +124,16 @@ struct eigen_tensor_helper<
 template <typename Type, bool ShowDetails, bool NeedsWriteable = false>
 struct get_tensor_descriptor {
     static constexpr auto details
-        = const_name<NeedsWriteable>(", flags.writeable", "") + const_name
+        = const_name<NeedsWriteable>(", \"flags.writeable\"", "") + const_name
               < static_cast<int>(Type::Layout)
-          == static_cast<int>(Eigen::RowMajor) > (", flags.c_contiguous", ", flags.f_contiguous");
+          == static_cast<int>(Eigen::RowMajor)
+                 > (", \"flags.c_contiguous\"", ", \"flags.f_contiguous\"");
     static constexpr auto value
-        = const_name("numpy.ndarray[") + npy_format_descriptor<typename Type::Scalar>::name
-          + const_name("[") + eigen_tensor_helper<remove_cv_t<Type>>::dimensions_descriptor
-          + const_name("]") + const_name<ShowDetails>(details, const_name("")) + const_name("]");
+        = const_name("typing.Annotated[")
+          + io_name("numpy.typing.ArrayLike, ", "numpy.typing.NDArray[")
+          + npy_format_descriptor<typename Type::Scalar>::name + io_name("", "]")
+          + const_name(", \"[") + eigen_tensor_helper<remove_cv_t<Type>>::dimensions_descriptor
+          + const_name("]\"") + const_name<ShowDetails>(details, const_name("")) + const_name("]");
 };
 
 // When EIGEN_AVOID_STL_ARRAY is defined, Eigen::DSizes<T, 0> does not have the begin() member
@@ -502,7 +505,10 @@ struct type_caster<Eigen::TensorMap<Type, Options>,
     std::unique_ptr<MapType> value;
 
 public:
-    static constexpr auto name = get_tensor_descriptor<Type, true, needs_writeable>::value;
+    // return_descr forces the use of NDArray instead of ArrayLike since refs can only reference
+    // arrays
+    static constexpr auto name
+        = return_descr(get_tensor_descriptor<Type, true, needs_writeable>::value);
     explicit operator MapType *() { return value.get(); }
     explicit operator MapType &() { return *value; }
     explicit operator MapType &&() && { return std::move(*value); }

include/pybind11/numpy.h

@@ -175,7 +175,6 @@ inline numpy_internals &get_numpy_internals() {
 PYBIND11_NOINLINE module_ import_numpy_core_submodule(const char *submodule_name) {
     module_ numpy = module_::import("numpy");
     str version_string = numpy.attr("__version__");
-
     module_ numpy_lib = module_::import("numpy.lib");
     object numpy_version = numpy_lib.attr("NumpyVersion")(version_string);
     int major_version = numpy_version.attr("major").cast<int>();
@@ -2183,7 +2182,8 @@ vectorize_helper<Func, Return, Args...> vectorize_extractor(const Func &f, Retur
 template <typename T, int Flags>
 struct handle_type_name<array_t<T, Flags>> {
     static constexpr auto name
-        = const_name("numpy.ndarray[") + npy_format_descriptor<T>::name + const_name("]");
+        = io_name("typing.Annotated[numpy.typing.ArrayLike, ", "numpy.typing.NDArray[")
+          + npy_format_descriptor<T>::name + const_name("]");
 };
 
 PYBIND11_NAMESPACE_END(detail)

tests/test_eigen_matrix.cpp

@@ -440,4 +440,8 @@ TEST_SUBMODULE(eigen_matrix, m) {
         py::module_::import("numpy").attr("ones")(10);
         return v[0](5);
     });
+    m.def("round_trip_vector", [](const Eigen::VectorXf &x) -> Eigen::VectorXf { return x; });
+    m.def("round_trip_dense", [](const DenseMatrixR &m) -> DenseMatrixR { return m; });
+    m.def("round_trip_dense_ref",
+          [](const Eigen::Ref<DenseMatrixR> &m) -> Eigen::Ref<DenseMatrixR> { return m; });
 }

tests/test_eigen_matrix.py

@@ -95,19 +95,20 @@ def test_mutator_descriptors():
     with pytest.raises(TypeError) as excinfo:
         m.fixed_mutator_r(zc)
     assert (
-        "(arg0: numpy.ndarray[numpy.float32[5, 6],"
-        " flags.writeable, flags.c_contiguous]) -> None" in str(excinfo.value)
+        '(arg0: typing.Annotated[numpy.typing.NDArray[numpy.float32], "[5, 6]",'
+        ' "flags.writeable", "flags.c_contiguous"]) -> None' in str(excinfo.value)
     )
     with pytest.raises(TypeError) as excinfo:
         m.fixed_mutator_c(zr)
     assert (
-        "(arg0: numpy.ndarray[numpy.float32[5, 6],"
-        " flags.writeable, flags.f_contiguous]) -> None" in str(excinfo.value)
+        '(arg0: typing.Annotated[numpy.typing.NDArray[numpy.float32], "[5, 6]",'
+        ' "flags.writeable", "flags.f_contiguous"]) -> None' in str(excinfo.value)
     )
     with pytest.raises(TypeError) as excinfo:
         m.fixed_mutator_a(np.array([[1, 2], [3, 4]], dtype="float32"))
-    assert "(arg0: numpy.ndarray[numpy.float32[5, 6], flags.writeable]) -> None" in str(
-        excinfo.value
+    assert (
+        '(arg0: typing.Annotated[numpy.typing.NDArray[numpy.float32], "[5, 6]", "flags.writeable"]) -> None'
+        in str(excinfo.value)
     )
     zr.flags.writeable = False
     with pytest.raises(TypeError):
@@ -201,7 +202,7 @@ def test_negative_stride_from_python(msg):
         msg(excinfo.value)
         == """
         double_threer(): incompatible function arguments. The following argument types are supported:
-            1. (arg0: numpy.ndarray[numpy.float32[1, 3], flags.writeable]) -> None
+            1. (arg0: typing.Annotated[numpy.typing.NDArray[numpy.float32], "[1, 3]", "flags.writeable"]) -> None
 
         Invoked with: """
         + repr(np.array([5.0, 4.0, 3.0], dtype="float32"))
@@ -213,7 +214,7 @@ def test_negative_stride_from_python(msg):
         msg(excinfo.value)
         == """
         double_threec(): incompatible function arguments. The following argument types are supported:
-            1. (arg0: numpy.ndarray[numpy.float32[3, 1], flags.writeable]) -> None
+            1. (arg0: typing.Annotated[numpy.typing.NDArray[numpy.float32], "[3, 1]", "flags.writeable"]) -> None
 
         Invoked with: """
         + repr(np.array([7.0, 4.0, 1.0], dtype="float32"))
@@ -634,37 +635,37 @@ def test_nocopy_wrapper():
     with pytest.raises(TypeError) as excinfo:
         m.get_elem_nocopy(int_matrix_colmajor)
     assert "get_elem_nocopy(): incompatible function arguments." in str(excinfo.value)
-    assert ", flags.f_contiguous" in str(excinfo.value)
+    assert ', "flags.f_contiguous"' in str(excinfo.value)
     assert m.get_elem_nocopy(dbl_matrix_colmajor) == 8
     with pytest.raises(TypeError) as excinfo:
         m.get_elem_nocopy(int_matrix_rowmajor)
     assert "get_elem_nocopy(): incompatible function arguments." in str(excinfo.value)
-    assert ", flags.f_contiguous" in str(excinfo.value)
+    assert ', "flags.f_contiguous"' in str(excinfo.value)
     with pytest.raises(TypeError) as excinfo:
         m.get_elem_nocopy(dbl_matrix_rowmajor)
     assert "get_elem_nocopy(): incompatible function arguments." in str(excinfo.value)
-    assert ", flags.f_contiguous" in str(excinfo.value)
+    assert ', "flags.f_contiguous"' in str(excinfo.value)
 
     # For the row-major test, we take a long matrix in row-major, so only the third is allowed:
     with pytest.raises(TypeError) as excinfo:
         m.get_elem_rm_nocopy(int_matrix_colmajor)
     assert "get_elem_rm_nocopy(): incompatible function arguments." in str(
         excinfo.value
     )
-    assert ", flags.c_contiguous" in str(excinfo.value)
+    assert ', "flags.c_contiguous"' in str(excinfo.value)
     with pytest.raises(TypeError) as excinfo:
         m.get_elem_rm_nocopy(dbl_matrix_colmajor)
     assert "get_elem_rm_nocopy(): incompatible function arguments." in str(
         excinfo.value
     )
-    assert ", flags.c_contiguous" in str(excinfo.value)
+    assert ', "flags.c_contiguous"' in str(excinfo.value)
     assert m.get_elem_rm_nocopy(int_matrix_rowmajor) == 8
     with pytest.raises(TypeError) as excinfo:
         m.get_elem_rm_nocopy(dbl_matrix_rowmajor)
     assert "get_elem_rm_nocopy(): incompatible function arguments." in str(
         excinfo.value
     )
-    assert ", flags.c_contiguous" in str(excinfo.value)
+    assert ', "flags.c_contiguous"' in str(excinfo.value)
 
 
 def test_eigen_ref_life_support():
@@ -700,25 +701,25 @@ def test_dense_signature(doc):
     assert (
         doc(m.double_col)
         == """
-        double_col(arg0: numpy.ndarray[numpy.float32[m, 1]]) -> numpy.ndarray[numpy.float32[m, 1]]
+        double_col(arg0: typing.Annotated[numpy.typing.ArrayLike, numpy.float32, "[m, 1]"]) -> typing.Annotated[numpy.typing.NDArray[numpy.float32], "[m, 1]"]
     """
     )
     assert (
         doc(m.double_row)
         == """
-        double_row(arg0: numpy.ndarray[numpy.float32[1, n]]) -> numpy.ndarray[numpy.float32[1, n]]
+        double_row(arg0: typing.Annotated[numpy.typing.ArrayLike, numpy.float32, "[1, n]"]) -> typing.Annotated[numpy.typing.NDArray[numpy.float32], "[1, n]"]
     """
     )
     assert doc(m.double_complex) == (
         """
-        double_complex(arg0: numpy.ndarray[numpy.complex64[m, 1]])"""
-        """ -> numpy.ndarray[numpy.complex64[m, 1]]
+        double_complex(arg0: typing.Annotated[numpy.typing.ArrayLike, numpy.complex64, "[m, 1]"])"""
+        """ -> typing.Annotated[numpy.typing.NDArray[numpy.complex64], "[m, 1]"]
     """
     )
     assert doc(m.double_mat_rm) == (
         """
-        double_mat_rm(arg0: numpy.ndarray[numpy.float32[m, n]])"""
-        """ -> numpy.ndarray[numpy.float32[m, n]]
+        double_mat_rm(arg0: typing.Annotated[numpy.typing.ArrayLike, numpy.float32, "[m, n]"])"""
+        """ -> typing.Annotated[numpy.typing.NDArray[numpy.float32], "[m, n]"]
     """
     )
 
@@ -817,3 +818,22 @@ def test_custom_operator_new():
     o = m.CustomOperatorNew()
     np.testing.assert_allclose(o.a, 0.0)
     np.testing.assert_allclose(o.b.diagonal(), 1.0)
+
+
+def test_arraylike_signature(doc):
+    assert doc(m.round_trip_vector) == (
+        'round_trip_vector(arg0: typing.Annotated[numpy.typing.ArrayLike, numpy.float32, "[m, 1]"])'
+        ' -> typing.Annotated[numpy.typing.NDArray[numpy.float32], "[m, 1]"]'
+    )
+    assert doc(m.round_trip_dense) == (
+        'round_trip_dense(arg0: typing.Annotated[numpy.typing.ArrayLike, numpy.float32, "[m, n]"])'
+        ' -> typing.Annotated[numpy.typing.NDArray[numpy.float32], "[m, n]"]'
+    )
+    assert doc(m.round_trip_dense_ref) == (
+        'round_trip_dense_ref(arg0: typing.Annotated[numpy.typing.NDArray[numpy.float32], "[m, n]", "flags.writeable", "flags.c_contiguous"])'
+        ' -> typing.Annotated[numpy.typing.NDArray[numpy.float32], "[m, n]", "flags.writeable", "flags.c_contiguous"]'
+    )
+    m.round_trip_vector([1.0, 2.0])
+    m.round_trip_dense([[1.0, 2.0], [3.0, 4.0]])
+    with pytest.raises(TypeError, match="incompatible function arguments"):
+        m.round_trip_dense_ref([[1.0, 2.0], [3.0, 4.0]])

tests/test_eigen_tensor.py

@@ -271,23 +271,46 @@ def test_round_trip_references_actually_refer(m):
 @pytest.mark.parametrize("m", submodules)
 def test_doc_string(m, doc):
     assert (
-        doc(m.copy_tensor) == "copy_tensor() -> numpy.ndarray[numpy.float64[?, ?, ?]]"
+        doc(m.copy_tensor)
+        == 'copy_tensor() -> typing.Annotated[numpy.typing.NDArray[numpy.float64], "[?, ?, ?]"]'
     )
     assert (
         doc(m.copy_fixed_tensor)
-        == "copy_fixed_tensor() -> numpy.ndarray[numpy.float64[3, 5, 2]]"
+        == 'copy_fixed_tensor() -> typing.Annotated[numpy.typing.NDArray[numpy.float64], "[3, 5, 2]"]'
     )
     assert (
         doc(m.reference_const_tensor)
-        == "reference_const_tensor() -> numpy.ndarray[numpy.float64[?, ?, ?]]"
+        == 'reference_const_tensor() -> typing.Annotated[numpy.typing.NDArray[numpy.float64], "[?, ?, ?]"]'
     )
 
-    order_flag = f"flags.{m.needed_options.lower()}_contiguous"
+    order_flag = f'"flags.{m.needed_options.lower()}_contiguous"'
     assert doc(m.round_trip_view_tensor) == (
-        f"round_trip_view_tensor(arg0: numpy.ndarray[numpy.float64[?, ?, ?], flags.writeable, {order_flag}])"
-        f" -> numpy.ndarray[numpy.float64[?, ?, ?], flags.writeable, {order_flag}]"
+        f'round_trip_view_tensor(arg0: typing.Annotated[numpy.typing.NDArray[numpy.float64], "[?, ?, ?]", "flags.writeable", {order_flag}])'
+        f' -> typing.Annotated[numpy.typing.NDArray[numpy.float64], "[?, ?, ?]", "flags.writeable", {order_flag}]'
     )
     assert doc(m.round_trip_const_view_tensor) == (
-        f"round_trip_const_view_tensor(arg0: numpy.ndarray[numpy.float64[?, ?, ?], {order_flag}])"
-        " -> numpy.ndarray[numpy.float64[?, ?, ?]]"
+        f'round_trip_const_view_tensor(arg0: typing.Annotated[numpy.typing.NDArray[numpy.float64], "[?, ?, ?]", {order_flag}])'
+        ' -> typing.Annotated[numpy.typing.NDArray[numpy.float64], "[?, ?, ?]"]'
     )
+
+
+@pytest.mark.parametrize("m", submodules)
+def test_arraylike_signature(m, doc):
+    order_flag = f'"flags.{m.needed_options.lower()}_contiguous"'
+    assert doc(m.round_trip_tensor) == (
+        'round_trip_tensor(arg0: typing.Annotated[numpy.typing.ArrayLike, numpy.float64, "[?, ?, ?]"])'
+        ' -> typing.Annotated[numpy.typing.NDArray[numpy.float64], "[?, ?, ?]"]'
+    )
+    assert doc(m.round_trip_tensor_noconvert) == (
+        'round_trip_tensor_noconvert(tensor: typing.Annotated[numpy.typing.NDArray[numpy.float64], "[?, ?, ?]"])'
+        ' -> typing.Annotated[numpy.typing.NDArray[numpy.float64], "[?, ?, ?]"]'
+    )
+    assert doc(m.round_trip_view_tensor) == (
+        f'round_trip_view_tensor(arg0: typing.Annotated[numpy.typing.NDArray[numpy.float64], "[?, ?, ?]", "flags.writeable", {order_flag}])'
+        f' -> typing.Annotated[numpy.typing.NDArray[numpy.float64], "[?, ?, ?]", "flags.writeable", {order_flag}]'
+    )
+    m.round_trip_tensor(tensor_ref.tolist())
+    with pytest.raises(TypeError, match="incompatible function arguments"):
+        m.round_trip_tensor_noconvert(tensor_ref.tolist())
+    with pytest.raises(TypeError, match="incompatible function arguments"):
+        m.round_trip_view_tensor(tensor_ref.tolist())

tests/test_numpy_array.cpp

@@ -586,4 +586,13 @@ TEST_SUBMODULE(numpy_array, sm) {
     sm.def("return_array_pyobject_ptr_from_list", return_array_from_list<PyObject *>);
     sm.def("return_array_handle_from_list", return_array_from_list<py::handle>);
     sm.def("return_array_object_from_list", return_array_from_list<py::object>);
+
+    sm.def(
+        "round_trip_array_t",
+        [](const py::array_t<float> &x) -> py::array_t<float> { return x; },
+        py::arg("x"));
+    sm.def(
+        "round_trip_array_t_noconvert",
+        [](const py::array_t<float> &x) -> py::array_t<float> { return x; },
+        py::arg("x").noconvert());
 }

tests/test_numpy_array.py

@@ -321,13 +321,13 @@ def test_overload_resolution(msg):
         msg(excinfo.value)
         == """
         overloaded(): incompatible function arguments. The following argument types are supported:
-            1. (arg0: numpy.ndarray[numpy.float64]) -> str
-            2. (arg0: numpy.ndarray[numpy.float32]) -> str
-            3. (arg0: numpy.ndarray[numpy.int32]) -> str
-            4. (arg0: numpy.ndarray[numpy.uint16]) -> str
-            5. (arg0: numpy.ndarray[numpy.int64]) -> str
-            6. (arg0: numpy.ndarray[numpy.complex128]) -> str
-            7. (arg0: numpy.ndarray[numpy.complex64]) -> str
+            1. (arg0: typing.Annotated[numpy.typing.ArrayLike, numpy.float64]) -> str
+            2. (arg0: typing.Annotated[numpy.typing.ArrayLike, numpy.float32]) -> str
+            3. (arg0: typing.Annotated[numpy.typing.ArrayLike, numpy.int32]) -> str
+            4. (arg0: typing.Annotated[numpy.typing.ArrayLike, numpy.uint16]) -> str
+            5. (arg0: typing.Annotated[numpy.typing.ArrayLike, numpy.int64]) -> str
+            6. (arg0: typing.Annotated[numpy.typing.ArrayLike, numpy.complex128]) -> str
+            7. (arg0: typing.Annotated[numpy.typing.ArrayLike, numpy.complex64]) -> str
 
         Invoked with: 'not an array'
     """
@@ -343,8 +343,8 @@ def test_overload_resolution(msg):
     assert m.overloaded3(np.array([1], dtype="intc")) == "int"
     expected_exc = """
         overloaded3(): incompatible function arguments. The following argument types are supported:
-            1. (arg0: numpy.ndarray[numpy.int32]) -> str
-            2. (arg0: numpy.ndarray[numpy.float64]) -> str
+            1. (arg0: numpy.typing.NDArray[numpy.int32]) -> str
+            2. (arg0: numpy.typing.NDArray[numpy.float64]) -> str
 
         Invoked with: """
 
@@ -528,7 +528,7 @@ def test_index_using_ellipsis():
     ],
 )
 def test_format_descriptors_for_floating_point_types(test_func):
-    assert "numpy.ndarray[numpy.float" in test_func.__doc__
+    assert "numpy.typing.ArrayLike, numpy.float" in test_func.__doc__
 
 
 @pytest.mark.parametrize("forcecast", [False, True])
@@ -687,3 +687,17 @@ def test_return_array_object_cpp_loop(return_array, unwrap):
     assert isinstance(arr_from_list, np.ndarray)
     assert arr_from_list.dtype == np.dtype("O")
     assert unwrap(arr_from_list) == [6, "seven", -8.0]
+
+
+def test_arraylike_signature(doc):
+    assert (
+        doc(m.round_trip_array_t)
+        == "round_trip_array_t(x: typing.Annotated[numpy.typing.ArrayLike, numpy.float32]) -> numpy.typing.NDArray[numpy.float32]"
+    )
+    assert (
+        doc(m.round_trip_array_t_noconvert)
+        == "round_trip_array_t_noconvert(x: numpy.typing.NDArray[numpy.float32]) -> numpy.typing.NDArray[numpy.float32]"
+    )
+    m.round_trip_array_t([1, 2, 3])
+    with pytest.raises(TypeError, match="incompatible function arguments"):
+        m.round_trip_array_t_noconvert([1, 2, 3])

tests/test_numpy_dtypes.py

@@ -373,7 +373,7 @@ def test_complex_array():
 def test_signature(doc):
     assert (
         doc(m.create_rec_nested)
-        == "create_rec_nested(arg0: int) -> numpy.ndarray[NestedStruct]"
+        == "create_rec_nested(arg0: int) -> numpy.typing.NDArray[NestedStruct]"
     )