Structured type for Numpy array

[zhqrbitee]

Following instructions on https://pybind11.readthedocs.io/en/stable/advanced/pycpp/numpy.html#structured-types , I write some examples codes:

  struct A {
    A(int x, double y): x{x}, y{y} {};
    int x;
    double y;
  };

  py::class_<A>(m, "A")
      .def(py::init<int, double>(), "x"_a, "y"_a);

  PYBIND11_NUMPY_DTYPE(A, x, y);

  m.def("from_A_array", [] (const py::array_t<A>& array) {
    auto r = array.unchecked();
    auto len = r.shape(0);
    for (py::ssize_t i = 0; i < len; i++) {
      A a = r(i);
      std::cout << "(x: " << a.x << ", y: " << a.y << ")" << std::endl;
    }
  });

  m.def("to_A_array", []() {
    std::vector<A> tmp = {A{1, 1.0}, A{2, 2.0}};
    py::array_t<A> array(tmp.size(), tmp.data());
    return array;
  });

It works if I write codes like these:

>>> array = np.ndarray(shape=2, dtype=[('x', '<i4'), ('y', '<f8')])
>>> from_A_array(array)
(x: 0, y: 5e-315)
(x: 1073741824, y: 1)

But what I really want is that Numpy can recognize A as a valid structure dtype (not Python Object), e.g.:

>>> array = np.ndarray(shape=2, dtype=A)
>>> from_A_array(array)
TypeError: from_A_array(): incompatible function arguments. The following argument types are supported:
    1. (arg0: numpy.ndarray[A]) -> None
Invoked with: array([None, None], dtype=object)

Is this possible?