Is there any way to directly pass numpy.float16 dtype to pybind?

[FeixLiu]

Is there any way to directly pass numpy.float16 dtype to pybind?

[jiwaszki]

Two ways of doing that:

#include <pybind11/pybind11.h>
#include <pybind11/numpy.h>

// ...

m.def("is_float16", [](py::dtype &type)  // use of dtype class from numpy.h
      {
        if (type.is(py::dtype("float16")))
        {
          py::print("First overload.");
        }
        else
        {
          py::print("Not float16!");
        }
      });

m.def("is_float16", [](py::object &type)  // pass python object directly
      {
        if (py::dtype::from_args(type).is(py::dtype("float16")))  // use helper function
        {
          py::print("Second overload.");
        }
        else
        {
          py::print("Not float16!");
        }
      });

In [1]: import numpy as np
   ...: import mymodule as m

In [2]: m.is_float16(np.dtype("float16"))
First overload.

In [3]: m.is_float16(np.dtype("float32"))
Not float16!

In [4]: m.is_float16(np.float16)
Second overload.

In [5]: m.is_float16(np.int32)
Not float16!

Hope it helps!:)

[FeixLiu]

great~ thx a lot~