numpy arrays scalar

[fbriol]

Hello I need different functions taking as arguments a value of « numpy arrays scalar » type.

It is possible to write prototypes like this:

PYBIND11_MODULE(np_scalar, m) {
    m.def("overloaded", [](py::array_t<double>) {
        return "double";
    });
    m.def("overloaded", [](py::array_t<std::int8_t>) {
        return "int8";
    });
    m.def("overloaded", [](py::array_t<std::int16_t>) {
        return "int16";
    });
}

However, this method does not allow to intercept scalars, for example:

>>> import np_scalar
>>> np_scalar.overloaded(numpy.array([1], dtype='int8'))
'int8'
>>> np_scalar.overloaded(numpy.int8(1))
'double'

How do we intercept numpy arrays scalar?

Thank you very much for your help.

[chhenning]

I'm not an expert but can you do:

m.def("overloaded", [](py::array a)
{
    auto buf = a.request();
    auto fmt = buf.format;

    if (fmt == "int32")
    {
        return "int32";
    }
    else if (fmt == "float64")
    {
        return "float64";
    }
    else
    {
        return "unknown";
    }
});

[fbriol]

This doesn't work because this prototype only accepts array, not scalar.

m.def("overloaded", [](py::array a)
{
    auto buf = a.request();
    auto fmt = buf.format;

    if (fmt == "i")
    {
        return "int32";
    }
    else if (fmt == "h")
    {
        return "int16";
    }
    else
    {
        return "unknown";
    }
});

>>> import np_scalar
>>> np_scalar.overloaded(numpy.array([1], dtype="int32"))
'int32'
>>> np_scalar.overloaded(numpy.int32(1))
.../...
TypeError: overloaded(): incompatible function arguments. The following argument types are 
supported:
    1. (arg0: array) -> str

[vanossj]

I was just doing a similar thing. It should work works (in numpy >= 1.15.0) with py::buffer instead of py::array
However numpy doesn't expose the buffer info of scalers correctly, numpy/numpy#10265, it always reports them as byte arrays.

m.def("overloaded", [](py::buffer a)
{
    auto buf = a.request();
    if (buf.ndim == 0)
    {
        return "scaler of type " + buf.format;
    }
    else
    {
        return "ndarray of type " + buf.format;
    }
});

Numpy < 1.15.0

>>> import np_scalar
>>> np_scalar.overloaded(numpy.array([1], dtype="int32"))
'ndarray of type i'
>>> np_scalar.overloaded(numpy.int32(1))
'ndarray of type B'

Numpy >= 1.15.0

>>> import np_scalar
>>> np_scalar.overloaded(numpy.array([1], dtype="int32"))
'ndarray of type i'
>>> np_scalar.overloaded(numpy.int32(1))
'scaler of type i'

As a workaround you could use ctypes to specify scalers

>>> import ctypes
>>> np_scalar.overloaded(ctypes.c_int32(1))
'scaler of type <i'

edited for numpy 1.15 release

[eacousineau]

@fbriol If you're still monitoring this issue, can you take a gander at the solutions (/ workarounds) in these two issues?

#1776
#1785

Both were for leveraging NumPy to gain access to "off the beaten path" floats, like float128 and float16. One of the workarounds I posted was making a npy_scalar_caster - however, that's only useful if you want to capture a specific type, and it could shadow (or error out) if you try and use it with already supported types like float or double:
https://github.com/eacousineau/repro/blob/5e6acf6/python/pybind11/custom_tests/test_numpy_issue1785.cc#L21-L47