inconsistent numpy dtype in python and C++

[zhangguy]

Hi, I'm trying to use the boost numpy bindings. If I pass two ndarray with different dtype to C++, they don't preserve the correct dtype. Here is a simple example: dummy.cpp:

#include <iostream>
#include <boost/python.hpp>
#include <boost/python/numpy.hpp>

namespace py = boost::python;
namespace np = boost::python::numpy;

void dummy(np::ndarray const& a, np::ndarray const& b) {

    char const* at = py::extract<char const* >(py::str(a.get_dtype()));
    char const* bt = py::extract<char const* >(py::str(b.get_dtype()));

    std::cout << at << std::endl;
    std::cout << bt << std::endl;
}

BOOST_PYTHON_MODULE(dummy) {
   Py_Initialize(); // need initialize in BOOST_PYTHON_MODULE, or will get segmentation fault and core dump 
   np::initialize();
   py::def("dummy", dummy);
}

compile

g++ -lboost_python38 -lboost_numpy38 -L$CONDA_PREFIX/lib -I$CONDA_PREFIX/include -I$CONDA_PREFIX/include/python3.8 -fPIC --shared -o dummy.so dummy.cpp

run_dummpy.py

import numpy as np
import dummy

a = np.array([[1, 2, 3], [4, 5, 6]], dtype = np.float32)
print("python", a.dtype)
b = np.array([[1, 2, 3], [4, 5, 6]], dtype = np.double)
print("python", b.dtype)

dummy.dummy(a, b)

The output is inconsistent. The first time I run it by python run_dummy.py it gives

python float32
python float64
__str__
float64

If I re-run it python run_dummy.py, it gives

python float32
python float64
float64
float64

Still not the right output. I'm using python 3.8.2, numpy 1.18.1, and boost 1.72.0. boost is installed from anaconda conda install -c conda-forge boost Could you help? Thanks