I'm attempting to bind a xt::pyarray<xt::pyarray<double>> to a Numpy array of arrays, but I am getting compilation errors instantiating the pyarray. The error is a bit long to include here, the relevant bit is

pybind11/include/pybind11/numpy.h:260:81: error: no member named 'dtype' in 'pybind11::detail::npy_format_descriptor<xt::pyarray<double, 16>, void>'
return detail::npy_format_descriptor<typename std::remove_cv::type>::dtype();

coming from the call to pybind11::dtype::of<T> in the return ofpyarray::raw_array_t. Unfortunately, I don't know enough about the inner workings of pybind11 to determine if the problem lies there, or whether this is a xtensor-python issue.

The problem I am trying to solve is to provide Python bindings to a code that uses xtensor xarray. Using xt::xarray<xt::xarray<double>> in a C++-only situation works as expected. Actually, my case is limited in that I always have three inner arrays, so in C++ I would prefer to use std::array<xt::xarray<double>, 3> (or perhaps xt::xtensor<xt::xarray<double>, 3>) but I do not know how to transparently (without copies) expose those as an ndarray of three ndarrays to Python.

Example code that produces the compilation error:

#include <numeric>
#include "pybind11/pybind11.h"
#include "xtensor/xarray.hpp"
#include "xtensor/xmath.hpp"
#include "xtensor-python/pyarray.hpp"

double sum_of_dim(xt::pyarray<xt::pyarray<double>> & m, const int dim)
{
    auto d = m(dim); 
    return std::accumulate(d.begin(), d.end(), 0.0);
}

PYBIND11_PLUGIN(example1_m)
{
    pybind11::module m("example1_m", "Array of array test");
    m.def("sum_of_dim", sum_of_dim, "Array of array function test");
    return m.ptr();
}