Cast NumPy array to/from custom C++ Matrix-class using pybind11

Question

I am trying to wrap my C++ code using pybind11. In C++ I have a class Matrix3D which acts as a 3-D array (i.e. with shape [n,m,p]). It has the following basic signature:

template <class T> class Matrix3D
{

  public:

    std::vector<T> data;
    std::vector<size_t> shape;
    std::vector<size_t> strides;

    Matrix3D<T>();
    Matrix3D<T>(std::vector<size_t>);
    Matrix3D<T>(const Matrix3D<T>&);

    T& operator() (int,int,int);

};

To minimize the wrapper code I would like to cast this class directly to and from a NumPy-array (copies are no problem). For example, I would like to directly wrap a function of the following signature:

Matrix3D<double> func ( const Matrix3D<double>& );

using the wrapper code

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

namespace py = pybind11;

PYBIND11_PLUGIN(example) {
  py::module m("example", "Module description");
  m.def("func", &func, "Function description" );
  return m.ptr();
}

Currently I have another function in-between that accepts and returns py::array_t<double>. But I would like to avoid having to write a wrapper function for each function by replacing it by some template.

This has been done for the Eigen-library (for arrays and (2-D) matrices). But the code is too involved for me to derive my own code from. Plus, I really only need to wrap only one, simple, class.

Answer 1

With the help of @kazemakase and @jagerman (the latter via the pybind11 forum) I have figured it out. The class itself should have a constructor that can copy from some input, here using an iterator:

#include <vector>
#include <assert.h>
#include <iterator>


template <class T> class Matrix3D
{
public:

  std::vector<T>      data;
  std::vector<size_t> shape;
  std::vector<size_t> strides;

  Matrix3D<T>() = default;

  template<class Iterator>
  Matrix3D<T>(const std::vector<size_t> &shape, Iterator first, Iterator last);
};


template <class T>
template<class Iterator>
Matrix3D<T>::Matrix3D(const std::vector<size_t> &shape_, Iterator first, Iterator last)
{
  shape = shape_;

  assert( shape.size() == 3 );

  strides.resize(3);

  strides[0] = shape[2]*shape[1];
  strides[1] = shape[2];
  strides[2] = 1;

  int size = shape[0] * shape[1] * shape[2];

  assert( last-first == size );

  data.resize(size);

  std::copy(first, last, data.begin());
}

To directly wrap a function of the following signature:

Matrix3D<double> func ( const Matrix3D<double>& );

the following wrapper code is needed

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

namespace py = pybind11;

namespace pybind11 { namespace detail {
  template <typename T> struct type_caster<Matrix3D<T>>
  {
    public:

      PYBIND11_TYPE_CASTER(Matrix3D<T>, _("Matrix3D<T>"));

      // Conversion part 1 (Python -> C++)
      bool load(py::handle src, bool convert) 
      {
        if ( !convert and !py::array_t<T>::check_(src) )
          return false;

        auto buf = py::array_t<T, py::array::c_style | py::array::forcecast>::ensure(src);
        if ( !buf )
          return false;

        auto dims = buf.ndim();
        if ( dims != 3  )
          return false;

        std::vector<size_t> shape(3);

        for ( int i = 0 ; i < 3 ; ++i )
          shape[i] = buf.shape()[i];

        value = Matrix3D<T>(shape, buf.data(), buf.data()+buf.size());

        return true;
      }

      //Conversion part 2 (C++ -> Python)
      static py::handle cast(const Matrix3D<T>& src, py::return_value_policy policy, py::handle parent) 
      {

        std::vector<size_t> shape  (3);
        std::vector<size_t> strides(3);

        for ( int i = 0 ; i < 3 ; ++i ) {
          shape  [i] = src.shape  [i];
          strides[i] = src.strides[i]*sizeof(T);
        }

        py::array a(std::move(shape), std::move(strides), src.data.data() );

        return a.release();

      }
  };
}} // namespace pybind11::detail

PYBIND11_PLUGIN(example) {
    py::module m("example", "Module description");
    m.def("func", &func, "Function description" );
    return m.ptr();
}

Note that function overloading is now also possible. For example if an overloaded function would exist with the following signature:

Matrix3D<int   > func ( const Matrix3D<int   >& );
Matrix3D<double> func ( const Matrix3D<double>& );

The following wrapper function definition would be needed:

m.def("func", py::overload_cast<Matrix3D<int   >&>(&func), "Function description" );
m.def("func", py::overload_cast<Matrix3D<double>&>(&func), "Function description" );

Answer 2

I am not familiar with pybind11 but became interested after reading this question. From the documentanion it looks like you will have to write your own type caster. This apparently is a rather advanced topic but seems doable with some effort.

Stripped from the documentation, this is the shell of such a converter for convertiong the C++ type inty:

namespace pybind11 { namespace detail {
    template <> struct type_caster<inty> {
    public:
        PYBIND11_TYPE_CASTER(inty, _("inty"));    

        // Conversion part 1 (Python->C++)
        bool load(handle src, bool);

        //Conversion part 2 (C++ -> Python)
        static handle cast(inty src, return_value_policy, handle);
    };
}} // namespace pybind11::detail

It seems all you have to do is to replace inty with Matrix3D<double> and implement load() and cast().

Let's see how they did it for Eigen (eigen.h, line 236 onward):

bool load(handle src, bool) {
    auto buf = array_t<Scalar>::ensure(src);
    if (!buf)
        return false;

    auto dims = buf.ndim();
    if (dims < 1 || dims > 2)
        return false;

    auto fits = props::conformable(buf);
    if (!fits)
        return false; // Non-comformable vector/matrix types

    value = Eigen::Map<const Type, 0, EigenDStride>(buf.data(), fits.rows, fits.cols, fits.stride);

    return true;
}

This does not look too difficult. First they make sure the input is of type array_t<Scalar> (probably array_t<double> in your case). Then they check the dimensions and some conformatibility (you can probably skip the latter). And finally create the Eigen matrix. Since copying is not a problem, at this point simply create a new Martix3D<double> instance and fill it with the data from the numpy array.

There are different implementations of the cast() function for different cases of l-value and const-ness. I guess it is sufficient to do only one implementation that creates a copy of the data in a new numpy array, if that is fine. See function eigen_array_cast() how to return an array as handle return type.

I have not tested any of this and there may be more to the process than it seems. Hopefully this will serve as a starting point.