Pybind11 如何从 python -> c++ 传递 n 维 numpy 数组

Question

I have a numpy array of numpy arrays of floats that I wish to pass to a c++ function that will read and modify the data as if it were a std::vector. I am struggling to figure out how to do this. What would the c++ argument type be for:

np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]], dtype=float)

This type declaration: py::array_t<py::array_t<float>>

produces this error: Attempt to use a non-POD or unimplemented POD type as a numpy dtype

and I can't see to get this type declaration to work: py:array

I have searched through documentation, github issues and examples I can find online but have not found anything that works or that I can make sense of, if anyone could help me with this, it would be much appreciated :)

Answer 1

Pull in the Eigen library on the C++ side and use that for your matrix operations. std::vector is not a good abstraction for 2d matrices... either you have to write your own 2d indexing into a 1d vector, or you need something silly like std::vector<std::vector<>> , which performs very badly and is ugly besides.

pybind11 knows the type conversions to automatically convert np.array to Eigen::MatrixXf , so you should be able to write your C++ function as taking a Eigen::MatrixXf &input , and call it as you've specified.