numpy array of arrays?

[cary-ilm]

With numpy, I can create a 2D array of arrays, where each entry has a unique length:

array_of_arrays = np.empty((3, 3), dtype=object)
for i in range(3):
    for j in range(3):
        array_of_arrays[i, j] = np.array([k for k in range(i*3+j)])

Is it possible to create such a thing in pybind11? When I try to create py::array_t<py::array_t<int>>, I get:

static assertion failed: Attempt to use a non-POD or unimplemented POD type as a numpy dtype

which makes it appear that py::array can't handle non-POD objects like py::array as a type, unless there's a way to specify it that I'm missing. The documentation also appears to imply this, but I'd like to confirm I'm understanding the limitation before resorting to an alternative design.

The specific application is deep image data: a 2D array of pixels, where each pixel is an array of samples, in case there's a better way of achieving that. A numpy array of lists would suffice, but presumably that has the same problem. The alternative of a list of lists of lists/arrays is less desirable.

[dkoo2]

What worked for me is a signature like int f(::pybind11::array my_array), where my_array can be an np.array of np.arrays where the inner arrays are different length.

Then you can access the inner arrays like

::pybind11::buffer_info buffer_info = my_array.request();
auto *ptr = static_cast<::pybind11::array*>(buffer_info.ptr);

If you just need a 2D np array then your type should be py_array_t<int> instead