C++ functional with Eigen pass-by-reference issue

[varunagrawal]

I am trying to create a C++ method that accepts a functional, which will then get passed a python function, e.g.

using CustomErrorFunction = std::function<Eigen::VectorXd(const Eigen::MatrixXd& values, std::vector<Eigen::MatrixXd> *)>;

class Model {
  CustomErrorFunction error_function_;

 public:
  Model(const CustomErrorFunction& func): error_function_(func) {}

  /// This method is not wrapped to python since it is only ever called by C++ classes by design
  Eigen::VectorXd error(const Eigen::MatrixXd& values,  boost::optional<std::vector<Eigen::MatrixXd>&> Jacobians) {
    if (Jacobians)
      Vector error = this->error_function_(values, Jacobians.get_ptr());
    else
      Vector error = this->error_function_(values, nullptr);

    return error;
  }
};

The idea behind the functional is to compute an error and return the jacobians as an output argument, with the error and jacobians being computed on the Python side of things:

def custom_error(x: np.ndarray, J: List[np.ndarray]):
  if J is not None:
    # some random jacobian(s) for illustration
    J[0] = np.eye(3)
    J[1] = -np.eye(3)

  return np.linalg.norm(x)  # return the L2 norm

model = Model(custom_error)
e = model.error(np.arange(3))

Then binding is done as follows:

py::class_<Model, boost::shared_ptr<Model>>(m_, "Model")
        .def(py::init<>())
        .def(py::init<const CustomErrorFunction&>(), py::arg("errorFunction"));

Note that we do not wrap Model::error since only the C++ side of things needs to know about it as per the design spec.

The issue we're facing is that the Jacobians vector is not getting populated with the values of J from the Python side of the code. We looked at both the writeable flag and the order (J[0].flags.f_contiguous) & J[1].flags.f_contiguous are both True) as detailed here and everything seems to be in order. Can someone please help us understand what the issue might be?

[varunagrawal]

Quick follow up: This seems to be a known issue #1200
I'll simply follow up there.