nanoeigenpy

This is a collection of tools for using Eigen together with nanobind, as a successor to the eigenpy support library. Its aim is to help the transition away from Boost.Python.

It reintroduces a few features (e.g. bindings for Eigen matrix decompositions) which are not in nanobind at time of writing.

Rationale

Eigenpy was based on Boost.Python, an aging, complex, heavily templated library with little community support. Support for many library features initially present in Boost, and which were added to the STL since C++11/14/17, for over a decade (nearly two), were just never added in Boost.Python. This includes support for {boost,std}::optional, {boost,std}::variant, {boost,std}::unique_ptr, proper support for map types… whereas they have been present in pybind11, and now nanobind, for years.

These features were finally added to eigenpy with a lot of developer effort. This created additional need for supporting these additional features ourselves, including many downstream consumers (mainly in the robotics community).

Features

nanoeigenpy provides the following features for helping you bind features from Eigen to Python:

• bindings for Eigen’s Geometry module - quaternions, angle-axis representations…
• bindings for Eigen’s matrix dense and sparse decompositions and solvers

Optional features

nanoeigenpy also provides bindings for Eigen’s Cholmod and Apple Accelerate modules.

[!NOTE] The Accelerate module is only available on Eigen’s master branch, and not in any official release yet (as of 3/28/2025).

Cholmod is part of the SuiteSparse algorithms library. It can be installed standalone from conda.

Example usage

The features included in nanoeigenpy are distributed in a Python module which can be imported, or through standalone headers which can be included in your own Python bindings code using a CMake target.

Using the nanoeigenpy headers (with CMake)

To directly use the tools in nanoeigenpy’s headers, link to it in CMake (or whichever build tool you have, but only CMake support is planned so far).

# look for the nanoeigenpy CMake package
find_package(nanoeigenpy REQUIRED)

nanobind_add_module(my_ext NB_STATIC my_ext.cpp)
target_link_libraries(my_ext PRIVATE nanoeigenpy::nanoeigenpy_headers)

Then, in your C++ extension module code, include the relevant headers and call functions to expose the required type:

#include <nanoeigenpy/geometry/quaternion.hpp>

namespace nb = nanobind;

void f(const Eigen::Quaterniond &quat) {
    // ...
}

NB_MODULE(my_ext, m) {
    nanoeigenpy::exposeQuaternion<double>(m, "Quaternion");
    m.def("f", f, nb::arg("quat"));
}

Using the compiled Python module

In the case above, nanoeigenpy’s Python extension module already includes bindings for Eigen::Quaternion with the double scalar type (AKA Eigen::Quaterniond). Then, we can simply get nanobind to import it in our extension module:

#include <Eigen/Geometry>

namespace nb = nanobind;

void f(const Eigen::Quaterniond &quat) {
    // ...
}

NB_MODULE(my_ext, m) {
    // import nanoeigenpy's module **here**
    nb::module_::import_("nanoeigenpy");
    m.def("f", f, nb::arg("quat"));
}

Alternatively, Python code which uses our extension my_ext can also bring in nanoeigenpy:

import nanoeigenpy
from nanoeigenpy import Quaternion
from my_ext import f

quat = Quaternion(0., 1., 0., 0.)
f(quat)

[!NOTE] If you have a specific scalar type (e.g. float16) with which you want to use Eigen::Quaternion, or matrix solvers, or other features in nanoeigenpy, you should refer to the first approach and use nanoeigenpy from C++ directly.

File truncated at 100 lines see the full file