glmnetpp

Native gaussian elnet (glmnet C++ core) binding for Python.

This package wraps the header-only C++ core of R's glmnet (glmnetpp, the

=4.0 reimplementation) behind a small extern "C" ABI and a ctypes binding. It fits ridge / elastic-net gaussian models that are numerically identical to R glmnet (including the intercept=FALSE refit path) without depending on the R runtime.

Installation

pip install glmnetpp

Heads-up — compiled extension. The numerical work is done by a native C++ extension (glmnetpp/_core.*), built from the vendored glmnetpp C++ core + Eigen (header-only). When a prebuilt wheel is available for your platform/Python, pip install glmnetpp Just Works. When it is not, pip builds from the sdist, which requires a C++17 compiler and the Eigen headers on your machine (see Building from source below). To ship wheels for every platform, build them in CI — see the Publishing wheels note at the end.

Why

R's glmnet(family="gaussian", alpha=0, ...) is the reference. The old glmnet_python wrapper ships a 2013 Fortran GLMnet.f whose intercept=FALSE path diverges from R's current (C++) glmnet by ~0.2–0.3%. The C++ core (glmnetpp) is the same code R 4.1-8 runs, so it matches exactly.

Use

from glmnetpp import fit_gaussian_ridge

intercept, coef, jerr = fit_gaussian_ridge(
    X, y, lam=0.1, alpha=0.0,
    lower_limits=[0.0]*p, upper_limits=[None]*p,
    standardize=True, intercept=True,
)

coef is on the original (unstandardized) scale, matching as.numeric(coef(glmnet(...))).

Building from source

This is a Poetry package. poetry-core owns packaging/metadata/dependencies; it does not compile C extensions itself — instead it runs scripts/build-extension.py, which drives setuptools' build_ext (Eigen discovery + compiler-conditional flags) and builds the extension in-place under src/glmnetpp/.

Build-only dependency: Eigen (header-only).

# either (with Poetry installed):
poetry install
# or (plain pip — uses the poetry-core backend, runs the build script):
pip install -e .

The build script searches for Eigen in this order and picks the first one that contains Eigen/Core:

1. third_party_eigen/eigen-<version>/ inside the project (what CI uses — drop the header-only Eigen release there and it builds anywhere, no system package needed)
2. $EIGEN_INCLUDE_DIR
3. /usr/include/eigen3 (libeigen3-dev)
4. /usr/lib/R/site-library/RcppEigen/include (r-cran-rcppeigen)
5. macOS Homebrew (/usr/local/include/eigen3, /opt/homebrew/include/eigen3)
6. Windows — vcpkg ($VCPKG_ROOT/installed/x64-windows/include/eigen3) or Conda ($CONDA_PREFIX/Library/include/eigen3)

Provenance

The vendored-from-R headers (the C++ glmnetpp core, copied from CRAN glmnet) and the full list of every part copied from R are documented in cpp/glmnetpp_include/README.md in the source tree.

Publishing wheels

Prebuilt wheels are produced by the GitHub Actions workflow in .github/workflows/python-publish.yml. It uses cibuildwheel to build manylinux (Linux), Windows and macOS (x86_64 + arm64) wheels for CPython 3.9–3.13, plus an sdist, and publishes them all to PyPI.

Because the C++ extension needs Eigen at build time, the workflow downloads the header-only Eigen release into third_party_eigen/ inside the project tree before building; scripts/build-extension.py finds it there relative to the repo root, so the same path works on every OS and inside cibuildwheel's manylinux container. (That directory is gitignored — it's a build input, not source.) Each wheel is smoke-tested (import + fit_gaussian_ridge) in CI before it ships, so end users of these platforms never need a C++ compiler or Eigen.

License

GPL-2.0-or-later, inherited from the upstream glmnet / glmnetpp C++ core (these headers are a derivative of that code). See the LICENSE file.