glmnet-rs

A port of glmnet's elastic-net coordinate descent to Rust, with a Python front end.

Ported from glmnetpp (the C++17 core of R glmnet >= 4.1), not the legacy Fortran, and validated against R glmnet 5.0.

Status: Gaussian, two-class binomial (logistic), and Poisson (dense X), plus sparse X for Gaussian and binomial (CSC). Cross-validation, summaries, and R-style plots are implemented. ~90 parity fixtures pass at ~1e-14 relative error with iteration counts (npasses) identical to R — the sparse ones against R's own sparse path — and coefficients match R to ~1e-13 on real long/wide datasets (see datasets/). What's still missing (multinomial, Cox, relax, offsets, sparse Poisson, …) is tracked in docs/ROADMAP.md; porting notes are in docs/PORTING.md.

Example: examples/glmnet_demo.ipynb is a runnable tour on two real datasets — Wine Quality (long, gaussian) and Leukemia gene expression (wide, binomial) — covering the path, cross-validation, plots, and prediction.

Layout

crates/glmnet-core/   pure Rust kernels (no Python, no C)
crates/glmnet-py/     PyO3 bindings, deliberately thin
python/glmnetrs/      the user-facing package
scripts/gen_fixtures.R   generates the R reference fixtures
tests/fixtures/       committed R glmnet output (tests run without R)

Two APIs, one solver

Faithful to R — the lambda path is the primitive, because it is what the algorithm actually computes:

from glmnetrs import glmnet

path = glmnet(X, y, alpha=1.0)     # alpha = elastic-net mixing (1 = lasso)
path.lambda_                       # (lmu,) descending
path.beta                          # (p, lmu)
path.coef(s=0.05)                  # interpolated, as in R's coef(fit, s=)
path.predict(X, s=0.05)
path.df                            # nonzeros per lambda

# logistic regression, same path object
lpath = glmnet(X, y01, family="binomial")
lpath.predict(X, s=0.05, type="response")   # class-1 probability

# poisson counts
ppath = glmnet(X, counts, family="poisson")
ppath.predict(X, s=0.05, type="response")   # expected count, exp(eta)

# sparse X (gaussian): pass a scipy sparse matrix, never densified
import scipy.sparse as sp
spath = glmnet(sp.csc_matrix(X), y)          # ~20x faster when p >> n and sparse

# cross-validation to pick lambda (matches R's cv.glmnet)
from glmnetrs import cv_glmnet
cv = cv_glmnet(X, y, family="gaussian", type_measure="mse", nfolds=10)
cv.lambda_min, cv.lambda_1se
cv.predict(X, s="lambda.1se")

# summaries, like R's print()
print(path)        # Df / %Dev / Lambda table
print(cv)          # lambda.min / lambda.1se with measure and SE
path.to_frame()    # optional pandas DataFrame

# plots, like R's plot.glmnet / plot.cv.glmnet (needs matplotlib)
path.plot(xvar="lambda")   # coefficient paths, Df on the top axis
cv.plot()                  # CV curve with error bars + min/1se lines

scikit-learn compatible, using scikit-learn's meaning of alpha:

from glmnetrs.sklearn import ElasticNet, Lasso, LogisticRegression

m = ElasticNet(alpha=0.1, l1_ratio=0.7).fit(X, y)   # alpha = penalty strength
m.coef_, m.intercept_

clf = LogisticRegression(C=1.0, penalty="l2").fit(X, y01)
clf.predict_proba(X)

The alpha trap. In glmnet alpha is the mixing parameter and lambda is the penalty strength. In scikit-learn alpha is the penalty strength and l1_ratio is the mixing. Worse, the two objectives are not related by a simple rename: glmnet rescales y to unit variance, which leaves the L2 term carrying a factor of 1/sd(y). glmnetrs.sklearn handles the conversion; the derivation is in docs/PORTING.md.

Develop

cargo test -p glmnet-core --release        # parity against committed fixtures
maturin develop --release --uv             # build the extension
python -m pytest tests/test_python.py      # end-to-end + sklearn agreement

Rscript scripts/gen_fixtures.R             # regenerate Gaussian fixtures (needs R + glmnet)
Rscript scripts/gen_fixtures_binomial.R    # regenerate binomial fixtures
Rscript scripts/gen_fixtures_poisson.R     # regenerate Poisson fixtures
Rscript scripts/gen_fixtures_sparse.R      # regenerate sparse Gaussian fixtures (needs Matrix)
Rscript scripts/gen_fixtures_sparse_glm.R  # regenerate sparse binomial/poisson fixtures

python scripts/bench.py                    # wall-clock vs R glmnet on identical data
python scripts/compare_datasets.py         # coefficients + timing vs R on real datasets
cargo run --release -p glmnet-core --example bench_core   # pure-core timings

Performance

Full-path wall clock vs R glmnet on identical data (Apple Silicon): Gaussian runs at ~0.6–0.85x of R, two-class logistic at ~0.7–1.1x (faster than R when n >> p). glmnet's compiled core is heavily tuned Eigen/SIMD, so parity-to-1.5x-slower is the expected range for a pure-Rust port. Inner products use four-accumulator reductions (matrix::dot4) that vectorize; see docs/PORTING.md.

License

GPL-2.0-only, matching upstream glmnet.