A portable PLS / NIRS engine in C++17 with a stable C ABI (libn4m) and thin first-class bindings for Python, R, MATLAB / Octave, plus JS / WebAssembly, Julia, JNI / Android. PoC bindings for Go, Rust, .NET, Ruby, Lua, Nim live under bindings/_archive/ and can be revived on demand.

The same numerical core powers every binding: a model trained in Python, R, MATLAB, a browser, or Android is checked against the same C++ parity contract and tolerance policy. Cross-checked against scikit-learn, R pls, ropls, mixOmics, ikpls, Octave plsregress.

Part of the open-source NIRS tools ecosystem: file readers, datasets, methods, browser modelling, reproducible pipelines, papers, benchmarks, and release dashboards for near-infrared spectroscopy.

Status: post-merge refactor in progress. See docs/REFACTOR_PLAN.md for the founding document and roadmap/ for phase-by-phase tracking. pls4all is now a packaging subset name (slim PLS-only re-export of n4m); the underlying binary is always libn4m.

Pulled from the cross-binding benchmark matrix (generated from the canonical benchmarks.parity_timing.registry method catalog; committed timing snapshots are regenerated, not hand maintained).

PLS external baselines in the same snapshot: ikpls 1.08 ms, R pls 5.22 ms, R mixOmics 7.77 ms, Octave plsregress 2.18 ms. Detailed methodology, parity gates and per-algorithm tables are in docs/benchmarks/.

git clone https://github.com/GBeurier/nirs4all-methods.git && cd nirs4all-methods
cmake --preset blas-omp && cmake --build --preset blas-omp -j
pip install -e bindings/python

import numpy as np
from n4m.estimators.regression.latent import PLS
from pls4all.sklearn import PLSRegression

X = np.random.randn(500, 200)
y = 2 * X[:, 3] - X[:, 6] + 0.5 * np.random.randn(500)

full = PLS(n_components=5).fit(X, y)              # nirs4all-methods -> import n4m
slim = PLSRegression(n_components=5).fit(X, y)    # slim PLS-only pls4all package
print(full.score(X, y), slim.score(X, y))         # sklearn-style estimators

Install the released binaries from R-universe (no C++/R toolchain to set up):

install.packages("n4m",     repos = c("https://gbeurier.r-universe.dev", getOption("repos")))
install.packages("pls4all", repos = c("https://gbeurier.r-universe.dev", getOption("repos")))  # slim PLS-only subset

Or build from this checkout:

N4M_INCLUDE_DIR=$PWD/cpp/include \
  N4M_LIB_DIR=$PWD/build/blas-omp/cpp/src \
  R CMD INSTALL bindings/r/n4m

library(n4m)

# Base R formula+S3
fit <- pls(y ~ ., data = df, ncomp = 5)
predict(fit, newdata = df)

# pls package style
fit <- plsr(y ~ ., data = df, ncomp = 10, method = "simpls", validation = "CV")
RMSEP(fit)
selectNcomp(fit)

# mdatools style
fit <- pls(X, y, ncomp = 10, center = TRUE, scale = FALSE, cv = 10)
predict(fit, x = Xnew)

addpath(genpath('bindings/matlab'))
mdl = n4m.fit("sparse_simpls", X, y, "NumComponents", 5, "Lambda", 0.05);
predict(mdl, Xnew)
% 18 classdefs available: Regression, SparsePlsRegression, EcrRegression,
% MbPlsRegression, GlmRegression, MirRegression, …
% (CI runs Octave; MATLAB-only divergences declared in bindings/matlab/COMPAT.md)

Cross-binding parity: every binding's SIMPLS predictions match the shared fixture within rmse_rel < 1e-12 for the committed smoke matrix. See docs/parity/ and the cross-binding benchmark.

This table condenses the cross-binding benchmark matrix into a "where do I find this method?" view. Python, R and MATLAB list external reference libraries observed in the benchmark; nirs4all is the sanctioned in-tree Python reference provider for methods that do not have a packaged external implementation.

• 🌐 Sphinx site : (configurer GitHub Pages — voir .github/workflows/docs.yml)
• 📊 Cross-binding benchmarks — parity + timing matrix generated from the canonical method registry
• 🧪 Cross-binding runbook — registry refs, clean sweeps, git-pinned refs, and legacy fixed-reference modes
• 🔬 Methodology — reference policy, tolerances, threading, hardware
• 🏗️ Architecture — memory model · error model · threading · serialization
• 📜 ABI reference — n4m_* C ABI surface, stability policy, changes log
• 🔌 Bindings — Python · R · MATLAB · JS / WASM · Android
• ✅ Parity methodology — every algorithm cross-checked against an external library
• 🛠️ Dev workflow · build · testing
• 📦 Releases & downloads — every version's Python wheels, R/CRAN source tarballs, source archive + SBOM, all attached to one GitHub Release
• 🚀 Release & registration process — per-binding publish, the exact files to upload, and CRAN / R-universe registration

• Methods (73) — PLS regression family (NIPALS, SIMPLS, kernel, wide-kernel, SVD, power, randomized SVD, PCR), OPLS / OPLS-DA, PLSCanonical, PLSSVD, sparse SIMPLS, CPPLS, ECR, MIR-PLS, ridge PLS, robust PLS (Huber IRLS), continuum regression, MB-PLS, LW-PLS, N-PLS, O2-PLS, missing-aware NIPALS, bagging / boosting / random-subspace ensembles, GPR-on-PLS, PLS-GLM (Gaussian / Poisson), PLS-DA / LDA / QDA / logistic / Cox, PDS / DS calibration transfer, AOM-PLS & POP-PLS (the scientific differentiator).
• Variable selection (24) — VIP / coefficient / SR rankers, SPA, CARS, GA-PLS, PSO, VISSA, IRIV, IRF, shaving, BVE, REP, IPW, ST-PLS, T2, WVC, EMCUVE, randomization, biPLS, siPLS, interval, stability, UVE, random frog, SCARS, VIP-SPA.
• Diagnostics — Hotelling T² · Q residuals · DModX · process monitoring with alarms · approximate PRESS · one-SE rule.
• Preprocessing pipeline — identity · center · autoscale · Pareto · SNV · MSC · EMSC · detrend · SG · ASLS · Norris-Williams · Haar wavelet · OSC · EPO.
• Validation — k-fold · LOO · holdout · external-fold · repeated k-fold · Monte-Carlo · Kennard-Stone · SPXY.
• Metrics — regression (RMSE/MAE/bias/R²/Q²/slope/intercept/RPD/RPIQ) · binary classification (sensitivity/specificity/balanced accuracy/precision/F1/MCC/AUC) · multiclass (macro/micro AUC) · calibration.

Full list at Overview.md ; canonical spec at Direction_Technique.md.

git clone https://github.com/GBeurier/nirs4all-methods.git && cd nirs4all-methods

# Reference build (single-thread, no BLAS — useful for parity)
cmake --preset dev-release
cmake --build --preset dev-release -j
ctest --preset dev-release --output-on-failure

# Production build (BLAS + OpenMP)
cmake --preset blas-omp
cmake --build --preset blas-omp -j

./build/blas-omp/cpp/cli/n4m_cli --version
./build/blas-omp/cpp/cli/n4m_cli --abi-info

Zero mandatory dependencies — no Eigen, no Boost, no BLAS, no pybind11, no Rcpp on the install path. Accelerated backends (BLAS / OpenMP / CUDA) plug in via CMake presets without changing the ABI.

# Install git-only external refs not released on PyPI yet.
python -m pip install -r benchmarks/cross_binding/requirements-git.txt

# Complete overnight canonical method/reference matrix.
# Default: registry cells + registry-declared external references only.
# Skips cells already present in results/full_matrix.csv.
# Env overrides: FORCE=1 RERUN_FAILED=1 THREADS="1 3 10" N_RUNS=40.
# N_RUNS is the adaptive total-run cap, including the warmstart.
benchmarks/cross_binding/run_overnight.sh

# Exhaustive stress matrix: all n4m bindings/tiers, all CPU libn4m
# backend variants, the default 11-size dataset sweep, threads 1/3/10,
# and the registry-declared external references.
FULL_MATRIX=1 benchmarks/cross_binding/run_overnight.sh

# Legacy fixed-reference cross-product too. This is useful for auditing
# external-library coverage, and may emit NOT_IMPLEMENTED for algorithms
# that sklearn / ikpls / ropls / mixOmics / plsregress do not support.
FULL_MATRIX=1 REFERENCE_BACKENDS=all benchmarks/cross_binding/run_overnight.sh

# Include the CUDA build too when CUDA is available.
FULL_MATRIX=1 LIBP4A_BUILD=all benchmarks/cross_binding/run_overnight.sh

# On the Pages branch (main), commit/push web data and trigger GitHub Pages.
PUBLISH_WEB=1 benchmarks/cross_binding/run_overnight.sh

# Exhaustive run, then publish the refreshed dashboard from main.
FULL_MATRIX=1 PUBLISH_WEB=1 benchmarks/cross_binding/run_overnight.sh

# From a work branch, publish the source files without deploying live Pages.
PUBLISH_WEB=1 DEPLOY_PAGES=0 benchmarks/cross_binding/run_overnight.sh

# Render an existing CSV only
python benchmarks/cross_binding/combine_and_render.py \
  --csvs benchmarks/cross_binding/results/full_matrix.csv \
  --out docs/benchmarks/cross_binding.md

Full re-run commands in docs/benchmarks/methodology.md.

nirs4all-methods is dual-licensed open-source — CeCILL-2.1 OR AGPL-3.0-or-later (your choice; CeCILL-2.1 is recognised by French law and GPL-family-compatible) — with an optional commercial license for closed-source / SaaS use. For any commercial use, contact nirs4all-admin@cirad.fr. See LICENSING.md, the texts under LICENSES/, and THIRD_PARTY_NOTICES.md.

@software{nirs4all_methods,
  author  = {Beurier, Grégory and contributors},
  title   = {nirs4all-methods: A portable Partial Least Squares engine with a stable C ABI},
  year    = {2026},
  url     = {https://github.com/GBeurier/nirs4all-methods},
  version = {1.0.1}
}

A machine-readable CITATION.cff is provided.

n4m keeps the detailed historical phase log in ROADMAP.md. The active roadmap is now focused on making the same C ABI catalog available through more language bindings.

Users can request a missing method through GitHub Issues. Please include the reference implementation or library, the paper / DOI, expected inputs and outputs, and any small reproducible dataset that can become a parity fixture.

n4m builds on lessons learned from the experimental aompls library and on the rich PLS ecosystem in scikit-learn, R pls / ropls / mixOmics / plsVarSel, and nirs4all.