pls4all

pls4all is the slim, PLS-only subset of nirs4all-methods — a thin Python binding over the portable libn4m C ABI (a C++17 PLS / NIRS engine). The wheel bundles the libn4m shared library, so pip install pls4all is self-contained; no separate native build is required. For the full method surface (preprocessing, selectors, diagnostics, augmenters, …) install the nirs4all-methods package instead and import it as n4m — both load the same libn4m.

The binding loads libn4m with ctypes.CDLL (so the GIL is released during native calls) and exposes:

• version() / abi_version() introspection,
• a Pythonic Context and Config (RAII lifecycle wrappers),
• the PLS fit/predict surface and a scikit-learn-compatible pls4all.sklearn.PLSRegression (and the other PLS-family estimators),
• a typed Pls4allError raised on any non-OK status, carrying the context's last_error message.

Quick start

import numpy as np
import pls4all
from pls4all.sklearn import PLSRegression

print(pls4all.version())      # e.g. "1.3.6+abi.2.19.0"
print(pls4all.abi_version())  # (2, 19, 0)

rng = np.random.default_rng(0)
X = rng.standard_normal((40, 12))
y = X @ rng.standard_normal(12)

model = PLSRegression(n_components=5).fit(X, y)
print(model.predict(X).shape)  # (40,)

Low-level lifecycle, if you need it:

with pls4all.Context() as ctx, pls4all.Config() as cfg:
    cfg.algorithm = pls4all.Algorithm.PLS_REGRESSION
    cfg.solver = pls4all.Solver.SIMPLS
    cfg.n_components = 5
    # ... drive a fit through the C ABI ...

scikit-learn is an optional dependency (only pls4all.sklearn needs it); the core import pls4all works with NumPy alone.

Verified affine-model migration

pls4all.export_linear_predictor_n4mm(...) is the public, PREDICT-only migration boundary for a converter that has already verified a source model's affine equation. It accepts finite, row-major coefficients with shape (n_features, n_targets) and a target-length intercept, then returns a raw N4MM payload. It neither reads pickle/joblib data nor retrains a model.

from pls4all import export_linear_predictor_n4mm

payload = export_linear_predictor_n4mm(
    coefficients=[[0.5], [-1.0]],
    intercept=[0.25],
    source_training_samples=120,
)
assert payload[:4] == b"N4MM"

The resulting imported-linear-predictor model supports native prediction only; it deliberately does not claim to reconstruct a source PLS latent transform. Callers remain responsible for attesting that the source model is exactly intercept + X @ coefficients before conversion.

Pure-native estimator selection in the full package

The full nirs4all-methods distribution (imported as n4m) exposes the native single-level finetuning driver. It is not part of the slim pls4all namespace. Build the validation plan and search space through public owning objects:

import numpy as np

from n4m.model_selection import (
    Algorithm,
    Metric,
    Sampler,
    SearchSpace,
    ValidationPlan,
    finetune_estimator,
)

rng = np.random.default_rng(42)
X = rng.normal(size=(60, 12))
y = X @ rng.normal(size=12) + rng.normal(scale=0.02, size=60)
fold_ids = np.arange(X.shape[0], dtype=np.int32) % 5

with ValidationPlan.from_fold_ids(fold_ids) as plan:
    with SearchSpace() as space:
        space.add_int("n_components", 1, 10)
        result = finetune_estimator(
            Algorithm.PLS_REGRESSION,
            X,
            y,
            plan,
            space,
            n_trials=30,
            sampler=Sampler.TPE,
            metric=Metric.RMSE,
            seed=42,
            timeout_seconds=20.0,
        )

print(result.best_params)       # for example: {"n_components": 6}
print(result.best_score)
print(result.timed_out)         # True only for a successful partial timeout
print(len(result.trials))       # owning completed/failed trial snapshots

The eligible algorithms are PLS_REGRESSION, PLS_CANONICAL, PLS_SVD, OPLS, SPARSE_PLS and PCR. The five non-sparse routes require the exact keyword n_components, a non-log integer axis. SPARSE_PLS accepts any non-empty subset of n_components and sparsity_lambda; omitted axes retain the native defaults 2 and 0.0. sparsity_lambda must stay in [0, 1) and may be declared with space.add_float(..., log=True) when low > 0.

This call selects a configuration and returns an owning trace; it does not fit a final estimator on all rows. A timeout before any completion raises N4MError with CANCELLED. A later timeout returns the best partial result with result.timed_out is True and fewer trials than result.requested_trials. See docs/methods/optimization.md for the exact schema, status and validation-plan contracts.

Relation to nirs4all conformal learning and robustness

The Python bindings are thin translation layers over the libn4m C ABI. They expose native kernels, scikit-learn-compatible PLS estimators and the single-estimator finetune_estimator(...) selection trace, but they do not own the higher-level nirs4all statistical lifecycle.

Use nirs4all.run(tuning=...) when the workflow needs final winner projection, workspace persistence, .n4a packaging, optional nirs4all.calibrate() / predict_calibrated() semantics, CalibratedRunResult, or nirs4all.robustness() / RobustnessReport artifacts. Those guarantees and invalidation reasons are produced by nirs4all, not by n4m or pls4all.

Binding consumers must not reimplement tuning/conformal/robustness guarantees from raw native predictions. A UI, Studio panel or alternate language binding that wants to display conformal intervals or robustness summaries should consume the public nirs4all artifact/schema surfaces instead of inferring guarantees locally from finetune_estimator(...) or PLSRegression.predict(...).

Loading libn4m

The bundled wheel ships libn4m inside pls4all/lib/, found automatically. For development against a local build the loader searches, in order:

1. $PLS4ALL_LIB_PATH — explicit path to libn4m for this package,
2. $N4M_LIB_PATH — shared libn4m override honoured by both pls4all and n4m,
3. pls4all/lib/libn4m* next to the installed package (wheel layout),
4. <repo-root>/build/dev-release/cpp/src/libn4m* (developer convenience),
5. the standard system search path (LD_LIBRARY_PATH, macOS rpath, Windows PATH).

Building libn4m from source (developers)

cmake --preset dev-release
cmake --build --preset dev-release --parallel

This produces build/dev-release/cpp/src/libn4m.so (.dylib / .dll on macOS / Windows), which the loader rules above pick up.

See https://github.com/GBeurier/nirs4all-methods for the full project.