Models

This section covers model training, comparison, and optimization in NIRS4ALL.

Overview

NIRS4ALL supports a wide range of machine learning models from different frameworks. This section covers model training workflows.

🎯 Model Training

Train models with cross-validation and evaluate performance.

Essential

Model Training

⚙️ Hyperparameter Tuning

Optimize models with Optuna integration and grid search.

Optimization

Hyperparameter Tuning

Native Tuning + Conformal

Use the native run(tuning=...) subset, persist tuning/conformal evidence, and apply calibrated intervals. Includes the smoke-tested tune→calibrate→predict→robustness example. Includes a native PLS example that tunes model.n_components.

Experimental

Native tuning and conformal calibration

🧠 Deep Learning

Use TensorFlow, PyTorch, and JAX models in pipelines.

Neural Networks

Deep Learning Models

AOM Models

Use AOM-PLS, AOM-Ridge, AutoSelector, Blender, and FastAOM with pipeline folds.

Spectroscopy

AOM Models

🤖 TabPFN for NIRS

Fixed-recipe TabPFN regressor that matches per-dataset HPO without it.

No HPO

TabPFN for NIRS — fixed recipe (no HPO)

Supported Model Frameworks

Framework	Import	Examples
scikit-learn	Direct	PLSRegression, RandomForest, SVM
TensorFlow/Keras	`nirs4all[tensorflow]`	Sequential, Custom architectures
PyTorch	`nirs4all[torch]`	nn.Module subclasses
JAX	`nirs4all[jax]`	Flax/Haiku models

Quick Example

from sklearn.cross_decomposition import PLSRegression
from sklearn.model_selection import KFold

result = nirs4all.run([
    MinMaxScaler(),
    KFold(n_splits=5),
    {"model": PLSRegression(n_components=10)}
], dataset="data/")
print(f"RMSE: {result.best_rmse:.4f}")

Built-in NIRS Models

NIRS4ALL includes specialized models optimized for spectroscopy:

PLS Variants

Model	Description
AOMPLSRegressor	Adaptive Operator-Mixture PLS — auto-selects best preprocessing from operator bank
AOMPLSClassifier	AOM-PLS for classification with probability calibration
AOMRidgeRegressor	AOM-Ridge with operator-bank selection and internal CV
AOMRidgeAutoSelector	Selects the best AOM-Ridge variant under outer CV
AOMRidgeBlender	Convex blend of AOM-Ridge variants; strongest general AOM-Ridge recipe
FastAOMPLSRidge	Fast screened operator-chain AOM PLS/Ridge
POPPLSRegressor	Per-Operator-Per-component PLS — different operator per component via PRESS
POPPLSClassifier	POP-PLS for classification with probability calibration
PLSDA	PLS Discriminant Analysis
OPLS / OPLSDA	Orthogonal PLS / OPLS-DA
MBPLS	Multi-Block PLS
DiPLS	Domain-Invariant PLS
IKPLS	Improved Kernel PLS
FCKPLS	Fractional Convolution Kernel PLS
LWPLS	Locally Weighted PLS
SparsePLS	Sparse PLS
SIMPLS	SIMPLS algorithm

from nirs4all.operators.models import AOMPLSRegressor, AOMRidgeBlender, POPPLSRegressor

# Native n4m AOM-PLS: select preprocessing and component count by CV
pipeline = [
    KFold(n_splits=5),
    {"model": AOMPLSRegressor(max_components=10, cv=5)}
]

# AOM-Ridge Blender: strong split-aware operator-mixture Ridge recipe
pipeline = [
    KFold(n_splits=5),
    {
        "model": AOMRidgeBlender(outer_cv=5, inner_cv=5),
        "train_params": {"use_pipeline_folds_for_aom": "required"},
    },
]

# Native n4m POP-PLS: selects a different operator per component
pipeline = [
    KFold(n_splits=5),
    {"model": POPPLSRegressor(max_components=15, cv=5)}
]

Deep Learning

nicon - 1D CNN for NIR classification/regression
decon - 1D CNN for NIR regression

from nirs4all.operators.models import nicon, decon

pipeline = [
    MinMaxScaler(),
    KFold(n_splits=5),
    {"model": decon(n_outputs=1)}
]