nirs4all 0.4.1

NIRS Analyses made easy.

NIRS4ALL is a comprehensive machine learning library specifically designed for Near-Infrared Spectroscopy (NIRS) data analysis. It bridges the gap between spectroscopic data and machine learning by providing a unified framework for data loading, preprocessing, model training, and evaluation.

What is Near-Infrared Spectroscopy (NIRS)?

Near-Infrared Spectroscopy (NIRS) is a rapid and non-destructive analytical technique that uses the near-infrared region of the electromagnetic spectrum (approximately 700-2500 nm). NIRS measures how near-infrared light interacts with the molecular bonds in materials, particularly C-H, N-H, and O-H bonds, providing information about the chemical composition of samples.

Key advantages of NIRS:

• Non-destructive analysis
• Minimal sample preparation
• Rapid results (seconds to minutes)
• Potential for on-line/in-line implementation
• Simultaneous measurement of multiple parameters

Common applications:

• Agriculture: soil analysis, crop quality assessment
• Food industry: quality control, authenticity verification
• Pharmaceutical: raw material verification, process monitoring
• Medical: tissue monitoring, brain imaging
• Environmental: pollutant detection, water quality monitoring

Notes:

NIRS4All is in active development; APIs and docs are subject to change. Pre-1.0 notice: interfaces and documentation may change without notice.

Features

NIRS4ALL offers a wide range of functionalities:

1. Spectrum Preprocessing:

   • Baseline correction
   • Standard normal variate (SNV)
   • Robust normal variate
   • Savitzky-Golay filtering
   • Normalization
   • Detrending
   • Multiplicative scatter correction
   • Derivative computation
   • Gaussian filtering
   • Haar wavelet transformation
   • And more...

2. Data Splitting Methods:

   • Kennard Stone
   • SPXY
   • Random sampling
   • Stratified sampling
   • K-means
   • And more...

3. Model Integration:

   • Scikit-learn models
   • TensorFlow/Keras models
   • Pre-configured neural networks dedicated to the NIRS: nicon & decon (see publication below)
   • PyTorch models (via extensions)
   • JAX models (via extensions)

4. Model Fine-tuning:

   • Hyperparameter optimization with Optuna
   • Grid search and random search
   • Cross-validation strategies

5. Visualization:

   • Preprocessing effect visualization
   • Model performance visualization
   • Feature importance analysis
   • Classification metrics
   • Residual analysis

Advanced visualization capabilities for model performance analysis

Installation

Basic Installation

pip install nirs4all

Install TensorFlow cpu support by default

With Additional ML Frameworks

# With PyTorch support
pip install nirs4all[torch]

# With Keras support
pip install nirs4all[keras]

# With JAX support
pip install nirs4all[jax]

# With all ML frameworks
pip install nirs4all[all]

Development Installation

For developers who want to contribute:

git clone https://github.com/gbeurier/nirs4all.git
cd nirs4all
pip install -e .[dev]

Installation Testing

After installing nirs4all, you can verify your installation and environment using the built-in CLI test commands:

# Basic installation test: checks required dependencies and versions
nirs4all --test-install

# Integration test: runs sklearn, tensorflow, and optuna pipelines on sample data
nirs4all --test-integration

# Check version
nirs4all --version

Each command will print a summary of the test results and alert you to any missing dependencies or issues with your environment.

Quick Start

Basic Pipeline Example

from sklearn.preprocessing import MinMaxScaler
from sklearn.model_selection import ShuffleSplit
from sklearn.cross_decomposition import PLSRegression
from sklearn.ensemble import RandomForestRegressor

from nirs4all.data import DatasetConfigs
from nirs4all.pipeline import PipelineConfigs, PipelineRunner
from nirs4all.operators.transforms import (
    StandardNormalVariate, SavitzkyGolay, MultiplicativeScatterCorrection
)

# Define your processing pipeline
pipeline = [
    MinMaxScaler(),                    # Scale features
    StandardNormalVariate(),           # Apply SNV transformation
    ShuffleSplit(n_splits=3),         # 3-fold cross-validation
    {"y_processing": MinMaxScaler()}, # Scale target values
    {"model": PLSRegression(n_components=10)},
    {"model": RandomForestRegressor(n_estimators=100)},
]

# Create configurations
pipeline_config = PipelineConfigs(pipeline, name="MyPipeline")
dataset_config = DatasetConfigs("path/to/your/data")

# Run the pipeline
runner = PipelineRunner(save_files=False, verbose=1)
predictions, predictions_per_datasets = runner.run(pipeline_config, dataset_config)

# Analyze results
top_models = predictions.top(n=5, rank_metric='rmse')
print("Top 5 models by RMSE:")
for i, model in enumerate(top_models):
    print(f"{i+1}. {model['model_name']}: RMSE = {model['rmse']:.4f}")

Advanced Pipeline with Feature Augmentation

from nirs4all.operators.transforms import (
    Detrend, FirstDerivative, Gaussian, Haar
)

# Define multiple preprocessing options
preprocessors = [Detrend, FirstDerivative, Gaussian, StandardNormalVariate]

# Advanced pipeline with feature augmentation
pipeline = [
    "chart_2d",  # Generate visualization
    MinMaxScaler(),
    {"y_processing": MinMaxScaler()},
    {
        "feature_augmentation": {
            "_or_": preprocessors,
            "size": [1, (1, 2)],  # Single and paired transformations
            "count": 7           # Generate 7 different combinations
        }
    },
    ShuffleSplit(n_splits=3, test_size=0.25),
]

# Add multiple PLS models with different components
for n_comp in range(5, 31, 5):
    pipeline.append({
        "name": f"PLS-{n_comp}_components",
        "model": PLSRegression(n_components=n_comp)
    })

# Run and analyze
pipeline_config = PipelineConfigs(pipeline, "AdvancedPipeline")
runner = PipelineRunner(save_files=False)
predictions, _ = runner.run(pipeline_config, dataset_config)

Neural Network Integration

from nirs4all.operators.models.tensorflow.nicon import nicon

# Pipeline with pre-configured neural network
pipeline = [
    MinMaxScaler(),
    StandardNormalVariate(),
    ShuffleSplit(n_splits=3),
    {"y_processing": MinMaxScaler()},
    {"model": PLSRegression(n_components=15)},
    {
        "model": nicon,  # Pre-configured convolutional neural network
        "name": "NICON-CNN",
        "train_params": {
            "epochs": 100,
            "patience": 20,
            "verbose": 1
        }
    }
]

pipeline_config = PipelineConfigs(pipeline, "NeuralNetworkPipeline")
runner = PipelineRunner(save_files=False, verbose=1)
predictions, _ = runner.run(pipeline_config, dataset_config)

# Compare neural network with traditional models
top_models = predictions.top(n=3, rank_metric='rmse')
for i, model in enumerate(top_models):
    print(f"{i+1}. {model['model_name']}: RMSE = {model['rmse']:.4f}")

Hyperparameter Optimization

# Pipeline with automated hyperparameter tuning
pipeline = [
    MinMaxScaler(),
    StandardNormalVariate(),
    ShuffleSplit(n_splits=3),
    {"y_processing": MinMaxScaler()},
    {
        "model": PLSRegression(),
        "name": "PLS-Optimized",
        "finetune_params": {
            "n_trials": 50,
            "verbose": 1,
            "approach": "single",  # "grouped" or "single"
            "model_params": {
                'n_components': ('int', 1, 30),
            },
        }
    }
]

pipeline_config = PipelineConfigs(pipeline, "OptimizedPipeline")
runner = PipelineRunner(save_files=False, verbose=1)
predictions, _ = runner.run(pipeline_config, dataset_config)

# Get the best optimized model
best_model = predictions.top(n=1, rank_metric='rmse')[0]
print(f"Best model: {best_model['model_name']} with RMSE: {best_model['rmse']:.4f}")

Visualization and Analysis

from nirs4all.data.prediction_analyzer import PredictionAnalyzer
import matplotlib.pyplot as plt

# Create analyzer for your predictions
analyzer = PredictionAnalyzer(predictions)

# Plot top performing models
fig1 = analyzer.plot_top_k_comparison(k=5, rank_metric='rmse')
plt.title('Top 5 Models Comparison')

# Create heatmap of model performance across preprocessing methods
fig2 = analyzer.plot_variable_heatmap(
    x_var="model_name",
    y_var="preprocessings",
    metric='rmse'
)
plt.title('Model Performance Heatmap')

# Candlestick plot for model variability
fig3 = analyzer.plot_variable_candlestick(
    filters={"partition": "test"},
    variable="model_name"
)
plt.title('Model Performance Variability')

plt.show(block=False)

Tutorials

NIRS4ALL provides comprehensive tutorials to help you master NIRS data analysis:

🚀 Tutorial 1: Beginner's Guide

Perfect for getting started with NIRS4ALL! This tutorial covers:

• Basic PLS Regression - Your first NIRS pipeline
• Enhanced Preprocessing - Spectral data preprocessing techniques
• Classification - Random Forest classification examples
• Model Persistence - Save and reuse trained models
• Multiple Datasets - Cross-dataset validation and analysis
• Data Visualization - Create meaningful plots and charts

Start here if you're new to NIRS analysis or the NIRS4ALL framework.

🔬 Tutorial 2: Advanced Analysis

For experienced users ready for sophisticated techniques:

• Multi-Source Analysis - Multi-target regression with single datasets
• Hyperparameter Optimization - Automated model tuning with Optuna
• Custom Components - Build your own transformers and models
• Configuration Generation - Dynamic pipeline customization
• Advanced Visualizations - Professional-grade analysis dashboards
• Neural Networks - Deep learning with pre-configured models (nicon, decon)
• Complete Workflows - End-to-end professional analysis

These tutorials demonstrate real-world workflows and best practices for production-ready NIRS analysis.

Examples

Ready-to-run example scripts demonstrating common NIRS workflows:

Basic Examples

• Q1_regression.py - Basic regression with PLS models and preprocessing combinations
• Q1_classif.py - Classification pipeline with Random Forest and preprocessing
• Q1_classif_tf.py - Classification with TensorFlow neural networks and confusion matrix visualization
• Q1_groupsplit.py - Group-based data splitting for maintaining sample integrity

Advanced Pipeline Techniques

• Q2_multimodel.py - Compare multiple model types (PLS, RF, SVM) in one run
• Q3_finetune.py - Hyperparameter optimization with Optuna
• Q4_multidatasets.py - Cross-dataset validation and transfer learning
• Q11_flexible_inputs.py - All possible input formats for PipelineRunner (configs, dicts, arrays, paths)
• Q12_sample_augmentation.py - Balanced sample augmentation for imbalanced classification datasets

Model Deployment & Prediction

• Q5_predict.py - Load saved models and predict on new data
• Q5_predict_NN.py - Prediction methods for neural network models
• Q14_workspace.py - Workspace management, library export, and global predictions database

Data Processing & Analysis

• Q6_multisource.py - Multi-target regression from single dataset
• Q7_discretization.py - Convert continuous targets to categorical
• Q8_shap.py - SHAP analysis for model interpretability
• Q9_acp_spread.py - PCA-based dataset analysis and visualization
• Q10_resampler.py - Wavelength resampling and interpolation techniques
• Q13_nm_headers.py - Working with nanometer (nm) wavelength headers instead of wavenumbers (cm⁻¹)

Custom Models

• custom_NN.py - Custom TensorFlow neural network architectures for NIRS
• custom_nicon.py - Custom NICON (NIRS Convolutional Network) model implementations

Run any example with: python examples/<example_name>.py t

Documentation

User Guide

• Preprocessing Guide - Complete reference of transformers (nirs4all, sklearn, scipy) with usage examples
• Preprocessing Cheatsheet - Quick reference for preprocessing operations
• Sample Augmentation Guide - Data augmentation techniques for NIRS

API Reference

• Data Module - Dataset handling and data loading APIs
• Pipeline Module - Pipeline configuration and execution APIs
• Workspace Module - Workspace management and organization

Specifications

• Pipeline Syntax - Complete pipeline configuration syntax
• Config Format - Pipeline configuration file format and structure
• Metrics - Available metrics and evaluation methods
• Nested Cross-Validation - Nested CV for unbiased hyperparameter tuning
• Cross-Dataset Metrics - Cross-dataset validation metrics
• Group Split - Group-based data splitting strategies
• Serialization - Pipeline serialization and deserialization

Explanations

• SHAP Explanation - Model interpretability with SHAP values
• Resampler - Wavelength resampling strategies
• SNV Explanation - Standard Normal Variate transformation
• PLS Study - Partial Least Squares regression analysis
• Metadata Usage - Working with dataset metadata

Reference

• Operator Catalog - Complete catalog of available operators
• Combination Generator - Feature augmentation and preprocessing combinations
• Writing Pipelines - Best practices for pipeline creation
• Outputs vs Artifacts - Understanding pipeline outputs
• Prediction Results - Understanding prediction results and metrics

Full documentation will be available at https://nirs4all.readthedocs.io/

Dependencies

• numpy (>=1.20.0)
• pandas (>=1.0.0)
• scipy (>=1.5.0)
• scikit-learn (>=0.24.0)
• PyWavelets (>=1.1.0)
• joblib (>=0.16.0)
• jsonschema (>=3.2.0)
• kennard-stone (>=0.5.0)
• twinning (>=0.0.5)
• optuna (>=2.0.0)

Optional Dependencies

• tensorflow (>=2.10.0) - For TensorFlow models
• torch (>=2.0.0) - For PyTorch models
• keras (>=3.0.0) - For Keras models
• jax (>=0.4.10) & jaxlib (>=0.4.10) - For JAX models

Research Applications

NIRS4ALL has been successfully used in published research:

Houngbo, M. E., Desfontaines, L., Diman, J. L., Arnau, G., Mestres, C., Davrieux, F., Rouan, L., Beurier, G., Marie‐Magdeleine, C., Meghar, K., Alamu, E. O., Otegbayo, B. O., & Cornet, D. (2024). Convolutional neural network allows amylose content prediction in yam (Dioscorea alata L.) flour using near infrared spectroscopy. Journal of the Science of Food and Agriculture, 104(8), 4915-4921. John Wiley & Sons, Ltd.

How to Cite

If you use NIRS4ALL in your research, please cite:

@software{beurier2025nirs4all,
  author = {Gregory Beurier and Denis Cornet and Camille Noûs and Lauriane Rouan},
  title = {nirs4all is all your nirs: Open spectroscopy for everyone},
  url = {https://github.com/gbeurier/nirs4all},
  version = {0.2.1},
  year = {2025},
}

License

This project is licensed under the CECILL-2.1 License - see the LICENSE file for details.

Acknowledgments

• CIRAD for supporting this research
• [LLMs] for providing fast documentation, nice charts, emojis in logs 😭, and plenty of useless tests, booby-trapped source code, and misleading specifications.