eusipco_2026

Research repository for EUSIPCO 2026 submission: Adaptive Geometric-Arithmetic-Harmonic (GAH) mean for SPD neural networks

This is a private paper repository for the second research phase focusing on the adaptive GAH method with simulation component.

Overview

This repository contains:

• Adaptive GAH callback: Custom PyTorch Lightning callback to track the learnable parameter evolution during training
• Simulation module: Complete simulation framework for studying BatchNorm mean type impact on SPD classification
• Experiment scripts: Shell scripts for running adaptive GAH experiments on various datasets
• Evaluation tools: Scripts for aggregating and analyzing experimental results

Architecture

The repository is organized to separate concerns:

• Data generation is provided by spdnet-datasets (synthetic data generation)
• Training framework is provided by spdnet-training (main training loop, callbacks, utilities)
• This repository contains paper-specific components:
  • Adaptive GAH parameter logger
  • Simulation training/testing logic
  • Paper-specific experiment configurations
  • Result aggregation and analysis scripts

Installation

This repository depends on three core modules that are installed automatically from GitHub:

Module	Role
yetanotherspdnet	SPD layers (BiMap, ReEig, LogEig, BatchNorm)
spdnet-datasets	Dataset loaders, synthetic data generators
spdnet-training	Lightning module, training loop, callbacks

# Clone the repository
git clone https://github.com/Yet-Another-Research-Organisation/eusipco_2026.git
cd eusipco_2026

# Create virtual environment (Python >= 3.11 required)
uv venv .venv --python 3.11
source .venv/bin/activate

# Install in development mode (fetches core modules automatically)
uv pip install -e ".[dev]"

# Set up environment variables
cp .env.example .env
# Edit .env to set DATA_ROOT to your dataset location

All three core modules are fetched and installed automatically during pip install. No manual cloning required.

Usage

Running Simulation Experiments

# Quick test
python -m eusipco_2026.simulation --quick --cpu

# Full Wishart/Inverse-Wishart experiments
python -m eusipco_2026.simulation --wishart-inverse --n-jobs 24 --output-dir results/wishart

# Custom configuration
python -m eusipco_2026.simulation --custom --matrix-size 16 --n-discriminant 2 --seeds 42 123 456

Running Adaptive GAH Experiments on Real Datasets

# HDM05 dataset
./scripts/adaptive_gah_experiment_hdm05.sh

# HyperLeaf dataset
./scripts/08_adaptative/adaptive_gah_experiment_hyperleaf.sh

# RICE S90 dataset
./scripts/08_adaptative/adaptive_gah_experiment_rices90.sh

Evaluating Results

# Aggregate results from multiple runs
python -m eusipco_2026.evaluate.aggregate_experiment_results \
    --results-dir results \
    --group spdnet_hdm05 \
    --name ADAPTIVE_GAH_HDM05

# Extract and generate LaTeX tables
python -m eusipco_2026.evaluate.extract_results results/spdnet_hdm05 -o tables/hdm05_results.tex

Dependencies

Core dependencies:

• spdnet-training: Training framework and utilities
• spdnet-datasets: Dataset loaders and synthetic data generation
• yetanotherspdnet: SPD neural network layers and operations
• PyTorch Lightning: Training framework
• Hydra: Configuration management

Project Structure

eusipco_2026/
├── src/eusipco_2026/
│   ├── callbacks/          # Adaptive GAH parameter logger
│   ├── simulation/         # Simulation training/testing logic
│   ├── evaluate/           # Result aggregation and analysis
│   └── train.py           # Training entry point
├── scripts/               # Experiment shell scripts
├── configs/               # Hydra configuration files
└── tests/                 # Unit tests

Citation

If you use this code in your research, please cite:

@inproceedings{gallet2026adaptive,
  title={Adaptive Geometric-Arithmetic-Harmonic Mean for SPD Neural Networks},
  author={Gallet, Matthieu and Mian, Ammar},
  booktitle={European Signal Processing Conference (EUSIPCO)},
  year={2026}
}

License

MIT License - see LICENSE file for details.

Authors

• Matthieu Gallet
• Ammar Mian (ammar.mian@univ-smb.fr)