ArchePy

Multi-Subject Archetypal Analysis (MS-AA) for fMRI data, in Python.

A Python port of the Multisubject Archetypal Analysis Toolbox originally developed in MATLAB at the Technical University of Denmark by Hinrich, Bardenfleth, Røge, Churchill, Madsen, and Mørup. Implements both spatial and temporal MS-AA with optional GPU acceleration via CuPy.

[!IMPORTANT] Academic / non-profit use only. This package is distributed under the same license as the upstream MATLAB toolbox, which restricts use to academic and non-profit institutions. Commercial users must obtain a separate license from DTU. See LICENSE for full terms.

[!NOTE] Related packages. ArchePy is one of several Python tools for archetypal analysis. If MS-AA isn't what you need, you may want:

• archetypes — general scikit-learn-compatible AA (BSD-3).
• py_pcha — Python implementation of PCHA, by one of MS-AA's upstream authors (Mørup).

ArchePy is specifically the multi-subject, heteroscedastic, GPU-accelerated variant for neuroimaging, ported from the DTU MATLAB toolbox.

Installation

# CPU only
pip install archepy

# With GPU acceleration (CUDA 12.x — requires a CUDA-capable NVIDIA GPU)
pip install archepy[gpu]

# With fMRI helpers (NIfTI loading)
pip install archepy[fmri]

# Everything
pip install archepy[all]

For development:

git clone https://github.com/the-brAIn-lab/archepy
cd archepy
pip install -e .[dev]
pytest

Requirements

• Python 3.9 or newer.
• Required: NumPy >= 1.23.
• Optional, for GPU acceleration: CuPy (CUDA 12.x via archepy[gpu]). CUDA 11.x users should install cupy-cuda11x manually.
• Optional, for fMRI helpers: nibabel >= 4.0 (via archepy[fmri]).
• Optional, for plotting: matplotlib >= 3.5 (via archepy[viz]).

The full dependency list is declared in pyproject.toml.

Quick start

import numpy as np
from archepy import Subject, multi_subject_aa

# Each subject's data: a (T, V) array — time points × voxels.
# (For temporal MS-AA, see archepy.multi_subject_aa_T which uses (V, T).)
rng = np.random.default_rng(0)
n_subjects, T, V = 3, 200, 1000

subjects = [
    Subject(
        X=rng.standard_normal((T, V)).astype(float),
        sX=rng.standard_normal((T, V)).astype(float),
    )
    for _ in range(n_subjects)
]

# Fit MS-AA with K=10 archetypes.
results, C, cost, varexpl, elapsed = multi_subject_aa(
    subjects,
    noc=10,
    opts={
        "maxiter": 100,
        "conv_crit": 1e-6,
        "use_gpu": False,           # set True if you installed archepy[gpu]
        "heteroscedastic": True,
        "rngSEED": 42,
    },
)

print(f"Variance explained: {varexpl * 100:.1f}%   ({elapsed:.1f}s, {len(cost)} iters)")
print(f"Shared generator C shape:        {C.shape}")
print(f"Subject 0 archetypal mixture S:  {results[0]['S'].shape}")
print(f"Subject 0 reconstructed XC:      {results[0]['sXC'].shape}")

For a fuller end-to-end example using real fMRI data, see examples/01_fit_msaa_flexible.ipynb.

What this package contains

Testing

ArchePy includes a smoke-test suite that exercises the FurthestSum initialization and runs end-to-end MS-AA fits on planted low-rank data. After installing the dev extras (pip install -e .[dev]), run:

pytest                                          # all tests
pytest --cov=archepy --cov-report=term-missing  # with coverage

The full suite finishes in seconds. CI runs the same suite on Linux, macOS, and Windows across Python 3.9-3.12 — see .github/workflows/ci.yml.

Contributing

Contributions are welcome — bug reports, feature requests, documentation improvements, and pull requests. Please:

• Read CONTRIBUTING.md for the dev-environment setup, code style, and PR workflow.
• Review the CODE_OF_CONDUCT.md before participating.
• For non-trivial changes, open an issue first to discuss.

By contributing, you agree that your contributions will be distributed under the same academic / non-profit terms as the rest of the package.

Citing

If you use ArchePy in a publication, please cite both the original paper and this port:

@article{hinrich2016archetypal,
  title   = {Archetypal Analysis for Modeling Multisubject {fMRI} Data},
  author  = {Hinrich, Jesper L\o{}ve and Bardenfleth, Sophia Elizabeth and
             R\o{}ge, Rasmus and Churchill, Nathan W. and Madsen, Kristoffer H.
             and M\o{}rup, Morten},
  journal = {IEEE Journal on Selected Topics in Signal Processing},
  volume  = {10},
  number  = {7},
  pages   = {1160--1171},
  year    = {2016},
  doi     = {10.1109/JSTSP.2016.2595103}
}

@software{archepy,
  author  = {Alex Shepherd},
  title   = {ArchePy: Multi-Subject Archetypal Analysis for fMRI in Python},
  url     = {https://github.com/the-brAIn-lab/archepy},
  version = {0.1.0},
  year    = {2026}
}

Acknowledgments

This package would not exist without the original MATLAB toolbox by Jesper L. Hinrich, Sophia E. Bardenfleth, and Morten Mørup at DTU's Section for Cognitive Systems. The Python port preserves their algorithms faithfully; any bugs in this port are mine alone.

Status

Alpha. The algorithm code is a direct port of the MATLAB reference and reproduces its outputs, but the Python API may evolve before 1.0. Pin to an exact version if you need stability.