pyamica

PyTorch implementation of AMICA (Adaptive Mixture Independent Component Analysis).

AMICA fits a mixture of ICA models to multi-channel time-series data. Each model has its own unmixing matrix and source densities; a per-sample posterior indicates which model most likely generated each data point. This makes it well-suited for EEG recordings where different experimental conditions have different source statistics.

Install

pip install pyamica            # core (no MNE)
pip install "pyamica[mne]"     # with MNE-Python integration

With uv:

uv add pyamica
uv add "pyamica[mne]"

Quick Start

import torch
from pyamica import AMICA, AmicaICA

# Low-level PyTorch API - X shape: (T, n_channels), float64
model = AMICA(n_models=2, max_iter=200, device="cuda")
model.fit(X)
print(model.gm_)           # global model weights
print(model.posteriors_)   # (n_models, T) per-sample posteriors

# MNE-Python wrapper
ica = AmicaICA(n_models=2, max_iter=200, device="cuda")
ica.fit(raw, picks="eeg")
ica.plot_model_dominance(smooth_s=1.0)
for m in range(ica.n_models):
    ica.review(raw, model_idx=m, eog_ch=["HEOG", "VEOG"], ecg_ch="ECG")
ica.apply(raw)

Background

pyamica is a pure Python reimplementation of AMICA built on PyTorch, with an MNE-Python wrapper to make it usable in modern EEG pipelines. It runs on CPU and GPU alike and slots in wherever mne.preprocessing.ICA would be used.

It started as a weekend project and a practical test of LLM-assisted code translation from the original Fortran implementation (sccn/amica). The translation was supported by Claude Sonnet 4.6 with a lot of back-and-forth, and turned out good enough (I think) to publish.

All credit for the algorithm, theory, and original implementation goes to Jason A. Palmer and the SCCN team. The canonical reference is:

Palmer, J. A., Kreutz-Delgado, K., & Makeig, S. (2012). AMICA: An Adaptive Mixture of Independent Component Analyzers with Shared Components. Technical Report, UC San Diego. [PDF]

The AMICA introduction on their wiki covers what it does, why it works, and how to interpret multi-model fits.

Development Install

git clone https://github.com/DerAndereJohannes/pyamica
cd pyamica/
pip install -e ".[mne,dev]"

With uv:

git clone https://github.com/DerAndereJohannes/pyamica
cd pyamica/
uv sync --extra mne --extra dev

Tests

# fast tests (synthetic data only)
pytest tests/ -v -m "not slow and not gpu"

# with coverage
pytest tests/ -v -m "not slow and not gpu" --cov=pyamica

# Fortran comparison (requires data/ directory)
pytest tests/ -v -m slow

With uv:

uv run pytest tests/ -v -m "not slow and not gpu"

Acknowledgements

This repository includes the pre-compiled amica15ub binary and default parameter file from the AMICA project by Jason Palmer and contributors, used here for development and testing only. They are not distributed as part of the installable package. That binary is licensed under the BSD 2-Clause License; see LICENSES/sccn-amica.txt for the full license text.