fastica

This is an implementation of the fixed-point FastICA algorithm presented by Aapo Hyvärinen and Erkki Oja in Independent Component Analysis: Algorithms and Applications with an emphasis on performance, machine resource utilization, and portability of data format.

fastica requires certain versions of MKL, Intel-OpenMP, Cython, and numpy, which are as of this writing (March 2025) somewhat out of date. There is a torch backed implementation in development at cml tools v2). Nevertheless, on Intel machines, especially multi-node NUMA machines, the implementation here is still 2x faster than the best torch or numpy backed pure-Python implementation we have been able to develop. Since the package provides a CLI tool which reads and writes numpy ndarray data files (e.g., file.npy), pre and post processing may easily be done in any manner you choose. Our recommendation is to run fastica as a standalone CLI tool from a dedicated Conda environment.

Installation

This is a Cython project which requires a C++ build toolchain in a conda environment. The easiest way to install is to make sure you have micromamba on your path, and run:

./build_env.sh  # (or `source build_env.sh`)
make && make install

If running $ ./build_env.sh fails because bash cannot find micromamba, you can try source build_env.sh instead. build_env.sh runs micromamba to install from environment.yaml, and additionally both adds a few hooks with useful environment variables and checks the installed versions.

make by itself will run the basic build and tests. If the build fails, you might start by checking that you have the correct versions of gcc and g++. make install will install the fastica library and command line tool into the activated Python environment. Running the jupyter notebook Visual Test will provide an additional check that the fastica algorithm implementation is working correctly.

Usage Notes

fastica is both a library and a command line tool. After installation run fastica --help to begin exploring the CLI. There are four subcommands: whiten, run, recover, and project. whiten preprocesses a data matrix by PCA projection and a whitening transform. run performs ICA analysis of a whitened data matrix to produce a matrix W, from which recover will produce the matrices A and S of the ICA formula X = AS. project will apply the learned preprocessing steps (centering and whitening) and the ICA mixing matrix to produce projections of new data through the ICA model. See also the example script example_fastica.sh.