This is a repository with package implementing two different NMF variants that allow and account for negative data with weights:
- Shift-NMF
- Nearly-NMF
The algorithms are doocumented in a paper (on the arxiv: https://arxiv.org/abs/2311.04855). Code related to plots and data in the paper are contained in their own repository at https://github.com/dylanagreen/nearly_nmf_paper.
The code is well documented, with extensive docstrings. The code provides both a direct interface (via a fit_NMF function) as well as an object-oriented approach (through an NMF object) that is designed to be familiar (but not identical) to those working with scikit-learn models.
The code is also optimized for gpu use, and cupy is a necessary dependency for this to work. Enable GPU running by passing the data arrays as cupy arrays rather than numpy arrays.
nearly_nmf can be installed with pip install ., which pulls in the necessary numpy dependency but will not install cupy, if you want to enable gpu use you must install cupy yourself.