sparseNMF

GPU-accelerated sparse non-negative matrix factorization with PyTorch.

sparseNMF factorizes very-large sparse non-negative matrices on the GPU without materializing a dense copy. Designed for biomedical data: gene-association counts, phenotype matrices, single-cell expression — anything that’s both sparse and too big to fit densely in VRAM.

Getting started

Reference

Project

Two flavors

Standalone sparse_nmf.SparseNMF: Drop-in replacement for sklearn.decomposition.NMF that runs on GPU and never densifies the input. Use it when you want W and H matrices directly.
Joint sparse_nmf.train_joint_model(): End-to-end training of an NMF + autoencoder pipeline. Use it when you want a low-dimensional embedding (e.g., 2-D for plotting, 192-D for downstream retrieval) rather than the full W matrix.

Indices and tables