Version 0.3 (2025‑07‑15)

• Improve UMAP via direct gradient computation and edge masking #198
• Support for torch.compile #194
• Automatically handle duplicates #188
• Standardize logging #187
• Make affinity_out optional in AffinityMatcher #186
• Implement PHATE algorithm #185
• Implement PACMAP algorithm #182
• COSNE support for hyperbolic embeddings #178
• Allow for any Torch optimizer or scheduler #174
• Ensure compatibility with Python 3.8+ #173

07 Feb 2025

• FAISS support for KNN PR #160.
• CIFAR examples with DINOv2 features PR #158.
• Fast linter and formatter with Ruff PR #151.
• Pre-commit hooks added for code quality and consistency checks PR #147.
• Incremental PCA PR #137.
• Clean citation style via sphinxcontrib-bibtex PR #143.
• Functionality to switch to keops backend if it is installed and an out-of-memory error is raised PR #130.
• Code of conduct PR #127.
• Pull request template PR #125.

17 Sep 2024

It provides the following features:

• Multiple basic affinities, including scalar product, Gaussian, and Student kernels.
• Affinities based on k-NN normalizations such as Self-tuning affinities and MAGIC.
• Doubly stochastic affinities with entropic and quadratic projections.
• Adaptive affinities with entropy control (entropic affinity) and its symmetric version.
• Input and output affinities of UMAP.
• A template object AffinityMatcher to solve DR with gradient descent and any input and output affinities.
• Neighbor embedding methods like SNE, t-SNE, t-SNEkhorn, UMAP, LargeVis, and InfoTSNE.
• Template objects for neighbor embedding methods.
• Spectral embeddings via eigendecomposition of the input affinity matrix (when applicable).
• KeOps compatibility for all components, except spectral embeddings.
• Silhouette score.