Skip to content

AGENTS.md

Latest commit

 

History

History
31 lines (25 loc) · 2.14 KB

AGENTS.md

File metadata and controls

31 lines (25 loc) · 2.14 KB

AGENTS

Repository overview

UMAP C# is a C# reimplementation of the UMAP (Uniform Manifold Approximation and Projection) algorithm for non-linear dimensionality reduction. It takes high-dimensional vectors and produces 2D/3D (or higher) embeddings suitable for visualization or downstream ML tasks.

Key concepts

  • Embedding: The lower-dimensional representation produced from your input vectors.
  • Neighbor graph / fuzzy manifold: UMAP builds a k-nearest-neighbor graph and optimizes an embedding that preserves local structure.
  • Epochs: Optimization iterations. The recommended number of epochs is returned by InitializeFit.
  • Distance function: Controls how similarity between vectors is measured (e.g., cosine distance).
  • Random generator: Controls stochastic behavior and optional multi-threading.

Usage styles

  • One-shot embedding: Initialize, run the recommended number of steps, then read the embedding.
  • Custom epochs / iterative refinement: Provide a custom epoch count in the constructor or call Step manually in a loop.
  • Custom distance function: Provide your own distance function to match your data characteristics.

Key methods and functionality

  • Umap.InitializeFit(float[][] vectors): Prepares the model and returns the recommended number of epochs.
  • Umap.Step(): Performs one optimization step; call repeatedly to reach the desired number of epochs.
  • Umap.GetEmbedding(): Returns the final embedding vectors in the same order as the input.
  • Constructor options: dimensions, distance function, random generator, number of neighbors, custom number of epochs, and progress reporter.

Typical workflow

  1. Create Umap with desired configuration.
  2. Call InitializeFit with your float[][] vectors to build the graph and get the epoch count.
  3. Call Step for the recommended (or custom) number of epochs.
  4. Call GetEmbedding to retrieve the reduced-dimension vectors.

Parallelization notes

If you provide a thread-safe random generator, optimization steps may run multi-threaded for performance. To force single-threaded execution, use the provided option to disable threading on the default generator.