Skip to content

Releases: noosphaerenbauer/tiny-knn

Release list

v0.2.0 — 2025-08-25

v0.2.0 — 2025-08-25 Pre-release
Pre-release

Choose a tag to compare

View all tags
@noosphaerenbauer noosphaerenbauer released this 25 Aug 05:53
v0.2.0
e77b1fb

Overview

  - Minor release with a simplified, faster API and CLI. exact_search now returns results directly and supports Torch/NumPy arrays or file paths. Benchmarks, tests, and docs updated.

Highlights

  - exact_search(arr1, arr2, k, metric) returns (indices, scores).
  - Accepts Torch tensors, NumPy arrays, or .pt/.npy paths.
  - Metrics: ip (inner-product) or cosine (with internal normalization).
  - CLI simplified with --metric and optional --output-path.

Breaking Changes

  - Removed file-writing API. Previously: returns output path; Now: returns indices and scores.
  - Removed low-level CLI flags (device/dtype/batch/chunk/autotune/etc.); use defaults.

New

  - NumPy support: inputs can be np.ndarray or .npy files; results come back as NumPy arrays.
  - Mixed precision maintained (fp32/fp16; bf16 for Torch tensors).
  - Results types mirror inputs (Torch → Torch; NumPy → NumPy).

CLI Changes

  - Command: tiny-knn path/to/queries.(pt|npy) path/to/docs.(pt|npy) --k 100 --metric ip --output-path results.pt
  - Output format:
    - .pt/.pth: Torch dict with keys indices, scores
    - .npz: NumPy arrays indices, scores

Migration

  - Library:
    - Before: result_path = exact_search(q_path=..., d_path=..., k=..., out_path=...)
    - After: indices, scores = exact_search(arr1, arr2, k, metric="ip")
  - CLI:
    - Replace --normalize with --metric cosine.
    - Use --output-path to persist results.

Usage

  - Torch:
    - indices, scores = exact_search(queries_t, docs_t, 100, metric="ip")
  - NumPy:
    - indices, scores = exact_search(queries_np, docs_np, 100, metric="cosine")
  - Paths:
    - indices, scores = exact_search("queries.pt", "docs.pt", 100, metric="ip")

Verification

  - Tests updated and passing; benchmarks adapted to new signature.
Loading