Skip to content

YOLOZU 4.8.0

Choose a tag to compare

View all tags
@thinksyncs thinksyncs released this 10 Sep 23:35
· 13 commits to main since this release
v4.8.0
bd51c77
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Summary

YOLOZU 4.8.0 adds a reusable labeled sample and improves first-run evaluation
and prediction-file compatibility.

Install from PyPI:

python -m pip install "yolozu[coco]==4.8.0"
yolozu demo dataset --run-dir sample-data --seed 0

Or download the ready-made labeled sample ZIP.
It includes images, YOLO labels, known predictions, a preview, and reuse instructions.

Changes

  • Generate eight synthetic images with YOLO bbox labels, a preview, known
    predictions, and a versioned manifest using yolozu demo dataset.
  • Keep samples portable across directories and older compatible readers.
  • Repair doctor proof images and improve installation and evaluation guidance.
  • Reuse proxy mAP preparation across thresholds; reject ambiguous COCO image
    matches while preserving exact-path and unique-basename behavior.
  • Reject unsupported prediction versions and invalid strict class IDs; keep
    metadata-free migration output valid.
  • Synchronize the CLI manifests, generated documentation, and TeX/PDF manual.

Testing

  • The final source-checkout suite ran 1,736 tests with 18 environment skips,
    zero failures, and zero errors; Ruff and whitespace checks passed.
  • The 4.8.0 candidate passed installed-wheel identity checks, pip check, strict
    sample validation, and official COCO evaluation before and after relocation.
  • Published 4.6.0 and 4.7.0 both read the same generated sample on macOS arm64 /
    Python 3.14.6 without changing files or evaluation results.
  • All eight candidate CI lanes passed: Linux Python 3.10–3.14, macOS Python
    3.14, Windows Python 3.12, and Linux Python 3.10 with declared core/COCO
    dependency floors.
  • A two-image CPU Mask R-CNN execution check passed eager/TorchScript parity
    with identical COCO metrics. This is not a performance benchmark.
  • The 175-page local manual passed qpdf/Ghostscript structural checks and
    all-page rendering with Poppler, PDFium, and PDFKit. The release rebuild is
    verified separately after publication.

Post-publication checks

  • The actual public wheel and source archive match PyPI hashes. A fresh macOS
    arm64 / Python 3.14.6 public-wheel installation passed both 13-check sample
    compatibility and 13-check first-use journeys, including official COCO
    evaluation and relocation. All 360 installed package files match the wheel.
  • All eight compatibility lanes passed again on the exact release commit.
  • The release ZIP was downloaded again and all 23 sample files matched.
  • The public Manual 4.8.0 matches the
    release workflow PDF. Its 175 pages passed structural checks and rendering
    with all three engines. All contact sheets and selected full-size pages were
    reviewed; no layout blocker was observed at those scales.

Notes

The sample's predictions come from its labels. Its perfect score is a workflow
check, not model accuracy. This release does not establish GPU/MPS performance,
support for every newer dependency, or the cause of low adoption.

Assets 4
Loading