Skip to content

Update detector runtime and evaluate YOLO26 models - #21

Merged
matteius merged 1 commit into
mainfrom
chore/python314-detector-models
Nov 13, 2026
Merged

matteius merged 1 commit into
mainfrom
chore/python314-detector-models

Conversation

@matteius

Copy link
Copy Markdown
Contributor

Summary

  • Upgrade the service and Docker runtime to Python 3.14 and pin current runtime dependencies from Pipfile.lock through requirements-runtime.txt.
  • Replace the retired tflite-runtime wheel with LiteRT, share the updated runtime base with the Kubernetes hosting image, and preserve its two-thread ONNX setting.
  • Add a YOLO26s NMS-free ONNX candidate, YOLOE trial and ONNX comparison scripts, export-shape validation, and CI checks for the model export and dependency pins.

Validation

  • pipenv verify and git diff --check
  • pipenv run python -m unittest discover -s tests -q (79 passed)
  • Built and smoke-tested runtime and hosting Docker targets on Python 3.14.10; checked /health and /api/v1/detect over HTTP.
  • Built the optional Moondream image and imported Torch, Transformers, and Accelerate.
  • Exported YOLO26s with (1, 300, 6) output; ran YOLO26s and YOLOE on a sample image.

Limits

Python 3.15.2 is the latest stable Python, but ONNX Runtime has no Python 3.15 wheel yet. Python 3.14.10 is the newest usable runtime for the default backend. OpenVINO builds currently need the documented Python 3.13 variant. A production model switch needs labeled camera frames; the single-image comparison does not establish recall or false-alarm rates. P2 training needs a camera dataset. EdgeTPU and OpenVINO hardware paths were not exercised locally.

@matteius
matteius merged commit b06a6d3 into main Nov 13, 2026
4 checks passed
@matteius
matteius deleted the chore/python314-detector-models branch November 13, 2026 04:22
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant