yolozu 1.0.1

Interface-contract-first evaluation toolkit for mitigating catastrophic forgetting and domain shift via reproducible workflows.

YOLOZU

日本語: Readme_jp.md

Contract-first evaluation toolkit for vision models (detection / segmentation / pose). Run inference anywhere, export outputs as a stable predictions.json (+ export_settings), then evaluate with a fixed protocol for reproducible, apples-to-apples comparisons.

Quickstart (run this first)

bash scripts/smoke.sh

Output artifact: reports/smoke_coco_eval_dry_run.json.

Docs index (start here): docs/README.md.

AI-friendly tool registry (source of truth): tools/manifest.json.

Tool list + args examples: docs/tools_index.md.

Learning features (training / continual learning / TTT / distillation): docs/learning_features.md.

Start here (choose 1 of 4 entry points)

• A: Evaluate from precomputed predictions (no inference deps) — predictions.json → validate → eval.
• B: Train → Export → Eval (RT-DETR scaffold + run interface contract / Run Contract) — run artifacts → ONNX → parity/eval.
• C: Interface contracts (predictions / adapter / TTT protocol) — schemas + adapter interface contract boundary + safe adaptation protocol.
• D: Bench/Parity (TensorRT / latency benchmark) — parity checks + pinned-protocol benchmarks.

All four entry points are documented (with copy-paste commands) in docs/README.md.

CLI note:

• yolozu ... is the pip/package CLI.
• python3 tools/yolozu.py ... is the repo wrapper CLI.
• For equivalent commands, swap only the executable (yolozu ↔ python3 tools/yolozu.py).

Key points

• Bring-your-own inference → stable predictions.json interface contract.
• Validators catch schema drift early.
• Protocol-pinned export_settings makes comparisons reproducible.
• Parity/bench quantify backend drift and performance.
• Tooling stays CPU-friendly by default (GPU optional).
• Apache-2.0-only ops policy is enforced in repo tooling.

Why YOLOZU?

YOLOZU standardizes evaluation around a predictions-first interface contract: run inference anywhere, export predictions.json (+ export_settings), then validate and evaluate with fixed protocols for reproducible comparisons.

Details: docs/yolozu_spec.md.

Install (pip users)

python3 -m pip install yolozu
yolozu --help
yolozu doctor --output -

Optional extras and CPU demos: docs/install.md.

Source checkout (repo users)

python3 -m pip install -r requirements-test.txt
python3 -m pip install -e .
python3 tools/yolozu.py --help
python3 -m unittest -q

Manual (PDF)

Printable manual source: manual/.

Support / legal

• Contact: [e-mail]
• © 2026 ToppyMicroServices OÜ Full support/legal: docs/support.md.

License

Code in this repository is licensed under the Apache License, Version 2.0. See LICENSE.