yutosaito0629/zzz

Score SN44 Scene Distribution Analyzer

Deep analysis tool for Score (Bittensor SN44) YOLO Element datasets — especially Detect-petrol-station — focused on the scene regimes that actually move mAP@50 / precision / recall.

Quick start

cd d:\work\Tools\image_analyzation
pip install -r requirements.txt
$env:PYTHONPATH = "d:\work\Tools\image_analyzation"
$env:PYTHONIOENCODING = "utf-8"

python cli.py ui
# -> http://127.0.0.1:8744

1. Analyze — paste eval folder → Analyze & cluster (also auto-prepares Gemini package from distribution)
2. Image Generation — uses analysis-prepared prompts/quotas (not hand-written package alone)
3. Test generate (1 image) — review in the gallery
4. Final generate (all) — all sources × prepared targets
5. Analyze generated — Score distribution on the generated run

YOLO layout: images/ + labels/ + data.yaml.

Gemini generation

After analyze, the tool writes outputs/score_active/gen_package/ with per-image prompts that include measured regimes, mix weights, gaps, and Gemini cues.

# one image smoke test (uses prepared package)
python cli.py gemini-test -s 1

# full run
python cli.py gemini-final

Outputs: outputs/generated/<mode>_<timestamp>/ with images/, draft labels/, data.yaml, run_manifest.json.

Put key in .env:

GEMINI_API_KEY=your_key_here

Other commands

scan, compare, snap, timeline, cluster, duplicates, outliers, embed, score-analyze, gen-dist-detail.