feat(visdrone): live browser demo (path 1) + preview ONNX weights

[aalvsz]

Summary

Closes the "Path 1: pending trained weights" stub from the original VisDrone PR. Paths 1 (browser) and 2 (Python) now both work against a real, MIT-licensed preview model.

What's live

🤗 Preview weights: https://huggingface.co/ander2221/visdrone-yolo9-preview
🌐 Browser demo: `visdrone-finetune/demo/index.html` (zero-install, fetches ONNX from HF on first visit, caches)

The demo:

• Self-contained single HTML file matching `blur-faces/demo/index.html` conventions
• Inference via onnxruntime-web — WebGPU first, WASM fallback
• Image upload + webcam input
• Boxes + class labels drawn on a canvas
• 10 VisDrone classes color-coded
• Override source repo via `?repo=org/name` URL param

What's in the PR

File	Role
`demo/index.html`	Browser demo (~300 LOC, vanilla JS module)
`src/export_and_push.py`	torch → ONNX (dynamic batch) + HF Hub upload + auto model card
`src/load_finetuned.py`	Helper reproducing libreyolo's `_rebuild_for_new_classes` path so the trained hybrid checkpoint loads cleanly
`src/train.py`	Drop hardcoded `nb_classes=10` — let the factory auto-detect from COCO weights and the trainer rebuild from `data.yaml`
`src/infer.py`	Use the new load helper
`README.md`	Path 1 banner flipped to "Live (preview)" with honest status
`.gitignore`	Ignore `logs/` and `export/` (transient)

Training honesty

These are preview weights, not production:

• Trained on a Mac M-series GPU (Apple Metal Performance Shaders) — not a real datacenter GPU
• 5 epochs, imgsz=384, batch=8, ~12 min wall clock
• Full Voxel51/VisDrone2019-DET train split (7766 images)
• Loss dropped 14.9 → 5.4

Detections are real and look right on held-out val images — e.g. 34 cars + 1 bus correctly identified on `9999938_00000_d_0000496.jpg` at conf 0.15+. Confidences are modest (0.2-0.6 typical) because the model is undertrained. A full ~50-epoch run on a real GPU would replace these.

The model card on HF Hub explicitly tags this as v0.1-preview with the same caveats.

Upgrade path

When real weights are trained:

1. Train on whatever GPU is available (script unchanged).
2. Run `python -m src.export_and_push --weights weights/visdrone.pt --repo-id LibreYOLO/visdrone-yolo9 ...`
3. Update the demo's default `HF_REPO` constant in `demo/index.html` (or just expect users to use the `?repo=` URL param).

Test plan

• `python -m src.train` runs end-to-end on MPS without errors
• `python -m src.infer` produces sensible detections on val images
• `python -m src.export_and_push` produces working ONNX and successfully uploads to HF Hub
• HF repo is publicly accessible (302 → CDN, content-type ok)
• ONNX loads in onnxruntime-cpu, output shape `(1, 14, 3024)` for imgsz=384, range looks right (logits, not probabilities)
• Browser demo end-to-end test on a fresh Chrome — pending; see screenshots in next iteration if needed

Related

• Original VisDrone PR: #1 (merged)
• Issue thread: LibreYOLO/libreyolo#99 (closed)

🤖 Generated with Claude Code