Skip to content

README.md

Latest commit

 

History

History
92 lines (67 loc) · 3.51 KB

README.md

File metadata and controls

92 lines (67 loc) · 3.51 KB
Raw

ComputerVisionAIHub

An open catalog of custom-trained YOLO models. Browse models, download the weights from the HuggingFace Hub, and run them locally with one Docker command.

How it works

Roboflow dataset ──► Google Colab (train YOLO model) ──► Hugging Face Hub (weights)
                                                            │
   GitHub repo (code) ──► GitHub Pages (catalog) ───────────┘
                                   │
                          user downloads + runs via Docker

Repo layout

docs/         
  index.html  structure
  styles.css  design tokens + the bounding-box card styling
  app.js      fetches models.json and renders the catalog
  models.json THE CATALOG — add a model by appending
notebooks/
  train_yolo26.ipynb   end-to-end training + publishing in HuggingFace
scripts/
  publish_to_hf.py     push existing weights trained locally to HuggingFace from your terminal
inference/
  Dockerfile  detect.py  app.py  requirements.txt   local model runner
LICENSE       AGPL-3.0

For users: run a model

Prerequisites:

Every catalog card has a copy-paste command. The model argument is the Hugging Face URL — the container downloads it on first run.

# build once (make sure you build from computervisionaihub directory)
docker build -t computervisionaihub:1.0 ./inference

# CLI: annotate one image
docker run -v $(pwd):/data computervisionaihub:1.0 \
  https://huggingface.co/youruser/yourmodel/resolve/main/best.pt \
  /data/test.jpg

# web UI: drag-and-drop at http://localhost:7860
docker run -p 7860:7860 --entrypoint python computervisionaihub:1.0 app.py

For maintainers: add a model

  1. Open notebooks/train_yolo26.ipynb in Google Colab, set runtime to a T4 GPU (free resource).
  2. Fill in your Roboflow + Hugging Face details and run all cells. It trains YOLO26, exports .pt + .onnx, pushes them to the Hub, and prints a ready-to-paste models.json entry.
  3. Paste that entry into the models array in docs/models.json, edit the human-readable fields (name, summary, dataset link, license), and commit.

That's it — no website code changes. (Trained locally instead - Use python scripts/publish_to_hf.py --repo youruser/model-id --pt best.pt --onnx best.onnx.)

Local preview

The site fetches models.json, so open it through a server (not file://):

cd docs && python -m http.server
# visit http://localhost:8000

A note on licensing

Ultralytics YOLO is AGPL-3.0, a strong copyleft license. Because the training notebook and Docker image use the Ultralytics library, this repo is AGPL-3.0 too. Two practical consequences:

  • Anyone redistributing or network-serving Ultralytics-based code inherits AGPL obligations. That's fine for an open project, but worth telling your users.
  • The ONNX export can be run with onnxruntime without Ultralytics, letting downstream users avoid AGPL in their own apps. That's why every model ships both formats.

Each model's underlying dataset has its own license (shown on each card). Check it before any commercial use.