seeni

Offline-capable mushroom identifier for Finnish species. Image classification runs entirely in the browser as a PWA, so there is no server and no uploads.

Seeni never tells you whether a mushroom is edible. It suggests candidate species, always shows their dangerous lookalikes, and walks you through checks that help tell them apart. Never eat a mushroom based on this app.

Status: early development. There is no trained model or published app yet. The planned address is https://luukalindgren.github.io/seeni/, later seeni.fi.

How it works

1. A model is trained on CC0 and CC-BY licensed mushroom photos from GBIF on a desktop GPU (training/).
2. It is exported to ONNX together with its class list and preprocessing settings (web/public/models/vN/).
3. The web app downloads the model once, caches it, and runs it in a Web Worker with onnxruntime-web (WebGPU, falling back to WASM). Photos stay on your device, and identification works offline.
4. Results are shown as candidates with their dangerous lookalikes and verification checks from hand-curated species data (species/). Every species entry must be reviewed by an expert before release.

Repository layout

Path	What
training/	Python (uv, PyTorch, timm): data download, training, evaluation, ONNX export
data-meta/	Species list, train/val/test split and attribution data (small files only)
species/	species.json: species information and verification checks
web/	Vite + TypeScript PWA, deployed to GitHub Pages
docs/PLAN.md	The full step-by-step plan (in Finnish)
docs/ssh.md	SSH from the Mac to the training desktop over Tailscale
CLAUDE.md	Binding project rules: image licenses, preprocessing, class ids, safety

Training images, datasets and checkpoints are never committed. They live in ~/seeni-data on the training machine. The only exceptions are the exported model, app and documentation images, and test fixtures, listed in CLAUDE.md hard rule 2.

Getting started

Training (Python 3.12 and uv; training itself needs Linux or WSL2 with an NVIDIA GPU, and development also works on an Apple Silicon Mac):

cd training
uv sync
uv run pytest

See training/README.md for the GPU check and the data directory.

Web app (Node 22.12 or newer):

cd web
npm ci
npm run dev

Plan and progress

The work is split into eleven phases described in docs/PLAN.md. Each phase is also a GitHub issue with the same number (phase 6 is #6), and the checklist in CLAUDE.md shows which are done.

License

• Code is licensed under the MIT License.
• Species data (species/, data-meta/lajilista.csv, data-meta/split.csv) is licensed under Creative Commons Attribution 4.0 International (CC BY 4.0), with credit to the sources listed in ATTRIBUTIONS.md.
• Model files (web/public/models/) will be released under CC BY 4.0, provided the licenses of the pretrained weights they start from allow it. Those licenses are checked before the first model is published (#6, #11).
• Attribution data in data-meta/ reproduces each photographer’s credit and license, and is not relicensed.
• Training images belong to their photographers and are used under their own CC0 or CC-BY licenses. They are not distributed in this repository.

Photographers, data sources, pretrained weights and third-party software are credited in ATTRIBUTIONS.md.

This site is open source. Improve this page.