CropIntel

Web app for crop leaf disease hints using EfficientNet → TensorFlow Lite, with a Next.js UI.

Run without training (no Kaggle)

Trained weights live under ml/models/ (gitignored). You do not need Kaggle or local training if you install models once:

Option A — Download a release zip (recommended)

1. Get a direct .zip URL from a maintainer (e.g. GitHub Releases).
2. From the repo root:

pip install -r ml/requirements-inference.txt
export CROPINTEL_MODELS_URL='https://example.com/cropintel-models.zip'
python3 -m ml.scripts.fetch_models

3. Install Node deps and start the app:

npm install
npm run dev

Open http://localhost:3050. The API runs scripts/predict.py with python3.

Option B — Docker

export CROPINTEL_MODELS_URL='https://example.com/cropintel-models.zip'   # required for first-time fetch
docker compose up --build

Models are stored in ./ml/models on your machine. Set NEXT_PUBLIC_GOOGLE_MAPS_API_KEY in .env if you use the outbreak map.

Train yourself (needs Kaggle data)

See ml/README.md: Kaggle API, download_datasets.py, and training scripts. Use pip install -r ml/requirements.txt (includes kaggle).

Maintainer: publish models for others

After training, package ml/models/:

pip install -r ml/requirements.txt   # or minimal env with ml on PYTHONPATH
python3 -m ml.scripts.package_models -o cropintel-models.zip

Upload cropintel-models.zip to a release and share the direct download URL as CROPINTEL_MODELS_URL.

Project layout

• app/ — Next.js (UI + /api/predict)
• scripts/predict.py — inference entrypoint for the API
• ml/ — training, tflite_predictor, config

License

MIT