# CropIntel ML stack: reproducible Python + TensorFlow without touching host Python.
# Data and trained weights live in mounted volumes (still gitignored on the host).
#
# Quick smoke test (no Kaggle):
#   docker compose -f docker-compose.ml.yml build
#   docker compose -f docker-compose.ml.yml run --rm ml \
#     python -m ml.scripts.create_synthetic_dataset --crop corn --force
#   docker compose -f docker-compose.ml.yml run --rm ml \
#     python -m ml.training.train_crop --crop corn --epochs 2 --no-fine-tune
#
# With Kaggle (mount token; dataset terms must be accepted on the website):
#   docker compose -f docker-compose.ml.yml run --rm -v "$HOME/.kaggle:/root/.kaggle:ro" ml \
#     python -m ml.scripts.download_datasets

services:
  ml:
    build:
      context: .
      dockerfile: Dockerfile.ml
    image: cropintel-ml:local
    working_dir: /app
    environment:
      PYTHONPATH: /app
    volumes:
      - ./ml/data:/app/ml/data
      - ./ml/models:/app/ml/models
      # Optional Kaggle token (only if downloading inside Docker):
      #   docker compose -f docker-compose.ml.yml run --rm -v ~/.kaggle:/root/.kaggle:ro ml ...