EO foundational model inference and fine tuning (self-directed training)

Objective

The governing objective is to learn how to run inference on finely-tuned models, and to fine-tune a foundational model. For this excercise I have chosen to use the IBM-NASA Prithvi Models Family

A more fine-grained list of objectives follows:

Checklist

• ✅ Understand foundational model architecture and capabilities
• ✅ Build a script to download HLS imagery from the Microsoft Planetary Computer archive
• Run inference on multiband HLS imagery for crop-coverage
  • ✅ Using Huggingface Docker file run locally
  • ❌ Using local Python environment
    • Failed thus far. Python 3.8 required, but it is not compatible with VS Code Python debugger. OpenMMLab API changed drastically since Python 3.8.
  • Need to update inference script to accommodate modern OpenMMLab API
  • Containerize new Python environment
  • Deploy onto cloud service with API
• Locally fine tune foundational model to predict 🤔
• Exlore could services to speed up fine tuning

Running the inference app in Docker

docker run -it --rm -v $PWD:/home/user/app -w /home/user/app -p 7860:7860 myapp

Acknowledgments

This project utilizes the Prithvi Models Family developed by IBM and NASA. Special thanks to the IBM-NASA Geospatial AI team for creating these foundational models for Earth observation tasks.

Notes