Connor Stone - AstroPhot: Fitting Everything Everywhere all at Once in Astronomical Images

AstroPhot and Caustics are two Python packages developed for analyzing astronomical images, particularly focusing on gravitational lensing and galaxy fitting

While GPU acceleration provided 10x speedup, it wasn’t always cost-effective compared to using multiple CPU cores - GPUs need to be 50x faster to justify their higher cost ($14,000 vs $5,000 for CPUs)

The development workflow includes:

• Automated testing across Windows, Linux, Mac
• Unit tests and notebooks for documentation
• Automatic website updates
• Deployment to pip and Conda
• Release notes generated from branch discussions

Small operations and branching statements/conditionals significantly reduce GPU performance benefits - optimal GPU usage requires bundling many operations together

The codebase is built on PyTorch, providing:

• GPU acceleration capabilities
• Automatic derivative tracking
• Backend numerical library support

The packages offer three interface levels:

• Low-level functional programming layer
• Object-oriented middle layer
• High-level simulator interface

Future telescopes will provide over 100,000 new gravitational lensing images, requiring scalable analysis tools

Code documentation includes direct links to equation numbers in referenced papers

The project demonstrates successful academic software development:

• User-friendly design
• Comprehensive documentation
• Professional software engineering practices
• Active community engagement

The tools support Bayesian inference and Markov chain analysis for astronomical image processing using score-based models