Code, models, and runtime tooling for machine learning on astronomical data.

We focus on practical bottlenecks in survey science and astrophysics: reading astro data directly into GPU tensors, differentiable data-driven sky modelling, computing pixel-level weight maps, quantifying model uncertainties properly, and running reproducible, multi-node agentic workflows on the CANFAR Science Platform.

• torchfits (docs) — High-performance, PyTorch-native FITS I/O powered by a multi-threaded C++ engine. Reads and writes images, compressed HDUs, and tables directly as tensors, arrow compatibility, with built-in dataset loaders, ready for streaming large catalogs and cutouts into training loops.
• torchsky — Torch-native astronomy stack for survey-scale coordinates, spherical geometry, catalog indexing, mapmaking, and joint modelling of sky observations.
• torchsky-harness — Scientific validation, upstream parity, benchmarks, and real-data comparisons for torchsky.
• xmatcher — Catalogue crossmatching from local sky regions to distributed full-sky matching.
• weightmask — Fast generation of per-pixel weight, mask (cosmics, streaks, saturation, bad pixels), inverse-variance maps for astronomical images.
• torchregress — PyTorch library for regression, uncertainty quantification, especially suited for real, difficult regression problems. Not specific to astro data!
• torchregress-harness — External-software benchmarking suite for torchregress.

• canfar-lab — In-session workbench CLI for the CANFAR Science Platform. Manages reproducible project environments, resume/save work, autoscaling distributed Ray clusters, and coding agent setups from a single interface.
• canfar-containers — Lean, pre-configured session images (images.canfar.net/astroai/*) for astronomy and ML development on CANFAR.
• canfar-skills — Agent skills and automated scientific workflows tuned for the CANFAR ecosystem (Now integrated into the canfar python client)

• ssl-unions — Self-supervised representation learning and similarity search for galaxies in the UNIONS survey.
• starnet — Convolutional neural networks and notebooks for extracting stellar parameters and features from spectra (e.g., SDSS APOGEE).

There are many updates to come, stay tuned!