[WIP with rare updates!] Geometric Deep Learning library in modern jax (with orbax, grain, metrax, hf datasets) that implements:
- Basic ChemProp model for common datasets like delaney
- ChemPropPred training and arrhenius property predictors (also fixes this issue)
- MARCEL learning for conformers on drugs-75k
- MolSets for a mixture property prediction
- BlueSky data collection, analysis and GraphSAGE recommender system
Modern courses like https://geometricdeeplearning.com/ or https://dmol.pub/ all start with basics on machine learning, linear algebra, equivariances etc. I'm a big fan of https://www.fast.ai/ and a layered API, where we first get attention by building useful models (in this case for chemistry, social networks, 3D models etc) and then peel layers and undestand concepts deeper and deeper