Machine Learning Primitives

Whenever I go to use a new machine learning algorithm / architecture / technique, I normally write it out myself first using plain Rust and NDArray. Doing all the differentiation and grunt work helps me appreciate this stuff more, and I like that nothing is hidden. I figured I would collect these handwritten architectures as I go, into a kind of mini-framework / primitives set.

Sections

• envs: Various simulators used to test stuff
• nn: Core neural net components, stuff like basic FFN, Attention, Layernorm, CTRNN, etc. TBD some experimentals in the nn/experimental dir.
• optim: Gradient collection and training, AdamW, SGD
• rl: A couple RL algos, SAC is still a bit sketch, but PPO & TD3 are solid.
• util: Graphing and benchmarking
• f: Function set