From‑scratch reverse‑mode autodiff engine (scalar & NumPy tensor) with optimizers and experiments
Provides a working reverse‑mode automatic differentiation engine built from first principles, offering both a scalar Value class and a NumPy‑backed tensor implementation. It bundles a full optimizer suite (SGD, AdaGrad, RMSProp, Adam) and gradient‑checking utilities, plus visualisations and experiments like learnable activation blends. Ideal for students, researchers, and ML hobbyists who want a transparent, extensible foundation for neural‑net experiments. Stands out by extending Karpathy's micrograd with a vectorised engine, thorough analysis, and unique activation‑mix experiments.
View on GitHub →SuvirRathore/micrograd-from-scratch