Deep Learning from Scratch, the NumPy Way (sql-hkr.github.io)

Syna is a minimalist, educational deep learning framework built from scratch on NumPy and modeled after DeZero. It implements a define-by-run (dynamic computation graph) execution model and even bundles a basic reinforcement learning toolkit — not as a separate package but as part of the same library. The project is aimed at beginners and researchers who want to learn how modern frameworks like PyTorch work under the hood, prioritizing clarity and teachability over runtime performance or production features. Technically, Syna exposes the core mechanics of autograd and dynamic graphs using plain NumPy arrays, making array operations and gradient flow explicit and easy to inspect. GPU support is intentionally omitted to keep the codebase compact and readable; this makes Syna ill-suited for large-scale training but ideal for debugging, prototyping algorithms, and pedagogy. The repo is available on GitHub and as a PyPI package, and it can visualize computation graphs via Graphviz. Licensed under MIT, Syna is a handy tool for anyone who wants a transparent, CPU-only reference implementation of deep learning primitives and a simple integrated RL environment to experiment with algorithmic ideas.