PyTurk (MiniTorch)

Minimal PyTorch-style autograd engine and neural network framework from scratch for education.

PyTurk is a minimal educational deep-learning framework inspired by micrograd and PyTorch.

Features

  • Simple scalar autograd (Value) with dynamic computation graphs and reverse-mode backpropagation through arbitrary DAG-structured forward passes.
  • PyTorch-like nn modules: Module, Linear, Sequential, MLP.
  • Optimizers: SGD, Adam, RMSProp with LR schedulers.
  • Lightweight datasets and DataLoader for experiments.
  • Validated via numerical gradient checking and exploding gradient diagnostics, confirming correctness against finite-difference approximations.

Links: GitHub