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
nnmodules:Module,Linear,Sequential,MLP. - Optimizers: SGD, Adam, RMSProp with LR schedulers.
- Lightweight datasets and
DataLoaderfor experiments. - Validated via numerical gradient checking and exploding gradient diagnostics, confirming correctness against finite-difference approximations.
Links: GitHub