Minimal Deep Learning Framework

A PyTorch-like framework with define-by-run autograd and CuPy-based GPU support.

I built this framework to understand the mechanics of deep learning systems from first principles.

Highlights

  • Implemented a dynamic computation graph and automatic differentiation engine from scratch.
  • Followed a define-by-run execution model to mirror modern deep learning frameworks.
  • Added GPU support through CuPy to keep the framework lightweight but practical.