/educational-autograd

Primary LanguageJupyter Notebook

A Tiny tensor library + an autograd engine

  • To run examples:
    PYTHONPATH="." python3 ./examples/or_gate.py
  • For MNIST: check ./mnist.ipynb

TODO:

  • A functional tensor library
  • Passing Backward passes with tests
  • MLP working
  • Basic CNN
  • CIFAR-10/MNIST working with the lib (MNIST working)
  • Add a different backend than numpy
    • Current plan is to convert the autograd graph into a set of basic ops and compile that to C/Metal/CUDA(not decided yet) kernels
    • Yes this is inspired from tinygrad

Warning

Solely for teaching myself how pytorch works