micrograd

micrograd is an exact reimplementation of micrograd by Andrej Karpathy, written for educational and experimental purposes.

This project features a minimalistic automatic differentiation engine and a tiny neural network module — all built from scratch and under 100 lines of code.

Features

• Reverse-mode autodiff (backpropagation)
• Scalar-based computation graph
• Tiny but expressive neural network (MLP)
• Easy to read and extend

Files

• train.py — sample training loop using the MLP class
• micrograd/ — contains the autodiff and neural net logic (copied exactly from micrograd)

Credits

This code is a faithful reproduction of micrograd by Andrej Karpathy.

License

Same license as the original micrograd — MIT.