edugrad 0.1.0

Minimal dynamic computation graph with backpropagation, for pedagogical purposes

edugrad

This is a library intended for pedagogical purposes illustrating a very minimal implementation of dynamic computational graphs with reverse-mode differentiation (backpropagation) for computing gradients. Three guidelines motivate design choices made in the implementation:

• Mimicking PyTorch's API as closely as possible.
• Simple forward/backward for operations (operating on numpy arrays).
• Dynamic computation graphs, built as operations are run.

The library has been inspired by several other similar projects. Specific acknowledgments are in the source where appropriate.

• micrograd by Karpathy
• autodidact: a pedagogical implementation of autograd
• joelnet

Installation

To simply use edugrad, you can pip install edugrad or uv add edugrad.

For an editable installation, clone the repo and run uv sync from the root directory.

Usage

In examples/toy_half_sum, you will find a basic use case. main.py exhibits a basic use case of defining a feed-forward neural network (multi-layer perceptron) to learn a basic function (in this case, y = sum(x)/2 where x is a binary vector). You can run it by running uv run python -m examples.toy_half_sum.main from the main directory of this repo.

Basics

There are a few important data structures:

• Tensor: this is a wrapper around a numpy array (stored in .value), which corresponds to a node in a computation graph, storing information like its parents (if any) and a backward method.
• Operator: an operator implements the forward/backward API and operates directly on numpy arrays. A decorator @tensor_op converts an Operator into a method that can be directly called on Tensor arguments, which will build the graph dynamically.
• nn.Module: as in PyTorch, these are wrappers for graphs that keep track of parameters, sub-modules, etc.