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

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 using python main.py from an environment with the packages from requirements.txt.

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.