GroundUpAI

A deep learning library built from basic operations to complex architectures


About

GroundUpAI is a simple to use and pythonic deep learning library. It is entirely implemented from scratch (even backpropagation) without the help from other libraries/frameworks with the exception of torch.Tensor due to the subpar FLOPS of plain python, hence the name "ground up".

Key Features

Key features of GroundUpAI include simplicity in making custom callbacks, generic optimizer, minimized boilerplate. If you are familiar with the implementation of PyTorch and FastAI, you would also likely spot numerous places where some great OOP practices are inspired by their source code.

Quick Start

Steps to get started

Demo

A simple demo to show how a few major building blocks work together

Building Blocks

Code documentation from basic math operations to generic optimizer in chronological file order

Advanced Architectures

Time to build the big names, exciting! Featuring ResNet, Transformers, and BERT

Other Info

Miscellaneous info

Acknowledgements

Credits and Academic papers that helped along the way