Final project for the Deep Learning Systems course (10-714) at CMU, Fall 2022.
We extend the needle library by adding sparse matrices and related operations and autodiff. It supports both CPU and CUDA backends. On top of that, we build the Graph Convolutional Network and the Graph Attention Network. We train and compare these two networks and a baseline fully-connected network on the Cora dataset, and show that graph networks can achieve higher accuracy without too much performance cost.
Source code is available in MashPlant/dlsys-proj Please request access to MashPlant (Chenhao Li)