pytorch-GAT

Why we included this project

This repository reads like a well-organized tutorial rather than a bare library drop, which makes it a solid pick if you actually want to understand graph attention networks instead of just running one. It implements the original GAT paper in PyTorch and covers both classic setups: transductive learning on the Cora citation graph and inductive learning on the PPI protein network, so you can watch the same model behave differently across the two. The real selling point for learners is the playground script, which renders the Cora graph, projects the learned embeddings with t-SNE, and plots attention and entropy histograms, so the self-attention math becomes something you can see rather than trust on faith. Code comments and profiling utilities are threaded through the training scripts, which helps you trace where attention coefficients come from and why the model makes the decisions it does.