graphormer-pyg

Microsoft Graphormer rewritten in PyTorch-Geometric

Implemented Layers

1. Centrality Encoding
2. Spatial Encoding
3. Edge Encoding
4. Multi-Head Self-Attention

Warning

This implementation differs from the original implementation in the paper in following ways:

1. No [VNode] ([CLS] token analogue in BERT)
2. The shortest path algorithm is not mentioned in the paper. This repository uses Floyd-Warshall