graphormer-pyg

Microsoft Graphormer rewritten in PyTorch-Geometric

!!! Dear Developers !!!

I currently do not have time to modify or work on this repository.
This implementation has one main disadvantage: IT's SLOW!
During the past few months I have recieved numerous issues concerning the execution speed.
If anyone wants to speedup my code or work on further development of this repository, you are MORE THAN WELCOME!
I will review all of your PRs and we can discuss any issues you are willing to discuss!

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