NWGraph

Graph library built on top of Pytorch to help the creation of GNNs.

Implementation based on (working draft) semantics defined at: https://www.overleaf.com/read/vfbqdgxtxnws

It does not define any trainer code (so no lightning), just graph semantics, edges, nodes and message passing. The training code is left to be done on a project by project basis.

Examples

• See ngclib trainer code, for example where they define a sequential way of training each edge independently. Upon training, the entire graph is loaded into memory to produce pseudo-labels, followed by a semi-supervised iteration. LME is used here for training.

• See mnist-ensemble-graph for a simple example where we train 5 edges in the same time. Each edge starts from a RGB image. Simple pytorch-lightning Trainer code is used here.