https://docs.dgl.ai/generated/dgl.nn.pytorch.gt.GraphormerLayer.html#dgl.nn.pytorch.gt.GraphormerLayer
Here, the graphormer layer from this paper is https://arxiv.org/pdf/2106.05234 given. However, an example usage is not provided. It only takes node features as the input and without any knowledge of connections, somehow, it seems to run. Source code here seem to be problematic to me (dgl.nn.pytorch.gt.graphormer — DGL 1.1 documentation), with the same reason I mentioned above. Does anyone have experience using this layer?
The information of node connections is passed to the layer through attn_bias. Here lists some modules for computing attn_bias, including PathEncoder and SpatialEncoder in the Graphormer paper.
We will provide an example to demonstrate the usage of these modules later. Feel free to elaborate any other confusioin on these modules.