Graphormer and Transformer vs Convolution in TransformerConv

[Nokimann]

In fact, TransformerConv is more like to be close to the convolutional type.

Unless the edge_index is the full connected graph with self-loops of all nodes (= all 1s in the adjacency matrix), we cannot call it graphormer or transformer as mostly seen in NPL or Vision.

What do you think about it?

[rusty1s]

Yes, we refer here to a SparseTransformer variant based on sparse-connections/an edge_index representation. As the underlying computation is the same, I think the name is still fitting. Wouldn't you agree?