Graphormer
Graphormer copied to clipboard
About atoms and nodes
What does atom stand for? Isn't it the node? What are the factors that set its size? Very confusing
And doesn't grapphormer support input of node features? Thx!
What does atom stand for? Isn't it the node? What are the factors that set its size? Very confusing
Yes, Graphormer uses nodes to represent atoms. A vector with certain size is used to embedding a atom. In my opinion, the dimention of the vector depends on the complexity of the problem.
What does atom stand for? Isn't it the node? What are the factors that set its size? Very confusing
Yes, Graphormer uses nodes to represent atoms. A vector with certain size is used to embedding a atom. In my opinion, the dimention of the vector depends on the complexity of the problem.
If I intend to apply such a powerful model to a non-molecular domain, and I only use nodes instead of atoms (nodes contain features), how should I go about changing this parameter? Thx!
What does atom stand for? Isn't it the node? What are the factors that set its size? Very confusing
Yes, Graphormer uses nodes to represent atoms. A vector with certain size is used to embedding a atom. In my opinion, the dimention of the vector depends on the complexity of the problem.
What nodes have to do with atoms is particularly puzzling to me, thanks for the answer.
I think you just need to replace the embedding with your feature, then it should works. And this would requires some studies on the code implementation of Graphormer. (Note graphormer also takes the edges into consideration, see the following suggestion)
In addition, Graphormer is designed to combine Graph representation of molecule and the architecture of transformer. If you only want to apply the model to a non-molecular domain, I think it will be more beneficial to go to transformer directly.
Have you tested the 'replacing embeddings with features'? and are there any benchmarks for non-molecular problems using Graphormer?
Sorry, I haven't try any tests yet and I also don't know any benchmarks for non-molecular problems.