I have the following strucuture for a Graph Neural Network (GNN), to construct the model, I use TensorFlow and sonnet. The model is used to do binary classification on particle physics collider events. This means for every collision event we construct a graph from the measured particles (with node values beeing energy, angles etc.). The network then predicts whether the event is a signal (Output 1) or a background (Output 0) event.
Since the encoder decoder strcuture is also used for transformer models, I was wondering, if it would be possible to implement a transformer for this task. Does anybody know if this was done before with tensorflow and sonnet or if there are any good resources that I could use? I would be very interested in implementing this and testing the performance for our appliations. I found this very nice tutorial but it is for translation tasks and I am really not sure what I need to change to be able to input graphs.