For a quick demo run best_model_for('Squirrel') or best_model_for('Chameleon') from best_model.py
The model's code is in EFGNN.py
MultiHopTransform.py just augments the data with some precomputed stuff, like CSR adjacencies, etc.
train.py contains a train function
AttLayer adds a learnable attention-based filter, but it doesn't seem to bring any performance benefits at this point.