I’m trying GAT code following the tutorial(https://docs.dgl.ai/en/latest/tutorials/models/1_gnn/9_gat.html), but I always get 0.6~0.7 accuracy. It seems so much different from the original paper. I try tuning the number of heads and epochs but it doesn’t work. What should I do if I would reproduce the results?
We use tutorials to introduce the ideas of a model, but they may not be carefully tuned due to the limitation of tutorial deployment. For tuned implementation, you may refer to our example code of GAT. I just checked it and it should reach roughly the accuracies reported in the paper.