This is an implementation of the GAT paper in tinygrad.
The ./app/train.py script uses the CORA dataset.
The model gets 80.39% accuracy on the test set (2/3rds of the dataset) after 1000 epochs.
Full training time was 51:11 minutes on a surface pro 9, CPU only.
Checkpoints for every 100 epochs are found in ./models/ along with the final
model.