qm9_nn_conv.py yields very large and inconsistent results

Describe the bug

Hi, I am trying to train GNNs for the QM9 dataset different targets using the qm9_nn_conv.py script in the examples folder, however, the training itself is very unstable and results does not match existing papers. For example, on target 1:

after few epochs, the test MAE grows very large, and in general the test MAE and validation MAE are not close to what the model should achieve. This happens across several tasks and I could not find why. It happens on the exacts script from the github repo with the exact hyperparameters. I would be glad for any help, thanks in advanc!

Versions