Reproducing results on Teubingen

[TomerGalanti]

I ran GNN with default parameters on Tub dataset with 10 epochs for train and 10 for test and nb_max_runs=5. Got AUC 54 (in https://arxiv.org/pdf/1711.08936.pdf it is specified that I should get a higher score). Am I doing something wrong?

Note: running 1000 epochs is infeasible, since it already takes more than 5 hours to run it with 10 epochs.

Thanks!!

[diviyank]

There should be a performance issue ; it should be really fast for 10 epochs.
The accuracy issue here comes from the number of epochs, you need those at least 1000 epochs for the neural networks to have learned something.

Can you give me your hardware setup ?

[diviyank]

You should try with more epochs. I will be closing this issue for inactivity, don't hesitate to reopen-it.

[TomerGalanti]

Actually, we found a small bug that prevented the algorithm to run on the gpu.
I can send you the corrections if you want.

[diviyank]

Sure, I would be glad to fix this !

[TomerGalanti]

When you want to run on one gpu and not multiple gpus, since you use a multi gpu wrapper, it switched to cpu by default. Therefore, we replaced the wrapper to the default single gpu wrapper.
Hope it helps!