Unfair comparison with other models.

[hengruizhang98]

In eval.py, train/test split follows a 90% / 10% mannner instead of that of public split. While the baseline models(e.g. DGI) use public split for evaluation.

[Linyxus]

Hi, Hengrui. The metrics reported for baseline models were obtained under the identical protocols used for GRACE. Note that the results reported for DGI in our paper (e.g. 82.6 on Cora) is different from the ones in the original paper (e.g. 82.3 in Cora).