Gap between the performance reported and that of this repo on QM9.

[hengruizhang98]

Hi, we notice that there is a gap between the performance on the QM9 dataset reported and that we reproduced using your code. e.g. for HOMO(2), the reported RMSE is 0.0060, while with your code it's 0.1560.

[sunfanyunn]

I conducted those experiments for the paper in 2019 and the version for pytorch-geometric==1.3.0 at the time (even though I later tested the code in this repo with pytorch-geometric==1.6.0) Apparently the code and source of QM9 dataset have been changed. Refer to

• https://pytorch-geometric.readthedocs.io/en/1.3.0/_modules/torch_geometric/datasets/qm9.html
• https://pytorch-geometric.readthedocs.io/en/1.6.1/_modules/torch_geometric/datasets/qm9.html#QM9

[sunfanyunn]

consider the performance reported by this paper in 2019: https://rlgm.github.io/papers/31.pdf

[hengruizhang98]

Got it, thank you very much!