Results for ogbg-molpcba

[chaitjo]

Hi, thanks for open-sourcing the code and releasing extensive experiments! GNN-specific normalization beyond BatchNorm and LayerNorm was definitely a missing component in the toolbox.

Did you experiment with GraphNorm on ogbg-molpcba (the larger and more challenging mol dataset from OGB) as well as the other graph classification tasks? Would love to know your experience or informal results.

[lsj2408]

Thanks for watching our work!
Indeed, we are performing further experiments on the large-scale datasets as you mentioned (ogbg-molpcba and ogbg-ppa). Moreover, we will try more GNN architectures to better understand the power of GraphNorm. The new results will be updated as soon as possible.

[chaitjo]

Awesome, keep me in the loop; looking forward to it.

[chaitjo]

If I want to use GraphNorm as a plug-and-play replacement for PyTorch's BatchNorm layer with DGL, which code snippet do you suggest?

• https://github.com/lsj2408/GraphNorm/blob/master/GraphNorm_ws/ogbg_ws/src/dgl_model/norm.py
• https://github.com/lsj2408/GraphNorm/blob/master/GraphNorm_ws/gnn_ws/gnn_example/model/Norm/norm.py

[lsj2408]

Both are fine~ You could take the second one for your tasks. Just try them for free.

There is also a thrid-part implementation by Microsoft ptgnn:

• https://github.com/microsoft/ptgnn/blob/master/ptgnn/neuralmodels/gnn/messagepassing/graphnorm.py

[felipemello1]

@lsj2408, thanks for sharing this repo.

It seems that your code here is missing alpha: https://github.com/lsj2408/GraphNorm/blob/master/GraphNorm_ws/ogbg_ws/src/dgl_model/norm.py

Did I miss it?

[hugovergnes]

@lsj2408, thanks for sharing this repo.

It seems that your code here is missing alpha: https://github.com/lsj2408/GraphNorm/blob/master/GraphNorm_ws/ogbg_ws/src/dgl_model/norm.py

Did I miss it?

I think so too, looks like in the microsoft implementation above they added