Weihua Hu
Weihua Hu
You need to tune the hyper-parameters. The default ones are one that is used in the paper. See the papers for the details.
I just suggest everyone stop using these old datasets! The Open Graph Benchmark (https://ogb.stanford.edu/) offers much better datasets, where GIN and more advanced models have been extensively benchmarked.
I did not record them :(
The hyper-parameters we tune for each dataset are: (1) the number of hidden units ∈ {16, 32} for bioinformatics graphs and 64 for social graphs; (2) the batch size ∈...
Hi! Thanks for your interest. GIN has been implemented by major libraries like Pytorch Geometric and DGL. The implementation is highly optimized and faster. I recommend using those libraries for...
We directly adopted the dataset files from https://github.com/muhanzhang/pytorch_DGCNN/tree/master/data
Hi, thanks for your interest. That's not surprising because we do not need expressive power in those node classification datasets. GIN is most useful when we really need expressive power...
Hi! Please use GIN and Cora in [PyG](https://github.com/pyg-team/pytorch_geometric).
Please use pytorch geometric or DGL for the implementation of GIN. They are much faster and easy to use :) PyG: https://pytorch-geometric.readthedocs.io/en/latest/modules/nn.html#torch_geometric.nn.conv.GINConv DGL: https://docs.dgl.ai/api/python/nn.pytorch.html#ginconv
Hi! Have you tried this https://github.com/weihua916/powerful-gnns/issues/13#issuecomment-824498186?