search

lightaime / deep_gcns_torch

Pytorch Repo for DeepGCNs (ICCV'2019 Oral, TPAMI'2021), DeeperGCN (arXiv'2020) and GNN1000(ICML'2021): https://www.deepgcns.org
MIT License
1.14k stars 155 forks source link

Model cannot converge problem #10

Closed im-wll closed 4 years ago

im-wll commented 4 years ago

Your work is very interesting. I use the network structure which introduced in your paper on the database of image classification. But I find that as the gcn network goes deeper, even use the Residual Connections of gcn , Convergence of networks get worse. when the depth of network is 2, the accuracy of train set can reach 80%, but when the depth of network grows to 7, the accuracy only drop to 50%. Is there also the problem of gradient disappearing?

lightaime commented 4 years ago

We haven't tried try to apply DeepGCNs to image yet. So not sure about the problem. But which is the gcn operator you used?

im-wll commented 4 years ago

I use the edgeConv,MRconv ,both of them have the problem

lightaime commented 4 years ago

You can try to combine CNN with GCN. I think using GCN solely for image classification is not a ideal choice.

im-wll commented 4 years ago

thank you for your advice. In fact, that's what I did. I first use the resnet18 to extract feature, then I input the feature to GCN net. But when I use your code to construct graph network, the network cann't convert even I train many epoches.

lightaime commented 4 years ago

Maybe you can try to Concat the resnet18 features at the end of your network.

On Wed, Nov 20, 2019 at 5:12 AM im-wll notifications@github.com wrote:

thank you for your advice. In fact, that's what I did. I first use the resnet18 to extract feature, then I input the feature to GCN net. But when I use your code to construct graph network, the network cann't convert even I train many epoches.

— You are receiving this because you commented.

Reply to this email directly, view it on GitHub https://github.com/lightaime/deep_gcns_torch/issues/10?email_source=notifications&email_token=AFUJTYCUTILCCYRUCXUTFIDQUSMITA5CNFSM4JOEE2P2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOEEQOD3Q#issuecomment-555803118, or unsubscribe https://github.com/notifications/unsubscribe-auth/AFUJTYCDLAQW6XPOFK7M7QLQUSMITANCNFSM4JOEE2PQ .