search

allenhaozhu / SSGC

Implementation for Simple Spectral Graph Convolution in ICLR 2021
Other
80 stars 17 forks source link

The code's learning rate hyperparameters seems inconsistent with the paper. #18

Open GooLiang opened 1 year ago

GooLiang commented 1 year ago

Hello, the learning rate of cora, citeseer, and pubmed is 0.02 in the paper. But it seems to be 0.2 in the implementation. Which one is correct? In addition, can you provide the code of reddit, or the learning rate of reddit.

allenhaozhu commented 1 year ago

It is! I'm sorry for not getting back to you sooner. I think you are right. As I recall, just keep the same hyperparameter as the SGC. You can have a try. Basically, the hyperparameters of SGC and SSGC are very similar.

liangyx @.***> 于2023年9月9日周六 00:35写道：

Hello, the learning rate of cora, citeseer, and pubmed is 0.02 in the paper. But it seems to be 0.2 in the implementation. Which one is correct? In addition, can you provide the code of reddit, or the learning rate of reddit.

— Reply to this email directly, view it on GitHub https://github.com/allenhaozhu/SSGC/issues/18, or unsubscribe https://github.com/notifications/unsubscribe-auth/AMPJZOEEMCOZHTV3E7T3NZLXZMUMXANCNFSM6AAAAAA4QPOMLA . You are receiving this because you are subscribed to this thread.Message ID: @.***>

GooLiang commented 11 months ago

Thank you for your reply. On the arxiv code, you provided two versions MLP.py and SSGC-MLP.py. From the implementation point of view, is MLP.py the actual SSGC+MLP of the paper? BTW, can you provide the reproducible hyperparameters of SSGC+MLP on the arxiv dataset？

allenhaozhu commented 11 months ago

Sorry, I have moved from Canberra to Sydney, and the record was missing. I remember that I kept the graph convolution result first, then trained on it with MLP. I did not use the specific parameter for the original raw features. Thus, it is an average pooling for different hops features. Then, I only tried different hyper-parameters on MLP.

liangyx @.***> 于2023年10月4日周三 03:18写道：

Thank you for your reply. On the arxiv code, you provided two versions MLP.py and SSGC-MLP.py. From the implementation point of view, is MLP.py the actual SSGC+MLP of the paper? BTW, can you provide the reproducible hyperparameters of SSGC+MLP on the arxiv dataset？

— Reply to this email directly, view it on GitHub https://github.com/allenhaozhu/SSGC/issues/18#issuecomment-1745310853, or unsubscribe https://github.com/notifications/unsubscribe-auth/AMPJZODS3RL3WHW4J4BS3HTX5Q3ERAVCNFSM6AAAAAA4QPOMLCVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTONBVGMYTAOBVGM . You are receiving this because you commented.Message ID: @.***>