Performance of 8-layer GCN #8

Activity

1. DropEdge commented on Apr 15, 2020

   

   DropEdge

   on Apr 15, 2020

   Last edited by DropEdge

   Owner

   More actions

   Hi ChandlerBang,
   As mentioned in #3 , To exactly reproduce the results, you need to satisfy the requirements listed in requirements.txt as well as python version.
2. ChandlerBang commented on Apr 15, 2020

   

   ChandlerBang

   on Apr 15, 2020

   Author

   More actions

   Thanks for your instant reply!
   I noticed Issue #3 before. I created a new virtual environment and installed all the required packages as suggested in README. So the environment should be the same as yours. The thing is, I can reproduce the results if I use the default hyper-parameters in scripts, but I meet problems if I change the number of layers.
3. XifengZhang commented on Apr 17, 2020

   

   XifengZhang

   on Apr 17, 2020

   More actions

   Have you fixed the problem? I've followed the exact same envrionment but I cannot reproduce the performance of 8, 16 or higher layer GCN (without DropEdge), either. They perform very badly on Cora in the semi-supervised setting.
4. txzhou-hust commented on Aug 20, 2020

   

   txzhou-hust

   on Aug 20, 2020

   More actions

   I have the same problem as you，It should be that the parameters of the multi-layer GCN are not given, and you cannot simply set --nbaseblocklayer 0 to --nbaseblocklayer 6 in script/semi-supervised/cora_GCN.sh

   👍React with 👍1Reacted by Wei Jin
5. wdzhong commented on Sep 28, 2020

   

   wdzhong

   on Sep 28, 2020

   More actions

   From the Table 6 of the Appendix, it seems that you need to randomly search the best parameters for different number of layers. Quote: "For each model, we try 200 different hyper-parameter combinations via random search and select the
   best test accuracy as the result." end quote. This is a pretty tricky and time-consuming process. Does this mean that you can get better results if you try more combinations?
6. DongXiang-CV commented on Dec 8, 2020

   

   DongXiang-CV

   on Dec 8, 2020

   More actions

   I also have the same problem. Have you guys fixed the problem?
7. txzhou-hust commented on Dec 13, 2020

   

   txzhou-hust

   on Dec 13, 2020 via email

   More actions

   You can try to modify these parameters，such as weight_decay, sampling_percent and dropout. For multi-layers model with no parameters provided, I get good results by tuning these parameters，but there is a small gap compared with the results of the paper

   …

   ------------------&nbsp;原始邮件&nbsp;------------------ 发件人: "DropEdge/DropEdge" <[email protected]&gt;; 发送时间:&nbsp;2020年12月8日(星期二) 晚上6:47 收件人:&nbsp;"DropEdge/DropEdge"<[email protected]&gt;; 抄送:&nbsp;"流觞曲水"<[email protected]&gt;;"Comment"<[email protected]&gt;; 主题:&nbsp;Re: [DropEdge/DropEdge] Performance of 8-layer GCN (#8) I also have the same problem. Have you guys fixed the problem? — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.
8. klchai commented on Dec 23, 2020

   

   klchai

   on Dec 23, 2020

   More actions

   You can try to modify these parameters，such as weight_decay, sampling_percent and dropout. For multi-layers model with no parameters provided, I get good results by tuning these parameters，but there is a small gap compared with the results of the paper
   …
   ------------------ 原始邮件 ------------------ 发件人: "DropEdge/DropEdge" <[email protected]>; 发送时间: 2020年12月8日(星期二) 晚上6:47 收件人: "DropEdge/DropEdge"<[email protected]>; 抄送: "流觞曲水"<[email protected]>;"Comment"<[email protected]>; 主题: Re: [DropEdge/DropEdge] Performance of 8-layer GCN (#8) I also have the same problem. Have you guys fixed the problem? — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

   What are your parameters? When I try to reproduce the multi-layer GCN, the nhiddenlayer is fixed to 1 and I can only modify nbaseblocklayer. Thanks in advance.