Class Incremental for Node Classification EWC and MAS cannot be reproduced

[WMX567]

Hi,

I ran the codes on A40 GPU, using the supplementary materials' hyperparameters.
EWC on CoraFull with 4.4±0.6 is far from 15.02 provided in the paper.

Thanks!

[QueuQ]

Hi,

I ran the codes on A40 GPU, using the supplementary materials' hyperparameters. EWC on CoraFull with 4.4±0.6 is far from 15.02 provided in the paper.

Thanks!

Hi,

I just run the experiments two times, and got the results AP: 16.03 AF: -80.11 ; AP: 17.28 AF: -78.92.

Please make sure that the memory strength is 100000 for ewc, backbone is GCN. The minibatch should be set as False, Using minibatch or not may cause difference in performance, which should have be explained in the paper or Appendix, Therefore, to fairly compare the methods, we adopt full batch training consistently for all methods on small NCGL datasets like CoraFull-CL. The ILmode shoud be set as classIL. Please also notice that --inter-task-edges should be False, this may also cause difference in the performance.

[SONGLEI-arch]

Hi,
I met the same problem. Could you please provide more detailed settings?