Accuracy of VN_DGCNN on the ModelNet40

[AkonLau]

Hi,
Thanks for your work and codes. However, when I reproduce your code with the VN_DGCNN model and default parameters, I just got 80.2% in the I/I case and 78.8% in the z/z case. The results are much less than the accuracy reported in the paper.
Are there any details I missed? Could you help us analyze the problem or give more details for discussion?
Thanks a lot!

[FlyingGiraffe]

Hi, thanks for the comment!

I was using the DGCNN code framework when generating the results in the paper, but later when publishing the code I merged the models into the PointNet framework for convenience. I quickly skimmed through the code and one reason might be: in DGCNN (and thus same for VN_DGCNN in the paper), the optimizer is SGD withlr=0.1 (see here), but the default optimizer here is Adam with lr=0.001.

I'll double-check with other settings. Thanks again for noticing this!

Best,
Congyue

[GostInShell]

Thanks for the wonderful code! With the above modification, I achieve the results 73.4% on SO3/SO3. Moreover, I find the training process rather unstable with large fluctuation, comparing to the normal DGCNN and PointNet. Is this usual? If possible, could you please share the training log?