End-to-end newtwork prediction parameters

[ideas-man]

Hey @Fang-Haoshu !

Can you please post the values of hyperparams you used when training your network presented in the paper (and the checkpoint's epoch value if you used earlier stop to get the scores presented)? I am trying to reproduce your results here for my work, and right now I am no where near with these values:

numer of points: 20000
num of views: 300
max epoch: 90
batch size: 4
learning rate: 0.001
weight decay: 0
bn decay step every: 10
bn decay rate: 0.5
learning rate decay steps: 40, 60, 80
learning rate decay rates: 0.1, 0.1, 0.1

Overfitting occurs quite early (at around 20th-25th epoch), while the paper states that you were training for at least 100 epochs. Am I missing something?

The initial learning rate is 0.001 and the batch size is 4. The learning rate is decreased to 0.0001 after 60 epochs and then decreased to 0.00001 after 100 epochs.

Thanks in advance!

[Fang-Haoshu]

Hi， the full network training and testing is now available at: https://github.com/graspnet/graspnet-baseline
Please check it out there

Best regards

[Fang-Haoshu]

Yes, we also only train for 10 epochs