Loss does not decrease?

[tangbohu]

In my experiments, the loss remains to be around 0.8-1.1 even after 300k iterations. Is it right?

[fabiopoiesi]

Yes, it is possible. Do you have some tensorboard plots? If you are interested, we are now using different training parameters to learn faster.

[tangbohu]

Thanks @fabiopoiesi ! I am very interested in this work. I have visualized the tensorboard, it seems OK. But if you have different training parameters, would you mind sharing with us? Besides, could you share us the test codes to get the results reported in table 1 and 2 of the paper.

Thanks again for your kind reply!

[fabiopoiesi]

Some tips:

• do not train on multiple gpus (for some reason the performance decreases), so avoid using nn.DataParallel
• train for 10 epochs using:

  optimizer = optim.SGD(net.parameters(), lr=1e-1, momentum=0.9, weight_decay=5e-5, nesterov=False)
  scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=3, gamma=0.1)

  you might achieve the same performance with this parameters, but a little quicker.

For the second request, I'll get back to you soon.

[tangbohu]

Thanks! I am looking forward to seeing the evalution codes.

[tangbohu]

Dear @fabiopoiesi , I wonder whether the test codes are available now?

[fabiopoiesi]

I haven't had time yet. But the code for Tab. 1 is already there, have you checked the file benchmark_3dmatch_pre.py? For Tab. 2, I have to prepare it.

[tangbohu]

Thanks. I will try it first.

[fabiopoiesi]

Have you had chance to try it?

[wenli4313]

In my experiments, the loss remains to be around 0.8-1.1 even after 300k iterations. Is it right?

Hello，How do you get the final_chkpt.pth?Thank you very much.