AssertionError: assert torch.all(torch.det(R)>0)

[gowda-95]

Hi again,
I am getting the AssertionError: assert torch.all(torch.det(R)>0) (reference image above). This was while training for the 3 different configurations in the cfgs folder and for the provided ModelNet datasets.

Removing the assert statement and training it might not give the desired results. Could you please advice what can be done in this case?

Also is any data pre-processing required?

Many thanks in advance.

[paul007pl]

Thank you for asking. It is great that you can train the model.
As for this reported error, I am very confused.
Because my code should have checked det(R) to make sure that all det(R) >0.
Otherwise, the negative det(R) should be manually corrected. Please check

GMCNet/src/models/gmcnet.py

Line 193 in ebe3005

V_neg[:, :, 2] *= -1

This is a common operation, and you can also check
https://github.com/yewzijian/RPMNet/blob/74f6531968bc084e44231e397da640432cd42a4b/src/models/rpmnet.py#L137

Sometimes, you may get the svd numerical errors (caused by PyTorch), but I do not know why the reported error could arise...

As for data preparation, please follow:
https://github.com/paul007pl/GMCNet#data
No additional pre-processing operations are required.