matrix not invertible

[yossilevii100]

Hi, I'm trying to reproduce your code.
Running python3 modelnetdataset.py on downloaded modelnet40_normal_resampled dataset yields the following error:

torch._C._LinAlgError: linalg.inv: The diagonal element 1 is zero, the inversion could not be completed because the input matrix is singular.

Any clue how to solve it?
Thanks a lot!

Here is the error log:

  File "modelnetdataset.py", line 105, in __getitem__
    blurs=plane_blur(pc[:3].unsqueeze(0),mask=None,single=False)
  File "/home/me.levi/igos/blur_utils.py", line 317, in plane_blur
    pc=plane_smoothing(pc,k=k,r=r,d=0,lamb=lamb,miu=miu,erosion=True,subset=idx)
  File "/home/me.levi/igos/blur_utils.py", line 285, in plane_smoothing
    x=project_to_line(nn_pos,x,normal,plane,erosion,lamb,miu)
  File "/home/me.levi/igos/blur_utils.py", line 228, in project_to_line
    uv_inverse=torch.inverse(uv) # 2 x 2
torch._C._LinAlgError: linalg.inv: The diagonal element 1 is zero, the inversion could not be completed because the input matrix is singular.

[arthurhero]

Hi. You can print and see whether the uv vectors are colinear. If they are, probabaly need to add some random noise.

[yossilevii100]

Thanks for the reply,
I can see that uv vectors are colinear e.g.:

tensor([[-0.9804,  0.1970],
            [0.1970,  0.9804]])

I run it on ModelNet 40, following your instructions.
I want to be precise with the implementation, how exactly should it be fixed?

[bunnyveil]

I seem to have solved the problem, please change this line bellow(location: blur_utils.py line 228)
uv_inverse = torch.inverse(uv) # 2 x 2
to:
if torch.linalg.matrix_rank(uv) < 2:
# Add random noise to break the collinearity
uv += torch.randn_like(uv) * 1e-3
uv_inverse = torch.inverse(uv) # 2 x 2
you can switch the random noise to a more properly value.