back-propagation issue about torch-vectorized HOT 4 CLOSED

Hello @zhan-xu , sorry for my late reply. I fixed the backprop for the vSymEig function by:

• Removing the inplace operation on the norm
• Adding an Epsilon value in the square root of the norm computation to avoid undefined gradients at 0.

I did an optimization test using an Adam optimizer and an L1 loss on the eigenvalues and eigenvectors and it looks good now:
`

from torch-vectorized.

Maybe it is not about the shape. Here is a bit more information. I tried the following code

import torch
from torchvectorized.utils import sym
from torchvectorized.vlinalg import vSymEig

# Random batch of volumetric 3x3 symmetric matrices of size 16x9x32x32x32
input = sym(torch.rand(16, 9, 32, 32, 32))
input.requires_grad=True

# Output eig_vals with size: 16x3x32x32x32 and eig_vecs with size 16,3,3,32,32,32
eig_vals, eig_vecs = vSymEig(input, eigen_vectors=True)

gt_vals = torch.randn((16, 3, 32, 32, 32))
gt_vecs = torch.randn((16, 3, 3, 32, 32, 32))

#loss = torch.nn.functional.mse_loss(eig_vals, gt_vals) # this line goes without error
loss = torch.nn.functional.mse_loss(eig_vecs, gt_vecs) # this line has error
loss.backward()

The problem seems happen when we add supervision to the eig_vecs.

from torch-vectorized.

The problem is perhaps related to https://github.com/banctilrobitaille/torch-vectorized/blob/master/torchvectorized/vlinalg.py#L55
When commenting out line 56, it works. btw why here do you replace 0 with 1? Shouldn't replace it with a small number like 1e-10?

from torch-vectorized.

As a walk-round, I change line 56 into

u0 = u0 / (norm+1e-10)
u1 = u1 / (norm+1e-10)
u2 = u2 / (norm+1e-10)

Seems working so far...

from torch-vectorized.