Jakob Unfried

typo in man/sectors.md

5

Stumbled upon this small typo while trying to understand your work

Typo

1

Hi Juraj, I was just reading your SVD code for reference on how people do the SVD backward pass. I found a typo at https://github.com/jurajHasik/peps-torch/blob/d56588c119392ba354de86318a2d7f9fdcbb3d07/linalg/svd_gesdd.py#L134 where you dont check `gu...

Gradient of SVD with degenerate singular values becomes NaN

11

If an SVD has degenerate singular values (multiple entries of S are exactly equal), the gradient pass will give `nan`. This is because the [AD formula](https://arxiv.org/pdf/1909.02659.pdf) contains something like `1...

``numpy.polyfit`` returns ``residuals.shape == (1,)`` in the case used here, where ``y`` is 1D