Improving symeig() function

autograd
1

Hi fellows,

I need to improve torch.symeig() in the following directions and like to get hints on where to start and how to do these:

  1. I need to make it more stable, like torch.svd(). I don’t know why svd function is more stable (maybe related to Issue #440).
  2. I need to implement a batch version of it on GPU. That means the input will be of size (B x D x D) where B is the batch size.
  3. I need to be able to implement top-k decomposition like scipy.linalg.eigh. For example, I should be compute only the top (dominant) eigenvalue and eigenvector.

Any hint about where to start is appreciated.

1 Like
2

Do you have examples of instability in symeig?

4

I don’t have a concrete example. I have a network implemented via both and the one using svd is much more stable. I will try to construct an example though.

Also, do you have an example of custom autograd function? Is there any guide available for this?

5

Please let me know when you do find a good example of unstable symeig!

Here’s an example of a custom autograd function: http://pytorch.org/docs/master/autograd.html#torch.autograd.Function

And here’s a guide for it:
http://pytorch.org/docs/master/notes/extending.html

6

@richard Thank you for the answers.
Regarding the batch version, is there a simple way to efficiently run a batch of symeig() on a single GPU? My matrices are relatively small.

7

I don’t think so, no. Please feel free to open a feature request at the github site for this.

1 Like
8

@richard I requested it at Issue #5354. Please edit it if it is not accurate.