Peter Hawkins comments

Results 291 comments of


                                            Peter Hawkins

Support autodiff of Eigendecomposition with repeated eigenvalues

I think we should leave these kinds of issues open. It makes them more easily searchable should someone else have the same problem; I'd rather have one issue than two.

Support autodiff of Eigendecomposition with repeated eigenvalues

I also observe that TF has the same limitation: https://github.com/tensorflow/tensorflow/blob/f33aa592f92e233aeb00198d0caab80eaa89afe9/tensorflow/python/ops/linalg_grad.py#L314

Use a checked cast() in accessor cast operator.

Hmmm. This is definitely something I don't want to do accidentally, and currently it's very easy to do by mistake. If the cost of doing these checks is too high,...

Use a checked cast() in accessor cast operator.

Nothing. But I'm wondering if I can do better and get more type safety in optimized builds, also. If you require an explicit cast, then the user can either `static_cast`...

Add support for the POWER architecture

Thanks for doing this! I'll plumb it through in our project and see if it makes the end user happy.

Add support for the POWER architecture

Happily for my use case we only support ppc64le anyway, so this part doesn't block me personally.

Add #include <exception> for std::exception_ptr

Could we please merge this?

jnp.linalg.norm() underflows

Rasmus pointed me to https://hal.archives-ouvertes.fr/hal-01511120/document which looks interesting.

jnp.linalg.norm() underflows

Another possibility is to rescale the norm by the element with the largest absolute value. This would require a second reduction, which perhaps rules it out for performance reasons.

Support (nonsymmetric) np.linalg.eig on GPU

JAX just calls scipy's copy of LAPACK. You can probably accelerate it by installing e.g., Intel's MKL scipy. Torch, as far as I know, also just calls LAPACK. It may...