Is there any way we could have complex-valued tensors in PyTorch… similar to dtype.complex in tensorflow…
If nit is there anyway I could implement it?
Hi,
Actually the work is in progress but there are some work arounds by third party libraries as you can find in this post:
Bests
Will check this out, thanks a lot.
Hi,
We are actually adding this right now.
Complex dtypes already exist (complex64 and complex128). Basic ops are working fine.
We have first draft of the autograd support with a limited set of functions right now.
Note that this will be in the upcoming 1.6 release.
Was there some specific features you were looking for?
Thank you for replying @albanD
Are they avaliable in the nightly builds ?
I was just looking to implement the paper deep complex networks, nothing specific though…
from there I thought would be cool to have complex backprop, gradients, optim etc etc…
Yes everything that will be in 1.6 is already in the nightly build.