Does PyTorch support complex data-type?

Poornima_Jain · February 12, 2020, 10:12am

Does PyTorch support complex valued variables/tensors yet?
If not, why is it not included in PyTorch?
Thanks for any comments on this.

ra_schoonhoven · February 12, 2020, 3:33pm

I am definitely not an expert but it seems people are working on it:

Poornima_Jain · February 12, 2020, 3:38pm

Yeah, I saw this post but it seems to be a bit old (2017). Complex data-type being a basic and simple feature, I was wondering why it is not being added with priority.

ra_schoonhoven · February 12, 2020, 3:40pm

Yes that is true but the latest updates are from a few days ago so it seems active. As to why it takes a lot of time or how its prioritized I don’t know.

Poornima_Jain · February 12, 2020, 3:43pm

Okay, I didn’t notice the recent updates. It means the feature is in pipeline at least. Thanks!

albanD · February 12, 2020, 4:38pm

Hi,

Yes, this is work in progress for this half

atithi8 · August 17, 2020, 6:41pm

What is the update on this project. Is torch_complex ready to be used? If yes, will it be compatible with Autograd for Complex Numbers?

albanD · August 17, 2020, 7:08pm

Hi,

This is still work in progress.
For gradients, there are still ongoing discussion on which convention we want to use: https://github.com/pytorch/pytorch/issues/41857

Poornima_Jain · September 11, 2020, 8:36am

Hi @albanD, thanks for the reference.
Could you please through some light on how soon complex backpropagation could be integrated into Pytorch. Actually I am working on implementing complex backpropagation using Wirtinger derivatives for my thesis work by overriding backward functions, using Pytorch hooks to modify gradients etc. As I am working single-handedly on this and the framework itself is just a prerequisite for the further experimentation on complex-valued data, I was thinking if Pytorch is going to integrate this framework in the near future, it would not be worth my single-handed efforts and time to create it from scratch.
Could you please advise me with this?

anjali411 · September 17, 2020, 3:34pm

Hi @Poornima_Jain thanks for reaching out! We are in the process of updating autograd formulas to work properly for complex (example PR: https://github.com/pytorch/pytorch/pull/43208) We are hoping to have complex autograd in beta in 1.8 release. You can check https://github.com/pytorch/pytorch/issues/33152 to see what’s being worked on for complex and add comments if you specifically need backward support for certain operators. Thank you!