search

flaport / fdtd

A 3D electromagnetic FDTD simulator written in Python with optional GPU support
https://fdtd.readthedocs.io
MIT License
454 stars 116 forks source link

is the pytorch backend differentiable? #45

Open ni-chen opened 2 years ago

ni-chen commented 2 years ago

Hi,

Thanks for making this amazing package, this could be useful. May I ask if the PyTorch backend is differentiable?

Thanks.

flaport commented 2 years ago

Hey @ni-chen ,

this has been on my todo list for too long, currently there are a few in-place operations preventing this... However, I don't think it should be too difficult to change, I expect the simulator to work slower, however...

Maybe when I find the time I might finally give it a shot soon. This would obviously be an awesome feature to have.

ni-chen commented 2 years ago

Thanks for your reply. I believe it would be of great interest for many applications. @flaport

simenhu commented 9 months ago

Would you be able to point to the places where the changes would have to be made? Could see if I could try to make a PR on it if it's not a very big change?

Would the general approach of preallocating the field tensors and write to a new index instead of updating an already populated tensor be the general solution you are thinking of?