Martin Schubert
Martin Schubert
Yes, we could update one of the examples to include multiple source planes. I anticipate the code changes will be entirely in the usage of the fmmax API, and that...
It would be nice to sketch out a "workflow" that involves all of these steps, to help inform what is developed. Let's discuss during our next developer's meeting.
Hi @smartalecH, we have some logic which automatically does some broadcasting, here: https://github.com/facebookresearch/fmmax/blob/main/fmmax/fmm.py#L942 I'm not sure your summary above is exactly right, so let me just repeat some of it....
Ah, I could have been more clear. There is actually nothing special about `angular_frequency` here -- we equivalently could provide: - `angular_frequency` with shape `[5]` - `in_plane_wavevector` with shape `[4,...
Use "differentiable optimization trick" for backpropagation through tangent vector field calculation
I am seeing some issues with super-long compile times in the optimization context, which are eliminated when we use a `stop_gradient` before the vector field calculation. I am thinking we...
Use "differentiable optimization trick" for backpropagation through tangent vector field calculation
I am looking at it a bit. It might actually be relatively straightforward. Here's a reference that seems nice, it even includes Jax code: https://implicit-layers-tutorial.org/implicit_functions/
Use "differentiable optimization trick" for backpropagation through tangent vector field calculation
@smartalecH @Luochenghuang I have things working here---all it needed was a bit of regularization. https://github.com/mfschubert/mewtax
Use "differentiable optimization trick" for backpropagation through tangent vector field calculation
I think we may want to put this on hold for now: the potential accuracy improvement is small, and there is a speed penalty. - I added a test with...
I haven't tried that here, but seems like it would be a good step. By the way, would you be willing to share some of that profiling info? It would...
Yes, it would be good to have more documentation here. I do think everything is consistent and logical, but we can revisit this if we think this is not the...