The fourier.NuFFT layer is currently implemented assuming that uu and vv coordinates are known at initialization. This is fine for cases where the whole dataset is being predicted, either in a one-off, or repeatedly in some optimization loop.
To implement a stochastic gradient descent, however, we need a NuFFT layer whose .forward method also accepts (potentially new) uu and vv points, e.g. .forward(cube, uu, vv).
This suggests a reorganization of the NuFFT layer such that both use-cases can be addressed.
The
fourier.NuFFT
layer is currently implemented assuming that uu and vv coordinates are known at initialization. This is fine for cases where the whole dataset is being predicted, either in a one-off, or repeatedly in some optimization loop.To implement a stochastic gradient descent, however, we need a NuFFT layer whose
.forward
method also accepts (potentially new) uu and vv points, e.g..forward(cube, uu, vv)
.This suggests a reorganization of the NuFFT layer such that both use-cases can be addressed.