SciML / OperatorLearning.jl

No need to train, he's a smooth operator
https://operatorlearning.sciml.ai/dev
MIT License
43 stars 8 forks source link

Use `fft` and `rfft` as fit #33

Open pzimbrod opened 2 years ago

pzimbrod commented 2 years ago

31 introduces Fourier Layer for higher dimensions. For now, rfft has been replaced by fft since it's simpler to implement.

However, rfft can save memory and computation time considerably (~2x), so it should be picked up again.

I think #28 plays a role here as well. If the input is complex, then rfft doesn't make sense. But since we're working with @generated anyway, maybe we can include a check where we can decide whether to use regular or real FFT depending on the input.