OperatorLearning.jl icon indicating copy to clipboard operation
OperatorLearning.jl copied to clipboard

Published 20 hours ago • verified publisher icon SciML

Add tests on FNO + DeepONet forward pass, gradients, and training

Open ChrisRackauckas opened this issue 3 years ago • 3 comments

The tests don't seem to cover the use and training of the operators, just a few properties. It would be good to get a few integration tests.

ChrisRackauckas avatar Feb 10 '22 10:02 ChrisRackauckas

Do these tests need to run on gpu?

ba2tro avatar Feb 16 '22 13:02 ba2tro

Yeah I think that makes sense at least for a subset of those, we also track that in a separate issue (#27).

Flux.jl also contains CUDA in the tests and since this package also exports layers and is based on Flux quite a bit, I would more or less stick to this.

pzimbrod avatar Feb 16 '22 14:02 pzimbrod

I would start by making the tests on CPU though, then create equivalent tests on GPU. GPU will have a separate set of issues and it would be easier to catch/debug the CPU issues (usually correctness) first.

ChrisRackauckas avatar Feb 16 '22 14:02 ChrisRackauckas