NeuralPDE.jl
NeuralPDE.jl copied to clipboard
Inverse Dirichlet Adaptive Loss
Adds inverse dirichlet adaptive loss function to address #500
The biggest thing is that this needs tests.
This seems to work ok for the 2D Poisson equation but it does not outperform the nonadaptive loss across several seeds.
Try the incompressible Navier-Stokes from the paper. I think this may only make a difference when it's "sufficiently hard", and Poisson is simple enough that non-adaptive is fine.
Some of the review comments haven't been addressed, but other than that I think this is pretty close to merging. Using the Poisson 2D as the test is fine for CI, but we should test it separately on something harder. That separating testing can then turn into a tutorial.
I've started working on a tutorial/test for the incompressible Navier-Stokes and I'll add the tutorial to docs/src/pinn. Should I add the test to adaptive_loss_tests.jl too?
Should I add the test to adaptive_loss_tests.jl too?
If it's not too long. That might get constrained by compute time when used in the test infrastructure.
This looks really good but I don't have time until my project deadline on Tuesday to review further.
This should get rebased due to changes in https://github.com/SciML/NeuralPDE.jl/pull/553. That would hopefully make it much cleaner too.