search

autoatml / autoplex

Code for automated fitting of machine learned interatomic potentials.
https://autoatml.github.io/autoplex/
GNU General Public License v3.0
47 stars 8 forks source link

mock ML fit? #64

Open QuantumChemist opened 5 months ago

QuantumChemist commented 5 months ago

This is an idea that is something that can be implemented in the long run, but I think it would be a good idea to implement a mock ML fit (similar to the mock VASP we use in the unit tests), because even the fits for the test data take a lot of RAM and also some time. What do you think?

MorrowChem commented 5 months ago

How about running a very cheap 2body fit with gap_fit? Only takes a second or two and very little RAM

QuantumChemist commented 5 months ago

How about running a very cheap 2body fit with gap_fit? Only takes a second or two and very little RAM

Good idea! I could at least do that for the unit tests that don't check SOAP settings.

naik-aakash commented 5 months ago

This is an idea that is something that can be implemented in the long run, but I think it would be a good idea to implement a mock ML fit (similar to the mock VASP we use in the unit tests), because even the fits for the test data take a lot of RAM and also some time. What do you think?

Thinking about it now a bit more, I think it would be good to add this, especially for fit routines that have non-Python dependencies (buildcell, Julia, GPUMD (Needed for NEP potential interface I wanted to add in future)).

This will also simplify the CI workflow and keep test suit run times less.

But before we do this, we need to ensure all the routines work with actual programs locally.

JaGeo commented 5 months ago

@naik-aakash sure. We can do this more complex workflows. However, i would like to have small actual fitting tests there as well. Just to make sure the fitting codes keep working.

JaGeo commented 5 months ago

I think we also need a mock run. At least for atomate2, I will likely need this.

YuanbinLiu commented 4 months ago

Fitting a potential is very fast if the training set is small and the model is not complex. So if writing a mock_fit is time-consuming, we can temporarily use the current unit test method. But agree we could do this later.