seg fault when predicting local uncertainty with trained sgp calculator

[aaronchen0316]

Describe the bug
I tried to predict the forces on a structure for testing purpose using an already trained sgp calculator/model. However, it threw a segmentation fault. I identified the seg fault happened at predict_local_uncertainties function in the sparse_gp.cpp.

My workflow is:

calculator, kernel = sgp_calculator.from_file('path-to-flare.json')
atoms = ASE.Atoms(species, positions, cells, pbc) # I also tried FLARE_Atoms.from_ase_atoms()
atoms.calc = calculator
atoms.calc.reset() # reset() because I saw this command frequently in otf.py; same seg fault without reset() 
force_pred = atoms.get_forces()

I tried to make a flare_pp.Structure object and predict_on_structure_descriptor() directly but it also threw a seg fault. Is the issue with "local" uncertainty estimate or there is a more efficient way to run a test set evaluation on existing sgp calculator/model? For example, should I supply a series of dft trajectories to OTF(trajectory) and use FakeMD to only predict properties

Thank you so much for your help!

NOTE: updated with better syntax

[YuuuXie]

Hi @aaronchen0316 , did you use this line to load the sgp model

calculator, kernels = sgp_calculator.from_file('path-to-flare.json')

Basically, make sure you have the "kernels" returned (even if not used) in the current work space.

If you just want to get the local uncertainties on a set of test structures, yes you can use FakeMD with a very large threshold and a warm start (by loading the SGP from file)

[aaronchen0316]

Hi @YuuuXie thank you for your suggestion as always!

Yes I used the correct line to load the sgp model: calculator, kernels = sgp_calculator.from_file() I corrected the syntax in the original issue.

Yes, I was able to get around with seg fault by using OTF() with FakeMD and either large min_steps_with_model or a large threshold.

Thank you for confirming that, but the seg fault persists. Is there a problem with my workflow or difficult to load a sgp calculator and predict on a few isolated atomic structures?

Thank you!

[YuuuXie]

Just realized I forgot to reply to this thread. The seg fault can happen in different places, we encountered in the kernel, the structure etc. which might need to debug in specific case

[Tinrry]

when i run otf_train.yaml，gnu compiler

this code cause segmentation fault, or run endless time.