Open mcaroba opened 11 months ago
Maybe it's worth looking at what causes the tests to fail. It's possible there is some more numerical discrepancy which is fine - just commit the new benchmarks or adjust the tolerance. I can also take a look if needed.
There's something about pip and sphinx: https://github.com/libAtoms/GAP/actions/runs/6525579900/job/17718407833. I think it's the same issue with the tests as was already there with your earlier commit. I don't know where to even start looking at this, I'm not familiar with that part of the code at all.
Once you give me thumbs up, I will update the GAP submodule in QUIP/src to add the new changes to the soap_turbo string handling.
I have significantly improved (simplified) the use of soap_turbo descriptors. Now one can use implicit definition of hyperparameters, e.g.,
atom_sigma=scalar
automatically sets theatom_sigma_r
andatom_sigma_t
arrays. Also, some of the more obscure hyperparameters don't need to be defined, e.g.,atom_sigma_r/t_scaling
, which now are set to zero by default and for which one can also use a shortcutatom_sigma_scaling=scalar
. There is compatibility with hypers inherited from soap: n_max, cutoff, cutoff_transition_width. I will add a comprehensive list as soon as I can.I would appreciate if people can test the implementation and the use of
add_species=T
, which was also previously missing and I have added, especially those who were already using soap_turbo before.Finally, @albapa the automated tests are failing, but they were already failing for the previous commit, so I'm not sure if that has anything to do with my updates. I have done some comprehensive tests and the code definitely compiles (at least with gfortran), both gap_fit binary and the quippy parts. All the tests I did for one and two species fits trying all the combinations of string shortcuts I could think of also seem to work.