Deployment of the MACE Model Without Using ASE about mace HOT 5 CLOSED

ASE calculator is not a must as long as you provide a dictionary that is correct for calling model.forward. Take a look at
https://github.com/ACEsuit/mace/blob/main/mace/data/atomic_data.py and how it is used in testing scripts etc.

If you want to use it in C++, maybe you can look at how it is used in lammps.

from mace.

Isn't the neighbor list calculated in python, which would therefore make it impossible to run in pure C++ without some significant supporting code?

from mace.

Yes. I would imagine you will have to write an interface for it though not something easy. (please correct me if I am wrong.)

from mace.

Hey @CheukHinHoJerry,

Yes it is totally possible to run MACE in C++ using torchscript. This exactly how we made the lammps interface, see here: https://github.com/ACEsuit/lammps/blob/mace/src/ML-MACE/pair_mace.cpp#L308.
As @CheukHinHoJerry you will just need to provide the right input dictionary. To construct the NL you should be able to use the C++ matscipy NN.

from mace.

I guess you're right. I find a code in eval_configs.py, which use output = model(batch.to_dict(), compute_stress=args.compute_stress) to obtain the force and energy. Is that means what I should do is to provide the model with right dictionary, does it make sense?

Furthermore, I am not quite sure about nblist you guys mentioned, should I provide the model with nblist?

from mace.