Feat: pairtab model pytorch

[anyangml]

The goal of this PR is to migrate the PairTabModel from TF to PT:

• Build a concrete AtomicModel class of PairTabModel based on this PR.
• Reimplement the C++ code (pair_tab.cc) in pytorch.
• Define an example input.json

Notes:

1. The fscale is not implemented in this method.

[wanghan-iapcm]

Could you please provide UT for the atomic model?

1. check the output energy is as expected. e.g. in the case you have a few atoms in the system you can manually calculate the energy and check if it is consistent with your code.
2. check if you code is jitable. see for example

   deepmd-pytorch/tests/test_se_e2_a.py

   Line 142 in 05c3f9d

   def test_jit(

[anyangml]

Could you please provide UT for the atomic model?

1. check the output energy is as expected. e.g. in the case you have a few atoms in the system you can manually calculate the energy and check if it is consistent with your code.
2. check if you code is jitable. see for example

   deepmd-pytorch/tests/test_se_e2_a.py

   Line 142 in 05c3f9d

   def test_jit(

optimized the for loop implementation, checked the calculation locally, going to figure out jit and unittests tomorrow.

[anyangml]

Could you please provide UT for the atomic model?

1. check the output energy is as expected. e.g. in the case you have a few atoms in the system you can manually calculate the energy and check if it is consistent with your code.
2. check if you code is jitable. see for example

   deepmd-pytorch/tests/test_se_e2_a.py

   Line 142 in 05c3f9d

   def test_jit(

I tested jit with the following code, no error msg

file_path = "test.txt"

model = PairTabModel(
            tab_file = file_path,
            rcut =  0.1,
            sel = 2
        )
scrp_md = torch.jit.script(model)
torch.jit.save(scrp_md,"test.pth",{})

[wanghan-iapcm]

What would be the behavior of your model

• if the r-range of the input table is smaller than the cut-off distance? is the interaction between the pairs beyond upper bound of the table and smaller than rcut zero?
• if the r-range of the input table is larger than the cut-off distance?

[wanghan-iapcm]

the key should be "energy", please check your output def.
you may want to use this decorator to ensure the correctness of your atomic model output

[anyangml]

Updated the key. Not sure about the decorator, this atomic model has no Fitting

[anyangml]

• range of the input table is smaller than the cut-off distance? is the interaction between the pairs beyond upper bound of the table and smaller than rcut zero?
• if the r-range of the input table is larger than the cut-of

@wanghan-iapcm

1. if r-range of the input table is smaller than the cut-off distance, there is not enough information to calculate the energy unless we want to do extrapolation. raise error
2. if the interaction between the pairs beyond upper bound of the table and smaller than rcut , the energy is not zero but we don't have information from the table to do the calculation. If we implement a check on 1, then this won't be a problem.
3. if the r-range of the input table is larger than the cut-off distance, I don't see a problem here. if the pairwise distance is beyond cutoff, it will be labeled as -1 in the nlist, and the energy for that pair is masked to 0.