The error : self.dataset_info == net_out_info_o.dataset_info means?

[newplay]

Dear developer,

I am attempting to replicate the 'Monolayer-MoS2' demo, and during the evaluation process, I encountered an error with the message self.dataset_info == net_out_info_o.dataset_info. The complete error traceback is as follows:

Traceback (most recent call last):
  File "/home/zjlin/DeepH-E3/deephe3-eval.py", line 29, in <module>
    kernel.eval(args.config, debug=args.debug)
  File "/home/zjlin/DeepH-E3/deephe3/kernel.py", line 373, in eval
    net: Net = self.load_model(os.path.join(model_path, 'src'), device=eval_config.device)
  File "/home/zjlin/DeepH-E3/deephe3/kernel.py", line 559, in load_model
    assert self.dataset_info == net_out_info_o.dataset_info
AssertionError

Upon inspecting the source code in kernel.py, I noticed that the information comes from the model_path/src/ directory. The content of the dataset_info.json file is as follows:

{"spinful": false, "index_to_Z": [16, 42], "orbital_types": [[0, 0, 1, 1, 2], [0, 0, 0, 1, 1, 2, 2]]}

The target_blocks.json is assumed to be correct, so I haven't provided the details.

I identified the lines where the error originates:

    def load_model(self, src_path, device='cpu'):
        assert os.path.isdir(src_path)
        sys.path.append(src_path)

        # load dataset_info and net_out_info
        net_out_info_o = NetOutInfo.from_json(src_path)
        if self.dataset_info is None:
            self.dataset_info = net_out_info_o.dataset_info
        else:
            assert self.dataset_info == net_out_info_o.dataset_info
        if self.net_out_info is None:
            self.net_out_info = net_out_info_o
        else:
            assert self.net_out_info == net_out_info_o

So, does this imply that the dataset_info from the model I trained is same as the dataset I intend to evaluate?

[Xiaoxun-Gong]

Yes, these lines in the code are to ensure that the user isn't trying to evaluate the model on materials that are not compatible with it. This is because the neural network model depends on specific details of the material, including atom types, "spinful", and angular momenta of atomic orbitals. Regarding your problem, what material are you trying to evaluate your model on? Please check these three properties and make sure they are the same as those specified in the dataset_info.json under model_path/src/.

[newplay]

I used the Monolayer_MoS2_Demo from your database to compare Openmx and DeepH-E3. When I ran a simple MoS2 example(unit cell), I encountered this error. However, when I ran the example for a 2x2 supercell of MoS2, it ran successfully. I initially thought you wanted to avoid potential issues arising from the accuracy of results due to pre-training in the training data (such as the correct law being overemphasized because it has been trained), but it seems I was mistaken.

[Xiaoxun-Gong]

Can you let me see the input file to the OpenMX calculation of the unit cell?

[newplay]

Sorry for my delayed response. Here is the content of the 'openmx_in.dat' file:

##################################################
System.Name                       MoS2
DATA.PATH                         /home/zjlin/openmx3.9/DFT_DATA19
HS.fileout                        on

Species.Number                    2
<Definition.of.Atomic.Species
  Mo  Mo7.0-s3p2d2 Mo_PBE19
  S  S7.0-s2p2d1 S_PBE19
Definition.of.Atomic.Species>
Atoms.Number                      3
Atoms.SpeciesAndCoordinates.Unit  FRAC
<Atoms.SpeciesAndCoordinates
   1  Mo      0.3333329856   0.6666669846   0.3719750881           7.0    7.0    0.0  0.0  0.0  0.0  0
   2  S       0.6666669848   0.3333329858   0.4502193077           3.0    3.0    0.0  0.0  0.0  0.0  0
   3  S       0.6666669848   0.3333329858   0.2937308686           3.0    3.0    0.0  0.0  0.0  0.0  0
Atoms.SpeciesAndCoordinates>

Atoms.UnitVectors.Unit            Ang
<Atoms.UnitVectors
    3.1903159618000001    0.0000000000000000    0.0000000000000000
   -1.5951581104000001    2.7628940436999998    0.0000000000000000
    0.0000000000000000    0.0000000000000000   20.0000000000000000
Atoms.UnitVectors>

scf.XcType                        GGA-PBE   # LDA/LSDA-CA/LSDA-PW/GGA-PBE
scf.ElectronicTemperature         0.0     # default=300 (K) SIGMA in VASP
scf.energycutoff                  300       # default=150 (Ry = 13.6eV)
scf.maxIter                       2000
scf.EigenvalueSolver              Band      # DC/DC-LNO/Krylov/ON2/Cluster/Band
scf.Kgrid                         6  6  1
scf.criterion                     4e-08     # (Hartree = 27.21eV)
scf.partialCoreCorrection         on

scf.SpinPolarization              off
scf.SpinOrbit.Coupling            off

scf.Mixing.Type                   RMM-DIISK
scf.Init.Mixing.Weight            0.3
scf.Mixing.History                30
scf.Mixing.StartPulay             6
scf.Mixing.EveryPulay             1

1DFFT.EnergyCutoff                3600
1DFFT.NumGridK                    900
1DFFT.NumGridR                    900

scf.ProExpn.VNA                   off

MD.Type                           Nomd      # Nomd (SCF) / NVT_NH (MD)
Band.dispersion                 on
Band.Nkpath                     3
<Band.kpath
30 0.000 0.000 0.000 0.500 0.000 0.000 \Gamma M
30 0.500 0.000 0.000 0.333333333 0.333333333 0.000 \M K
30 0.333333333 0.333333333 0.000 0.000 0.000 0.000 \K Gamma
Band.kpath>

### END ###
########################################

Additionally, the training dataset is from the original database in your connection, Database 1 Best regards, TzuChing

[Xiaoxun-Gong]

I tried running a test calculation using your input file and I didn't find any problem with DeepH-E3, except that I changed the System.Name entry in your input file from MoS2 to openmx. This is because DeepH-E3 parses openmx.scfout file to get Hamiltonian and overlap matrices by default. And do not forget to run cat openmx.out >> openmx.scfout before running DeepH-E3. If the problem still persists, please let me know.

[newplay]

Thank you for your assistance. I will attempt this case again. I believe you are correct; I forgot to execute cat openmx.out >> openmx.scfout before get overlap-matrix and evalue by DeepH-E3.