Attempted to add ZBL training in finetune but failed

[JiangXiaoMingSan]

When using deepmdkit-b3 and scripts from GitHub for finetune, an error is reported. When using the se_2_a descriptor, an error is reported: ValueError: LinearAtomicModel only supports AtomicModel of mixed type, the following models are not mixed type: [DPAtomicModel(
(descriptor): DescrptSeA(
(sea): DescrptBlockSeA(
(emask): PairExcludeMask()
(filter_layers): NetworkCollection(
(networks): ModuleList(
(0-1): 2 x EN(
(layers): ModuleList(
(0-2): 3 x MLPLayer(
(activate): ActivationFn()
)
)
)
)
(_networks): ModuleList(

    (0-1): 2 x EN(
        (layers): ModuleList(
          (0-2): 3 x MLPLayer(
            (activate): ActivationFn()
          )
        )
      )
    )
  )
)

)
(fitting_net): EnergyFittingNet(
(emask): AtomExcludeMask()
(filter_layers): NetworkCollection(
(networks): ModuleList(
(0-1): 2 x FN(
(layers): ModuleList(
(0-3): 4 x MLPLayer(
(activate): ActivationFn()
)
)
)
)
(_networks): ModuleList(
(0-1): 2 x FN(
(layers): ModuleList(
(0-3): 4 x MLPLayer(
(activate): ActivationFn()
)
)
)
)
)
)
)].

I don't quite understand this error, the model I am using is OpenLAM_2.2.0_27heads_beta3.pt, What type is a mixed type?
But when I tried to change the descriptor to hybrid and dpa2, I encountered another error about rcut:
Traceback (most recent call last):
File "/opt/deepmd-kit-3.0.0b3/bin/dp", line 10, in
sys.exit(main())
File "/opt/deepmd-kit-3.0.0b3/lib/python3.10/site-packages/deepmd/main.py", line 923, in main
deepmd_main(args)
File "/opt/deepmd-kit-3.0.0b3/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/init.py", line 346, in wrapper
return f(*args, **kwargs)
File "/opt/deepmd-kit-3.0.0b3/lib/python3.10/site-packages/deepmd/pt/entrypoints/main.py", line 562, in main
train(FLAGS)
File "/opt/deepmd-kit-3.0.0b3/lib/python3.10/site-packages/deepmd/pt/entrypoints/main.py", line 326, in train
trainer = get_trainer(
File "/opt/deepmd-kit-3.0.0b3/lib/python3.10/site-packages/deepmd/pt/entrypoints/main.py", line 190, in get_trainer
trainer = training.Trainer(
File "/opt/deepmd-kit-3.0.0b3/lib/python3.10/site-packages/deepmd/pt/train/training.py", line 277, in init
self.model = get_model_for_wrapper(model_params)
File "/opt/deepmd-kit-3.0.0b3/lib/python3.10/site-packages/deepmd/pt/train/training.py", line 1258, in get_model_for_wrapper
_model = get_single_model(
File "/opt/deepmd-kit-3.0.0b3/lib/python3.10/site-packages/deepmd/pt/train/training.py", line 1250, in get_single_model
model = get_zbl_model(deepcopy(_model_params)).to(DEVICE)
File "/opt/deepmd-kit-3.0.0b3/lib/python3.10/site-packages/deepmd/pt/model/model/init.py", line 131, in get_zbl_model
model_params["descriptor"]["rcut"],
KeyError: 'rcut'

Specific scripts can be found on GitHub, and I have added ZBL to all the scripts.( https://github.com/JiangXiaoMingSan/finetune_zbl_test.git )

[njzjz]

File "/opt/deepmd-kit-3.0.0b3/lib/python3.10/site-packages/deepmd/pt/model/model/init.py", line 131, in get_zbl_model
model_params["descriptor"]["rcut"],
KeyError: 'rcut'

@anyangml please take a look at this issue. The descriptor should not be assumed to be dpa1

[njzjz]

descriptor.get_rcut() and descriptor.get_sel() should be used instead.

[anyangml]

In the example we provided, we did not give the pairtable model a separate rcut and sel . We simply used the the rcut and sel from the dp model. Technically, pairtable model has its own rcut and sel, and when combining into a ZBL model, the rcut and sel are processed within the ZBL model. If we want the flexibility of using a different rcut and sel, we would need to add these information in the input.json at the model_params level. If we don't need this flexibility, we can simply update the get_zbl_model method.

[anyangml]

fixed in #4335