[BUG] Incorrect results from DPLR model in python inference interface

[ChiahsinChu]

Bug summary

In deepmd v3, the inference results from the DPLR model are incorrect when using the Python interface. In deepmd-v3, the long-range correction (i.e., the data modifier) is not calculated and added to the final results. It should be relevant to the change in architecture in the inference module.

DeePMD-kit Version

DeePMD-kit v3.0.1

Backend and its version

TensorFlow v2.18.0-rc2-4-g6550e4bd802

How did you download the software?

pip

Input Files, Running Commands, Error Log, etc.

Test python script for v2:

from ase import build
import numpy as np

from deepmd.infer import DeepPot


dp = DeepPot("graph.pb")
dm = dp.dm

type_map = ["O", "H"]

atoms = build.molecule("H2O")
atoms.set_cell([10, 10, 10])
atoms.center()

atype = np.array([type_map.index(a) for a in atoms.symbols])


out = dp.eval(
    atoms.get_positions().reshape(1, -1),
    atoms.get_cell().reshape(1, -1),
    atype,
)
print(out[0])

out_dm = dm.eval(
    atoms.get_positions().reshape(1, -1),
    atoms.get_cell().reshape(1, -1),
    atype,
)
print(out_dm[0])

Output:

[[-494.9365821]]
[46.573121]

Test python script for v3:

from ase import build
import numpy as np

from deepmd.tf.infer.data_modifier import DipoleChargeModifier
from deepmd.infer import DeepPot


dp = DeepPot("graph.pb")
dm = DipoleChargeModifier(
    "dw_model.pb",
    model_charge_map=[-8],
    sys_charge_map=[6, 1],
    ewald_h=0.5,
    ewald_beta=0.4,
)

type_map = ["O", "H"]

atoms = build.molecule("H2O")
atoms.set_cell([10, 10, 10])
atoms.center()

atype = np.array([type_map.index(a) for a in atoms.symbols])


out = dp.eval(
    atoms.get_positions().reshape(1, -1),
    atoms.get_cell().reshape(1, -1),
    atype,
)
print(out[0])

out_dm = dm.eval(
    atoms.get_positions().reshape(1, -1),
    atoms.get_cell().reshape(1, -1),
    atype,
)
print(out_dm[0])

Output:

[[-541.5097031]]
[46.573121]

In both cases, the energies from data modifier are identical. However, the total energy in v3 = the total energy in v2 + the energy of data modifier.

Steps to Reproduce

Please download the attached models and run the scripts above with deepmd v2/v3.

Further Information, Files, and Links

H2O_dplr_model.zip

[njzjz]

deepmd-kit/deepmd/tf/infer/deep_eval.py

Line 764 in 70bc6d8

if self.modifier_type is not None and isinstance(self.model_type, DeepPot):

This line seems wrong; it should be issubclass(self.model_type, DeepPot).

[njzjz]

However, I don't understand why this test can pass:

https://github.com/deepmodeling/deepmd-kit/blob/70bc6d89d057d348558b87865ff2f225cae677aa/source/tests/tf/test_dplr.py

[ChiahsinChu]

This line seems wrong; it should be issubclass(self.model_type, DeepPot).

Yes, but there should be more than this. I tried temporality to remove this condition but it still did not work. The problem seems to come from the impropriate initialisation of the DeepEval class, as the result of which _init_tensors and _init_attr methods are not called.

[ChiahsinChu]

However, I don't understand why this test can pass:

Where did the reference data come from? When I tried to run this test in v2.2.11, I got the error about the missing dipole model, which makes sense to me:

Traceback (most recent call last):
  File "/home/jxzhu/workspace/test/test.py", line 7, in <module>
    dp = DeepPot("lrmodel.pb")
  File "/home/jxzhu/apps/deepmd/2.2.11/deepmd/infer/deep_pot.py", line 205, in __init__
    self.dm = DipoleChargeModifier(
  File "/home/jxzhu/apps/deepmd/2.2.11/deepmd/infer/data_modifier.py", line 61, in __init__
    DeepDipole.__init__(
  File "/home/jxzhu/apps/deepmd/2.2.11/deepmd/infer/deep_dipole.py", line 58, in __init__
    DeepTensor.__init__(
  File "/home/jxzhu/apps/deepmd/2.2.11/deepmd/infer/deep_tensor.py", line 71, in __init__
    DeepEval.__init__(
  File "/home/jxzhu/apps/deepmd/2.2.11/deepmd/infer/deep_eval.py", line 64, in __init__
    self.graph = self._load_graph(
  File "/home/jxzhu/apps/deepmd/2.2.11/deepmd/infer/deep_eval.py", line 189, in _load_graph
    graph_def.ParseFromString(f.read())
  File "/home/jxzhu/apps/miniconda3/envs/deepmd-v2.2.11/lib/python3.10/site-packages/tensorflow/python/lib/io/file_io.py", line 114, in read
    self._preread_check()
  File "/home/jxzhu/apps/miniconda3/envs/deepmd-v2.2.11/lib/python3.10/site-packages/tensorflow/python/lib/io/file_io.py", line 76, in _preread_check
    self._read_buf = _pywrap_file_io.BufferedInputStream(
tensorflow.python.framework.errors_impl.NotFoundError: lrdipole.pb; No such file or directory

In other words, the v3 code does not try to load the dipole model...

[njzjz]

Where did the reference data come from?

It comes from #2545.

[ChiahsinChu]

Where did the reference data come from?

It comes from #2545.

I see. I checked the orginal test in v2, and found that the reference data used in the UT in v3 is actually the sr term. That explains it.

[njzjz]

I see. Ideally it should reproduce the following result.

deepmd-kit/source/lmp/tests/test_dplr.py

Lines 128 to 138 in bab9258

	expected_e_lr_efield_constant = -40.57352302
	expected_f_lr_efield_constant = np.array(
	[
	[0.47635071, 0.15088380, -1.20378471],
	[-0.52759976, -0.01182856, -0.72773815],
	[-0.70317794, 0.29171446, 1.22375302],
	[0.20500683, -0.50210283, -0.04263579],
	[0.82463041, 0.04231172, 0.47856560],
	[-0.27521024, 0.02902140, 0.27184004],
	]
	)