search

deepmodeling / deepmd-kit

A deep learning package for many-body potential energy representation and molecular dynamics
https://docs.deepmodeling.com/projects/deepmd/
GNU Lesser General Public License v3.0
1.45k stars 499 forks source link

[BUG] `atom_ener` doesn't work in v2.0.0.b0 #712

Closed njzjz closed 3 years ago

njzjz commented 3 years ago

Summary

atom_ener doesn't work in v2.0.0.b0, even if I set both model->fitting_net->atom_ener and model->descriptor->set_davg_zero. The input can be seen in #658. However, it works if the model is trained and frozen in v1.3.3 and converted to v2.0 by #633.

Deepmd-kit version, installation way, input file, running commands, error log, etc.

The model frozen in v2.0.0.b0

Steps to Reproduce

The input is as the same as #658. After freezing the model, predict the energy of a single HW or OW atom using Python API. It's not zero.

from deepmd.infer import DeepPot
import numpy as np
dp = DeepPot('graph.000.pb')
coord = np.array([[0,0,0]]).reshape([1, -1])
atype = [2]
e, f, v = dp.eval(coord, None, atype)
print(e)

Result for the model trained and frozen in v2.0.0.b0: [[0.15377867]] (not zero)

Result for the model trained and frozen in v1.3.3 and converted using #633: [[-1.11022302e-15]] (zero)

njzjz commented 3 years ago

By the way, it's better to add a unittest for it (although I don't know how to do it).