[FEATURE_REQUEST] "dp compress" supports only the compression of energy model?

[salinelake]

Summary

For the recent release (deepmd-kit-2.0.3-cuda11.3_gpu-Linux-x86_64.sh), "dp compress" will call the "init_variables" method of the fitting neural network, through

deepmd-kit/deepmd/train/trainer.py

Line 286 in 159e45d

self.fitting.init_variables(get_fitting_net_variables(self.model_param['compress']['model_file']))

.

But this method is only implemented in deepmd/fit/ener.py, not in deepmd/fit/dipole.py.

The compression of dipole used to be supported in an earlier devel version of DP2.0. So I suppose this is a bug due to the recent introduction of "init_variables"?

[wanghan-iapcm]

@denghuilu Could you please take a look?

[denghuilu]

Could you please provide a training data sets as well as the training script for the dipole type fitting net, so that I can test it in my local environment.

[njzjz]

Fixed in #1228.