Chore: refactor InvarFitting

[anyangml]

No description provided.

[codecov]

Codecov Report

Attention: 4 lines in your changes are missing coverage. Please review.

Comparison is base (43f17da) 74.98% compared to head (e125be1) 74.99%.

Files	Patch %	Lines
deepmd/pt/model/task/fitting.py	97.14%	4 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##            devel    #3266   +/-   ##
=======================================
  Coverage   74.98%   74.99%           
=======================================
  Files         384      384           
  Lines       33522    33547   +25     
  Branches     1604     1604           
=======================================
+ Hits        25136    25157   +21     
- Misses       7525     7529    +4     
  Partials      861      861           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[anyangml]

@wanghan-iapcm, this is only a draft, not tested, but it's what i have in mind. feels like we can reuse a large portion of the code from InvarFitting, we only need to change the output dim from 1 to 3or 9, and transform the output with gr. If that's the case, we might want to create a GeneralFitting to encapsulate the common methods, and refactor InvarFitting as a child class.

[wanghan-iapcm]

please note that the predictions of dipole and polar are rotationally equivariant.

[anyangml]

please note that the predictions of dipole and polar are rotationally equivariant.

I thought this handles the rotational equivariance,

deepmd-kit/deepmd/pt/model/task/dipole.py

Line 203 in 9bc9900

torch.bmm(outs, gr).squeeze(-2).view(nframes, nloc, 3)

[anyangml]

@wanghan-iapcm UT 'test_make_hessian_model` fails on CUDA, not sure if it was tested before.