Added all activation functions for model compression.

[liangadam]

Now the model with activation functions such as 'relu', 'relu6', 'softplus', 'sigmoid' in descriptor can be compressed, and the output results have been well tested.

Modified related document with activation function.

[codecov-commenter]

Codecov Report

Merging #1283 (dc60e1b) into devel (abec07e) will decrease coverage by 0.08%.
The diff coverage is 0.00%.

@@            Coverage Diff             @@
##            devel    #1283      +/-   ##
==========================================
- Coverage   75.97%   75.89%   -0.09%     
==========================================
  Files          91       91              
  Lines        7406     7414       +8     
==========================================
  Hits         5627     5627              
- Misses       1779     1787       +8     

Impacted Files	Coverage Δ
deepmd/utils/tabulate.py	85.23% <0.00%> (-2.98%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update abec07e...dc60e1b. Read the comment docs.

[njzjz]

Why insert here?

[liangadam]

Because there is a lack of guidance on available activation functions, neither in DP-GUI nor for editing 'input.json' by oneself.
You can tell me if there is a more suitable place to add.

[njzjz]

I think it's mentioned, though.

[liangadam]

Fine~ It's better to left document work to you because it‘s hard for me to think about it comprehensively.

[wanghan-iapcm]

How was this implementation tested?

[liangadam]

How was this implementation tested?

By using dp test to compare the results of models (trained with different activation functions) before and after compression. The test results show that all compressed models are very accurate.
Also，I deliberately add different types of errors to test the error reporting mechanism. The results are sensible.

Moreover, using lammps to run those models to make sure there are no abnormal result.

The following shows an test example:

[wanghan-iapcm]

How was this implementation tested?

By using dp test to compare the results of models (trained with different activation functions) before and after compression. The test results show that all compressed models are very accurate. Also，I deliberately add different types of errors to test the error reporting mechanism. The results are sensible.

Moreover, using lammps to run those models to make sure there are no abnormal result.

The following shows an test example:

It looks great. Have you tested all types of activation functions you implemented?

[liangadam]

How was this implementation tested?

By using dp test to compare the results of models (trained with different activation functions) before and after compression. The test results show that all compressed models are very accurate. Also，I deliberately add different types of errors to test the error reporting mechanism. The results are sensible.
Moreover, using lammps to run those models to make sure there are no abnormal result.
The following shows an test example:

It looks great. Have you tested all types of activation functions you implemented?

Yes, All of them have been tested~