support type embedding changing for frozen .pb

[iProzd]

No description provided.

[codecov]

Codecov Report

Patch coverage: 9.52% and project coverage change: -0.06% ⚠️

Comparison is base (0d5737f) 75.45% compared to head (08556fe) 75.40%.

Additional details and impacted files

@@            Coverage Diff             @@
##            devel    #2827      +/-   ##
==========================================
- Coverage   75.45%   75.40%   -0.06%     
==========================================
  Files         242      242              
  Lines       24205    24226      +21     
  Branches     1537     1541       +4     
==========================================
+ Hits        18265    18267       +2     
- Misses       5023     5042      +19     
  Partials      917      917              

Files Changed	Coverage Δ
deepmd/infer/deep_eval.py	82.48% <9.52%> (-13.21%)	⬇️

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

[njzjz]

To my understanding of this PR, the embedding network has the following requirements:

• ResNet does not apply;
• activation function should be an identity function;
• the number of layers should be 1.

These requirements should be documented.

A better way might be to directly update o_typeebd to a constant instead of updating the NN.