support training with -f or -t argument

[njzjz]

Fix #1122.

[codecov]

Codecov Report

Attention: 3 lines in your changes are missing coverage. Please review.

Comparison is base (9b44711) 48.54% compared to head (3ee219b) 48.64%.

Additional details and impacted files

@@            Coverage Diff             @@
##            devel    #1340      +/-   ##
==========================================
+ Coverage   48.54%   48.64%   +0.09%     
==========================================
  Files          82       82              
  Lines       14659    14681      +22     
==========================================
+ Hits         7116     7141      +25     
+ Misses       7543     7540       -3     

Files	Coverage Δ
dpgen/generator/arginfo.py	100.00% <100.00%> (ø)
dpgen/generator/run.py	67.75% <88.00%> (+0.33%)	⬆️

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

[njzjz]

@wanghan-iapcm Discussion is required: whether these flags should only work in iteration 0 so that they can work with training_reuse_iter without modifying input files.

[wanghan-iapcm]

@wanghan-iapcm Discussion is required: whether these flags should only work in iteration 0 so that they can work with training_reuse_iter without modifying input files.

Yes, they should only work for iteration 0. Actually the finetune does two things: 1. statistic the training data and set the new energy bias in the model, 2. train the new model with the --init-model mode.

You may want to check the implementation in dpgen2 deepmodeling/dpgen2#152