Retraining DeepMD Model: Should I Use -f graph.pb or --init-model model.ckpt?

[NoorAldinAlz]

I want to retrain an existing DeepMD model with new data. I see two options:

1. dp train -f graph.pb input.json
2. dp train --init-model model.ckpt input.json

Which of these is the correct or recommended approach for retraining a model? What are the key differences between these two methods when used in the context of transfer learning or retraining with new data?

[wanghan-iapcm]

They should be equivalent. The difference is the file format of the model.

1. a frozen model file (generated by dp freeze)
2. a model checkpoint