Model Reuse in dpgen#

dpgen supports two related but distinct mechanisms for reusing previously trained models across active learning iterations:

1. training_init_model — a transfer-learning / warm-start approach where model weights from the previous iteration are carried forward as the starting point for the next training run.
2. training_reuse_* family — a set of parameters that tune how the reuse training proceeds (data sampling ratio, number of steps, learning rate, and loss prefactors).

All parameters live in param.json under the top-level training section and are documented in dpgen/generator/arginfo.py. The application logic is in make_train_dp() and run_train_dp().

training_init_model#

When training_init_model: true, the model trained at iteration N−1 is passed to DeePMD-kit via --init-model old/model.ckpt, preserving all learned weights (embedding network, fitting network) instead of starting from random initialization 1. At iteration 0, it falls back to initializing from paths listed in training_iter0_model_path 2.

The actual --init-model flag injection happens in run_train_dp(). If a checkpoint already exists (e.g., a restarted run), dpgen uses --restart instead, so --init-model only fires on a fresh start 3.

Related options — for iteration 0 only:

• training_init_frozen_model — initializes from a pre-existing frozen model (.pb / .pth) via --init-frz-model 4 5.
• training_finetune_model — fine-tunes from a frozen model via --finetune 6 7.

All three (training_init_model, training_init_frozen_model, training_finetune_model) are mutually exclusive; dpgen raises a RuntimeError if more than one is set 8.

training_reuse_* — Reuse Mode#

These parameters are only active when training_reuse_iter is set and the current iteration index ≥ training_reuse_iter. At that point, training_init_model is also forced to true 9.

Alias: training_reuse_stop_batch is a legacy alias for training_reuse_numb_steps 16.

training_reuse_old_ratio values#

dpgen uses this ratio to set a prob_sys_size sampling probability, splitting data directories into an "old" group (iterations 0 … N−2) and a "new" group (iteration N−1) 17:

• float (e.g., 0.3) — 30% old data, 70% new data. Values of 0.3–0.5 are recommended when the new dataset is much smaller than the accumulated old data 18.
• "auto:f" — automatic ratio where the new-to-old sampling ratio is f; dpgen counts frames in all data systems and computes ratio = n_old / (n_old + f × n_new) 19.
• "auto" — equivalent to "auto:10" (10× preference for new data) 11.

Example configuration#

"training_init_model": true,
"training_reuse_iter": 5,
"training_reuse_old_ratio": 0.3,
"training_reuse_numb_steps": 400000,
"training_reuse_start_lr": 1e-4,
"training_reuse_start_pref_e": 0.1,
"training_reuse_start_pref_f": 100

This is representative of a scenario where a large initial dataset is supplemented by smaller per-iteration FP data 20.

Key source files#