Questions about "training_init_model" and "training_reuse_*"

[roger13231]

Dear developer,

My question is similar to #1485 that I have a large original dataset, but the new dataset is small. In #1485, Dr. Zeng replied that we can use "training_reuse_old_ratio". I have addtional question.

1. For the meaning of "training_init_model". In the dpgen document, the function of "training_init_model" is to keep the model parameters. Do the model parameters mean the trained results, such as learned weights and biases of the neural networks in the DP model, the descriptor/embedding network, and the fitting network? In other words, does training_init_model: true mean that the model in the current iteration is initialized from the trained model of the previous iteration, instead of being trained from random initialization?

2. For the "training_reuse_*" setting. Should I use these parameters together with "training_init_model" for my new iteration?
   such as "training_init_model": true,
   "training_reuse_old_ratio": 0.3,
   "training_reuse_iter": 135,
   "training_reuse_numb_steps": 400000,
   "training_reuse_start_lr": 0.001
   And what is the recommendation for "training_reuse_old_ratio", 0.2, 0.3, or some other ratio?

[dosubot[bot]]

Hi @roger13231! I'm Dosu and I'm helping the DeepModeling team.

1. Regarding training_init_model:

Yes, your understanding is correct. When training_init_model: true, the model in the current iteration is initialized from the trained model of the previous iteration — preserving all learned weights, biases, descriptor/embedding network, and fitting network parameters — rather than being trained from random initialization [1]. When iter == 0, it initializes from training_iter0_model_path instead [2].

2. Regarding training_reuse_* parameters:

Yes, these should be used together with training_init_model: true. The system works as follows [3]:

• training_reuse_iter — the iteration index at which model reuse mode activates [4]
• training_reuse_old_ratio — the probability of sampling old data during training. A value of 0.3 means 30% old data, 70% new data [5]
• training_reuse_numb_steps — overrides the number of training steps during reuse iterations [6]
• training_reuse_start_lr — overrides the starting learning rate for reuse iterations [7]

Your example configuration looks reasonable:

"training_init_model": true,
"training_reuse_old_ratio": 0.3,
"training_reuse_iter": 135,
"training_reuse_numb_steps": 400000,
"training_reuse_start_lr": 0.001

Regarding the recommended training_reuse_old_ratio:

For your scenario (large original dataset + small new dataset), a value in the range of 0.3–0.5 is reasonable to boost the probability of new data being selected during training [8]. Note that:

• A lower value (e.g., 0.3) gives more weight to new data (70% new, 30% old)
• You can also use "auto" which applies an automatic ratio calculation (default 10:1 new-to-old), or "auto:f" where f is a custom new-to-old ratio factor [5]

A test case in the codebase uses training_reuse_old_ratio=0.8 with training_reuse_start_lr=1e-4 [9], but that's for a scenario where you want to keep most old data. Since your new dataset is small relative to the original, using 0.3 as you proposed should help the model learn from the new configurations effectively.

To reply, just mention @dosu.

---

Docs are dead. Just use Dosu.

[njzjz-bot]

Dosu's answer has one important error: training_reuse_* does not need to be combined with an explicit "training_init_model": true.

Your understanding of training_init_model is correct: it initializes the descriptor/fitting parameters from the preceding iteration's checkpoint instead of a random initialization. However, once iter_index >= training_reuse_iter, DP-GEN sets training_init_model = True itself, and later runs DeePMD with --init-model old/model.ckpt. Therefore, setting it explicitly is redundant.

For training_reuse_old_ratio:

• A float r assigns total sampling probability r to the old data and 1-r to data from the newest completed iteration.
• The default "auto" is "auto:10". DP-GEN computes N_old / (N_old + 10 * N_new), so each new frame receives approximately 10 times its ordinary per-frame sampling weight. This is implemented here.

There is no generally correct choice among 0.2 or 0.3. I would start with "auto", inspect the generated iter.XXXXXX/00.train/000/input.json to confirm the actual auto_prob, and choose a different ratio only if held-out tests show forgetting of old configurations or insufficient learning of the new ones.

Coding agent: Codex
Codex version: codex-cli 0.149.0
Model: gpt-5.6-sol
Reasoning effort: xhigh