Skip to content

dpa4.md

Latest commit

 

History

History
110 lines (83 loc) · 3.67 KB

dpa4.md

File metadata and controls

110 lines (83 loc) · 3.67 KB
Raw

DPA4 training reference

Read this file only after the user chooses DPA4/SeZM, or when it is the best fit for the task. Keep shared data checks and the train/monitor workflow in ../SKILL.md; this file records DPA4-specific choices.

When to choose DPA4

Choose DPA4 when the user explicitly requests DPA4/SeZM or wants its SO(3)-equivariant message-passing architecture and accepts a GPU-oriented, PyTorch-only workflow. The aliases DPA4, SeZM, and sezm select the same implementation.

DPA4 is not selected merely because a checkpoint ends in .pt. Inspect an existing checkpoint with:

dp --pt show model.pt descriptor fitting-net type-map

Minimal model configuration

Start from the maintained example at examples/water/dpa4/input.json. A minimal model section is:

{
  "model": {
    "type": "dpa4",
    "type_map": [
      "O",
      "H"
    ],
    "descriptor": {
      "rcut": 6.0
    }
  }
}

model.type: "dpa4" selects the DPA4/SeZM descriptor and its default energy fitting network. DPA4 defaults to float32; double precision is unnecessary and not recommended for the normal workflow.

Parameters to choose deliberately

  • rcut sets the local environment cutoff.
  • On the conservative energy path, sel is an initial neighbor-search capacity that grows on demand; it does not truncate the neighbor list. It may also be set to auto or auto:factor from training data.
  • lmax/l_schedule and mmax/m_schedule control angular resolution and are primary accuracy-cost levers.
  • n_blocks controls depth; channels and n_radial control width.
  • n_focus and n_atten_head control aggregation.

Use documented defaults or a maintained example unless the user has evidence for changing these parameters. Do not copy DPA3 descriptor parameters into DPA4.

Train and monitor

Use the PyTorch backend:

dp --pt train input.json

Monitor lcurve.out, validation metrics, checkpoint creation, and non-finite values. DPA4 also supports advanced property, spin, denoising, ZBL, multitask, and LoRA configurations; follow the DPA4 documentation and examples rather than combining those features from memory. For checkpoint adaptation and LoRA, use the deepmd-finetune-dpa4 skill.

Freeze and test

DPA4 checkpoints are .pt, but deployment uses an AOTInductor .pt2 archive:

dp --pt freeze -c model.ckpt.pt -o frozen_model
dp test -m frozen_model.pt2 -s /path/to/test_system -n 30

The command detects DPA4/SeZM and appends .pt2. DPA4 does not use the ordinary TorchScript .pth freeze path and does not support model compression. Validate the exported archive in the target inference or LAMMPS environment.

If the checkpoint is multi-task, inspect its branches and pass the selected head during export:

dp --pt show model.ckpt.pt model-branch descriptor type-map
dp --pt freeze -c model.ckpt.pt -o frozen_model --head SELECTED_BRANCH

The frozen .pt2 is a selected single-head artifact.

DPA4 checklist

  • The PyTorch backend is available.
  • model.type is dpa4/sezm, or the stored checkpoint configuration proves it.
  • type_map, data labels, and train/validation systems are consistent.
  • Parameter changes are based on DPA4 documentation, not DPA3 defaults.
  • Training and validation metrics are finite.
  • The selected checkpoint is exported to .pt2 and tested.
  • dp compress is not used for DPA4.

References