跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
设计与 UI

deepmd-finetune-dpa3

Fine-tune a DPA3 model in DeePMD-kit using the PyTorch backend. Use when the user wants to adapt a pre-trained DPA3 model to a new downstream dataset. Supports fine-tuning from a self-trained DPA3 model (.pt checkpoint), from a multi-task pre-trained model, or from a built-in pretrained model downloaded via `dp pretrained download` (e.g., DPA-3.1-3M, DPA-3.2-5M, DPA-3.3-1M). Covers single-task and multi-task fine-tuning workflows.

jinzhezenggroup · 已有 AI 审核记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add jinzhezenggroup/computational-chemistry-agent-skills --skill deepmd-finetune-dpa3

技能说明


name: deepmd-finetune-dpa3 description: Fine-tune a DPA3 model in DeePMD-kit using the PyTorch backend. Use when the user wants to adapt a pre-trained DPA3 model to a new downstream dataset. Supports fine-tuning from a self-trained DPA3 model (.pt checkpoint), from a multi-task pre-trained model, or from a built-in pretrained model downloaded via dp pretrained download (e.g., DPA-3.1-3M, DPA-3.2-5M, DPA-3.3-1M). Covers single-task and multi-task fine-tuning workflows. compatibility: Requires deepmd-kit with PyTorch backend installed. GPU strongly recommended. license: LGPL-3.0-or-later metadata: author: iProzd version: '1.0' repository: https://github.com/deepmodeling/deepmd-kit

DeePMD-kit Fine-tuning: DPA3

Fine-tune a pre-trained DPA3 model on a downstream dataset. This skill covers three scenarios:

  1. Fine-tuning from a self-trained single-task DPA3 model
  2. Fine-tuning from a multi-task pre-trained DPA3 model
  3. Fine-tuning from a built-in pretrained model (e.g., DPA-3.1-3M, DPA-3.2-5M, DPA-3.3-1M) downloaded via dp pretrained download

Quick Start

# Fine-tune from a self-trained model
dp --pt train input.json --finetune pretrained.pt --use-pretrain-script

# Fine-tune from a built-in pretrained model
dp pretrained download DPA-3.2-5M
dp --pt train input.json --finetune /path/to/DPA-3.2-5M.pt --use-pretrain-script --model-branch OMat24

Agent Responsibilities

  1. Determine the fine-tuning scenario:
    • Does the user have a self-trained .pt model?
    • Does the user want to use a built-in pretrained model (DPA-3.1-3M, DPA-3.2-5M, DPA-3.3-1M, etc.)?
    • Is the pre-trained model single-task or multi-task?
  2. If using a built-in pretrained model, download it first with dp pretrained download.
  3. Collect the downstream training data paths and element types.
  4. Generate the fine-tuning input.json.
  5. Run fine-tuning and monitor the learning curve.
  6. Freeze and test the fine-tuned model.

Scenario 1: Fine-tune from a Self-trained Single-task Model

When you have trained a DPA3 model yourself and want to adapt it to new data.

Step 1: Prepare input.json

When using --use-pretrain-script, the model architecture is inherited from the pre-trained model. You only need to specify type_map, data paths, and training parameters:

{
  "model": {
    "type_map": [
      "O",
      "H"
    ],
    "descriptor": {},
    "fitting_net": {}
  },
  "learning_rate": {
    "type": "exp",
    "decay_steps": 5000,
    "start_lr": 0.0001,
    "stop_lr": 3e-06
  },
  "loss": {
    "type": "ener",
    "start_pref_e": 0.2,
    "limit_pref_e": 20,
    "start_pref_f": 100,
    "limit_pref_f": 60,
    "start_pref_v": 0.02,
    "limit_pref_v": 1
  },
  "optimizer": {
    "type": "AdamW",
    "weight_decay": 0.001
  },
  "training": {
    "training_data": {
      "systems": [
        "./downstream_data/train_0",
        "./downstream_data/train_1"
      ],
      "batch_size": 1
    },
    "validation_data": {
      "systems": [
        "./downstream_data/valid_0"
      ],
      "batch_size": 1
    },
    "numb_steps": 200000,
    "gradient_max_norm": 5.0,
    "seed": 10,
    "disp_file": "lcurve.out",
    "disp_freq": 100,
    "save_freq": 2000
  }
}

Fine-tuning tips:

  • Use a smaller start_lr (e.g., 1e-4) than training from scratch (1e-3).
  • Use fewer numb_steps since the model is already pre-trained.
  • The elements in the downstream data must be a subset of the pre-trained model's type_map.

Step 2: Run Fine-tuning

dp --pt train input.json --finetune pretrained.pt --use-pretrain-script

The --use-pretrain-script flag tells DeePMD-kit to inherit the model architecture from the pre-trained model, so the descriptor and fitting_net sections in input.json can be empty.

Without --use-pretrain-script, the model section in input.json must exactly match the pre-trained model's architecture.

Scenario 2: Fine-tune from a Multi-task Pre-trained Model

When the pre-trained model was trained with multiple datasets (multi-task training), you can select a specific branch to fine-tune from.

Check Available Branches

dp --pt show multitask_pretrained.pt model-branch

Run Fine-tuning from a Specific Branch

dp --pt train input.json --finetune multitask_pretrained.pt --model-branch CHOSEN_BRANCH --use-pretrain-script

If --model-branch is not set or set to RANDOM, a randomly initialized fitting net will be used.

Multi-task Fine-tuning (Prevent Forgetting)

To retain knowledge from the pre-trained datasets during fine-tuning, use multi-task fine-tuning. Prepare a multi-task input script:

{
  "model": {
    "shared_dict": {
      "type_map_all": [
        "O",
        "H",
        "C",
        "N"
      ],
      "dpa3_desc": {
        "type": "dpa3",
        "repflow": {}
      }
    },
    "model_dict": {
      "pre_data_1": {
        "type_map": "type_map_all",
        "descriptor": "dpa3_desc",
        "fitting_net": {}
      },
      "pre_data_2": {
        "type_map": "type_map_all",
        "descriptor": "dpa3_desc",
        "fitting_net": {}
      },
      "downstream": {
        "finetune_head": "pre_data_1",
        "type_map": "type_map_all",
        "descriptor": "dpa3_desc",
        "fitting_net": {}
      }
    }
  },
  "learning_rate": {
    "type": "exp",
    "decay_steps": 5000,
    "start_lr": 0.0001,
    "stop_lr": 3e-06
  },
  "loss_dict": {
    "pre_data_1": {
      "type": "ener",
      "start_pref_e": 0.2,
      "limit_pref_e": 20,
      "start_pref_f": 100,
      "limit_pref_f": 60
    },
    "pre_data_2": {
      "type": "ener",
      "start_pref_e": 0.2,
      "limit_pref_e": 20,
      "start_pref_f": 100,
      "limit_pref_f": 60
    },
    "downstream": {
      "type": "ener",
      "start_pref_e": 0.2,
      "limit_pref_e": 20,
      "start_pref_f": 100,
      "limit_pref_f": 60
    }
  },
  "training": {
    "model_prob": {
      "pre_data_1": 0.3,
 
阅读完整来源 ↗