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?

1. If using a built-in pretrained model, download it first with dp pretrained download.
2. Collect the downstream training data paths and element types.
3. Generate the fine-tuning input.json.
4. Run fine-tuning and monitor the learning curve.
5. 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,
      "pre_data_2": 0.3,
      "downstream": 1.0
    },
    "data_dict": {
      "pre_data_1": {
        "training_data": {
          "systems": [
            "./pre_data_1/train"
          ],
          "batch_size": 1
        }
      },
      "pre_data_2": {
        "training_data": {
          "systems": [
            "./pre_data_2/train"
          ],
          "batch_size": 1
        }
      },
      "downstream": {
        "training_data": {
          "systems": [
            "./downstream/train"
          ],
          "batch_size": 1
        },
        "validation_data": {
          "systems": [
            "./downstream/valid"
          ],
          "batch_size": 1
        }
      }
    },
    "numb_steps": 200000,
    "gradient_max_norm": 5.0,
    "disp_file": "lcurve.out",
    "disp_freq": 100,
    "save_freq": 2000
  }
}

Key points:

• "finetune_head": "pre_data_1" specifies which branch the downstream task fine-tunes from.
• model_prob controls the sampling probability for each dataset.
• Pre-trained branches continue training in init-model mode; the downstream branch fine-tunes from the selected head.

Run:

dp --pt train multi_input.json --finetune multitask_pretrained.pt

Freeze a specific branch:

dp --pt freeze -o model_downstream.pth --head downstream

Scenario 3: Fine-tune from Built-in Pretrained Models

DeePMD-kit provides built-in pretrained models that can be downloaded directly.

Step 1: Check Available Models

dp pretrained download -h

Currently available models include:

• DPA-3.3-1M — 1M parameter DPA3 pretrained model
• DPA-3.2-5M — latest large-scale pretrained model
• DPA-3.1-3M — 3M parameter DPA3 pretrained model
• DPA3-Omol-Large — large organic molecule model

Step 2: Download the Model

# Download to default cache directory
dp pretrained download DPA-3.1-3M

# Download to a custom directory
dp pretrained download DPA-3.1-3M --cache-dir ./models

The command prints the local path of the downloaded model file on success.

Step 3: Check Model Branches (if multi-task)

dp --pt show /path/to/DPA-3.1-3M.pt model-branch

Step 4: Prepare input.json and Run Fine-tuning

The input.json is the same as Scenario 1. Use --use-pretrain-script to inherit the model architecture:

{
  "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": [
        "./my_data/train_0",
        "./my_data/train_1"
      ],
      "batch_size": 1
    },
    "validation_data": {
      "systems": [
        "./my_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
  }
}

The meaning of each parameter can be generated through dp doc-train-input. Considering the output RST documentation on the screen is very long, use grep to find the documentation of a specific parameter:

dp doc-train-input | grep -A 7 training/numb_steps

Run fine-tuning:

# Single-task fine-tuning from a specific branch
dp --pt train input.json --finetune /path/to/DPA-3.1-3M.pt --model-branch CHOSEN_BRANCH --use-pretrain-script

# If the pretrained model is single-task, --model-branch is not needed
dp --pt train input.json --finetune /path/to/DPA3-Omol-Large.pt --use-pretrain-script

Step 5: Freeze and Test

dp --pt freeze -o finetuned_model.pth
dp --pt test -m finetuned_model.pth -s /path/to/test_system -n 30

Fine-tuning Command Reference

Agent Checklist

• [ ] Pre-trained model file exists (downloaded or self-trained)
• [ ] Downstream data elements are a subset of the pre-trained model's type_map
• [ ] --use-pretrain-script is used if model architecture is unknown
• [ ] Learning rate is reduced compared to training from scratch (e.g., 1e-4 vs 1e-3)
• [ ] For multi-task pretrained models, the correct --model-branch is selected
• [ ] Training completes without NaN in lcurve.out
• [ ] Fine-tuned model is frozen and tested
• [ ] Test RMSE values are reported to the user

References

• Fine-tuning documentation
• Pretrained model download
• Multi-task training
• DPA3 descriptor documentation
• DeePMD-kit GitHub