add-descriptor

Adding a New Descriptor to deepmd-kit

Follow these steps in order. Each step lists files to create/modify and patterns to follow.

Step 1: Implement in dpmodel

Create deepmd/dpmodel/descriptor/<name>.py

Inherit from NativeOP and BaseDescriptor. Register with decorators:

from deepmd.dpmodel import NativeOP
from .base_descriptor import BaseDescriptor


@BaseDescriptor.register("your_name")
@BaseDescriptor.register("alias_name")  # optional aliases
class DescrptYourName(NativeOP, BaseDescriptor): ...

Key requirements:

• __init__: initialize cutoff, sel, networks, davg/dstd statistics
• call(coord_ext, atype_ext, nlist, mapping=None): forward pass returning (descriptor, rot_mat, g2, h2, sw)
• serialize() -> dict: save with @class, type, @version, @variables keys
• deserialize(cls, data): reconstruct from dict
• Property/getter methods: get_rcut, get_sel, get_dim_out, mixed_types, etc.
• __getitem__/__setitem__ for davg/dstd access via multiple key aliases

All dpmodel code must use array_api_compat for cross-backend compatibility (numpy/torch/jax/paddle). See references/dpmodel-implementation.md for full method table, array API pitfalls, and utilities.

Reference implementations:

• Simple: deepmd/dpmodel/descriptor/se_e2_a.py
• Three-body: deepmd/dpmodel/descriptor/se_t.py
• Attention-based: deepmd/dpmodel/descriptor/dpa1.py

Step 2: Register

Edit deepmd/dpmodel/descriptor/__init__.py — add import and __all__ entry.

Edit deepmd/utils/argcheck.py — register descriptor arguments:

@descrpt_args_plugin.register("your_name", alias=["alias"], doc="Description")
def descrpt_your_name_args() -> list[Argument]:
    return [
        Argument("sel", [list[int], str], optional=True, default="auto", doc=doc_sel),
        Argument("rcut", float, optional=True, default=6.0, doc=doc_rcut),
        Argument("rcut_smth", float, optional=True, default=0.5, doc=doc_rcut_smth),
        Argument(
            "neuron", list[int], optional=True, default=[10, 20, 40], doc=doc_neuron
        ),
        # ... add all constructor parameters
    ]

Step 3: Wrap for JAX backend

Create deepmd/jax/descriptor/<name>.py

Pattern: @flax_module decorator + custom __setattr__ for attribute conversion.

from deepmd.dpmodel.descriptor.your_name import DescrptYourName as DescrptYourNameDP
from deepmd.jax.common import ArrayAPIVariable, flax_module, to_jax_array
from deepmd.jax.descriptor.base_descriptor import BaseDescriptor
from deepmd.jax.utils.exclude_mask import PairExcludeMask
from deepmd.jax.utils.network import NetworkCollection


@BaseDescriptor.register("your_name")
@flax_module
class DescrptYourName(DescrptYourNameDP):
    def __setattr__(self, name, value):
        if name in {"davg", "dstd"}:
            value = to_jax_array(value)
            if value is not None:
                value = ArrayAPIVariable(value)
        elif name in {"embeddings"}:
            if value is not None:
                value = NetworkCollection.deserialize(value.serialize())
        elif name == "env_mat":
            pass  # stateless
        elif name == "emask":
            value = PairExcludeMask(value.ntypes, value.exclude_types)
        return super().__setattr__(name, value)

For nested sub-components, define wrapper classes bottom-up. See deepmd/jax/descriptor/dpa1.py for example.

Edit deepmd/jax/descriptor/__init__.py — add import and __all__ entry.

Step 4: Wrap for pt_expt backend

Create deepmd/pt_expt/descriptor/<name>.py

The @torch_module decorator handles everything automatically:

• Auto-generates forward() delegating to call() (and forward_lower() from call_lower())
• Auto-generates __setattr__ that converts numpy arrays to torch buffers and dpmodel objects to pt_expt modules via a converter registry
• Any unregistered NativeOP assigned as an attribute will raise TypeError — register it first

Simple descriptors (no custom sub-components) need only an empty body:

from deepmd.dpmodel.descriptor.your_name import DescrptYourName as DescrptYourNameDP
from deepmd.pt_expt.common import torch_module
from deepmd.pt_expt.descriptor.base_descriptor import BaseDescriptor


@BaseDescriptor.register("your_name")
@torch_module
class DescrptYourName(DescrptYourNameDP):
    pass

Standard dpmodel sub-components (NetworkCollection, EmbeddingNet, PairExcludeMask, EnvMat, TypeEmbedNet) are pre-registered in deepmd/pt_expt/utils/ and converted automatically. No __setattr__ override needed.

For custom sub-components (e.g., a new block class inheriting NativeOP), create a separate wrapper file and register bottom-up with register_dpmodel_mapping:

# deepmd/pt_expt/descriptor/your_block.py
from deepmd.dpmodel.descriptor.your_block import YourBlock as YourBlockDP
from deepmd.pt_expt.common import register_dpmodel_mapping, torch_module


@torch_module
class YourBlock(YourBlockDP):
    pass


register_dpmodel_mapping(
    YourBlockDP,
    lambda v: YourBlock.deserialize(v.serialize()),
)

Then import this module in deepmd/pt_expt/descriptor/__init__.py for its side effect (the registration must happen before the parent descriptor is instantiated).

Reference: deepmd/pt_expt/descriptor/se_t_tebd.py + se_t_tebd_block.py

Edit deepmd/pt_expt/descriptor/__init__.py — add import and __all__ entry.

Step 5: Hard-code for PT backend (if needed)

Create deepmd/pt/model/descriptor/<name>.py

PT descriptors are fully reimplemented in PyTorch (not wrapping dpmodel). They inherit from BaseDescriptor and torch.nn.Module. Must implement forward(), serialize(), deserialize().

Edit deepmd/pt/model/descriptor/__init__.py — add import.

Reference: deepmd/pt/model/descriptor/se_a.py

Step 6: Hard-code for TF backend (if needed)

Create deepmd/tf/descriptor/<name>.py

TF descriptors are fully reimplemented in TensorFlow. They inherit from BaseDescriptor and implement the TF computation graph.

Edit deepmd/tf/descriptor/__init__.py — add import.

Reference: deepmd/tf/descriptor/se_a.py

Step 7: Hard-code for PD backend (if needed)

Same as PT but using Paddle. Inherit from BaseDescriptor and paddle.nn.Layer.

Edit deepmd/pd/model/descriptor/__init__.py — add import.

Reference: deepmd/pd/model/descriptor/se_a.py

Step 8: Write tests

Eight test categories. See references/test-patterns.md for full code templates.

pt_expt tests use pytest.mark.parametrize (not itertools.product), do not inherit from unittest.TestCase, and use setup_method (not setUp).

Step 9: Write documentation

Create doc/model/<name>.md

Each descriptor needs a documentation page in doc/model/. Use MyST Markdown format with Sphinx extensions. List supported backends using icon substitutions.

Template:

# Descriptor `"your_name"` {{ pytorch_icon }} {{ dpmodel_icon }}

:::{note}
**Supported backends**: PyTorch {{ pytorch_icon }}, DP {{ dpmodel_icon }}
:::

Brief description of what the descriptor is and its theoretical motivation.

## Theory

Mathematical formulation using LaTeX:

```math
    \mathcal{D}^i = ...
```

## Instructions

Example JSON configuration:

```json
"descriptor": {
    "type": "your_name",
    "sel": [46, 92],
    "rcut_smth": 0.50,
    "rcut": 6.00,
    "neuron": [10, 20, 40],
    "resnet_dt": false,
    "seed": 1
}
```

Explain key parameters and link to the argument schema using `{ref}` directives,
e.g. `{ref}rcut <model[standard]/descriptor[your_name]/rcut>`.

Available backend icons: {{ tensorflow_icon }}, {{ pytorch_icon }}, {{ jax_icon }}, {{ paddle_icon }}, {{ dpmodel_icon }}. Only list backends that actually support this descriptor.

Edit doc/model/index.rst — add the new page to the toctree:

.. toctree::
   :maxdepth: 1

   ...
   <name>

Reference docs: doc/model/train-se-e2-r.md (simple), doc/model/dpa2.md (modern)

Verification

# dpmodel self-consistency
python -m pytest source/tests/common/dpmodel/test_descriptor_<name>.py -v

# pt_expt unit tests
python -m pytest source/tests/pt_expt/descriptor/test_<name>.py -v

# Cross-backend consistency
python -m pytest source/tests/consistent/descriptor/test_<name>.py -v

# PT/PD unit tests (if hard-coded)
python -m pytest source/tests/pt/model/test_descriptor_<name>.py -v
python -m pytest source/tests/pd/model/test_descriptor_<name>.py -v

# Quick smoke test
python -c "
from deepmd.dpmodel.descriptor import DescrptYourName
d = DescrptYourName(rcut=6.0, rcut_smth=1.8, sel=[20, 20])
d2 = DescrptYourName.deserialize(d.serialize())
print('Round-trip OK:', d.get_dim_out() == d2.get_dim_out())
"

Files summary