Exporting models

Exporting models to work with any metatensor-compatible simulation engine is done with the MetatensorAtomisticModel class. This class takes in an arbitrary torch.nn.Module, with a forward functions that follows the ModelInterface. In addition to the actual model, you also need to define some information about the model, using ModelMetadata and ModelCapabilities.

class metatensor.torch.atomistic.ModelInterface

Bases: Module

Interface for models that can be used with MetatensorAtomisticModel.

There are several requirements that models must satisfy to be usable with MetatensorAtomisticModel. The main one is concerns the forward() function, which must have the signature defined in this interface.

Additionally, the model can request neighbor lists to be computed by the simulation engine, and stored inside the input System. This is done by defining the optional requested_neighbor_lists() method for the model or any of it’s sub-module.

MetatensorAtomisticModel will check if requested_neighbor_lists is defined for all the sub-modules of the model, then collect and unify identical requests for the simulation engine.

forward(systems: List[System], outputs: Dict[str, ModelOutput], selected_atoms: Labels | None) → Dict[str, TensorMap]

This method should run the model for the given systems, returning the requested outputs. If selected_atoms is a set of Labels, only the corresponding atoms should be included as “main” atoms in the calculation and the output.

outputs will be a subset of the capabilities that where declared when exporting the model. For example if a model can compute both an "energy" and a "charge" output, the simulation engine might only request one them.

The returned dictionary should have the same keys as outputs, and the values should contains the corresponding properties of the systems, as computed for the subset of atoms defined in selected_atoms. Some outputs are standardized, and have additional constrains on how the associated metadata should look like, documented in the Standard model outputs section. If you want to define a new output for your own usage, it name should looks like "<domain>::<output>", where <domain> indicates who defines this new output and <output> describes the output itself. For example, "my-package::foobar" for a foobar output defined in my-package.

The main use case for selected_atoms is domain decomposition, where the System given to a model might contain both atoms in the current domain and some atoms from other domains; and the calculation should produce per-atom output only for the atoms in the domain (but still accounting for atoms from the other domains as potential neighbors).

Parameters:

• systems (List[System]) – atomistic systems on which to run the calculation

• outputs (Dict[str, ModelOutput]) – set of outputs requested by the simulation engine

• selected_atoms (Labels | None) – subset of atoms that should be included in the output, defaults to None

Returns:

properties of the systems, as predicted by the machine learning model

Return type:

Dict[str, TensorMap]

requested_neighbor_lists() → List[NeighborListOptions]

Optional method declaring which neighbors list this model requires.

This method can be defined on either the root model or any of it’s sub-modules. A single module can request multiple neighbors list simultaneously if it needs them.

It is then the responsibility of the code calling the model to:

1. call this method (or more generally MetatensorAtomisticModel.requested_neighbor_lists()) to get the list of the requested neighbor lists.;

2. compute all neighbor lists corresponding to these requests and add them to the systems before calling the model.

Return type:

List[NeighborListOptions]

class metatensor.torch.atomistic.MetatensorAtomisticModel(module: ModelInterface, metadata: ModelMetadata, capabilities: ModelCapabilities)

MetatensorAtomisticModel is the main entry point for atomistic machine learning based on metatensor. It is the interface between custom, user-defined models and simulation engines. Users should export their models with this class, and use save() to save the exported model to a file. The exported models can then be loaded by a simulation engine to compute properties of atomistic systems.

When exporting a module, you should declare what the model is capable of (using ModelCapabilities). This includes what units the model expects as input and what properties the model can compute (using ModelOutput). The simulation engine will then ask the model to compute some subset of these properties (through a ModelEvaluationOptions), on all or a subset of atoms of an atomistic system.

The exported module must follow the interface defined by ModelInterface, should not already be compiled by TorchScript, and should be in “eval” mode (i.e. module.training should be False).

For example, a custom module predicting the energy as a constant value times the number of atoms could look like this

>>> class ConstantEnergy(torch.nn.Module):
...     def __init__(self, constant: float):
...         super().__init__()
...         self.constant = torch.tensor(constant).reshape(1, 1)
...
...     def forward(
...         self,
...         systems: List[System],
...         outputs: Dict[str, ModelOutput],
...         selected_atoms: Optional[Labels] = None,
...     ) -> Dict[str, TensorMap]:
...         results: Dict[str, TensorMap] = {}
...         if "energy" in outputs:
...             if outputs["energy"].per_atom:
...                 raise NotImplementedError("per atom energy is not implemented")
...
...             dtype = systems[0].positions.dtype
...             energies = torch.zeros(len(systems), 1, dtype=dtype)
...             for i, system in enumerate(systems):
...                 if selected_atoms is None:
...                     n_atoms = len(system)
...                 else:
...                     n_atoms = len(selected_atoms)
...
...                 energies[i] = self.constant * n_atoms
...
...             systems_idx = torch.tensor([[i] for i in range(len(systems))])
...             energy_block = TensorBlock(
...                 values=energies,
...                 samples=Labels(["system"], systems_idx.to(torch.int32)),
...                 components=torch.jit.annotate(List[Labels], []),
...                 properties=Labels(["energy"], torch.tensor([[0]])),
...             )
...
...             results["energy"] = TensorMap(
...                 keys=Labels(["_"], torch.tensor([[0]])),
...                 blocks=[energy_block],
...             )
...
...         return results

Wrapping and exporting this model would then look like this:

>>> import os
>>> import tempfile
>>> from metatensor.torch.atomistic import MetatensorAtomisticModel
>>> from metatensor.torch.atomistic import (
...     ModelCapabilities,
...     ModelOutput,
...     ModelMetadata,
... )
>>> model = ConstantEnergy(constant=3.141592)
>>> # put the model in inference mode
>>> model = model.eval()
>>> # Define the model capabilities
>>> capabilities = ModelCapabilities(
...     outputs={
...         "energy": ModelOutput(
...             quantity="energy",
...             unit="eV",
...             per_atom=False,
...             explicit_gradients=[],
...         ),
...     },
...     atomic_types=[1, 2, 6, 8, 12],
...     interaction_range=0.0,
...     length_unit="angstrom",
...     supported_devices=["cpu"],
...     dtype="float64",
... )
>>> # define metadata about this model
>>> metadata = ModelMetadata(
...     name="model-name",
...     authors=["Some Author", "Another One"],
...     # references and long description can also be added
... )
>>> # export the model
>>> model = MetatensorAtomisticModel(model, metadata, capabilities)
>>> # save the exported model to disk
>>> with tempfile.TemporaryDirectory() as directory:
...     model.save(os.path.join(directory, "constant-energy-model.pt"))

module: ModelInterface

The torch module exported by this MetatensorAtomisticModel.

Reading from this attribute is safe, but modifying it is not recommended, unless you are familiar with the implementation of the model.

Parameters:

• module (ModelInterface) – The torch module to export.

• capabilities (ModelCapabilities) – Description of the model capabilities.

• metadata (ModelMetadata)

capabilities() → ModelCapabilities

Get the capabilities of the exported model

Return type:

ModelCapabilities

metadata() → ModelMetadata

Get the metadata of the exported model

Return type:

ModelMetadata

requested_neighbor_lists() → List[NeighborListOptions]

Get the neighbors lists required by the exported model or any of the child module.

Return type:

List[NeighborListOptions]

forward(systems: List[System], options: ModelEvaluationOptions, check_consistency: bool) → Dict[str, TensorMap]

Run the exported model and return the corresponding outputs.

Before running the model, this will convert the systems data from the engine unit to the model unit, including all neighbors lists distances.

After running the model, this will convert all the outputs from the model units to the engine units.

Parameters:

• systems (List[System]) – input systems on which we should run the model. The systems should already contain all neighbors lists corresponding to the options in requested_neighbor_lists().

• options (ModelEvaluationOptions) – options for this run of the model

• check_consistency (bool) – Should we run additional check that everything is consistent? This should be set to True when verifying a model, and to False once you are sure everything is running fine.

Returns:

A dictionary containing all the model outputs

Return type:

Dict[str, TensorMap]

export(file: str, collect_extensions: str | None = None)

Export this model to a file that can then be loaded by simulation engine.

Warning

export() is deprecated. Use save() instead.

Parameters:

• file (str) – where to save the model. This can be a path or a file-like object.

• collect_extensions (str | None) – if not None, all currently loaded PyTorch extension will be collected in this directory. If this directory already exists, it is removed and re-created.

save(file: str | Path, collect_extensions: str | None = None)

Save this model to a file that can then be loaded by simulation engine.

Parameters:

• file (str | Path) – where to save the model. This can be a path or a file-like object.

• collect_extensions (str | None) – if not None, all currently loaded PyTorch extension will be collected in this directory. If this directory already exists, it is removed and re-created.