DeepDFTModel

Included in QATK.MLDFT

class DeepDFTModel(model_dir, model_type=None, cutoff=None, num_interactions=None, node_size=None, num_components=None, scaling_factor=None, range_type=None, gaussian_sigma=None)

Class for a model for predicting electron densities that holds its specific hyperparameters. The model directory will contain the model parameters obtained from the training, the state of the training optimizer, and metadata.

Parameters:

  • model_dir (str) – Absolute or relative path to the directory containing the model files.

  • model_type (DeepDFTModelType) – The type of the model (SchNet/PaiNN).

  • cutoff (PhysicalQuantity of type length.) – The cutoff radius used for determining bonds between atoms.
    Default: 4.0 * Angstrom

  • num_interactions (int) – The number of propagation steps “T” in message passing.
    Default: 3

  • node_size (int) – The size of the embedded feature vector for each node.
    Default: 128

  • num_components (int) – The number of grid values components (e.g. spin, for polarized calc.) in the model.
    Default: 1

  • scaling_factor (float) – The output grid values are divided (scaled down) by this number.
    Default: 1.0

  • range_type (str or None) – The range type for range-separated models (“short-range” or “long-range”).
    Default: None

  • gaussian_sigma (PhysicalQuantity of type length.) – The sigma of the Gaussian (in Bohr units) for the convolution with the density used when making range-separated density models.
    Default: 3.0 * Bohr

absoluteModelDir()
Returns:

The absolute path to the model directory.

Return type:

str

absoluteModelDirs()
Returns:

A list of absolute paths to all model directories in the parent directory.

Return type:

list[str]

absoluteModelFile()
Returns:

The absolute path to the model file.

Return type:

str

atomsToGraphDict(configuration)

Create a dict for the graph representation for the atom nodes. The output is a dictionary of tensors that is used as input to the DeepDFT model.

Parameters:

configuration – The bulk configuration to create the graph from.

Returns:

The dictionary of torch.tensor for atom nodes, edges, etc.

copyHyperparameters(model)

Copy the hyperparameters from another model.

Parameters:

model – The model to copy the hyperparameters from.

cutoff()

Return the cutoff radius.

gaussianSigma()

Return the sigma of the Gaussian for convolution with density (for range separation).

classmethod initializeFromGridValuesModel(model)

Initialize the DeepDFTModel from a GridValuesModel, using its model directory and hyperparameters within.

Parameters:

model – The GridValuesModel to initialize from.

Returns:

A DeepDFTModel initialized from the given GridValuesModel.

Return type:

DeepDFTModel

initializeModel()

Initialize an internal DeepDFT model class for training.

modelDir()

Get the directory containing the model files.

Returns:

The model directory.

Return type:

str

modelType()

Return the type of the model (SchNet/PaiNN).

nlinfo()
Returns:

The nlinfo.

Return type:

dict

nodeSize()

Return the node size the model was trained with.

numComponents()

Return the number of (spin) components the model was trained with.

numInteractions()

Return the number of interactions the model was trained with.

predictGridValuesWithModel(configuration, grid_descriptor, probe_count, gpu_acceleration)

Use the defined model to predict the grid property of a configuration that the model was trained for.

Parameters:

  • configuration – The configuration for which to predict the grid values.

  • grid_descriptor – The grid descriptor to use for the prediction.

  • probe_count – Number of probe graph nodes to use.

  • gpu_acceleration – Flag to indicate whether to use GPU acceleration.

Returns:

The predicted grid values.

rangeType()

Return the range type for range-separated models (“short-range” or “long-range”, or None).

saveHyperparameters()

Save the hyperparameters of the model to the json file in the model directory

scalingFactor()

Return the scaling factor used for the targets during the training of the model.

static torchDevice(gpu_acceleration)

Resolve the flag for GPU acceleration and make a torch device.

:rtype : torch.device

uniqueString()

Return a unique string representing the state of the object.