kliff.legacy.calculators.calculator_torch

class kliff.legacy.calculators.calculator_torch.CalculatorTorch(model, gpu=None)

A calculator for torch based models.

Parameters:

• model (ModelTorch) – torch models, e.g. NeuralNetwork.

• gpu (Union[bool, int, None]) – whether to use gpu for training. If int (e.g. 0), will trained on this gpu device. If True will always train on gpu 0.

implemented_property = ['energy', 'forces', 'stress']

create(configs, use_energy=True, use_forces=True, use_stress=False, fingerprints_filename='fingerprints.pkl', fingerprints_mean_stdev_filename=None, reuse=False, use_welford_method=False, nprocs=1)

Process configs to generate fingerprints.

Parameters:

• configs (List[Configuration]) – atomic configurations

• use_energy (bool) – Whether to require the calculator to compute energy.

• use_forces (bool) – Whether to require the calculator to compute forces.

• use_stress (bool) – Whether to require the calculator to compute stress.

• fingerprints_filename (Union[Path, str]) – Path to save the generated fingerprints. If reuse=True, Will not generate the fingerprints, but directly use the one provided via this file.

• fingerprints_mean_stdev_filename (Union[str, Path, None]) – Path to save the mean and standard deviation of the fingerprints. If reuse=True, Will not generate new fingerprints mean and stdev, but directly use the one provided via this file. If normalize is not required by a descriptor, this is ignored.

• reuse (bool) – Whether to reuse provided fingerprints.

• use_welford_method (bool) – Whether to compute mean and standard deviation using the Welford method, which is memory efficient. See https://en.wikipedia.org/wiki/Algorithms_for_calculating_variance

• nprocs (int) – Number of processes used to generate the fingerprints. If 1, run in serial mode, otherwise nprocs processes will be forked via multiprocessing to do the work.

get_fingerprints()

Return a list of fingerprints of the configurations.

Return type:

List[dict]

get_compute_arguments(batch_size=1)

Return the dataloader with batch size set to batch_size.

set_fingerprints(fingerprints)

Update the fingerprints of the calculator. The fingerprints input argument should be in the same format as the output of meth:~kliff.legacy.descriptors.descriptor.load_fingerprints, which is a list of dictionaries.

Parameters:

fingerprints (List[dict]) – A list of fingerprints.

fit()

compute(batch)

property model

Get the underlying torch model

save_model(epoch, force_save=False)

Save the model to disk.

When to save a model is dependent on epoch and a model’s metadata for save.

Parameters:

• epoch (int) – current optimization epoch.

• force_save (bool) – save the model, ignoring epoch and save metadata.

get_energy(batch)

get_forces(batch)

get_stress(batch)

get_size_opt_params()

Return the size of the parameters.

Returns:

Each element in the list gives the shape of each type of parameter

tensors, containing, e.g., weights and biases, for each layer.

nelements: Number of elements of each parameter tensor. nparams: Total number of parameters

Return type:

sizes

get_num_opt_params()

Return the total number of parameters.

Return type:

int

get_opt_params(flat=True)

Retrieve the parameters, i.e., weights and biases.

Parameters:

flat (bool) – A flag to return a flat, 1D array.

Return type:

Union[List, array]

Returns:

Parameters, i.e., weights and biases. If flat=True, a 1D np.ndarray will be returned. Otherwise, nested lists will be returned, where each list contain the weights and biases for each layer.

update_model_params(parameters)

Update the model parameters from a 1D array.

Parameters:

parameters (array) – New parameter values to set. It needs to be a 1D array.

class kliff.legacy.calculators.calculator_torch.CalculatorTorchSeparateSpecies(models, gpu=None)

A calculator supporting models of difference species.

Parameters:

• models (Dict[str, NeuralNetwork]) – {species: model} with species specifying the chemical symbol for the model.

• gpu (Union[bool, int, None]) – whether to use gpu for training. If int (e.g. 0), will trained on this gpu device. If True will always train on gpu 0.

compute(batch)

property model

Get the underlying torch model

save_model(epoch, force_save=False)

Save the models to disk.

When to save a model is dependent on epoch and a model’s metadata for save.

Parameters:

• epoch (int) – current optimization epoch.

• force_save (bool) – save the model, ignoring epoch and save metadata.

exception kliff.legacy.calculators.calculator_torch.CalculatorTorchError(msg)