autoemulate.emulators.base

class Emulator(x=None, y=None, **kwargs)

Bases: ABC, ValidationMixin, ConversionMixin, TorchDeviceMixin

Base class for all emulators.

This class provides the basic structure and methods for emulators in AutoEmulate. It includes methods for fitting, predicting, and handling device management.

is_fitted_ = False

supports_grad = False

scheduler_cls = None

x_transform = None

y_transform = None

supports_uq = False

fit(x, y)

Fit the emulator to the provided data.

classmethod model_name()

Return the full name of the model.

classmethod short_name()

Return a short name for the model.

Take the capital letters of the class name and return them as a lower case string. For example, if the class name is GaussianProcess, this will return gp.

predict(x, with_grad=False)

Predict the output for the given input.

Parameters:

• x (TensorLike) – Input tensor to make predictions for.

• with_grad (bool) – Whether to enable gradient calculation. Defaults to False.

Returns:

The predicted output.

Return type:

OutputLike

predict_mean(x, with_grad=False, n_samples=100)

Predict the mean of the target variable for input x.

Parameters:

• x (TensorLike) – Input tensor of shape (n_batch, n_features) for which to predict the mean.

• with_grad (bool) – Whether to compute gradients with respect to the input. Defaults to False.

• n_samples (int) – Number of samples to draw when using sampling-based predictions. Defaults to 100.

Returns:

Mean tensor of shape (n_batch, n_targets).

Return type:

TensorLike

predict_mean_and_variance(x, with_grad=False, n_samples=100)

Predict the mean and variance of the target variable for input x.

Parameters:

• x (TensorLike) – Input tensor of shape (n_batch, n_features) for which to predict the mean and variance.

• with_grad (bool) – Whether to compute gradients with respect to the input. Defaults to False.

• n_samples (int) – Number of samples to draw when using sampling-based predictions. Defaults to 100.

Returns:

A tuple containing: - Mean tensor of shape (n_batch, n_targets). - Variance tensor of shape (n_batch, n_targets) if model supports UQ

otherwise None.

Return type:

tuple[TensorLike, TensorLike | None]

abstract static is_multioutput()

Flag to indicate if the model is multioutput or not.

static get_tune_params()

Return a dictionary of hyperparameters to tune.

The keys in the TuneParams must be implemented as keyword arguments in the __init__ method of any subclasses.

e.g.

tune_params: TuneParams = {

“lr”: list[0.01, 0.1], “batch_size”: [16, 32], “mean”

}

model_params: ModelParams = {

“lr”: 0.01, “batch_size”: 16

}

class MySubClass(Emulator):

def __init__(lr, batch_size):

self.lr = lr self.batch_size = batch_size

classmethod get_random_params()

Return a random set of params for the model.

classmethod scheduler_params()

Return a random parameters for the learning rate scheduler.

This should be added to the get_tune_params() method of subclasses to allow tuning of the scheduler parameters.

scheduler_setup(kwargs=None)

Set up the learning rate scheduler for the emulator.

Parameters:

kwargs (dict | None) – Keyword arguments for the model. This should include scheduler_kwargs.

class DeterministicEmulator(x=None, y=None, **kwargs)

Bases: Emulator

A base class for deterministic emulators.

supports_uq = False

predict(x, with_grad=False)

Predict the output for the given input.

Parameters:

• x (TensorLike) – Input tensor to make predictions for.

• with_grad (bool) – Whether to enable gradient calculation. Defaults to False.

Returns:

The emulator predicted output for x.

Return type:

TensorLike

predict_mean_and_variance(x, with_grad=False, n_samples=100)

Predict the mean and variance of the target variable for input x.

Parameters:

• x (TensorLike) – Input tensor of shape (n_batch, n_features) for which to predict the mean and variance.

• with_grad (bool) – Whether to compute gradients with respect to the input. Defaults to False.

• n_samples (int) – Number of samples to draw when using sampling-based predictions. Defaults to 100.

Returns:

A tuple containing: - Mean tensor of shape (n_batch, n_targets). - Variance tensor as None since the model does not support UQ.

Return type:

tuple[TensorLike, None]

class ProbabilisticEmulator(x=None, y=None, **kwargs)

Bases: Emulator

A base class for probabilistic emulators.

supports_uq = True

predict(x, with_grad=False)

Predict the output distribution for the given input.

Parameters:

• x (TensorLike) – Input tensor to make predictions for.

• with_grad (bool) – Whether to enable gradient calculation. Defaults to False.

Returns:

The emulator predicted distribution for x.

Return type:

DistributionLike

predict_mean_and_variance(x, with_grad=False, n_samples=100)

Predict the mean and variance of the target variable for input x.

Parameters:

• x (TensorLike) – Input tensor of shape (n_batch, n_features) for which to predict the mean and variance.

• with_grad (bool) – Whether to compute gradients with respect to the input. Defaults to False.

• n_samples (int) – Number of samples to draw when using sampling-based predictions. Defaults to 100.

Returns:

A tuple containing: - Mean tensor of shape (n_batch, n_targets). - Variance tensor of shape (n_batch, n_targets).

Return type:

tuple[TensorLike, None]

class GaussianEmulator(x=None, y=None, **kwargs)

Bases: ProbabilisticEmulator

A base class for Gaussian emulators.

supports_grad = True

predict(x, with_grad=False)

Predict the Gaussian distribution for the given input.

Parameters:

• x (TensorLike) – Input tensor to make predictions for.

• with_grad (bool) – Whether to enable gradient calculation. Defaults to False.

Returns:

The emulator predicted Gaussian distribution for x.

Return type:

GaussianLike

class GaussianProcessEmulator(x=None, y=None, **kwargs)

Bases: GaussianEmulator

A base class for Gaussian Process emulators.

predict(x, with_grad=False)

Predict the Gaussian distribution for the given input.

Parameters:

• x (TensorLike) – Input tensor to make predictions for.

• with_grad (bool) – Whether to enable gradient calculation. Defaults to False.

Returns:

The emulator predicted Gaussian distribution for x.

Return type:

GaussianLike

class PyTorchBackend(*args, **kwargs)

Bases: Module, Emulator

PyTorchBackend provides a backend for PyTorch models.

The class provides the basic structure and methods for PyTorch-based emulators to enable further subclassing and customization. This provides default implementations to simplify model-specific subclasses by only needing to implement:

• .__init__(): the constructor for the model

• .forward(): the forward pass of the model

• .get_tune_params(): the hyperparameters to tune for the model

batch_size = 16

shuffle = True

epochs = 10

loss_history = []

verbose = False

loss_fn = MSELoss()

optimizer_cls

alias of Adam

optimizer

supports_grad = True

lr = 0.1

scheduler_cls = None

supports_uq = False

loss_func(y_pred, y_true)

Loss function to be used for training the model.

class SklearnBackend(x=None, y=None, **kwargs)

Bases: DeterministicEmulator

SklearnBackend provides a backend for sklearn models.

The class provides the basic structure and methods for sklearn-based emulators to enable further subclassing and customization. This provides default implementations to simplify model-specific subclasses by only needing to implement:

• .__init__(): the constructor for the model

• .get_tune_params(): the hyperparameters to tune for the model

model

normalize_y = False

y_mean

y_std

supports_grad = False

class DropoutTorchBackend(*args, **kwargs)

Bases: PyTorchBackend

PyTorch backend model that is able to support dropout.