Uncertainty Estimators

class alpaca.ue.UE(net, *, nn_runs=25, keep_runs: bool = False)

Abstract class for all uncertainty estimation method implementations

Parameters

• net (torch.nn.Module) – Neural network on based on which we are calculating uncertainty region

• nn_runs – A number of iterations

• keep_runs (bool) – Whenever to save iteration results

Examples

>>> # This could be used to create custom
>>> # uncertainty estimation strategy
>>> class CustomUE(UE):
>>>     def __init__(self, ...):
>>>         ...
>>>     def __call__(self, X_pool: torch.Tensor):
>>>         ...
>>> estimator = CustomUE(model, ...)
>>> predictions, estimations = estimator(x_batch)

last_mcd_runs() → torch.Tensor

Return model prediction for the last uncertainty estimation

class alpaca.ue.MCDUE(*args, num_classes=0, **kwargs)

MCDUE constructor. Depending on the provided num_classes argument, the constructor will initialize MCDUE_regression or MCDUE_classification classes.

Other Parameters

num_classes (int) – Integer that sets the number of classes for prediction

Examples

>>> import alpaca
>>> model : nn.Module = ... # define a torch nn.Model
>>> model = train_model(...) # train the model
>>> estimator = MCDUE(model, nn_runs=100, num_classes=10)
>>> predictions, estimations = estimator(x_batch)

class alpaca.ue.EnsembleMCDUE(*args, acquisition: Optional[Union[str, Callable]] = None, reduction=None, **kwargs)

Estimate uncertainty for samples with Ensemble and MCDUE approach