incerto.calibration.get_uncertainty_from_evidence

incerto.calibration.get_uncertainty_from_evidence#

incerto.calibration.get_uncertainty_from_evidence(evidence, num_classes)[source]#

Compute uncertainty measures from evidential outputs.

Parameters:

  • evidence (Tensor) – Non-negative evidence values (N, C)

  • num_classes (int) – Number of classes

Returns:

  • alpha: Dirichlet parameters (N, C)

  • belief: Predicted probabilities (N, C)

  • uncertainty: Total uncertainty / vacuity (N, 1)

  • epistemic: Epistemic uncertainty (N, 1)

Return type:

Dictionary with

Example

>>> evidence = F.softplus(model(x))
>>> uncertainty = get_uncertainty_from_evidence(evidence, num_classes=10)
>>> total_unc = uncertainty['uncertainty']
Contents