VariationRatioCriterion#

class torch_uncertainty.ood_criteria.VariationRatioCriterion[source]#

OOD criterion based on variation ratio.

This criterion computes the variation ratio from ensemble predictions. Higher variation ratio values indicate greater uncertainty.

Given ensemble predictions where \(n_{\text{mode}}\) is the count of the most frequently predicted class among \(K\) predictions, the variation ratio is computed as:

\[\text{VR} = 1 - \frac{n_{\text{mode}}}{K}\]
Variables:

  • ensemble_only (bool) – Requires ensemble predictions.

  • input_type (OODCriterionInputType) – Expected input type is estimated probabilities.

forward(inputs)[source]#

Compute variation ratio from ensemble predictions.

Parameters:

inputs (Tensor) – Tensor of ensemble probabilities with shape (ensemble_size, batch_size, num_classes).

Returns:

Variation ratio for each sample.

Return type:

Tensor