credal_dro

probly.method.credal_dro(base: Predictor[In, Out], num_members: int, reset_params: bool = True) → CredalDROPredictor[In, Out][source]

Create a credal DRO predictor from a base predictor based on [WFC+26].

Structurally identical to credal_wrapper; the methods differ only in training. Member i is trained on the CVaR cross-entropy at level credal_dro_deltas(delta_g, num_members)[i] (see probly.losses.torch.cvar_ce_loss).

Parameters:

  • base – The base classifier to replicate into an ensemble.

  • num_members – Number of ensemble members.

  • reset_params – Whether to reset the parameters of each member. Default is True.

Returns:

The credal DRO predictor outputting a ProbabilityIntervalsCredalSet.

Examples using probly.method.credal_dro

Credal DRO Output Visualization

Credal DRO Output Visualization