incerto.conformal.aps

incerto.conformal.aps(model, calib_loader, alpha)

Adaptive Prediction Sets (APS) — Romano, Patterson, and Candes, NeurIPS 2020.

Produces variable-sized sets by thresholding randomized cumulative probability mass, calibrated on held-out data. Randomized scores yield tight (non-conservative) coverage guarantees.

Parameters:

• model (Module)

• calib_loader (DataLoader)

• alpha (float)

Return type:

Callable[[Tensor], List[Tensor]]