ott.tools.conformal.OTCP

class ott.tools.conformal.OTCP(model, nonconformity_fn=<built-in function sub>, sinkhorn_output=None, sampler=None, offset=0.0, scale=1.0, calibration_scores=None)

Optimal transport conformal prediction [Klein et al., 2025].

Parameters:

• model (Callable[[Array], Array]) – Fitted model.

• nonconformity_fn (Callable[[Array, Array], Array]) – Multivariate nonconformity score function with a signature (target, prediction) -> score.

• sinkhorn_output (Optional[SinkhornOutput]) – Sinkhorn output computed in fit_transport().

• sampler (Optional[Callable[[PRNGKey, Tuple[int, int]], Array]]) – sampler function used to sample points from a reference measure. sobol_ball_sampler() is used by default.

• offset (Array) – Offset used when re-scaling the data.

• scale (Array) – Scale when re-scaling the data.

• calibration_scores (Optional[Array]) – Nonconformity calibration scores computed in calibrate().

Methods

calibrate(x, y)	Compute calibration scores.
fit_transport(x, y[, epsilon, n_target, ...])	Fit the transport map.
get_scores(x, y)	Compute nonconformity scores.
nonconformity_fn(b, /)	Same as a - b.
predict(x[, y_candidates, alpha])	Conformalize the model's prediction.

Attributes

calibration_scores
offset
sampler
scale
sinkhorn_output
target_measure	Target measure of shape [n_target, dim_y].
model