Conformal Prediction API¶

deepuq.methods.conformal ¶

Conformal Prediction methods for distribution-free uncertainty quantification.

AdaptiveConformalPredictor ¶

Online adaptive conformal predictor with coverage guarantees.

Parameters:

Name	Type	Description	Default
`model`	`Module`	Trained model.	required
`target_coverage`	`float`	Target coverage level (default 0.9).	`0.9`
`gamma`	`float`	Step size for online threshold updates.	`0.01`

calibrate ¶

calibrate(cal_X: Tensor, cal_y: Tensor) -> None

Initial calibration on held-out data.

predict_set ¶

predict_set(x: Tensor) -> tuple[torch.Tensor, torch.Tensor]

Predict intervals using current threshold.

update ¶

update(x_new: Tensor, y_new: Tensor) -> None

Online update of threshold based on observed coverage.

Parameters:

Name	Type	Description	Default
`x_new`	`Tensor`		required
`y_new`	`Tensor`		required

CQRPredictor ¶

Bases: BaseConformalPredictor

Adaptive conformal intervals using quantile regression.

Requires a model that outputs (lower_quantile, upper_quantile) as a tensor of shape (N, 2) or a tuple.

ConformalClassifier ¶

Bases: BaseConformalPredictor

Prediction sets with marginal coverage guarantee.

Supports Adaptive Prediction Sets (APS) and Regularized APS (RAPS).

ConformalUQWrapper ¶

Bases: BaseConformalPredictor

Calibrate intervals from any predict_uq()-compatible method.

Wraps Laplace, Ensembles, MC Dropout, etc. to guarantee coverage.

SplitConformalRegressor ¶

Bases: BaseConformalPredictor

Distribution-free prediction intervals for any regression model.

Guarantees P(Y in [lower, upper]) >= 1 - alpha for exchangeable data.

WeightedConformalPredictor ¶

Conformal predictor with importance-weighted calibration.

Parameters:

Name	Type	Description	Default
`model`	`Module`	Trained model (callable on input tensors).	required
`score_fn`	`Callable[..., Tensor] \| None`	Callable `(model, x, y) -> scores`. Defaults to absolute residual.	`None`

calibrate ¶

calibrate(
    cal_X: Tensor, cal_y: Tensor, weights: Tensor
) -> None

Calibrate using weighted nonconformity scores.

Parameters:

Name	Type	Default
`cal_X`	`Tensor`	required
`cal_y`	`Tensor`	required
`weights`	`Tensor`	required

predict_set ¶

predict_set(
    x: Tensor, alpha: float = 0.1
) -> tuple[torch.Tensor, torch.Tensor]

Compute prediction intervals.

Parameters:

Name	Type	Description	Default
`x`	`Tensor`		required
`alpha`	`float`		`0.1`

Returns:

Type	Description
`Tuple of (lower, upper) bounds.`

absolute_residual_score ¶

absolute_residual_score(
    y_pred: Tensor, y_true: Tensor
) -> torch.Tensor

Absolute residual: |y - y_hat|.

check_coverage ¶

check_coverage(
    y_true: Tensor, lower: Tensor, upper: Tensor
) -> float

Compute empirical coverage fraction.

conformal_quantile ¶

conformal_quantile(scores: Tensor, alpha: float) -> float

Compute the conformal quantile of nonconformity scores.

Returns the ceil((n+1)*(1-alpha))/n quantile of the scores.

normalized_residual_score ¶

normalized_residual_score(
    y_pred: Tensor, y_true: Tensor, sigma: Tensor
) -> torch.Tensor

Normalized absolute residual: |y - y_hat| / sigma.

quantile_score ¶

quantile_score(
    y_true: Tensor, q_lo: Tensor, q_hi: Tensor
) -> torch.Tensor

CQR score: max(q_lo - y, y - q_hi).

signed_residual_score ¶

signed_residual_score(
    y_pred: Tensor, y_true: Tensor
) -> torch.Tensor

Signed residual: y - y_hat.