neuralop.losses.data_losses.PointwiseQuantileLoss

class neuralop.losses.data_losses.PointwiseQuantileLoss(alpha, reduction='sum')[source]

PointwiseQuantileLoss computes Quantile Loss.

It is used for uncertainty quantification of Neural Operators, as described in [1].

Parameters:

alphafloat

value, between 0 and 1, of the proportion of points in the output domain expected to fall within predicted quantiles

reductionstr, optional

whether to reduce across the batch and channel dimensions by summing (‘sum’) or averaging (‘mean’)

.. warning::

PointwiseQuantileLoss always averages over the spatial dimensions. reduction only applies to the batch and channel dimensions.

Methods

__call__(y_pred, y[, eps])	y_pred : torch.tensor
reduce_all(x)	reduce x across the batch according to self.reduction

References

[1]

Ma, Z., Pitt, D., Azizzadenesheli, K., Anandkumar, A., (2024). Calibrated Uncertainty Quantification for Operator Learning via Conformal Prediction TMLR 2024, https://openreview.net/pdf?id=cGpegxy12T

reduce_all(x)[source]

reduce x across the batch according to self.reduction

Parameters:

x: torch.Tensor

inputs