credit.losses.kcrps

credit.losses.kcrps#

Classes#

KCRPSLoss

Adapted from Nvidia Modulus.

Module Contents#

class credit.losses.kcrps.KCRPSLoss(reduction, biased: bool = False)#

Bases: torch.nn.Module

Adapted from Nvidia Modulus.

Estimate the CRPS from a finite ensemble

Computes the local Continuous Ranked Probability Score (CRPS) by using the kernel version of CRPS. The cost is O(m log m).

Creates a map of CRPS and does not accumulate over lat/lon regions. Approximates:

\[CRPS(X, y) = E[X - y] - 0.5 E[X-X']\]

with

\[sum_i=1^m |X_i - y| / m - 1/(2m^2) sum_i,j=1^m |x_i - x_j|\]
Parameters:

  • pred (Tensor) – Tensor containing the ensemble predictions. The ensemble dimension is assumed to be the leading dimension

  • obs (Union[Tensor, np.ndarray]) – Tensor or array containing an observation over which the CRPS is computed with respect to.

  • biased (bool) –

    When False, uses the unbiased estimators described in (Zamo and Naveau, 2018):

    E|X-y|/m - 1/(2m(m-1)) sum_(i,j=1)|x_i - x_j|

Unlike ``crps`` this is fair for finite ensembles. Non-fair ``crps`` favors less
dispersive ensembles since it is biased high by E|X- X'|/ m where m is the
ensemble size.

biased = False#
batched_forward#
forward(target, pred)#
single_sample_forward(target, pred)#

Forward pass for KCRPS loss for a single sample.

Parameters:

  • target (torch.Tensor) – Target tensor.

  • pred (torch.Tensor) – Predicted tensor.

Returns:

CRPS loss values at each lat/lon

Return type:

torch.Tensor

_kernel_crps_implementation(pred: torch.Tensor, obs: torch.Tensor, biased: bool) → torch.Tensor#

An O(m log m) implementation of the kernel CRPS formulas