cayleypy.train.PinballLoss

class cayleypy.train.PinballLoss(tau: float = 0.5)

Pinball (quantile) loss, which penalizes underestimation and overestimation differently.

A model trained with this loss predicts the tau-quantile of the target distribution instead of its mean. With tau > 0.5, underestimating a distance costs more than overestimating it by the same amount; with tau < 0.5, the other way round, which pushes predictions towards a lower bound on the true distance. At tau = 0.5 this is exactly half of the mean absolute error.

Example:

>>> import torch
>>> from cayleypy.train import PinballLoss
>>> loss = PinballLoss(0.9)
>>> float(loss(torch.tensor([1.0, 3.0]), torch.tensor([2.0, 2.0])))
0.5

__init__(tau: float = 0.5)

Initializes PinballLoss.

Parameters:

tau – Quantile of the target distribution to predict. Must be strictly between 0 and 1.

Methods

__init__([tau])	Initializes PinballLoss.
elementwise(predictions, targets)	Computes pinball loss for every element.

elementwise(predictions: Tensor, targets: Tensor) → Tensor

Computes pinball loss for every element.

Parameters:

• predictions – Predicted distances.

• targets – Target distances, of the same shape as predictions.

Returns:

Tensor of the same shape as predictions, with pinball losses - tau*(target-prediction) where the target is underestimated, and (1-tau)*(prediction-target) where it is overestimated.