peal.training.criterions.FocalLoss

class peal.training.criterions.FocalLoss

Bases: Module

Binary focal loss on logits (Lin et al., 2017).

Parameters:

• alpha (float, optional) – Constant weight of every term (default 0.25).

• gamma (float, optional) – Focusing exponent; 0 recovers plain BCE (default 1).

• reduction ({"mean", "sum", "none"}, optional) – How the per-element losses are aggregated.

Methods

forward

Return alpha * (1 - p_t) ** gamma * BCE reduced as configured.

__init__(alpha=0.25, gamma=1, reduction='mean')

Store the focal-loss hyperparameters.

forward(inputs, targets)

Return alpha * (1 - p_t) ** gamma * BCE reduced as configured.