Overview

• This is an implementation of multi-class focal loss in PyTorch.

Brief description

• This loss function generalizes multiclass cross-entropy by introducing a hyperparameter gamma(focusing parameter) that allows to focus on hard examples. The Focal loss: $FL(p_t)=-\alpha_t(1-p_t)^\gamma\log(p_t)$, where $\alpha_t$ is a weighting facrot, $p_t$ is a model's estimated probability, $\gamma$ is a focusing parameter and $-\log(p_t)$ is the cross entropy loss in this case.

Prerequisites

• torch

Parameters

• gamma(int): The focusing parameter (Must be non-negative).
• weight(Tensor, Optional): Weighting factor for each of the n classes.

References

• Focal loss: Lin, T. Y., Goyal, P., Girshick, R., He, K., & Dollár, P. (2017). Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision (pp. 2980-2988).