Label Smoothing Regularization

[lorinczszabolcs]

Support label smoothing regularization

Motivation
Label smoothing has been shown to boost performance in Inception-V2, as well as in Revisiting ResNets papers.

Related resources
PyTorch Cross-Entropy Loss already supports it, by passing label_smoothing parameter between 0.0 and 1.0. The Cross-Entropy implementation in mmsegmentation could take additional kwargs and pass them forward to the pytorch functions.

[MengzhangLI]

Hi, it is a good suggestion. We would consider to add this feature in cross entropy loss.