May, 2020: See my updated implementation in github.com/daveboat/dy_common, which adds support for label smoothing.
This is a pytorch implmentation of focal loss (https://arxiv.org/abs/1708.02002), meant to be understandable and easily swappable with nn.CrossEntropyLoss and F.cross_entropy. This implementation includes the ignore_index parameter like pytorch's cross entropy functions.
Everything is in focalloss.py and should be well-commented enough to be self-explanatory. You can either use the regular function version or the nn.Module-wrapped version.