python version: python3.6
os version: Ubuntu 16.04 Kernel 4.4.0-81-generic
Objective：focal_loss.py is created for caculate pixel-wise focal loss
Attention: the code hasn't been used to train network. Maybe there are some bugs!
Paper:https://arxiv.org/pdf/1708.02002.pdf
Thanks to https://github.com/clcarwin/focal_loss_pytorch
How to test: run "python3 focal_loss.py" in shell

ToDo:

• figure out whether FocalLoss is numerically stable
• use the code to train network

Test Result:

• 20171124: The code has been use for tranning network(gamma = 2).The result is about 92% accuracy in local dataset but when I use cross entropy loss the result is about 93% accuracy. So maybe the FocalLoss is unsuitabe for my dataset or the FocalLoss is something wrong. And when I set gamma = 0 for FocalLoss,the trainning dosen't work!