about circle loss

I modified the get_symbol() function in train.py in recognition in order to train recognition model with circle loss, with reference to another repository https://github.com/zjysnow/CircleLoss , but the training result is stange, when m=0.25,s=64, the acc is always 0, and when the loss value didn't decrese any longer, the ver acc of lfw/cfp/agedb is still low(99.03/87.88/91.93). What‘s even worse, when I set s=256, the loss value keep always 'nan'. Can anyone help me to see what's the problem of my code? Thanks sooo much.