loss nan bug

[wade0604]

loss = loss.sum([1,2,3,4])/masks.sum([1,2,3,4])
if masks.sum([1,2,3,4]) contains 0, then the loss will be NaN
so i simply change the upper code to loss = loss.sum([1,2,3,4])/(masks.sum([1,2,3,4])+1e-6), Is this appropriate?