How to caculate the Dice loss for one-hot encode label , espacially for multi-label class?

1

My Dice loss for multi-segment in training processing graually converge to a negative number. There is any problem?