Defining class weights for BCE with multi-channel gaussian masks

How to define class weights for multiclass BINARY_CROSS_ENTROPY_WITH_LOGITS with gaussian masks as targets?