Should custom loss returns a tensor whose length is the same as the sequence's instead of a scaler when using masking layer?

Question

First of all, I used Keras to build a model. I ask this question because my model cannot be compiled when custom loss returns a scaler, but can be complied when custom loss