Is there a way of ignoring zero values in y_true such that y_pred can learn from non-zero values?

Question

I\'m building a NN that creates a soundwave-esque sequence given a text input. The problem is that y_true requires padding in some cases, and thus the majority of the true s