Tensorflow 2 custom loss return nan

Question

I have a model, I compile it using binary_crossentropy, the training process goes well, the loss is printed.

model = MyModel()
model.compile(opti