How to handle the batch axis in loss function of tensorflow since the shape is always “None”?

Question

I\'m writing a customized loss function in Keras with tensorflow as backend. I found it very difficult to handle the batch axis because it\'s always "None".