(This is in continuation to my earlier question here.)
Suppose I have a custom layer which computes the loss for me, using external trainable variables using TF
