Tensorflow Gradients do not exist for near identical custom losses

Question

I am trying to code the custom loss functions in this paper, in particular the MFTR and MCFD on page 6. The application being to use this custom loss in an LSTM for time ser