Tensorflow - NaN or Close to Zero loss while training a model with BinaryCrossEntropy

Question

I am trying to implement a model, that will take input as a (q, a) pair where q is Question and a is Answer and both q and a are positional encoded. The output will be how real