Tensorflow lambda layer to differentiate predictions across batches

Question

I am working on a LSTM Recurrent Neural Network and want to differentiate the outputs/predictions of the last dense layer I have.

Consider a simple network-