LSTM+Attention in Keras Error when Initialising states to zeros

Question

I am trying to implement this paper specifically, the Encoder with input attention section. In essence, this is manipulating the input sequence with attention b