Hey :)
The figures in section Explained: 1. Controller raise some questions:
I can't find an explanation what the inputs of the LSTM cells should be. You use g_emb, Anchors and w_emb. Which values do these variables have and why do you alternate them?
Why do you run seperate LSTM cells just to fill anchors instead of reusing the hidden states next_h, when determining index and operation?
Why does b_soft have dimension [64,5] and b_soft_no_learn dimension [64,1], and if they are not bias vectors, which role do they play?
Hey :) The figures in section Explained: 1. Controller raise some questions: I can't find an explanation what the inputs of the LSTM cells should be. You use g_emb, Anchors and w_emb. Which values do these variables have and why do you alternate them?
Why do you run seperate LSTM cells just to fill anchors instead of reusing the hidden states next_h, when determining index and operation?
Why does b_soft have dimension [64,5] and b_soft_no_learn dimension [64,1], and if they are not bias vectors, which role do they play?