bidirectional_readout

[parshakova]

why do you need bidirectional_readout? because in the paper they use outputs from attention_match_rnn() directly as the input to pointer_network()

[ghost]

Hi @digitnumber, if you look at the original paper, they added an extra bidirectional layer before the pointer network. Please check figure 1 of the paper. " After the original self-matching layer of the passage, we utilize bi-directional GRU to deeply integrate the matching results before feeding them into answer pointer layer. It helps to further propagate the information aggregated by self-matching of the passage." from section 4.2 Main Results.

[parshakova]

Thanks for a quick reply! Could you please also explain why you used indef call() instead ofdef __call__() in gated_attention_Wrapper

because I was getting an error __call__ raise NotImplementedError("Abstract method")

[ghost]

I think RNN cell instance in tf has def call() function. But not sure, changed to __call__ thanks!