Bahdanau Decoder Implementation

[sidoki]

Hi @spro,

Thanks for really great explanation of decoder, especially for Bahdanau decoder. But, i'm little bit confuse about code in init function of BahdanauAttnDecoderRNN class.

self.attn = GeneralAttn(hidden_size)

I can't find any class that define GeneralAttn. This is built-in class? Can you please elaborate for this? Thanks again!

[spro]

Good catch, it was originally split out as 3 separate attention modules (GeneralAttn, DotAttn, ConcatAttn) instead of one with an argument to choose the strategy. Further, they actually used the "concat" strategy. So this should be self.attn = Attn("concat", hidden_size)

[sidoki]

Cool, Thanks for the clarification!

[rafaelvalle]

Can you please change that line on the notebook?

[poweihuang17]

Still not changed. Hope somebody could do it.

[anantzoid]

#119 fixes this and some more issues with Bahdanau decoder.

[kyquang97]

@anantzoid it's still not fixes in tutorial.