Multiple inputs in seq2seq

I want to do beam search while testing seq2seq. So say, if the top 2 labels are [0,1] i want to feed them to the next decoder lstm’s input.


But now i want to feed multiple inputs at once ,so that i can have softmax over vocab for all relevant inputs in one go. But then input and hidden dimension is a mismatch.


input = [0] hidden =(1X1XDIM, 1X1XDIM)

Now i want

input = [1,2] hidden = ?

Once i get this, i will have the softmax over the vocab for both the inputs and can set the backpointer for each word in vocab and continue