Open AmitMY opened 6 years ago
Yes this is something that isn't super easy to do right now. One solution is to compute all your token scores separately and concatenate them in copy_vec
later. Not sure if that's applicable in your case.
This is something that we should add though.
@pmichel31415 that is my current fix, but adding that made my model train 10 times slower on cpu, 60 times slower on gpu.. not 100% sure if it’s this thing or the way I’m calculating the scores, but yeah that sucks
I'm trying to write a seq2seq with out-of-vocab copy, for which I create a zeros input vector in the size of the out-of-vocab tokens:
But after I value each token seperately (conditional copy mechanism) and have a score for the token, I want to add to that word's scrore so I do:
Which is deriveable as it's a sum, and should just change the expression in that dimension to be a sum of expressions, as far as I understand.
Error: