Open adakum opened 7 years ago
I have the exact same question, have you figured it out yet?
@adakum @johnsonice
There is a bug fixes for beam search in tensorflow (https://github.com/tensorflow/tensorflow/commit/18f89c81d288f191abd56501ec6f86fe29265bdd), maybe you can try to see if the top beam gets better result?
You will need to install the latest tf-nightly.
In
decode_and_evaluate
function inutils/nmt_utils.py
, in case ofbeam_width > 0
, I see thenmt_outputs=nmt_outputs[0]
. LinkIs
nmt_outputs[0]
contains the best translations ornmt_outputs[beam_width-1]
?After training my models I see in general
nmt_outputs[beam_width-1]
contains better output thatnmt_outputs[0].