Mulitple translations (top-K) for a given input sentence in transformer Text2Text

[repoloper]

Description

In the transformer-translation task, is there a way to get multiple alternate translations?

I've been trying to get multiple translations for an input sentence. I have set the num_decodes variable to 5 to no effect. I am using decode-from-dataset. The command I am using is along the lines of:

t2t-decoder --t2t_usr_dir ./usrdir --data_dir ~/datadir --problem translate_ende_wmt32k
 --model transformer --hparams_set transformer_base --output_dir ~/outputdir --decode_to_file translations.en 
--decode_hparams beam_size=8,alpha=0.9,num_decodes=5

I see that it is being used in the code here, but I don't understand if decode_once means that there's only one output coming out of the decoder even if there is a loop.

https://github.com/tensorflow/tensor2tensor/blob/master/tensor2tensor/utils/decoding.py#L173

Can provide more information as needed. ...

Environment information

OS: Ubuntu

$ pip freeze | grep tensor
tensor2tensor==1.7.0
tensorboard==1.8.0
tensorflow==1.8.0

$ python -V
Python 3.5.2

[zealseeker]

Refer to #1334

t2t-decoder \
   --data_dir=$DATA_DIR \
   --problem=$PROBLEM \
   --model=$MODEL \
   --hparams_set=$HPARAMS \
   --output_dir=$TRAIN_DIR \
   --decode_hparams="beam_size=$BEAM_SIZE,alpha=$ALPHA,return_beams=True" \
   --t2t_usr_dir=$USR_DIR

return_beams=True is important to get the alternatives. beam_size is the number of examples you want to get.