No such file or directory: 'wmt14_en_de/tmp/dev.de.ori'? about flowseq HOT 7 CLOSED

Please contact fairseq for this issues.
Thanks.

from flowseq.

@XuezheMax - this is not a fairseq issue - prepare-wmt14en2de.sh does not create any files called *.ori. Because of the way these files are prepared by moses, it is difficult to recreate them (the bpe files are cleaned, meaning we cannot just use the raw {train,dev,test}.{en, de} files). Would you be able to share how you produced these files?

from flowseq.

In the meantime, I first cleaned the validation data (without peforming bpe segmentation on it) using this line https://github.com/pytorch/fairseq/blob/master/examples/translation/prepare-wmt14en2de.sh#L138

and then ran the BPE encoding script. This meant that both the word tokenized version and the bpe tokenized version get the same data filtered out (as it was done before). I then just copied the files to valid.{lang}.ori. I don't know if this is officially correct, because the data that is filtered is different depending on whether you run the filtering script before or after doing BPE, but 🤷‍♂

from flowseq.

@DeNeutoy *.ori are just the original tokenized files, we use them to compute the BLEU score. You can just put your original dev/test original tokenized files in the corresponding directories.

from flowseq.

@violet-zct you actually can't do that, I don't think - because the prepare-wmt14en2de.sh script does two things,

1. BPE encodes {train, dev, test}.
2. cleans the BPE encoded files, using the moses script, which reduces the number of examples in {train,dev}.

So, once the final dataset is produced, it does not have the same number of dev/test data points as in the original (because it has been cleaned), meaning that just putting the original files in the right place doesn't work.

from flowseq.

@DeNeutoy You can simply sed -r 's/(@@ )|(@@ ?$)//g' to restore the original segmentation of ``the '' dev file you are using from the BPE version.

from flowseq.

Oh great, thanks!

from flowseq.