duyvuleo / transformer-dynet Goto Github PK

This is a great repo.
Can this code support multi-GPU training?
I wonder if it can achieve the same performance as tensor2tensor on wmt14-en-de corpus.
Thanks.

when i run that command
./build_gpu/transformer-decode --model-path model/iwslt-envi --beam 5 -T sample-data/tst2013.en.capped
i got that error:
....
Reading test examples from sample-data/tst2013.en.capped
Segmentation fault (core dumped)
...
I find many hours about it. Friend, hope you can write back when you see that. thanks!

Do the reported results of single model (not ensemble) come from single checkpoint or your average checkpoints before translating? Thanks.