train_eval for non-Horovod multi-gpu broken for transformer

NVIDIA / OpenSeq2Seq

Toolkit for efficient experimentation with Speech Recognition, Text2Speech and NLP

https://nvidia.github.io/OpenSeq2Seq

Apache License 2.0

1.54k stars 369 forks source link

Closed borisgin closed 5 years ago

borisgin commented 5 years ago

train_eval mode does not work in multi-gpu non horovod mode: transformer-big script exits right after very first evaluation; step remains 0

okuchaiev commented 5 years ago

done