Move TRT args, params to config file instead of hardcoding them.

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

Move TRT args, params to config file instead of hardcoding them. #281

Closed trevor-m closed 5 years ago

trevor-m commented 5 years ago

Also allows TRT infer to work when using horovod. There are still bugs in TF-TRT which are preventing most models from gaining any speedup, however progress is being made.

These changes will also allow us to include the models in our QA to track progress.