Should I use OpenNMT-py or OpenNMT-tf

Hello, I want to build model for language translation and there are two version of OpenNMT (based on pytorch and tenserflow).

In this video, a people say that OpenNMT.py is oriented for research and OpenNMT.tf is oriented on production but this page Features - OpenNMT say that Copy attention and Coverage attention is only aviable on OpenNMT.py.

So we can deduce that for better accuracy on translation, is it recommended to use OpenNMT.py but for production, we can use OpenNMT.tf but with less accuracy (but more easy to use).

Do I understand the difference between these two toolkits?
Sincerely.

Hi,

For machine translation, both frameworks will reach a similar accuracy when training a Transformer model with the recommended settings.

In most cases your choice is about PyTorch vs. TensorFlow and which one seems simpler to use and integrate for you.

These days you don’t even need to consider how to run inference with these frameworks, because you’ll probably want to use CTranslate2 instead.

3 Likes

Okay, thank you for your answer!