Gluon-nlp: [Discussion] Support Multiple Backends

[sxjscience]

Many DL frameworks are moving to use numpy as their default way for writing DL models. Thus, we may try to support multiple-backends for the model-zoo in GluonNLP. As long as the backend supports numpy-like interface, e.g., JAX , tf.experimental.numpy, and the ongoing effort in PyTorch: https://github.com/pytorch/pytorch/issues/38349, we are able to reuse most of the existing model implementations and make it work for different DL frameworks. One advantage is that the user is able to try out TF-2.0, Jax and PyTorch very easily.

What do you think? @dmlc/gluon-nlp-committers

[leezu]

We can extend the Gluon interface/create a independent modeling library to enable switching backends

[sxjscience]

I think after some rounds of discussion, the plan will be to add an additional PyTorch-based backend in GluonNLP.

[sxjscience]

Why close this?

[szha]

As you stated above that the discussion has concluded that we will support pytorch, I thought we can close the discussion thread and start listing the tasks.

[szha]

Feel free to reopen if you think this decision needs to be revisited with more discussion.