RNN-Transducer Loss support

[Times125]

I found that there are no RNN-Transducer loss in the new version. I have to use third party pkg like wrap-transducer So, official support in both tf1.x and tf2 is therefore expected. Thanks！

[dynamicwebpaige]

Though this is potentially out of the scope of TensorFlow core, it might make sense to add to TensorFlow Addons as a loss function.

Rerouting the request to addons; @seanpmorgan for visibility.

[seanpmorgan]

I'm not very familiar with this so I'll defer to @qlzh727 as RNN subpackage owner to evaluate if this makes sense within addons.

[qlzh727]

Sorry for the very late reply.

I am not very familiar with that too. Could u provide a reference to the paper for it?

Also, for the addon package, we are targeting TF 2.0, which means 1.x is probably not supported.

[WindQAQ]

Paper is here https://arxiv.org/abs/1211.3711. It's quite similar to CTC loss.

[Times125]

Sorry for the very late reply.

I am not very familiar with that too. Could u provide a reference to the paper for it?

Also, for the addon package, we are targeting TF 2.0, which means 1.x is probably not supported.

Paper is here https://arxiv.org/abs/1211.3711

[will-rice]

This issue is pretty old, but since it is still open does that mean I could work on an implementation? I work on ASR and would benefit from an implementation with TensorFlow ops. From what I can tell, everyone is still using the warp-transducer version, and the build process adds additional complexity to projects.

[qlzh727]

Sure, any contribution is welcome.

[iamjanvijay]

@will-rice I have an implementation of RNNT (link to the repository) using TensorFlow ops. We can work on this together, if you are interested?

[will-rice]

@iamjanvijay sounds good!

[will-rice]

@iamjanvijay do you want me to just port your ops to the addons loss format and tag you in the PR?

[iamjanvijay]

@will-rice Yes. That can be a good way to start then we can do iterations over it. What do you suggest?

[will-rice]

@will-rice Yes. That can be a good way to start then we can do iterations over it. What do you suggest?

that works. I'll put it together and then tag you.

[seanpmorgan]

TensorFlow Addons is transitioning to a minimal maintenance and release mode. New features will not be added to this repository. For more information, please see our public messaging on this decision: TensorFlow Addons Wind Down

Please consider sending feature requests / contributions to other repositories in the TF community with a similar charters to TFA: Keras Keras-CV Keras-NLP

[Times125]

您好，我已经收到您的来信，会尽快处理。