tf.addons.text ops coverage about keras-nlp

[mattdangerw]

Thanks for bringing this up! This is helpful!

I am unsure of how we could monitor for usage of the TFA utilities, but can poke around. That could help us prioritize well here.

Re CRF, did you have a design in mind?

I think the highest priority symbol of the bunch might be the skip_gram* functions. We have these in TFA, and another similar version in keras.preprocessing. This would be better off as a preprocessing layer that could replace both of those usages. word2vec is definitely a common used model we would like to enable easy support for with this library.

I think a skip gram preprocessing layer, with the goal of replacing the tf.keras.preprocessing.sequence.skipgrams usage in the official tensorflow word2vec tutorial, is a good goal.

[bhack]

I am unsure of how we could monitor for usage of the TFA utilities, but can poke around. That could help us prioritize well here.

We have a quite old and long thread about this at:
tensorflow/addons#236

Re CRF, did you have a design in mind?

I think that @howl-anderson, our new CRF codeowner, could give use some feedback.

I think the highest priority symbol of the bunch might be the skip_gram* functions. We have these in TFA, and another similar version in keras.preprocessing. This would be better off as a preprocessing layer that could replace both of those usages.

This has an (old) custom op in TFA so I hope that you can handle it here with a pure compositional python/TF solution.

We have also an rnn and a seq2seq namespace that you could evaluate in this context:
https://www.tensorflow.org/addons/api_docs/python/tfa/rnn
https://www.tensorflow.org/addons/api_docs/python/tfa/seq2seq

/cc @seanpmorgan @yarri-oss @theadactyl

[Stealth-py]

Hey @mattdangerw! I wanted to ask if the keras team is still interested in implementing a preprocessing layer for models such as the skipgram model.

[bhack]

/cc @seanpmorgan @yarri-oss

[bhack]

keras-team/keras-cv#74 (comment)