Implement Conditional Random Field

[seanpmorgan]

Per the RFC, we need to move CRF from contrib to addons:

This will involve moving the code and modifying test cases to run in TF2.

[a6802739]

@seanpmorgan, may I give some help on this?

[seanpmorgan]

Absolutely! Please let us know if you run into any issues. As noted in the other issue, one thing to note is that the RNN cells are unified under Keras RNN in TF2.

[a6802739]

@seanpmorgan, thanks for your help, I will ping you if I met some issues.

[a6802739]

@seanpmorgan, Sorry for late interaction. So what I should do now is move CRF from contrib to the directory addons/tensorflow_addons, and change the corresponding test in tensorflow, right?

[seanpmorgan]

Hmmm move the code and change the tests/code in addons to match the TF2 syntax would be how I would word that.

[a6802739]

@seanpmorgan, Thanks for you help, and could you please give me some PR reference that I could follow?

As my understanding, I should move the code to addons and add crf_test.py within adding/crf and guarantee the unit test style match the TF2 syntax, for example using test.TestCase and add py_test in BUILD file, right?

And how I should deal with the function defined in contrib/crf, should I use class to implement a wrapper or just move the code directly?

I'm appreciate for your kind comments.

[seanpmorgan]

Hi @a6802739

As my understanding, I should move the code to addons and add crf_test.py within adding/crf and guarantee the unit test style match the TF2 syntax, for example using test.TestCase and add py_test in BUILD file, right?

Yes that is correct.

And how I should deal with the function defined in contrib/crf, should I use class to implement a wrapper or just move the code directly?

This is where I'm confused. You'll need to move the code from contrib, but then modify it to use the unifed RNN class in TF2. This will come with some changes to the code so that it runs correctly. To be honest, this is not the most straightforward of all the issues we need help with. It might be easier to move something that already has an example PR we can point you to. For example moving an optimizer and then you can match how LazyAdam was moved.

[a6802739]

@seanpmorgan , Thanks for you kind help, I will have a try then. Could you please point me the example PR and the ongoing issue about optimizer.

[HuggingLLM]

How is it going on about this issue? I want to use tf.keras(2.0), but there is no CRF Layer, so that I can only use keras & keras_contrib.