Extensions to the TensorFlow ecosystem have found a new, well-structured home in the TensorFlow Addons repository, with a number of maintainers making sure it’s kept tip top.
Before TensorFlow 2.0, new functionality from community members often landed in the tf.contrib module. But since the code in it often wasn’t very well documented or even maintained, it became harder to use and a sort of graveyard for some additions. For that reason, the TensorFlow team decided to get rid of tf.contrib, moving important code into separate modules and dumping code no one stepped up for.
To make sure nothing of importance was lost in the process, however, the SIG looked into the functionality available in tf.contrib. If something wasn’t otherwise available in TF, garnered sufficient community interest, and conformed to an established API pattern, it should end up in the Addons repository now.
While the new extension home might not be immune to out-of-date code, the SIG has been busy installing measures to make sure it doesn’t end up like its predecessor. The repository is structured in a modular way, with designated maintainers for all subpackages and submodules.
Maintainers have quite a job with tasks such as “periodically review and deprecate old and unused code, manage graduate candidate for the TensorFlow Core, ensure API conformity and test code quality, manage issues, and review changes”.
Meanwhile, the documentation issue is going to be tackled by “providing examples for all functionalities through Google Colab Notebooks”. Having implementation examples like this available sounds like a good step towards TF 2.0’s general goal to get more newbies interested in the project, and might in turn motivate some to start contributing themselves.
If you’re interested in the new SIG, there is a mailing list for general discussions and monthly video conference for talking strategy. Quick questions can be posed via Gitter, while the place to inform about code issues and bugs is still the project’s GitHub repo.