What are some of the trade-offs between CLU and Jaxline?

[marcvanzee]

CLU and Jaxline are both libraries for simplifying some of the boiletplate code that exists in ML.

[marcvanzee]

CLU

• Open-sourced: https://github.com/google/CommonLoopUtils
• Doesn't (and never will) contain a training loop. They believe that the training loop should be written and controlled by the user.
• Contains a few utilities making your code and training loop short and readable (e.g. periodic_actions, deterministic_data).

Jaxline

• Open-sourced: https://github.com/deepmind/jaxline
• Focuses on the "running a synchronous step function in a loop on TPUs"
• Has all the boilerplate code for checkingpoint / logging / etc.

[kamalkraj]

Hi @marcvanzee,
Is there any documentation available for CLU, Other than the clu_synopsis.ipynb

What is the function of the platform module. ? https://github.com/google/flax/blob/e30b7f5fff03df0840e7da40a9f8923aee6fb42b/examples/mnist/main.py#L54:L57

[andsteing]

Hi @kamalkraj
clu_synopsis.ipynb covers most of the functionality and is our only documentation existing so far (other than the pydocs).

The platform module is a bit special because "work unit" merely call absl.logging.info() externally, but interact with Google's experiment scheduling system internally. So this is very useful for keeping the same code internally and externally, but the module currently does not add any value externally, that's also why it's not documented better.

[JonathanRaynerMonumo]

@andsteing Any timeline on when clu might be opened for contributions or documentation will be developed? It seems that

• not much has happened in terms of documentation in 2 years since this thread
• clu is recommended in core examples in jax and flax, including "Quickstart"
• but things can't be improved, because clu doesn't accept PR's or opening issues

this doesn't seem like a good state of affairs for the ecosystem if wanting to grow jax/flax outside of Google...