TensorFlow Addons Meeting Minutes: May2020

19 views
Skip to first unread message

Sean Morgan

unread,
May 5, 2020, 1:18:46 PM5/5/20
to SIG Addons, Karmel Allison, Joana Carrasqueira

Agenda

    • Dropping custom-ops from activation functions

      • Rationale

      • XLA compiled composite ops

      • To prevent re-tracing on while using tf.function we may want to convert python to constant tensor 

    • Future of custom-ops in Addons / TF

      • What is the outlook for custom-ops non activation functions? (cause part of the rationale on removing custom ops and related issues is in common with non activation functions)

    https://github.com/tensorflow/addons/tree/master/tensorflow_addons/custom_ops

    https://github.com/google/mlir-npcomp/issues/1

    We have only an approved policy on the SIG side. What is the policy for the Tensorflow team/community for processing/accept/refuse a promotion RFC? Having a clear policy on TF side could help the community to avoid to generate noise or stalled RFC that are not compliant with TF evaluation metrics. Cause if an RFC is accepted in TF core then the code overship will be passed to TF team so we could have metrics that don’t overlap with the SIG ones that we defined. https://github.com/tensorflow/community/issues/239

    • Popularity

    • Stability

    • Complexity

    • Maintainability (related to Sponsor of the RFC?)

    • Disambiguation between promotion process and direct contribution:

    If we clarify the previous point we could enable to route new operators in SIG Addons (or in other SIGs i.e. like IO etc.) and then upstreaming with a default defined  process. This could improve the community process and will avoid something like seems an “arbitrary decision” on what third_party is candidate to go to TFcore directly and what needs to pass through Addons. See https://github.com/tensorflow/tensorflow/pull/37937#issuecomment-604694529


    • Intern projects?

      • If we could create well defined projects for addons we could staff 

      • Timeline? Starting ~Ending of May




    Action Items:

    • Create a proposed policy for accepting RFCs for upstreaming code (Karmel)

      • Adoption? Metrics? Testing etc.?

    • Can we create instructions for triaging new contributions that clash with Addons (SIG)

    • Create `tensorflow-addons/` namespace on dockerhub (SIG)

      • Can use docker cache instead of image for BUILD env

      • Secret Sharing may be the best practice 


    Mark Daoust

    unread,
    May 5, 2020, 1:37:31 PM5/5/20
    to Sean Morgan, SIG Addons, Karmel Allison, Joana Carrasqueira
    User adoption: Now averaging 12k page-views/week on tensorflow.org

    Screen Shot 2020-05-05 at 10.28.05 AM.png

    --
    You received this message because you are subscribed to the Google Groups "SIG Addons" group.
    To unsubscribe from this group and stop receiving emails from it, send an email to addons+un...@tensorflow.org.
    To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/addons/DM5PR07MB38787DCEA43EB5120AE20147DDA70%40DM5PR07MB3878.namprd07.prod.outlook.com.

    Gabriel de Marmiesse

    unread,
    May 5, 2020, 1:57:22 PM5/5/20
    to Mark Daoust, Sean Morgan, SIG Addons, Karmel Allison, Joana Carrasqueira
    Happy to see that we're having more and more users! :)

    To view this discussion on the web visit https://groups.google.com/a/tensorflow.org/d/msgid/addons/CAL1kWUO%2B3DWuUA0H0nfTKvLs%3DUS_QVXy5xOvnOVC31PuNpxhSg%40mail.gmail.com.
    Reply all
    Reply to author
    Forward
    0 new messages