TensorFlow Addons Meeting Minutes: March2020

16 views
Skip to first unread message

Sean Morgan

unread,
Mar 9, 2020, 10:00:06 PM3/9/20
to SIG Addons, Joana Carrasqueira, Karmel Allison

Agenda

  • Review previous action items

    • Fix custom CUDA install: #770 (SIG)

    • Issue for GPU test script

    • Create draft PR of Philip’s F1 implementation / Tag Karmel in the issue to discuss with Keras team.


  • SIG Build notes:

    • Custom-op docker images

    • Requested release candidate images

    • Can we use 2.1-cpu for build on 2.2.?

      • Probably shouldn’t rely on it

    • 2.2 image should be out ~immediately following 2.2 release

    • Python3.8 available on nightly for linux

      • How do we want to handle (macos may not be supported in 2.2)?

      • Okay with just building is available

  • Profiling custom ops

    • https://github.com/tensorflow/addons/issues/1156

    • Likely not a universal suite for all ops

      • Define a set of models / microbenchmarks for regression testing

      • Nvprof may be helpful for benchmarking GPU ops

      • TFprofiler GUI based method

      • Build a representative model (example per second)

      • Check if there are external microbenchmarking code

      • Can we make custom GPU code with python op CPU mode

    • Gelu example (may update request for merge in Keras)

    • XLA

      • Non-static shapes?

        • Perhaps send an email to discuss

  • Pytest as default test runner

    • https://github.com/tensorflow/addons/issues/1177

    • https://github.com/tensorflow/addons/pull/1207

    • SIG Build would like to test across SIGs

      • Would like Bazel tests if possible

      • We can run all builds with the pytest command + one build with the bazel command to ensure that googlers can still use it when they want to (we might forget bazel tests like before though, since we have no safeguard against that) and use Jason’s idea to run pytest with bazel so that we can use pytest’s syntax and fixtures. I’ll update my pull request to reflect that.

  • Documentation questions

    • Can we add nightly docs like TF-Core?

    • Can we make it so type hints show up in the docs? (Not sure of current behavior)

  • Notifications for proxy maintainers

    • Hooray!!

    • Missing proxy maintainers

      • image/interpolate_spline.py

      • image/sparse_image_warp.py

      • losses/metric_learning

      • losses/triplet*.py

  • Registering ops and Keras objects: 

  • Backporting fixes!

  • F1 Metric

  • Open Floor

    • TFA OSS Usage analyzer

    • GSoC

      • AutoAugment

      • Issue tagging of proxy maintainers

    • Caching pip packages on CI: https://github.com/actions/cache

      • pip ENV variable to cache somewhere else

    • Large pull requests / strategy for reviewing

      • Many stale PRs we need to review

      • Large PRs make for very difficult reviews

      • Single feature per PR (even similar layers should be split into PRs)

      • PR op before layer

      • Split documentation into the body of the PR

        • Can even split documentation strings out to make reviews easier

      • Similar to linux patches have a flag to disable implementation that are not complete as we split up PRs.



Action Items

  • Fix custom CUDA install: #770 (SIG)

  • Investigate the recent segfaults on tf-nightly (SIG)

    • Is this just because of different mangled names that we’re linking against

  • Can we have CUDA kernels (GPU) and composite python ops for CPU (SIG)

    • Check LSTM recent consolidation

  • XLA non-static update (SIG)

    • Email developers@tensorflow

Reply all
Reply to author
Forward
0 new messages