TensorFlow Addons Meeting Minutes: September2020
Sean Morgan
2020-09-03
Thursday, September 3rd, 11am-12pm PT
Agenda
Review previous action items
Development containers (SIG)
Enable Codespaces / IDEs
Discuss tensorflow testing utilities as a stand alone repo (SIG)
TFA OSS Usage Analysis (SIG)
Streamline dev containers and custom-op containers (SIG)
Support build from source for several CUDA versions (SIG)
TF 2.3rc testing (SIG)
Congrats on ~1K stars!
1.5M downloads
TFA 0.11 Release
0.11.2 with updated doc formatting and doctests
Tensorflow.org has been updated
Pressure from TF vertical grouping (keras-cv, etc.)
Want to understand our position in the ecosystem
Addons graduates to Model Garden now?
Problems w/ Model Garden OSS Interactions
Reusable building blocks should not go into model garden
Model garden to depend on keras-core, keras-cv, keras-nlp, addons
Who can we reach out to when there are issues with reusability
Many teams are working within the tf/model repo
Zhenyu can be tagged, or Tomer, Francois, Paige
@tomerk, @tanzhenyu, @dynamicwebpaige, @fchollet
Agreement that we need coordination
Solutions
Graduation and deprecation process would be handled by TF team?
Keras team to be involved in any upstreams to keras-cv, keras-nlp
Feature request ambassadors
Is something that will be superseded in the future?
Draft a flow chart
Where should the PR land
Tf-addons, tf-core, keras-core, keras-cv, keras-nlp, model-garden?
Single centralized place for issues would be great
TF.org contributing page?
People often try tf-core first
Grace period on Addons issues to be able to notify internal team to decide where the PR should go
Paige willing to be internal contact for looking at feature requests that land on our repo
@dynamicwebpaige
Stale maintainership
Need to formalize a process
Gabriel’s bot currently gets poor response from old CODEOWNERS
No need to notify stale maintainers because it gets less response
Publish dev container
Stefano SIG Build Meetings
Bazel remote cache for TF-Core development was released on GCP
Missing crosstool functionality in the tf-devel docker image so remote cache isn’t always usable
SIG Build is doing inventory on all of the docker images
Streamline the docker images is on-going
Working with NVIDIA to see if we can publish CUDA images on something other than nvidia registry / DockerHub
Okay with the process?
Vscode integration is merged for Tensorflow and Addons (Addons rely on this image):
Aligns with SIG IO: https://github.com/tensorflow/io/issues/1087
Pytest in dev container: https://github.com/tensorflow/addons/pull/1909
Misc
Open Floor
TF-Core and TF-Addons are the only C++ repositories
TPU custom-op C++ compatibility is only available for TF-core since it has to be compiled alongside the core package
Will Keras-CV, Core, NLP place all custom-ops in TF-Core / Addons?
Using composite ops only?
TF serving often doesn’t use a python interpreter (SavedModel may solve that?)
Long term plan is relying on compiler tech
Possible C++ custom ops will go away in the future
TFRT has experimental SavedModel inference support
Keras historically had some dependency on PIL
New preprocessing is fully integrated with TF ops
Keras CV will need custom-ops landed in TF core.
Is TF-Serving still maintained?
Part of TFX team
Not responding to GitHub PRs
SavedModel C++ inference on artifact
Serving endpoints still needed
Action Items:
- Update our repos CONTRIBUTING / GRADUATION md files (Sean + SIG)
Edit the Keras RFCs to say model garden isn’t the intermediate between addons and Keras (Zhenyu)
Begin the grace period and notifications for feature requests (SIG)
Discuss with model team that should be using modular packages (Paige)
Publish dev containers (SIG)
Test the health of current CODEOWNERS (SIG)
CRF Layer tutorial (SIG)