WebML CG Teleconference – 29 October 2020

Meeting minutes

WebNN API explainer

Update the webnn explainer docs (PR #109)

Chai: Considered alternatives still missing content

anssik: perhaps mention Model Loader and MLIR-inspired proposals

Chai: those are actually already discussed in other sections
… we haven't had any U-turns on the API design
… I can do some cleanup and fill the acknowledgements sections

Considered alternatives

anssik: "One of the most important things you can do in your design process is to catalog the set of roads not taken. As you iterate on your design, you may find that major choices in your approach or API style will be revisited and enumerating the full space of alternatives can help you apply one (or more) of them later, may serve as a “graveyard” for u-turns in your design, and can give reviewers and potential users

confidence that you’ve got your ducks in a row."

Rafael: Maybe we could delete this section
… only alternative we could ask here is Model Loader, or why don't we just use WebGL/GPU
… refer to sections providing answers to those question

ningxin_hu: as discussed before, agree perf and efficiency are to be highlighted in here

anssik: after completing the considered alternatives and ack section, Anssi to initiate TAG review

Support the execution of the sub-graph scenario

Support the execution of the sub-graph scenario (issue #105)

ningxin_hu: this should be covered in the PR #94
… also related issue #105

Ping's comment: "With this API, the computation is tied with the compilation. should the compile method have the inputs/outputs pair and execution can only execution the graph compiled with the input/output pair? Otherwise, the compilation might not support the execution of the sub-graph scenario?"

Chai: two ways to do this, described in a PR:

https://github.com/webmachinelearning/webnn/pull/94#discussion_r498289215

anssik: what is Ping's reaction to the proposal?

Chai: no comments yet

ningxin_hu: Ping would need to elaborate the use case to understand the gap better

Proposed new ops Mirrorpad, SquaredDifference, Pow, TransposeConv

anssik: Please review the issue and provide feedback

WebNN API is not support Mirrorpad, SquaredDifference, Pow, TransposeConv (issue #108)

Chai: this issue is about how do we know what models we need to support?
… we should look at the models people want, this is one such a request
… I'm open to brainstorming what models we should look at and map into the spec
… personally I don't prefer expanding the first wave models, but we need a way to accept extensions
… we could have a separate repo e.g. "models" under "webmachinelearning" GH org

<Chai> examples: ONNX model zoo https://github.com/onnx/models

<Chai> TensorFlow model garden: https://github.com/tensorflow/models

anssik: What type of questions we should ask from folks proposing new ops to be added?

Chai: Style Transfer is pretty known model and supported in e.g. ONNX Model Zoo, so I think the model we might not support are those that are used in classical ML for example
… any Deep Neural Network would be good to look into supporting
… nothing wrong in Style Transfer

anssik: can you respond to Calvin maybe providing platform support information?

ningxin_hu: can do that, also cross-check between ONNX and XLA-HLO
… re model criteria, maybe ask folks to look at the use cases in the WebNN spec, if the use case is not included need to start from there

https://webmachinelearning.github.io/webnn/#usecase-style-transfer

ningxin_hu: linking models to use cases could be the precondition

anssik: style transfer indeed one of the use cases

<Chai> + 1 on ningxin's

anssik: you can implement only a subset of the use cases with the first-wave ops?