– DRAFT –
WebML CG Teleconference – 11 April 2019
11 April 2019
me/ anssik shall be scribing details in this irc window.
we have a new scribe, please be nice and talk clearly and slowly
we can use more than 30 minutes today, we have three topic.
anssik: agenda topic proposed by Greg from Microsoft
<anssik> 1) Discuss & review concrete API proposals (WebNN Chromium POC, JIT proposal, others)
<anssik> https://github.com/webmachinelearning/webnn/issues/11
<anssik> WebNN Chromium POC
<anssik> JIT proposal
anssik: kickoff next stage api sketch.
gregwhitworth talking
<gregwhitworth> https://github.com/DanielMazurkiewicz/WebAI/blob/master/README.md
greg: speaking on webnn proposal, use cases.
<gregwhitworth> gregwhitworth: Microsoft reviewed WebNN, WebAI, jit, and our own proposal
<gregwhitworth> gregwhitworth: We ultimately decided based on the use cases identified we propose use the WebNN Intel POC example
<gregwhitworth> gregwhitworth: we would like to use that as the foundation spec that we can evolve to meet the need for L1
<gregwhitworth> gregwhitworth: I'm fine allowing folks to review it but I'd like to make progress on it
<gregwhitworth> Nikhil: I agree folks are talking passed each other but that's due to the goals not being known
<gregwhitworth> Nikhil: we do know that clients want to execute ML on the client
<anssik> [ https://webmachinelearning.github.io/charter/ attempts to explain the target audience ]
<gregwhitworth> Nikhil: the angle I was coming from, it would be cool if this was an authoring platform in the browser
Nikhil_Thorat: programming world moving towards eager based approach
<gregwhitworth> Nikhil: I do think jit is on the trend and is entirely eager based, it's not about performance
<anssik> PROPOSED RESOLUTION: start "v0 spec" as a graph builder and graph-executor API and in "v1" explore direct usage of native calls in the browser as outlined in the JIT-based approach for improved web developer ergonomics
<chrisheilmann> +1 on that. I'd love to have the JIT earlier than later, from what I had as feedback people like the idea of using pre-defined models on device.
<gregwhitworth> gregwhitworth: yes that proposed resolution is what we'd like
tomoyuki speaking
<anssik> tomoyuki: graph builder and graph-executor API as first step sounds good
<Daniel_Smilkov> +1 to proposed resolution
<NingxinHu> +1 to proposed resolution
<tomoyuki> +1 to proposed resolution
Resolved: start "v0 spec" as a graph builder and graph-executor API and in "v1" explore direct usage of native calls in the browser as outlined in the JIT-based approach for improved web developer ergonomics
<anssik> 2) Kick off concrete API sketching at https://webmachinelearning.github.io/webnn/
<Nikhil> no objections to the graph-based proposal v0 spec. I wanted to make sure we understand the loss of usability in favor of the simpler implementation (exposing the graph to the user) and it sounds like everybody understands those constraints to happy to move forward. We can iterate on the JIT / eager-based API in v1! :)
<anssik> PROPOSED RESOLUTION: Kick off concrete API sketching at https://webmachinelearning.github.io/webnn/ based on WebNN Chromium POC
<gregwhitworth> Proposed Resolution: Evolve specification using https://github.com/intel/webml-polyfill/blob/master/docs/api.md as foundation specification
anssik: thin wrapper over android nn api.
<anssik> PROPOSED RESOLUTION: Evolve WebNN API specification using https://github.com/intel/webml-polyfill/blob/master/docs/api.md as foundation specification
<Daniel_Smilkov> Starting with the WebNN API and evolving it sounds good
<gregwhitworth> gregwhitworth: Nikhil or Daniel_Smilkov - issues with that resolution proposal?
<Daniel_Smilkov> No issues
<anssik> [hearing no objections from anyone]
<Nikhil> +1 to Daniel, no issues
anssik: wait a sec, I'll reconnect to WebEc
Resolved: Evolve WebNN API specification using https://github.com/intel/webml-polyfill/blob/master/docs/api.md as foundation specification
<NingxinHu> WebNN POC API starts as a thin wrapper of NNAPI. However it has main interfaces for a graph-builder and graph-executor, including Model, Compilation and Execution interfaces. There are many operations listed in the POC API doc. For WebNN foundation API, we can just start a very basic set of operations.
anssik: thank you Venkat
anssik: what we want to do at the TPAC meeting
anssik: it is scheduled for Sep 16-20
<anssik> https://www.w3.org/2019/09/TPAC/
anssik: it's in Japan
anssik: I did some research here
anssik: WebGPU group does not have f2f at TPAC, and web assembly they will meet on Thursday
anssik: so I propose that we would meet f2f on Tuesday 17th in the afternoon
anssik: 1:30PM - 3:30, 3:30 - 5PM
anssik: my plan is to merge these two slots into 4 hours of face time
anssik: that may be a lot to ask since we're a CG
<anssik> PROPOSED RESOLUTION: WebML CG will meet F2F at TPAC Tuesday 17 September 2019 - 13:30-15:30 or 15:30-17:30, with preference to merge the two slots for a 4 hour session.
anssik: Friday is another slot, but folks will be tired, but more energy on Tuesday
anssik: describes what TPAC is
RESOLOLVED: WebML CG will meet F2F at TPAC Tuesday 17 September 2019 - 13:30-15:30 or 15:30-17:30, with preference to merge the two slots for a 4 hour session.
Resolved: WebML CG will meet F2F at TPAC Tuesday 17 September 2019 - 13:30-15:30 or 15:30-17:30, with preference to merge the two slots for a 4 hour session.
anssik: silence I think means consent
Resolved: Have bi-weekly telecons
NingxinHu: I heard from Jebs at Google it was a little early
NingxinHu: a half an hour earlier is fine with me
NingxinHu: half an hour later is fine with me
chrisheilmann: maybe we can alternate times
anssik: yes we can do that possibly
<tomoyuki> +1 to ningxin’s opinion :-)
anssik: I would like to have active contribs on the call
<NingxinHu> I heard from James at Google it was a little earlier
Resolved: push the meeting 30 minutes later for PDT ease of waking up