Meeting minutes
WebNN API recent new feature requests
Request input layout and resize only on height and width for Resample
Anssi: we discussed this in our last meeting
… the PR was updated accordingly
… it has been reviewed
… I think we should be ready to merge this, after confirmation we're happy with the change, incl from Rama
Rama: fine with me
proposed RESOLUTION: Merge Change resample to resample2d PR #205
Resolution: Merge Change resample to resample2d PR #205
Support for configuring rounding type in pooling operations
Anssi: updated too following our last meeting
commit "Support outputSizes of pooling operations"
Anssi: Chai, wanted to get your input on this?
Chai: I thought I had provided feedback on this, but maybe not directly on the PR
Anssi: wanted to check you're OK with the latest commit in it
… wanted an explicit signal
Chai: I'll follow up on the PR
reduceLogSum, reduceLogSumExp, reduceSumSquare are not supported on OV/NNAPI
Anssi: do we still want to include the informative text on emulation in the spec for consistency sake
https://
<chai> the pool2d PR looks good to me. I've approved the PR.
chai: where would the informative text be added?
anssi: within the context of ops that can be used to implement this
… reduction operations section
https://
chai: for these 3 ops, to add this pseudo-code?
… yes, we can do that
anssi: I think in the end it will help us
<chai> +1
Support iohw filterLayout for conv2d op
<chai> [need to step away for a minute]
anssi: a new request to be triaged
Anssi: Ninxgin submitted a PR for this
anssik: a new request informed by fast style transfer ONNX model
… ConvTranspose op needs "iohw" filter layout:
… [input_channels/groups, output_channels, height, width]
<chai> [back now]
Anssi: in the spec use cases we quote two other style transfer models
… are there concerns in expanding the layout options as proposed?
chai: style transfer can work with any type of layout
… the proposal makes it easier to implement *a* version of style transfer
Anssi: Ningxin, have you talked with Wanming about the specifics of the proposal
Ningxin: I did
… he's been using the ONNX format of the said model
… and has been trying to map that to WebNN ops
… as part of development of WebNN Samples
… the ONNX ConvTranspose uses a filterLayout that isn't supported by WebNN Conv2D
… he proposes to support that with a WebNN Conv2D specifically for the ConvTranspose usage
… and to have better support for the ONNX ConvTranpose op
… The PR itself is fairly straightforward
… the underlying implementation can support that with not much effort
… it can be transposed to the native transpose operation
… would like feedback from the users-side of the API
Rama: this seems fine, we'll take a closer look
chai: thinking to our decision on convolution transpose a while back, this topic was raised
… I would like to see what our resolution was
… I want to make sure that manipulating the filterLayout is a good to do it now that we have quite a few of them
… I'll need to look more into it
anssi: let's keep the PR open until we get reviews from Chai & Rama
TPAC registration and agenda building
Anssi: first, Public Service Announcement: registration for TPAC has opened https://
… it's a very quick registration process
… you should receive to the virtual attendee hub that will help you join the right meeting during TPAC week
dom: this is pending a technical glitch in our system, should be fixed soon
Anssi: Another announcement: we have 3 confirmed meetings during TPAC, 26-28 at 14:00 UTC
… For that meeting, please raise suggestion of topics via github https://
Integrating an open-source cross-platform implementation of the Web Neural Network API into a web engine
Anssi: we've received this proposal that I think should be of interest for folks working on Web engines
… in particular, we would invite non-WG participants as guests to that session
… e.g. folks from the Dawn project
… a cross-platform webgpu implementation
… WebNN-native shares quite a bit of infrastructure (inc a code generator) with the Dawn project
Anssi: who should we invite from Chromium for such a topic? Kai?
RafaelCintron: I would be interested to join that session
Anssi: absolutely - maybe also other Microsoft folks not generally involved
dom: maybe this could be a breakout session?
anssi: great idea; logistically, this timeslot works for this WG which makes it easier
… but we could have it as a breakout instead or in addition
… the first step is to identify who to have at that session
Anssi: we would love to get Chrome engineers involved in Dawn in this discussion
Ningxin: last year, during the workshop, Corentin the WebGPU co-chair gave us very good input on the WebGPU & Dawn
… we can check with Corentin about his interests & bandwidth
… when we started with WebNN, Corentin provided us valuable feedback from the Dawn usage perspective
… another aspect for the Web specs, we actually define the interop between WebNN & WebGPU buffers with Chai's input
… Wrt Dawn's usage, it might be a good discussion point for WebNN/WebGPU interop
… if they use a single infrastructure, would it make it easier for implementors to provide interop between the two systems
… that would be an interesting topic
anssi: please let us know if there are people outside this WG interested in this discussion
other TPAC topics proposals
Anssi: our group is set to work on ethical issues - I've started a document to bootstrap that work that I could present at TPAC
… Dom & I have started to reach out to guests to join that effort
… I know most companies working in this space have dedicated people looking at ethical issues
… it would be much appreciated if you could connect these folks with us
Dom: +1 on that plan
dom: having something out will encourage more contributions
anssi: what I've learned is that I know very little about it, and it's a very cross-discipline topic
… there is a lot to learn
… we don't want to reinvent the wheel, but develop perspectives in the lens of the Web platform
… based on its specificities - e.g. the Web has the broadest reach of all platforms, making ML/AI that much more powerful when available there
… I would love to have someone outside of this WG educate us on this topic
… I'll get a starting point in a repo
… for our TPAC discussions, I expect we'll have someone to lead a given topic to seed the discussions with an explanation of where we are
… similar (but less formally structured) as the workshop
Anssi: going through the list of proposed topics
… "Rationale/criteria for adding new ops to the WebNN API"
… this would help people to contribute to the work - anyone willing to lead this?
… e.g. some frameworks have similar criteria - could they be adapted to WebNN?
… next topic "Versioning and web compatibility"
… it would be great to provide a Web perspective on this
dom: someone from the TAG would be good person to kick off this discussion
dom: interesting topic in general, with a specific angle for this technology
… not entirely sure how to structure this conversation
anssi: getting someone from the TAG involved would be interesting
… next topic "How should WebNN API deal with fusion"
… again, is someone passionate about this topic? or is there an expert we should reach out to?
… IIRC Daniel Smirkow from Google raised this topic early on
… another topic is "ML JS framework performance" - ensuring that WebNN fulfills its promises to that primary category of customers
… i.e. WebNN as backed to JS frameworks
Ningxin: I've been working with Chai and Ping into ONNX-Web and TF-Lite
… we're trying to develop a design doc (in progress), with use cases and running code based on WebNN-native implementations
… to share some data
… we should be able to share the status of that WIP and that data
Anssi: sounds great
Ningxin: Chai has a lot of experience with DirectML integration with ONNX, Ping with TF
dom: please remember to register to TPAC and if you want to propose topics outside this WG submit breakout sessions, soft close for breakouts Oct 8
Ningxin: I'd like to propose a topic of discussion with regard to integrating with Web platform tests
… there has been progess happening in the polyfill repo
… the maintainer has recently converted his mocha-based tests into WPT test harness tests
… incl work on IDL harness
… this may be a good topic for discussion for the WG, at TPAC or otherwise
Anssi: good suggestion! do you if Bruce need any specific help?
https://
Ningxin: may need to invite someone from WPT to help with some of the open issues
… Another aspect is based on what Chai had shared during the workshop on test granularity
… op-level vs …
… it may be a good time to discuss these principles
… and existing gaps in the test suite as a roadmap for our work in this space
Conformance testing of ML APIs for the Web
anssik: Happy National Day holiday (1-7 Oct) to our participants in China!