Meeting minutes
Model Execution API
anssik: chai noted by introducing the notion of topology in the API, it acts as an agent to the context that owns the transient state of the topology being constructed.
… this allows for more efficient garbage collection of nodes
anssik: Rama mentions how the proposed API might allow the API to act as both a graph builder and an eager API that evals ops as they are constructed
anssik: Ningxin proposed an 'eager' arg for getNeuralNetworkContext
anssik: proposal to track eager model in a separate issue
… considered an implementation variant of the same context interface
Chai: this is a long discussion, chatting with Ningxin offline that we should handle with this in separate issues
… secondly, I have pushed a PR out that hopefully addresses the issue
Update model execution API PR#94
Chai: key point around how we separate different states, neuralnetworkcontext should keep the state, model builder is the place where the implementation can keep the state that is specific to the graph building session
… with proper separation of states, GC easier, once you produce a model you no longer need them and they came be CG'd
Chai: I think PR #94 ffixes this issue #87
… eager has a separate issue so be can piggy-pack on that one
Executing operations (issue #11)
anssik: issue #11 for eager mode follow ups
<ningxin_hu> https://github.com/webmachinelearning/webnn/pull/22
<ningxin_hu> and https://github.com/webmachinelearning/webnn/issues/16
ningxin_hu: PR #94 LGTM, apart from a separate issue for scalar tensor
<ping_yu> morning, I will do that
anssik: anything else to discuss for issue #87?
ningxin_hu: a comment re model loader, may need Jonathan's input on that one
Rama: just a minor questions, probably a different issue re model loader
… the only question is whether we need indirection to load two different models over on a single call
… if having a separate model loader makes no sense, we could have the WebNN and Model Loader under the same API
<zkis> The model loader API could have an `options` optional arg for specifying constraints
Chai: I don't see a drawback with Rama's proposal, but we need to think this a bit more and open a separate issue for merge proposals, with input from Jonathan
anssik: should understand the scope of WebNN API at the time of WG transition with is expected to happen by the end of the year
https://github.com/webmachinelearning/webnn/issues/90
zkis: for model loader, options arg would make sense, to be able to pass constraints
anssik: API ergonomics improvement?
WebNN code editor
anssik: Review and solicit feedback on the WebNN code editor:
Add the WebNN code example (issue #4)
ningxin_hu: this feature is inspired by TF.js docs, discussed with Ping already
… we may want to run the example code in our spec with this feature
… also make the example code in the spec editable
… by default read-only, can turn into edit mode to turn the code snippet into editable code
… considering adding a mechanism to switch between examples from within the code editor
<Chai> this is an excellent idea. it'll help with making sure that the samples and the spec stay consistent.
ningxin_hu: later we can add graph builder LeNet example
<ping_yu> looks great and very useful
<ping_yu> It would be good to integrate this with the spec doc
anssik: everyone is in agreement this code editor is a great feature, LGTM!
WebNN API explainer
anssik: TAG review readiness depends on the explainer document
anssik: Discuss how to approach the following explainer sections:
… Key scenarios
… Detailed design discussion
… Considered alternatives
<Chai> i can take a peak and see how to help. is there any timeline on this?
anssik: no timeline per se
<zkis> Sangwhan can probably proxy this spec in the TAG, but IMHO other technical members would wonder if the API could be made more JS-like
zkis: Kenneth of TAG provided feedback on making the API more JS-like
<ningxin_hu> i can help as well
<zkis> https://w3ctag.github.io/design-principles/
https://github.com/webmachinelearning/webnn/issues/18
https://github.com/webmachinelearning/webnn/issues/74
<zkis> OK, I can take a look if I would have any useful suggestions on this
Workshop proceedings at TPAC 2020
anssik: WebML CG Virtual Meeting at TPAC 2020 discusses the W3C Workshop on Web and Machine Learning key outcomes with a focus on proposed near-term and long-term standardization directions and next steps.
… WebML CG participants are invited.
… meeting info:
… WebML CG Virtual Meeting at TPAC 2020
… 22 October 2020 - 14:00-15:00 UTC+0
… Agenda: https://github.com/webmachinelearning/meetings/tree/master/2020-10-22-tpac2020
Proposals for future work
anssik: Per workshop feedback, new repo created to capture Web & ML proposals for future work that are not in scope of any group currently:
Adjourn
<ping_yu> bye