Meeting minutes
Announced: W3C Virtual Workshop on Web and Machine Learning
anssik: W3C announced the Web & Machine Learning Workshop now being organized as a virtual event in September 2020.
anssik: The event will be organized as a combination of pre-recorded talks (to be submitted in July 2020) followed by a series of live sessions in September, organized around 4 main themes:
… Opportunities and Challenges of Browser Based Machine Learning
… Web Platform Foundations for Machine Learning
… Machine Learning Experiences on the Web: A Developer’s Perspective
… Machine Learning Experiences on the Web: A User’s Perspective
anssik: The event is free and open to anyone with relevant perspectives on the topic to register for the event. For more information on the workshop, please see the workshop details and submission instructions.
Workshop details and submission instructions
anssik: Deadline to submit a proposal for a talk is 3 July 2020 and registration is open through 14 August 2020.
anssik: Questions?
LeakyReLU and min ops
Add leakyRelu and min (PR #63)
anssik: PR reviewed and merged, thanks!
… Ningxin reports: Per the first wave models, leakyRelu is required by TinyYOLOV2 model. leakyRelu can be implemented by element-wise add, mul, max and min where min is being added in this PR.
… Any comments?
ningxin_hu: Thanks Chai and Rama for review.
anssik: I'd like to note we should add a more complex example to the spec that makes use of these new ops we've added to the spec recently. We could have a simple "Hello World!" example and in addition a more advanced one within a reasonable LOC limit.
ningxin_hu: I support the idea to add a more complex example to the spec
anssik: I'll make an issue for that
2D pooling and reduce ops
2D pooling and reduce ops (PR #64)
anssik: This PR defines the 2D version of pooling and relevant reduce ops
… other changes to make convolution's strides, padding, and dilations optional params for convenience
… also this PR received adequate review and was merged.
… Chai anything you want to mention about this PR?
Chai: I decided to work on these ops together, since they're connected, and tried to make them consistent in terms of params and name a few optional
… pooling related to reduction so in one chart defined all the ops supported by all the frameworks, 7 types of reductions
… some frameworks were inconsistent among themselves (?)
… if we need to add a new reduction in the future can append this newly added section to keep consistent
… GlobalMaxPool is a specialized version of MaxPool and therefore it needs not be defined explicitly
Chai: thanks Rama and Ningxin for your extensive review
anssik: thanks for the work!
ningxin_hu: Some design considerations, conventions introduced in this PR, grouped related ops into one section such as reduction and pooling, in the spirit of this PR I put together a PR for softmax, in that PR the group was element-wise binary, add etc. in one section
… convenient way to avoid spec definition duplication
… also element-wise unary for softmax, it is a separate PR, just making a point following the conventions
Grouped conv2d
Support grouped conv2d (PR #65)
anssik: This PR adds grouped convolution support to the existing conv2d definition, or convolutions in parallel said differently
… as with the previous PRs, our active participants reviewed also this PR ahead the call and it was merged
… great work!
as a related matter, I'd like us to keep https://webmachinelearning.github.io/webnn/#acknowledgements updated
ningxin_hu: I'd like to add Rama to ack section
anssik: +1 to add Rama
… Ningxin anything to bring for discussion regarding this PR?
ningxin_hu: I would like to highlight Chai's suggestion we call out depthwise conv op as a variant of grouped conv2d used in MobileNet first-wave model, a very good use of this spec
anssik: any further comments?
Noise suppression
Evaluate noise suppression models for required ops (issue #66)
anssik: anyone interested in investigating this issue? Belem perhaps?
anssik: some proposed tasks include:
… - Evaluate RNNoise model for required ops, to be added to the first-wave models
… - Gauge PoC interest, possible collaboration with related groups e.g. WebRTC
Chai: This is timely, Google Meet launched noise cancelling as a feature, an important scenario
… if nobody takes up the work, happy to look into RNNoise