13:58:05 <RRSAgent> RRSAgent has joined #webmachinelearning
13:58:05 <RRSAgent> logging to https://www.w3.org/2020/04/30-webmachinelearning-irc
13:58:11 <Zakim> Zakim has joined #webmachinelearning
13:58:15 <anssik> RRSAgent, make logs public
13:58:19 <anssik> Meeting: WebML CG Teleconference – 30 April 2020
13:58:33 <anssik> Chair: Anssi
13:58:37 <anssik> Agenda: https://github.com/webmachinelearning/meetings/blob/master/telcons/2020-04-30-agenda.md
13:58:41 <anssik> Scribe: Anssi
13:58:45 <anssik> scribeNick: anssik
13:58:57 <anssik> Present+ Anssi_Kostiainen
13:59:00 <anssik> Present+ Rafael_Cintron
13:59:04 <anssik> Present+ Rafael_Cintron
13:59:33 <ningxin_hu> ningxin_hu has joined #webmachinelearning
14:00:13 <anssik> Present+ Ganesan_Ramalingam
14:00:24 <daniel_smilkov> daniel_smilkov has joined #webmachinelearning
14:00:32 <anssik> Present+ Ningxin_Hu
14:00:43 <anssik> Regrets+ Daniel_Smilkov
14:00:58 <anssik> Present+ Paul_McDaniel
14:00:58 <RafaelCintron> RafaelCintron has joined #webmachinelearning
14:00:59 <gabe_intel> gabe_intel has joined #webmachinelearning
14:01:35 <Jonathan> Jonathan has joined #webmachinelearning
14:01:37 <anssik> Present+ Gabe_Esteven
14:01:39 <Rama> Rama has joined #webmachinelearning
14:02:05 <anssik> Present+ Jonathan_Bingham
14:02:34 <anssik> Present+ Ping_Wu
14:02:43 <anssik> RRSAgent, draft minutes v2
14:02:43 <RRSAgent> I have made the request to generate https://www.w3.org/2020/04/30-webmachinelearning-minutes.html anssik
14:02:44 <ningxin_hu> Present+ ningxin_hu
14:03:22 <paul_mcdaniel_msft> paul_mcdaniel_msft has joined #webmachinelearning
14:03:28 <Nikhil> Nikhil has joined #webmachinelearning
14:03:51 <anssik> RRSAgent, make logs public
14:04:01 <anssik> RRSAgent, draft minutes v2
14:04:01 <RRSAgent> I have made the request to generate https://www.w3.org/2020/04/30-webmachinelearning-minutes.html anssik
14:04:24 <anssik> TOPIC: WebNN first wave models and ops
14:04:39 <anssik> anssik: Discuss WebNN first wave models and ops, action review
14:04:46 <anssik> anssik: Review PR
14:04:51 <anssik> -> https://github.com/webmachinelearning/webnn/pull/52 Add first_wave_models.md #52
14:05:05 <anssik> -> https://github.com/webmachinelearning/webnn/blob/df7d9b0b5bfa5f9426e8b0a897e9fdf46a97d537/op_compatibility/first_wave_models.md first_wave_models.md (HTML preview)
14:05:20 <anssik> anssik: models in this proposal target image classification and object detection use cases
14:05:29 <anssik> ... models are SqueezeNetV1.1, MobileNetV2, ResNetV2, TinyYOLOV2
14:05:41 <anssik> ... 13 ops: Add, AveragePool, BatchNormalization, Concat, Conv, Gemm, GlobalAveragePool, LeakyRelu, MaxPool, Mul, Relu, Reshape, Softmax
14:05:53 <anssik> anssik: related actions:
14:06:06 <anssik> ... 1. Compare with XLA to see what ops, if any, are too high level (Nikhil & Daniel)
14:06:15 <anssik> ... 2. Review opsets from a year ago to see what couldn't be ran on ONNX opsets (Daniel)
14:06:27 <anssik> ... 3. Intersection between XLA & ONNX (Paul)
14:06:41 <pingyu> pingyu has joined #webmachinelearning
14:06:52 <anssik> anssik: Daniel's review comments to fix action 1 was to compare which ops are too high level for XLA HLO (see PR#52 for the context), feedback:
14:07:02 <anssik> Daniel: AveragePool, GlobalAveragePool, and MaxPool can be expressed in XLA HLO via ReduceWindow which is lower level because it can take an arbitrary function f: T->T of type XLAComputation.
14:07:07 <anssik> Daniel: Gemm can be covered by Matmul and Add. If the underlying hardware supports Gemm, the fusion of matmul+add --> gemm should be done by an internal optimization pass.
14:07:13 <anssik> Daniel: LeakyRelu, Relu and Softmax can all be lowered into other XLA HLO ops.
14:08:55 <anssik> https://github.com/pytorch/xla/blob/master/torch_xla/csrc/ops/softmax.cpp#L15
14:09:05 <anssik> https://github.com/pytorch/xla/blob/5bc5a0fd522845595961d42f377bde566b4ec94a/torch_xla/csrc/softmax_builder.cpp#L83
14:09:55 <Chai> q+
14:10:02 <anssik> ack Chai
14:11:28 <anssik> Chai: I think no need to map 1:1 to XLA-HLO, but to make sure the set is not too large, should consider break them up
14:11:51 <anssik> ... compiler is specific to the browser, so my concern is how to make sure the compiler has enough context to handle the big group as one
14:13:08 <anssik> ... want to understand how to preserve this context, maybe fusion container could be contained within the graph
14:13:59 <anssik> ... with this information compiler can use it as a clue to check whether the hardware actually support fusion, otherwise fall back and handle ops one by one
14:15:05 <anssik> ... maybe that'd be a good compromise between XLA and ONNX, allow retaining context to allow optimizations
14:15:29 <anssik> daniel_smilkov: that's a great point, just for a full disclosure, I did not comment on batch normalization
14:15:58 <anssik> ... I like the label idea that gives a safe fallback
14:16:12 <anssik> ... no need for complex heuristics
14:16:22 <anssik> ... I looked at ONNX and other high-level op sets
14:20:02 <anssik> Chai: not obvious at authoring time, what kind of hardware would run the model, even with the composite idea there's a chance that compiler walks the graphs and can do a second level fusion, on the heuristics side of the compiler
14:20:52 <anssik> Paul: I also like the idea of labels
14:21:09 <anssik> ... having heuristics is so difficult, we've learned that hard way
14:21:40 <anssik> ... on the path of labels, they probably need to be structured, have some semantics with it, the bottom knows what's on top
14:21:52 <anssik> s/with it/with them/
14:22:01 <pingyu> From tooling point of view, MLIR has a multi-level representation of ML graph. If WebML can be defined as a dialect, the composite can be presented as function of low level HLO ops
14:22:16 <anssik> Chai: for the label, we need to define what the label means, it cannot be arbitrary
14:22:54 <anssik> ... need to define what the subgraph is, no need to define what to do if we fail, the graph is the source of truth, label is just a hint for fusion
14:24:24 <pingyu> TFLite uses labeling before for fusing but it has been bit hacky.
14:25:34 <anssik> ACTION Chai to open an issue for the label hint for defining semantics for op fusion
14:25:45 <ningxin_hu> q+
14:25:57 <anssik> ACTION: Chai to open an issue for the label hint for defining semantics for op fusion
14:26:02 <anssik> ack ningxin_hu
14:26:10 <anssik> ningxin_hu: thanks for discussion
14:27:05 <anssik> ningxin_hu: regarding the PR, daniel commented on particular ops, how about we go forward with elementwise add and multiply?
14:27:20 <anssik> ... concat and reshape also well supported
14:27:45 <anssik> ... these primitives supported in both XLA and ONNX
14:28:28 <anssik> q?
14:29:06 <anssik> PROPOSED RESOLUTION: Add elementwise add, elementwise multiply, concatenation, and reshape ops to the WebNN API
14:29:30 <daniel_smilkov> looks good
14:29:37 <anssik> ningxin_hu: LGTM
14:29:45 <Chai> LGTM
14:29:57 <anssik> RESOLUTION: Add elementwise add, elementwise multiply, concatenation, and reshape ops to the WebNN API
14:31:19 <anssik> paul_mcdaniel_msft: for the ntersection between XLA & ONNX, I have something but maybe better for me to share those on our next call?
14:31:36 <anssik> anssik: SGTM
14:32:43 <Jonathan> @paul, at google we can open xls files :)
14:32:58 <paul_mcdaniel_msft> :)
14:33:04 <paul_mcdaniel_msft> lol
14:33:09 <Jonathan> interop ftw
14:33:44 <anssik> anssik: are we ready to merge your PR?
14:34:29 <anssik> ningxin_hu: OK to keep this PR open and iterate on it a bit, then come back with a revised version
14:34:58 <anssik> Present+ Jonathan_Bingham
14:36:12 <anssik> RRSAgent, draft minutes v2
14:36:12 <RRSAgent> I have made the request to generate https://www.w3.org/2020/04/30-webmachinelearning-minutes.html anssik
14:37:43 <anssik> TOPIC: Model Loader API
14:37:53 <anssik> anssik: Review spec strawman, please open issues for your feedback
14:37:57 <anssik> -> https://webmachinelearning.github.io/model-loader/ Spec strawman
14:38:02 <anssik> -> https://github.com/webmachinelearning/model-loader GitHub repo
14:38:06 <Rama> Elementwise ops can also benefit by introducing a single higher-order op "elementwise"
14:38:16 <anssik> anssik: thanks Jonathan for volunteering to edit the spec
14:38:19 <Rama> which takes a scalar function f as a parameter
14:38:31 <anssik> ... explainer is in a good shape, documents use cases, FAQ, discusses design considerations
14:38:52 <ningxin_hu> @Rama, SGTM
14:39:06 <anssik> ... this repo to be used to iterate on the API and evolve examples with it
14:39:20 <anssik> ... repo welcomes issues and PRs
14:39:43 <anssik> ... Jonathan, do you have some open questions in mind you'd like to ask the group for feedback? (sees zero open issues)
14:40:17 <anssik> Jonathan: thanks Ningxin for proposing extending ML namespace
14:40:29 <anssik> ... first thing to do is to review the spec examples section
14:40:40 <anssik> ... when first drafted, it was pseudo code
14:40:55 <anssik> ... especially interested from folks how to improve it
14:41:56 <anssik> ... another update, I've talked to Googlers working on MLIR, and we're trying to figure out how to get engineering staffing behind this
14:42:58 <paul_mcdaniel_msft> welcome ping !!
14:43:05 <pingyu> thanks
14:43:07 <anssik> ... discussed with Android and TF teams as well
14:43:19 <ningxin_hu> welcome ping!
14:43:24 <anssik> anssik: pingyu welcome!
14:43:31 <anssik> pingyu: working with Nikhil on TF.js
14:43:51 <anssik> ... looking forward to contributing to this group
14:44:03 <anssik> Nikhil: excited to have pingyu in this group, a lot of experience with MLIR
14:44:43 <RafaelCintron> q+
14:46:18 <anssik> Nikhil: with Daniel we're moving from TF.js to another project in around 6 months
14:46:36 <anssik> daniel_smilkov: nothing changes in TF.js we've already reduced our direct involvement in TF.js over the last months so no slowdowns for that project
14:47:49 <anssik> Jonathan: for next steps, need to have discussion in GH on engineering details, as Ping ramps up, my hope is he can coordinate with Ningxin
14:47:50 <pingyu> Jonathan, thanks for the suggestion
14:48:05 <anssik> ningxin_hu: SGTM, happy to work with Ping on that
14:48:47 <anssik> Jonathan: also the prototype, but also the future of how this should look like
14:50:18 <anssik> TOPIC: Virtual workshop
14:50:30 <anssik> anssik: wanted to discuss virtual workshop prepwork status
14:50:44 <anssik> ... on our last call heard interest for a session to discuss XLA and MLIR
14:50:53 <anssik> ... I created a "Domain-specific compilers for ML" high-level topic to discuss these topics, if not accurate name help me perfect it
14:52:30 <anssik> Nikhil: high-level topic name looks good to me
14:52:48 <anssik> ... I can poke people for lightning talks
14:53:22 <anssik> -> https://bit.ly/webml-workshop-talks Review workshop agenda details
14:53:32 <anssik> -> https://bit.ly/webml-workshop-survey Review topics survey
14:54:26 <anssik> anssik: survey maps 1:1 to the proposed agenda
14:54:32 <anssik> ... asks to rate high-level topics on scale 1-5 from not important to very important
14:54:48 <anssik> ... the idea is to send to through your personal networks to gather feedback on scoping. Couple of open questions at the end.
14:55:09 <anssik> ... The Program Committee would build an agenda based on its assessments of what topics would benefit to be covered (starting from https://bit.ly/webml-workshop-talks).
14:55:15 <anssik> ... To gather feedback from the community on the workshop scoping and focus areas, we prepare a survey (starting from https://bit.ly/webml-workshop-survey) to be sent to the prospective participants.
14:56:51 <anssik> anssik: does this survey look good to you? would you respond? too long, too short, asking wrong questions? LGTM?
14:57:16 <anssik> anssik: we'll discuss this on the workshop program committee call right after this meeting, you're feedback is extremely valuable.
14:58:04 <Jonathan> Sure, happy to introduce the model-loader API
14:59:17 <anssik> TOPIC: Adjourn
14:59:40 <anssik> RRSAgent, draft minutes v2
14:59:40 <RRSAgent> I have made the request to generate https://www.w3.org/2020/04/30-webmachinelearning-minutes.html anssik
16:03:54 <anssik> Present+ Ping_Yu
16:03:58 <anssik> RRSAgent, draft minutes v2
16:03:58 <RRSAgent> I have made the request to generate https://www.w3.org/2020/04/30-webmachinelearning-minutes.html anssik
16:12:52 <myles> myles has joined #webmachinelearning