14:58:19 <RRSAgent> RRSAgent has joined #webmachinelearning
14:58:19 <RRSAgent> logging to https://www.w3.org/2020/03/19-webmachinelearning-irc
14:58:27 <Zakim> Zakim has joined #webmachinelearning
14:58:33 <anssik> RRSAgent, make logs public
14:58:39 <anssik> Meeting: WebML CG Teleconference – 19 March 2020
14:58:44 <anssik> Chair: Anssi
14:58:50 <anssik> Agenda: https://github.com/webmachinelearning/meetings/blob/master/telcons/2020-03-19-agenda.md
14:58:55 <anssik> Scribe: Anssi
14:59:00 <anssik> scribeNick: anssik
14:59:05 <anssik> Present+ Anssi_Kostiainen
14:59:32 <anssik> Present+ Nikhil_Thorat
15:00:11 <anssik> Present+ Rafael_Cintron
15:00:37 <Jonathan> Jonathan has joined #webmachinelearning
15:00:50 <Ningxin_hu> Ningxin_hu has joined #webmachinelearning
15:01:07 <Ningxin_Hu> Ningxin_Hu has joined #webmachinelearning
15:01:37 <RafaelCintron> RafaelCintron has joined #webmachinelearning
15:01:42 <walrusmcd> walrusmcd has joined #webmachinelearning
15:01:43 <anssik> Present+ Jonathan_Bingham
15:02:34 <anssik> Present+ Paul_McDaniel
15:02:42 <anssik> Present+ Chai_Chaoweeraprasit
15:02:43 <Ningxin_Hu> Present+ Ningxin_Hu
15:02:46 <Chai> Chai has joined #webmachinelearning
15:03:08 <anssik> Present+ Ningxin_Hu
15:03:15 <Rama> Rama has joined #webmachinelearning
15:04:03 <daniel_smilkov> daniel_smilkov has joined #webmachinelearning
15:04:20 <anssik> Present+ Ganesan_Ramalingam
15:04:28 <anssik> Present+ Daniel_Smilkov
15:04:37 <anssik> RRSAgent, draft minutes v2
15:04:37 <RRSAgent> I have made the request to generate https://www.w3.org/2020/03/19-webmachinelearning-minutes.html anssik
15:04:43 <walrusmcd> anssi - you might be dropping out audio
15:06:12 <anssik> TOPIC: matMul op definition
15:06:19 <anssik> anssik: ask for the group was to review PR:
15:06:24 <anssik> -> https://github.com/webmachinelearning/webnn/pull/49 Add matmul PR #49
15:07:21 <anssik> Ningxin_Hu: PR is based on the matMul signature proposal from Nikhil
15:08:45 <anssik> ... the proposal is strongly numpy inspired
15:09:25 <anssik> [Ningxin_Hu walking thought the details of the matMul op definition]
15:10:22 <anssik> Chai: might be a good idea to explain the output
15:10:33 <gregwhitworth> Present+ Greg Whitworth
15:11:00 <anssik> Ningxin_Hu: returns section is a brief description of the output, need more details?
15:11:42 <anssik> Chai: right, input in this case can be inclusive of all the cases enumerated, impacting the output, e.g. if dimension greater than 2 the output should match accordingly
15:11:49 <anssik> ... happy to help clarify that
15:12:41 <anssik> anssik: any other feedback on the matMul op definition?
15:13:07 <anssik> TOPIC: Handling unsupported OperandType
15:13:14 <anssik> -> https://github.com/webmachinelearning/webnn/issues/36 Handling unsupported OperandType issue #36
15:13:21 <anssik> -> https://github.com/webmachinelearning/webnn/pull/50 How to handle a failing compilation PR #50
15:14:06 <anssik> anssik: Chai submitted a PR #50 to fix issue #36 (thanks!). This PR proposes to update CompilationPreference to follow WebGL convention for context creation. Update the code sample to show how a certain compilation could fail such as when the model uses data type not supported by the hardware, and how to recover from the failure.
15:15:42 <anssik> Chai: wanted to clarify, I looked at WebGL, there's also a concept of context attribute, WebGLPowerPreference: :high perf, low power, 3rd_one_script_did_not_catch
15:16:10 <anssik> ... matches what we'd probably want to have for this API as well e.g. in cases when we fallback to CPU
15:16:44 <anssik> ... if float16 not possible on GPU, then can fallback to CPU, but if low power or high perf preference, we'd try that
15:17:22 <anssik> ... another preference in WebGL allows to create a high performance context, but that could fail(?) if not supported
15:17:43 <anssik> ... power preference of WebGL is something we should try to align with
15:19:24 <anssik> RafaelCintron: agree we need a default
15:19:38 <anssik> ... in WebGL if you ask high perf, you always get something
15:19:40 <anssik> ... never fails
15:20:16 <anssik> ... you always get something back regardless of the preference, it just sorts the list of GPUs
15:21:57 <anssik> Chai: in this case we have a model, that might not be supported by the GPU, so how we interpret high perf is more absolute in the case of WebNN API, similar to WebGL preference + flag
15:22:35 <anssik> RafaelCintron: in this case high perf preference is not power setting, just whether it is supported?
15:23:47 <anssik> Chai: if we try to compile a model, and GPU does not support it, we want the preference to say "we cannot create a high perf compilation of this model"
15:24:01 <anssik> ... semantics different from those of the WebGL behavior
15:24:25 <anssik> RafaelCintron: so you're saying, if the model requires a lot of copies then high perf pref would fail?
15:24:36 <anssik> Chai: right
15:25:03 <anssik> Chai: if developers do not care, they could retry compilation with CPU
15:25:25 <anssik> RafaelCintron: is there a threshold?
15:25:48 <anssik> Chai: you don't want to mix two devices midstream
15:26:06 <anssik> RafaelCintron: if 6 out of 10 ops can be done, does that qualify as a high perf GPU?
15:26:17 <anssik> Chai: hard to say how we define runtime performance
15:27:09 <anssik> ... given what we know w/o executing, we want to compile with high perf preferences, there could be some policy as long as it is not too slow, if the browser needs to walk the graph to determine which is fast and which is slow then that's too much
15:27:55 <anssik> ... the idea of high perf when user is certain the scenario/use case needs high fps, e.g. computer vision usages need high perf for good UX
15:27:59 <Ningxin_Hu> q+
15:29:04 <anssik> RafaelCintron: I see your point, if device-specific there's a fingerprinting concern, maybe there's no other way to specify this than how it's now spec'd
15:29:14 <anssik> Chai: important topic, should offline
15:31:09 <anssik> anssik: WebGL API fingerprintable surface is way larger
15:31:15 <anssik> Ningxin_Hu: thanks Chai for the PR
15:31:34 <anssik> ... ask for clarification, compilation preference: low power, high perf, fast answer
15:31:47 <anssik> ... in current spec not clear fast answer is the default
15:32:17 <anssik> ... fast answer is about responsiveness
15:32:53 <anssik> ... high perf is for use cases that need successive frames such as input from camera frames
15:33:18 <anssik> ... Android NN API has similar preferences
15:34:00 <anssik> ... on previous meeting, we wanted to handle the failure similar to WebGL, and define error codes to indicate the reason of failure
15:34:11 <anssik> q?
15:34:17 <anssik> ack Ningxin_Hu
15:36:52 <anssik> Chai: general observation, if we know what to recommend to the caller given a specific error code, then we should define a error code(s) for such a failure path
15:38:32 <anssik> Chai: any cases when we do not want to caller to fall back to another preference, not just run the model?
15:40:55 <anssik> RafaelCintron: in WebGL you can ask preference, high power, low power before context creation, after the context creation can ask if can if e.g. can do 16-bit float textures or not
15:41:01 <anssik> ... similarly for WebGPU
15:43:24 <anssik> anssik: do we have extension points in the API to add error codes later on? maybe have a separate issue for error codes?
15:44:21 <anssik> Chai: right
15:44:38 <Ningxin_Hu> q+
15:44:42 <anssik> anssack Ningxin_Hu
15:44:50 <anssik> ack Ningxin_Hu
15:45:06 <anssik> Ningxin_Hu: agree we want to separate error code
15:46:27 <anssik> ... can we create a separate issue for compilation preference change?
15:48:13 <anssik> anssik: do you want to offline this topic?
15:48:17 <anssik> Chai: sounds good
15:48:25 <anssik> Ningxin_Hu: similarly, ok with that
15:49:02 <anssik> TOPIC: Inference API to load and run a model
15:49:40 <anssik> anssik: I review WebML CG charter to figure out what changes if any are needed to allow us to incubate the load and run a model API proposal alongside the current graph builder API as a result of that review
15:50:11 <anssik> ... proposed resolution: add the load and run a model API to the group's scope, the group does not attempt to mandate a specific ML schema or format. No changes to the charter with this clarification.
15:50:24 <anssik> anssik: comments? concerns?
15:52:03 <anssik> Greg: when does the charter expire?
15:52:16 <anssik> anssik: does not expire, since this is a CG
15:52:48 <anssik> Greg: I think we should go ahead and circle back to charter if we hit a roadblock with charter definition
15:54:01 <gregwhitworth> >thanks
15:55:08 <anssik> Jonathan: would like to record a decision in the respective GH issue
15:57:21 <anssik> ... maybe someone could pull out the model loader part of WebNN into a separate spec?
15:57:52 <anssik> anssik: how about starting with all in one spec and split later?
15:58:00 <anssik> Jonathan: sounds okay too
15:58:13 <anssik> anssik: proposed resolution: add the load and run a model API to the group's scope, the group does not attempt to mandate a specific ML schema or format. No changes to the charter with this clarification.
15:59:07 <anssik> anssik: any concerns?
15:59:12 <anssik> RafaelCintron: LGTM
15:59:30 <anssik> Jonathan: LGTM
15:59:44 <anssik> Ningxin_Hu: LGTM
15:59:55 <anssik> RESOLUTION: Add the load and run a model API to the group's scope, the group does not attempt to mandate a specific ML schema or format. No changes to the charter with this clarification.
16:00:02 <anssik> RRSAgent, draft minutes v2
16:00:02 <RRSAgent> I have made the request to generate https://www.w3.org/2020/03/19-webmachinelearning-minutes.html anssik
16:01:21 <anssik> TOPIC: Adjourn
16:01:29 <anssik> anssik: Take care y'all
16:01:47 <anssik> RRSAgent, draft minutes v2
16:01:47 <RRSAgent> I have made the request to generate https://www.w3.org/2020/03/19-webmachinelearning-minutes.html anssik
16:19:31 <zkis> zkis has joined #webmachinelearning
16:42:35 <zkis> zkis has joined #webmachinelearning
18:18:35 <Zakim> Zakim has left #webmachinelearning
19:20:50 <zkis> zkis has joined #webmachinelearning
20:02:47 <zkis> zkis has joined #webmachinelearning