IRC log of webmachinelearning on 2019-09-17
Timestamps are in UTC.
- 04:26:06 [RRSAgent]
- RRSAgent has joined #webmachinelearning
- 04:26:06 [RRSAgent]
- logging to https://www.w3.org/2019/09/17-webmachinelearning-irc
- 04:26:11 [Zakim]
- Zakim has joined #webmachinelearning
- 04:26:17 [anssik]
- RRSAgent, make logs public
- 04:26:28 [anssik]
- Meeting: WebML CG F2F Day 1 – 17 September 2019
- 04:26:32 [anssik]
- Chair: Anssi
- 04:26:36 [anssik]
- Agenda: https://github.com/webmachinelearning/meetings/blob/master/2019-09-17-fukuoka/
- 04:26:45 [niwamoto]
- niwamoto has joined #webmachinelearning
- 04:26:48 [anssik]
- Scribe: Anssi
- 04:26:50 [anssik]
- scribeNick: anssik
- 04:27:10 [anssik]
- Present+ Anssi_Kostiainen
- 04:27:26 [ningxin_hu]
- Present+ Ningxin_Hu
- 04:27:37 [takio]
- takio has joined #webmachinelearning
- 04:27:43 [niwamoto]
- Present+ Narifumi_Iwamoto
- 04:27:55 [Youngsun_Ryu]
- Present+ Youngsun_Ryu
- 04:27:58 [David_Marsh]
- Present+ David Marsh
- 04:28:10 [riju]
- riju has joined #webmachinelearning
- 04:28:20 [HelloFillip]
- Present+ Phil_Laszkowicz
- 04:28:24 [Kangchan]
- present+ Kangchan_Lee
- 04:28:25 [riju]
- Present+ Rijubrata Bhaumik
- 04:28:26 [takio]
- Present+ Takio_Yamaoka
- 04:32:15 [anssik]
- RRSAgent, draft minutes v2
- 04:32:15 [RRSAgent]
- I have made the request to generate https://www.w3.org/2019/09/17-webmachinelearning-minutes.html anssik
- 04:33:04 [anssik]
- TOPIC: Welcome and intros
- 04:33:24 [anssik]
- anssik: welcome to the WebML CG's 2nd F2F, happy to see both new and old faces around
- 04:33:58 [Bruce]
- Bruce has joined #webmachinelearning
- 04:34:37 [jdarpinian]
- jdarpinian has joined #webmachinelearning
- 04:34:37 [anssik]
- ... on the agenda today on Day 1: intros, custom operations, MLIR (Multi-Level Intermediate Representation) exploration, Operation set
- 04:34:48 [anssik]
- ... on Friday Day 2 exploratory topics, standards track next steps, W3C workshop planning
- 04:35:00 [anssik]
- anssik: Let's do a roundtable 30-sec intros: your affiliation & interests toward the group
- 04:35:37 [dino]
- dino has joined #webmachinelearning
- 04:35:43 [dino]
- present+
- 04:35:45 [anssik]
- anssik: I'm the chair working for Intel
- 04:36:06 [anssik]
- nikhil: working for Google, Deeplearn.js co-author, want to bring the ecosystem forward, not familiar with W3C
- 04:36:23 [anssik]
- ningxin_hu: Intel, CV and ML interest, OpenCV.js background
- 04:36:42 [anssik]
- kenneth: Intel architect, W3C TAG rep, overseeing the architecture of the Web
- 04:37:02 [jdarpinian]
- jdarpinian has joined #webmachinelearning
- 04:37:22 [Cortiz]
- Cortiz has joined #webmachinelearning
- 04:37:33 [anssik]
- Yongsun: Samsung, interested in ML in general
- 04:37:47 [anssik]
- s/replacethis/withthis/
- 04:38:13 [whsieh]
- whsieh has joined #webmachinelearning
- 04:38:13 [kimwooglae_]
- kimwooglae_ has joined #webmachinelearning
- 04:38:13 [anssik]
- Dave: Payments network with many members, just interested in ML
- 04:38:43 [anssik]
- Chimiming: University affiliated
- 04:38:48 [Chunming]
- Chunming has joined #webmachinelearning
- 04:39:31 [anssik]
- Dean: Apple, interested in everything the group does, not ML specialist but I'll do my best connecting Apple experts, I work on WebKit project also Safari
- 04:40:11 [anssik]
- Philip: Omnijar, working with DL for 13 years, with large companies, automotive, NVIDIA, ARM, interest to continue move commercial project to the Web
- 04:40:44 [anssik]
- Riju: Intel, Chromium developer, sensors, NFC, media capture, OpenCV, not using ML currently
- 04:41:15 [anssik]
- Kangchan: ETRI Korea, working on standards in ITU, ML as a Service
- 04:42:04 [anssik]
- Wenson: Apple, WebKit, interest in ML
- 04:42:26 [anssik]
- Diogo: Brazil W3C office, NLP background and interest
- 04:42:38 [Bruce]
- Bruce has joined #webmachinelearning
- 04:42:58 [anssik]
- Takio: Yahoo Japan, sensor processing, transcoding, interest in CV w/ ML
- 04:43:34 [anssik]
- Sangwhan: TAG member, used to work for Opera, CV startup not affiliated with Web, I also do NLP
- 04:43:53 [anssik]
- Frank: Inria France, curious of the group
- 04:43:59 [tung]
- tung has joined #webmachinelearning
- 04:44:40 [anssik]
- Belem: Intel, responsible to WebML polyfill
- 04:45:16 [anssik]
- James: Google, working on Chrome, WebGL/GPU, interested in ML in Chrome
- 04:46:08 [anssik]
- TOPIC: Custom operations
- 04:49:13 [hyojin]
- hyojin has joined #webmachinelearning
- 04:49:15 [Big_Screen]
- https://docs.google.com/presentation/d/1KGRc1RnnYt_1JK2Pk6r2xRkD60v4F8jc4beHMv0crng/edit#slide=id.p
- 04:50:44 [anssik]
- q+ to say something
- 04:50:53 [anssik]
- ack anssik
- 04:50:53 [Zakim]
- anssik, you wanted to say something
- 04:51:17 [anssik]
- [ningxin presents the slides]
- 04:53:17 [anssik]
- ningxin_hu: ML field is fast moving. Model architecture and the ops are evolving quickly. This leads JS ML frameworks usually have big op set (e.g. TF.js has over 200 ops)
- 04:53:25 [anssik]
- ... Today’s framework’s ops are implemented in WebGL and WASM, and WebGPU
- 04:53:32 [anssik]
- ... WebNN’s built-in op set that focuses on hardware acceleration will be small and grow slowly
- 04:53:52 [anssik]
- ... Problem: It demands a way for library authors to write ops that can interop with built-in ops.
- 04:54:08 [anssik]
- Options: WebNN built-in ops interop with framework ops in WASM and WebGL/WebGPU (focus of this investigation)
- 04:54:49 [anssik]
- Kenneth: can you mix Wasm and WebNN ops?
- 04:55:14 [anssik]
- Shangwan: there's a GPU-CPU transfer with a performance cost
- 04:55:38 [anssik]
- ... WebNN provides a way to write custom op by a domain specific language (e.g. Kai’s proposal) (future exploration)
- 04:55:57 [anssik]
- ningxin_hu: next subtopic, WebNN-WebGPU Interop
- 04:56:18 [anssik]
- [showing example code of Conv + Add + Relu by TF.js WebGPU]
- 04:56:42 [anssik]
- RRSAgent, draft minutes v2
- 04:56:42 [RRSAgent]
- I have made the request to generate https://www.w3.org/2019/09/17-webmachinelearning-minutes.html anssik
- 04:57:41 [anssik]
- [showing example of compile WebNN op for WebGPU device]
- 04:59:48 [anssik]
- [scribe sees ~30 participants, not all names recorded in minutes]
- 05:00:10 [HelloFillip]
- HelloFillip has joined #webmachinelearning
- 05:00:25 [anssik]
- [showing example of execute WebNN's op with WebGPU op]
- 05:00:29 [junwei]
- junwei has joined #webmachinelearning
- 05:02:16 [anssik]
- -> https://docs.google.com/presentation/d/1KGRc1RnnYt_1JK2Pk6r2xRkD60v4F8jc4beHMv0crng/ WebNN Interop Investigation slides
- 05:03:16 [junwei]
- junwei has joined #webmachinelearning
- 05:03:31 [anssik]
- [ningxin showing a demo on his laptop]
- 05:03:52 [anssik]
- ningxin_hu: custom build of Chromium on macOS
- 05:04:03 [lisha]
- lisha has joined #webmachinelearning
- 05:05:08 [kimwooglae]
- kimwooglae has joined #webmachinelearning
- 05:05:20 [ningxin_hu]
- https://docs.google.com/presentation/d/1KGRc1RnnYt_1JK2Pk6r2xRkD60v4F8jc4beHMv0crng/edit?usp=sharing
- 05:05:40 [Franck]
- Franck has joined #webmachinelearning
- 05:05:44 [ningxin_hu]
- conv input dims: [1,100,100,100] and filter dims: [3,3,100,100] WebGPU conv2d/add/relu elapsed time: 60.81 ms WebNN conv2d interops with WebGPU add/relu via ArrayBuffer elapsed time: 39.67 ms WebNN conv2d interops with WebGPU add/relu via WebGPUBuffer elapsed time: 22.11 ms WebNN conv2d with fused add/relu elapsed time: 21.11 ms
- 05:06:32 [anssik]
- [above pasted text is an output of test case of TF.js sets backend as WebGPU]
- 05:07:50 [anssik]
- sangwhan: is the Chromium source available?
- 05:07:57 [anssik]
- ningxin_hu: that's available
- 05:08:58 [anssik]
- nikhil: how fast is the readback?
- 05:09:08 [anssik]
- ningxin_hu: not yet tested that
- 05:09:24 [anssik]
- dino: you can't use MPS, why is that?
- 05:09:40 [anssik]
- ningxin_hu: different memory layout internally
- 05:10:15 [anssik]
- dino: can you show conv operations, what they are doing?
- 05:10:24 [yoshiaki_]
- yoshiaki_ has joined #webmachinelearning
- 05:10:33 [anssik]
- ... I was expected to see a custom op, i.e. shader code
- 05:10:45 [anssik]
- ningxin_hu: shader code is inside TF.js
- 05:11:47 [anssik]
- ningxin_hu: subtopic, POC Implementation on MPS
- 05:11:57 [anssik]
- ... Reuse the same MTLDevice associated with WebGPUDevice.
- 05:12:07 [anssik]
- ... Get the MTLBuffer associated with input and output WebGPUBuffer.
- 05:12:14 [anssik]
- ... Allocate MPSImage for inputs with MTLDevice.
- 05:12:21 [anssik]
- ... Create MTLCommandBuffer from MTLQueue associated with WebGPUDevice.
- 05:12:28 [anssik]
- ... Encode a compute shader that copies and reorders data from MTLBuffer to MPSImage (MPSImage layout).
- 05:12:48 [anssik]
- dino: is this a custom WebGPU implementation? Where you decide you MPS?
- 05:13:15 [anssik]
- ... TF.js running on top of WebGPU
- 05:13:25 [anssik]
- ... this is an impl of WebNN, not TF on for of Chromium
- 05:13:44 [anssik]
- ... using WebGPU infra underneath it has platform implementation e.g. MPS
- 05:13:55 [anssik]
- ningxin_hu: Encode MPSNNGraph/MPSCNNKernel to MTLCommandBuffer
- 05:14:02 [anssik]
- ... Encode a compute shader that copies and reorders data from output MPSImage to output MTLBuffer.
- 05:14:10 [anssik]
- ... Commit MTLCommandBuffer.
- 05:14:37 [anssik]
- ningxin_hu: Performance Summary
- 05:15:24 [anssik]
- ... Inference time (ms)
- 05:16:00 [anssik]
- ... WebGPU conv/add/relu 61.31
- 05:16:14 [anssik]
- ... WebNN conv interops with WebGPU add/relu via ArrayBuffer 43.42
- 05:16:28 [anssik]
- ... WebNN conv interops with WebGPU add/relu via WebGPUBuffer 23.06
- 05:16:34 [anssik]
- ... WebNN conv with fused add/relu 21.25
- 05:16:52 [anssik]
- ningxin_hu: Copying/Reordering Optimization
- 05:17:01 [anssik]
- ... Inference time (ms)
- 05:17:16 [anssik]
- WebGPU conv x2 112.96
- 05:17:22 [anssik]
- WebNN conv + WebGPU conv 67.33
- 05:17:38 [anssik]
- ... WebNN conv x2 with reordering 24.53
- 05:17:54 [anssik]
- s/ WebGPU conv x2 112.96/... WebGPU conv x2 112.96/
- 05:18:00 [yoshiaki]
- yoshiaki has joined #webmachinelearning
- 05:18:03 [anssik]
- s/WebNN conv + WebGPU conv 67.33/... WebNN conv + WebGPU conv 67.33/
- 05:18:16 [yoshiaki_]
- yoshiaki_ has joined #webmachinelearning
- 05:18:45 [anssik]
- sangwhan: with this design, vendors targeting a single type of accelerator, what are the implications?
- 05:19:12 [anssik]
- ... if you were to implement this in a general browser, not OS bound, you'd have multiple accelerators, what's the story?
- 05:19:41 [anssik]
- ... you'd need to have compilers for every accelerator
- 05:19:51 [anssik]
- ... implementability question
- 05:20:20 [anssik]
- ... if you'd use the platform APIs, it'd be fine, but they can be limited in terms of support
- 05:20:48 [anssik]
- dino: Apple's perspective is we want to offload to the hardware as much as possible
- 05:21:26 [anssik]
- sangwhan: when testing the POC, did the inference affect the ref(?)
- 05:21:47 [anssik]
- dino: same issue with WebGL/GPU
- 05:22:20 [anssik]
- ... issue if the background task freezes the computer
- 05:22:44 [anssik]
- ... battery and perf benefit for going to ML hardware
- 05:23:01 [anssik]
- sangwhan: would be nice if everyone had these purpose-built accelerators
- 05:23:11 [anssik]
- ... curious of implications of that
- 05:23:24 [anssik]
- dino: not sure what Android devices have AI accelerators
- 05:23:48 [anssik]
- sangwhan: based on testing, could be NEON accelerated, or GPU, whatever the vendor had time to do
- 05:24:20 [anssik]
- nikhil: also good to benchmark readback times from those accelerators
- 05:24:22 [yoshiaki]
- yoshiaki has joined #webmachinelearning
- 05:25:17 [anssik]
- [skipping slides to Summary of WebNN-WASM interop slide]
- 05:25:28 [anssik]
- ningxin_hu: WebNN ops allow to access vendor specific CPU acceleration
- 05:25:36 [anssik]
- ... Interop between WASM ops and WebNN op has overhead
- 05:25:42 [anssik]
- ... Memory copying between WASM heap and WebNN backend
- 05:25:49 [anssik]
- ... Memory reordering, e.g. MKL-DNN blocked layout
- 05:25:58 [anssik]
- ... Execute WebNN ops chain with opaque operands can avoid unnecessary overhead
- 05:26:24 [anssik]
- ningxin_hu: Proposal
- 05:26:43 [anssik]
- ... Support key ops that access hardware acceleration (#17) E.g. conv2d and matmul
- 05:26:57 [anssik]
- ... Support compiling and executing ops for devices (new issue?) CPU or GPU
- 05:27:09 [anssik]
- ... Support interop with WebAssembly and WebGPU compute shader
- 05:27:18 [anssik]
- ... Sharing ArrayBuffer with WASM op
- 05:27:27 [anssik]
- ... Sharing WebGPUBuffer with WebGPU op (new issue?)
- 05:28:00 [anssik]
- ... Support interop with WebAssembly and WebGPU compute shader
- 05:28:07 [anssik]
- ... - Sharing ArrayBuffer with WASM op
- 05:28:17 [anssik]
- ... - Sharing WebGPUBuffer with WebGPU op (new issue?)
- 05:28:25 [anssik]
- ... Support executing ops chain with opaque operands (#11)
- 05:28:33 [anssik]
- ... - Leverage device optimized memory layout and avoid unnecessary memory reordering
- 05:28:41 [anssik]
- ... Explore custom op support by DSL (new issue?)
- 05:30:03 [anssik]
- dino: how do these numbers compare with true native frameworks, CoreML, TensorFlow native?
- 05:31:29 [anssik]
- ningxin_hu: 10% WebNN overhead over native
- 05:31:48 [anssik]
- nikhil: TensorFlow/WebGL vs. CUDA, CUDA 10x faster
- 05:32:06 [anssik]
- ???: what kind of model do you use?
- 05:32:46 [anssik]
- ningxin_hu: we have multiple models for this experiment, we use conv kernels, MobileNet, Inception, ResNet50
- 05:33:12 [anssik]
- ... on our website we have bigger models, the model size constraints us
- 05:34:09 [anssik]
- nikhil: CPU and non-CPU accelerators an issue, how to consider them in the context of custom ops, understand readbacks
- 05:34:40 [anssik]
- ???: what is the focus in terms of hardware targets of this group?
- 05:34:58 [anssik]
- ningxin_hu: we have experience on Android phone with an AI accelerator, close to native perf
- 05:35:18 [yoshiaki]
- yoshiaki has joined #webmachinelearning
- 05:35:59 [anssik]
- ???: what is the scope of this work? Recommendation to define a higher level abstraction to be flexible
- 05:36:25 [anssik]
- [hearing no concerns for the proposed tasks to investigate further]
- 05:36:55 [anssik]
- ningxin_hu: I'm willing to take "Support compiling and executing ops for devices (new issue?)" task
- 05:37:28 [anssik]
- ... maybe Kai could help with "Explore custom op support by DSL (new issue?)"
- 05:38:34 [anssik]
- dino: Apple could look at "Support key ops that access hardware acceleration (#17)" and provide feedback for that
- 05:39:15 [anssik]
- nikhil: just filed issues for conv2d and matmul
- 05:39:28 [anssik]
- https://github.com/webmachinelearning/webnn/issues/27
- 05:39:34 [anssik]
- https://github.com/webmachinelearning/webnn/issues/28
- 05:39:57 [anssik]
- ... will move forward with issues #27 and #28
- 05:40:45 [anssik]
- Topic: MLIR
- 05:41:07 [anssik]
- nikhil: disclaimer, I'm not a compiler person, but talked with Google experts on that field
- 05:41:22 [anssik]
- RRSAgent, draft minutes v2
- 05:41:22 [RRSAgent]
- I have made the request to generate https://www.w3.org/2019/09/17-webmachinelearning-minutes.html anssik
- 05:42:07 [anssik]
- nikhil: we'll not proposing MLIR, just exploring this area
- 05:42:14 [jdarpinian]
- do you have a link to the slides?
- 05:43:00 [anssik]
- -> https://docs.google.com/presentation/d/1vv-pFsTqAVITtx3RwmEs-g7YRK1PD9APSIuice88aSI/ MLIR slides by Nikhil
- 05:43:20 [anssik]
- [nikhil presenting MLIR slides]
- 05:44:49 [anssik]
- ???: XLA compiler spits out LLVM IR already?
- 05:44:54 [anssik]
- nikhil: correct
- 05:45:04 [yoshiaki]
- yoshiaki has joined #webmachinelearning
- 05:46:11 [anssik]
- ... Domain specific optimizations, progressive lowering
- 05:46:34 [anssik]
- ... The TensorFlow compiler ecosystem has many “Graph” IRs, each with challenges
- 05:47:47 [anssik]
- ... Domain Specific IRs, Great! High-level domain-specific optimizations; Progressive lowering encourages reuse between levels
- 05:48:24 [anssik]
- ... Not great!
- 05:48:29 [anssik]
- ... Huge expense to build this infrastructure
- 05:48:34 [anssik]
- ... Reimplementation of all the same stuff:
- 05:48:40 [anssik]
- ... pass managers, location tracking, use-def chains, inlining,
- 05:48:47 [anssik]
- ... constant folding, CSE, testing tools, ….
- 05:48:51 [anssik]
- ... Innovations in one community don’t benefit the others
- 05:49:19 [anssik]
- nikhil: let's talk about what is MLIR
- 05:50:21 [anssik]
- ... TensorFlow
- 05:50:21 [anssik]
- ... "An open source machine learning framework for everyone"
- 05:50:21 [anssik]
- ... Multi-Level Intermediate Representation
- 05:50:21 [anssik]
- ... "An open source program optimization framework for ... everyone"
- 05:50:21 [anssik]
- ... Abstraction Building Toolkit
- 05:50:22 [anssik]
- ... Reusable set of compiler passes for higher abstractions
- 05:50:22 [anssik]
- ... Targeting analysis/program optimization/code generation
- 05:50:22 [anssik]
- ... Open governance and part of LLVM
- 05:50:48 [anssik]
- nikhil: MLIR has wide support across industry
- 05:51:09 [yoshiaki_]
- yoshiaki_ has joined #webmachinelearning
- 05:51:19 [anssik]
- nikhil: Extensible Operations Allow Multi-Level IR
- 05:52:37 [jc]
- jc has joined #webmachinelearning
- 05:52:43 [anssik]
- ... MLIR “Dialects”: Families of defined operations
- 05:53:16 [anssik]
- ... Example Dialects:
- 05:53:16 [anssik]
- ... TensorFlow, LLVM IR, XLA HLO, TF Lite, Swift SIL…
- 05:53:16 [anssik]
- ... Dialects can define:
- 05:53:16 [anssik]
- ... Sets of defined operations
- 05:53:16 [anssik]
- ... Entirely custom type system
- 05:53:16 [anssik]
- ... Customization hooks
- 05:53:16 [anssik]
- ... Constant folding, decoding
- 05:53:18 [anssik]
- ... Operation can define:
- 05:53:18 [anssik]
- ... Invariants on # operands, results, attributes, etc
- 05:53:18 [anssik]
- ... Custom parser, printer, verifier, …
- 05:53:37 [anssik]
- nikhil: MLIR Type System - some examples
- 05:53:39 [yoshiaki]
- yoshiaki has joined #webmachinelearning
- 05:53:58 [anssik]
- ... Scalars:
- 05:53:58 [anssik]
- ... f16, bf16, f32, … i1, i8, i16, i32, … i3, i4, i7, i57, …
- 05:53:58 [anssik]
- ... Vectors:
- 05:53:58 [anssik]
- ... vector<4 x f32> vector<4x4 x f16> etc.
- 05:53:59 [anssik]
- ... Tensors, including dynamic shape and rank:
- 05:53:59 [anssik]
- ... tensor<4x4 x f32> tensor<4x?x?x17x? x f32> tensor<* x f32>
- 05:53:59 [anssik]
- ... Others: functions, memory buffers, quantized integers, other ... TensorFlow stuff, ...
- 05:53:59 [anssik]
- ... Extensible!!
- 05:55:58 [anssik]
- nikhil: Applications of MLIR
- 05:56:05 [anssik]
- ... TensorFlow Lite Converter
- 05:56:30 [anssik]
- ... One of the focusses: Usability
- 05:56:45 [anssik]
- ... Usability of TOCO top complaint among TFLite users
- 05:56:46 [anssik]
- ... Debugging
- 05:56:53 [anssik]
- ... Report why a model failed to convert
- 05:57:01 [anssik]
- ... Dialect types enable more checking & better reporting
- 05:57:03 [yoshiaki]
- yoshiaki has joined #webmachinelearning
- 05:58:51 [anssik]
- ... [MLIR] for the Web?
- 05:59:08 [anssik]
- ... Some facts from MLIR investigations
- 05:59:14 [anssik]
- ... Operator expansion is about 25% YoY for TensorFlow
- 05:59:20 [anssik]
- ... Hardware vendors will implement dialects
- 05:59:50 [anssik]
- ... Open governance
- 06:00:29 [anssik]
- riju: regarding operator expansion, is there a fallback mechanism, even if with performance penalty?
- 06:00:37 [anssik]
- nikhil: we'd need to use e.g. a Wasm polyfill
- 06:01:36 [anssik]
- ... MLIR dialect on the web -- thoughts
- 06:02:13 [anssik]
- ... No backwards compatible guarantees today from MLIR
- 06:02:13 [anssik]
- ... A dialect could be invented that is backwards compatible
- 06:02:13 [anssik]
- ... What does maintaining this look like?
- 06:02:13 [anssik]
- ... Web sourcemaps => python code
- 06:02:13 [anssik]
- ... Immediately tells you whether python code will execute in browser
- 06:02:28 [anssik]
- kenneth: web needs backwards compat, and we do not really do versioning on the Web
- 06:02:46 [anssik]
- nikhil: how maintaining backwards compatibility could happen?
- 06:03:00 [jc]
- jc has joined #webmachinelearning
- 06:03:42 [anssik]
- dino: LLVM IR is a well-suited as a web transport format
- 06:04:20 [yoshiaki_]
- yoshiaki_ has joined #webmachinelearning
- 06:04:20 [whsieh]
- ^ *not* well-suited?
- 06:04:31 [anssik]
- ... a lot of lowering, what is the improvement?
- 06:04:49 [anssik]
- s/well-suited/not well-suited/
- 06:05:18 [anssik]
- RRSAgent, draft minutes v2
- 06:05:18 [RRSAgent]
- I have made the request to generate https://www.w3.org/2019/09/17-webmachinelearning-minutes.html anssik
- 06:07:05 [anssik]
- dino: what is the scope of the group, all models interop with all devices?
- 06:07:22 [sushrajaMSFT]
- sushrajaMSFT has joined #webmachinelearning
- 06:08:39 [anssik]
- ... we could start with a set of ops everyone supports
- 06:09:05 [anssik]
- nikhil: initially we wanted to support all ops
- 06:09:24 [anssik]
- ... then understood better growing the set slowly is a better approach
- 06:10:14 [anssik]
- dino: our fear is, and I can be wrong, if the ecosystem becomes skewed toward TF models, so that those get hardware acceleration while some other models might not
- 06:10:31 [anssik]
- nikhil: as a group we can grow that set so that it does not happen
- 06:10:57 [anssik]
- dino: TF is growing fast, how's hardware adding ops?
- 06:11:20 [anssik]
- nikhil: I think hardware vendors add new ops more slowly
- 06:11:34 [anssik]
- kenneth: do any ops go away with time?
- 06:12:02 [anssik]
- riju: any kind of ranking within these ops, what are used the most?
- 06:12:07 [jc]
- jc has joined #webmachinelearning
- 06:12:15 [anssik]
- nikhil: TF has it, not sure if can make that data public
- 06:13:56 [anssik]
- Philip: Swift for TF was good experience from usability perspecticve
- 06:14:13 [anssik]
- ... ML not a domain of data scientists for any longer, need good dev ergonomics
- 06:14:30 [Franck]
- Franck has joined #webmachinelearning
- 06:14:43 [anssik]
- ningxin_hu: on which level of abstraction would the Web dialect of MLIR sit on?
- 06:15:35 [HelloFillip]
- HelloFillip has joined #webmachinelearning
- 06:15:39 [anssik]
- nikhil: lower level things would evolve more slowly, but not sure at this point on which level the web dialect should be at
- 06:16:01 [anssik]
- dino: generally Apple's position is that a high-level abstraction works well on the Web since it allows implementations to optimize
- 06:16:18 [anssik]
- ... we don't have a huge dataset, but JS is a good example
- 06:16:34 [anssik]
- ... no enough data yet how Wasm goes
- 06:16:53 [anssik]
- ... if we did a Web dialect, it would be something like that, but we'd make it a bit more higher-level than LLVM IR
- 06:17:37 [anssik]
- nikhil: I'm wondering whether there's a level of abstraction between ops and LLVM IR we should target
- 06:19:44 [zkis]
- zkis has joined #webmachinelearning
- 06:20:30 [anssik]
- anssik: what would be good next steps for the group re MLIR tasks?
- 06:20:48 [anssik]
- nikhil: talking to MLIR people, it seems a bit too early still since moving target
- 06:21:46 [anssik]
- ... concretely, I can try to figure out which ops are used, how many times an op is called
- 06:22:30 [yuta]
- yuta has joined #webmachinelearning
- 06:22:38 [anssik]
- RRSAgent, draft minutes v2
- 06:22:38 [RRSAgent]
- I have made the request to generate https://www.w3.org/2019/09/17-webmachinelearning-minutes.html anssik
- 06:24:01 [anssik]
- RRSAgent, draft minutes v2
- 06:24:01 [RRSAgent]
- I have made the request to generate https://www.w3.org/2019/09/17-webmachinelearning-minutes.html anssik
- 06:24:24 [HelloFillip]
- HelloFillip has joined #webmachinelearning
- 06:24:30 [HelloFillip]
- The link to Chris's talk on Swift for TensorFlow can be found here (as an example for other languages): https://www.youtube.com/watch?v=s65BigoMV_I
- 06:25:56 [anssik]
- we'll defer Day 1 3rd topic "operation set" to Day 2 on Friday
- 06:26:09 [anssik]
- thanks for attending, we'll see again on Friday!
- 06:26:15 [anssik]
- Topic: Adjourn
- 06:26:17 [anssik]
- RRSAgent, draft minutes v2
- 06:26:17 [RRSAgent]
- I have made the request to generate https://www.w3.org/2019/09/17-webmachinelearning-minutes.html anssik
- 06:26:22 [belem]
- Thanks Anssi!
- 06:29:10 [anssik]
- Present+ Nikhil_Thorat
- 06:29:12 [anssik]
- RRSAgent, draft minutes v2
- 06:29:12 [RRSAgent]
- I have made the request to generate https://www.w3.org/2019/09/17-webmachinelearning-minutes.html anssik
- 06:31:03 [jc]
- jc has joined #webmachinelearning
- 06:32:09 [anssik]
- Present+ Heejin_Chung
- 06:32:53 [anssik]
- Present+ Philip_Laszkowicz
- 06:33:11 [anssik]
- Present+ Diogo_Cortiz
- 06:33:36 [anssik]
- Present+ Dean_Jackson
- 06:34:05 [anssik]
- Present+ Wooglae_Kim
- 06:34:25 [anssik]
- RRSAgent, draft minutes v2
- 06:34:25 [RRSAgent]
- I have made the request to generate https://www.w3.org/2019/09/17-webmachinelearning-minutes.html anssik
- 06:36:01 [anssik]
- Present+ David_Marsh
- 06:36:25 [anssik]
- Present+ Kenneth_Christiansen
- 06:37:15 [anssik]
- Present+ Wenson_Hsieh
- 06:37:38 [anssik]
- Present+ A
- 06:37:42 [dino]
- dino has joined #webmachinelearning
- 06:37:55 [anssik]
- Present+ Takio_Yamaoka
- 06:38:02 [anssik]
- s/Present+ A//
- 06:38:25 [anssik]
- Present+ Sangwhan_Moon
- 06:39:36 [anssik]
- Present+ Belem_Zhang_(remote)
- 06:39:54 [anssik]
- Present+ James_Darpinian_(remote)
- 06:39:59 [anssik]
- RRSAgent, draft minutes v2
- 06:39:59 [RRSAgent]
- I have made the request to generate https://www.w3.org/2019/09/17-webmachinelearning-minutes.html anssik
- 06:41:02 [anssik]
- Present+ Frank_?
- 06:41:13 [anssik]
- RRSAgent, draft minutes v2
- 06:41:13 [RRSAgent]
- I have made the request to generate https://www.w3.org/2019/09/17-webmachinelearning-minutes.html anssik
- 06:50:29 [jc]
- jc has joined #webmachinelearning
- 06:50:38 [jc]
- jc has joined #webmachinelearning
- 06:57:35 [jc]
- jc has joined #webmachinelearning
- 06:58:29 [yoshiaki]
- yoshiaki has joined #webmachinelearning
- 07:02:33 [jc]
- jc has joined #webmachinelearning
- 07:03:55 [whsieh]
- whsieh has joined #webmachinelearning
- 07:07:18 [dino]
- dino has joined #webmachinelearning
- 07:11:16 [whsieh]
- whsieh has joined #webmachinelearning
- 07:11:35 [whsieh]
- whsieh has left #webmachinelearning
- 07:18:22 [yoshiaki]
- yoshiaki has joined #webmachinelearning
- 07:38:12 [jc]
- jc has joined #webmachinelearning
- 07:41:15 [yoshiaki]
- yoshiaki has joined #webmachinelearning
- 07:44:09 [Chunming]
- Chunming has joined #webmachinelearning
- 07:44:43 [jc]
- jc has joined #webmachinelearning
- 07:54:45 [jc]
- jc has joined #webmachinelearning
- 08:12:09 [yoshiaki_]
- yoshiaki_ has joined #webmachinelearning
- 08:13:59 [jc]
- jc has joined #webmachinelearning
- 08:34:52 [jc]
- jc has joined #webmachinelearning
- 08:47:20 [jc]
- jc has joined #webmachinelearning
- 08:51:01 [jc]
- jc has joined #webmachinelearning
- 09:02:10 [jc]
- jc has joined #webmachinelearning
- 09:02:24 [Zakim]
- Zakim has left #webmachinelearning
- 09:22:30 [yoshiaki]
- yoshiaki has joined #webmachinelearning
- 09:47:21 [yoshiaki]
- yoshiaki has joined #webmachinelearning
- 10:21:34 [zkis]
- zkis has joined #webmachinelearning
- 11:58:14 [zkis]
- zkis has joined #webmachinelearning
- 12:11:12 [Chunming]
- Chunming has joined #webmachinelearning
- 12:30:16 [Chunming]
- Chunming has joined #webmachinelearning
- 12:32:46 [zkis_]
- zkis_ has joined #webmachinelearning
- 12:46:19 [dino]
- dino has joined #webmachinelearning
- 13:35:15 [zkis]
- zkis has joined #webmachinelearning
- 13:55:14 [Chunming]
- Chunming has joined #webmachinelearning
- 14:33:24 [yoshiaki]
- yoshiaki has joined #webmachinelearning
- 17:49:47 [zkis]
- zkis has joined #webmachinelearning
- 18:21:17 [zkis_]
- zkis_ has joined #webmachinelearning
- 18:29:54 [zkis_]
- zkis_ has joined #webmachinelearning
- 18:47:57 [zkis__]
- zkis__ has joined #webmachinelearning
- 19:05:15 [zkis__]
- zkis__ has joined #webmachinelearning
- 19:24:46 [zkis_]
- zkis_ has joined #webmachinelearning