TSC Voting Members
Company | Name | |
TSC members | ZTE | Tao Liu, Shiying Jin, Liya Yuan |
China Mobile | Qi Sun | |
China Unicom | Tengfei Liu, Ao Li |
TSC Meeting Info
Adlik Technical Steering Committee meetings are open to the public and held every two weeks on Tuesday from 12/03/2019. You can see Adlik TSC calender and subscribe to its mailing list adlik-tsc@lists.lfai.foundation which will send you TSC meeting information on https://lists.lfai.foundation/g/adlik-tsc.
The TSC meeting agenda is published prior to the meeting. If you have a topic that you'd like to discuss, please email your requested agenda item to adlik-tsc@lists.lfai.foundation to be added to an upcoming meeting.
Zoom Details
Topic: Adlik TSC Meeting (Bi-Weekly)
Time: 03:00 PM Beijing, Shanghai
Every 2 weeks on Tuesday until Oct 18, 2021
Join Zoom Meeting
https://zoom.us/j/470930313
04/21
- Zoom meeting:TSC_20200421.mp4
Roll Call
Company
Attendee
TSC members
ZTE
Tao Liu, Shiying Jin, Liya Yuan, Bintao Han, Chengcan Wang
China Mobile
Qi Sun, Xiang Li China Unicom
Tengfei Liu
Other attendees
Status of open action items
- Action/ Further study on possible solution for mapping, procedure to apply Adlik in ORAN.
- Come up with a solution on next TSC meeting.
- Action/ Further study on possible solution for mapping, procedure to apply Adlik in ORAN.
- Follow up on Adlik+ORAN discussion
- Presentation on integrating model compiling(which can be done by Adlik) in ML workflow in ORAN- by Bingtao Han
- Before deployment, a model needs to go through the model optimization and compiling pipeline.
- Action/ Propose a CR for ORAN standardization.
- Discussion on ORAN previous proposal for model deployment - by Qi Sun
- Three possible solutions, mapping to the cloud, edge and device solutions of Adlik.
- Use case for release C need to be further discussed.
- Presentation on integrating model compiling(which can be done by Adlik) in ML workflow in ORAN- by Bingtao Han
Work report and plans
- Model optimizer:
- Coding and testing for quantilization and pruning are almost done, document in progress. Code will be pushed to github in two weeks. Will implement more algorithms afterwards.
- Serving engine: Performance test for TF lite and tensorflow 2.1 runtime.
- In progress: 1) Compiler for lite; 2) Some new features, e.g. interfaces for model onboarding.
- Plan to test Adlik on ARM.
- Model optimizer:
- Open discussion
04/08
Roll Call
Company
Attendee
TSC members
ZTE
Tao Liu, Shiying Jin, Liya Yuan, Bintao Han, Wei Meng
China Mobile
Qi Sun, Xiang Li China Unicom
Tengfei Liu, Ao Li
Other attendees
Status of open action items
- Invited presentation from ORAN - Qi Sun
- Introduced projects in ORAN B release https://wiki.o-ran-sc.org/pages/viewpage.action?pageId=1179662.
- Introduced ML flow involved in traffic steering use case.
- Introduction to Adlik and possible solution of integration in ORAN open source
- Adlik can be used to optimize ML models before deploying to RIC platform.
Work report and plans
- 2 features were merged in master: Support for TF lite runtime and ML runtime.
- Updated data of ResNet50 benchmark test.
- Open discussion
- To deploy ML/DL models in RIC platform, the following need to be taken into consideration:
- Platform expansion to support resource allocation
- Generic xAPP that supports interaction with other xAPP
- Will further discuss about proposing a new use case in ORAN Release C. A demo for the use case is expected to be delivered based on ORAN release B.
- Action/ Further study on possible solution for mapping, procedure to apply Adlik in ORAN.
- To deploy ML/DL models in RIC platform, the following need to be taken into consideration:
03/24
Roll Call
Company
Attendee
TSC members
ZTE
Tao Liu, Shiying Jin, Liya Yuan
China Mobile
China Unicom
Ao Li, Tengfei Liu
Other attendees
Status of open action items
Action/ GitHub documentation for releases.
Follow up on Adlik+MEC discussion
The pull request of Benchmark code is on Github.
Work report and plans
- Benchmark test.
- New runtime supporting ML algorithms.
- Pruning is 80% finished, will be pushed to Optimizer repo.
- Open discussion
- Questions from Tengfei Liu:
- Difference between Adlik and other inference framework
- Adlik is product driven and focus more on quick deployment of models in production environment, which enables heterogeneous computing better model scheduling, less resource consumption.
- What can we do?
- Provide your requirements derived from your practices in concrete cases.
- Use it in your projects and do techinical verification.
- Difference between Adlik and other inference framework
- Questions from Tengfei Liu:
03/10
Roll Call
Company
Attendee
TSC members
ZTE
Shiying Jin, Liya Yuan
China Mobile
China Unicom
Ao Li, Tengfei Liu
Other attendees
Status of open action items
Action/ Article about heterogeneous computing in deep learning inference- Published at https://www.infoq.cn/article/eg4KWZd1UoFwjSsUzfgt
Action/ GitHub documentation for releases
- Will be updated soon.
Follow up on Adlik+MEC discussion
Benchmark code in progress.
- Ao Li from China Unicom presented benchmark-test China Unicom.pptx, discussed possible scenarios, testing metrics and use cases.
Work report and plans
- developing features about support for tensorflow lite runtime, support for ML models.
02/25
Roll Call
Company
Attendee
TSC members
ZTE
Tao Liu, Shiying Jin, Liya Yuan
China Mobile
China Unicom
Tengfei Liu, Ao Li
Other attendees
Status of open action items
Action/ Usability and inference performance related issues to be created on Github.Benchmark feature has been created as an issue.
Action/ Article about heterogeneous computing in deep learning inference
A draft version will be delivered before this weekend.
Action/ GitHub documentation for releases
Under discussion.
Action/ Enable DCO on Github
Follow up on Adlik+MEC discussion
Benchmark Test presented by Tao Liu benchmarkTest-cn.pptx,about the procedure, metrics, environment of the performance test. Code to be expected on github by next TSC meeting.
Models like MobileNet, Inception V3 can be added as new test models.
Metrics like latency can be added as new metrics.
Work report and plans
Shiying Jin: planning support for machine learning models, enhanced interface for inference, etc.
02/11
- Roll Call
- ZTE: Tao Liu, Shiying Jin, Liya Yuan
Status of open action items
Action/ Further discussion on ingration of Adlik in MEC. ----- Discussion materials will be provided and discussed via emails first.Action/ Documentation for detailed information about how to use the compiler. ----- Will be included as a new feature.- Work report and plans
- Taoliu: Discussed with China Unicom and concluded some directions to work on: a. usability(ease the use of Adlik and provide automated optimzation of parameters); b. inference performance( testbed, standard testing including model, framework and metrics). Corresponding issues will be created on Github.
- Shiying Jin: planning RSE(Realtime smart engine) that supports the management of machine learning models.
- Liya Yuan: Since several meetups and outreach activities are yet to come, detailed document need to be prepared in advance.
12/17
Roll Call
Company
Attendee
TSC members
ZTE
Tao Liu, Shiying Jin, Liya Yuan
China Mobile
Qi Sun
China Unicom
Tengfei Liu
Other attendees
Adlik TSC chair election
TSC voted and approved Liya Yuan as Adlik TSC chair.
Status of open action items
Further discussion on ingration of Adlik in MEC. ----- Discussion materials will be provided and discussed via emails first.
Documentation for detailed information about how to use the compiler. ----- Will be included as a new feature.
Work report and plans
Roadmap is updated on Github.
FPGA compiler and runtime will be developed first.
Open Discussion
Call for more developers to contribute to Adlik, e.g. to enable support for more op.
12/03
Self introdution of TSC members and developers
Company
Attendee
TSC members
ZTE
Tao Liu, Shiying Jin, Liya Yuan
China Mobile
Qi Sun
China Unicom
Tengfei Liu, Ao Li
Other attendees
Vishnu Ram
Adlik code walkthrough
Introduction to the two sub projects, model compiler and model serving(Tao Liu and Shiying Jin).
Q: How to give input to the model compiler? A: It's shown in the simpleMNIST demo.
Q: I have to always use the name mnist.h5 and manually input the layers in this compile_model.py? A: Further documentation will show how to get these parameters and pass them to compiler.
Adlik work plans
New features in progress: support for ARM CPU and FPGA; support for quantilization(Tao Liu).
Open Discussion
Adlik intergration with MEC(Ao Li)
The interface requirements need further discussion.
Requirements for APIs can be written collaboratively with ITU-T FG ML5G.
Actions
Further discussion on ingration of Adlik in MEC.
Documentation for detailed information about how to use the compiler.