While Machine Learning is becoming more and more important in everyday life, designing efficient ML Systems and deploying them in the real world is becoming increasingly challenging, time consuming and costly. Researchers and engineers must keep pace with rapidly evolving software stacks and a Cambrian explosion of hardware platforms from the cloud to the edge. Such platforms have their own specific libraries, frameworks, APIs and specifications and often require repetitive, tedious and ad-hoc optimization of the whole model/software/hardware stack to trade off accuracy, latency, throughout, power consumption, size and costs depending on user requirements and constraints.
The Collective Knowledge framework (CK) is our attempt to develop a common plug&play infrastructure that can be used by the community similar to Wikipedia to learn how to solve above challenges and make it easier to co-design, benchmark, optimize and deploy Machine Learning Systems in the real world across continuously evolving software, hardware and data sets (see our ACM TechTalk for more details):
-
CK aims at providing a simple playground with minimal software dependencies to help researchers and practitioners share their knowledge in the form of reusable automation recipes with a unified Python API, CLI and meta description:
-
CK helps to organize software projects and Git repositories as a database of above automation recipes and related artifacts based on FAIR principles as described in our journal article (shorter pre-print). See examples of CK-compatible GitHub repositories:
We collaborated with the community to reproduce 150+ ML and Systems papers and implement the following reusable automation recipes in the CK format:
-
Portable meta package manager to automatically detect, install or rebuild various ML artifacts (ML models, data sets, frameworks, libraries, etc) across different platform and operating systems including Linux, Windows, MacOS and Android:
-
Portable manager for Python virtual environments: CK repo.
-
Portable workflows to support collaborative, reproducible and cross-platform benchmarking:
-
Portable workflows to automate MLPerf™ benchmark:
- End-to-end submission suite used by multiple organizations to automate the submission of MLPerf inference benchmark
- MLperf inference v1.1 results: MLCommons press-release, Datacenter results, Edge results
- Reproducibility studies for MLPerf inference benchmark v1.1 automated by CK
- Design space exploration of ML/SW/HW stacks and customizable visualization
- End-to-end submission suite used by multiple organizations to automate the submission of MLPerf inference benchmark
Please contact Grigori Fursin if you are interested to join this community effort!
- CK automations for unified benchmarking
- CK-based MLPerf inference benchmark automation example
- CK basics
The latest version of the CK automation suite supported by MLCommons™:
We plan to develop a new version of the CK framework (v3) within the MLCommons' Design Space Exploration workgroup - please contact Grigori Fursin to join this community effort!
- Automating MLPerf(tm) inference benchmark and packing ML models, data sets and frameworks as CK components with a unified API and meta description
- Developing customizable dashboards for MLPerf™ to help end-users select ML/SW/HW stacks on a Pareto frontier: aggregated MLPerf™ results
- Providing a common format to share artifacts at ML, systems and other conferences: video, Artifact Evaluation
- Redesigning CK together with the community based on user feedback: incubator
- Other real-world use cases from MLPerf™, Qualcomm, Arm, General Motors, IBM, the Raspberry Pi foundation, ACM and other great partners;
Follow this guide to install CK framework on your platform.
CK supports the following platforms:
As a host platform | As a target platform | |
---|---|---|
Generic Linux | ✓ | ✓ |
Linux (Arm) | ✓ | ✓ |
Raspberry Pi | ✓ | ✓ |
MacOS | ✓ | ± |
Windows | ✓ | ✓ |
Android | ± | ✓ |
iOS | TBD | TBD |
Bare-metal (edge devices) | - | ± |
Here we show how to pull a GitHub repo in the CK format and use a unified CK interface to compile and run any program (image corner detection in our case) with any compatible data set on any compatible platform:
python3 -m pip install ck
ck pull repo:mlcommons@ck-mlops
ck ls program:*susan*
ck search dataset --tags=jpeg
ck pull repo:ctuning-datasets-min
ck search dataset --tags=jpeg
ck detect soft:compiler.gcc
ck detect soft:compiler.llvm
ck show env --tags=compiler
ck compile program:image-corner-detection --speed
ck run program:image-corner-detection --repeat=1 --env.MY_ENV=123 --env.TEST=xyz
You can check output of this program in the following directory:
cd `ck find program:image-corner-detection`/tmp
ls
processed-image.pgm
You can now view this image with detected corners.
Check CK docs for further details.
We have prepared adaptive CK containers to demonstrate MLOps capabilities:
You can run them as follows:
ck pull repo:mlcommons@ck-mlops
ck build docker:ck-template-mlperf --tag=ubuntu-20.04
ck run docker:ck-template-mlperf --tag=ubuntu-20.04
You can create multiple virtual CK environments with templates to automatically install different CK packages and workflows, for example for MLPerf™ inference:
ck pull repo:mlcommons@ck-venv
ck create venv:test --template=mlperf-inference-main
ck ls venv
ck activate venv:test
ck pull repo:mlcommons@ck-mlops
ck install package --ask --tags=dataset,coco,val,2017,full
ck show env
All CK modules, automation actions and workflows are accessible as a micro-service with a unified JSON I/O API to make it easier to integrate them with web services and CI platforms as described here.
We have developed the cKnowledge.io portal to help the community organize and find all the CK workflows and components similar to PyPI:
- Search CK components
- Browse CK components
- Find reproduced results from papers
- Test CK workflows to benchmark and optimize ML Systems
The community provides Docker containers to test CK and components using different ML/SW/HW stacks (DSE).
- A set of Docker containers to test the basic CK functionality using some MLPerf inference benchmark workflows: https://github.com/mlcommons/ck-mlops/tree/main/docker/test-ck
Users can extend the CK functionality via CK modules or external GitHub reposities in the CK format as described here.
Please check this documentation if you want to extend the CK core functionality and modules.
Note, that we plan to redesign the CK framework to be more pythonic (we wrote the first prototype without OO to be able to port it to bare-metal devices in C but eventually we decided to drop this idea).
Please contact Grigori Fursin to join this community effort.
- Grigori Fursin (OctoML/MLCommons/cTuning foundation)
We would like to thank all contributors and collaborators for their support, fruitful discussions, and useful feedback! See more acknowledgments in the CK journal article.