Integration of Caffe2 to Collective Knowledge workflow framework to provide unified CK JSON API for AI (customized builds across diverse libraries and hardware, unified AI API, collaborative experiments, performance optimization and model/data set tuning):
Python Shell Batchfile Arc
Switch branches/tags
Nothing to show
Clone or download

Unification of AI for collaborative experimentation and optimization using Collective Knowledge workflow framework with common JSON API

logo logo License

Note that Caffe2 moved to PyTorch GitHub source tree, so some packages here may not work correctly.


After spending most of our "research" time not on AI innovation but on dealing with numerous and ever changing AI engines, their API, and the whole software and hardware stack, we decided to take an alternative approach.


We started adding existing AI frameworks including Caffe, Caffe2, TensorFlow, MXNet, CK-PyTorch, CNTK and MVNC.

to the non-intrusive open-source Collective Knowledge workflow framework (CK). CK allows to plug in various versions of AI frameworks together with libraries, compilers, tools, models and data sets as unified and reusable components with JSON API, automate and customize their installation across Linux, Windows, MacOS and Android (rather than using ad-hoc scripts) and provide simple JSON API for common operations such as prediction and training (see demo).

At the same time, CK allows us to continuously optimize (1, 2) the whole AI stack (SW/HW/models) across diverse platforms from mobile devices and IoT to supercomputers in terms of accuracy, execution time, power consumption, resource usage and other costs with the help of the community (see public CK repo and reproducible and CK-powered AI/SW/HW co-design competitions at ACM/IEEE conferences).

See for more details.

Coordination of development

Example of Caffe and Caffe2 unified CPU installation on Ubuntu via CK

Note that you need Python 2.x!

$ sudo apt install coreutils build-essential make cmake wget git python python-pip
$ sudo pip install jupyter pandas numpy scipy matplotlib scikit-image scikit-learn pyyaml protobuf future google
$ sudo pip install --upgrade beautifulsoup4
$ sudo pip install --upgrade html5lib

$ sudo pip install ck

$ ck pull repo --url=
$ ck pull repo:ck-caffe2

$ ck install package --tags=lib,caffe,vcpu --env.CAFFE_BUILD_PYTHON=ON
$ ck install package --tags=lib,caffe2,vcpu

Using compiled Caffe (with python) as a virtual environment:

$ ck virtual env --tags=lib,caffe
$ ipython

Dependencies for Windows

We tested it with Anaconda Python 2.x (should be in path for pip). Note that only stable package currently works - 0.8.1 (there is an issue with custom protobuf in the master branch)

$ pip install jupyter pandas numpy scipy matplotlib scikit-image scikit-learn pyyaml protobuf future google
$ pip install flask glog graphviz  pydot python-nvd3 pyyaml requests scikit-image setuptools tornado
$ pip install future hypothesis six
$ pip install --upgrade beautifulsoup4
$ pip install --upgrade html5lib

Example of Caffe2 unified CPU installation on Raspberry Pi 3+

$ sudo apt install coreutils build-essential make cmake python python-pip libblas-dev python-scipy 
$ sudo apt install python-numpy python-pandas python-matplotlib python-scikit-image
$ sudo pip install scikit-image pyyaml protobuf future google
$ sudo pip install --upgrade beautifulsoup4
$ sudo pip install --upgrade html5lib

$ sudo pip install ck

$ ck pull repo:ck-caffe2

$ ck install package --tags=lib,caffe2,vcpu --env.CK_HOST_CPU_NUMBER_OF_PROCESSORS=2 --env.CAFFE2_CPU_FLAGS="-mfpu=neon -mfloat-abi=hard"

Example of Caffe and Caffe2 unified CUDA installation on Ubuntu via CK

If you have CUDA-compatible GPGPU with drivers, CUDA and cuDNN installed, you can install Caffe and Caffe2 for GPGPU via CK as follows (CK will automatically find your CUDA installation):

$ ck install package:lib-caffe-bvlc-master-cuda-universal --env.CAFFE_BUILD_PYTHON=ON
$ ck install package:lib-caffe-bvlc-master-cudnn-universal --env.CAFFE_BUILD_PYTHON=ON
$ ck install package:lib-caffe2-master-eigen-cuda-universal --env.CAFFE_BUILD_PYTHON=ON

You can find detailed instructions to install Caffe (CPU, CUDA, OpenCL versions) via CK on Ubuntu, Gentoo, Yocto, Raspberry Pi, Odroid, Windows and Android here. Caffe2 detailed instructions about customized CK builds are coming soon!

Example of Caffe and Caffe2 unified classification on Ubuntu via CK

$ ck run program:caffe --cmd_key=classify
$ ck run program:caffe2 --cmd_key=classify

Note, that Caffe2, besides some very useful improvements, also changed various support programs and API. However, our approach helped our collaborators hide these changes via CK API and thus protect higher-level experimental workflows!

You can find and install additional Caffe and Caffe2 models via CK:

$ ck search package:* --tags=caffemodel
$ ck search package:* --tags=caffemodel2

$ ck install package:caffemodel-bvlc-googlenet
$ ck install package:caffemodel-bvlc-alexnet

$ ck install package:caffemodel2-deepscale-squeezenet-1.1
$ ck install package:caffemodel2-resnet50

Collaborative and unified benchmarking of DNN

Additional motivation to use CK wrappers for DNN is the possibility to assemble various experimental workflows, crowdsource experiments and engage with the community to collaboratively solve complex problems (notes). For example, we added basic support to collaboratively evaluate various DNN engines via unified CK API:

$ ck crowdbench caffe --env.BATCH_SIZE=5
$ ck crowdbench caffe2 --env.BATCH_SIZE=5 --user=i_want_to_ack_my_contribution

Performance results are continuously aggregated in the public CK repository, however they can also be aggregated only on your local machine or in your workgroup - you just need to add flag "--local".

Unified, multi-dimensional and multi-objective autotuning

It is now possible to take advantage of our universal multi-objective CK autotuner to optimize Caffe. As a first simple example, we added batch size tuning via CK. You can invoke it as follows:

$ ck autotune caffe
$ ck autotune caffe2

All results will be recorded in the local CK repository and you will be given command lines to plot graphs or replay experiments such as:

$ ck plot graph:{experiment UID}
$ ck replay experiment:{experiment UID} --point={specific optimization point}

Collaborative and unified optimization of DNN

We are now working to extend above autotuner and crowdsource optimization of the whole SW/HW/model/data set stack (paper 1, paper 2).

We would like to thank the community for their interest and feedback about this collaborative AI optimization approach powered by CK at ARM TechCon'16 and the Embedded Vision Summit'17 - so please stay tuned ;) !


Unified DNN API via CK

We added similar support to install, use and evaluate TensorFlow via CK:

$ ck pull repo:ck-tensorflow

$ ck install lib-tensorflow-1.1.0-cpu
$ ck install lib-tensorflow-1.1.0-cuda

$ ck run program:tensorflow --cmd_key=classify

$ ck crowdbench tensorflow --env.BATCH_SIZE=5

$ ck autotune tensorflow

Online demo of a unified CK-AI API

  • Simple demo to classify images with continuous optimization of DNN engines underneath, sharing of mispredictions and creation of a community training set; and to predict compiler optimizations based on program features.

Realistic/representative training sets

We provided an option in all our AI crowd-tuning tools to let the community report and share mispredictions (images, correct label and wrong misprediction) to gradually and collaboratively build realistic data/training sets:

Next steps

We would like to improve Caffe2 installation via CK on Android similar to CK-Caffe.

Long term vision

CK-Caffe, CK-Caffe2, CK-Tensorflow are part of an ambitious long-term and community-driven project to enable collaborative and systematic optimization of realistic workloads across diverse hardware in terms of performance, energy usage, accuracy, reliability, hardware price and other costs (ARM TechCon'16 talk and demo, DATE'16, CPC'15).

We are working with the community to unify and crowdsource performance analysis and tuning of various DNN frameworks (or any realistic workload) using Collective Knowledge Technology:

We continue gradually exposing various design and optimization choices including full parameterization of existing models.

Open R&D challenges

We use crowd-benchmarking and crowd-tuning of such realistic workloads across diverse hardware for open academic and industrial R&D challenges - join this community effort!

Related Publications with long term vision

    title = {{Collective Knowledge}: towards {R\&D} sustainability},
    author = {Fursin, Grigori and Lokhmotov, Anton and Plowman, Ed},
    booktitle = {Proceedings of the Conference on Design, Automation and Test in Europe (DATE'16)},
    year = {2016},
    month = {March},
    url = {}

    url = {},
    title = {{Collective Mind, Part II: Towards Performance- and Cost-Aware Software Engineering as a Natural Science.}},
    author = {Fursin, Grigori and Memon, Abdul and Guillon, Christophe and Lokhmotov, Anton},
    booktitle = {{18th International Workshop on Compilers for Parallel Computing (CPC'15)}},
    publisher = {ArXiv},
    year = {2015},
    month = January,
    pdf = {}

Testimonials and awards


Get in touch with CK-AI developers here. Also feel free to engage with our community via this mailing list: