Skip to content
Public scenarios to crowdsource experiments (such as DNN crowd-benchmarking and crowd-tuning) using Collective Knowledge Framework across diverse mobile devices provided by volunteers. Results are continuously aggregated at the open repository of knowledge:
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
.cm
experiment.scenario.mobile
.ckr.json
CHANGES
README.md

README.md

CK-powered experiment crowdsourcing scenarios

compatibility DOI License

This repository contains public scenarios (software and data sets) in the Collective Knowledge Format to let the community participate in collaborative deep learning optimization including DNN engines (Caffe, TensorFlow) and various models across diverse mobile devices using universal experiment crowdsourcing Android app.

The results (performance, mispredictions, etc) are continuously aggregated in the open CK repository of knowledge.

Normally, you use this repository only for development. Whenever ready, this repository is synced at the cKnowledge.org/repo to make scenarios available for update in the Android app - just do not forget to select "Information" -> "Update Scenarios"!

Prerequisites

  • Collective Knowledge framework (@GitHub)
  • python (with imaging and matplotlib)
  • git client
  • Android NDK

Authors

License

  • BSD, 3-clause

Installation

$ sudo apt-get install python python-pip git
$ sudo pip install ck
 $ ck pull repo:ck-crowd-scenarios

Android application

Android application let volunteers participate in collaborative benchmarking and optimization of deep learning algorithms. It also sends benchmarking and optimization statistics as a JSON blob to the module 'experiment.bench.caffe.mobile' (function "process") from the ck-caffe repository to be aggregated in the live CK repository.

Updating existing scenarios

If you would like to update existing scenarios (generate new libcaffe.so and classification binary for Android via CK), you should copy them to new entries with a version extension.

For example, you can copy entry "bvlc-caffenet-android-recognize-image-v2" to "bvlc-caffenet-android-recognize-image-v3" via

 $ ck cp experiment.scenario.mobile:bvlc-caffenet-android-recognize-image-v2 ck-crowd-scenarios::bvlc-caffenet-android-recognize-image-v3

Then you should add key "outdated":"yes" to the meta.json of "bvlc-caffenet-android-recognize-image-v2" entry. In such case this entry will be in an "achive" state, will not be updated and will not be visible for Android app.

You should then update "program" key in a meta.json of the new entry with the new version, say 3.0.0. You also need to substitute v2 with v3 in all keys in meta.json.

After that you should build Caffe libs for all scenarios that will be updated.

For example, for CPU version of Caffe you should run:

$ ck install package:lib-caffe-bvlc-master-cpu-universal --target_os=android21-arm64
$ ck install package:lib-caffe-bvlc-master-cpu-universal --target_os=android21-arm-v7a

For Caffe OpenCL you should do the following:

ck install package:lib-caffe-bvlc-opencl-clblast-universal --target_os=android21-arm64 --env.DISABLE_DEVICE_HOST_UNIFIED_MEMORY=ON
ck install package:lib-caffe-bvlc-opencl-clblast-universal --target_os=android21-arm-v7a --env.DISABLE_DEVICE_HOST_UNIFIED_MEMORY=ON

For Caffe CPU you should invoke the following:

$ ck install package:lib-caffe-bvlc-master-cpu-universal --target_os=android21-arm64
$ ck install package:lib-caffe-bvlc-master-cpu-universal --target_os=android21-arm-v7a --env.OPTFLAGS="-O2 -march=armv7-a -mfloat-abi=softfp -mfpu=neon"

For TFLite CPU you should run the following:

$ ck install package:lib-tflite-1.7.0-src-static --target_os=android21-arm64
$ ck install package:lib-tflite-1.7.0-src-static --target_os=android21-arm-v7a

Note that while Caffe can currently run on Android 5+, you can build and run Caffe on older Android 4.2+ via:

$ ck install package:lib-caffe-bvlc-master-cpu-universal --target_os=android19-arm64
$ ck install package:lib-caffe-bvlc-master-cpu-universal --target_os=android19-arm-v7a --env.OPTFLAGS="-O2 -march=armv7-a -mfloat-abi=softfp -mfpu=neon"

Additional info:

  • building Caffe via CK workflow framework with various libraries for Android: notes.
  • building TensorFlow via CK workflow framework with various libraries for Android: notes.

Note that we suggest you to have a clean installation of all libs. You can do it by deleting CK env for all software via

$ ck rm -f env:*

and then removing files from '$USER/CK_TOOLS' directory.

Now you are ready to update new scenarios. You can do it as follows:

$ ck generate experiment.bench.dnn.mobile

Normally, all outdated libcaffe.so will be automatically deleted and updated ones will be copied to the new entries. You can also update only scenarios for a specific engine via

$ ck generate experiment.bench.dnn.mobile --prune_engine="Caffe CPU"
 and/or
$ ck generate experiment.bench.dnn.mobile --prune_engine="Caffe OpenCL"
 and/or
$ ck generate experiment.bench.dnn.mobile --prune_engine="TensorFlow CPU"
 and/or
$ ck generate experiment.bench.dnn.mobile --prune_engine="TFLite CPU"
 and/or
$ ck generate experiment.bench.dnn.mobile --prune_engine="ArmCL OpenCL"

Finally, you can now automatically update length of files, their MD5 and URLs for all scenarios as follows:

 $ ck process experiment.scenario.mobile

or for specific ones such as tflite as follows:

 $ ck process experiment.scenario.mobile:tflite*

Now, new scenarios should be ready to be used by this Android app if updated at the cKnowledge.org/repo server - contact authors for more details.

Public results

Public benchmarking and optimization results collected from Android devices are continuously aggregated in the live CK repository.

Related publications

See all CK publications with BibTex here.

Further discussions

You can’t perform that action at this time.