Skip to content

Latest commit

 

History

History
113 lines (92 loc) · 3.87 KB

README.md

File metadata and controls

113 lines (92 loc) · 3.87 KB
Raw

ElasticDNN

This repo includes the implementation of IONN (Incremental Offloading of Neural Network), which was proposed by the paper titled "IONN: Incremental Offloading of Neural Network Computations From Mobile Devices to Edge Servers" published in SoCC (ACM Symposium on Cloud Computing) 2018.

The repo merely consists of two submodules: IONN-client and IONN-server. Each submodule runs on the mobile device (ARM) and the server (x86), respectively.

Installation

Client-side

  1. Clone our project (IONN-client)
$ git clone https://github.com/jinevening/IONN-client.git
  1. Build our project according to the caffe installation instructions.

Server-side

  1. Clone our project (IONN-server)
$ git clone https://github.com/jinevening/IONN-server.git

The header lines related to hdf5.h may have to be modified (from hdf5/serial/hdf5.h to hdf5.h) or a symbolic link from "/usr/include/hdf5/serial/" to "/usr/include/" created.

  1. Build our project according to the caffe installation instructions.

Execution

First, we have to run IONN-server on the server by running the following command at the repository directory

./build/examples/partitioning_server/classification.bin

Then, logs like following should be printed on the shell.

GPU mode
Upload Server Started on Port 7675
Execution Server Started on Port 7676

While IONN-server is running on the server, IONN-client can offload DNN execution to the server by running the following command at the client device

./build/examples/partitioning/classification.bin \
  *path to the prototxt file* \
  *path to the caffemodel file* \
  *path to the mean file* \
  *path to the label file* \
  *profile of the execution time in the client* \
  *prediction function for the execution time in the server* \
  *network speed in Mbps* \
  *path to the image file* \
  *value of K in the IONN paper* \
  incremental \
  *optimization target* \
  *ip_address*

Example

./build/examples/partitioning/classification.bin \
  models/bvlc_alexnet/deploy.prototxt \
  models/bvlc_alexnet/bvlc_alexnet.caffemodel \
  data/ilsvrc12/imagenet_mean.binaryproto \
  data/ilsvrc12/synset_words.txt \
  models/bvlc_alexnet/prediction_model.txt \
  server_prediction_model.txt \
  80 \
  cat.jpg \
  0.8 \
  incremental \
  time \
  123.45.678.910

Acknowledgements

We implemented IONN based on caffe, an open source project to train and test DNN. We extended caffe to support the collaborative execution (inference) between a client and a server. To transmit the activation data between the client and the server, we used boost.asio library.

License and Citation

ElasticDNN is released under the BSD 2-Clause license. Please cite following papers in your publications if it helps your research:

@inproceedings{jeong2018ionn,
  title={Ionn: Incremental offloading of neural network computations from mobile devices to edge servers},
  author={Jeong, Hyuk-Jin and Lee, Hyeon-Jae and Shin, Chang Hyun and Moon, Soo-Mook},
  booktitle={Proceedings of the ACM Symposium on Cloud Computing},
  pages={401--411},
  year={2018},
  organization={ACM}
}

@inproceedings{shin2019enhanced,
  title={Enhanced Partitioning of DNN Layers for Uploading from Mobile Devices to Edge Servers},
  author={Shin, Kwang Yong and Jeong, Hyuk-Jin and Moon, Soo-Mook},
  booktitle={The 3rd International Workshop on Deep Learning for Mobile Systems and Applications},
  pages={35--40},
  year={2019},
  organization={ACM}
}

Project Team Members

Hyuk Jin Jeong ([email protected]) and Kwang Yong Shin ([email protected])