README.md

December 8, 2018 · View on GitHub

Neurosurgeon Implementation (Single Image Adaptation)

We aim to use UDP socket programming as well as the Tensorflow API to measure the composite end-to-end delay of offloading Image Classification.

Motivation

This testbench is being used as part of ongoing research into Machine Learning Task Partitioning by Noah Johnson, Xiangqi Kong, and Dr. Bin Li of the University of Rhode Island Smart Networking and Computing Lab.

Tech/framework used

Built with

To Do List

  • Adapt to use UDP instead of TCP (TCP is too slow)

Environment Installation / Setup - TensorFlow CPU

Windows

  • Install Anaconda and add it to $PATH if you dont already have it.
  • Install Tensorflow CPU and all dependencies: 
    conda create --name tf python=3.5
activate tf
conda install scipy
pip install tensorflow
  • If everything is installed correctly then this test code 
    >> import tensorflow as tf
>> hello = tf.constant('Hello, TensorFlow!')
>> sess = tf.Session()
>> print(sess.run(hello))
    Should return 
    Hello, TensorFlow!

Linux

  • Install Anaconda and add it to $PATH if you dont already have it.
  • Install Tensorflow CPU and all dependencies: 
    conda create --name tf python=3.5
source activate tf
conda install scipy
pip install tensorflow
  • If everything is installed correctly then this test code 
    >> import tensorflow as tf
>> hello = tf.constant('Hello, TensorFlow!')
>> sess = tf.Session()
>> print(sess.run(hello))
    Should return 
    Hello, TensorFlow!

Environment Installation / Setup - TensorFlow + NVIDIA GPU

Disclaimer

Setting Up TensorFlow GPU will take a long time, and there are no guarantees. The code will run much better on the GPU Version, however installation of TF-GPU on Windows is tedious at best and maddening at worst. Use this guide at your own risk.

Windows

  • Ensure you have Microsoft C/C++ Build Tools 2015 installed. If you have Visual Studio 2017 installed, these tools are not installed by default. Make sure these are in your $PATH as well.
  • Ensure that your NVIDIA Base Drivers are up to date. At time of writing, the current driver set is Release 390. Reboot after this step.
  • Install the NVIDIA Cuda Toolkit. This must be version 9.0, version 9.1 is not yet supported. Add the bin folder and enclosing CUDA folder to $PATH.
  • Install CuDNN version 7.0 and add that to $PATH.
  • Reboot. If you got all of this installed correctly, the tedious part is over.
  • Install Anaconda and add it to $PATH if you dont already have it.
  • Install Tensorflow GPU and all dependencies: 
    conda create --name tf-gpu python=3.5
activate tf-gpu
conda install scipy
pip install tensorflow-gpu
  • If everything is installed correctly then this test code 
    >> import tensorflow as tf
>> hello = tf.constant('Hello, TensorFlow!')
>> sess = tf.Session()
>> print(sess.run(hello))
    Should return 
    Hello, TensorFlow!
  • Go take a walk because you're done and you deserve it.

Linux

  • See this guide, it does a great job of explaining the protocol for installing.

Project Dependencies (Need to update this part still)

A full list of dependencies can be found in env.yml. Anaconda maintains that an equivalent enviroment to mine can be built using:

conda env create -f env.yml -n tf-gpu

I have been unable to independenly verifiy this functionality. If anyone finds success with this method, let me know and I will update this to reflect any new information.

Credits

Inspired by TensorFlow Model Zoo

License

MIT License. Provided AS IS.

MIT © Noah Johnson