epfl-dcsl/ptf-system

The main PTF repository for the TF core.

PTF System

Repository contents

• This contains the TensorFlow-based components for PTF and Persona (the components necessary for this application to work).
• tensorflow/contrib/gate contains the code for the PTF components.
• tensorflow/contrib/persona contains the Persona-components.
• The rest of the code is standard TensorFlow, with some additions for building / compiling we added for our own convenience.

Arxiv submission of PTF work

• This is a paper describing the architectural details of this repository’s work and contributions.
• To get a full picture of this work, check this paper out in addition to the application below that uses this repository.

Building PTF Persona System

• The easiest way to do this is in Docker container.
• In order to replicate the steps outside of the Docker container, you may do the steps in the Dockerfile manually.

./build_container.sh

• This builds the container with a specific name that is needed by PTF Persona (the Python application).

Using Docker for only building

• It is possible to just use only the final wheel file built in the Docker container and merely copy that out.
• To do so, it is recommended that you adjust the base image (Ubuntu Bionic by default) to whatever image of Ubuntu you want to use for binary compatibility. You should do this for both the build and run images.
• Modify the Dockerfile to remove the second non-build image (the executor one).
• Make sure your base system has all the packages listed in RUNDEPS in the executor image environment variable.
• Also make sure you do the pip3 install step too so that you have all of the dependencies.
• Then run the following command to get the wheel file out of the build image.

docker run --rm -it -v "$(pwd)":/host ptf-system "cp /ptf/pip_pkg/*.whl /host"

You can adjust the mount to copy the wheel file to another location.

Accompanying PTF Persona Application

Link to Python-based application that uses this repository