TFMin - Minimal TensorFlow

A library to export light weight c++ implementations of TensorFlow graphs, for use on micro-processor and embedded systems.

Overview

The TFMin library allows you to convert a TensorFlow graph within a python script into a c++ implementation with only standard library dependencies. This allows the produced c++ code to be compiled on small computer systems and embbedded systems. Unlike the standard c++ implementation of TensorFlow which is already available the binaries produced by TFMin do not have dependencies on large shared object libraries. These dependencies can make implementing this code on embedded systems, difficult or impossible.

There are two parts to this package, a python library that is used to analyse and export the flow graph to c++ code and a header only c++ library containing the operations that is needed to compile the generated code. An example in provided where an MNIST classifier is trained using TensorFlow python then exported to c++ and built into a native binary.

This open source software has been developed during my PhD research at the Surrey Space Centre, at Surrey University and made possible by support from Airbus.

Installation Instructions

Clone this repository and run the install.bash script. This simply adds the required locations to your PYTHONPATH and CPLUS_INCLUDE_PATH environment variables, the script will need write access to your .bashrc file.

Licence

This software is (c) 2019 Pete Blacker, Surrey Space Centre & Airbus Defence and Space Ltd. It is licenced under the GPL v3 license for more details see the LICENCE file.

Airbus Commercial TFMin - LEON3 Support

A commercial extenstion to the TFMin tool is available from Airbus, which has full support for the LEON family of radiation hardened processors. This includes tensor operations optimised for these processors and V&V assurance to industry standards. For further information about features and licencing please contact: ilke.karsli@airbus.com, arthur.scharf@airbus.com or carlos.hervasgarcia@airbus.com

Publications

If you use this software for research purposes please cite the following publication in any published work.

• Blacker, P., Bridges, C.P., Hadfield, S., 2019, July. Rapid Prototyping of Deep Learning Models on Radiation Hardened CPUs. In Thirteenth NASA/ESA Conference on Adaptive Hardware and Systems (AHS 2019). IEEE.

Developer Guides

The attached wiki contains an installation guide and a set of tutorials to introduce new developers to TFMin and walks through how to export and build C++ implementations of existing TensorFlow models.