Home

TFMin : The TensorFlow to Minimal C++ Converter

Introduction

The TFMin library can be used to convert a TensorFlow graph within a python script into c++ 11 code with dependencies on only the standard libraries and the Eigen linear algebra library. This allows the produced code to be compiled on small computers and embedded systems. Unlike the standard c++ implementation of TensorFlow 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.

This project was developed by Pete Blacker during his PhD in the On-board Data Handling Group at Surrey Space Centre, while researching deep learning approaches for common challenges on-board Mars rovers. This work is sponsored and supported by Airbus.

If you use this library in further research please cite the following publication:

• 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.