This converter was created to convert my own (in Keras with tensorflow as backend trained) Model to Openvx Graph. The reason being that a direct comparison of the performance between openvx on AMD Hardware and TF_inference on Nvidia GPUs.
This implementation is still not fully implemented. However I tried to create a expandable structure to make it easy to expand the source code.
Since a caffe converter already exists, I tried to keep the structure similar for easy integration into amdovx-modules.
To get run this project you have to download and build tensorflow. This should be resolved in the future by using NNEF
In the main function you have to define the model path "*.pb" and the NCHW input size.
After successful run this project extract weights and biases of the input Model simultaneously generate an GDF (Graph Descriptor File) and c++ ( with make file).
Please follow the instructions here to execute the generated OpenVX Graph.
MIT License (c) reger-men