[VTA][FPGA] Running a Tensorflow graph on VTA

Questions
#1

First I followed the VTA Installation Guide and tested with the 2Dconvolution provided in it. After that I followed the tutorial named “Deploy Pretrained Vision Model from MxNet on VTA” also.

Now I am finding it very difficult to run a basic tensorflow model created by me on vta in the same manner. I couldn’t find any directly related tutorial even. If anyone has done running a pre-trained tensorflow model for inference on a xilinx pynq fpga using vta, please help me on how to do it. If you can send me a python source code of any tensorflow model running on vta, that would be great.

#2

@nalithlakshan at the moment we don’t offer a direct path from Tensorflow model to VTA; in fact even the MxNet importer at this point is quite brittle.

I believe there is more work to be done on those ends; specifically how an input graph needs to be massaged to support VTA and other accelerator backends to do graph partitioning, rule based fusion, and quantization etc. This is something that will benefit VTA among other accelerators.

In terms of supporting/exercising other forms of model importing (e.g,. Tensorflow), I welcome contributors who are interested to work on this.

1 Like