TF Keras to Xilinx Zynq (Ultra96v2) design flow

This repo contains all files for compilation and deployment. (Note: check out this project for more detailed explanation.)

List of files:

File name	description	comment
hw_files	hardware description files for DPU on Ultra96 v2	-
docker_run.sh	starts the docker for VITIS AI with required parameters	! `conda activate vitis-ai-tensorflow`
h5topb.py	loads weights (hdf5 format) and converts to TensorFlow .pb format	run inside VITIS AI docker for TF version 1.15
freeze.sh	script to freeze the converted model	run inside docker
quantize.sh	script to quantize the frozen model	run inside docker
unet.py	Unet model from dhkim0225	-
compile.sh	script for compilation	inside docker

Step 1: Convert Model to .pb format

h5topb.py -> provide file path to hdf5 encoded weights

! run inside VITIS AI docker or with tensorflow v1.15 and keras==2.2.4

Step 2: Freeze the converted Model

freeze.sh -> changed input and output node names accordingly

Step 3: Quantization

quantize.sh -> provide a subset of the image dataset previously used for training or validation and adjust path in graph_input_fn.py accordingly. This python script is required for calibration of the quantized model.

! Make sure that output and input labels are correct for quantize.shand graph_input_fn.py

Step 4: Compilation

compile.sh -> DPU hardware files can be retrieved through DPU-PYNQ build.

Troubleshooting

Freezing fails:

IndexError: list index out of range

Make sure the docker image is modified such that conda has correct keras==2.2.4 package installed:

 ./docker_run.sh xilinx/vitis-ai-cpu:latest
 sudo su
 conda activate vitis-ai-tensorflow
 conda install keras==2.2.4  # compatible with TF 1.15
 conda deactivate
 exit
 conda activate vitis-ai-tensorflow

Stage your modifications for the next docker restart:

 sudo docker commit -m "<message>" <image_id> xilinx/vitis-ai-cpu:latest

swappad/TFModeltoEdge