Cityscapes Semantic Segmentation on FPGA with FINN

This repository contains source files to build and deploy on Alveo U250 board a quantized ResNet18-Unet semantic segmentation model trained on Cityscapes dataset using the FINN framework from https://github.com/Xilinx/finn.

Requirements:

• Vitis 2022.2
• XRT 2.13
• Docker
• Python >= 3.8
• Pynq >= 3.0.1

A pre-synthesized XCLBIN file is available in /deploy/driver/bitfile/ and can be used along with the pynq-based driver to deploy and evaluate the model.

To re-generate binaries, do:

#clone FINN repository at correct commit
get-finn.sh
cd finn/
#launch the build
./run-docker.sh build_custom ../scripts

Newly generated bitfile should be in /outputs/bitfile/.