LPYOLO: Low Precision YOLO for Face Detection on FPGA

Contents

• Introduction
• Requirements
• Installation
• Usage
• Citation

Introduction

This repo contains evaluation and deploy scripts for official LPYOLO: Low Precision YOLO for Face Detection on FPGA paper. Models are trained by Brevitas which is a PyTorch research library for quantization-aware training (QAT) and exported to ONNX. FINN which is an experimental framework from Xilinx Research Labs to explore deep neural network inference on FPGAs is used for deploying models on a PYNQ-Z2 board.

Requirements

• Finn == 0.7
• Pytorch >= 1.8.1

Installation

First, download Pytorch armv7 and Torchvision armv7.

Then, run

pip install torch-1.8.1-cp38-cp38-linux_armv7l.whl
pip install torchvision-0.9.1-cp38-cp38-linux_armv7l.whl

on PYNQ-Z2 board.

Also, install bitstream

pip install bitstream

Usage

First, connect a PYNQ-Z2 board and open a terminal.Then, clone the project:

git clone git@github.com:sefaburakokcu/finn-quantized-yolo.git

Then, download one of the deploy.zip file from the table below and extract. Copy finn-accel.bit, finn-accel.hwh and scale.npy into src/deploy/bitfile/.

Model Definition

Definition of LPYOLO architecture is given below.

Pretrained Models

Models	ONNX	Deploy
2W4A	2w4a.onnx	2w4a_deploy.zip
3W5A	3w5a.onnx	3w5a_deploy.zip
4W2A	4w2a.onnx	4w2a_deploy.zip
4W4A	4w4a.onnx	4w4a_deploy.zip
6W4A	6w4a.onnx	6w4a_deploy.zip
8W3A	8w3a.onnx	8w3a_deploy.zip

Note: All models have 8 bits input and 8 bits output precisions. xWyA indicates x bits for weights and y bits precision for activations.

Training

Please use the following quantized-yolov5 repository for training quantized models.

Deployment

Please follow the steps in End-to-End Deployment of LPYOLO: Low Precision YOLO for Face Detection on FPGA.

Inference

Open inference.ipynb for inference on images, video_demo.ipynb for inference on a video under finn-quantized-yolo/src/deploy/ in jupyter-notebook.

Citation

@inproceedings{Gunay_2022, series={EECSS’22},
   title={LPYOLO: Low Precision YOLO for Face Detection on FPGA},
   ISSN={2369-811X},
   url={http://dx.doi.org/10.11159/mvml22.108},
   DOI={10.11159/mvml22.108},
   booktitle={Proceedings of the 8th World Congress on Electrical Engineering and Computer Systems and Science},
   publisher={Avestia Publishing},
   author={Gunay, Bestami and Okcu, Sefa Burak and Bilge, Hasan Sakir},
   year={2022},
   month=jul, collection={EECSS’22} }