Deployment of neural networks through PYNQ

Abstract

The PYNQ platform provides a python interface for accessing FPGA resources, which gives us the opportunity to efficiently deploy neural network models on FPGA to achieve high-performance and real-time image classification and target detection tasks. This hardware-accelerated approach can provide faster inference speed and lower power consumption than software-accelerated approach. In this research and development project, our main research objective is to deploy neural networks on PYNQ. I have used the PYNQ-Z1 development board for experiments. Four type of networks have been deployed, namely: a YOLO network, a BNN network, a ResNet network and a MobileNetv2 network. After deployment, I have compared their accuracy and measured their execution time on hardware, achieving promising results for a resource-constrained device as the Z1 board.