ZyNet: Automating Deep Neural Network Implementation on Low-Cost Reconfigurable Edge Computing Platforms

Prevalence of internet of things (IoT) enabled applications provide a new opportunity to low-cost FPGA devices to act as edge computing neural network nodes. Although FPGA vendors provide neural network development environments, they often target high-end devices. At the same time these development platforms are not as user friendly as their software counterparts. In this work we introduce ZyNet, a Python package, which enables faster implementation of deep neural networks (DNNs) targeting low-cost hybrid FPGA platforms such as the Xilinx Zynq. Based on hardware-software co-design approach, this platform supports pre-trained or on-board trained networks with development environment very similar to the popular TensorFlow. Implementation results show that the DNNs generated by the platform achieve accuracy very close to software implementations at the same time gives throughput by an order of magnitude compared to other edge computing devices at lower energy footprint. The platform is integrated with Xilinx development tools and is distributed as open source.