A library-based tool to translate high level DNN models into hierarchical VHDL descriptions

Abstract

This work presents a tool to convert high level models of deep neural networks into register transfer level designs. In order to make it useful for different target technologies, the output designs are based on hierarchical VHDL descriptions, which are accepted as input files for a wide variety of FPGA, SoC and ASIC digital synthesis tools. The presented tool is aimed to speed up the design and synthesis cycle of such systems and provides the designer with certain capability to balance network latency and hardware resources. It also provides a clock domain crossing to interface the input layer of the synthesized neural networks with sensors running at different clock frequencies. The tool is tested with a neural network which combines convolutional and fully connected layers designed to perform traffic sign recognition tasks and synthesized under different hardware resource usage specifications on a Zynq Ultrascale+ MPSoC development board.