Research On MobileNet Acceleration Methods Based On HLS

With the hot development of deep learning,convolutional neural networks are playing an increasingly important role in unmanned driving,biometrics and other fields.At present,the development trend of convolutional neural networks is to obtain more accurate networks through deeper and more complex networks.However,due to limited hardware resources,an embedded system needs a convolutional neural network model with lower computation and parameters.However,MobileNet is a kind of lightweight convolutional neural network.Compared with traditional convolutional neural network,MobileNet has lower computation and parameters,so it is suitable for deployment in embedded systems.At present,most MobileNet hardware acceleration is carried out on GPU.Although GPU can meet real-time processing,its high power consumption and high cost cannot meet the requirements of some low-power and low-cost embedded systems.However,FPGA is a hardware programmable accelerator with strong parallel capability and low power consumption.Therefore,it has certain advantages to use FPGA for hardware acceleration of MobileNet convolutional neural network.Traditional FPGA development is to use hardware description language directly to complete the optimization design,which has the disadvantages of high design difficulty and long development cycle.High Level Synthesis?HLS?can complete the conversion from C/C++ and other high-level languages to hardware languages,which can greatly shorten the development cycle of FPGA.However,for complex algorithms,the performance of HLS automatically converted circuit is usually not ideal,so it must be optimized.Therefore,based on the HLS tool,this paper studies the structure of MobileNet network of lightweight convolutional neural network and the acceleration optimization method of FPGA.The main work done is:?1?The network structure of mobilenet_v1₁.0₂24 network model is adjusted to reduce part of the computation and to facilitate the subsequent system architecture design.?2?The overall architecture of the adjusted MobileNet network model was designed,and the top-level function design was carried out for the most time-consuming part of the standard convolutional layer in MobileNet convolution from the perspectives of reducing external I/O operations and satisfying memory resources and memory.?3?Based on particle swarm optimization?pso?,a general HLS automation method is designed to find the optimal scheme of the top-level function.Finally,the optimization scheme of the top layer function was successfully found through experiments,which effectively reduced the latency index and achieved the effect of system acceleration.