Design And FPGA Verification Of CNN Accelerator Based On Weight Combination

Deep convolutional neural networks are widely used in image recognition,target tracking and other fields.The deep convolutional neural network is deployed on the equipment with strict requirements of power consumption and real-time performance,so it is necessary to solve the problem that the network is too intensive in computation and storage.Therefore,the research of the convolutional neural network accelerator based on algorithm and hardware co-design is of great significance.The basic structure of the convolutional neural network and the common optimization methods of the convolutional neural network accelerator are summarized in this thesis.In order to solve the problem of parameters and calculations,the network parameters are quantified by the incremental quantization algorithm in this thesis.The quantization accuracy is 5bit.The low-bit width index data is used to replace the weight parameters,which eases the burden of data transmission and storage.The quantization steps include weight grouping,quantization and retraining.For the repeatability and sparseness of the quantized parameters,the number of redundant computations is reduced due to the convolution calculation method based on weight combination.The automated systolic array structure and parametric design module were adopted.It can accelerate the neural network of different size well.Different data reuse patterns are used to optimize the accelerator data flow mapping and reduce the energy consumption of accelerator access memory.After quantization,Alex Net and VGG-16 were tested based on the ILSVRC2012 dataset,the Top-1 error rate of Alex Net only increased by 0.37%,the Top-1 error rate of VGG-16 only increased by 0.51%,and the parameter scale decreased by 62.5%.In this thesis,a convolutional neural network accelerator based on weight combination is verified on the Xilinx Virtex-7 FPGA development board.The results show that under the working frequency of150 MHz,when the test network is Alex Net,the average throughput of the accelerator is 197.41 GOPS,and when the test network is VGG-16,the average throughput of the accelerator is 214.71 GOPS and the power is 7.1W.The convolutional neural network accelerator based on weight combination can configure parameters to accelerate different networks,and have good real-time performance and low power consumption.