| In recent years,Deep Learning techniques which represented by Convolutional Neural Networks have been applied in more and more scenarios.However,the volume and computational complexity of Convolutional Neural Networks are usually quite large,and they have high requirements for memory,bandwidth and computing resources.Although the performance of various hardware devices is also improving,it still cannot keep up with the development speed of large networks,which severely limits the further promotion of Convolutional Neural Networks.Therefore,it is an urgent problem to be solved that how to compress and accelerate the Convolutional Neural Networks while maintaining their performance.Many researchers have studied the parameters quantization,pruning,low-rank decomposition of weights and hardware implementation of Convolutional Neural Networks.Some of these works have also combined multiple techniques.This paper also proposes an effective solution that combine the quantization technique and the pruning technique in a cascaded manner,and use knowlodge distillation technique to make up for the performance loss caused by the compression,the main work is as follows:In order to solve the problem that the Convolutional Neural Networks are difficult to be quantized to extremely low precision,the mixed-precision quantization scheme is used,which allows each layer to use different quantization bits;and a gradual quantization strategy is proposed,which can select the optimal accuracy for each layer in a short time.Experimental results show that this algorithm can modify the Convolutional Neural Networks more flexibly,and effectively reduce the volumes of their parameters.Aiming at the accuracy and promotion of pruning,a channel importance scoring method combining global information and local information is proposed,a pruning strategy automatically determines the pruning ratios of each layer based on the scoring results is designed,and the pruning methods for different types of Convolutional Neural Networks are also studied.Experimental results show that this algorithm has effect on a variety of Convolutional Neural Networks,and can simultaneously reduce the volume and the amount of calculation.For the feasibility of the above solution,this paper deployed the Convolutional Neural Networks that before and after compression on the FPGA platform,comprares the amount of hardware resources and calculation delay required by these two situations.Experimental results show that this solution can also achieve effective compression and acceleration of Convolutional Neural Networks on FPGA,but there are some differences between the actual and theoretical effect. |
