基于FPGA的卷积神经网络优化压缩技术研究

Optimal Compression Technology of Convolutional Neural Network Based on FPGA

针对现有卷积神经网络(CNN)模型计算效率低、内存带宽浪费大等问题,提出了一种基于现场可编程门阵列(FPGA)优化压缩策略。对预先训练好的CNN模型进行分层剪枝,采用基于新型的遗传算法进行信道剪枝,同时设计了两步逼近适应度函数,进一步提高了遗传过程的效率。此外,通过对剪枝CNN模型进行数据量化,使得卷积层和全连接层的权值根据各自的数据结构以完全不同的方式存储,从而减少了存储开销。实验结果表明,在输入4 000个训练图像进行压缩过程中,该方法所耗压缩时间仅为15.9 s。

Aiming at the low computational efficiency of the existing convolutional neural network(CNN) model and large waste of memory bandwidth, an optimized compression strategy based on field programmable gate array(FPGA) is proposed. The pre-trained CNN model is hierarchically pruned, and a new genetic algorithm is used for channel pruning. At the same time, a two-step approximation fitness function is designed to further improve the efficiency of the genetic process. In addition, by quantizing the data of the pruned CNN model, the weights of the convolutional layer and the fully connected layer are stored in completely different ways according to their respective data structures, thereby reducing storage overhead. The experimental results show that when 4000 training images are input for compression, the compression time consumed by this method is only 15.9 seconds.