基于PCA的深度可分离卷积滤波器剪枝算法研究

Research on Pruning Algorithm of Depth Separable Convolutional Filter Based on PCA

为了解决DCNN计算量大的问题,文章提出了一种基于PCA的深度可分离卷积的滤波器剪枝方法。首先,采用深度可分离卷积代替ResNet中的普通卷积。先使用逐通道卷积在空间维度上进行分离,以增加网络宽度并扩大特征提取范围,再利用逐点卷积降低普通卷积操作的计算复杂度。其次,采用PCA降维区分堆叠的相似过滤器,不仅缓解维度灾难而且压缩数据的同时最小化信息损失。实验结果表明,该方法能显著提升深度卷积神经网络模型的计算速度和准确度,并进一步压缩模型大小。在cifar10上,减少了ResNet56上约41%的参数量,并且极大的缩短了模型运行的时间。

In order to solve the problem of large amount of calculation in DCNN,this paper proposes a PCA-based deep separable convolution filter pruning method.Firstly,use depthwise separable convolution instead of ordinary convolution in ResNet.Specifically,use depthwise convolution to separate the ordinary convolution in spatial dimensions to increase the network width and expand the range of feature extraction.Then pointwise convolution is used to reduce the computational complexity of ordinary convolution operation.Secondly,PCA dimensionality reduction is used to distinguish stacked similar filters,which not only alleviates the disaster of dimensionality,but also minimizes information loss while compressing data.Experimental results show that this method can significantly improve the calculation speed and accuracy of the deep convolutional neural network model,and further compress the model size.On cifar10,the amount of parameters on ResNet56 is reduced by about 41%,and the running time of the model is greatly shortened.