一种基于通道域自注意力的特征图剪枝方法

A Feature Map Pruning Method Based on Channel Domain Self-attention Mechanism

为同时保证网络剪枝方法的准确率和稳定性,提出一种基于通道域自注意力的特征图剪枝方法。该方法采用主成分分析(principle component analysis,PCA)算法降低噪声干扰,引入通道域注意力为特征图自动分配不同权重,移除低权重通道的滤波器,并对网络进行重训练和精调,以减少网络精度损失。在公开数据集上对VGG-16网络模型展开的实验表明,当剪枝率为60%时,达到视觉几何组模型Top-5的准确率为89.23%;当剪枝率逐渐增加到80%时,仍保持73%准确率。相较于同类方法,本文提出的方法更能保证剪枝时模型的准确率和稳定性。

A channel domain attention-based feature graph pruning method is proposed to guarantee the accuracy and stability of network pruning.The method uses principal component analysis(PCA)algorithm to reduce noise interference,introduces filters that automatically assign different weights to feature maps based on channel domain attention to remove low-weight channels,and retrains and fine-tunes the network to reduce network accuracy loss.The proposed method is experimentally validated on public dataset for VGG-16 network model,and the experimental results show that the model achieves 89.23%accuracy of VGG model Top-5 at a pruning rate of 60%,and 73%accuracy at a gradually increased pruning rate of 80%,which guarantees better accuracy and stability when pruning over similar methods.