基于层融合特征系数的动态网络结构化剪枝

剪枝是一种减少模型复杂度的有效方式,现有的剪枝方法仅考虑卷积层对特征图的影响,无法准确判断冗余滤波器,文中提出基于层融合特征系数的动态网络结构化剪枝方法.同时考虑卷积层和批规范化(BN)层对输出特征图的影响,利用多动态参数确定滤波器的重要性,动态寻找冗余滤波器,获取最优网络结构.在CIFAR-10、CIFAR-100数据集上的实验表明,无论是残差网络还是轻量型网络,文中方法在采用较大剪枝率时仍能保持较高精度.

Dynamic Network Structured Pruning via Feature Coefficients of Layer Fusion

Pruning is an effective way to reduce the complexity of the model. In the existing pruning methods, only the influence of the convolutional layer on the feature map is taken into account,and therefore the redundant filter cannot be determined accurately. In this paper, a dynamic network structured pruning method based on layer fusion feature coefficients is proposed. Considering the influence of convolutional layer and Batch Norm layer on the feature map, the importance of the filter is determined by multiple dynamic parameters, and the redundant filter is dynamically searched to obtain the optimal network structure. Experiments on the standard datasets of CIFAR-10 and CIFAR-100 shows that both the residual network and the lightweight network maintain high accuracy while using large pruning rates.