Abstract: Most of the existing channel pruning methods utilize a single evaluation rule to identify redundant channels, making it challenging to accurately assess their importance. Some pruning methods attempt to enhance accuracy by iteratively or hierarchically filtering channels multiple times, but this approach increases the time cost of pruning. To address these issues, this paper proposes an enhanced structured pruning method based on channel classification contribution and feature scaling coefficient. This method achieves a sparse convolutional neural network model with sparse parameters by combining L1 and L2 regularization of the original model. Furthermore, the network model is structurally pruned by combining the channel contributions to classification task and feature scaling coefficients in order to comprehensively and accurately screen redundant parameters within the network. On the CIFAR-10 dataset, the proposed method achieves a precision of 93.52% for the fine-tuned VGG-16 network model compressed with a 76.6% reduction in FLOPs, which is 1.65?percentage points higher than that achieved by compensation-aware pruning (CaP) method with a 65.6% reduction in FLOPs. The experimental results demonstrate that this pruning method can effectively maintain or even improve the accuracy of neural network models while significantly reducing their size.

Key words: model compression, structured pruning, channel pruning, convolutional neural network

摘要： 现有的通道剪枝方法大多采用单一的评判规则来筛选冗余通道，难以准确评估通道的重要性。一些剪枝方法尝试通过迭代或分层修剪并多次筛选通道来提升准确率，但却增加了剪枝的时间成本。针对以上问题，提出了一种改进的基于通道分类贡献和特征缩放系数的结构化剪枝方法。该方法通过对卷积神经网络模型进行L1、L2正则化相结合的稀疏正则化，得到参数较稀疏的深度卷积神经网络模型，并结合通道对分类任务的贡献度以及特征缩放系数两种因素对网络模型进行结构化剪枝，以更全面、准确的方式筛选网络中的冗余参数。在CIFAR-10数据集上，使用所提方法压缩的VGG-16网络模型在FLOPs减少76.6%的情况下，微调后的模型精度达到93.52%，比FLOPs减少65.6%的补偿感知剪枝（compensation-aware pruning，CaP）方法高出1.65个百分点。实验结果表明，该剪枝方法在大幅度压缩神经网络模型的同时，能够更有效地保持甚至提升模型的精度。

关键词: 模型压缩, 结构化剪枝, 通道剪枝, 卷积神经网络