基于相似度感知的深度卷积神经网络剪枝方法

Similarity-aware Pruning Method for Deep Convolutional Neural Networks

随着卷积神经网络规模的不断扩大,由于其庞大的计算量和参数量,终端智能设备的部署及发展面临着巨大的挑战,因此如何保持模型精度的同时尽可能地压缩和加速模型至关重要.目前已有工作提出的压缩方法仍然存在压缩算法实现、压缩效果、压缩效率等方面的缺陷.为此,本文提出了一种基于通道相似性的卷积神经网络剪枝方法.具体而言,首先探究了卷积神经网络特征通道间的相似冗余,引入了一种高效的相似性指标来量化特征通道之间的相似性;其次,通过相似性排序算法移除整个网络中冗余的通道从而实现剪枝;再次,加载保留的通道参数通过微调减少由于剪枝操作造成对模型分类性能的影响.为了提高压缩效率,本文采用一次性剪枝策略,满足时间复杂度更低的要求.最后,在CIFAR-10、CIFAR-100数据集上对VGG-16、ResNet-56、ResNet-110、GoogLeNet模型的实验结果表明,与现有方法相比本文所提方法可以更高效地压缩模型且模型依然保持良好精度.

As the scale of convolutional neural networks continues to expand,the deployment and development of terminal intelligent devices face significant challenges due to their substantial computational and parameter requirements.Therefore,it is crucial to compress and accelerate models while maintaining their accuracy as much as possible.Existing compression methods still have deficiencies in terms of compression algorithm implementation,compression effectiveness,and compression efficiency.In this paper,we propose a convolutional neural network pruning method based on channel similarity.Specifically,we first explore the similarity redundancy between feature channels in convolutional neural networks and introduce an efficient similarity metric to quantify the similarity between feature channels.Secondly,we achieve pruning by removing redundant channels throughout the network using a similarity ranking algorithm.Thirdly,we fine-tune the loaded parameters of the remaining channels to mitigate the impact of pruning on the model's classification performance.To improve compression efficiency,we adopt a one-shot pruning strategy to meet the requirement of lower time complexity.Finally,experimental results on the CIFAR-10 and CIFAR-100 datasets with VGG-16,ResNet-56,ResNet-110 and GoogLeNet models demonstrate that the proposed method in this paper can compress models more efficiently while maintaining good accuracy compared to existing methods.