Linearly Replaceable Filters for Deep Network Channel Pruning

Convolutional neural networks (CNNs) have achieved remarkable results; however, despite the development of deep learning, practical user applications are fairly limited because heavy can be used solely with latest hardware and software supports. Therefore, network pruning is gaining attention for general in various fields. This paper proposes a novel channel method, Linearly Replaceable Filter (LRF), which suggests that filter approximated by linear combination other filters replaceable. Moreover, an additional method called Weights Compensation proposed to support LRF method. technique effectively reduces output difference caused removing via direct weight modification. Through experiments, we confirmed our achieves state-of-the-art performance several benchmarks. In particular, on ImageNet, LRF-60 approximately 56% FLOPs ResNet-50 without top-5 accuracy drop. Further, through extensive analyses, proved effectiveness approaches.