Slice Operator for Efficient Convolutional Neural Network Architecture

Van-Thanh Hoang, Kang-Hyun Jo

Abstract: Convolutional neural networks (CNNs) have shown remarkable performance in various computer vision tasks in recent years. However, the increasing model size has raised challenges in adopting them in real-time applications as well as mobile and embedded vision applications. A number of efficient architectures have been proposed in recent years, for example, MobileNet, ShuffleNet, MobileNetV2, and ShuffleNetV2. This paper describes an improved version of ShuffleNetV2, which uses the Channel Slice operator with slice-step parameters to make information interaction between two channels, instead of using Channel Shuffle and Channel Split operators. Because the Channel Slice and Channel Split operators are similar and the proposed architecture does not have Channel Shuffle operator, it has lower memory access cost than ShuffleNetV2. The experiments on ImageNet demonstrate that the proposed network is faster than ShuffleNetV2 while still achieves similar accuracy.