Practical Analysis on Running Speed for Efficient CNN Architecture Design

Convolutional neural networks (CNNs) have shown significant performance in solving various artificial intelligence tasks in recent years. However, the increasing model size has raised challenges in adopting them in limited-resource applications. Recently, many research works try to build efficient networks which are as small as possible and have small computation time while still have acceptable performance. The state-of-the-art architectures are ShuffleNetV2 and MobileNet. They use Depthwise Separable Convolution (DWConvolution) in place of standard Convolution to reduce the model size. Their designs are very efficient which follows many practical guild-lines. However, the ShuffleNetV2 has higher memory access cost (MAC) than the MobileNet. This paper evaluates these two networks on GPU and CPU with and without tuning step to show the advantages of each network. The experiments show that the MobileNet is faster on high-computational-optimization devices like GPU, whereas the ShuffleNetV2 is faster on low-computational-optimization devices like CPU when they have similar FLOPs.