GhostNet: More Features From Cheap Operations

TLDR

Deploying CNNs on embedded devices is challenging because of limited memory and computation, and the redundancy of feature maps—an important characteristic of successful CNNs—has rarely been explored in architecture design. This work introduces a Ghost module that generates additional feature maps using inexpensive operations. The Ghost module derives many ghost feature maps from a small set of intrinsic maps via cheap linear transformations, can be inserted into existing networks, and is stacked into Ghost bottlenecks to build the lightweight GhostNet. Experiments on ImageNet show that GhostNet attains 75.7 % top‑1 accuracy, outperforming MobileNetV3 while maintaining comparable computational cost. Code is available at https://github.com/huawei-noah/ghostnet.

Abstract

Deploying convolutional neural networks (CNNs) on embedded devices is difficult due to the limited memory and computation resources. The redundancy in feature maps is an important characteristic of those successful CNNs, but has rarely been investigated in neural architecture design. This paper proposes a novel Ghost module to generate more feature maps from cheap operations. Based on a set of intrinsic feature maps, we apply a series of linear transformations with cheap cost to generate many ghost feature maps that could fully reveal information underlying intrinsic features. The proposed Ghost module can be taken as a plug-and-play component to upgrade existing convolutional neural networks. Ghost bottlenecks are designed to stack Ghost modules, and then the lightweight GhostNet can be easily established. Experiments conducted on benchmarks demonstrate that the proposed Ghost module is an impressive alternative of convolution layers in baseline models, and our GhostNet can achieve higher recognition performance (e.g. 75.7% top-1 accuracy) than MobileNetV3 with similar computational cost on the ImageNet ILSVRC-2012 classification dataset. Code is available at https://github.com/huawei-noah/ghostnet.