[Paper Review] GhostNet: More Features from Cheap Operations

Kai Han, Yunhe Wang+4 more|arXiv (Cornell University)|Nov 27, 2019

Advanced Neural Network Applications63 references112 citations

TL;DR

This paper introduces the Ghost module to generate more feature maps from cheap linear operations, enabling GhostNet to outperform similar-cost lightweight models while using fewer parameters. It demonstrates plug-and-play applicability to existing CNNs and strong results on ImageNet and COCO.

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

Motivation & Objective

• Motivate efficient CNN design for mobile/devices by exploiting redundancy in feature maps.
• Propose a Ghost module to generate many feature maps from a few intrinsic maps with cheap operations.
• Demonstrate that Ghost modules can replace standard convolutions in existing architectures to form a lightweight yet accurate network (GhostNet).
• Validate performance gains on ImageNet classification and MS COCO object detection compared with state-of-the-art compact models.

Proposed method

• Split a standard convolution into a primary convolution producing intrinsic feature maps and a set of cheap linear transformations to generate ghost feature maps.
• Use two Ghost modules per Ghost bottleneck, where the first expands channels and the second reduces to match the shortcut, forming a Ghost bottleneck (G-bneck).
• Replace bottleneck blocks in MobileNetV3-like architectures with Ghost bottlenecks to build GhostNet.
• In GhostNet, set the primary convolution kernel size to 1x1, use s=2 (generating n/s intrinsic maps per block), and d=3 for the depthwise transformations.
• Optionally apply Squeeze-and-Excite (SE) modules to certain residual paths.
• Provide a width multiplier alpha to scale GhostNet complexity.

Experimental results

Research questions

• RQ1 Can a two-stage Ghost module (intrinsic maps plus cheap linear transformations) reduce computation while preserving accuracy?
• RQ2 Do Ghost bottlenecks integrated into a MobileNetV3-like backbone yield superior accuracy-efficiency trade-offs on ImageNet and COCO compared with state-of-the-art compact models?
• RQ3 How do hyperparameters s (expansion of intrinsic maps) and d (kernel size of linear transformations) affect accuracy and FLOPs on benchmark datasets?

Key findings

• Ghost module can significantly reduce FLOPs and parameters while preserving or improving accuracy on CIFAR-10, ImageNet, and COCO benchmarks.
• On ImageNet classification, GhostNet achieves higher top-1 accuracy than MobileNetV3 at similar computational cost across multiple complexity levels (e.g., GhostNet-1.0x achieves 73.9% top-1 with ~141 MFLOPs).
• GhostNet with width multipliers up to 1.3x achieves 75.7% top-1 accuracy at ~226 MFLOPs, outperforming several compact architectures.
• For mobile-ready speed, GhostNet yields faster actual inference than comparable MobileNetV3 implementations at similar accuracy levels on ARM/TF Lite benchmarks.
• In COCO object detection, GhostNet-based backbones deliver competitive mAP (e.g., 26.6-26.9% with ~164-300 MFLOPs) against MobileNetV2/V3 backbones.
• GhostNet demonstrates strong feature-map efficiency, achieving substantial gains with a partial replacement of traditional convolutions by cheap linear transformations.