Source-linked AI summary
GhostNet: More Features from Cheap Operations
Kai Han, Yunhe Wang, Qi Tian, Jianyuan Guo, Chunjing Xu, Chang Xu
TL;DR
Deploying accurate CNNs on resource-constrained mobile devices is challenging, while feature-map redundancy remains underused in architecture design. GhostNet addresses this with cheap transformations that generate additional feature maps, and consistently outperforms competing small networks across computational-complexity levels, including 0.5% higher top-1 accuracy than MobileNetV3 at the same latency.
Problem
Efficient CNN deployment is difficult because mobile devices have limited resources, while feature-map redundancy has rarely been investigated in neural architecture design.
Method
The Ghost module uses fewer filters to create intrinsic feature maps, then applies cheap linear transformations to generate additional ghost features within Ghost bottlenecks and GhostNet.
Results
GhostNet consistently outperforms competing small networks across computational-complexity levels, including 0.5% higher top-1 accuracy than MobileNetV3 at the same latency.
Takeaways & Limitations
The Ghost module provides a plug-and-play way to convert conventional CNNs into compact models while maintaining comparable performance.
Abstract
from arXiv · showhide
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
1. Introduction
The introduction motivates efficient CNN design by highlighting the high cost of conventional networks and the redundancy present in their feature maps. It then proposes Ghost modules, which generate additional feature maps from controlled intrinsic features using cheap linear operations.
- Motivation: 25.6M parameters and 4.1B FLOPs make ResNet-50 costly for processing a 224 × 224 image.These costs motivate the recent focus on efficient deep neural network design.
- Related approaches: Compact-network methods include pruning, low-bit quantization, knowledge distillation, and efficient architectural units.MobileNet uses depthwise and pointwise convolutions, while ShuffleNet explores channel shuffle operations.
- Feature redundancy: Well-trained CNNs contain abundant and redundant feature-map information, including similar feature-map pairs that can reveal related features.The introduction treats this redundancy as an important characteristic of successful deep neural networks rather than something to eliminate outright.
- Ghost module: The Ghost module splits convolution into controlled ordinary convolutions followed by simple linear operations that generate more feature maps from intrinsic feature maps.It is introduced as a way to generate more features using fewer parameters while preserving the output feature-map size.
2. Related Work
Related work on lightweight neural networks spans model compression and compact model design. Compression reduces computation, energy, and storage through methods such as pruning, quantization, decomposition, and distillation, while compact architectures use efficient convolutional designs and automated optimization.
- Model compression: Model compression reduces computation, energy, and storage costs through pruning connections and channels, quantization, tensor decomposition, and knowledge distillation.Pruning removes unimportant connections or useless channels; other methods exploit discrete values, low-rank redundancy, or larger teacher models.
- Model compression: Binarization uses 1-bit values and efficient binary operations to accelerate neural networks.The passage describes binarization as a specific quantization approach.
- Compact model design: Compact model design targets embedded deployment with efficient architectures including Xception, MobileNets, MobileNetV2, and MobileNetV3.Xception and MobileNets use depthwise convolution variants, MobileNetV2 introduces inverted residual blocks, and MobileNetV3 adds AutoML technology.
3. Approach
The approach exploits redundancy in convolutional feature maps by generating a small set of intrinsic features and expanding them with cheap linear operations. Ghost bottlenecks stack these modules into GhostNet, with width multipliers enabling uniform network scaling.
- Ghost Module: Ghost modules reduce resource use by treating redundant output feature maps as “ghosts” of a smaller set of intrinsic features.The method argues that similar feature maps need not be generated independently with large numbers of FLOPs and parameters.
- Ghost Module: Cheap linear operations applied channel-wise to each intrinsic feature generate ghost features, producing n = m · s output feature maps.The final operation preserves each intrinsic feature through identity mapping, while the other operations generate additional ghost maps at much lower cost than ordinary convolution.
- Ghost Module: Using linear operations of a common size within each module improves efficient implementation on CPU or GPU hardware.The paper notes that heterogeneous operation shapes could obstruct online inference and therefore recommends uniform sizes such as 3 × 3 or 5 × 5.
- Ghost Bottlenecks: Ghost bottlenecks consist mainly of two stacked Ghost modules, with the first serving as an expansion layer that increases channel count.For stride=2, the shortcut uses downsampling and a depthwise convolution with stride=2 is inserted between the Ghost modules.
- GhostNet: GhostNet replaces MobileNetV3 bottleneck blocks with Ghost bottlenecks and builds the network as a stack of these modules.The architecture follows MobileNetV3 as a basic design, while further tuning or architecture search may improve performance.
- Width Multiplier: A width multiplier α uniformly scales the number of channels at every layer to obtain smaller, faster models or higher-accuracy variants.The resulting model is denoted GhostNet-α×.
4. Experiments
Experiments evaluate Ghost modules as replacements for convolutional layers and test GhostNet across image-classification and object-detection benchmarks. Results show strong feature-map redundancy, favorable accuracy–efficiency tradeoffs, and improved mobile inference performance.
- Datasets and Settings: Experiments replace convolutional layers with Ghost modules and evaluate GhostNet on CIFAR-10, ImageNet ILSVRC 2012, and MS COCO.CIFAR-10 is used to analyze method properties, while ImageNet and MS COCO provide larger-scale classification and detection benchmarks.
- Toy Experiment: Extremely small MSE values show strong correlations between deep feature maps, supporting generation of redundant maps from intrinsic feature maps.A small depthwise convolution learns the linear mapping between feature-map pairs, with kernel sizes d ranging from 1 to 7.
- Analysis on Hyper-parameters: d = 3 performs better than smaller or larger kernels, so it is adopted for subsequent experiments because 1 × 1 kernels lack spatial information while larger kernels overfit and increase computation.The kernel-size analysis is conducted on the CIFAR-10 validation set with s = 2.
- Analysis on Hyper-parameters: Increasing s significantly reduces FLOPs while gradually decreasing accuracy, reflecting the expected tradeoff between compression and recognition performance.With d = 3, s controls the number of intrinsic feature maps and the resulting computational cost.
- ImageNet Classification: About 2× acceleration and compression are achieved by Ghost-ResNet-50 with s=2 while maintaining the original ResNet-50 accuracy, outperforming several state-of-the-art methods at that setting.The comparison uses ImageNet and includes Thinet, NISP, Versatile filters, and Sparse structure selection.
- Actual Inference Speed: 0.5% higher top-1 accuracy than MobileNetV3 is obtained by GhostNet at the same latency, while requiring less runtime to achieve similar performance.Inference is measured on an ARM-based mobile phone with TFLite in single-threaded mode and batch size 1.
5. Conclusion
The paper presents a Ghost module for building efficient neural architectures with reduced computational cost. It generates intrinsic feature maps using fewer filters, then applies cheap transformations to produce additional ghost feature maps.
- Conclusion: The Ghost module reduces computational costs by splitting an original convolutional layer into two parts.It uses fewer filters to generate intrinsic feature maps before producing additional features through cheap transformations.
- Conclusion: Cheap transformation operations efficiently generate ghost feature maps from intrinsic feature maps.This design provides the basis for building efficient neural architectures.