We introduce an extremely computation-efficient CNN architecture named ShuffleNet, which is designed specially for mobile devices with very limited computing power (e.g., 10-150 MFLOPs). The new architecture utilizes two new operations, pointwise group convolution and channel shuffle, to greatly reduce computation cost while maintaining accuracy. Experiments on ImageNet classification and MS COCO object detection demonstrate the superior performance of ShuffleNet over other structures, e.g. lower top-1 error (absolute 7.8%) than recent MobileNet on ImageNet classification task, under the computation budget of 40 MFLOPs. On an ARM-based mobile device, ShuffleNet achieves ~13x actual speedup over AlexNet while maintaining comparable accuracy.
核心贡献 · Key contributions
提出逐点分组卷积以减少 1x1 卷积的计算量。 Proposes pointwise group convolution to reduce computation of 1x1 convolutions.
引入通道混洗操作以实现跨组信息流动。 Introduces channel shuffle operation to enable cross-group information flow.
构建 ShuffleNet 架构，在低算力预算下实现卓越精度。 Builds ShuffleNet architecture achieving superior accuracy under low FLOPs budgets.
在 40 MFLOPs 下，相比 MobileNet 实现 7.8%的绝对 top-1 错误率降低。 Demonstrates 7.8% absolute top-1 error reduction over MobileNet at 40 MFLOPs.
在基于 ARM 的移动设备上实现约 13 倍于 AlexNet 的实际加速。 Achieves ~13x actual speedup over AlexNet on ARM-based mobile devices.
在 MS COCO 目标检测上验证泛化能力，取得显著提升。 Validates generalization on MS COCO object detection with significant gains.
局限 · Limitations
通道混洗操作可能引入额外的内存访问开销。 Channel shuffle operation may introduce additional memory access overhead.
由于滤波器退化，较大分组数下的性能增益减弱。 Performance gains diminish for larger group numbers due to filter degradation.
深度可分离卷积在低功耗移动设备上因内存访问比而效率低下。 Depthwise convolution inefficient on low-power mobile devices due to memory access ratio.
由于实现开销，实际加速低于理论值。 Actual speedup lower than theoretical due to implementation overheads.
评估仅限于 ImageNet 和 COCO，未探索其他任务。 Evaluation limited to ImageNet and COCO; other tasks not explored.
论文章节 · Sections（共 13）
摘要Abstract
1 引言1 Introduction
2 相关工作2 Related Work
3 方法3 Approach
3.1 分组卷积的通道混洗3.1 Channel Shuffle for Group Convolutions
3.2 ShuffleNet 单元3.2 ShuffleNet Unit
3.3 网络架构3.3 Network Architecture
4 实验4 Experiments
4.1 消融研究4.1 Ablation Study
4.2 与其他结构单元的比较4.2 Comparison with Other Structure Units
4.3 与 MobileNets 及其他框架的比较4.3 Comparison with MobileNets and Other Frameworks