作者:Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, Jian Sun
作者单位:Megvii Inc (Face++),旷视
发布时间:2018
发布期刊/会议:CVPR
**论文全称:**ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices
论文地址:
CVPR 2018 Open Access Repository
论文代码:
https://github.com/megvii-model/ShuffleNet-Series
DeepLearning_tutorials/ShuffleNet.py at master · xiaohu2015/DeepLearning_tutorials
地位:针对移动端边缘计算的极致轻量化卷积神经网络
We introduce an extremely computation-effificient CNN architecture named ShufflfleNet, 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 maintainingaccuracy. Experiments on ImageNet classifification and MSCOCO object detection demonstrate the superior performance of ShufflfleNet over other structures, e.g. lower top-1 error (absolute 7.8%) than recent MobileNet [12] on ImageNet classifification task, under the computation budget of 40 MFLOPs. On an ARM-based mobile device, ShufflfleNet achieves ∼13× actual speedup over AlexNet while maintaining comparable accuracy.