UltraGroup: Towards Deploying MobileNet to Ultra-Resource-Bounded Edge Devices
Wang Jihe, Huang Shu, Jiaxiang Zhao, Wang Danghui · 2019
Though MobileNet achieves significant memory and computing reduction by depth-wise separable convolutions, some grouping methods have showed further capabilities those resource-bounded devices on edge. Thereby we propose UltraGroup, an ultra-flexible parameter grouping method, to fit MobileNet into various kinds of resource-bounded embedded systems. Inspired by the ShuffleNet, this work largely scales the parameter quantities through Gapping and Overlap methods to search an optimized deploying balance between resource limitations and the model accuracy. Our experiment shows that 1) Gapping method makes the amount of parameters reduce by nearly half compared with original group convolution and 2) Overlap method increases the accuracy of the model, and the best accuracy has improved 1.3 times compared with the original group convolution.