Exploring the Capabilities of Mobile Devices Supporting Deep Learning
Exploring the Capabilities of Mobile Devices Supporting Deep Learning
复制标题
DOI：
10.1145/3220192.3220460
复制
发表时间：
2018-06
期刊：
影响因子：
--
通讯作者：
Yitao Chen;Saman Biookaghazadeh;Ming Zhao
中科院分区：
文献类型：
--
作者：
Yitao Chen;Saman Biookaghazadeh;Ming Zhao
借助越来越强大的移动设备，可以在设备上执行更深入的学习任务，并且在设备（例如个性化和效率）上学习也具有重要的优势。但是，通常缺乏对现代移动设备进行深度学习的能力的良好理解。为了解决知识的这一差距，本文介绍了一项有关在移动设备上执行深神网络（DNNS）的培训和推断的全面研究。这项研究基于TensorFlow+，这是广泛使用的Tensorflow框架的扩展，该框架框架使其能够在设备上训练DNN，并使用可用的GPU加速学习。我们研究最重要的结果是：1）网络的大小不仅要满足设备的内存限制，而且对于训练性能至关重要； 2）硬件加速对设备的学习速度很重要。通过使用设备的GPU加速前进和向后路径，我们的扩展张量可以将训练时间减少44.8％； 3）比较CPU，内存和电池使用情况，内存大小是对设备上训练网络的最严重限制。
With the increasingly more powerful mobile devices, it becomes possible to perform more deep learning tasks on the devices, and there are also important advantages of learning on devices, such as personalization and efficiency. However, a good understanding of the capabilities of modern mobile devices for deep learning is generally lacking. To address this gap in knowledge, this paper presents a comprehensive study on performing training and inference of deep neural networks (DNNs) on mobile devices. This study is based on TensorFlow+, an extension of the widely used TensorFlow framework that enables it to train DNNs on devices and use the available GPUs to accelerate the learning. The most significant results of our study are: 1) The size of the network is crucial not only to meet the device's memory constraint but also for training performance; 2) Hardware acceleration is important to the learning speed on devices. By accelerating both the forward and backward path with the device's GPU, our extended TensorFlow can cut down the training time by 44.8%; 3) Comparing CPU, memory, and battery usages, memory size is the most serious constraint to training networks on devices.