编辑雷竞技app下载苹果版

2018年机器学习:NumPyCNNAndroid:图书馆明确执行卷积神经系统为android设备-艾哈迈德穆罕默德Fawzy迦得Menoufia大学

文摘

另一个名为NumPyCNNAndroid建议限制的开源库构建和运行的开销卷积神经系统android设备。图书馆是用Python写3。它利用Kivy用于构建应用程序接口和numerical python构建系统本身。图书馆支撑最知名的层。与一般认识深刻学习库,NumPyCNNAndroid维护战略距离的额外开销使系统适合手机上运行。测试结果通过对图书馆的执行通过观察结果提出图书馆和TensorFlow依赖意味着最高失误。NumPyCNNAndroid是一种事业,为Android设备组装卷积神经系统利用NumPy Kivy。应用程序的目的是处理每三进步conv-relu-pool层,展示他们的收益率,返回客户端可以执行以下三层通过点击一个捕获屏幕的底部。过去结果之前攻将用于额外的处理。这种风险取决于过去的事业称为NumPyCNN然而NumPyCNNAndroid目前在Android。最近移动芯片系统(SoC)技术的发展,便携式Android deviceshas的性能增加了在过去的几年中倍数。 With theirmulti-core processors, dedicated GPUs, and gigabytes ofRAM, the capabilities of current smartphones have alreadygone far beyond running the standard built-in phone applica-tions or simple mobile games. Whereas their computationalpower already significantly exceeds the needs of most every-day use cases, artificial intelligence algorithms still remainchallenging even for high-end smartphones and tablets.Many recent developments in deep learning are, however,tightly connected to tasks meant for mobile devices. One no-table group of such tasks is concerned with computer visionproblems like image classification,image enhance-ment and super-resolution ,optical characterrecognition, object tracking ,visual scene under-standing, face detection and recognition ,gazetracking etc. Another group of tasks encompasses vari-ous natural language processing problems such as natural lan-guage translation ,sentence completion , sen-tence sentiment analysis ,or interactive chatbots .A separte group deals with on-line sensor data processing forhuman activity recognition from accelerometer data ,gesture recognition or sleep monitoring .Severalother deep learning problems on smartphones are related tospeech recognition, virtual reality and many other tasks.Despite the rising interest in deep learning for mobile ap-plications, the majority of AI algorithms are either not avail-able on smartphones or are executed on remote servers dueto the aforementioned phones’ hardware limitations. The lat-ter option is also not flawless, causing: a) privacy issues; b) dependency on an internet connection; c) delays associatedwith network latency; d) bottleneck problems — the numberof possible clients depends on the servers’ computational ca-pabilities. To overcome these issues, there were a number ofattempts to port separate algorithms or whole machine learn-ing libraries to mobile platforms with added hardware accel-eration (HA) using GPUs or DSPs. In the authors imple-mented a mobile neural network classification engine capableof sensor inference tasks on Qualcomm’s Hexagon DSP .Though they achieved very impressive energy consumptionresults, the DSP was able to run only very simple CNN modelsdue to its small program and memory space. In the au-thors presented a GPU-accelerated library CNNdroid for par-allel execution of pre-trained CNNs on mobile GPUs

传记:

埃及Menoufia大学学习

穆罕默德艾哈迈德Fawzy迦得

阅读全文下载全文