【摘要】：針對(duì)受限波爾茲曼機(jī)處理大數(shù)據(jù)時(shí)存在的訓(xùn)練緩慢問(wèn)題,在Hadoop云計(jì)算平臺(tái)和GPU并行加速的基礎(chǔ)上設(shè)計(jì)了基于Hadoop-GPU框架的的RBM加速計(jì)算實(shí)現(xiàn)方法.通過(guò)對(duì)MapReduce機(jī)制和RBM訓(xùn)練過(guò)程的分析,將RBM訓(xùn)練分割為采用Map端實(shí)現(xiàn)吉布斯采樣,Reduce端實(shí)現(xiàn)參數(shù)更新,并通過(guò)GPU實(shí)現(xiàn)運(yùn)算并行加速的方法組合.最后通過(guò)MNIST手寫數(shù)字識(shí)別集實(shí)驗(yàn)證明,在大規(guī)模數(shù)據(jù)下,Hadoop-GPU平臺(tái)對(duì)RBM的訓(xùn)練具有良好的可行性,加速比達(dá)到20以上,并且隨著數(shù)據(jù)規(guī)模的增加,加速比呈現(xiàn)更為顯著的增長(zhǎng).
[Abstract]:Aiming at the problem of slow training of constrained Boltzmann machine in big data processing, a RBM accelerated computing implementation method based on Hadoop-GPU framework is designed on the basis of Hadoop cloud computing platform and GPU parallel acceleration. Based on the analysis of MapReduce mechanism and RBM training process, the RBM training is divided into two parts: the Gibbs sampling and reduce end is used to realize the parameter updating and the GPU is used to realize the method combination of parallel computation acceleration. Finally, the experiment of MNIST handwritten numeral recognition set shows that the Hadoop-GPU platform is feasible for RBM training under large scale data, and the speedup ratio is more than 20, and the speedup increases more significantly with the increase of data scale.
【作者單位】： 海軍航空工程學(xué)院信息融合研究所;山西美佳礦業(yè)裝備有限公司;
【基金】：國(guó)家自然科學(xué)基金(61032001)
【分類號(hào)】：TP332

【參考文獻(xiàn)】

相關(guān)期刊論文 前2條

1 江務(wù)學(xué);張t，

本文編號(hào)：2139363