LP-SVM在CUDA架构上的加速实现  

作　　者：兰远东[1] 

机构地区：[1]惠州学院计算机科学系,广东惠州516007

出　　处：《数字技术与应用》2014年第4期130-132,134,共4页Digital Technology & Application

摘　　要：为了能在PC机上处理大规模数据集问题,提出了使用CUDA架构对LP-SVM的加速实现方法。该方法针对PC机内存小的缺点,对SVM的分解算法进行改编,得到求解LP-SVM的分解算法。LP-SVM分解算法每次只需要求解一个小规模的线性规划问题,避免一次性把所有训练数据都装进内存。同时把求解线性规划中比较耗时的矩阵运算,移植到CUDA上进行,提高了求解效率。实验结果表明:LP-SVM算法在经过CUDA加速以后,算法的执行效率提高了10-35倍。In order to handle the problem of large-scale data processing on PC, an acceleration method which uses the CUDA architecture to accelerate the implementation of LP-SVM is proposed. Considering the limit memory of PC, the decomposition algorithm of SVM is modified to solve the LP-SVM. Each time the new decomposition algorithm of only solves a small-scale linear programming problem and avoids loading all the training data into memory. The comparative time-consuming parts of linear programming are migrated to CUDA platform, which has greatly improved calculation efficiency. Experimental results show that the efficiency of LP-SVM algorithm has been speeded up by 10 to 35 times.

关 键 词：机器学习 线性规划 支持向量机 图形处理器 统一计算设备架构 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]