Is possible to accelerate sequentialfs using a GPU?

Question

Roberto Herrera-Lara 2012-3-25

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/33406-is-possible-to-accelerate-sequentialfs-using-a-gpu

Using SVM, Treebagger, Naive Bayes as a function handle but processing is too slow. About 7 hours in an array of double points 200x60000.

My idea is to use single precision.

2 个评论
显示无隐藏无

Walter Roberson 2012-3-25

Have you used to profiler to determine which part of the operations are slowest?

Roberto Herrera-Lara 2012-3-26

The problem is, sequentialfs, is very slow .... had intended to stop working with double and work with for gain speed....

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Ilya 2012-3-25

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/33406-is-possible-to-accelerate-sequentialfs-using-a-gpu#answer_42001

TreeBagger and sequentialfs can be run in a parallel mode using parfor. See 'help parallelstats' for more info. Executing these two utilities in the parallel mode would let you use multiple cores on your machine and, if you have MATLAB Distributed Computing Server, on multiple nodes in your computing cluster. You cannot use GPU's for these utilities, at least not in an obvious way. See the User Guide section in Parallel Computing Toolbox for descriptions of parfor and GPU computing.

Either way, if you have 60k predictors, computation is going to be quite slow.

If you have 12a, I recommend regularized discriminant analysis available from ClassificationDiscriminant class in the Statistics Toolbox. It is well suited for wide data with thousands of predictors. It provides a way of discarding poor predictors too.