[mlpack-git] [mlpack] mlpack KMeans class much slower than armadillo kmeans() (#514)

Ryan Curtin notifications at github.com
Thu Feb 4 15:31:22 EST 2016


Thanks for doing this extra testing.  I'm going to reopen this bug then, since it seems like there is good reason to believe that there actually is an efficiency issue (and also now that I'm more awake than when I originally responded to this).  The things that should happen are

 - improvement of the current `NaiveKMeans` code to at least be roughly equivalent to the Armadillo code, which may mean using the internal Armadillo code, depending
 - OpenMP at least for the `NaiveKMeans` strategy, possibly also for `ElkanKMeans` and `HamerlyKMeans`; the tree-based ones are a lot harder so OpenMP may not be as useful there...
 - production of some nice benchmarks and maybe addition of k-means to the benchmarking system (which unfortunately is down at the moment, but with any luck might come back up next week), in order to verify that mlpack's naive k-means implementation is at least as good as Armadillo's.

Anyway, it may be a while until I'm able to look into this more fully, but I will, when I have a chance...

---
Reply to this email directly or view it on GitHub:
https://github.com/mlpack/mlpack/issues/514#issuecomment-180033568
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.cc.gatech.edu/pipermail/mlpack-git/attachments/20160204/9260a946/attachment-0001.html>


More information about the mlpack-git mailing list