10:00 < cult-> rcurtin: oh ok thanks. what is your oppinion on my issue. is it affecting only me, or are you going to propose a patch soon?

12:47 < rcurtin> cult-: I think random initialization is better, so I will go ahead and patch today

13:19 < cult-> rcurtin: that's awesome

14:19 < sai___> hello sir..i wish to represent your organisation in GSOC 2017.can you help me how to start with your organisation.

21:09 < cult-> is there a list of mlpack usage in commercial and/or free softwares?

21:10 < rcurtin> cult-: probably the best you can do is search for it on google scholar and see what is citing mlpack

21:10 < rcurtin> I occasionally get emails from people who are using mlpack in their projects or work, but I never bothered to try and keep a comprehensive list

21:11 < cult-> rcurtin: may I ask why SVM isn't implemented?

21:12 < rcurtin> a couple things: lack of interested contributor

21:12 < rcurtin> and if we did implement SVMs, then we would need to outperform existing toolkits, so the implementation quality would either need to be high

21:13 < rcurtin> or the optimization would need to be implemented with better-scaling algorithms

21:13 < cult-> we could just take opencv or libsvm implementation but with armadillo

21:14 < rcurtin> I'd prefer to keep the list of dependencies small when possible, I'm not a huge fan of wrapping other ML libraries inside of mlpack

21:15 < cult-> sure, i take the code there and refactor it into mlpack. its likely not that huge work with arma.

21:15 < cult-> *i mean take ...

21:15 < rcurtin> there are possible licensing issues with doing that

21:15 < cult-> they are all open source afaik

21:17 < cult-> opencv is based on libsvm

21:17 < rcurtin> yes, that's right, but I haven't looked into the license details of either library. taking the code directly may modify the license that we have to distribute for mlpack, and there may be other complexities there

21:17 < cult-> ok, just an idea.

21:19 < rcurtin> yeah, definitely an idea, but I do think that a reimplementation from scratch can provide better runtime and also better flexibility

21:19 < rcurtin> I think libsvm is limited in the types of kernels that it can accept, for instance

21:19 < cult-> i agree, but their code may act as a blueprint?

21:19 < cult-> anyway its 3k lines

21:21 < rcurtin> yeah, a blueprint for reimplementation could be an ok place to start, although if we were going to do that we should be sure to note that our implementation is based on libsvm

21:21 < rcurtin> but there are almost certainly opportunities for optimization with armadillo

21:22 < rcurtin> one nice thing is that we have an automatic benchmarking system, so once the scripts are setup to benchmark libsvm and other implementations it will be very easy to see which is faster :)

21:22 < cult-> awesome

21:22 < cult-> however i am aware that there are gpu versions of it, and for some of the algorithms opencv also accelerate gpu

21:25 < rcurtin> if we can reduce svm to linear algebra operations with armadillo then nvblas can be used to do those computations on the GPU in some cases

21:27 < rcurtin> that might not be as good as a direct GPU implementation but it may also be pretty close :)

21:28 < cult-> probably good enough yes

21:28 < rcurtin> the Armadillo maintainer has told me he wants to write a GPU version of Armadillo

21:29 < rcurtin> so I am hoping that this can be used as a plug and play replacement in mlpack

21:31 < cult-> great