13:27 < apir8181> Hi, I don't understand why mlpack need the class mlpack/core/util/NullOutputStream. If not in debug mode, we can just pass true to constructor argument ignoreInput.

13:30 < apir8181> At present, if DEBUG mode is enabled, core/util/Log::Debug would be type of PrefixedOutputStream. If not, it would be type of NullOutputStream.

14:02 < naywhayare> apir8181: the NullOutputStream causes every single call to Log::Debug to be optimized out entirely

14:02 < naywhayare> although it's possible that a smart compiler could realize that ignoreInput was true and Log::Debug was an unnecessary class, I don't think we can reasonably assume that every compiler will do that

14:04 < apir8181> naywhayare: That sounds cool.

18:35 < naywhayare> stephentu: sorry about the delay on the generalized eigensolver

18:35 < naywhayare> the SuperLU wrapping is going super slow

18:35 < naywhayare> the documentation isn't all that great and the package is really hard to work with... (that plus my time is pretty limited)

18:35 < naywhayare> but if I had to guess, you're pretty busy too :)

19:14 < stephentu> naywhayare: ya i'm trying to grind through some idea

19:14 < stephentu> naywhayare: while doing psets during the weekend

19:14 < stephentu> no rush man

19:14 < stephentu> i'm excited for this to be a gsoc project though

19:21 < naywhayare> yeah, I think it will be a good one

19:26 < stephentu> when do the apps start coming in?

19:37 < naywhayare> March 2 is when they announce accepted organizations

19:37 < naywhayare> if mlpack is accepted (it probably will be -- google likes machine learning), then the mailing list will explode

19:38 < naywhayare> applications start a few weeks later (March 16)

19:38 < naywhayare> the first year we got 50 or so, but last year we got fewer (I think because we made the applications a decent amount more difficult and were more realistic with students about what they needed to do to have a competitive chance)

19:38 < naywhayare> the quality of applicants last year was a lot higher

19:38 < naywhayare> (or, the average quality, that is)

21:07 < stephentu> oh god screening 50 applications sounds horrible

21:08 < stephentu> haha lets make it super restrictive: "must be intimiately familiar with nesterov's volumes on convex opt"

21:08 < stephentu> :)

21:08 < naywhayare> the applications are generally bimodal

21:09 < naywhayare> and most of the applications that are seriously considered are from students we've already been in touch with and who have already made a decent number of contributions

21:09 < naywhayare> so throwing away the ones that are crap is easy

21:09 < naywhayare> I got one the first year where the applicant had said "I didn't have time to finish the application, please click this link in a few days"

22:36 < lezorich> naywhayare: should I contribute first in order to be a GSoC student? I know that's the ideal, but the project I'm interested in is more research oriented (Fast k-centers algorithm & implementation)

22:49 < naywhayare> lezorich: a contribution isn't required, so don't take what I wrote earlier to heart as "I need to have 3 or more contributions to be accepted!"

22:49 < naywhayare> contributions are definitely helpful in showing us that you are capable of writing good code

22:49 < naywhayare> for the k-centers algorithm, you're right, it is more research oriented

22:50 < naywhayare> even so, the mlpack abstractions for dual-tree algorithms are complex, so it's certainly worth poking around with them and playing with them, and I wouldn't be surprised if you ended up contributing in the course of learning about the framework mlpack has

22:56 < vlad_gl> naywhayare: Hi! I tried to do #406, but have a different neighbours for X.col(i) and H.col(i). I build query for H.col([users list]) and then just compute AllkNN for H and query. Is that right?

23:01 < naywhayare> hi vlad, I'm still working on writing up the document for single-tree GMM training :)

23:01 < naywhayare> I hope to have it done in a day or two...

23:01 < naywhayare> anyway...

23:03 < naywhayare> if we use H instead of X, it should give us the same neighborhood of users unless my derivation is wrong

23:05 < naywhayare> so if I run AllkNN(H) (which uses H as both the queries and references), the resulting neighborhood should be identical to if we ran AllkNN(X) (with X as both the queries and the references)

23:05 < naywhayare> you're saying that you're doing that, but the results are different?

23:07 < vlad_gl> I try run AllkNN(H, query), there query is the matrix of H.col(user) columns.

23:08 < vlad_gl> where* :)

23:10 < naywhayare> yeah; and this gives different results than AllkNN(X, query) (where query is built from X.col(user))?

23:11 < vlad_gl> yes

23:12 < naywhayare> oh, right, I think my derivation is wrong

23:12 < naywhayare> X.col(i) = W * H.col(i)

23:13 < naywhayare> that part is correct

23:13 < vlad_gl> yeah :)

23:13 < naywhayare> but it is not true that the distance d(X.col(i), X.col(j)) = d(H.col(i), H.col(j))

23:15 < vlad_gl> d(x1,x2) = d(h1,h2)? i think you mean just order.

23:16 < naywhayare> yeah, that statement can't be true:

23:16 < naywhayare> d(X.col(i), X.col(j)) = d(W * H.col(i), W * H.col(j))

23:16 < naywhayare> but d(W * H.col(i), W * H.col(j)) != d(H.col(i), H.col(j))

23:16 < vlad_gl> ok

23:17 < naywhayare> so I guess I need to rethink my derivation a little bit

23:17 < naywhayare> the GMMs are higher on my priority list, though, so I'll get to those first...

23:18 < naywhayare> thanks for pointing that out :)

23:21 < lezorich> naywhayare: I agree with you that contributions are helpful :) And yes, Ryan Curtin also suggested me to look implementations of dual-tree algorithms in mlpack, so may be there I can contribute in order to make a strong application

23:22 < vlad_gl> naywhayare: not at all : )

23:30 < vlad_gl> so, while you working gmm document, I also looked at #345. gmm_diag only supports from armadillo version 4.40. and it's does not support real probabilities as parameter. So, I can try to implement this.

23:31 < vlad_gl> or it is depends on what are you working now?