00:56 < marcosirc> rcurtin: Thanks for that comment about Argentina. I hope we win the game tomorrow! :)

01:00 < marcosirc> zoq: Are benchmarks updated somewhere?

01:01 < marcosirc> I mean, is the benchmark system regularly executed?

01:01 < marcosirc> I can see some messages in IRC.

06:01 < rcurtin> marcosirc: I hope to be able to watch the game tomorrow; I saw the third placeatch (or some of it) today. the US did better against colombia than argentina but not good enough for a win I suppose :)

06:01 < rcurtin> *third place match

06:07 < nilay> zoq: ok, i am fixing these now.

09:11 < nilay> how do we ensure whether FullConvolution or NaiveConvolution is performed?

10:35 < zoq> nilay: I think you mean, ValidConvolution and FullConvolution: NaiveConvolution<ValidConvolution>::Convolution(...) or NaiveConvolution<FullConvolution>::Convolution(...)

10:55 < nilay> yes i don't know why it doesn't work for me

10:55 < zoq> nilay: do you get some error?

10:55 < nilay> i am writing this: mlpack::ann::NaiveConvolution::Convolution<FullConvolution>(InImage, k_mat, Output);

10:56 < nilay> and the previous one that you wrote

10:56 < nilay> mlpack::ann::NaiveConvolution<FullConvolution>::Convolution(InImage, k_mat, Output);

10:57 < nilay> do we need any other include other than naive_convolution.hpp

10:59 < zoq> #include <mlpack/methods/ann/convolution_rules/border_modes.hpp>

11:02 < zoq> That's where FullConvolution and ValidConvolution is defined.

11:05 < nilay> do i need anything else also?

11:06 < zoq> no, that's it

11:07 < nilay> i get some weird linking error now

11:08 < nilay> why do we have to include border_modes.hpp when its already included in naive_convolution.hpp

11:09 < zoq> oh, you are right, it's already included in naive_convolution.hpp

11:11 < zoq> maybe you can show me the error?

11:12 < nilay> http://pastebin.com/68HTRxQK

11:14 < nilay> i am writing this new function to test time: http://pastebin.com/QYukRaAU

11:14 < nilay> haven't changed anything else much

11:15 < zoq> hm, do you need mlpack::ann::... or does ann::... work? Also does it work if you remove the two lines?

11:24 < nilay> hm, it doesn't work after removing them too, this is weird

11:25 < nilay> i am checking it, i should be able to figure it out.

11:25 < nilay> thanks for your help.

11:25 < zoq> sure, here to help

18:28 < zoq> marcosirc: Sorry for the slow response, it takes weeks, to run the complete benchmark suite. So, we usually do this once there is a new release. The results you see in the channel, are from the commit benchmark, which runs a benchmark on a smaller subset. I guess, what we could do, is to integrate the results into the nighly benchmark which take about 13 hours, that could be interesting.

18:28 < zoq> http://www.mlpack.org/benchmark-daily.html ... what do you think?

19:05 < zoq> nilay: My timings: https://gist.github.com/zoq/85fff46864bc4c55ba6567bd6856584c

19:12 < zoq> ah, wait for NumTree = 1, let's see how it looks like for NumTree=8

19:17 < zoq> okay, similar results: https://gist.github.com/zoq/280d9de9eab7d58c18808990dab8e780

19:19 < nilay> did you do any major changes?

19:20 < zoq> no, just fixed the errors mentioned in the PR, let's see how we can increase the performance

19:22 < nilay> naiveconvolution and sobel filters?

19:23 < zoq> https://github.com/mlpack/mlpack/pull/703/files#r68506617 and https://github.com/mlpack/mlpack/pull/703/files#r68494923

19:26 < zoq> I wasn't able to run the executable, before I fixed the errors; about 10 seconds just for the PCA step ... hm

19:33 < nilay> these links show all the files

19:35 < zoq> oh, okay ... can you see the last comment, about 2 hours ago?

19:35 < nilay> only the comments that you made

19:35 < nilay> yes ok

19:36 < zoq> the one about -DBL_MAX and rowvec

19:43 < nilay> ok

20:11 < marcosirc> zoq: Thanks for your reply.

20:11 < marcosirc> Ok, but I am not sure if it would be useful, because I can see that nighly benchmark only consider small datasets for allknn

20:12 < marcosirc> it includes: isolet, corel-histogram, covtype and cloud.

20:12 < marcosirc> which are the same I included in my comparison.

20:13 < marcosirc> I think it would be useful to compare with bigger datasets, and also, running flann and ann too.

20:15 < marcosirc> I can do it in my computer, and then upload it to github.io. I think it would take some days...

20:17 < marcosirc> What do you think is the best option?

20:41 < zoq> marcosirc: I could just run the neighbor search benchmarks and update the results by using UPDATE=True, instead of running the complete benchmark. I think that's the best option.

20:41 < zoq> That would save us a lot of time.

20:42 < marcosirc> zoq: yeah. That would be great!

20:42 < marcosirc> Is it possible to run flann and ann too ?

20:42 < marcosirc> You should recompile their command line tools.

20:42 < zoq> yes, I'll do that tomorrow

20:45 < marcosirc> Ok, perfect. Thanks.

20:47 < nilay> zoq: http://pastebin.com/Yi361chf

20:48 < zoq> looks like using ann::FullConvolution instead of FullConvolution would solve the issue

20:49 < nilay> yes ok thanks

20:51 < nilay> its better to use namespace i guess

20:51 < zoq> yeah

21:45 < nilay> zoq: in the ConvTriangle function, using submat is faster than using NaiveConvolution

21:46 < nilay> here are the times i got: http://pastebin.com/LBJcZkYA

21:48 < nilay> here are the 2 functions, one using naiveconvolution, and one not: http://pastebin.com/yjE6n97S

21:54 < nilay> i haven't done submat after doing fullconvolution

22:13 < nilay> should i use the sepFilter2D function for sobel filter too?

22:33 < zoq> hm, could you start the timer before the actual NaiveConvolution function, because we could use a pregenerated kernel matrix

22:34 < nilay> i was generating the kernel in the sepfilter2d also, so i did to equate them.

22:34 < zoq> ah, okay

22:34 < nilay> anyways i don't think that would take more than 0.01 s

22:36 < nilay> what does armadillo do, so that this is coming fast?

22:37 < zoq> it uses template meta-programming techniques: http://arma.sourceforge.net/armadillo_joss_2016.pdf

22:38 < zoq> if your ConvTriangle is faster, I don't see a reason to use the conv method

22:39 < nilay> do you think same vectorized code could be written in naiveconvolution

22:40 < nilay> also, when we write a vectorized step, how is it implemented internally

22:42 < zoq> if, the SepFilter2D turns out to be faster, I think it's a good idea to change the naive convolution implementation. We should run more test with larger kernels, to see if it's always faster.

22:44 < zoq> I guess, in that case you have to take a look at the armadillo code.

22:45 < nilay> ok

22:49 < zoq> nilay: Can you also test FFTConvolution?

22:49 < zoq> I'm just curious :)

22:49 < nilay> ok yes

22:49 < zoq> thanks :)

22:57 < nilay> oops, FFTConvolution takes a ridiculous amount of time

22:58 < zoq> yeah, it should be faster when using a larger kernel

23:00 < nilay> http://pastebin.com/sGg0nd0d

23:01 < zoq> okay, thanks for the results ... so using SepFilter2D sounds like a good idea, right?

23:01 < nilay> yes, i think so

23:01 < zoq> okay, good

23:02 < nilay> maybe i should change the name of this function

23:03 < zoq> good idea

23:03 < nilay> should i name it Convolution

23:04 < zoq> Convolution sounds good to me

23:04 < nilay> :)