15:19 < travis-ci> mlpack/mlpack#734 (master - e0e7ba5 : Ryan Curtin): The build is still failing.

15:19 < travis-ci> Change view : https://github.com/mlpack/mlpack/compare/1417d45d7ca3...e0e7ba5353e5

15:19 < travis-ci> Build details : https://travis-ci.org/mlpack/mlpack/builds/122808007

16:28 < uniden> hello, anybody here?

16:30 < zoq> too late

17:33 < palashahuja> hi, zoq

17:33 < palashahuja> I couldn't work on the benchmarks thing

17:34 < palashahuja> I was really busy these days

18:32 < wasiq> lol its like challenge on the irc to catch the people who ask the questions before they leave.

19:01 < Dean_> Greetings all! I have a question regarding the mlpack_nmf command line program. I am receiving the output in Matrices W and H "-nan" in all elements of the csv files. I am told that the number of random columns is more than available in matrix V, "weird results may ensue!" V is of size [2][512]. How do I specify the dimensions of W and H? Or what is happening here please? - Massive respect for your work. Thank you for your time. - Dean

19:02 < rcurtin> hey dean, you can specify the rank of the decomposition with --rank

19:02 < rcurtin> that will control the size of W and H

19:02 < rcurtin> are you srue that your csv file isn't transposed?

19:03 < rcurtin> if you have 512 points, then it should be 512 rows in your csv

19:03 < Dean_> It is at 2. (Just testing as I'm still trying to get my head around how the rank affects the output)

19:04 < Dean_> Hmm ok.

19:04 < rcurtin> yeah; a higher rank will mean a better reconstruction, generally

19:04 < rcurtin> your V matrix probably represents some data points, right?

19:05 < Dean_> I see. What is the recommended range of rank then please? Matrix V is representing Spectrogram data

19:05 < Dean_> I am building a Source Separation tool

19:07 < rcurtin> yeah, so in this case I am not sure why you are decomposing the matrix V; it is already only of rank 2 (at maximum)

19:08 < Dean_> Ok, thank you. I guess I need to read into the rank of decomposition.

19:09 < Dean_> Thanks for your time.

19:09 < rcurtin> sure, feel free to ask more questions :)

19:09 < rcurtin> I think maybe another good paper to read is the original NMF paper:

19:10 < rcurtin> http://www.public.asu.edu/~jye02/CLASSES/Fall-2007/NOTES/lee01algorithms.pdf

19:10 < rcurtin> in case you haven't seen that one already

19:10 < rcurtin> basically, NMF is well-suited to problems where you have a large data matrix V, and you want to represent it as two smaller matrices W and H

19:10 < rcurtin> but one of the keys of the decomposition is that your big matrix V is actually low-rank

19:11 < rcurtin> it looks like your data matrix is small, with size 2x512, so I don't know if NMF is applicable

19:11 < rcurtin> if the dataset was more like, I dunno, 100x512, I think you would have a good case for using NMF

19:13 < Dean_> thanks for the recommendation! I will read that paper. Well I had briefly come to understand that NMF would be the best for solving the task of Audio Noise removal. This is my undergraduate thesis project. I am building a tool to accomplish this. Tbe

19:13 < Dean_> The size of V is the result of taking a very short recording and its STFT data

19:15 < Dean_> I'm going to go away, read the paper, and think about this all. Thank you very much for your time!

19:15 < rcurtin> sure, I hope the paper is helpful

19:15 < rcurtin> :)

19:16 < Dean_> :) thanks, me too !

19:16 < rcurtin> I do think that NMF will be a good tool, but I suspect your V matrix should have each STFT as its own column

19:16 < rcurtin> so, e.g., maybe you have 1000 STFT frames and each of those have, say, 512 frequency bins in them

19:16 < rcurtin> then your V matrix is 512 x 1000

19:17 < rcurtin> when you read through the paper, think about H being a matrix of "prototype" STFT frames

19:17 < rcurtin> and every actual STFT frame in V is made up of some linear combination of those "prototypes" in H

19:17 < rcurtin> so when you choose the rank of the decomposition, what you are really choosing is the number of "prototypes"

19:17 < Dean_> Hmm ok

19:20 < Dean_> I will bare this in mind. Very helpful!

19:20 < Dean_> I have been trying to self-teach all of this so I appreciate the advice. Thanks a lot. Take care

19:21 < rcurtin> sure, it is a lot to learn, so it's normal if it seems overwhelming :)

19:22 < Dean_> hehe :) yeah, it is quite!

19:49 < travis-ci> mlpack/mlpack#735 (master - 77532c8 : Marcus Edel): The build is still failing.

19:49 < travis-ci> Change view : https://github.com/mlpack/mlpack/compare/e0e7ba5353e5...77532c84c6f4

19:49 < travis-ci> Build details : https://travis-ci.org/mlpack/mlpack/builds/122874713

20:20 < zoq> palashahuja: Hello, don't worry about the benchmarks.