verne.freenode.net changed the topic of #mlpack to: http://www.mlpack.org/ -- We don't respond instantly... but we will respond. Give it a few minutes. Or hours. -- Channel logs: http://www.mlpack.org/irc/
wasiq has joined #mlpack
skon46 has quit [Ping timeout: 248 seconds]
mrbean has quit [Ping timeout: 276 seconds]
mrbean has joined #mlpack
skon46 has joined #mlpack
skon46 has quit [Ping timeout: 244 seconds]
skon46 has joined #mlpack
Nilabhra has joined #mlpack
wasiq has quit [Ping timeout: 276 seconds]
Mathnerd314 has quit [Ping timeout: 250 seconds]
wasiq has joined #mlpack
Nilabhra has quit [Ping timeout: 264 seconds]
Nilabhra has joined #mlpack
ranjan123 has joined #mlpack
ranjan123 has quit [Ping timeout: 250 seconds]
u8sand has quit [Ping timeout: 252 seconds]
u8sand has joined #mlpack
Nilabhra has quit [Remote host closed the connection]
skon46 has quit [Ping timeout: 250 seconds]
skon46 has joined #mlpack
slardar has quit [Ping timeout: 244 seconds]
slardar has joined #mlpack
wasiq has quit [Ping timeout: 244 seconds]
u8sand has quit [Ping timeout: 250 seconds]
u8sand has joined #mlpack
wasiq has joined #mlpack
Nilabhra has joined #mlpack
travis-ci has joined #mlpack
< travis-ci>
mlpack/mlpack#737 (master - f73d898 : Ryan Curtin): The build has errored.
< Guest24490>
Hello. My name's Dean. I was here yesterday asking a question about the mlpack_nmf program
< Guest24490>
I am having trouble when running the program on matrix V[512][100] that the resulting W and H are filled with NAN. What could be the cause of this issue I wonder? V is initialised with 0, it then later holds Spectrogram data. 512 bins, 100 Frames
Mathnerd314 has joined #mlpack
ank_95_ has joined #mlpack
slardar has quit [Ping timeout: 252 seconds]
skon46 has quit [Read error: Connection reset by peer]
Nilabhra has quit [Ping timeout: 252 seconds]
slardar has joined #mlpack
Nilabhra has joined #mlpack
< rcurtin>
Guest24490: are you sure the input matrix does not have any NaNs in it?
< Guest24490>
Hello again :) Well I was under the impression that it contains all zeros except for where there is STFT data. I strongly believe there's no NaNs
< Guest24490>
I have checked that there are no NaNs in the input CSV
< Guest24490>
the input is filled with floats for 2 of its columns, then the rest of columns are 0
< rcurtin>
hm, is there any chance you can make that file available so I can try to reproduce the issue?
< Guest24490>
I really appreciate you trying to reproduce the issue. Thanks a lot!
< Guest24490>
also I am running with -- $ mlpack_nmf -H h.csv -i v.csv -r 5 -W w.csv
< rcurtin>
okay, I ran a quick test
< rcurtin>
the problem is that by default mlpack_nmf uses the multiplicative update rules
< rcurtin>
which divide by W * H * H.t()
< rcurtin>
so if W or H have a large number of zeros, this can cause big problems and introduce lots of NaNs
< rcurtin>
but your V matrix is mostly zeros, so W and H end up with a lot of zeros
< rcurtin>
I would say the right thing to do here is to use a different set of update rules... so you can also specify -u als
< Guest24490>
Oh right
< rcurtin>
with the ALS update rules it seems to converge to an actual solution instead of dividing by 0
< Guest24490>
Yeah that makes much sense
< Guest24490>
I was rather confused with this problem that has plagued me all day
< Guest24490>
running it now...
< Guest24490>
Ah yes!!
< Guest24490>
Thank you so much.
< Guest24490>
If I could buy you a beer I would man hahaha
< rcurtin>
:)
< rcurtin>
yeah, so one of the difficulties is that the Lee and Seung paper I linked you to yesterday doesn't really consider what happens when there are zeros in the W or H matrices
< rcurtin>
so this leads to weird behavior sometimes
< Guest24490>
I see! Yeah I did spend the day pretty baffled with it. And I didn't manage to find anything in the paper. Very happy to have overcame this issue
< rcurtin>
unfortunately it's computationally difficult to check for 0s which will cause this NaN issue in the update rules
< rcurtin>
I might consider switching the default rules to the ALS rules, but sometimes they can be much slower than the multiplicative distance update rules or the multiplicative divergence update rules
< Guest24490>
I see. I will definitely be able to consider this in future now. Thanks for the knowledge
< rcurtin>
sure, I am glad we could get the issue worked out :)
< Guest24490>
Yes :) there aren't many that I know I can turn to for NMF advice. Excellent! Have a great day !
travis-ci has joined #mlpack
< travis-ci>
mlpack/mlpack#740 (master - 637809f : Ryan Curtin): The build has errored.
< rcurtin>
you too, I hope your noise suppression technique is successful :)
< zoq>
looks like travis can't fetch the gpg key, hopefully this is just temporary, if not we could force apt-get update to install the package :)
< Guest24490>
Thank you :)
< rcurtin>
zoq: yeah, it seems like it's been down since this morning. I was going to wait a few days before looking into it further and hope it fixed itself
< rcurtin>
I saw some days ago that the AppVeyor build also couldn't access some packages with nuget, but that looks like it's fixed now
< zoq>
rcurtin: yeah, right, not sure, since the AppVeyor build fails every time, because of this ridiculous time constraints, we could delete the config.
< zoq>
rcurtin: Btw, I'm not even sure we need to add the ppa repo to build mlpack.
< rcurtin>
yeah, maybe we should remove the AppVeyor config entirely
< rcurtin>
hmm, let me try something first
< rcurtin>
ok, let's see if 62a9489 does anything
travis-ci has joined #mlpack
< travis-ci>
mlpack/mlpack#739 (master - 3781ccb : Ryan Curtin): The build was broken.
Nilabhra has quit [Remote host closed the connection]
tsathoggua has joined #mlpack
tsathoggua has quit [Client Quit]
< rcurtin>
as a last attempt to try and use appveyor, I sent the team an email to see if anything can be done
< rcurtin>
I'm hoping maybe they will make an exception and allow us 90 minutes of build time, but I am not too hopeful...
< zoq>
great :), I'm looking into some msbuild flags to speed up the build process, hopefully, the appveyor team can provide some insights and can provide some hints how to speed things up.
< zoq>
wow
< rcurtin>
yeah, nice, more CPU speed
< rcurtin>
hopefully it helps...
< zoq>
Hopefully it's for the next build, I'd like to see if switching to debug mode helps.
mrbean has quit [Ping timeout: 250 seconds]
mrbean has joined #mlpack
ank_95_ has quit [Quit: Connection closed for inactivity]
ank_95_ has joined #mlpack
travis-ci has joined #mlpack
< travis-ci>
mlpack/mlpack#744 (master - e047bcf : Marcus Edel): The build is still failing.