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/
curiousguy13 has quit [Ping timeout: 240 seconds]
curiousguy13 has joined #mlpack
curiousguy13 has quit [Ping timeout: 276 seconds]
dataVinci has joined #mlpack
awhitesong has joined #mlpack
awhitesong has left #mlpack []
stephentu has quit [Ping timeout: 245 seconds]
curiousguy13 has joined #mlpack
curiousguy13 has quit [Ping timeout: 272 seconds]
curiousguy13 has joined #mlpack
curiousguy13 has quit [Ping timeout: 245 seconds]
curiousguy13 has joined #mlpack
dataVinci has quit [Quit: Leaving]
curiousguy13 has quit [Ping timeout: 272 seconds]
jbc_ has quit [Ping timeout: 272 seconds]
jbc_ has joined #mlpack
nice_guy has joined #mlpack
nice_guy has quit [Client Quit]
stephentu has joined #mlpack
< naywhayare>
stephentu: sorry for the delay in reply... I was out of town this weekend
govg has quit [Quit: leaving]
< naywhayare>
was participating in a 24 hour go-kart race... very sore now :)
govg has joined #mlpack
govg has quit [Remote host closed the connection]
govg has joined #mlpack
govg is now known as Guest65831
Guest65831 has quit [Ping timeout: 245 seconds]
govg has joined #mlpack
govg has quit [Client Quit]
< zoq>
/msg naywhayare Hey, I talked with udit yesterday. He is interested in working on deep learning.
< zoq>
Maybe I should work on my irc skills ...
< naywhayare>
hey, you do better than me usually :)
< stephentu>
naywhayare: wait do you actually drive a go-kart for 24 hrs?
< stephentu>
naywhayare: i thought go karts were things you drove when you went to the amusement park lol
< stephentu>
also FYI my mlpack involvement will probably substnatially decrease in the next few weeks
< stephentu>
but i'm still happy to get the occasional PR in
< stephentu>
it was a lot of fun to hack on this stuff :)
< naywhayare>
stephentu: it's a team race -- I did about 6 hours of the driving. we ended up 3rd, so I got to come home with a moderately sized trophy :)
< naywhayare>
they go pretty fast, up to 100mph on some tracks
< stephentu>
wtf are you serious
gabor has joined #mlpack
< stephentu>
hahahah thats cool
< stephentu>
its like mario kart IRL
< gabor>
Hi guys :)
< naywhayare>
yeah, that's basically what it is :)
< stephentu>
too bad you cant drop mushrooms
< naywhayare>
gabor: hello! :)
< stephentu>
for your competitors
< naywhayare>
maybe banana peels and stuff... I've considered it...
< naywhayare>
will you have time to finish the interior point solver? I know how busy it gets during the semester, so, no pressure or anything, but I'm happy to help out where I can
< stephentu>
oh ya i will
< stephentu>
i meant i probably wont be doign anymore ambitious PRs
< gabor>
just a quick question :) It seems mlpack does not have decision tree algorithms, is that true?
< naywhayare>
gabor: you are correct
< naywhayare>
stephentu: okay; hopefully your research will use some of your SDP contributions. and if not, I'll eventually get around to building something on top of it to provide LMNN/MVU implementations and stuff like that
< gabor>
is there a specific reason for that, or just simply no one volunteered to implement one for you?
< naywhayare>
gabor: nobody's implemented one, that's all
< naywhayare>
Udit Saxena implemented a decision stump last year as part of his GSoC project, but we don't have CART or anything
< stephentu>
i think DT would make mlpack infinitely more appealign b/c apparenlty DT is all they use in industry
< naywhayare>
stephentu: I agree with that. for me it's simply an issue of lacking the necessary time
< naywhayare>
I honestly don't think it would be very hard to provide a fast, robust decision tree implementation
< stephentu>
ya i also want to get a generalized empiricaly risk minimziation solver in
< stephentu>
that takes arbitrary convex loss function
< stephentu>
(which woudl subsume SVM)
< naywhayare>
one of the difficulties is handling categorical or non-numeric data; there's not really any framework for that at the moment
< stephentu>
but no time
< naywhayare>
since we assume arma::mat (or similar) for all data
< gabor>
cool :) I am looking for challenges for myself and this can be a suitable one for me :) also would like to participate on the next GSoC
< stephentu>
well we coudl use arma::umat
< stephentu>
for categorical
< naywhayare>
yeah, or arma::mat with integer columns or something. there needs to be a conversion step first, to map things like textual labels to numeric identifiers, but that's not too hard
< gabor>
currently doing my PhD at York, UK and completely new to mlpack
< naywhayare>
:)
< naywhayare>
I have still yet to travel to the UK. can't seem to get any papers in conferences there...
< gabor>
I have implemented a lots of ML algorithms (e.g. linear/logistic regression, neural net, decision tree, k-means), but not in C++ so far
< stephentu>
gabor: what is your weapon of choice?
< gabor>
Well, I have a laptop :) and have good knowledge on linux :)
< stephentu>
haha i meant language
< gabor>
:D :D
< gabor>
mlpack is in c++, isn't it? that leaves me only one choice.
< stephentu>
naw i'm just curious
< naywhayare>
I think he meant the language you usually use for implementation :)
< stephentu>
ya sorry
< stephentu>
weapon of choice is a bit vague
< gabor>
ohhh, sorry :) Java
< stephentu>
ouch
< stephentu>
no scala?
< gabor>
I was always a Java guy, a bit Scala nowadays
< stephentu>
i'ma ctually trying to find a library to show you
< stephentu>
that i think really did a great job w/ the type system
< stephentu>
for DTs
< stephentu>
which might serve as some inspiration
< stephentu>
i really like what they did w/ the type system here
< stephentu>
it might be too functional for mlpack
< stephentu>
i dont think our users care about monoids
< naywhayare>
I'm not big on functional programming but that's more due to ignorance than anything
< naywhayare>
the vast majority of users will want the simplest possible option, though, so I think most mlpack users use the command-line programs or make basic modifications to the code in the tutorials
< naywhayare>
I'll poke through this... thanks for the link
< stephentu>
ya thats the thing i think our users
< stephentu>
dont want to deal w/ this stuff
< stephentu>
but i like the abstractions here
< naywhayare>
yeah; I think building flexibility into our designs is extremely important, though
< naywhayare>
but I generally have accomplished that through templates. I'm interested to see how it's done in brushfire, and I suspect I will be learning some scala in order to understand it...
< gabor>
brushfire seems a very interesting project. I will pick up some general knowledge on your project, check out the code and run some example and analyse it. Any advice on this?
< stephentu>
naywhayare: one cool thing they do is they use an algebraic data type to represent categorical data
< stephentu>
so it can be like Ordered
< stephentu>
Continuous
< stephentu>
(well in this case its not categorical)
< stephentu>
and then you have these type traits basically that say what the type is capable of
< stephentu>
and you can define different comparators and stuff
< naywhayare>
stephentu: yeah, I'll be paying close attention to how they represent different data types, and think about how it might be done for mlpack in a way that won't be slow
< stephentu>
ya the problem there is every element is wrapped by a jvm object
< naywhayare>
yeah
< stephentu>
and you ahve to essentially swtich on the type
< stephentu>
for every operation
jbc_ has quit [Quit: jbc_]
< naywhayare>
Weka is super slow in part because of the flexibility it gives in that respect... their Dataset objects (I think that's what they are called) are just super slow and super heavy
< stephentu>
sigh its always this
< stephentu>
great abstractions
< stephentu>
too slow
< naywhayare>
yeah... :(
< naywhayare>
with C++ you can often get both, but making the compiler actually provide something fast is very hard
< naywhayare>
and if something goes wrong you get 10 billion lines of errors...
< stephentu>
which is why i switched to clang
< stephentu>
clang has awesome error messages
< gabor>
:)
< stephentu>
also have you been following this rms discussion
< stephentu>
about gcc asts?
< naywhayare>
not at all
< naywhayare>
sounds like I should look it up...
< gabor>
I still like all the services what the JVM can give for the Java programs, but I have to agree that objects are super heavy and have a lots of stuff what usually we don't need
< gabor>
I really like the RAII concept of the C languages which I think is much better than the Java's GC idea
< naywhayare>
yeah... I'm not a fan of GCs really
< naywhayare>
at least not when I am writing code where I care about the performance
< gabor>
I played a bit with Weka in the past and I felt its speed... it's just simply slow...
< naywhayare>
yeah, when I have benchmarked weka it is orders of magnitude slower than other libraries