Open in new window / Try shogun cloud
--- Log opened Wed Jan 18 00:00:19 2012
blackburnsonney2k_: w/o page on
blackburnevery toolkit has one..01:30
-!- in3xes [~in3xes@] has joined #shogun05:37
-!- in3xes [~in3xes@] has quit [Ping timeout: 252 seconds]06:10
-!- blackburn [~qdrgsm@] has quit [Ping timeout: 272 seconds]07:45
sonney2k_blackburn *with* one08:20
CIA-27shogun: Soeren Sonnenburg master * rbaa7f11 / src/interfaces/modular/SGBase.i : fix hdf5 serialization check mismatch -
shogun-buildbotbuild #272 of python_modular is complete: Failure [failed test_1]  Build details are at  blamelist: sonne@debian.org09:29
-!- karlnapf [~heiko@] has joined #shogun11:57
-!- karlnapf [~heiko@] has quit [Ping timeout: 252 seconds]12:02
-!- blackburn [~qdrgsm@] has joined #shogun13:18
blackburnsonne|work: I didn't get you :( how can I have mloss page with shogun code but with 'EDRT' label?13:20
sonne|workyou do EDRT label and point to your shogun subpage13:21
blackburnsonne|work: but what about source?13:22
blackburnand binaries13:22
blackburnor so13:22
sonne|workpoint to shogun13:23
blackburnsonne|work: ok thanks, I thought it would be strange..13:24
-!- blackburn [~qdrgsm@] has quit [Quit: Leaving.]13:53
-!- blackburn [~qdrgsm@] has joined #shogun14:05
-!- karlnapf [~heiko@] has joined #shogun15:14
-!- karlnapf [~heiko@] has quit [Ping timeout: 245 seconds]15:50
-!- sonne|work [~sonnenbu@] has quit [Read error: Connection reset by peer]16:10
-!- sonne|work [~sonnenbu@] has joined #shogun16:47
CIA-27shogun: Soeren Sonnenburg master * r299493d / src/interfaces/modular/SGBase.i :17:18
CIA-27shogun: fallback to original pickle/serialization routines if objects are not17:18
CIA-27shogun: instances of SGObject -
blackburnsonney2k_: I just got shogun serialized object with size of 7gb :D17:42
-!- Ram108 [~amma@] has joined #shogun18:07
blackburnRam108: hey how are you?18:15
Ram108hey am fine18:15
Ram108working on neural networks nw :)18:15
Ram108hw r u? :)18:15
blackburnwhat do you do with NN?18:16
Ram108classify....... working on backpropagation algorithm18:16
Ram108fairly easy but slow18:17
blackburnI see18:17
Ram108hmmm :) i got a faint idea on hw real brain works nw18:17
Ram108lol gona use that on my robot18:17
Ram108built 118:18
Ram108 lol :P18:18
blackburnRam108: well that's exactly *not* how the brain works :)18:19
blackburntoo simplified..18:19
shogun-buildbotbuild #273 of python_modular is complete: Success [build successful]  Build details are at
Ram108oh thats true :) this is a very crude representation18:20
blackburnNN is very useful in OCR and visual pattern recognition though18:20
blackburnRam108: sorry I can't remember if you did something with MNIST dataset?18:21
Ram108no that was the other guy18:21
blackburnheh okay18:21
blackburnthen you could check MNIST page18:21
Ram108sure :) looking forward to it :)18:22
blackburnbest results are computed with convolution neural nets commitee18:22
Ram108god this is huge!18:22
blackburnwith error of 0.2718:22
blackburnjust check results to compare methods18:22
Ram108hmmm i ll try it out :) definitely with shoguns svm methods and the code which i am working on currently lol :P18:23
blackburnbut in other tasks NNs could act worse..18:24
Ram108thanks a lot again........ well initially i was overwhelmed but now i seem to have a quaint idea of hw it works :)18:24
Ram108nn is slow and more often than necessary the code keeps running out of memory18:24
Ram108thats 1 thing i have to fix18:25
Ram108bye the way do u mind u could tell me abt urslf?18:25
blackburnwhat exactly you want to know? :)18:25
blackburnI guess we are of same age, I'm a student of Samara State Aerospace University in Samara, Russia18:26
Ram108oh hmmm :)18:26
Ram108nice :)18:27
blackburnI'm also a software engineer @ netcracker18:27
Ram108aerospace univ?18:27
Ram108cool :)18:27
blackburnlast year I was a GSoC student, I did impelement 12 dimensionality reduction methods18:27
blackburnand am still interested in nonlinear dimensionality reduction :)18:28
Ram108yeah u told me that........ any chance i could try getting in? just asked18:28
blackburnas GSoC student?18:28
blackburnwell if we get accepted18:29
blackburnand if GSoC 2012 is announced :)18:29
blackburnyou definitely should try to some organization18:29
Ram108well mostly shogun ll be right?18:29
blackburnwhat do you mean?18:30
Ram108as in ML is one of the frontier fields18:30
Ram108so why wuldnt it?18:30
blackburnI meant you can choose one organization you would like to work with18:31
Ram108i guess am no1 to comment upon this.......18:31
Ram108well i really want to learn this out18:31
Ram108the thing is i was into robotics for so long..... and i got saturated with it. and i want to impart intelligence to most of the machines i am building18:32
Ram108lol i feel GOD wen i do that lol :P18:33
blackburnwe would have to choose some students, I can't promise anything but we would prefer students that know how to work with shogun18:33
Ram108hmmm :) i see18:33
blackburnI guess that's one of the reasons you started to learn shogun :)18:34
Ram108well u mean equipping my gadgets with intelligence?18:34
blackburnokay I have to study optics a little heh18:35
blackburnwill have exam on saturday18:36
Ram108oh thats a drag.... go on :)18:36
blackburnno idea why I have to learn this18:36
Ram108dads a physicist i kno why18:37
Ram108lol :P18:37
blackburnbtw I don't know anything about 'aerospace' :D18:37
Ram108yeah i know i feel out of the soup as well........ took the wrong branch......... found myself liking robotics more than communication engineering18:38
Ram108in the soup*18:39
Ram108actually i was gona type fish out of water18:39
blackburnwell I'm just major in mathematics and nothing with neither aero nor space ;)18:39
blackburnI've never been studying anything related18:40
Ram108math..... thats a pain......... never liked math for itself... fells like it just loses touch with reality18:40
blackburncan't agree, there are a lot of branches18:41
Ram108i did make a plane wen i was in my 6th grade though....... i have it with me wana see?18:41
blackburnwell even machine learning is math 99% :)18:41
Ram108hate tensors......... and statistics18:42
Ram108yeah agreed but i can really see the touch with the real world18:42
blackburnoh tensors :) I guess you had a lot of mechanics18:42
Ram108yeah i did......... fluid mechanics, Bernoullis theorems etc etc18:43
Ram108was way over my head18:43
Ram108just was not cut out for that kind of thing18:43
Ram108pretty cool like u guys :)18:44
blackburnsure, orange has good gui and is easier18:46
Ram108yeah i saw that18:47
blackburnbut we have a lot of features they has not18:47
Ram108thats y i chose shogun over that18:47
blackburnscikits-learn has a lot of nice features18:47
blackburnI like it pretty much18:47
Ram108lol saw that too :P18:48
Ram108was just wondering18:52
Ram108why cant the algorithms be classified18:52
Ram108based on parameters like datatypes time required and processing requirements18:53
blackburnwhich algos?18:54
Ram108then probably the end user wont have to kno which algorithm to use...... all he has to specify is the parameters and their limits......18:54
Ram108could act like an AI engine18:54
blackburnwell nobody knows how good some algorithm will work with some data18:55
Ram108algorithms like SVM NN etc18:55
Ram108oh erm is there no way of predicting that? just curious18:55
blackburnwell experts do know a little more, but generally you would have to test whether it works well or not18:56
Ram108hmmm so the possibility of automating that is slim right?18:57
Ram108okay :) thanks for ur time :)18:57
blackburnthere are some estimates of probabilities but they are too much general18:57
blackburnit is being researched in COLT18:57
blackburnComputation Learning Theory18:58
Ram108oh hmmm18:58
Ram108oh i see :) i ll google that up18:58
blackburnI don't know much about COLT, too complex for practice18:58
blackburnanyway choosing right methods is not the way how AI should work19:00
Ram108hw else?19:01
blackburnshould be unified..19:01
blackburnsomething of interest is a deep learning19:02
Ram108well cant we use unsupervised learning methods to classify the given data?19:02
Ram108thats an idea.....19:02
blackburnunsupervised? well we could but results would be worse19:03
Ram108oh hmmm......... i have to try that someday19:04
blackburnwell some tasks are supervised by nature19:04
blackburnif you want to recognize road signs you know classes, etc19:05
blackburnno need to do that unsupervised19:05
Ram108well for now am working on unmanned flights and vehicles19:06
blackburnunfortunately you have not many choices19:07
Ram108as in?19:07
blackburnI mean it would be strange to use KNN or Ridge Regression to ride a vehicle :)19:07
Ram108could u be a little more specific pls?19:07
Ram108hmmm lol well gona use NN19:07
blackburnusually robots are powered by NN19:08
Ram108yeah lol :P19:08
Ram108read abt em19:08
blackburncause nature of riding a vehicle is not really formalized :)19:08
Ram108thanks :)19:08
Ram108i ll post a youtube video of it soon........ lol but i have to go through a lot of logistics and paper filling to procure the essential funding for that19:09
blackburnI see19:09
Ram108i ll get started on the hardware soon :)19:10
Ram108am pretty good with digital circuits19:10
Ram108:) guess that wont be a problem19:10
blackburnhah I don't give a fuck about it19:10
blackburnU = IR that's the only thing I know19:10
Ram108hmmm well microcontrollers are not that bad19:11
Ram108besides i firmly believe innovation lies in the merging point between fields.... so lol determined to learn this out19:12
Ram108lol :)19:12
Ram108guess i gotta crash for the night.......... really tired.......19:13
Ram108thanks for ur time :)19:13
blackburnah you have different timezone :)19:13
blackburnwhat time is it?19:13
Ram108yeah apparantly i do lol :P19:13
Ram108its just 11 45 but i woke up at 3 am this morn19:14
Ram108and i had a badminton match which lasted for 3 hrs.......19:15
blackburnwow you +1.30 from me?19:15
blackburnyeah I guess so19:15
blackburnUTC + 5.30?19:15
Ram108lol yeah :)19:15
Ram108hmmm the time there is 7 15 pm right?19:16
blackburnno, I'm in UTC+419:16
blackburnMoscow timezone19:16
Ram108i can barely open my eyes........ lol :P19:16
blackburnthen go sleep :)19:16
Ram108good night :)19:17
blackburngood night19:17
blackburnsee you19:17
Ram108u too thanks for ur time :)19:17
Ram108bye :)19:17
-!- Ram108 [~amma@] has quit [Quit: Ex-Chat]19:17
blackburnsonney2k_: can one-versus-all ROC be considered as good performance measure?20:20
CIA-27shogun: Soeren Sonnenburg master * rc9e6d08 / src/interfaces/modular/SGBase.i : pickle_ascii should be 1 if pickle version is 0 -
blackburnsonney2k_: did you receive my mail? :)21:54
sonney2k_blackburn, yes22:22
-!- sonney2k_ is now known as sonney2k22:22
blackburnany thought?22:22
-!- mode/#shogun [+o sonney2k] by ChanServ22:23
@sonney2kI found some potential mentors... but was/am busy with other stuff22:24
blackburnah I see22:24
@sonney2kbut yes we need to come up with some more polished list22:24
blackburnsonney2k: I'm unavailable for thinking till 21st :)22:25
blackburnbut I hope to come up with something to the end of January22:25
@sonney2kWell then after 21st. it looks like I won't progress before that either...22:25
@sonney2kat least the buildbot is happy again :)22:26
@sonney2kblackburn, what did you serialize that is *that big*22:26
@sonney2kwas that with pickle?22:26
blackburnsonney2k: it looks like you are going to be busier and busier22:26
blackburnsonney2k: yes, pickle22:27
blackburnsonney2k: LibSVM with 39202 train vectors and ~12k test ones of dim 8122:27
@sonney2kwell kids that don't sleep kill any shogun time...22:27
@sonney2kyeah but that should just be 30M22:28
@sonney2khow many classes?22:28
blackburn43, OVO22:28
@sonney2kI suspect features are saved multiple times22:29
blackburnI guess so22:29
blackburnsonney2k: but libsvm do not support OvA :)22:29
@sonney2ksvmlight does :)22:29
blackburnsonney2k: does svmlight support multiclass?22:29
@sonney2kahh no you need to do it manually of course22:30
blackburnoh no I don't like it22:30
@sonney2kbut yes all that multiclass stuff would be a great gsoc project22:30
blackburnsonney2k: do you think it would be cool to have MulticlassMachine (ECOC,OvA,OvO)?22:30
blackburnbased on linear machines22:30
@sonney2ksuch that any classifier can be turned into *22:30
blackburnI would say it is not a GSoC project22:31
blackburnbut week of work..22:31
@sonney2kthere is this advanced stuff from nips last year22:31
blackburni'm on vacation starting from 25, Jan to 5, Feb22:31
@sonney2kthat would certainly make it so22:31
blackburnwill try to start22:31
@sonney2kwould be cool22:32
blackburnI need it just now, so I would have to implement22:32
@sonney2kbtw, we don't even have some linear regression methods in shogun22:32
blackburnat least simplest concept22:32
blackburnsonney2k: we have bad feature list now22:32
@sonney2kno LS-regression22:32
blackburnI'm getting upset looking on scikit-learn methods22:32
blackburnand comparing with us22:32
@sonney2k(which I need now)22:32
@sonney2kthey are much broader sure22:33
blackburnthey have one billion authors22:33
blackburnauthors list in jmlr paper do not fit in one page!22:33
blackburnand two billion contributors :)22:33
@sonney2kyeah but we had the same problem w shogun22:34
blackburnsonney2k: aha but where are these guys? :)22:34
@sonney2ktrue all not really active right now. I think shogun does what they needed now and so they are just users22:35
@sonney2ktime to relax a bit22:36
blackburnsonney2k: sleep? :)22:36
-!- karlnapf [] has joined #shogun23:01
-!- karlnapf [] has quit [Ping timeout: 240 seconds]23:16
-!- karlnapf [] has joined #shogun23:28
blackburnsonney2k: that's how my time measure script does output :)23:53
blackburnsee you23:54
-!- blackburn [~qdrgsm@] has quit [Quit: Leaving.]23:57
--- Log closed Thu Jan 19 00:00:19 2012