Open in new window / Try shogun cloud
--- Log opened Tue Mar 29 00:00:36 2011
-!- alesis-novik [~alesis@] has quit [Quit: I'll be back...]00:27
-!- blackburn [~qdrgsm@] has joined #shogun00:52
-!- blackburn [~qdrgsm@] has quit [Quit: Leaving.]01:02
-!- seviyor [c1e20418@gateway/web/freenode/ip.] has quit [Ping timeout: 252 seconds]01:21
-!- aifargonos [~aifargono@] has left #shogun []02:04
-!- serialhex [] has quit [Remote host closed the connection]02:38
-!- serialhex [] has joined #shogun02:53
-!- shelhamer [] has joined #shogun02:55
-!- sploving [~root@] has left #shogun []02:57
-!- sonney2k [] has quit [Ping timeout: 276 seconds]03:04
-!- sonney2k [] has joined #shogun03:17
-!- serialhex [] has quit [Remote host closed the connection]03:29
-!- ^knrrrd^ [] has joined #shogun05:00
-!- Netsplit *.net <-> *.split quits: cwidmer, yoh, yin, l0uis_, siddharth, sonney2k, meanerelk, @knrrrd, shelhamer05:07
-!- Netsplit over, joins: siddharth, meanerelk, l0uis_, sonney2k05:20
-!- Ryaether [] has joined #shogun05:20
-!- Netsplit over, joins: yoh, shelhamer05:20
-!- Netsplit over, joins: cwidmer, yin05:20
-!- sploving [~root@] has joined #shogun07:16
-!- alesis-novik [~alesis@] has joined #shogun07:30
-!- sploving [~root@] has left #shogun []07:32
-!- eppz [~epps@] has joined #shogun07:39
-!- eppz [~epps@] has quit [Changing host]07:39
-!- eppz [~epps@unaffiliated/epps] has joined #shogun07:39
-!- eppz [~epps@unaffiliated/epps] has quit [Client Quit]07:44
-!- dvevre [b49531e3@gateway/web/freenode/ip.] has joined #shogun07:47
-!- Tanmoy [75d35896@gateway/web/freenode/ip.] has joined #shogun07:48
-!- Tanmoy [75d35896@gateway/web/freenode/ip.] has quit [Client Quit]07:49
-!- sploving [~root@] has joined #shogun09:19
-!- Tanmoy [75d35896@gateway/web/freenode/ip.] has joined #shogun09:29
splovinghello sonney2k09:53
sonney2ksploving, good morning09:54
splovinggood morning. I find that jblas do not support sparse matrix09:55
sonney2ksploving, not many packages do - not even R has support for sparse matrices09:56
splovingokay. then we should not write these typemaps09:57
sonney2ksploving, let me check - I recall that there is one big java matrix package around09:57
-!- aifargonos [~aifargono@] has joined #shogun09:57
sonney2kthat one probably
sonney2kit can indeed do sparse matrices
splovingthis lib is maturate?10:02
splovingI have never heard that10:02
splovingit indeed10:02
sonney2ksploving, I think it is very mature10:02
splovingok. sonney2k, another question, I took at the typemap.i in the python_modular and know what thatis meaning. But I do not understand the two interface file in pyton dir. "pythoninterface.h" "pythoninterface.cpp"10:04
splovingwhat are they for?10:05
splovingare they equal the wrapper file that swig generate?10:05
sonney2ksploving, sorry where are these files?10:05
sonney2k$ ls python_modular/10:06
sonney2ksg_print_functions.cpp  swig_typemaps.i10:06
sonney2kis all I have there10:06
sploving$ls python/10:06
sonney2kyou will only need to modify swig_typemaps.i10:06
sonney2ksploving, ahh I understand. That is the legacy static interface10:07
sonney2klet me explain10:07
sonney2kshogun has two kinds of interfaces: static ones (do not require swig) and modular ones (require swig)10:07
splovingthis i know.10:07
splovingwhat is the difference?10:08
sonney2kthe static ones are in the directories octave, r, python, matlab ...10:08
sonney2kmodular swig based ones in the *_modular10:08
splovingthe static interface files equal the wrapper file that swig generate?10:09
sonney2kthe difference is that for static ones one has to write interfaces / for swig they are automatically created *AND* maybe more important the static ones are not object oriented and have interfaces for only a very limit amount of functions10:09
sonney2kDid I make myself clear?10:11
splovingokay. swig could also generated static wrapper file10:11
sonney2ksploving, well no...10:11
sonney2kswig exports all of shoguns classes directly into python etc10:11
sonney2kso we have e.g. CClassifier -> Classifier in python10:12
sonney2kthe static interfaces have just some 'sg' command that takes a few options e.g. sg('new_classifier', 'LibSVM')10:13
sonney2kso totally different10:13
splovingi took at the examples. and clear now. then I also need write the static java file?10:13
sonney2ksploving, no10:13
sonney2konly java_modular interface10:14
splovingi am glad to hear that. but why not support static?10:15
sonney2ksploving, because functionality is very limited in static anyways10:15
splovingi know~10:15
sonney2kso I expect new users to only ever use a modular interface10:15
splovingthen for jave support, i need write typemap.i for ujmp.10:17
sonney2ksploving, btw, I am just now modifying the configure script etc to enable swig for lua, ruby, java - such that you can get started much easier then10:17
sonney2kswig_typemaps.i yes - I think you should start writing wrappers for jblas first and then maybe later for ujmp ... simply because mikio braun can mentor how this works (he is the author of jblas) and also a mentor for this10:18
splovingsonney2k, i know. thx.10:20
-!- Ryaether [] has left #shogun []10:20
splovingas there are 12 weeks, what do you think better.  java coding ->java test,doc; lua coding->lua test,doc; or java coding, lua coding->java test,doc, lua test doc10:22
splovingi just need write the proposal plan. although they are not difficult to implement10:22
-!- dvevre [b49531e3@gateway/web/freenode/ip.] has quit [Ping timeout: 252 seconds]10:25
sonney2ksploving, I don't think it will take you that much time but I guess it is better to finish one before attempting the other.10:26
sonney2kI guess you will easily be able to also do ruby_modular but I am fine with it not being mentioned in the proposal10:27
-!- aifargonos [~aifargono@] has quit [Ping timeout: 240 seconds]10:28
splovingruby never used:(10:29
splovingoverestimate me :)10:30
sonney2ksploving, doesn't matter ... you know enough languages already to easily read that code10:30
sonney2kthe difficulty lies only in getting the typemaps to work, not the language - see
splovingthen are you familiar with one matrix lib?10:32
splovingif support ruby which lib we use?10:33
sonney2ksploving, the matrix package I think.10:38
-!- jabbok [8d550074@gateway/web/freenode/ip.] has joined #shogun10:38
-!- aifargonos [~aifargono@] has joined #shogun10:39
-!- siddharth [~siddharth@] has quit [Ping timeout: 246 seconds]10:43
-!- ^knrrrd^ is now known as knrrrd10:46
-!- mode/#shogun [+o knrrrd] by ChanServ10:46
splovingsonney2k, for a language, the final file i need write is : swig_typemaps.i sg_print_functions.cpp, and testsuite files. the doc and the examples are not need. am i right?10:46
-!- mode/#shogun [+o sonney2k] by ChanServ10:46
@sonney2kwell some examples at least10:50
-!- knrrrd [] has left #shogun []10:51
@sonney2kgtg, l8r10:53
-!- sonney2k [] has quit [Quit: Leaving]10:53
-!- yin [~chatzilla@] has quit [Quit: ChatZilla [Firefox 3.6.16/20110319135224]]10:55
-!- jabbok_ [8d550074@gateway/web/freenode/ip.] has joined #shogun10:55
-!- jabbok [8d550074@gateway/web/freenode/ip.] has quit [Ping timeout: 252 seconds]10:56
-!- jabbok_ [8d550074@gateway/web/freenode/ip.] has quit [Client Quit]10:56
-!- jabbok [8d550074@gateway/web/freenode/ip.] has joined #shogun10:56
-!- yin [~chatzilla@] has joined #shogun10:56
-!- knrrrd [] has joined #shogun11:20
knrrrdwhen did sonne leave?11:20
-!- mode/#shogun [+o knrrrd] by ChanServ11:21
-!- shelhamer [] has quit [Quit: Computer has gone to sleep]11:42
-!- shelhamer [] has joined #shogun12:06
-!- sonney2k [~sonney2k@] has joined #shogun12:10
sonney2ksploving, hmmhh it is actullay not totally straight forward to even get the basic swig thingies to compile12:17
sonney2ksploving, I will try a few more hours but then I guess you have to take over...12:18
sonney2kthere are things happening like: java package names cannot have the same name like a class in the package12:18
sonney2ksploving, it seems like only lua works out of the box12:20
splovingsonney2k, you mean lua works well12:20
sonney2klua at least compiles12:20
splovingcould we use %rename java?12:21
sonney2ksploving, yes but it is somehow stupid - I mean in all languages it is named differntly? stupid!12:21
sonney2kbut yes, for now I will rename Classifier -> ClassifierBase etc12:22
-!- siddharth [~siddharth@] has joined #shogun12:23
splovingsonney2k, when you develop with python, r , not have these problems?12:26
sonney2ksploving, well with R I had some problems too, but octave & python were easy, but give me a few more moments...12:27
-!- shelhamer [] has quit [Quit: Computer has gone to sleep]12:28
sonney2ksploving, regarding lua12:31
sonney2kwhat how are binary extension named?12:31
sonney2kI mean do the end with ?12:31
sonney2kand how do I load an extension?12:32
splovingrequire ("example")12:32 is the binary12:32
sonney2ksploving, segfault!12:33
sonney2kdammed, this train is crowded with kids calling me a hacker! a hacker *LOL*12:33
sonney2kok so I guess something lua wise is not correct12:33
sonney2kbut at least it compiles ...12:34
splovingsonney2k compiles is a good start.12:35
sonney2ksploving, I will commit later today so you can have a look...12:35
splovingsonney2k, i am writing the proposal now12:35
sonney2ksploving, sure that has priority12:36
splovingsoney2k, okay12:41
-!- aifargonos [~aifargono@] has quit [Ping timeout: 240 seconds]12:43
-!- sonney2k [~sonney2k@] has quit [Ping timeout: 246 seconds]12:44
-!- siddharth [~siddharth@] has quit [Read error: Connection reset by peer]12:45
-!- Tanmoy [75d35896@gateway/web/freenode/ip.] has quit [Ping timeout: 252 seconds]12:50
-!- siddharth [~siddharth@] has joined #shogun12:58
-!- sonney2k [] has joined #shogun13:02
-!- seviyor [c1e20418@gateway/web/freenode/ip.] has joined #shogun13:11
-!- jabbok [8d550074@gateway/web/freenode/ip.] has quit [Ping timeout: 252 seconds]13:18
-!- Tanmoy [75d35896@gateway/web/freenode/ip.] has joined #shogun14:24
yinon the second page of the coffin paper, the last formula I(\Phi(x)_k) should be \Phi(z)??14:35
sonney2kyin, ehh could be :)14:37
yin:) as x should be the mapped vector and z is the original feature vector.14:37
sonney2ktrue then14:39
sonney2kyin, I think I have more papers to be proof red :D14:39
sonney2ksploving, java compiles now too...14:41
splovingsonney2k, congratulation~ thx for your work14:41
Tanmoy@yin...feature map shld be on i/p vectors14:41
sonney2know off to ruby14:42
-!- aifargonos [~aifargono@] has joined #shogun14:43
splovingsonney2k, good luck~14:43
splovingwaiting for good news.14:43
splovingsonney2k, the typemaps.i just map the in/out para, where are the *.i files that include all the functions that pytho,etc calls. for example, in pythin example, svm=LibSVM(..), where is the wrapper file for libsvm function14:49
sonney2ksploving, in the modular/ dir we have all the *.i files that do the svm=LibSVM() magic etc14:51
yinsonney2k, in section 2.1.2 computing \Phi for a variety of kernels. are the three kernels implemented in shogun now?14:52
sonney2kthe typemaps just do the conversion e.g. numpy matrix -> set_matrix(float64_t* matrix, int dim1, int dim2)14:52
splovingthat is the interface file for general language14:52
sonney2kyin, spectrum / weightedegree, polynomial ? yes14:52
yinsonney2k, yes. these three.14:53
splovingsonney2k, i know ant take a look at it14:53
-!- skydiver [4deac315@gateway/web/freenode/ip.] has joined #shogun15:09
yinsonney2k, i am reading CFeatures and its children now. These classes should contain all the data structures that hold data samples. Are there any other class outside this family that are relevant to data samples?15:40
splovingsonney2k, there are mainly four arrays in python typemap: input output in-place argout. but the octave, r modular just have input argout array.15:40
splovingDo r and octave not support output and in-place array?15:41
-!- siddharth [~siddharth@] has quit [Remote host closed the connection]15:42
sonney2ksploving, within shogun I've never used inplace argout15:52
sonney2kyin, no15:52
yinsonney2k, thanks:)15:53
sonney2kyin, these classes can be pretty big /messy15:53
splovingsonney2k, besides the arrays, there are also CTstring feature, What are they for?15:53
yinsonney2k, i am diving into the source now.15:53
yinhope i can figure some structure out and i will add to my "some design patterns in shogun" report.15:54
sonney2ksploving, regarding features we support strings (that is what these are for), sparse matrices and dense matrices15:54
yini think analyzing this part of code is important as we have a project for designing online feature class.15:55
yinsonney2k, i also found that CAttributeFeatures and some others are never used by a classifier or clustering algorithm.15:58
sonney2kyin, I hope it does not turn into a 'clean up the mess' project ...16:00
-!- aifargonos [~aifargono@] has quit [Ping timeout: 250 seconds]16:01
sonney2kyin, yes they are not used. I was only drafting this - but had no time to finish... basically what I wanted is something like in C structs as features16:01
yinsonney2k, i see. btw, there is a "feature name" concept for cdotfeatures?16:04
yine.g. the first dimension of feature vector x means the "length", the 2nd dimension means "age"16:05
yinok. but if we add decision trees, this information is better encoded into its feature class.16:06
-!- aifargonos [~aifargono@] has joined #shogun16:10
sonney2kyin, well not really16:10
sonney2kit is nice to have I agree16:10
yini've sent a email to the mailing list earlier today about the opencv library16:11
yindo you have any comments?16:11
sonney2kYes, I will reply today16:11
sonney2ksploving, OK lua,java,ruby now at least compile with swig interfaces...16:31
splovingsonney2k, good news~16:31
-!- jabbok [8d550074@gateway/web/freenode/ip.] has joined #shogun16:31
sonney2ksploving, I've pushed the changes to github already ...16:32
sonney2kso if you have time try things out...16:32
splovingokay. I will16:32
-!- sonney2k [] has quit [Ping timeout: 246 seconds]17:04
-!- skydiver [4deac315@gateway/web/freenode/ip.] has quit [Quit: Page closed]17:05
-!- aifargonos [~aifargono@] has quit [Ping timeout: 248 seconds]17:16
splovingsonney2k, thx. I submitted the proposal. when you have time please take a look at it.17:29
splovingbye. good night~17:29
-!- aifargonos [~aifargono@] has joined #shogun17:43
-!- fredt [] has joined #shogun17:47
* fredt hrrrrr :)17:49
fredtthanks for the reports :) bye17:53
-!- fredt [] has quit [Quit: thats it :)]17:53
-!- epps [~epps@unaffiliated/epps] has joined #shogun18:55
-!- alesis-novik [~alesis@] has quit [Remote host closed the connection]19:10
-!- alesis-novik [~alesis@] has joined #shogun19:11
-!- epps [~epps@unaffiliated/epps] has quit [Ping timeout: 264 seconds]19:20
-!- epps [~epps@unaffiliated/epps] has joined #shogun19:48
-!- jabbok [8d550074@gateway/web/freenode/ip.] has quit [Ping timeout: 252 seconds]20:13
-!- hiteshk [~hitesh_na@] has joined #shogun20:35
hiteshkhi , are you a mentor for this project20:40
hiteshkor any project20:40
alesis-novikNo, I'm a student20:41
hiteshkOk , is Vojtech Franc available on irc20:42
alesis-novikI don't think so. What idea are you interested in?20:43
hiteshkI am interested in new machine learning algorithms , and you20:44
alesis-novikImplementing EM20:44
hiteshkSo the mentor is Vijtech only , did you talk to him20:45
hiteshkis any mentor available on the irc20:47
alesis-novikI did send him an e-mail today20:53
alesis-novikSo what are you studying?20:56
hiteshkI am a Computer Science undergraduate in India21:07
hiteshkwhat about you21:09
alesis-novikArtificial Intelligence masters student from UK21:11
-!- Tanmoy [75d35896@gateway/web/freenode/ip.] has quit [Ping timeout: 252 seconds]21:11
@knrrrdHi, Guss21:14
alesis-novikHello knrrrd21:15
hiteshkalesis-novik , did he reply to your mail21:15
alesis-novikI only sent it recently, so no. Hopefully tomorrow.21:16
@knrrrdThe official deadline is 8th, right?21:17
hiteshkhave you subscribed to the mailing list21:17
alesis-novikyes knrrrd21:17
alesis-novikNot yet hiteshk21:18
@knrrrdAh ok. I am looking forward to the applications.21:18
alesis-novikIf we submit them earlier, will anyone comment on them before the deadline?21:19
@knrrrdalesis-novik: I am not sure if this would be fair21:20
alesis-novikI meant asking to clarify something or similar comments, but either way is fine21:22
alesis-novikJust so I know that I can refine it until the last day and then submit21:22
@knrrrdAs far as I know, the mentors will have a discussion on all applications after the deadline21:23
-!- siddharth_ [~siddharth@] has joined #shogun21:45
-!- sonney2k [] has joined #shogun21:47
-!- alesis-novik [~alesis@] has quit [Quit: I'll be back]21:56
sonney2k*sigh* my router just died21:56
sonney2k4 hours without internet - hard to believe that I survived this nightmare ;-)21:57
aifargonos...I can imagine...21:58
aifargonostorture for a programmer ...21:58
-!- davidcg [d8a57e6f@gateway/web/freenode/ip.] has joined #shogun22:14
davidcgNot sure if anyone is there, but just wanted to point out that SVMLin is behaving strangely.  It generates decent classifiers in very little time, but sometimes it reverses the labels (e.g. generates a classifier with 20% accuracy instead of 80% accuracy).22:18
sonney2kdavidcg, do you see any pattern in that?22:19
sonney2kI mean like the first label is always -1 in such cases?22:19
davidcgIt happens about 50% of the time, I'd wager it's being assigned randomly.22:19
davidcgHm, not sure if it's dependent on what the first label is.22:19
sonney2kdavidcg, you could check if liblinear behaves in the same way22:21
sonney2kif not then it could be really a bug22:22
davidcgIt may be due to the first label....22:23
davidcgJust checked and if the first label was -1, it would generate the opposite classifier, when I changed only the first label to 1 it produced the correct result.22:23
sonney2kdavidcg, heh ... that code looks suspicous:22:24
sonney2k    float64_t sgn=train_labels[0];22:24
sonney2k    for (int32_t i=0; i<num_feat+1; i++)22:24
sonney2k        Weights.vec[i]*=sgn;22:24
sonney2kperfectly correlates with your observation22:25
sonney2kdavidcg, would you mind commenting lines 100-102 in SVMLin.cpp and see if it fixes the problem?22:25
sonney2ksiddharth_, hi...22:25
siddharth_i would like to work on "Implement Olivier Chapelles fast newton based SVM solver"22:27
davidcgHaha, yeah, that looks like it would cause it.22:27
siddharth_so how should i apply through GSOC?22:27
davidcgBtw, using the python modular interface - don't seem to be able to use LibLinear (get a "Sorry, not yet implemented." when attempting to train).22:28
siddharth_Btw sonney2k are u Soeren Sonnenburg?22:29
davidcgThanks for the help on SVMLin though - I'm sure that changing those lines will fix it.22:31
sonney2kdavidcg, so can you confirm it?22:31
sonney2ksiddharth_, I am about to send out an email with details22:31
siddharth_so should i apply after ur email?22:32
siddharth_There is one more thing22:34
siddharth_i was not able to clearly understand the member function "Calibrate" in SVMSGD22:34
sonney2ksiddharth_, that is Leon Bottou black magic :)22:36
davidcgI'm sorry, don't have the permissions to recompile part of shogun on my server - I've used several other classifiers without seeing this behavior though, and that first line would definitely explain it.22:36
davidcg(float64_t sgn=train_labels[0])22:36
davidcgThanks again!22:37
-!- davidcg [d8a57e6f@gateway/web/freenode/ip.] has quit [Quit: Page closed]22:37
siddharth_ I went through the slides  of Leon Bottou "learning with large datasets" where he has explained SGD...but from the code its difficult to understand22:40
-!- davidcg [d8a57e6f@gateway/web/freenode/ip.] has joined #shogun22:48
davidcgJust wanted to add - probably obvious, but SVMLin inverts the sign for b as well as w in the svm if the first training label is -1.22:50
-!- davidcg [d8a57e6f@gateway/web/freenode/ip.] has quit [Client Quit]22:50
-!- epps [~epps@unaffiliated/epps] has quit [Ping timeout: 250 seconds]22:51
-!- shelhamer [] has joined #shogun22:53
sonney2ksiddharth_, he simply has a lot of SGD experience22:55
sonney2kI think he estimates step sizes for bias and w independently and also has some heuristics how often to update one but not the other22:56
siddharth_yes and he using the sparse features22:58
sonney2ksiddharth_, yes22:59
siddharth_sonney2k, for the primal SVM solver project,we have to convert matlab code to c++,in the overloaded train() function22:59
siddharth_before jumping to that,can u suggest some SVm code to begin with23:00
siddharth_so that I can get a feel of it23:00
sonney2kthe liblinear in shogun maybe?23:02
siddharth_sonney2k, ok i will try them...Also for the proposal do we have to give the timeline or just the project name?23:05
sonney2ksiddharth_, timeline is better. plan conservatively and maybe add additional projects in some later months/outlook... if you accidentally finish early :)23:06
siddharth_sonney2k, thanks...i will keep this in mind...Also i want to ask whether SVMSMO has been implemented in Shogun?23:09
sonney2ksiddharth_, what is SVMSMO? I mean that sounds like libsvm - no?23:10
sonney2ksiddharth_, potentially more interesting is a GPU based svm ...23:10
siddharth_sonney2k, Support Vector Machines Using Sequential Minimal Optimization23:10
sonney2ksiddharth_, but libsvm does that alread...23:10
sonney2kthat would be pretty cool I think ...23:11
siddharth_yeah :)23:14
siddharth_sonney2k, well I am currently going into CUDA thing for my 3D texture modeling project23:16
sonney2ksiddharth_, well then the plan is clear :D23:17
-!- epps [~epps@unaffiliated/epps] has joined #shogun23:17
-!- hiteshk [~hitesh_na@] has left #shogun []23:26
-!- jabbok [8d550074@gateway/web/freenode/ip.] has joined #shogun23:42
--- Log closed Wed Mar 30 00:00:36 2011