IRC logs of #shogun for Thursday, 2016-04-28

« Wednesday, 2016-04-27 Index Friday, 2016-04-29 »
--- Log opened Thu Apr 28 00:00:32 2016
-!- besser82 [~besser82@fedora/besser82] has quit [Ping timeout: 252 seconds]00:24
-!- WiggleRoom_ [49bd0ad0@gateway/web/freenode/ip.73.189.10.208] has joined #shogun01:40
WiggleRoom_Hey I have a machine learning question!01:41
-!- WiggleRoom__ [49bd0ad0@gateway/web/freenode/ip.73.189.10.208] has joined #shogun01:42
WiggleRoom__Is anyone there01:42
WiggleRoom__My model is huuuuge like 100GB AND GROWING! How can I store and use it? It won't even load into memory!01:42
-!- WiggleRoom_ [49bd0ad0@gateway/web/freenode/ip.73.189.10.208] has quit [Ping timeout: 250 seconds]01:46
-!- WiggleRoom__ [49bd0ad0@gateway/web/freenode/ip.73.189.10.208] has quit [Quit: Page closed]01:50
-!- lupinix [~quassel@fedora/lupinix] has quit [Ping timeout: 240 seconds]01:50
-!- lupinix [~quassel@v22014041761818086.yourvserver.net] has joined #shogun01:51
-!- lupinix [~quassel@v22014041761818086.yourvserver.net] has quit [Changing host]01:51
-!- lupinix [~quassel@fedora/lupinix] has joined #shogun01:51
-!- mizari [~mizari@95-174-213-100.nts.su] has joined #shogun05:13
-!- arianepaola [~ariane@unaffiliated/arianepaola] has quit [Excess Flood]06:21
-!- arianepaola [~ariane@unaffiliated/arianepaola] has joined #shogun06:22
-!- Saurabh7 [~Saurabh7@103.225.100.51] has quit [Quit: Leaving]06:38
-!- besser82 [~besser82@fedora/besser82] has joined #shogun08:03
-!- mode/#shogun [+o besser82] by ChanServ08:03
@wikinglisitsyn: pingpong09:34
sonne|workWiggleRoom__ wrong timezone :P09:38
@wiking:)09:39
@wikingsonne|work: did you guys ever wondered to support mixed types in Features? :)09:39
sonne|workwiking: shogun does that pretty well already...09:40
sonne|workand I use this quite often09:40
@wikingwhere?09:40
@wikingwhich Features?09:40
sonne|workCombinedFeatures09:40
sonne|workstrings09:40
sonne|workrealfeatures09:40
@wikingdoh09:40
sonne|worketc09:40
@wikingso you put combinedFeatures everything09:41
@wiking?09:41
@wikingi've used this only with combinedKernel09:41
sonne|workwe even have combined dotfeatures09:41
sonne|workso you can train e.g. a linear svm with all kinds of DotFeatures attached to it09:41
sonne|worksure09:41
@wikingnever thought of actually09:41
sonne|worksome methods required RealFeatures though09:42
@wikingusing that as combination of string and real fatures09:42
@wikingyeah09:42
sonne|workthat is one of the strengths of shogun09:42
@wikingthat's ok09:42
sonne|workbut it is not just strings09:42
sonne|workit is boolean/int/float/double ... you can mix that at will09:42
@wikingyeah yeah09:42
@wikingbut yeah we should ahve an example for this09:43
@wikingbecause i tooootally had no clue09:43
@wikingthat actually the combinedfeatures09:43
@wikingcould be used like this09:43
@wikingmaybe that's only me09:43
@wikingbut then again09:43
@wikingmmm09:44
@wikingthe only thing would be good to have i guess09:44
@wikingand IO method09:44
@wikingthat actually can create a CombinedFeatures09:44
@wikingfrom a csv09:44
@wikingdetecting the types09:44
@wikingright?09:44
@wikingsince basically we can say that CombinedFeatures = DataFrame09:44
@wikingbut now you would have to manually do the loading of a csv with various types of features09:45
@wikinginto a CombinedFeatures09:45
@wikingOXPHOS could work on this09:46
@wikingdear god, all aws emails are ending up on gmail in spam09:47
@wikingg really does not want people to use aws :)09:47
-!- besser82 [~besser82@fedora/besser82] has quit [Ping timeout: 276 seconds]10:16
-!- wiking [~wiking@huwico/staff/wiking] has quit [Remote host closed the connection]10:17
-!- wiking [~wiking@huwico/staff/wiking] has joined #shogun10:36
-!- mode/#shogun [+o wiking] by ChanServ10:37
@wikinghaha11:06
@wikingjust realised our welcome string for #shogun: "[#shogun] Welcome to #shogun - The official channel of the shogun machine learning toolbox - http://www.shogun-toolbox.org/. We participate in GSoC 2014 http://bit.ly/16LaK5Z. This channel is logged and logs are available from http://bit.ly/eQDk8Q.11:07
@wikingnote: "We participate in GSoC 2014"11:07
@wiking:)11:07
lisitsynlol11:07
lisitsynwiking: change it, you're an op11:07
lisitsynI've lost my registration actually11:07
@wikingwhere? :)11:08
@wikingi mean where did you lost it11:08
@wiking?11:08
lisitsynnickserv11:08
@wikingok done11:09
@wikingif you check the link of We participate in GSoC 201411:09
@wikingit is actually pointing to a 2013 page11:09
@wiking:)11:09
@wikingso it's really good :11:10
@wiking:>11:10
lisitsyn2013!11:10
lisitsynрфрф11:10
lisitsynhaha11:10
lisitsynoops11:10
@wikinganyhow now it's updated11:10
lisitsynit is hard to live with multiple keyboard layouts :D11:10
@wikingdarn cyrillic :)11:10
lisitsynwiking: there was an initiative11:11
lisitsynto use latin letters11:11
@wikingit's easy11:11
@wikingthere's a 1-1 mapping11:11
lisitsynexactly11:11
@wiking(at least we have in serbian)11:11
@wikingwhen i told this to an american11:11
lisitsynyeah there are hard cases though11:11
@wikinghe tought i'm crazy11:11
@wiking:)11:11
@wikingbut then again11:11
@wiking:)))))11:11
lisitsyniirc you can get some ambiguity with sh sch11:12
@wikingliving with 26 characters in your alphabet is quite weak :P11:12
@wikingwe have 44 in hugnarian11:13
@wiking33 in russian right?11:13
lisitsynyes11:13
lisitsyn3311:13
lisitsynmost of them are not that useful though11:13
lisitsynь/ъ11:14
@wikingso yeah it's easy :)11:14
lisitsynlegacy stuff11:14
lisitsynе/ё11:14
@wikingif you have a normal alphabet to cover all of them11:14
@wiking:P11:14
lisitsynwiking: american alphabet must be of like ten letters actually11:14
lisitsynthey like to simplify things :P11:14
@wiking:)11:15
@wikingyeah all the vowels are out11:15
@wiking:>11:15
@wikingflckr11:15
@wikingrzr11:15
@wiking:)11:15
lisitsyntwttr11:15
@wikingyeap pretty much11:15
lisitsynggl11:15
-!- HeikoS [~heiko@p57972783.dip0.t-ipconnect.de] has joined #shogun11:15
-!- mode/#shogun [+o HeikoS] by ChanServ11:15
lisitsynhks11:15
@wiking:DDDD11:16
lisitsynh hks11:16
lisitsynhk!11:16
lisitsynhll11:16
lisitsynH!11:16
lisitsynv lst m vwls11:16
lisitsyndmn11:17
@wikingwll y knw n hbrw thrs n vwls jst dts11:17
@wiking:)11:17
lisitsynxctl11:17
lisitsynHeikoS: H!11:18
-!- lisitsyn is now known as lstsn11:19
lstsnHkS11:19
-!- HeikoS [~heiko@p57972783.dip0.t-ipconnect.de] has quit [Ping timeout: 260 seconds]11:21
lstsndmn11:22
-!- HeikoS [~heiko@p57972783.dip0.t-ipconnect.de] has joined #shogun11:26
-!- mode/#shogun [+o HeikoS] by ChanServ11:26
@wiking:>11:27
@HeikoSwiking: have an appointment to sign things next Tuesday11:30
@HeikoSwiking: they will check all the docs now, I brought them in yesterday11:30
@HeikoSsonney sent me the original signatures per post-mail as they are needed11:30
@HeikoSSo looking good11:30
lstsncl11:40
lstsnHkS11:45
@HeikoSlstsn: ?12:02
lstsnHeikoS: v lst m vwls12:38
CaBahi13:02
CaBacan one adjust the number of threads used by libsvm somehow?13:03
@HeikoSCaBa: I dont think so13:03
@HeikoSCaBa: if you precompute the kernel matrix, then multiple threads are used13:04
@HeikoSlstsn: !13:04
lstsnHeikoS: wazzzzup13:04
@HeikoSlstsn: !!!13:04
lstsnwzzzzzuppp13:04
CaBaHeikoS: i'm wondering about finetuning "multiple" ;) it seems to use the number of physical CPUs rather than logical CPUs to decide how many threads to create... I wonder if hyperthreading would boost this further...13:05
@HeikoSCaBa: Shogun always uses the number of cores available to the OS13:06
@HeikoShyperthreading just gives you 4 cores from 2 physical cores13:06
@HeikoSbut the OS always "sees" 413:06
CaBaHeikoS: hm okay then HT simply doesn't kick in that well in case of kernel training..13:40
@HeikoSCaBa: I think LibSVM is simply not parallelised code13:40
@HeikoSOnly if you pre-compute the kernel matrix, this is parallelised13:41
@HeikoSThis gets important if you cross-validate13:41
CaBaHeikoS: well i have a 12-core machine with HT and during train() the process uses ~1200% cpu ressources... so train() seems parallelized...13:42
@HeikoSoh ok, I see13:42
@HeikoSmaybe someone added that :D13:42
@HeikoStry sg_parallel13:42
CaBa;)13:42
@HeikoSyou can change the number of cores used within there13:43
@HeikoSany_sg_object_instance->get_global_parallel()->set_num_threads(n)13:44
CaBaHeikoS: it really seems that HT doesn't scale well here. Calling that function works, I can e.g. limit to 4 threads or so, but anything above 12 won't increase the load much13:49
@HeikoSThat makes sense I think13:49
@HeikoSsolving SVM is a sequential problem13:49
@HeikoSLibSVM is not designed formultiple cores13:50
@HeikoSI think there are SVM solvers that can exploit more13:50
CaBaHeikoS: What are the criteria for choosing the backend anyway? Should I be using another more efficient packend?13:50
CaBa*backend13:50
@HeikoSLibSVM is a standard library13:50
@HeikoSnot the fastest13:50
@HeikoSit depends on your problem13:50
@HeikoSalso, if you need to tune parameters, this is where most of the cpu cycles go (x-validation, grid search, etc)13:50
@HeikoSso if you parallelise that, you get almost linear gains13:50
@HeikoSin the number of threads, as completely independent13:51
CaBaindeed, maybe i shouldn't be optimizing here... right now i'm not doing any model selection, just playing with a single model13:51
CaBabut since that won't stay that way i can just rely on the model selection parallelization later...13:52
CaBaHeikoS: is there like an overview of the different backends and how to choose it? like how does it depend on my problem?13:53
@HeikoSI am sure there is13:53
@HeikoSnot so much in Shoguns doc though13:54
@HeikoSDepends a bit on your problem13:54
@HeikoShow much N, D, non-linearity you have?13:54
CaBaright now, i'm working with ~80 numerical (real) features on ~54k observations... i might however also want to use or adopt some of the string kernels in shogun soon13:54
CaBaright now i just played with a simple, plain linear kernel. however, nobody tells me that this data is linearly separable, hence i'll also try other kernels like poly or rbf...13:56
CaBai take it there is no general doitall workhorse library backend? i really didn't "choose" libsvm, i just saw it in one of the examples and plugged it in ;)13:57
@HeikoSI see14:03
@HeikoSso RBF kernel seems like a good idea14:03
@HeikoShowever, if you go for a linear kernel, then LibLinear ist *much* faster14:03
@HeikoSthere is a multi-core version of it available, but we haven't integrated it yet unfortunately14:03
@HeikoSsonney2k_, sonne|work might have comments14:04
CaBaHm. LibLinear... doesn't show up in the SVM class inheritance diagram ^^14:04
@HeikoShttp://shogun-toolbox.org/doc/en/latest/classshogun_1_1CLibLinear.html14:18
@HeikoShttp://shogun-toolbox.org/doc/en/latest/classshogun_1_1CLinearMachine.html14:18
@HeikoSSVMLight should also be faster than libsvm14:19
@HeikoSand this guy is also fast I think: SVMOCAS 14:19
CaBaI'll play around and also check how easily interchangable they are, thanks :)14:21
sonne|workCaBa: well try with a small subsample of your data and the rbf kernel first then once you know which parameter range (kernel width, C) seem to do well use a bigger set14:41
@wikingmooo14:53
sonne|workwiking: holy cow?15:06
sonne|workwiking:  apt-get moo ?15:06
CaBasonne|work: will do :)15:17
CaBa(subsample that is... not that much into mooing, though i totally appreciate super cow powers...)15:19
@wikingyes15:28
@wikingsonne|work: hoooooooly holly cow15:28
@wikingso what do you say about an IO func for CombinedFeatures?15:29
sonne|workwiking: sure do it... but no idea how15:30
CaBareplacing LibSVM by SVMLight in my python script gets me segfaults \o/ :)16:10
-!- besser82 [~besser82@fedora/besser82] has joined #shogun16:17
-!- mode/#shogun [+o besser82] by ChanServ16:17
@wiking:)16:51
-!- OXPHOS [8ca3fe9e@gateway/web/freenode/ip.140.163.254.158] has joined #shogun16:53
OXPHOSHey @ Heiko, wiking, I'll be analyzing data basically the whole day. We can discuss about the project anytime as long as my PI doesn't jump out.17:02
-!- besser82 [~besser82@fedora/besser82] has quit [Ping timeout: 246 seconds]17:08
@HeikoSOXPHOS: jojo17:45
OXPHOSHeikoS: hello17:46
@HeikoSOXPHOS: so I think first step is to share a googledoc with your proposal17:47
@HeikoSit should have two parts17:47
@HeikoS1. schedule of things we want to do and how long they take. This is "what and when"17:47
@HeikoS2. A detailed list of issues that we have to solve. Dividing tasks into subtasks, figuting out details etc. This is "how exactly things are done"17:47
@HeikoScould you start a doc (doesnt have to be perfect for now, proposal is ok) and share it?17:48
OXPHOSSure. Thought the current one is already shared doc. Be back in 1 sec17:49
OXPHOSIt has to be .word right?17:49
OXPHOS.doc17:49
@HeikoSmaybe we should create a new doc17:49
@HeikoSand paste your proposal in, we dont want to change the proposal doc17:49
@HeikoSgoogledoc .doc yes17:49
OXPHOSYep make sense17:49
@HeikoSOXPHOS: cool17:50
@HeikoSOXPHOS: one thing we will definitely need to do is to prototype a serialization using ceral. This will be done best in a separate environment, i.e. not in the shogun source tree but in a standalone file to isolate things as good as possible17:51
@HeikoSone other thing to to discuss a few linalg internal changes with wiking and lambday17:51
OXPHOSCopy that17:52
OXPHOSI sent you the docs link @ HeikoS17:53
@HeikoSyep saw it17:53
@HeikoSI suggest we keep going slowly but steadily17:53
@HeikoSdont need to make a lot of progress every day, but we should make constant progress, even if slow and we are busy17:54
OXPHOSAgree17:54
OXPHOSAnd also the readme files? XD17:54
@HeikoSreadme files?17:55
OXPHOSThere's a readme.linalg, with several 'coming soon'17:55
@HeikoSah yes17:56
@HeikoSdefinitely should look at those17:56
@HeikoSafter GSOC, it would be good to have some developer docs that are complete on this topic :)17:56
OXPHOSYes I guess things are going to change a lot17:56
-!- mizari [~mizari@95-174-213-100.nts.su] has quit [Quit: Leaving]17:57
@HeikoSdefinitely17:57
@HeikoSoh and the final big thing is the smart pointers17:57
@HeikoSthis also needs some prototype that applies the principle to an isolated part of shogun, i.e. one single class17:58
@HeikoSre io, we need to talk to wiking17:58
@HeikoSneed to use proper modern c++ for this, the code is horribly old ;)17:58
OXPHOSwiking once said he's almost done with the I/O17:58
OXPHOSStupid question - Is it possible I use the same class/environment/whatever for serialization/smart pointer test?17:59
@HeikoStest in what sense?18:00
OXPHOSI mean set up the prototype18:03
@HeikoSah i see18:04
@HeikoSI think it is necessary in fact18:04
@HeikoScereal is based on smart pointers18:04
@HeikoScannot deal with raw pointers18:04
@HeikoSso in fact these two depend on each other18:04
OXPHOSWow. I like that18:05
OXPHOSfor no reason18:05
@HeikoShehe18:06
@HeikoSthat is good18:06
@HeikoSmakes so well suited for that18:06
@HeikoSsince it will be a mess :D18:06
@HeikoSbut it IS quite cool 8-)18:07
@HeikoSgetting rid of all this 90s code ;)18:07
OXPHOSEVOLUTION!18:08
@HeikoSexactly18:08
@HeikoSbut some shogun classes will go extinct hehe18:08
OXPHOSlike human in some day. meh.18:09
@HeikoSbetter not make shogun too powerful then18:09
@HeikoSto buy time18:09
OXPHOShaha18:10
-!- sanuj [0e8bc402@gateway/web/freenode/ip.14.139.196.2] has joined #shogun18:50
-!- sanuj [0e8bc402@gateway/web/freenode/ip.14.139.196.2] has quit [Ping timeout: 250 seconds]18:55
-!- sanuj [0e8bc402@gateway/web/freenode/ip.14.139.196.2] has joined #shogun18:57
sanujhello everyone!19:04
sanujHeikoS: are you making shogun accept donations in germany?19:04
sanujjust curious :)19:04
@wiking:>19:11
-!- besser82 [~besser82@fedora/besser82] has joined #shogun19:14
-!- mode/#shogun [+o besser82] by ChanServ19:14
@wikingOXPHOS: sorry was away19:15
sanujlstsn: what happened to your name? :P19:41
CaBathe latest release is 4.1.0, right? But there is no online doc?19:57
@wikinghe removed all the unnecessary vowels :)19:57
@wikingCaBa: there was a bug in that release19:58
@wikingyou can use the docs of 4.2.0 (development)19:58
@wikingit shouldn't be too different19:58
CaBawiking: i just started to play with model selection and the 4.2 doc says that the GaussianKernel has a "log_width" param, which it doesn't in 4.1.0 ;-)20:00
lstsnsanuj: I've lost vowels20:05
sanujohkay20:05
@wikingCaBa: why dont u use develop branch from the git repo?20:05
CaBawiking: i prefer letting only homebrew touch my /usr/local and installing it anywhere else would require me to fiddle with DYLD_LIBRARY_PATH etc... but yes, i'm tempted to switch to the latest and greatest shogun at some point, maybe i'll look into altering the homebrew formula to checkout from upstream git instead of fetching a stable tarball...20:11
CaBawiking: you don't have nightly tarballs or such thing, right?20:11
@wiking:>20:13
CaBamaybe i can make use of githubs tarballs...20:13
@wikingyou can use github20:13
@wikingand you can install it to any prefix20:13
@wikingbtw can you tell what's the use case? :)20:14
@wikingbecause actually we dont know what shogun is being used for :)20:14
@wiking:P20:14
CaBayeah i know it's a standard cmake or autoconf project, i remember compiling it from source a while ago. it's just trouble to use shared libaries that aren't in the standard locations the system uses to look for them...20:16
@wikingyeah indeed20:17
-!- sanuj [0e8bc402@gateway/web/freenode/ip.14.139.196.2] has quit [Quit: Page closed]20:17
CaBasure :) i'm a bioinformatician, my features are usually DNA sequences (originally) and my field of research is immunogenetics, not far from what gunnar and sören have applied ML for already20:18
CaBawhat's your background?20:19
@wikingah i see20:19
@wikingi did computer vision20:19
@wikingunfortunately not for a while now20:20
CaBaand now?20:20
@wikingwent to industry20:21
CaBastill involved in ML or are you spending your free time with shogun?20:22
@wikingfree time shogun20:22
@wikingbut yeah more or less still involved with ml20:22
@wikingalthough nothing with shogun directly20:23
CaBawhere are you from?20:23
OXPHOSwiking: me too. Just came back from lunch : )20:29
@wikingCaBa: huh that's a complicated answer20:29
@wikingOXPHOS: cool... yeah we should do a quick demo with cereal :)20:30
OXPHOSwiking: Sure. Lemme read more about it.20:32
CaBawiking: :D sounds interesting20:33
@wikinghttp://motherfuckingwebsite.com/20:57
CaBa^^21:17
-!- HeikoS1 [~heiko@p4FED52C3.dip0.t-ipconnect.de] has joined #shogun21:35
-!- HeikoS [~heiko@p57972783.dip0.t-ipconnect.de] has quit [Ping timeout: 250 seconds]21:36
-!- OXPHOS [8ca3fe9e@gateway/web/freenode/ip.140.163.254.158] has quit [Quit: Page closed]22:56
-!- HeikoS1 [~heiko@p4FED52C3.dip0.t-ipconnect.de] has quit [Ping timeout: 250 seconds]23:39
--- Log closed Fri Apr 29 00:00:34 2016
« Wednesday, 2016-04-27 Index Friday, 2016-04-29 »

Generated by irclog2html.py 2.10.0 by Marius Gedminas - find it at mg.pov.lt!